Datasets:

nickypro

/

minipile-split

like 0

[**Isotropic stars in higher-order torsion scalar theories**]{} [**Gamal G.L. Nashed**]{} *Center for Theoretical Physics, British University in Egypt* *Sherouk City 11837, P.O. Box 43, Egypt [^1]* e-mail:[email protected] \ \ \ \ \ Two tetrad spaces reproducing spherically symmetric spacetime are applied to the equations of motion of higher-order torsion theories. Assuming the existence of conformal Killing vector, two isotropic solutions are derived. We show that the first solution is not stable while the second one confirms a stable behavior. We also discuss the construction of the stellar model and show that one of our solution capable of such construction while the other cannot. Finally, we discuss the generalized Tolman-Oppenheimer-Volkoff and show that one of our models has a tendency to equilibrium. Introduction ============ It is well known that in $f(T)$ gravity inflation [@FF] and late time cosmic acceleration can be realized in the early universe [@Le0]–[@CDDS1]. Recently there are many models constructed to describe dark energy without the use of cosmological constant (for more details see review [@BCNO] and references therein). The gravitational field equation of $f(T)$ gravity is second order as general relativity (GR). $f(T)$ gravity suffers from non-invariance of local Lorentz transformation [@LSB]–[@KS], non-minimal coupling of teleparallel gravity to a scalar field [@GLSW]–[@GLSW2] and non-linear causality [@OINC]. Recently, number of $f(T)$ gravitational theories have been proposed [@BF]–[@NBSP]. The structures of neutron and quark stars in $f(T)$ theory have been discussed [@KHR]. The anisotropic behavior, regularity conditions, stability and surface redshift of the compact stars have been checked [@AAZ]. Under those theories it is shown that $f(T)$ are not dynamically identical to teleparallel action plus a scalar field [@Yr11]. It has been shown that investigations of $f(T)$, using observational data, are compatible with observations (see e.g. [@ZH; @FF7] and references therein). A new type of $f(T)$ theory was proposed in order to explain the acceleration phase of the universe [@Yr]. Also it has been shown that the well-known problem of frame dependence and violation of local Lorentz invariance in the formulation of $f(T)$ gravity is a consequence of neglecting the role of spin connection [@KS]. $f(T )$ theory coupled with anisotropic fluid has been examined for static spacetimes with spherical symmetry and many classes of solutions have been derived [@DRH2]. It has been shown that some conditions on the coordinates, energy density and pressures, can produce new classes of anisotropic and isotropic solutions. Some of new black holes and wormholes solutions have been derived by selecting a set of non-diagonal tetrads [@DRH3]. It has been shown that relativistic stars can exist in $f(T)$ [@YLZ14]. A special analytic vacuum spherically symmetric solution with constant torsion scalar, within the framework of $f(T)$, has been derived [@Ngrg]. D-dimensional charged flat horizon solutions has been derived for a specific form of $f(T)$, i.e., $f(T)=T+\alpha T^2$ [@CGSV1]. A complete investigation of the Noether symmetry approach in $f(T )$ gravity at FRW and spherical levels respectively has been investigated [@PBSCADT]. In the framework of $f(T)$ gravitational theories there are many solutions, spherically symmetric [@Nap], spherically symmetric charged [@Nprd], homogenous anisotropic [@RHGR], stability of the Einstein static closed and open universe [@LLG]. Some cosmological features of the $\Lambda$CDM model in the framework of the $f(T)$ are investigated [@SRKHT]. However, till now, no spherically symmetric isotropic solution, [*using non-diagonal tetrad fields*]{}, derived in this theory. It is the aim of the present study to find an analytic, isotropic spherically symmetric solution in higher-order torsion scalar theories. The arrangement of this study are as follows: In Section §2, ingredients of $f(T)$ gravitational theory are provided. In Section §3, two different tetrad spaces having spherical symmetry are applied to the field equations of $f(T)$. Assuming the conformal Killing vector (CKV), we derived two non-vacuum spherically symmetric solutions in §3. The physics relevant to the derived solutions are analyzed in §4. The energy conditions are satisfied for the two solutions provided that the constants of integration be positive. In addition, the stability condition, the nature of the star and Tolman-Oppenheimer-Volkoff (TOV) equation are shown to be satisfied for one solution. The results obtained in this study are discussed in final section. Ingredients of f(T) gravitational theory ======================================== Another description of Einstein’s general relativity (GR) of gravitation is done through the employ of what is called teleparallel equivalent of general relativity (TEGR). The ingredient quantity of this theory is the vierbein (tetrad) fields[^2] $\{h^{a}{_{\mu}}\}$ alternative to metric tensor fields $g_{\mu \nu}$. The associated metric $g_{\mu\nu}=\eta_{ab}h^{a}{_\mu}h^{b}{_\nu}$ with $\eta_{ab}=\textmd{diag}(1,-1,-1,-1)$ being Minkowskian metric, thus Levi-Civita symmetric connection $\mathring{\Gamma}^{\alpha}{_{\mu \nu}}$ is constructed from the metric and its first derivative [@MTW]. Within TEGR, it is possible to build a nonsymmetric connection, Weitzenböck, $\Gamma^{\alpha}{_{\mu\nu}}=h^{a}{_{\mu}}\partial_{\nu}h_{a}{^{\alpha}}=-h_{a}{^{\alpha}}\partial_{\nu}h^{a}{_{\mu}}$. Tetrad space has a main merit that is the null of the vierbein’s derivative, i.e. $\nabla_{_{\nu}}h^{a}{_{\mu}}\equiv 0$, where $\nabla$, regarding the nonsymmetric Weitzenböck connection. Therefore, the vanishing of the vierbein’s covariant derivative recognizes auto-parallelism or absolute parallelism condition. Actually, the $\nabla$ operator is not invariant under local Lorentz transformations (LLT). The metric $g_{\mu \nu}$ is not able to guess one set of vierbein fields; thus extra freedom need to be determined so as to determine unique frame. Because of the absolute parallelism condition, it can be shown that the metricity condition is satisfied. The Weitzenböck connection is curvatureless while it has a non vanishing torsion tensor $T$ given as $$\label{torsion} {T^\lambda}_{ \mu \nu} := {h_a}^{\lambda}(\partial_\mu {h^a}_{\nu}-\partial_\nu {h^a}_{\mu}),$$ and contortion tensor $K$ $$\label{contortion} {K^{\mu \nu}}_\alpha = -\frac{1}{2}\left({T^{\mu \nu}}_\alpha-{T^{\nu \mu}}_\alpha-{T_\alpha}^{\mu \nu}\right).$$ The torsion scalar of TEGR theory is given by $$\label{Tor_sc} \mathrm{T} := {T^\alpha}_{\mu \nu} {S_\alpha}^{\mu \nu},$$ with $S$ defined as $$\label{Stensor} {\mathrm{S}_\alpha}^{\mu \nu} := \frac{1}{2}\left({\mathrm{K}^{\mu\nu}}_\alpha+\mathrm{\delta}^\mu_\alpha{\mathrm{T}^{\beta \nu}}_\beta-\mathrm{\delta}^\nu_\alpha{\mathrm{T}^{\beta \mu}}_\beta\right).$$ Equation (\[Stensor\]) shows skewness in $\mu$ and $\nu$. Like $f(R)$, we could establish Lagrangian of $f(T )$ like $$\begin{aligned} \label{q7} & & {\cal \mathrm{L}}({\mathrm{h}^a}_\mu, \mathrm{\Phi}_A)=\int d^4x h\left[\frac{1}{16\pi}f(\mathrm{T})+{\cal L}_{Matter}(\mathrm{\Phi}_A)\right], \quad \textrm {where} \quad \mathrm{h}=\mathrm{\sqrt{-g}}=det\left({\mathrm{h}^a}_\mu\right), \nonumber\\ & & \mathrm{\Phi}_A \quad \textrm{are} \quad \textrm{ matter} \quad \textrm {fields}.\end{aligned}$$ In this study we postulate the units in which $G = c = 1$. Lagrangian (\[q7\]) can consider as a function of the fields ${h^a}_\mu$. Variation of Lagrangian (\[q7\]) with respect to the tetrad field ${h^a}_\mu$ we obtain the following field equations [@BF; @CGSV1] $$\label{q8} {S_\mu}^{\rho \nu}\, T_{,\rho} \ f(T)_{TT}+\left[h^{-1}{h^a}_\mu\partial_\rho\left(h{h_a}^\alpha {S_\alpha}^{\rho \nu}\right)-{T^\alpha}_{\lambda \mu}{S_\alpha}^{\nu \lambda}\right]f(T)_T-\frac{1}{4}\delta^\nu_\mu f(T)=-4\pi {\mathfrak{T}_{\mu}}^{\nu},$$ where $T_{,\rho}=\displaystyle\frac{\partial T}{\partial x^\rho}$, $\mathrm{f(T)_T}=\displaystyle\frac{\partial \mathrm{f(T)}}{\partial T}$, $\mathrm{f(T)_{TT}}=\displaystyle\frac{\partial^2 \mathrm{f(T)}}{\partial T^2}$ and ${\mathfrak{T}_{\mu}}^{\nu}$ denotes the energy-momentum tensor of the anisotropic fluid which is defined as $$\label{em} {\mathfrak{T}_{\mu}}^{\nu}=(\rho+p_t)u_\mu u^\nu-p_t{\delta_\mu}^\nu+(p_r-p_t)\eta_\mu\eta^\nu,$$ with $p_r$ represents the radial pressure, $p_t$ the tangential pressure and $$u_\mu u^\mu=-\eta_\mu \eta^\mu=1, \qquad {and} \qquad u^\mu \eta_\mu=0.$$ Equations (\[q8\]) are the field equations of$f(T)$ gravitational theory. Non-vacuum spherically symmetric solutions in higher-order torsion scalar theories ================================================================================== In this section, we are going to apply two, non-diagonal, different tetrad fields having spherical symmetry to the field equations (\[q8\]). First tetrad ------------ The equation of motion of GR supply rich field to use symmetries which link geometry and matter in a natural way. Collineations are symmetries which come either from geometrical viewpoint or physical relevant quantities. The importance of collineations is the CKV which provides a more information of the construction of the spacetime geometry. The employs of the CKV simplifies the equations of motions of GR. The CKV is defined as $$\label{kv} {\mathbf L}_\zeta g_{i j}=\zeta_{i;j}+\zeta_{j;i}=\psi g_{i j},$$ with ${\mathbf L}$ being the Lie derivative and the $\psi$ being the conformal factor. One can assume the vector $\zeta$ which creates the conformal symmetry and makes the metric conformally mapped onto itself through $\zeta$. One must note that $\zeta$ and $\psi$ not necessary be static even supposing a static metric [@BHL]. In addition, one must be careful about the following:\ (i) if $\psi = 0$, then (\[kv\]) leads to a Killing vector,\ (ii) if $\psi=constant$, then (\[kv\]) leads to homothetic vector\ (iii) if $\psi =\mathrm{ \psi(x, t)}$ then (\[kv\]) leads to conformal vectors. Furthermore, if $\psi$ is vanishing then the spacetime behaves as asymptotically flat and one has a null Weyl tensor. Thus, to have more understanding of the spacetime geometry one must take into account the CKV. Essentially, the Lie derivative ${\mathbf L}$ shows the inner field of gravity of a stellar configuration related to the vector field $\zeta$. The first tetrad field having a spherical symmetry takes the shape [@Nmpla] $$\label{te1} \begin{tabular}{l} \hspace{-0.3cm} $\left( {h^i}_\mu \right)=$ \\[3pt] $\left( \begin{array}{cccc} \displaystyle\frac{{\cal F}_1(r)}{{\cal F}_2(r)} &{\cal F}_2(r)& 0 & 0\\[9pt] \sin\theta \cos\phi &{\cal F}_1(r)\sin\theta \cos\phi&r\cos\theta \cos\phi & -r\sin\theta \sin\phi \\[9pt] \sin\theta \sin\phi&{\cal F}_1(r)\sin\theta \sin\phi&r\cos\theta \sin\phi & r\sin\theta \cos\phi \\[9pt] \cos\theta & {\cal F}_1(r)\cos\theta&-r\sin\theta & \\[9pt] \end{array} \right)$, \end{tabular}$$ where ${\cal F}_1(r)$, and ${\cal F}_2(r)$ are two unknown functions of the radial coordinate, $r$. The associated metric of (\[te1\]) takes the form $$\label{m1} ds^2=-\frac{{{\cal F}_1}^2-{{\cal F}_2}^2}{{{\cal F}_2}^2}dt^2+({{\cal F}_1}^2-{{\cal F}_2}^2)dr^2+d\Omega, \qquad d\Omega=r^2(d\theta^2+\sin^2\theta d\phi^2),$$ which is a static spherically symmetric spacetime admits one parameter group of conformal motion. Equation (\[m1\]) is conformally mapped onto itself along $\zeta$. Therefore, (\[kv\]) leads to $$\begin{aligned} \label{m2} && 2[{\cal F'}_1{{\cal F}_1}{{\cal F}_2}-{{\cal F}_1}^2{{\cal F'}_2}]\zeta^1=\psi(r)[ {{\cal F}_1}^2{\cal F}_2-{{\cal F}_2}^3], \nonumber\\ && \zeta^0=c, \qquad \qquad \qquad \zeta^1=\frac{\psi(r) r}{2},\nonumber\\ && 2\zeta^1[{\cal F}_1 {\cal F'}_1-{\cal F}_2 {\cal F'}_2]+2\zeta'^1[{{\cal F}_1}^2-{{\cal F}_2}^2]=\psi(r)[{{\cal F}_1}^2-{{\cal F}_2}^2], \end{aligned}$$ where 0 and 1 refer to the temporal and spatial coordinates. Equation (\[m2\]) leads to $$\label{so1} {\cal F}_1=\frac{\sqrt{1+{c_0}^2r^2}{\cal F}_3}{c_0 r}, \qquad {\cal F}_2=\frac{{\cal F}_3}{rc_0}, \qquad \zeta^i=c_1{\delta^i}_0+\frac{\psi(r) r}{2}{\delta^i}_1, \qquad {{\cal F}_1}\neq{{\cal F}_2}, \qquad {{\cal F}_3}=\frac{c_2}{\psi(r)},$$ with $c$, $c_0$, $c_1$ and $c_2$ are constants of integration. Using (\[so1\]), tetrad (\[te1\]) is rewritten as $$\label{te2} \hspace{-0.3cm}\begin{tabular}{l} $\left( {h^i}_\mu \right)=$\\[3pt] $\left( \begin{array}{cccc} \sqrt{1+{c_0}^2r^2} & \displaystyle\frac{{\cal F}_3}{rc_0}& 0 & 0\\[9pt] \sin\theta \cos\phi &\displaystyle\frac{{\cal F}_3\sqrt{1+{c_0}^2r^2}\sin\theta \cos\phi}{rc_0}&r\cos\theta \cos\phi & -r\sin\theta \sin\phi \\[9pt] \sin\theta \sin\phi&\displaystyle\frac{{\cal F}_3\sqrt{1+{c_0}^2r^2}\sin\theta \sin\phi}{rc_0}&r\cos\theta \sin\phi & r\sin\theta \cos\phi \\[9pt] \cos\theta & \displaystyle\frac{{\cal F}_3\sqrt{1+{c_0}^2r^2}\cos\theta}{rc_0}&-r\sin\theta & 0 \\[9pt] \end{array} \right)$, \end{tabular}$$ Tetrad field (\[te2\]) has the following associated metric $$ds^2=-{c_0}^2r^2dt^2+{{\cal F}_3}^2dr^2+d\Omega, \qquad d\Omega=r^2(d\theta^2+\sin^2\theta d\phi^2),$$ Using (\[te2\]) in (\[Tor\_sc\]) we get $$\label{ts} \mathrm{T}=2\frac{2(1+2{c_0}^2r^2){\cal F}_3(r)-rc_0\sqrt{1+{c_0}^2r^2}(3+{{\cal F}_3}^2(r))}{c_0r^3\sqrt{1+{c_0}^2r^2}{{\cal F}_3}^2(r)},$$ Using Eqs. (\[ts\]) and (\[te2\]) in the field equations (\[q8\]) we get the following non-vanishing components: $$\begin{aligned} \label {fe} & & 4\pi {{\cal T}_0}^0=4\pi \rho=-\frac{\sqrt{1+{c_0}^2r^2}{\cal F}_3-rc_0}{r^2c_0{{\cal F}_3}^2}T'f_{TT}\nonumber\\ & &+\frac{rc_0\sqrt{1+{c_0}^2r^2}(2{{\cal F}_3}-r{\cal F'}_3)-{{\cal F}_3}^2(1+2{c_0}^2r^2)}{r^3c_0\sqrt{1+{c_0}^2r^2}{{\cal F}_1}^3}f_{T}+\frac{f}{4},\nonumber\\ & & 4\pi {{\cal T}_1}^0=-\frac{T'f_{TT}}{r^3{c_0}^2},\nonumber\\ & & -4\pi {{\cal T}_1}^1=4\pi p_r=-\frac{{{\cal F}_3}(1+2r^2{c_0}^2)-3rc_0\sqrt{1+{c_0}^2r^2}}{r^3c_0{{\cal F}_1}^2\sqrt{1+{c_0}^2r^2}}f_{T}+\frac{f}{4},\nonumber\\ & & -4\pi {{\cal T}_2}^2=-4\pi {{\cal T}_3}^3=4\pi p_t=-\frac{\sqrt{1+{c_0}^2r^2}{\cal F}_3-2rc_0}{2r^2c_0{{\cal F}_3}^2}T'f_{TT}\nonumber\\ & & +\frac{rc_0\sqrt{1+{c_0}^2r^2}({{\cal F}_3}^3+4{{\cal F}_3}-2r{\cal F'}_3)-2{{\cal F}_3}^2(1+2{c_0}^2r^2)}{2r^3c_0\sqrt{1+{c_0}^2r^2}{{\cal F}_3}^3}f_{T}+\frac{f}{4}.\end{aligned}$$ Second equation of (\[fe\]) leads to $f_{TT}= 0$, or $T=constant$. The case $T=constant$ gives a constant function and this is out the scope of the present study. Therefore, we seeking solutions make constrain on the form of $f(T)$ to have the form $$\label{ft} f(T)=T, \qquad \quad \Rightarrow f_{TT}=0.$$ Assuming the isotropic condition $$\label{co} p_r=p_t=p,$$ and using (\[co\]) in (\[fe\]), we get: $$\begin{aligned} \label{co1} & & {\cal F}_3(r)=\frac{2}{\sqrt{2+4r^2c_3}},\nonumber\\ & & T=-\frac{6c_3c_0r^3\sqrt{1+r^2{c_0}^2}+5rc_0\sqrt{1+r^2{c_0}^2}-2(1+2{c_0}^2r^2)\sqrt{2+4r^2{c_3}}} {r^3c_0\sqrt{1+r^2{c_0}^2}},\nonumber\\ & & 16 \pi \rho=\frac{6c_3r^2-1}{r^2}, \qquad 16 \pi p=\frac{6c_3r^2+1}{r^2}, \qquad \psi(r)=\frac{c_2\sqrt{2+4r^2c_3}}{2}.\end{aligned}$$ The sound velocity ${v_s}^2$ is defined as ${v_s}^2:=\frac{dp}{d\rho}$. Using (\[co1\]) we get the sound velocity in the form $${v_s}^2=-1.$$ Second tetrad ------------- The second tetrad space having a stationary and spherical symmetry takes the form [@Nprd] $$\label{te3} \hspace{-0.3cm}\begin{tabular}{l} $\left( {h^i}_\mu \right)=$\\[5pt] $\left( \begin{array}{cccc} {\cal F}_4(r) &0& 0 & 0\\[9pt] 0 &{\cal F}_5(r)\sin\theta \cos\phi&r\cos\theta \cos\phi & -r\sin\theta \sin\phi \\[9pt] 0&{\cal F}_5(r)\sin\theta \sin\phi&r\cos\theta \sin\phi & r\sin\theta \cos\phi \\[9pt]0 & {\cal F}_5(r)\cos\theta&-r\sin\theta & 0 \\[9pt] \end{array} \right)$, \end{tabular}$$ where ${\cal F}_4(r)$ and ${\cal F}_5(r)$ are unknown functions. Using the same procedure applied to tetrad (\[te1\]) we get the following equations of CKV of tetrad (\[te3\]) $$2{\cal F}'_4\xi^1=\psi(r) {\cal F}_4, \qquad \xi^0=c, \qquad \xi^1=\frac{\psi(r) r}{2}, \qquad 2\xi^1{\cal F}'_5+2{\xi^1}_{,1}{\cal F}_5=\psi(r){\cal F}_5,$$ The above set of equations imply $$\label{so5} {\cal F}_4=c_4 r, \qquad {\cal F}_5=\frac{c_5}{\psi(r)}, \qquad \xi^i=c_6{\delta^i}_0+\frac{\psi(r) r}{2}{\delta^i}_1.$$ where $c_4$, $c_5$ and $c_6$ are constants of integration . Using (\[so5\]), tetrad (\[te3\]) can be rewritten as $$\label{te4} \hspace{-0.3cm}\begin{tabular}{l} $\left( {h^i}_\mu \right)=$\\[5pt] $\left( \begin{array}{cccc} c_4r & 0& 0 & 0\\[9pt] 0 &{\cal F}_5\sin\theta \cos\phi&r\cos\theta \cos\phi & -r\sin\theta \sin\phi \\[9pt] 0&{\cal F}_5\sin\theta \sin\phi&r\cos\theta \sin\phi & r\sin\theta \cos\phi \\[9pt] 0 & {\cal F}_5\cos\theta&-r\sin\theta & 0 \\[9pt] \end{array} \right)$. \end{tabular}$$ Using (\[te4\]), the torsion scalar (\[Tor\_sc\]), takes the form $$\label{ts3} T=\frac{2(3-4{\cal F}_5+{{\cal F}_5}^2)}{r^2{{\cal F}_5}^2}.$$ Inserting (\[ts3\]) and the components of the tensors ${S^{\nu \mu}}_\rho$ and ${T^{\nu \mu}}_\rho$ in the field equations (\[q8\]) we obtain $$\begin{aligned} \label{fe3} & & 4\pi {{\cal T}_0}^0=4\pi\rho=\frac{(1-{{\cal F}_5})}{r{{\cal F}_5}^2}T'f_{TT}-\frac{2{{\cal F}_5}^2-2{\cal F}_5+r{\cal F}'_5}{r^2{{\cal F}_5}^3}f_{T}+\frac{f}{4},\nonumber\\ & &- 4\pi {{\cal T}_1}^1=4\pi p_r=\frac{3-2{\cal F}_5}{r^2{{\cal F}_5}^2}f_{T}+\frac{f}{4},\nonumber\\ & & -8\pi {{\cal T}_2}^2= -8\pi {{\cal T}_3}^3=4\pi p_t=\frac{(2-{{\cal F}_5})}{2r{{\cal F}_5}^2}T'f_{TT}-\frac{2{{\cal F}_5}^2-2{\cal F}_5+r{\cal F}'_5-{{\cal F}_5}^3}{2r^2{{\cal F}_5}^3}f_{T}+\frac{f}{4},\nonumber\\ & &\end{aligned}$$ The above system cannot be solved without assuming some a specific constraint on the form of $f(T)$. Therefore, we are going to use the constraint (\[ft\]) in (\[fe3\]) and obtain the following $$\begin{aligned} \label{so3} & & {\cal F}_5(r)=\frac{2}{\sqrt{2+4r^2c_7}},\nonumber\\ & & T=\frac{5\sqrt{2+4r^2c_7}-8-16r^2c_5+6c_7\sqrt{2+4r^2c_7}}{r^2\sqrt{2+4r^2c_7}},\nonumber\\ & & 16 \pi \rho=\frac{9+18r^2c_7-8\sqrt{2+4r^2c_7}}{r^2}, \qquad 16 \pi p=\frac{11+18r^2c_7-8\sqrt{2+4r^2c_7}}{r^2},\nonumber\\ & & \psi(r)=\frac{c_5\sqrt{2+4r^2c_7}}{2}. {\end{eqnarray}}Using (\ref{so3}), the sound velocity ${v_s}^2$ takes the form \begin{equation} \frac{dp}{d\rho}=\frac{16+32r^4{c_7}^2+48r^2c_7-22r^2c_7\sqrt{2+4r^2c_7}-11\sqrt{2+4r^2c_7}} {(8c_7\sqrt{2+4r^2c_7}-18r^2c_7-9+8\sqrt{2+4r^2c_7})\sqrt{2+4r^2c_7}}.\end{equation} \section{Physics relevant to the models} \underline{\bf Energy conditions:}\vspace{0.2cm}\\ Energy conditions are essential tools to understand cosmology and general results related to strong gravitational fields. These tools are three energy conditions, null energy (NEC), the strong energy (SEC) and weak energy conditions (WEC) \cite{HE}--\cite{ZW}. Such conditions have the following inequalities \begin{eqnarray}\label{fe4} & & NEC: \rho+p_r\geq 0, \qquad \rho+p_t\geq 0,\nonumber\\ & & SEC: \rho+p_r\geq 0, \qquad \qquad \rho+p_r+2p_t\geq 0,\nonumber\\ & & WEC: \rho\geq 0, \qquad \qquad \rho+p_r\geq 0,\qquad \qquad \rho+p_t\geq 0.\end{aligned}$$ The broken of (\[fe4\]) leads to ghost instabilities.\ \ Let us apply the above procedure of the energy conditions given by (\[fe4\]) to the derived solutions given in the previous section. For the case of isotropic, i.e., $p_r=p_t$, we can see from figures 1 and 2: The density has a positive value and the conditions $\rho+p\geq 0 $ $\rho+3p\geq 0 $ are satisfied when the constant $c_3>0$ for the first model and $c_7>0$ for the second model. This means that NEC, SEC and WEC conditions are satisfied for the above two models. Also it is interesting to note that the density and pressure of both solutions do not depend on the constants $c_0$ and $c_4$.\ \ To study the stability issue of the above two models we use the cracking mechanism [@Hl] in which the squared of speed sound must lies in the range \[0, 1\], i.e., $0\leq {v_s}^2 \leq 1$. Figure 3 (a) does not show the positivity criterion i.e., ${v_s}^2\leq0$. However, Fig. 3 (b) satisfies the criterion of stability i.e., ${v_s}^2 \geq0$ within the matter distribution provided any value of the constant $c_7$, in figure 3 (b), and thus second solution preserves stability.\ \ To understand the star behavior we use a plot to indicate the radius of the stellar model for the second solution. Fig. 4, shows the cut on r-axis is approximately 1 km (Fig. 4). This value is a small value and shows a compact star [@BRRC; @DRGR]. The value $R\sim 1$ km produces us to find the surface density of the system. As $r$ approximately vanishing, density approximates $\infty$ and thus, the central density is far from the aim of the present study. Only, we can inspect the surface density by close the values of the Newtonian constant, $G$, and the speed of light, $c$ in the expression of density which gives the numerical value as $13 gm/cm^3$. This is a normal energy density in which the radius $R = 1 km$ is very small. This shows that the second solution of $f(T)$ describes an ultra-compact star [@Rr]–[@HR]. The first model is not a physical one because to find the cutting of the pressure with the r-axis the constant $c_3<0$ which produces a contradiction with the energy conditions. \ \ The TOV equation has the shape $$ds^2=-e^{\nu(r)}dt^2+e^{\lambda(r)}dr^2+r^2(d\theta^2+sin^2\theta d\phi^2),$$ can be written in the form [@BRRC] $$\label{tov} -\frac{M_{G(r)}(\rho+p_r)e^{\frac{\lambda(r)-\nu(r)}{2}}}{r^2}-\frac{dp_r}{dr}+\frac{2(p_t-p_r)}{r}=0,$$ where $M_{G(r)}$ is mass of gravity in a sphere with radius $r$ which has the form $$\label{gm} M_{G(r)}=\frac{r^2 \nu'e^{\frac{\lambda(r)-\nu(r)}{2}}}{2}.$$ Using (\[gm\]) in (\[tov\]), we obtain in the isotropic case $$\label{tov1} -\frac{\nu'(\rho+p_r)}{2}-\frac{dp_r}{dr}=0,$$ Equation (\[tov1\]) demonstrates the equilibrium of the configuration under distinct forces. As an equilibrium condition we write (\[tov1\]) in the form: $$F_g+F_h=0,$$ where $$F_g=-\frac{\nu'(\rho+p_r)}{2}, \qquad \qquad F_h=-\frac{dp_r}{dr}.$$ Using (\[te2\]), (\[co1\]), (\[te4\]) and (\[so3\]) we plot the feature of TOV equation for the above two models in Figure 5. Conclusion and discussion ========================= In this study we have used two non diagonal different tetrad fields having spherical symmetry and reproduce the same associated metric. These tetrads are connected by local Lorentz transformation. We have used the CKV mechanism to reduce the highly nonlinear partial differential equations. We have applied the field equations of $f(T)$ to the first tetrad and have obtained anisotropic system consists of four non linear differential equations. One of these deferential equations put a constraint on the form of $f(T)$. This constraints make the form of $f(T)$ to be $f(T)=T$. Using this form and the isotropic condition, i.e., $p_r=p_t$, we get an isotropic solution. For the second tetrad we have obtained anisotropic system that consists of three non linear differential equations. We cannot solve this system without any constrains on the form of $f(T)$. Using the constraint of $f(T)$ applied to the first tetrad, i.e. $f(T)=T$ and the condition of isotropy, we get another solution. We have studied the physics relevant to each solution and have shown that the first and second tetrads satisfied the energy conditions provided that the two constants of integration involved in these solutions be positive. We have shown that the first tetrad is not stable one because the sound speed is negative, i.e., $\frac{dp}{d\rho}<0$ [@Hl]. However, the second model has confirmed stable manner and has shown a dynamical behavior. We have indicated that the first tetrad is not suitable to construct a stellar model because the radius has an imaginary quantity. In meanwhile the second model has illustrated a stellar model that has a radius about one $Km$ and the density is not a dense on the surface. Finally we have shown that the figures of the widespread TOV equation indicate that static equilibrium has been achieved by distinct forces. Figure 5b show that the second model has a tendency toward equilibrium while the first one did not show such equilibrium. [99]{} R. Ferraro and F. Fiorini, [*Phys. Rev.*]{} [**D 75**]{} (2007). E. V. Linder, [*Phys. Rev.*]{} [**D 81**]{} (2011), 127301. P. Wu and H. W. Yu, [*Eur. Phys. J.*]{} [**C 71**]{} (2011), 1552. K. Bamba, C. Q. Geng, C. C. Lee and L. W. Luo, [*JCAP*]{} [**1101**]{} (2011), 021. K. Bamba, S. D. Odintsov, D. S.-Gomez, [*Phys. Rev.*]{} [**D 88**]{} (2013), 084042. J. B. Dent, S. Dutta and E. N. Saridakis, [*JCAP*]{} [*1101*]{} (2011), 009. K. Bamba, R. Myrzakulov, S. Nojiri and S. D. Odintsov, [*Phys. Rev.*]{} [**D 85**]{} (2012), 104036. A. Aviles, A. Bravetti, S. Capozziello and O. Luongo, [*Phys. Rev.*]{} [**D 87**]{} (2013), 064025. M. Jamil, D. Momeni and R. Myrzakulov, [*Eur. Phys. J.*]{} [**C 72**]{} (2012), 2075. M. Sharif and S. Rani, [*Astrophys. Space Sci.*]{} [**345**]{} (2013), 217. R. Ferraro and F. Fiorini, [*Phys. Lett.*]{} [**B 702**]{} (2011), 75. R. Ferraro and F. Fiorini, [*Int. J. Mod. Phys. Conf. Ser.*]{} [**3**]{} (2011), 227. P. Wu and H. W. Yu, [*Phys. Lett.*]{} [**B 693**]{} (2010), 415. I. G. Salako, M. E. Rodrigues, A. V. Kpadonou, M. J. S. Houndjo and J. Tossa, [*JCAP*]{} [**060**]{} (2013), 1475. Z. Haghani, T. Harko, H. R. Sepangi and S. Shahidi, [*JCAP*]{} [**1210**]{} (2012), 061. Z. Haghani, T. Harko, H. R. Sepangi and S. Shahidi, [*Phys. Rev.*]{} [**D 88**]{} (2013), 044024. T. Shirafuji and G. G. L. Nashed, [*Prog. Theor. Phys.*]{} [**98**]{} (1997), 1355. K. Bamba, S. Nojiri and S. D. Odintsov, [*Phys. Lett.*]{} [**B 725**]{} (2013), 368. K. Bamba, J. de Haro and S. D. Odintsov, [*JCAP*]{} [**1302**]{} (2013), 008. S. Nojiri and S. D. Odintsov, [*Phys. Rept.*]{} [**505**]{} (2011), 59. G. G. L. Nashed, [*Int. J. Mod. Phys.*]{} [**A 21**]{} (2006), 3181. S. Nojiri and S. D. Odintsov, [*eConf*]{} [**C 0602061**]{} (2006), 06. S. Nojiri and S. D. Odintsov, [*Int. J. Geom. Meth. Mod. Phys.*]{} [**4**]{} (2007), 115. S. Capozziello and V. Faraoni, [Beyond Einstein Gravity (Springer, 2010)]{}; S. Capozziello and M. De Laurentis, [*Phys. Rept.*]{} [**509**]{} (2011), 167. A. de la Cruz-Dombriz and D. Saez-Gomez, [*Entropy*]{} [**14**]{} (2012), 1717. S. -H. Chen, J. B. Dent, S. Dutta and E. N. Saridakis, [*Phys. Rev.*]{} [**D 83**]{} (2011), 023508. Y. -P. Wu and C. -Q. Geng, [*JHEP*]{} [**1211**]{} (2012), 142. K. Bamba, S. Capozziello, S. Nojiri, and S. D. Odintsov, [*Astrophysics and Space Science*]{} [**342**]{} (2012), 155. B. Li, T. P. Sotirious and J. D. Barrow, [*Phys. Rev.*]{} [**D 83**]{} (2011), 064035. T. P. Sotirious, B. Li, J. D. Barrow, [*Phys. Rev.*]{} [**D 83**]{} (2011), 104030. M. Kššárk, E. N. Saridakis, [*1510.08432*]{}. C. -Q. Geng, C. -C. Lee, E. N. Saridakis and Y. -P. Wu, [*Phys. Lett.*]{} [**B 704**]{} (2011), 384. C. -Q. Geng, J. -A. Gu and C. -C. Lee, [*Phys. Rev.*]{} [**D 88**]{} (2013), 024030. C. -Q. Geng, C. -C. Lee and E. N. Saridakis, [*JCAP*]{} [**1201**]{} (2012), 002. Y. C. Ong, K. Izumi, J. M. Nester and P. Chen, [*Phys. Rev.*]{} [**D 88**]{} (2013), 024019. G. R. Bengochea and R. Ferraro, [*Phys. Rev.*]{} [**D 79**]{} (2009), 124019. C. Xu, E. N. Saridakis, and G. Leon, [*JCAP*]{} [**1207**]{} (2012), 005. Y.-F. Cai, S.-H. Chen, J. B. Dent, S. Dutta, and E. N. Saridakis, [*Class. Quantum Grav.*]{} [**28**]{} (2011), 215011. A. Einstein, [*Sitzungsber. Preuss. Akad. Wiss. Phys. Math. Kl.*]{}, (1928) 217 (1930) 401. J. Yang, Y.-L. Li, Y. Zhong and Y. Li, [*Phys. Rev.*]{} [**D 85**]{} (2012), 084033. K. Karami and A. Abdolmaleki, [*JCAP*]{} [**04**]{} (2012), 007. G. G. L. Nashed, [*Eur. Phys. J.* ]{} [**C 49**]{} (2007), 851. K. Atazadeh and F. Darabi, [*Eur. Phys. J.*]{} [**C 72**]{} (2012), 2016. H. Wei, X.-J. G. and L.-F. Wang, [*Phys. Lett.*]{} [**B 707**]{} (2012), 298. K. Karami, A. Abdolmaleki, [*Phys. Rev.*]{} [**D 88**]{} (2013), 084034. P. A. Gonzalez, E. N. Saridakis and Y. Vasquez, [*JHEP*]{} [**2012**]{} (2012), 53. S. Capozziello, V. F. Cardone, H. Farajollahi and A. Ravanpak, [*Phys. Rev.*]{} [**D 84**]{} (2011) 043527. R.-X. Miao, M. Li and Y.-G. Miao, [*JCAP*]{} [**2011**]{} (2011), 033. X.-H. Meng and Y.-B. Wang, [*Eur. Phys. J.*]{} [**C 71**]{} (2011), 1755. T. Shirafuji, G. G. L. Nashed and Y. Kobayashi, [*Prog. Theor. Phys.*]{} [**96**]{} (1996), 933. H. Wei, X.-P. Ma and H.-Y. Qi, [*Phys. Lett.*]{} [**B 703**]{} (2011),74. M. Li, R.-X. Miao and Y.-G. Miao, [*JHEP*]{} [**1107**]{} (2011), 108. S. Chattopadhyay and U. Debnath, [*Int. J. Mod. Phys.*]{} [**D 20**]{} (2011), 1135. P. B. Khatua, S. Chakraborty and U. Debnath, [*Int. J. Theor. Phys.* ]{} [**51**]{} (2012), 405. R.-J. Yang, [*Europhys. Lett.*]{} [**93**]{} (2011), 60001. D. Liu and M. J. Reboucas, [*Phys. Rev.*]{} [**D 86**]{} (2012) 083515. K. Atazadeh, M. Mousavi, [*Eur. Phys. J.*]{} [**C 73**]{} (2013), 2272. G. G. L. Nashed, [*Eur. Phys. J.* ]{} [**C 51**]{} (2007), 377. R.-J. Yang, [*Eur. Phys. J.*]{} [**C 71**]{} (2011), 1. R.-J. Yang, [*Eur. Phys. Lett.*]{} [**93**]{} (2011), 60001. S. Nesseris, S. Basilakos, E. Saridakis, and L. Perivolaropoulos, [*Phys. Rev.*]{} [**D 88**]{} (2013), 103010. A. V. Kpadonou, M. J. S. Houndjo, M. E. Rodrigues, [*arXiv:1509.08771*]{}. G. Abbas, A. Kanwal and M. Zubair, [*Astro. Phys. Space Sci.*]{} [**357**]{} (2015), 109. R. Zheng and Q.-G. Huang, [*JCAP*]{} [**1103**]{} (2011), 002. R. Ferraro and F. Fiorini, [*Int. J. Mod. Phys. Conf. Ser.*]{} [*3*]{} (2011), 227. M. H. Daouda, M. E. Rodrigues and M. J. S. Houndjo, [*Eur. Phys. J.*]{} [**C 72**]{} (2012) 1890. M. H. Daouda, M. E. Rodrigues and M. J. S. Houndjo, [*Phys. Lett.*]{} [**B 715**]{} (2012), 241. C. G. Böehmer, A. Mussa and N. Tamanini, [*Class. Quant. Grav.*]{} [**28**]{} (2011), 245020. G. G. L. Nashed, [*Gen. Relat. Grav.*]{} [**45**]{} (2013), 1887. S. Capozzielloa, P. A. Gonzálezc, E. N. Saridakise, Y. Vásquez, [*JHEP*]{} [**1302**]{} (2013), 039. A. Paliathanasis, S. Basilakos, E.N. Saridakis, S. Capozziello, K. Atazadeh, F. Darabi and M. Tsamparlis, [*Phys. Rev.*]{} [**D 89**]{} (2014), 104042. G. G. L. Nashed, [*Astrophysics and Space Science*]{} [**330**]{} (2010), 173. G. G. L. Nashed [*Phys. Rev.*]{} [**D 88**]{} (2013), 104034. M. E. Rodrigues, M. J. S. Houndjo, D. Sáez-Gómez and F. Rahaman, [*Phys. Rev.*]{} [**D 86**]{} (2012), 104059. J.-T. Li, C.-C. Lee and C.-Q. Geng, [*Eur. Phys. J.*]{} [**C73**]{} (2013), 2315. I. G. Salako, M. E. Rodrigues, A. V. Kpadonou. M. J. S. Houndjo and J. Tossa, [*JCAP*]{} [**060**]{} (2013), 1475. C. W. Misner, K. S. Thorne and J. A. Wheeler, [*Gravitation (Freeman, San Francisco, 1973)*]{}. C. G. B$\ddot{o}$hmer, T. Harko and F. S. N. Lobo, [*Phys. Rev.*]{} [**D 76**]{} (2007), 084014; [*Class. Quantum Gravit.*]{} [**25**]{} (2008), 075016. G.G.L. Nashed, [*Gen. Relat. Grav.*]{} [**34**]{} (2002), 1047. S. W. Hawking and G. E. R. Ellis, [*The Large Scale Structure of Spacetime. Cambridge University Press, Cambridge (1973)*]{}. S. Carroll, [*Spacetime and Geometry: An Introduction to General Relativity. Addison-Wesley, Reading (2004)*]{}. M. Zubair and S. Waheed, [*Astrophys. Space Sci.*]{} [**355**]{} (2015), 361. L. Herrera, [*Phys. Lett.*]{} [**A 165**]{} (1992), 206. P. Bhar, F. Rahaman, S. Ray and V. Chatterjee, [*Eur. Phys. J.*]{} [**C 75**]{} (2015), 190. A. Das, F. Rahaman B. K. Guha, and S. Ray, [*Astrophys.Space Sci.*]{} [**358**]{} (2015), 36. R. Ruderman, [*Rev. Astr. Astrophys.*]{} [**10**]{} (1972), 427. N. K. Glendenning, [Compact Stars: Nuclear Physics, Particle Physics and General Relativity (Springer-Verlag, New York, p. 70, 1997)]{}. M. Herjog and F. K. Roepke, [*Phys. Rev.*]{} [**D84**]{} (2011) 083002. [^1]: Mathematics Department, Faculty of Science, Ain Shams University, Cairo, 11566, Egypt\ Egyptian Relativity Group (ERG) URL: http://www.erg.eg.net [^2]: Greek letters $\alpha,\beta,...$ indicate spacetime indices while Latin indices $i,j,...$ run from $0$ to $3$ describe Lorentz indices.

ArXiv

--- abstract: 'We present Very Large Array observations of the 33 GHz radio continuum emission from 22 local ultraluminous and luminous infrared (IR) galaxies (U/LIRGs). These observations have spatial (angular) resolutions of 30–720 pc (0$\,\farcs$07-0$\,\farcs$67) in a part of the spectrum that is likely to be optically thin. This allows us to estimate the size of the energetically dominant regions. We find half-light radii from 30 pc to 1.7 kpc. The 33 GHz flux density correlates well with the IR emission, and we take these sizes as indicative of the size of the region that produces most of the energy. Combining our 33 GHz sizes with unresolved measurements, we estimate the IR luminosity and star formation rate per area, and the molecular gas surface and volume densities. These quantities span a wide range (4 dex) and include some of the highest values measured for any galaxy (e.g., $\mathrm{\Sigma_{SFR}^{33GHz} \leq 10^{4.1}~M_{\odot}~yr^{-1}~kpc^{-2}}$). At least $13$ sources appear Compton thick ($\mathrm{N_{H}^{33GHz} \geq 10^{24}~cm^{-2}}$). Consistent with previous work, contrasting these data with observations of normal disk galaxies suggests a nonlinear and likely multi-valued relation between SFR and molecular gas surface density, though this result depends on the adopted CO-to-H$_{2}$ conversion factor and the assumption that our 33 GHz sizes apply to the gas. 11 sources appear to exceed the luminosity surface density predicted for starbursts supported by radiation pressure and supernovae feedback, however we note the need for more detailed observations of the inner disk structure. U/LIRGs with higher surface brightness exhibit stronger \[[Cii]{}\] 158$\mu$m deficits, consistent with the suggestion that high energy densities drive this phenomenon.' author: - 'L. Barcos-Muñoz' - 'A. K. Leroy' - 'A. S. Evans' - 'J. Condon' - 'G. C. Privon' - 'T. A. Thompson' - 'L. Armus' - 'T. Díaz-Santos' - 'J. M. Mazzarella' - 'D. S. Meier' - 'E. Momjian' - 'E. J. Murphy' - 'J. Ott' - 'D. B. Sanders' - 'E. Schinnerer' - 'S. Stierwalt' - 'J. A. Surace' - 'F. Walter' title: 'A 33 GHz Survey of Local Major Mergers: Estimating the Sizes of the Energetically Dominant Regions from High Resolution Measurements of the Radio Continuum' --- Introduction ============ Luminous and ultraluminous infrared (IR) galaxies (LIRGs: $10^{11}~L_{\odot} \leq L_{IR}$ \[8-1000$\mu$m\] $< 10^{12}~L_{\odot}$, ULIRGs: $L_{IR} \geq 10^{12}~L_{\odot}$) host some of the most extreme environments in the local universe. Local U/LIRGs are primarily triggered by galaxy interactions and mergers . During this process, large amounts of gas are funneled into the central few kpc. There, the gas fuels prodigious star formation and/or AGN activity. This activity is heavily embedded in dust and gas, which reprocesses the emergent light into the IR, giving rise to the high IR luminosities of U/LIRGs. Their enormous gas surface densities, gas volume densities, energy densities, and high star formation rates make the local U/LIRGs crucial laboratories to understand the physics of star formation and feedback in an extreme regime. Indeed, these systems have among the highest SFR and gas surface densities measured for any galaxy population in the local universe . These extreme conditions may lead U/LIRGs to convert gas into stars in a mode that is distinct from what we find in main-sequence galaxies like the Milky Way, and more similar to extreme starbursts observed at high redshift. In this scenario, U/LIRGs and their high redshift counterparts produce a higher rate of star formation per unit gas mass compared to “main sequence galaxies” at both low and high redshift [e.g., @Daddi10; @Genzel10]. The combination of high opacity, high gas surface density, and on-going star formation also makes these galaxies key testbeds for theories exploring the balance between feedback and gravity [e.g., @Murray05; @Shetty11]. For example, @Thompson05 have argued that the most extreme local U/LIRGs may represent “Eddington limited” star-forming systems or “maximal starbursts”, which produce stars at the maximum capacity allowed for the considered feedback mechanism, i.e., radiation pressure on dust. Exploring the physics of U/LIRGs requires knowing their intensive properties, i.e., the luminosity, or mass, per unit area or volume. The extreme nature of these systems is most evident when the high luminosity is viewed in the context of the very small area from which it emerges. In turn, measuring these intensive quantities requires knowing the size of the region where star formation is on-going. This is a challenging measurement. Even the nearest U/LIRGs are quite distant ($50$–$150$ Mpc) compared to prototypes of more quiescent main-sequence galaxies. Thus very high angular resolution is required to study them. Compounding the challenge, U/LIRGs host enormous amounts of dust (e.g., A$_{V}$ $\sim$ 1000 for Arp 220 @Lutz96), rendering them optically thick at optical and even infrared wavelengths. They are also opaque at very long radio wavelengths due to free-free absorption [e.g., @Condon90], leaving them transparent only over a limited regime, from radio to sub-millimeter wavelengths (for the extreme case of Arp 220, see @BM15). Interferometric radio imaging is the ideal, and almost only, way to measure the sizes of the energetically dominant regions in the centers of local U/LIRGs. Radio interferometers make it possible to achieve the high angular resolution required to resolve the compact central starbursts, while cm-wave photons penetrate the high dust columns that prevent measurements of the inner regions at optical wavelengths. The two dominant radio continuum emission mechanisms at cm wavelengths, free-free (“thermal”) and synchrotron (“nonthermal”) emission, both trace the distribution of recent star formation and can indicate AGN activity, if present. Following this logic, @Condon90 and @Condon91 used the old (pre-upgrade) Very Large Array (VLA) to study the energetically dominant regions in U/LIRGs at 1.49 GHz [angular resolution $\geq$1$\,\farcs$5, @Condon90] and 8.44 GHz [angular resolution $\geq0\,\farcs25$, @Condon91]. Their constraints on the sizes of the star-forming/AGN dominated regions in these systems are still some of the strongest measurements twenty five years later. Because the VLA has fixed antenna array configurations, higher frequency observations provide the logical pathway to better angular resolution, and so better size constraints for the local U/LIRGs. However, the spectral index of radio emission from galaxies is negative over the range $\nu \sim 1{-}50$ GHz, so that galaxies are fainter at higher frequencies. The sensitivity of the historic VLA receivers was also lower at high frequency. As a result, efforts to imaging these systems using the historic VLA at $\nu \gtrsim 10$ GHz were limited. With the upgrade from the old VLA to the Karl G. Jansky Very Large Array (VLA), this situation changed. Both the bandwidth and receiver sensitivity improved, thereby improving the ability of the VLA to image the radio continuum from U/LIRGs at high frequency (and thus high angular resolution). Given the current VLA capabilities, the Ka band (26.5 $-$ 40 GHz) offers the ideal balance between low opacity in the source, high angular resolution, and good sensitivity. We demonstrated this capability in @BM15, where we used the VLA at Ka band to make the sharpest-ever image that recovered all of the flux density of the nuclear disks of Arp 220. Here we extend the work of @BM15 to a sample of 22 of the most luminous northern U/LIRGs. This is the first high resolution, high sensitivity, 33 GHz continuum survey of local U/LIRGs. The angular resolution (beam size) of the VLA at $\nu = 33$ GHz improves compared to the $8.44$ GHz of @Condon91 by at least a factor of two. The paper proceeds as follows. In Section \[sec:red\], we describe the survey and the data reduction process. In Section \[sec:results\], we present the measurements. We explore the physical implications of these measurements in Section \[sec:impl\_radio\]. In Section \[sec:discussion\], we discuss the nature of the energy emission at 33 GHz, the implied physical conditions in these systems, the implications of our measurements for star formation scaling relations, and whether the systems are maximal starbursts. We summarize our conclusions in Section \[sec:conclusions\], and the Appendix presents detailed notes on individual systems. Throughout this paper, intrinsic quantities are derived by adopting the cosmology H$_{0}$ = 73 km s$^{-1}$ Mpc$^{-1}$, $\mathrm{\Omega_{vacuum}=0.73}$ and $\mathrm{\Omega_{matter}=0.27}$, with recessional velocities corrected to the frame of the cosmic microwave background. Sample, Observations, and Data Reduction {#sec:red} ======================================== Observations ------------ We used the Karl G. Jansky Very Large Array (VLA) to observe radio continuum emission from the most luminous nearby LIRGs and ULIRGs. Our sample (see Table \[table:tbl-1\]) consists of 22 sources from the IRAS Revised Bright Galaxy Sample [RBGS; @Sanders03]. These galaxies have infrared luminosities $\mathrm{L_{IR}[8-1000 \mu m] = 10^{11.6} - 10^{12.6}~L_{\odot}}$ and were selected to be northern enough to be observed by the VLA, i.e., $\delta > -15^{\circ}$. These systems are also a subset of the Great Observatories All-sky LIRG Survey [GOALS; @Armus09], for which multiwavelength data are available. As part of the resident shared risk project AL746, we observed the radio continuum emission from each source at C band (4–8 GHz) and Ka band (26.5–40 GHz). For each observation we used dual polarization mode with two 1 GHz-wide basebands. Each band was split into eight 128 MHz spectral windows (spw’s) with 64 channels each. We centered the 1 GHz basebands at $\sim 4.7$ and $7.2$ GHz in C band and $\sim 29$ and $36$ GHz in Ka band. In order to recover emission across a wide range of angular scales, we observed our sample in each frequency range in separate sessions using each of the four VLA configurations (A, B, C and D, from highest to lowest angular resolution). Observations spanned the period 2010 August 2 to 2011 August 16. In the D and C configurations, we observed each source for five minutes. In the B configuration, we observed each source for ten minutes split between two five-minute scans. In the A configuration we observed most sources for 20 minutes, split into four five-minute scans. Due to scheduling constraints, eight sources were not observed in the A configuration at Ka band; these are identified with an asterisk in Table \[table:tbl-3\]. Thus the total time on source for most targets was $\sim$ 40 minutes per band. At the beginning of each session, we observed either 3C 48 or 3C 286, which was used to set the flux density scale and calibrate the bandpass. Through the rest of the session we alternated between observations of science targets and a secondary calibrator within a few degrees of each science target. We used observations of these secondary calibrators to measure phase and gain variations due to atmospheric/ionospheric and instrumental fluctuations. Table \[table:tbl-2\] summarizes the calibrators used for each science target. These data have also appeared in @Leroy11 and @BM15. @Leroy11 presented first results from our observations at both C and Ka band but used only observations from the two most compact VLA configurations. @BM15 presented C and Ka band observations using all four configurations for the specific case of Arp 220. In this paper, we report on the full survey, emphasizing the Ka band observations and the combination of all four array configurations. These represent the highest resolution, highest sensitivity radio observations for these galaxies published to date. The C band observations combining all four array configurations will be reported in an upcoming paper focused on the resolved spectral energy distribution, i.e., across the disks of the systems in our sample (Barcos-Muñoz et al. in preparation). Data Processing --------------- We used the Common Astronomy Software Application [CASA, @McM07] to calibrate, inspect, and analyze the data. We followed a standard VLA reduction procedure, including calibrating the bandpass, phase, and amplitude response of each antenna. We set the overall flux density scale using “Perley-Butler 2010" models for the primary calibrators and assuming that the Ka band emission shares the same structure as the VLA-provided Q-band model. Once the data were calibrated, we imaged each science target. To do this, we used the task `CLEAN` in mode `mfs` (multi-frequency synthesis) [@SaW94], with *Briggs* weighting setting `robust=0.5`. For each array configuration, we imaged each baseband independently. Whenever possible, we iterated this imaging with phase and amplitude self calibration. The number of self calibration iterations varied from zero to eight based on the signal to noise of the data, with four iterations typical. After several iterations of phase-only self calibration, when possible, we also performed amplitude self calibration. We always solved for only relative variations in the amplitudes gains among the antennas (`solnorm=True` in CASA’s `gaincal`), and so avoided forcing the flux of the source to some value. After self calibrating the two basebands independently, we combined both into a single image using `clean`’s multifrequency synthesis mode and `nterms=2`. The latter allows us to model the frequency dependence of the sky emission with two Taylor coefficients. After the described combination we ended up with four images per source (one per array configuration). Finally, we jointly imaged all self-calibrated data, combining all eight measurement sets (four configurations and two basebands). This combined image represents our best data product, using all of our observations with sensitivity to a wide range of angular scales. In the highest signal-to-noise cases, for example UGC 08058 (Mrk 231) and UGC 09913 (Arp 220), we performed further self calibration during this final imaging step. These final images have a nominal frequency $\nu = 32.5$ GHz[^1] and a typical rms noise $\mathrm{26~\mu Jy~beam^{-1}}$. Table \[table:tbl-3\] reports the exact beam size and rms noise for the combined image for each target. Additional Data {#sec:add_data} --------------- We combine our survey with previous observations of our sample at 1.49 GHz (beam FWHM $\sim$ 15$\arcsec$) from @Condon90. We also use the 5.95 GHz flux densities (beam FWHM $\sim$ 0.4$\arcsec$) from @Leroy11 and CO (1$-$0) flux densities, obtained using the ARO 12-m antenna (FWHM = 1’), the latter of which will be reported in Privon et al. (in preparation). We present a compilation of the flux densities at these different frequencies, along with 32.5 GHz flux densities measured from our new images, in Table \[table:tbl-3\]. The uncertainty in the 1.49 GHz flux density values are assumed to be dominated by flux density calibration errors ($\sim$ 5%, see Section III in @Condon90). Five of our sources lack flux density measurements at 1.49 GHz. For three of these — VII Zw 031, CGCG 448-020 and IRAS F23365+3608 — we use the 1.4 GHz flux density from the NVSS catalog [@Condon98]. For IRAS 19542+1110 and IRAS 21101+5810, we take the values at 1.425 GHz measured by @Condon96. We use these flux densities interchangeably with the $1.49$ GHz fluxes, but assign them a larger ($10\%$) uncertainty in these cases to reflect some uncertainty in the spectral index. Results {#sec:results} ======= In Figure \[fig:fig1\], contour and color maps show new VLA $\nu = 32.5$ GHz images for our sample of 22 local U/LIRGs. These are the first $33$ GHz images of these systems that have both high resolution and sensitivity to a wide range of angular scales. We use them to measure: (1) the area of the energetically dominant region in each galaxy, (2) the integrated flux density of each target at 33 GHz, and (3) the contribution (by area and flux density) of compact regions to the integrated properties of each system. In Tables \[table:tbl-3\] and \[table:tbl-4\] we report the measured areas and integrated flux densities at 33 GHz, along with the integrated flux densities from the literature that we use to study the spectral index, and so the nature of the radio emission.   Flux Densities at $\nu = 32.5$ GHz {#sec:Iflux} ---------------------------------- We measure integrated flux densities for each source from the lowest angular resolution observations, which were obtained using the VLA in its D configuration. The maximum recoverable scale for the D configuration, $\approx 22''$, corresponds to $\sim$ 16 kpc at the 165 Mpc median distance of our sample. This would cover most of the star forming activity in a local disk galaxy [e.g., @Schruba11]. U/LIRGs are observed to be much more compact, with sizes less than a few kpc based on previous radio [@Liu15], near-IR [@Haan11], mid-IR [@DS10], and far-IR observations [@Lutz16]. Therefore, we expect negligible missing flux in the D configuration-based flux densities. Confirming this expectation, most of our targets appear unresolved in the D configuration-only images, which have beam sizes $\approx2\,\farcs7$. We obtained the flux densities reported in Table \[table:tbl-3\] using CASA task `imfit` to fit two dimensional Gaussians to these mostly unresolved point sources. A few targets, including NGC 3690, CGCG 448-020, IRAS 17132+5313, VII Zw 031, VV 250, VV 340, and VV 705, showed some extent or multiple components in the D configuration maps. In most of these cases, we tapered the D configuration data to a lower resolution until the morphology became a single point-like source. Then we fit a 2D-Gaussian to this degraded image. NGC 3690 and VV 250 show well separated components that can only be fit using two Gaussians, even in the tapered images. We report the sum of these two components as the integrated flux density. The uncertainties that we report sum (in quadrature) the statistical error calculated by `imfit` with uncertainty in the calibration of the flux scale, which we estimated to be $\sim$ 12% in @BM15. For the two faintest galaxies in our sample, UGC 04881 and IRAS 08572+3915, the signal-to-noise ratio of the D configuration data only was not high enough to recover integrated flux densities. For these two systems, we instead report results from the combined data using all configurations, which we tapered until we recovered point-like structures that could be fit using Gaussians. Spectral Indices Involving $\nu = 32.5$ GHz {#sec:alpha} ------------------------------------------- In addition to our new $32.5$ GHz flux densities, Table \[table:tbl-3\] reports literature flux densities for our sources at $\nu =1.49$ and $5.95$ GHz. We combine these with the $\nu=32.5$ GHz measurements to calculate the galaxy-integrated spectral index between 1.49 GHz and 5.95 GHz ($\alpha_{1.5-6}$) and between 5.95 GHz and 32.5 GHz ($\alpha_{6-33}$). Here, we define the spectral index, $\alpha$, by $\mathrm{S_{\nu} \propto \nu^{\alpha}}$. Note that because we use the flux density integrated over the whole galaxy, we do not expect the different angular resolutions at different frequencies to affect these calculations. In Figure \[fig:fig2\], we show the derived spectral indices. We plot $\alpha_{1.5-6}$ as a function of $\alpha_{6-33}$. Here the solid line shows equal spectral indices for both pairs of bands, which we would expect if a single spectral index holds across the entire radio regime (from 1.5 to 33 GHz). Dashed lines show $\alpha = -0.8$, a typical spectral index for synchrotron emission without any opacity effects [e.g., @Condon92]. {width="3.5in"} {width="3.5in"} {width="3.5in"} Size of the Radio Emission {#sec:sizes} -------------------------- A main goal of our study is to measure the extent of the radio continuum emission in our targets with the purpose of constraining the size of the energetically dominant region. To do this, we analyzed the final images combining data from all the array configurations. These high resolution images are sensitive to the brightest compact cores, but they have lower surface brightness sensitivity than the D configuration data that we used to determine the total flux density. Therefore, they may miss extended, low surface brightness emission. To take this into account, we measure the size of the energetically dominant region from the half-light area (A$_{50}$). This is the area enclosed by the highest intensity isophote that includes half of the total integrated flux density of the system, which we measured from the lower resolution data above and expect to be complete. Note that this approach measures the observed A$_{50}$, which reflects the true size of the source convolved with the synthesized beam of the array. We require the intensity of the isophote enclosing the half-light area, or C$_{50}$, to be at least $5$ times the rms noise in the image. If C$_{50}$ would be less than 5$\sigma$ in the combined image, we interpret this to indicate an important component of extended, low surface brightness emission. In order to recover this emission, we measure A$_{50}$ for these systems from lower resolution versions of the data that have better surface brightness sensitivity. In these cases, we first tried using `natural` weighting instead of `Briggs` (see Section \[sec:red\]). If we still could not recover half of the light within a S/N$>5$ contour, then we produced progressively lower resolution images by applying larger and larger $u-v$-tapers to the data. We stepped the size of the taper by 0$\,\farcs$2 and used `Briggs` weighting schemes with `robust` parameter 0.5 at each step. In this way, we measure A$_{50}$ from the highest resolution image where C$_{50}$ can be reliably measured, i.e., where C$_{50}$ $\geq$ 5$\sigma$. The following systems showed extended, low surface brightness emission and required $u-v$-tapering: CGCG 436-030, CGCG 448-020, IRAS 21101+5810, IRAS 17132+5313, VV 340a and VV 705. For NGC 3690, the natural weighting approach was sufficient. Once we identified a reliable half-light contour, C$_{50}$, we calculated A$_{50}$ by multiplying the number of pixels within the C$_{50}$ contour by the pixel area. Figure \[fig:fig1\] shows the images that were used to measure A$_{50}$ and the C$_{50}$ contour (in red) for each source. {width="3.0in"} {width="3.0in"} {width="3.0in"} {width="3.0in"} Many of our sources show sizes close to that of the synthesized beam. We show this in Figure \[fig:fig3\]. There, we plot the ratio of the observed A$_{50}$ to the beam area, $\mathrm{A_{beam}}$, as a function of the beam area in units of arcsec$^{2}$ (top left panel) and kpc$^{2}$ (top right panel). The quantity of physical interest is the true size of the 33 GHz emission with the beam deconvolved, $\mathrm{A_{50d}}$. In the top, and bottom left panels of Figure \[fig:fig3\], a dashed line indicates a value of $\mathrm{2\times A_{beam}}$, which we consider a practical threshold for the emission to be viewed as resolved. Here $\mathrm{A_{beam}=\frac{\pi \theta_{\rm maj} \theta_{\rm min}}{4}}$ with $\mathrm{\theta_{\rm maj}~and~\theta_{\rm min}}$ the FWHM of the synthesized beam along its major and minor axis. In this definition, $A_{beam}$ refers to the area expected to enclose half the total power in the beam. This definition is consistent with our measured area $A_{50}$, and appropriate for deconvolution. We treat the sources that show extent larger than the beam but size smaller than $2\times A_{beam}$ as marginally resolved (region between the solid and dashed lines in Figure \[fig:fig3\]). In these cases, we assume that the intrinsic shape (deconvolved size) of the source follows a Gaussian distribution. We then estimate the deconvolved size of the source by $\mathrm{A_{50}(deconvolved)=A_{50}(observed)-A_{beam} \equiv A_{50d}}$, equivalent to deconvolving the FWHM in quadrature. In the top panels of Figure \[fig:fig3\], two sources lie below the solid line, indicating an observed size smaller than the beam. These are IRAS F08572+3915 and UGC 04881NE. Although statistical fluctuations could produce this situation, the signal-to-noise of the data appear to be too high for this explanation to hold. The most likely culprit is a calibration issue when combining observations using the different array configurations. We adopt a conservative upper limit of $\mathrm{A_{50d} < A_{beam}}$ for these two systems[^2]. In order to determine the best estimate of A$_{50d}$ for “resolved” sources with $\mathrm{A_{50} > 2\times A_{beam}}$, we inspected the shape of the C$_{50}$ contour (red in Figure \[fig:fig1\]) to determine if the source exhibits a Gaussian shape. If it did, then we apply the same approach used for the marginally resolved sources to each component and summed the results to find the total A$_{50d}$. This tended to be the case when more than one component is present, such as VV 705 and CGCG 448-020. If C$_{50}$ showed a more complex morphology, our simple Gaussian treatment becomes invalid. In these cases we instead assume that the measured, not deconvolved $A_{50}$ represents an upper limit to the true size. This is true for the following galaxies: IRAS 19542+1110, IRAS F23365+3604, UGC 08387, VII Zw 031, VV 250a and VV 340a. For two sources, UGC 04881 and VV250, a second, faint component could be recovered only in the low resolution map used to assess the integrated flux density. In both cases, the individual components are unresolved in this integrated map. Here, we had to lower our conservative limit of 5$\sigma$ in order to recover the half-light area. In these two systems, we measure C$_{50}$ from a contour with $S/N \approx 3$ and treat the size estimate as an upper limit (see Table \[table:tbl-4\]). For NGC 3690 and IRAS 17132+5313, one component of the C$_{50}$ contour shows a Gaussian distribution while others show more complex morphology. In both cases, we performed the deconvolution on the Gaussian components. Then we have a partially deconvolved estimate, $\mathrm{A_{50d} < A_{50} (observed)}$, which is still an upper limit because of the un-deconvolved more complex structure. We report values for A$_{50d}$ and C$_{50}$ in Table \[table:tbl-4\], along with an equivalent R$_{50d}$ value where $\mathrm{A_{50d}=\pi R_{50d}^{2}}$. We caution, however, that $R_{50d}$ is only a representative number reflective of the upper limit to the area in these cases. In Table \[table:tbl-4\], we also report the degree of Gaussianity, defined as the ratio between the flux density level of the C$_{50}$ contour and the peak flux density. For a two dimensional Gaussian, this value is 0.5. Compact Sources Decomposition {#sec:comp} ----------------------------- In addition to the integrated flux density and a characteristic size, we measured the contribution of compact sources to the overall flux density of each target using the maps of Figure \[fig:fig1\]. For our purposes, compact sources are those that clearly belong to the system and show a Gaussian morphology. For each target, we identify these sources by eye and fit them using `imfit`, providing estimates of the rms noise and reasonable starting guesses for the sizes, and peak intensity and position. The locations of the fit compact sources appear as white crosses in Figure \[fig:fig1\]. Their sizes, which are often comparable to the size of the beam, are shown in the bottom left panel of Figure \[fig:fig3\]. We also calculated the flux density that is originating from all the compact sources in a system, and compared it to the integrated flux density (see top panels in Figure \[fig:fig4\]). We note that such comparisons may be affected by the different physical resolutions achieved from the observations, however we find no trend relating the fraction of flux in compact sources to beam physical area. In the bottom panel of Figure \[fig:fig4\] we show instead the contribution of each point source – especially important when more than one is present – to the integrated flux density at 33 GHz. We identified compact sources in each of our targets except the north-east component in IRAS F17132+5313, which shows mostly extended emission. For the cases of the faint components in the systems UGC 04881 and VV250, the Gaussian fit was performed on the low resolution image that was used to obtain the integrated flux density of the system. A subset of our sources show most of their emission concentrated into a very small area, consistent with a point source producing much of the flux density even at our highest angular resolution. To make the strongest possible measurement of the compactness of these targets, we used our highest resolution images. This is usually the A configuration image ($\sim0\,\farcs1$), except in those cases with B as the longest baseline array configuration observed ($\sim0\,\farcs2$; see Table \[table:tbl-3\]). From this highest resolution image, we measured the flux density detected at S/N$\geq5$, which corresponds well to the total flux density in the compact core of the image. We compared this flux density in the bright core at the highest resolution to the integrated flux density of the system, $f_{A~(or~B)}$. Most of the U/LIRGs in this sample show single bright point sources in the highest resolution image, although a few, including NGC 3690, UGC 08387, Arp 220, and VV 705, show more than one compact core. We also measured the size of the 33 GHz emission showing significant detection, as set by the 5$\sigma_{A~(or~B)}$[^3] contour, at this highest angular resolution image. We report the beam size of the A, or B, array configuration images along with the sizes of the 5$\sigma_{A~(or~B)}$ contour and $f_{A~(or B)}$ of each system in Table \[table:tbl-5\]. We highlight those sources with most of their emission being contributed by a single bright compact source, being good potential AGN candidates. These include: IRAS F01364-1042, III Zw 035, and IRAS 15250+3609. Arp 220 should also be on this list as it shows $f_{A~(or~B)}>50\%$, however we refer the reader to a more exhaustive discussion on the morphology of its 33 GHz emission presented in @BM15. There are six other sources with $f_{A~(or~B)}>50\%$, but unfortunately the highest resolution achieved was only $\sim0\,\farcs2$ and the constraint on their compactness is then weaker. However, note that Mrk 231, a known AGN [e.g., @Ulvestad99; @Lonsdale03], belongs to that group. In Table \[table:tbl-5\] we also note two systems, VII Zw 031 and VV 340a, with $f_{A~(or B)}<1\%$ indicating most of their emission at $0\,\farcs1$ resolution is filtered out, and then is mostly extended in nature. {width="3.0in"} {width="3.0in"} {width="3.0in"} Implications of the Radio Sizes {#sec:impl_radio} =============================== From the 33 GHz images, we either measure or strongly constrain the size of the energetically dominant regions in our targets. Radio interferometers are almost unique in their ability to peer through heavy dust extinction while also achieving very high angular resolution. As a result, similar sizes are difficult to obtain at other wavelengths. Here, we assume that the energetically dominant region traced by the radio data has approximately the same size as the region bearing the mass or emitting the light at other wavelengths. This allows the calculation of intensive (per unit area or volume) quantities. Our method to do this, in general, is to assume that half of the flux at some other wavelength of interest (e.g., 1.4 GHz, IR\[8–1000$\mu$m\] and CO emission) is enclosed in the $33$ GHz half-light area, $A_{\rm 50,d}$. We then calculate the surface brightness and related parameters (surface density, volume density) implied by this assumption. Note that in several cases, we expect optical depth to play a key role (e.g., at 1.4 GHz or in the IR). In this case, the $\tau \approx 1$ photosphere may lie outside the calculated size [e.g., see @BM15]. In other cases, our assumption that the radio structure indicates the structure at other wavelengths may break down (e.g., if an AGN contributes significant IR but weak radio emission or if gas traced by CO decouples from star formation). We discuss these cases in the individual sections and report the derived values in Table \[table:tbl-6\]. Brightness Temperatures {#sec:Tb} ----------------------- For a resolved or nearly resolved source, where beam filling is a minor consideration, the brightness temperature, $T_{b}$, offers the prospect to constrain the emission mechanism and opacity of the source [e.g., @Condon91]. At radio frequencies, the brightness temperature, T$_{\rm b}$, follows the Rayleigh-Jeans approximation where $$\label{eq:Tb} T_{b} = \Big(\frac{S_{\nu}}{\Omega_{source}}\Big)\frac{c^{2}}{2 k_{B} \nu^{2}}~,$$ with $S_{\nu}$ the flux density at frequency $\nu$ and $\Omega_{source}$ the area of the source. Most of our targets are resolved. Thus an “averaged nuclear T$_{\rm b}$” at 32.5 GHz can be derived using $\Omega_{source}$ = A$_{\rm 50d}$ and S$_{\nu}$ = 0.5 $\times$ S$_{32.5}$ (see above for the explanation of the aperture correction). We also calculate T$_{\rm b}$ from the point of highest intensity in the highest resolution image for each target, peak T$_{\rm b}$, where $\Omega_{source} = \Omega_{beam}$ in that case. Figure \[fig:fig5\] shows histograms of these peak and averaged nuclear T$_{\rm b}$ at $\nu = 32.5$ GHz. {width="3.5in"} The averaged nuclear $T_{\rm b}$ for our targets is typically a few 10s of Kelvin to a few times 100 K, reaching up to a few thousand Kelvin in the brightest targets. For only free-free emission filling the beam, we would expect $T_{\rm b}$ for optically thick emission to approach $T_{\rm e}$ for the [Hii]{} regions. For physical conditions like those present in our sample, the expected electron temperature, T$_{\rm e} \sim 5000{-}10^{4}$ K . In metal rich environments, such as the central regions of ULIRGs [@Veilleux09], the cooling is more efficient and T$_{\rm e}$ may tend towards the low end of this range, $\sim 5000$ K [e.g., @Puxley89], though note that @Ana00 found T$_{\rm e}$ of 7500 K for Arp 220 from integrated measurements of radio recombination lines. In Figure \[fig:fig5\] we observe T$_{\rm b}$ does not exceed either 10$^{4}$ K or 5000 K for any galaxy. In theory the unresolved, or marginally resolved, sources could be optically thick and highly clumped at scales much smaller than the beam size. However, both the observed spectral index (which would be positive with $\alpha \sim 2$ for the free-free emission if optically thick) and the relative smoothness of the images argue against such a scenario. Instead, low opacity at $32.5$ GHz appears to be the natural explanation for the $T_{\rm b}$ that we observe. In Figure \[fig:fig5\] we observe the peak brightness temperatures do not exceed the likely $T_{\rm e}$. However, T$_{\rm b}$ (peak) should be treated as a lower limit for the unresolved and marginally resolved sources. Are these sources likely to be optically thick? Excluding the case of Mrk 231 since it hosts an AGN, the lower limits for the peak T$_{\rm b}$ go from 20 K up to 690 K, with the unresolved case, UGC 04881, having a temperature of 22 K. In the marginally resolved cases, we can gain insight into the likely size of the source by contrasting the peak and average T$_{\rm b}$. Figure \[fig:fig3\] shows that for most of these marginally resolved with T$_{b}$(peak) $<$ T$_{b}$(average), we would expect to be able to resolve them with a beam area that is 2 times smaller at most. This would imply a true $T_{\rm b}$ peak $\sim 2$ times larger than what we measure, still not enough for these sources to reach the optically thick regime. In these marginally resolved cases, in particular, the substructure of the emission remains unclear. Our data offer limited insight into whether the data may be structured into smaller optically thick regions beneath the beam. For the unresolved source, the situation is less clear. With the size unconstrained, the source could be optically thick at 33 GHz and heavily beam diluted. However, we note again that the spectral index that we observe does not appear consistent with optically thick free-free emission. We proceed assuming that we observe optically thin 33 GHz emission for this source. The flux densities of many of our targets have been measured at $1.4$ GHz (Table \[table:tbl-3\]), but even in its most extended configuration, the VLA reaches only $\approx 1\arcsec$ resolution at this frequency. Using the measured $1.4$ GHz flux densities, we calculate the averaged nuclear T$_{\rm b}$ at $1.4$ GHz assuming that the $32.5$ GHz sizes also describe the true extent of the $1.4$ GHz emission. These span 10$^{3}$ up to 10$^{6.5}$ K. These are high values. Values of T$_{\rm b}>5\times10^{3}$ or 10$^{4}$ K, imply that the emission at 1.49 GHz is mostly synchrotron in nature, because the source function of the free-free emitting ionized gas is a black body at $T \sim 5\times10^3$ - 10$^{4}$ K, as explained above. Dominant synchrotron emission may be expected at $1.4$ GHz, but the values that we find may in fact be too high for the standard mixture of free-free and synchrotron emission seen in starburst galaxies. Considering such a mixture, @Condon91 suggested a maximum T$_{\rm b}$ for a starburst of $\sim$10$^{4.6}$ K at 1.49 GHz (their Equation 9, using $\mathrm{T_{e}\sim5000~K}$). At least 12 sources in the sample show T$_{\rm b~1.4GHz}>10^{4.6}$ K when we combine the 33 GHz sizes and the 1.49 GHz flux densities (see Table \[table:tbl-6\]). This could imply that the 1.49 GHz emission from these sources includes a significant AGN contribution. One of those sources, Mrk 231, is well known to be dominated by an AGN, which explains why it has the highest predicted averaged nuclear T$_{\rm b}$ at 1.49 GHz. Based on this line of argument, for these high brightness $1.49$ GHz sources, we would expect much of the flux density to be confined to an unresolved core in VLA $1.49$ GHz imaging. In Mrk 231, most of the emission is unresolved at 1.49 GHz, however, other sources show resolved emission at 1.49 GHz. In these cases, the 33 GHz sizes, which are small compared to the $\sim 1.5''$ VLA beam, may not be representative of the true $1.49$ GHz emission. Indeed, we might worry that the 33 GHz size will underestimate the size at 1.49 GHz if the system is optically thick at these lower frequencies. In such case, the emission will emerge from a photosphere larger than the emitting (optically thin) region at $33$ GHz and the true brightness temperature at $1.49$ GHz will be lower than our estimate. Another alternative is that an extended synchrotron component may contribute to the integrated flux density. This component would have to have a spectrum steep enough that it does not contribute much to the flux density at $33$ GHz, implying substantial variations in the resolved spectral index. On the other hand, several sources have high $T_{b}$ and remain barely resolved even at 8.44 GHz [see maps in @Condon91]: IRAS 08572+3915, IRAS 17132+5313, IRAS 15250+3609 and III Zw 035. These are our best AGN candidates based on $T_{b}$ arguments. Here extra information is needed to determine whether they are powered by an AGN and/or starbursts. In an upcoming paper, we investigate this possibility by combining the current observations with the lower frequency ($\nu = 4{-}8$ GHz) part of our survey (L. Barcos-Muñoz et al. in preparation). Star Formation Rate and IR Surface Density {#sec:Ssfr} ------------------------------------------ Infrared luminosity, L$_{\rm IR}[8-1000\mu m]$, and radio emission both trace recent star formation in starburst galaxies. IR luminosity reflects reprocessed light from young stars, while the 33 GHz continuum predominately captures a mix of synchrotron and thermal emission, both of which originate indirectly from young stars. Considering a mix of synchrotron and thermal emission, @Murphy12 relate the recent star formation rate to the 33 GHz luminosity, $L_{\rm 33GHz}$, via $$\label{eq-sfrrad} \begin{split} \left(\frac{\rm SFR_{\nu}}{M_{\sun}\,{\rm yr^{-1}}}\right) &= 10^{-27} \left[2.18 \left(\frac{T_{\rm e}}{10^{4}\,{\rm K}}\right)^{0.45} \left(\frac{\nu}{\rm GHz}\right)^{-0.1}\right. + \\ %&\left.1300 \left(\frac{q_{\rm SNR}}{\rm 0.0116~yr^{-1}}\right) \left(\frac{\nu}{\rm GHz}\right)^{-\alpha^{\rm NT}}\right]^{-1}\\ &\left.15.1 \left(\frac{\nu}{\rm GHz}\right)^{\alpha^{\rm NT}}\right]^{-1} \left(\frac{L_{\nu}}{\rm erg~s^{-1}~Hz^{-1}}\right). \end{split}$$ where $T_{e}$ is the electron temperature and $\alpha_{\rm NT}$ is the nonthermal spectral index. @Murphy12 relate the infrared luminosity to the recent star formation rate via $$\label{eq:sfrir} \mathrm{\left(SFR_{IR} \over {\rm M_\odot~yr^{-1}} \right) = 3.15\times 10^{-44} \left ( L_{IR}[8-1000\mu m] \over {\rm ergs~s^{-1}} \right)}~.$$ In the left panel of Figure \[fig:fig6\] we compare IR-based and 33 GHz based SFRs estimated for each U/LIRG in our sample. Following @Murphy12, we adopt T$_{\rm e}=10^4$ K and $\alpha_{\rm NT} = -0.8$ at $\nu=32.5$ GHz, but note both as a source of uncertainty. If we use T$_{\rm e}$ = 5000 K, the SFR based on 33 GHz increases by $\sim$37%. The left panel of Figure \[fig:fig6\] shows that these two simple SFR estimates agree in our sample. The strong outlier, source \#10, is Mrk 231. This system is known to be dominated by an AGN that appears to contribute substantially to the 33 GHz emission. The other sources are consistent with a simple radio-infrared correlation that has a normalization in agreement with the @Murphy12 relations. {width="3.5in"} {width="3.5in"} If the assumption is made that the 33 GHz size, A$_{\rm 50,d}$, reflects the distribution of star formation, we can derive a star formation rate surface density, $\Sigma_{\rm SFR_{IR}^{33GHz}}$. As above, we take $\mathrm{\Sigma^{33GHz}_{SFR_{IR}} = 0.5\times SFR_{IR}/A_{50,d}}$[^4]. The right panel in Figure \[fig:fig6\] shows our calculated $\Sigma_{\rm SFR_{IR}}^{\rm 33GHz}$. These span from $\mathrm{10^{0.6}}$ up to $\mathrm{10^{4.1}~M_{\odot}~yr^{-1}~kpc^{-2}}$ (right panel, bottom axis, in Figure \[fig:fig6\]). The high end of this range represents the highest $\Sigma_{\rm SFR}$ found for any galaxy in the local universe. The wide range indicates diverse conditions. Even though we have observed the brightest and closest U/LIRGs, these span about four orders of magnitude in $\Sigma_{\rm SFR_{IR}}^{\rm 33GHz}$. The IR surface brightness is also of interest. In local U/LIRGs, most of the bolometric luminosity is emitted in the 8$-$1000 $\mu$m range. By assuming that half of L$_{\rm IR}$ is concentrated within A$_{50,d}$, we estimate $\mathrm{\Sigma^{33GHz}_{IR}}$ for this inner region. For our approach from Equation \[eq:sfrir\], $\mathrm{\Sigma^{33GHz}_{\rm IR}}$ is identical to $\mathrm{\Sigma^{33GHz}_{\rm SFR_{IR}}}$ within a constant factor. Therefore we show the $\mathrm{\Sigma^{33GHz}_{\rm IR}}$ axis along the top of the right panel of Figure \[fig:fig6\]. The U/LIRGs in this sample have $\mathrm{\Sigma^{33GHz}_{\rm IR}}$ ranging from $\mathrm{10^{10.5}~to~10^{14.1}~L_{\odot}~kpc^{-2}}$. The high end of this range is of particular interest. The dashed vertical line in Figure \[fig:fig6\] indicates $\mathrm{\Sigma^{33GHz}_{IR}=10^{13}~L_{\odot}~kpc^{-2}}$. This value of $\mathrm{\Sigma^{33GHz}_{\rm IR}}$ has been argued to correspond to the characteristic Eddington limit set by radiation pressure on dust in self-regulated, optically thick disks [@Thompson05]. Some sources in our sample show $\mathrm{\Sigma^{33GHz}_{IR} \geq 10^{13}~L_{\odot}~kpc^{-2}}$, indicating they may be Eddington-limited starbursts (see Section \[sec:edd\] for further discussion). Note that for systems that are optically thick in the infrared, the $\tau_{\rm IR} \sim 1$ photosphere may be larger than the 33 GHz size. In this case, the $\mathrm{\Sigma^{33GHz}_{IR}}$ that we calculate would never be observed, even if very high resolution FIR observations were available. This does not mean that this quantity lacks physical meaning, however. These systems are incredibly opaque to UV and optical light, which we expect to be generated in the region of active star formation traced by our 33 GHz data. This will then be quickly reprocessed into IR light, which then scatters out to the photosphere before leaving the system. In this case, that inner region captured by the 33 GHz emission and $\mathrm{\Sigma^{33GHz}_{\rm IR}}$, and $\mathrm{\Sigma^{33GHz}_{SFR_{IR}}}$ are the quantities directly related to the region of most intense feedback and the immediate sites of star formation. Are our sources optically thick in the IR? Infrared observations are limited to relatively coarse resolution, so direct size measurements in this range provide only modest constraints. @DS10 and @Lutz16 measured infrared sizes sizes[^5] for the systems in our sample at 13 $\mu$m ([*Spitzer*]{}) and 70 $\mu$m ([*Herschel*]{})[^6]. For most systems, the sizes at 13 $\mu$m are larger than those at 70 $\mu$m, and the latter are larger than those we measured at 33 GHz. A more powerful constraint comes from comparing our measured size to that implied by the measured dust temperature and luminosity. To do this, we consider the emission emitted in the IR, specifically between 8 and 1000 $\mu$m, and the dust temperature found comparing 63 $\mu$m and 158 $\mu$m emission (for more details, see Diaz-Santos et al. submitted). For an optically thick black body of temperature $T_{dust}$, $$0.5 \times L_{IR}[80-1000\mu m] = 4 \pi R_{50IR}^{2}\sigma T^{4}_{dust}~,$$ where $\mathrm{L_{IR}}$ is shown in Table \[table:tbl-1\]. $L_{\rm IR}$ and $T_{dust}$ are measured, and this approach allows for the size expected for a photosphere with $T_{\rm dust}$ to produce $L_{\rm IR}$. In Figure \[fig:fig7a\], we compare the sizes measured at 13 $\mu$m, 70 $\mu$m, and 33 GHz, and those calculated for a black body (assuming $\mathrm{A^{IR}_{50}=\pi R^{2}_{50IR}}$). We see that at least half of the sources in our sample are optically thick at infrared wavelengths, with our measured 33 GHz size smaller than the blackbody size. Thus IR opacity appears significant in our sample, which might be expected considering we are studying the most obscured systems in the local universe. {width="3.5in"} {width="3.3in"} Gas Surface and Column Density {#sec:Smol} ------------------------------ Our sample consists of gas-rich mergers. In these systems, large masses of gas are funneled to the center, where they become mostly molecular [e.g., @Larson16]. The surface and volume densities of this gas relate closely to its self-gravity and ability to form stars. Again, we assume that the 33 GHz size is characteristic of the system and by combining this with half of the integrated CO (1$-$0) measurements, we estimate these quantities for the sample. Both the assumption of the 33 GHz characteristic size and the conversion between CO luminosity and mass (“conversion factor”) introduce uncertainties into the calculation. Our calculation assumes that the gas shares a characteristic size with the star formation traced by the radio. If our targets harbor large amounts of non-star forming gas or the internal relationship between gas and star formation is strongly non-linear, e.g., with stars forming much faster in a subset of very dense gas, the calculation will yield biased results. We do expect the approximation to hold, at least to first order. On larger scales star formation traced by IR and CO emission do track one another approximately one-to-one in major mergers [@Daddi10]. More, interferometric CO measurements find that nearly half of the total CO mass is enclosed in the central few kpc in local U/LIRGs . For a starburst $\mathrm{\alpha_{\rm CO} = 0.8~M_{\odot}~(K~km~s^{-1} pc^{2})^{-1}}$ (including helium) and coexisting gas and radio emission, we infer values for the molecular gas surface density, $\Sigma_{\rm mol}^{\rm 33GHz}$, from $\mathrm{10^{2.5}~to~10^{5.3}~M_{\odot}~pc^{-2}}$. Even the low end of this range corresponds to source-averaged surface densities in excess of many Local Group molecular clouds [e.g., @Bolatto08; @Fukui10]. The high end is far in excess of $\sim$ 1 g cm$^{-2}$, which is commonly invoked as an immediate precondition for star formation considering dense substructure inside molecular clouds. Here this gas column density is the average value across the whole energetically dominant area of a galaxy. These values obviously depend on the mass-to-light ratio adopted to convert CO luminosity to mass. The appropriate conversion factor for starburst galaxies has been a matter of debate, with suggestions ranging from approximately Galactic [e.g., @Pap12; @Scoville14] to low and highly environment-dependent [@Shetty11] values. To see the effect of a higher, Milky Way, $\alpha_{\rm CO}$ one should multiply our nominal surface and volume densities by $\approx 5.4$. Also note, that this assumption of matched $\Sigma_{\rm gas}$ and $\Sigma_{\rm SFR}$ distributions does not hold for some ULIRGs. For example, for IRAS 13120-5453 the measured starburst size derived from sub-mm continuum is found to be more compact than the emission from dense [@Privon17] and more diffuse [@Sliwa17] molecular gas tracers. More, recent high resolution observations of the CO emission in Arp 220 [@Scoville17] suggest that the gas is distributed in a larger area compared to the star formation area traced by the 33 GHz emission (see Figure \[fig:fig1\] and @BM15). Only $\sim 20\%$ of the total CO emission is coming from the nuclei. At the moment, Arp 220 is uniquely well-studied. These results argue that high resolution interferometric observations of the gas to match our SFR-tracing continuum will yield important information on how the SFR-per-unit gas varies across the system. Lacking such information, we proceed assuming matched gas and SFR. If these ULIRGs represent the general case, the reader may think of our $\Sigma_{\rm gas}$ as an upper limit, with 10s of percent of the material in an extended, comparatively non-star forming disk. This will imply even higher SFR per unit gas mass in the nuclear regions than we calculate below. One class of models considers the total mass surface density a main driver of the conversion factor, largely via its effect on the line width [e.g., @Shetty11; @Narayanan12]. Our measured sizes give us the opportunity to illustrate the effect of such a dependence on derived surface densities. To do this, we use the prescription in @Bolatto13 [their equation 31] which follows @Shetty11. Neglecting any metallicity dependence and considering only the regime where $\Sigma > 100$ M$_\odot$ pc$^{-2}$, their prescription is $$\label{eq:alpha_co} \left( {\alpha_{\rm CO} \over {{\rm M}_\odot \over [{\rm K~km~s^{-1}~pc^2]}}} \right) \approx 4.35 \left( \frac{\Sigma_{\rm total}}{100~\mathrm{M_\odot~pc}^{-2}} \right)^{-0.5},$$ where $\Sigma_{\rm total}$ is the total mass surface density driving the potential well. We will assume the systems studied here to be gas-dominated in the main CO-emitting region and take $\Sigma_{\rm total} \sim \Sigma_{\rm gas}$. The overall gas mass fraction in local U/LIRGs is closer to $\sim 30\%$ [@Larson16]. However we expect the gas to be concentrated relative to the stars, so that we can assume the systems to be locally gas dominated in the emitting region. We assume that in the dense, well shielded central regions of U/LIRGs, the HI content is negligible, and we consider $\Sigma_{\rm gas} \sim \Sigma_{\rm mol}$. We calculate the conversion factor from equation \[eq:alpha\_co\] iteratively, because $\Sigma_{\rm mol}$ changes as $\alpha_{\rm CO}$ changes. Numerically iterating, we reach a value of $\alpha_{\rm CO}$ that converges to within 0.1%. These values go from 0.2 up to 1.65, with a median value of 0.43 for our sample. We report the gas properties derived using this surface-density dependent $\alpha_{\rm CO}$ in brackets, along with $\alpha_{\rm CO}$ for each source, in Table \[table:tbl-6\]. The effect of applying this correction is to narrow the range of derived gas surface densities, as the high surface density systems have low $\alpha_{\rm CO}$. The gas surface density values derived here translate to average Hydrogen column densities that range from $\mathrm{10^{22.6}~cm^{-2}~to~10^{25.4}~cm^{-2}}$ when using $\mathrm{\alpha_{CO}=0.8}$, and $\mathrm{10^{22.9}~cm^{-2}~to~10^{24.8}~cm^{-2}}$ when using the surface-density dependent conversion factor. Assuming a Galactic dust-to-gas ratio [@Bohlin78], which may be roughly appropriate [@Rupke08; @Iono09], these column densities imply line of sight extinctions of $A_V \sim 22$ to $12,000$ mag, for a starburst conversion factor, and $A_V \sim$ 48 to $3,200$ mag, for a surface-density dependent conversion factor. Gas Volume Density ------------------ The gas volume density, and the corresponding free fall time, are central quantities for many theories of star formation [e.g., @Krumholz12]. We estimate the gas volume density from the measured sizes and the integrated CO luminosities. This requires additional geometric assumptions. We consider the most basic approach and assume that our sources are three dimensional Gaussians. In this case, $\sim 30$% of the mass exists inside the FWHM of the Gaussian[^7], R$_{50,d}$. Adopting this geometry, we find n$_{\rm H_{2}}$ from $\mathrm{10^{0.5}~cm^{-3}~to~10^{4.9}~cm^{-3}}$ for a fixed starburst $\alpha_{\rm CO}$. Using the variable, surface-density dependent $\alpha_{\rm CO}$, we find a narrower range of $\mathrm{10^{0.9}~cm^{-3}~to~10^{4.3}~cm^{-3}}$. The free fall collapse times associated with these densities range from $6{-}100$ ($2{-}190$) Myr with the fixed (variable) $\alpha_{CO}$. Discussion {#sec:discussion} ========== The 33 GHz sizes reported in this paper represent the best measurements to date of the energetically dominant regions in this set of bright, nearby U/LIRGs. These sizes, combined with the integrated flux density measurements allow us to study the physical properties of the nuclear regions in the sample. Here, we discuss the implications of these measurements for the nature of the 33 GHz emission, star formation scaling relations, optical depth, and radiation pressure feedback. Nature of the $33$ GHz Radio Emission {#sec:nature} ------------------------------------- In models like those of @Condon92 and @Murphy12, the radio SED reflects a mixture of thermal and nonthermal emission. What powers the emission that we observe from U/LIRGs at 33 GHz? In the @Condon92 model for a starburst galaxy like M82, about 50% of the total 33 GHz continuum is produced by free-free (“thermal”) emission; for comparison, $<12\%$ of the emission is expected to be produced by free-free emission at 1.5 GHz. We have several constraints on the nature of the emission mechanism in our targets: the SED shape, the brightness temperature, and the comparison with the SFR implied by the IR. Together, these indicate some 10s of percent contribution of thermal emission to the 33 GHz flux density, with the balance being synchrotron. However, a detailed understanding of the emission mechanism will need to wait for better coverage of the radio SED in these targets (L. Barcos-Muñoz et al. in preparation). [*Brightness Temperature and Optical Depth:*]{} The brightness temperature of optically thick free-free emission is expected to be $\sim 5\times10^3$ K - 10$^{4}$ K. If the 33 GHz $T_b$ exceeded this value, this would provide evidence that synchrotron dominates the emission. Figure \[fig:fig5\] shows that the averaged nuclear T$_{\rm b}$ does not exceed this limit. Either the emission is patchy within the beam, or the emission at $33$ GHz is optically thin. Thus, the brightness temperature in the sources allows for a normal mix of emission mechanisms and is consistent with optically thin free-free emission making up a large part (or all) of the observed 33 GHz flux density. If we neglect filling factor effects and assume that $\approx 50\%$ of the total T$_{\rm b}$ is due to thermal emission, then we can estimate the optical depth of the free-free emission. We derive $\mathrm{\tau_{thermal} \sim T_{b}/T_{e} \leq 0.2}$ for all our sample. This number is still less than 1, therefore optically thin, even if we assume 100% of the 33 GHz flux density is due to thermal emission. [*Spectral Index:*]{} For a mixture of synchrotron (“non-thermal”) emission and optically thin free-free emission, give the following approximation to the fraction of emission that is thermal, $$\label{eq:thermal} \frac{S}{S_{\rm T}} \sim 1+10 \left( \frac{\nu}{\rm GHz} \right)^{0.1+\alpha_{\rm NT}}~.$$ Here $S$ is the total flux density, S$_{\rm T}$ is the flux density from thermal emission, and $\alpha_{\rm NT}$ is the typical non-thermal spectral index $\sim -0.8$. The formula assumes a power-law spectral energy distribution for the non-thermal emission. We combine equation \[eq:thermal\] with the S$_{5.95}$ from Table \[table:tbl-3\] to calculate $S/S_T$ at $5.95$ GHz. Then, knowing that $\mathrm{S_{T} \propto \nu^{-0.1}}$ we predict the spectral index between $\alpha_{6-33}$. Based on this, we expect an average $\alpha_{6-33} = -0.53$. We expect $\alpha_{6-33}$ to approach $\alpha_{\rm NT} = -0.8$ as the thermal fraction decreases to zero, while if the thermal fraction is higher than this estimate, $\alpha_{6-33}$ will be $> -0.53$. Figure \[fig:fig2\] shows that 17 out of the 22 systems in our sample have $\alpha_{6-33}<-0.53$, implying that in most of our sample, non-thermal emission is stronger relative to thermal emission than predicted by Equation \[eq:thermal\]. @BM15 found a similar result comparing 6 and 33 GHz emission in Arp 220. We caution that our assumed $\alpha_{\rm NT}$ affects this result and that we cannot, at present, distinguish between variations in the thermal fraction and $\alpha_{\rm NT}$ from only two frequencies. Indeed, multi-frequency observations, particularly at high frequency, suggest curvature in the radio SED [e.g., see @Clemens08; @Clemens10; @Leroy11; @Marvil15] so that the power-law assumption for the non-thermal emission model in Equation \[eq:thermal\] represents a simplification. Observations that cover a wide band will allow for a more complex treatment for a better disentanglement of the contribution of the two components at these frequencies (Linden et al. in prep). [*Spectral Index and Implied Opacity at Lower Frequencies:*]{} Following the same approach, we use Equation \[eq:thermal\] and the flux at $5.95$ GHz to predict an integrated $\alpha_{1.5-6}$ of $-0.71$. However, less than half of the sample show spectral indices that agree with this predicted value. Most of our targets show shallower spectral indices. This is most likely due to opacity affecting the low frequency emission, especially the observations at 1.49 GHz where free-free absorption is known to play a major role in compact starbursts [see, e.g., @Condon91; @Murphy13a]. In fact, in Figure \[fig:fig2\] we also observe a change in slope as frequency increases for several sources, from shallower to steeper in most cases. Mrk 231 even shows a change from positive $\alpha_{1.5-6}$ to a negative $\alpha_{6-33}$. For a compact starburst this would indicate that $\tau_{\rm thermal}$ becomes one at some frequency between 1.5 and 33 GHz[^8]. However, we know Mrk 231 has a very compact core [e.g., @Lonsdale03; @Helmboldt07], which suggest instead the change in slope is most likely due to synchrotron self-absorption at low frequencies. In addition, it is also possible that the flattening in the observed $\alpha_{1.5-6}$ could be caused by ionization and bremsstrahlung losses [@Thompson06; @Lacki10], which become important at low frequencies in high density environments such as those found in our sample (see Section \[sec:Smol\]). {width="3.5in"} Several systems show the opposite trend, exhibiting steep $\alpha_{1.5-6}$ and a shallower $\alpha_{6-33}$. The simplest explanation for these measurements is that these systems have a higher thermal fraction than the other targets. Alternatively, some other source may contribute to the 33 GHz emission, e.g., anomalous dust emission [@Draine98; @Ali09; @Murphy10]. More detailed SED coverage could confirm this interpretation. Another possible explanation includes contribution from thermal dust, which is normally important only at much higher frequencies, $\gtrsim$ 100 GHz. Again, better frequency coverage will play a key role. In Figure \[fig:fig2\], we find a tentative correlation between $\mathrm{\alpha_{1.5-6GHz}}$ and $\mathrm{A_{50,d}}$, showing a shallower spectral index for more compact sources. This trend makes sense if more compact sources are also more opaque. In this case, $1.5$ GHz emission in more opaque systems will be suppressed due to a higher opacity at $1.5$ GHz relative to $6$ GHz. Integrated spectral indices only give us a partial view of the processes that are powering star formation in our sample. We require more detailed spectral index maps to dissect the distribution of the radio emission. We will report resolved spectral index maps between 6 and 33 GHz in a future paper (Barcos-Munõz et al. in prep). These results will be greatly complemented by spectral indices maps between 1.49 and 8.44 GHz reported in @Vardoulaki15 using the @Condon90 and @Condon91 observations. [*Expectations from IR-Based SFRs:*]{} The contrast of the 33 GHz flux density with the total IR emission also sheds some light on the emission mechanism. Inasmuch as the IR tells us about the star formation rate, it also makes a prediction for the expected thermal emission, along with some simplifying assumptions. We derive the expected free-free emission, S$_{\rm T}$, and then thermal fraction, S$_{\rm T}$/S, at 33 GHz by assuming that all the IR luminosity is due to star formation and that none of the ionizing photons (that will potentially produce free-free emission) are absorbed by dust. Note that if an AGN is present and contributes significantly to the IR luminosity, then the SFR derived by this method will be overestimated [see @Armus07; @Petric11 for an estimation of the AGN contribution to L$_{IR}$ in local U/LIRGs]. We use equation \[eq:sfrir\] and the thermal SFR from Table 8 in @Murphy12, which relates SFR and the thermal luminosity, L$^{\rm T}$, by the following equation, $$\label{eq-sfrt} \begin{split} \left(\frac{\rm SFR_{\nu}^{T}}{M_{\sun}\,{\rm yr^{-1}}}\right) &= 4.6\times10^{-28}\\ &\left(\frac{T_{\rm e}}{10^{4}\,{\rm K}}\right)^{-0.45} \left(\frac{\nu}{\rm GHz}\right)^{0.1} \left(\frac{L_{\nu}^{\rm T}}{\rm erg~s^{-1}~Hz^{-1}}\right), \end{split}$$ where we assume T$_{\rm e}\sim10^{4}$ K (see Section \[sec:Tb\] and \[sec:Ssfr\] for further discussion on this assumption). In this way, we predict the thermal radio emission expected given the IR luminosity. Comparing it to L$_{\rm IR}$, we derive the thermal fractions shown in the top right panel in Figure \[fig:fig2\]. We see no clear trend, however note that T$_{\rm e}$ is uncertain, and the derived thermal fractions depend on it. Lower values of T$_{\rm e}$, or higher thermal optical depths, imply lower thermal fractions. We also observe that Mrk 231 shows the lowest predicted thermal fraction in our sample. This is expected since it does not follow the radio-IR correlation (see Figure \[fig:fig6\]), with SFR$_{\rm 33GHz}$ being $\sim$ 4 times higher than SFR$_{\rm IR}$. By comparing the thermal fractions shown in Figure \[fig:fig2\] with the radio-IR correlation shown in Figure \[fig:fig6\], we see that all 11 sources with low thermal fraction (i.e., thermal fractions $< 60\%$) are below the equality line in Figure \[fig:fig6\]. This is consistent with Equation \[eq-sfrrad\] underestimating the SFR due to a more dominant non-thermal component (i.e., a plausible shallower $\alpha_{\rm NT}$) than what is assumed for the equation (-0.8). From our analysis of the spectral index, we expect thermal fractions $\leq$ 50%. Figure \[fig:fig2\] shows that, based on the prediction from the IR, most of the sources have thermal fractions $\approx 50$–$100\%$. We expect that this is the combination of three effects. First, even if the IR is all powered by star formation, some of the ionizing photons produced by young stars that could otherwise produce free-free emission will be absorbed dust and thus not produce free-free emission. These should not be counted in our prediction for the thermal emission, and the true thermal fraction would be accordingly smaller. We highlighted a similar situation in Arp 220, where the predicted thermal fraction is $\sim$ 50% but SED analysis shows it should be closer to 35% [see @BM15]. Second, as noted above, the SED-based estimates remain hampered by the lack of sensitive, wide-band coverage of the spectral energy distribution. As long as the adopted non-thermal spectral index (or SED) remains uncertain, so will do the thermal fractions estimated in this way. Third, if an AGN contributes a substantial amount to the IR emission, then the thermal fraction would be overestimated because the AGN will not contribute to the free-free emission in the same way as star formation. Two sources, UGC 04881NE and IRAS 08572+3915 show thermal fractions $>$ 100%, meaning that they have very high ratios of IR to radio emission (see Figure \[fig:fig6\]). This IR excess has been reported before for IRAS 08572+3915 [see discussion in @Yun04], and this system was already noted as an interesting source in discussion of first results from this survey [@Leroy11]. See the Appendix for further discussion on these two sources. [*Radio-FIR Correlation at 33 GHz:*]{} As a more observational restatement of the previous result, we derive $q_{33}$, the ratio of FIR flux (between $\sim$ 42 and $\sim$ 122 $\mu$m) to radio flux density at 33 GHz: $$\label{eq:q} \mathrm{q_{\rm \nu} = Log_{10}((S_{FIR}/3.75\times10^{12}~Hz)/S_{\nu})}~.$$ Here, $S_{\nu}$ is the flux density at frequency $\nu$ in units of $\mathrm{W~m^{-2}~Hz^{-1}}$, and $\mathrm{S_{\rm FIR}[42-122\mu m] = 1.26\times10^{-14}~(2.58~S_{60\mu m}+ S_{100 \mu m})}$, in units of $\mathrm{W~m^{-2}}$, is the far-infrared flux, with the flux density at 60 and 100 $\mu$m measured in Jy. We show a histogram of $q_{32.5GHz}$ in Figure \[fig:fig7\]. We find a median $q_{33}\approx3.32$ and a dispersion of $0.19$ dex. $q_{33}$ is similar to that found by @Rabidoux14 studying regions in local star forming galaxies, but we find a tighter correlation. Their measured dispersion is $0.1$ dex larger than ours. In fact, the 0.19 dex in dispersion we observed for q$_{33}$ is similar to that found in @Condon91 at 1.49 GHz. The tighter dispersion found for our global measurements compared to the local ones of @Rabidoux14 appears to corroborate the global nature of the IR-radio correlation. Note as well that $q_{32.5GHz}$ does not appear to correlate with $\Sigma_{SFR}$. Physical Conditions at the Heart of Local Major Mergers ------------------------------------------------------- Our size estimates imply that a large part of the star forming activity, and so presumably also the gas that fuels it, is concentrated in areas with half-light radii from 30 pc up to 1.7 kpc[^9]. Applying these sizes to global quantities using the proper aperture corrections, we estimate $\Sigma_{\rm SFR}$, $\Sigma_{\rm IR}$, N$_{\rm H}$, and $\mathrm{n_{mol}}$. The resulting values span a wide range, typically $4$ dex. The high end of the range for each property is among the highest average gas, SFR, or luminosity surface density measured for any galaxy. The low end of the range is still high compared to values found in “normal” disk galaxies: the lowest density systems have $\mathrm{\Sigma_{mol}^{33GHz} \sim 10^2}$–$10^3$ M$_\odot$ pc$^{-2}$ and $\mathrm{\Sigma_{SFR_{IR}}^{33GHz} \sim 10^0}$–$10^1$ M$_\odot$ yr$^{-1}$ kpc$^{-2}$. These already resemble the highest kpc-resolution values (which come from active galaxy centers) found in @Leroy13 (see bottom panel in Figure \[fig:fig9\]). Moreover, the gas surface densities in our sample, even the lowest values, resemble those found for individual molecular clouds, but here they extend over the whole energetically dominant region of a galaxy. This implies average interstellar gas pressures that match or exceed those found inside individual clouds. Because of this high pressure, a Milky Way GMC dropped into any of the targets would not remain an isolated, self-gravitating object. Self-gravitating, overpressured clouds in these targets must be more extreme and denser than clouds in normal galaxies, a conjecture born out by observations of nearby starburst galaxies [e.g., @Keto05; @Wei12; @Leroy15; @Johnson15]. About half (13) of the 22 targets studied here show galaxy-averaged $\mathrm{\Sigma_{SFR_{IR}}^{33GHz} \geq 10^{2.7}~M_{\odot} yr^{-1} kpc^{2}}$. This corresponds to $\geq$ 2 times higher than the $\Sigma_{\rm SFR}$ that would be inferred based on the IR emission from the Orion core [@Soifer00]. Several (7) sources show $\mathrm{\Sigma_{SFR_{IR}}^{33GHz} >10^{3}~M_{\odot} yr^{-1} kpc^{-2}}$, corresponding to $\mathrm{\Sigma_{IR}^{33GHz}>10^{13}~L_{\odot} kpc^{-2}}$. This value has been put forward as the characteristic Eddington limit for $\Sigma_{\rm SFR}$ in a radiation pressure-supported, optically thick disk [@Scoville03; @Thompson05] (see Section \[sec:edd\] for further discussion). The high column densities obscure the energetically dominant regions at non-radio wavelengths. Assuming a “starburst" conversion factor, $13$ U/LIRGs show hydrogen column densities consistent with being Compton-thick, $\mathrm{N_{H}>1.5\times10^{24} cm^{-2}}$ [e.g., @Comastri04], which would directly affect the ability of X-ray diagnostics to detect the presence of AGN in these systems. As mentioned above, the implied optical extinctions are extreme, 22$-$12,000 mag for our sample assuming a Galactic dust-to-gas ratio. Even infrared wavelengths, at which a normal star-forming galaxy is usually optically thin, will show significant opacity for these dust columns. At 100 $\mu$m, for a mass absorption coefficient of $\mathrm{\kappa_{100} = 31.3~cm^2~g^{-1}}$ , the dust opacity of these targets is $\tau_{100} \sim $0.02$-$12, with those same 13, but one, Compton-thick sources also being optically thick at 100 $\mu$m, i.e., $\tau_{100}>1$. The [\[]{}[Cii]{}[\]]{} Deficit ------------------------------- Several studies have reported a “deficit” in the \[C II\] 158$\mu$m-to-far infrared luminosity (from 40 to 120 $\mu$m) ratio, $L_{\rm [\sc C II]}/L_{\rm FIR}$, in U/LIRGs relative to lower luminosity star-forming galaxies [e.g., @Malhotra01; @DS13; @Lutz16]. The $L_{\rm [C II]}/L_{\rm FIR}$ decreases with increasing dust temperature, mid-IR opacity, star formation efficiency ($L_{\rm IR} / M_{\rm H_2}$) and infrared surface density (where [*Spitzer*]{} and [*Herschel*]{} data are utilized to measure sizes). The deficit arises because the collisional energy required to produce \[[Cii]{}\] is suppressed in the compact, dense starburst environments of U/LIRGs, and/or because the infrared luminosity is increased. The sizes used to gauge the IR surface brightness in @DS13 come from IR space telescopes, which have much coarser angular resolution than our maps. In Figure \[fig:fig9\] (top left panel), we plot $L_{\rm [C II]}/L_{\rm FIR}$ from @DS13 as a function of the star formation rate surface density inferred using our sizes, $\Sigma^{\rm 33 GHz}_{\rm SFR_{\rm IR}}$. The plot shows clear, strong anti-correlation between $L_{\rm [C II]}/L_{\rm FIR}$ and $\mathrm{\Sigma^{33GHz}_{SFR_{IR}}}$. The top right panel in Figure \[fig:fig8\] shows $L_{\rm [C II]}/L_{\rm FIR}$ as a function of $\mathrm{A_{50d}}$. Both plots show that more compact systems with more locally intense star formation show stronger $L_{\rm [C II]}/L_{\rm FIR}$ deficits (lower $L_{\rm [C II]}/L_{\rm FIR}$). This is strong corroboration, using the best size measurements to date, of the correlation found by @DS13 of higher deficit for systems with higher luminosity densities. {width="3.5in"} {width="3.5in"} {width="3.5in"} The spectral index between 1.5 and 6 GHz may give some indication of the opacity at low frequencies. In the bottom panel of Figure \[fig:fig8\], we plot $L_{\rm [C II]}/L_{\rm FIR}$ as a function of this spectral index, $\alpha _{\rm 1.5-6GHz}$. $L_{\rm [C II]}/L_{\rm FIR}$ is lower, and thus the \[[Cii]{}\] deficit is larger, for systems with flatter (more nearly $0$) spectral indices. This flattening is believed to be due to increasing opacity [e.g., see @Murphy13a], so the bottom panel of Figure \[fig:fig8\] shows that the $L_{\rm [C II]}/L_{\rm FIR}$ ratio is lowest for U/LIRGs that are most obscured at radio, as well as infrared, wavelengths. With the exception of IRAS F08572+3915, the five U/LIRGs (Mrk 231, IRAS 15250+3608, III Zw 035, IRAS F01364-1042, and Arp 220) with the flattest $\alpha _{\rm 1.5-6GHz}$, and among the largest \[[Cii]{}\] deficit, also have the lowest estimated thermal fraction at 33 GHz in our sample. These results are broadly consistent with our detailed study of Arp 220 [@BM15], where we presented evidence of suppressed 33 GHz thermal emission and speculated that the suppression is due to the absorption of ionizing UV photons by dust concentrated within the HII regions [see also @Luhman03; @Fischer14]. Such scenario would also imply a lack of heating of photodissociated regions (PDR) and thus a suppression of the amount of collisional energy available to produce \[[Cii]{}\]. Implications for Star Formation Scaling Relations {#sec:relations} ------------------------------------------------- {width="3.5in"} {width="3.5in"} {width="3.5in"} The observed scaling between star formation rate surface density, $\Sigma_{\rm SFR}$, and gas surface density, $\Sigma_{\rm gas}$, is often used as a main diagnostic of the physics of star formation in galaxies [e.g., @Kennicutt98]. @Kennicutt98 fit a scaling between galaxy-averaged $\Sigma_{\rm SFR}$ and $\Sigma_{\rm gas}$ that describes both normal disk galaxies and starbursts. The starbursts in @Kennicutt98 have high $\Sigma_{\rm SFR}$ and $\Sigma_{\rm gas}$ and include U/LIRGs like those studied here. The contrast between the normal disks (low $\Sigma_{\rm SFR}$, $\Sigma_{\rm gas}$) and the starburst galaxies (high $\Sigma_{\rm SFR}$, $\Sigma_{\rm gas}$) played a main role in driving the best overall fit of @Kennicutt98, $\Sigma_{\rm SFR} \sim \Sigma_{\rm gas}^{1.4}$. This contrast depends on the sizes adopted for the starburst galaxies. Changing the size affects both surface densities by the same factor, but because the overall relationship between $\Sigma_{\rm SFR}$ and $\Sigma_{\rm gas}$ is non-linear, the adopted size affects the slope. In Figure \[fig:fig9\] we place each of our targets in the $\Sigma_{\rm SFR_{IR}}^{\rm 33GHz}$-$\Sigma_{\rm gas}^{\rm 33GHz}$ (or $\Sigma_{\rm SFR_{IR}}^{\rm 33GHz}$-$\Sigma_{\rm mol}^{\rm 33GHz}$) plane (see Section \[sec:Ssfr\] and \[sec:Smol\] for details on the derivation of $\Sigma_{\rm SFR_{IR}}^{\rm 33GHz}$ and $\Sigma_{\rm mol}^{\rm 33GHz}$). In the top left panel, we show only the U/LIRGs from our sample and adopt a fixed $\alpha_{\rm CO}$ = 0.8 M$_\odot$ pc$^{-2}$ (K km s$^{-1}$). These U/LIRGs show high surface densities and an approximately linear relationship. A non-linear least-squares fit[^10] yields $$\label{eqn:ks} \mathrm{log_{10}(\Sigma_{SFR_{IR}}^{33GHz})} = (1.02 \pm 0.10)~\mathrm{log_{10}(\Sigma_{mol}^{33GHz})}-(1.33 \pm 0.47)~.$$ This slope is in good agreement with the results found by @Liu15 for disk galaxies and for U/LIRGs. @Genzel10 also noted that the internal relationship for starburst galaxies was more nearly linear than the relationship using both types of galaxies, giving rise to the idea of “two sequences” of star formation. A similar conclusion of “two sequences” of star formation is also derived by @Daddi10, although they obtained a steeper slope ($\sim$1.4) for each type of galaxies that approaches unity within the uncertainty of their measurements. With a slope close to unity, another way to express Equation \[eqn:ks\] is that for a “starburst” conversion factor, we find a typical gas depletion time, $\tau_{\rm dep} \equiv M_{\rm mol}/{\rm SFR}$, of $\tau_{\rm dep} \sim 20$ Myr for the targets studied here. Note that this short timescale would potentially lead to a relatively flat-spectrum radio source inconsistent with the observed FIR/radio correlation (see section \[sec:nature\]), however the uncertainty in the calculated $\tau_{\rm dep}$ is at least a factor of a few. In addition to the size, the adopted conversion factor can have a large effect on the results. Because we find an approximately linear relationship within our sample, shifting from one constant $\alpha_{\rm CO}$ to another will not affect the slope. For example, if we use a Galactic $\alpha_{\rm CO}$ = 4.35 M$_\odot$ pc$^{-2}$ (K km s$^{-1}$) instead, the coefficient would shift to -2.08$\pm$0.55, raising the depletion time to $\tau_{\rm dep} = 125$ Myr. For comparison, @Leroy13 find a significantly longer $\tau_{\rm dep}$, $\sim 1.6$ Gyr, in the disks of nearby normal galaxies. Several suggestions posit a continuous variation in $\alpha_{\rm CO}$ that depends on surface density (see Equation \[eq:alpha\_co\]). Adopting such prescription affects the slope of the derived relation. If we adopt the surface density-dependent slope discussed in Section \[sec:Smol\], the best fit shifts to $$\label{eqn:ks2} \mathrm{log_{10}(\Sigma_{SFR_{IR}}^{33GHz})} = (1.52 \pm 0.16)~\mathrm{log_{10}(\Sigma_{mol}^{33GHz})}-(3.09 \pm 0.66)~.$$ The top right panel in Figure \[fig:fig9\] shows our data for two cases: a fixed “starburst” conversion factor and the mass surface-density dependent value. Internal to the starburst sample, the linearity or non-linearity of the slope depends entirely on the treatment of the conversion factor, and the assumption of the cospatiality between CO and radio emission; the apparent relationship between $\Sigma_{\rm SFR}$ and CO luminosity surface brightness is approximately linear. As mentioned above, the contrast between normal disk galaxies and starbursts played a large role in determining the @Kennicutt98 fit. The bottom panel of Figure \[fig:fig9\] explores this contrast. There, we compare our results to those found for kpc-size regions drawn from $30$ nearby disk galaxies by @Leroy13. Individual regions appear as green squares and the median and scatter in $\Sigma_{\rm SFR}$, in bins of fixed $\Sigma_{\rm mol}$, appear as red points with error bars. Note that, in contrast to @Kennicutt98, we consider only the molecular gas component of the ISM, and, in the normal galaxies, we consider individual kpc-sized regions. @Kennicutt98 include atomic gas and consider whole-disk averages. We chose our approach to focus on star-forming (molecular) gas in comparable sized regions in order to contrast the ability of gas to form stars in the two types of systems. Figure \[fig:fig9\] shows a significant contrast between disks and our starburst sample, even for matched $\alpha_{\rm CO}$ (a similar contrast was seen when comparing $\tau_{\rm dep}$). In that case, $\alpha_{\rm CO}=4.35$ M$_\odot$ pc$^{-2}$ (K km s$^{-1}$) for both samples, a fit to our sample and the @Leroy13 bins yield: $$\label{eqn:ks4} \mathrm{log_{10}(\Sigma_{SFR_{IR}}^{33GHz})} = (1.35 \pm 0.04)~\mathrm{log_{10}(\Sigma_{mol}^{33GHz})}-(3.85 \pm 0.13)~.$$ Meanwhile, adopting the starburst $\alpha_{\rm CO} = 0.8$ M$_\odot$ pc$^{-2}$ (K km s$^{-1}$) for our sample only yields: $$\label{eqn:ks3} \mathrm{log_{10}(\Sigma_{SFR_{IR}}^{33GHz})} = (1.63 \pm 0.07)~\mathrm{log_{10}(\Sigma_{mol}^{33GHz})}-(4.18 \pm 0.22)~.$$ In both cases, the data appear to support the “two sequences” idea, at least to some degree, with internal relationships in the two sub-samples that are more nearly linear, and a steep slope when contrasting both populations (but see below). This is particularly the case when we use a starburst conversion factor for our sample. Adopting $\mathrm{\alpha_{CO} \propto \Sigma_{mol}^{-0.5}}$ (see equation \[eq:alpha\_co\]) we find instead $$\label{eqn:ks5} \mathrm{log_{10}(\Sigma_{SFR_{IR}}^{33GHz})} = (1.87 \pm 0.06)~\mathrm{log_{10}(\Sigma_{mol}^{33GHz})}-(4.63 \pm 0.19)~.$$ In this case we find an even steeper slope when fitting the combined data, from the U/LIRGs studied here and the normal spirals from @Leroy13, than when we use a starburst conversion factor for our sample only, and even more so when we fit either sample alone. To some degree, this reinforces the “two sequences” view, but with a strong caveat. Our results are consistent with the idea that the depletion time is multi-valued at a fixed gas surface density, but they do [*not*]{} offer any strong evidence regarding a true bimodality. The data that we use are not complete in any meaningful sense. Therefore, the absence of intermediate $\tau_{\rm dep}$ points near where the two samples would overlap can easily be a selection effect. That is: there may be plenty of parts of galaxies that fill in apparently empty space in Figure \[fig:fig9\], our samples are simply not constructed to reveal this. Indeed, @Saintonge11 [@Huang15; @Genzel15] and others have convincingly shown that a continuous range of gas depletion times appear to exist within the population [see also @Scoville16 for further discussion on continuous and bi-modal star formation scaling relations]. Our results do strongly reinforce the idea that the disk-starburst contrast is essential to probe the non-linear nature of star formation scaling relations. We also show, following a number of others that the adopted conversion factor, in addition to the starburst sizes, plays a large role in the results. We summarize all the different fits to the gas star formation law using the different conversion factors in Table \[table:tbl-7\]. [*Efficiency per Free Fall Time:*]{} A popular class of models posits an approximately fixed fraction of gas converts to stars per gravitational free fall time, $\mathrm{\tau_{ff}^{\rm mol} = \sqrt{3\pi/(32G\rho_{mol})}}$ [e.g., @Krumholz05; @Krumholz12]. If we adopt a simple, spherical, with radius R$_{\rm 50,d}$, view of the geometry of the systems, we can estimate $\tau_{\rm ff}^{\rm mol}$. For a three dimensional Gaussian, this implies an aperture correction of $\sim 1/3.4$ for the total gas mass (or SFR) within that volume. Comparing $\tau_{\rm ff}^{\rm mol}$ to the depletion time of the molecular gas mas, $\mathrm{\tau_{dep}^{mol}}$, we estimate the efficiency of the conversion of the gas mass into stars per free fall time, or $\epsilon_{\rm ff}$ = $\tau_{\rm ff}^{\rm mol}/\tau_{\rm dep}^{\rm mol}$. We find median values for $\mathrm{\tau_{ff}^{mol}}$ of 1.1, 1.5, and 0.5 Myr for “starburst", surface-density dependent, and Galactic conversion factors. These numbers imply median $\epsilon_{\rm ff}$ of 8%, 15%, and 0.6%. The first two numbers appear high compared to the universal $\sim 1\%$ $\epsilon_{\rm ff}$ assumed in the @Krumholz12 model, and in more agreement with a non-universal star formation efficiency [@Semenov16], but we emphasize the uncertainty in the adopted geometry. Are Local Major Mergers Eddington-Limited Starbursts? {#sec:edd} ----------------------------------------------------- {width="3.5in"} The high density of star formation and luminosity in the inner parts of our targets undoubtedly creates strong feedback on the gas. This can suppress or even halt ongoing star formation, and in equilibrium we might expect this feedback to counter-balance the force of gravity, leading to some degree of self-regulation. Radiation pressure on dust has been proposed as the main feedback mechanism for compact, optically thick starbursts . Momentum injection by supernova explosions and cosmic ray pressure [e.g., @Socrates08; @FG13] also likely play a key role. The high $\mathrm{\Sigma^{33GHz}_{IR}}$ values derived for our targets and their very dusty nature makes them excellent candidates to be “Eddington-limited” starbursts. In such a system, the star formation surface density will increase until it yields a radiation pressure on dust that balances the force of gravitational collapse. Because we expect that the force from radiation pressure must be present, then if a source shows a luminosity surface density above this equilibrium value, then some other assumption in the calculation must break down. This could be the assumption of equilibrium, as the pressure exerted by radiation might temporarily or permanently suppress star formation and/or expel gas from the system in a galactic wind. Alternatively, the source of the luminosity could be something other than star formation. One common inference when this “maximal starburst” case is exceeded is that an appreciable part of the luminosity in the system may arise from an AGN. Alternatively, the assumptions about disk structure used to calculate the force of gravity may be wrong. For example, in the models of @Thompson05 the gas fraction and velocity dispersion play a key role. We have already seen some evidence that this case may apply to our systems. @Thompson05 noted an infrared luminosity surface density of $\mathrm{\Sigma_{IR} \sim 10^{13}~L_\odot~kpc^{-2}}$ as characteristic for dense, optically thick Eddington-limited starbursts. We showed above that a subset of our targets exhibit $\mathrm{\Sigma_{IR}^{33GHz}}$ near, or even above, this limit. In detail, the exact limiting $\mathrm{\Sigma^{33GHz}_{IR}}$ depends on the detailed structure of the starburst disk, including its size, stellar velocity dispersion ($\sigma$), gas mass fraction (f$_{\rm g}$), dust-to-gas ratio, and the Rosseland mean opacity ($\kappa$) of the system. Thus, the Eddington limit varies from source to source. Taking this in to account, we compare our inferred $\mathrm{\Sigma^{33GHz}_{IR}}$ (or F$_{\rm obs}$) for each target to the predicted Eddington flux. For hydrostatic equilibrium in a disk, the Eddington flux, $F_{\rm edd}$ is: $$\label{eqn:hidro} \mathrm{F_{edd}} = \frac{4 \pi G c \Sigma}{\kappa}~,$$ where $\Sigma$ is the surface density of the mass that dominates the gravitational potential involved in the star forming region and $\kappa$ is the effective opacity. The effective opacity depends on the characteristics of the system under study. Following @Thompson05 and , for systems that are optically thick to the UV radiation, but optically thin to the re-radiated far infrared emission, $\mathrm{\kappa (thin)\sim\Sigma_{\rm gas}^{-1}}$. For systems that are optically thick to the re-processed far infrared emission, i.e., when $\mathrm{\Sigma_{gas}~\gtrsim~1~g~cm^{-2}}$, $\mathrm{\kappa(thick)\approx\kappa_{o}T^{2}}$, where T is the temperature of the central star forming disk and $\mathrm{\kappa_{o}~\approx2.4\times10^{-4}~cm^{2}~g~K^{-2}}$ [@Semenov03]. The transition between regimes is expected to occur when $\mathrm{\Sigma_{gas}~\sim~1~g~cm^{-2}}$. Note that in systems without large IR optical depths, the momentum and turbulence from supernovae is expected to dominate support of the disk, rather than radiation pressure. For a Milky Way gas-to-dust ratio and $\Sigma = \Sigma_{mol}/f_{g}$ a version of Equation \[eqn:hidro\] that captures all three possible regimes is $$\label{eqn:fedd} {\rm F_{edd}} = \frac{\pi G c \Sigma_{\rm mol}^{2}}{\rm {f_{g}(1 + \tau_{IR} + 10n_{mol}^{-1/7} -exp[-\tau_{UV}])}}~.$$ Here $\mathrm{\Sigma_{mol}=\Sigma_{mol}^{33GHz}}$ is the gas surface density (see Table \[table:tbl-6\]), and f$_{g}$ is the gas mass fraction in the core of the galaxy. $\tau_{\rm IR}=\kappa_{\rm IR}(\rm T)\Sigma_{\rm mol}/2$ is the infrared optical depth and $\mathrm{\tau_{UV}\sim500-1000(cm^{2}~g^{-1})\times\Sigma_{mol}/2}$ the ultraviolet optical depth. We approximate the contribution to support by supernovae (SNe) as $10n_{\rm mol}^{-1/7}$ (see Appendix of @FG13, and ); the numerical prefactor can vary by a factor of several, up to $\sim$30n$_{\rm mol}^{-1/7}$), where $\mathrm{n_{mol}\equiv n_{mol}^{33GHz}}$ is the number density of the gas (see Table \[table:tbl-6\]). In order to derive $\kappa_{\rm IR}(\rm T)$ we assume equation (40) from @Thompson05 describes the relation between $T$, $T_{\rm eff}$, and the vertical IR optical depth. We then solve the implicit equation for $T$ assuming that, $$\frac{\kappa_{IR}(T)}{\rm (cm^{2}~g^{-1})} = \begin{cases} 2.4\times10^{-4}~T^{2}, & \mbox{if } T\mbox{ $<$ 180 K} \\ 2.4\times10^{-4}~180^{2}\approx7.8, & \mbox{if } T\mbox{ $\geq$ 180 K} \end{cases}$$ where $\mathrm{\Sigma^{33GHz}_{IR} = \sigma_{SB} T_{eff}^{4}}$, and $\sigma_{SB}$ is the Stefan-Boltzmann constant. In Figure \[fig:fig10\] we show the resulting $\mathrm{F_{obs}/(f_{g}*F_{edd})}$ estimated for each galaxy. We draw lines at the Eddington limit where $\mathrm{F_{obs} = F_{edd}}$ for different gas fractions, and above which the systems are considered super-Eddington. We include the cases for which we consider SNe feedback (black solid circles) and where we do not (open red circles). If we assume the gas fraction in the center of the sources is closer to 1 and neglect supernovae feedback, we observe 5 systems that are super-Eddington including Mrk 231 (UGC 08058), Arp 220, and CGCG 448-020. As noted earlier, Mrk 231 is known to host a strong AGN, and if this drives the IR luminosity then this Eddington calculation for a starburst disk does not apply. Arp 220 has mid-IR evidence of energetic AGN – based on low PAH equivalent widths (= 0.03$-$0.17) and/or high 30-to-15 $\mu$m flux density ratios (= 10$-$20) indicative of very warm, Seyfert-like mid-IR dust emission [@Stier13]. Arp 220 is a special case, where the mid-IR diagnostics potentially break due to the high dust opacity of the system. This is also applicable to CGCG 448-020 for which the source dominating the emission at IR and radio wavelengths (north-east component, see Figure \[fig:fig1\] and the Appendix for more information) is highly obscured. The sources in our sample are extreme starbursts for which SNe feedback is most likely important, especially in systems that are more extended and warm (T $<$ 180 K). If we include SNe feedback in the calculation of $\mathrm{F_{edd}}$ (solid black circles in Figure \[fig:fig10\]) we observe that for a gas fraction of 1, 11 out of 22 systems in our sample show super-Eddington values, including the systems mentioned above. Assuming a more conservative gas fraction of 0.3, which is about the system averaged gas mass fraction based on @Larson16, we find that 5 systems are super-Eddington. Note that CGCG 448-020 is a special case since it shows super-Eddington values independent of the gas fraction, indicating the potential presence of an AGN. We note that our results highly depend on the adopted gas fraction, and while we might expect some funneling of gas to the center to raise the gas fraction to higher values locally, the best way to further improve our constraints are resolved observations of the disk dynamics, which can yield the total (dynamical) mass, velocity dispersion, and gas mass. Conclusions {#sec:conclusions} =========== We present a high resolution imaging survey of 33 GHz continuum emission from local U/LIRGs. Using all four VLA configurations and a bandwidth of $\Delta \nu = 2$ GHz, we achieve very high resolutions of $0\,\farcs07$–$0\,\farcs67$, or 30-720 pc at the distance of these sources, while still retaining sensitivity to emission on large scales. This is the first such survey at such high frequencies (for the VLA). As a result, we improve on the resolution of previous work by @Condon91 and @Condon90 by a factor of $4$. Because of the steep spectral index of galaxies in this range, the improved sensitivity gained from the VLA upgrade was a key element in the survey. Using these data, we find: 1. Most of the 33 GHz emission observed at low resolution arises from sources that appear compact in the highest resolution maps. For the majority of the U/LIRGs studied here, more than 50% of the integrated flux density at 33 GHz arises from sources with Gaussian-like morphologies at high resolution and extend typically a few times the size of the beam. 2. The 33 GHz emission reflects a mixture of synchrotron and free-free emission. For different approaches, we achieve slightly different results, but within the uncertainties approximately equal fractions of thermal and nonthermal emission could contribute at 33 GHz. To improve on this uncertain number, improved coverage of the radio SED, especially achieving reliable flux densities at many frequencies in the range $15$–$50$ GHz, will be extremely helpful. Unless the emission is highly clumped within the recovered high resolution beam, brightness temperature arguments suggest that all of the observed U/LIRGs are optically thin at 33 GHz. 3. By making use of the 33 GHz size to indicate the active, star-forming region, we provide estimates for the surface densities of gas, star formation, and infrared emission. These quantities are more extreme than those found in typical star-forming galaxies but also vary strongly across the sample, spanning a range of $\approx$ 4 dex. The highest values in our sample are among the highest measured for any galaxies. 4. We also make use of the measured 33 GHz sizes of the sample to estimate their star formation rate surface densities, $\mathrm{\Sigma_{SFR_{33 GHz}}}$. We find that $L_{\rm [C II]} / L_{\rm FIR}$ decreases with increasing $\Sigma_{\rm SFR_{33GHz}}$, increasing opacity (as measured via the flattening of the radio spectral index between 1.5 and 6 GHz), and increasing compactness. These measurement agree with prior studies which used infrared sizes measured at coarser resolution to estimate $\Sigma_{\rm IR}^{\rm 33GHz}$ (or $\Sigma_{\rm SFR_{IR}}^{\rm 33GHz}$). They confirm that the \[[Cii]{}\] ‘deficit’ is more pronounced in the most compact and obscured U/LIRGs. 5. We consider the implications for star formation scaling relations from $\Sigma_{\rm SFR_{IR}}^{\rm 33GHz}$ and $\Sigma_{\rm gas}^{\rm 33GHz}$ derived combining the 33 GHz size estimates with unresolved CO (1$-$0) and IR observations. For any single, fixed conversion factor and considering only the U/LIRGs, we find a slope near unity ($\approx 1.02$) relating the two. However, the U/LIRGs studied here contrast with results for normal spiral galaxies from [@Leroy13], and a nonlinear slope is needed to relate the two different populations [consistent with @Kennicutt98; @Liu15]. 6. The exact value of the power law index needed to fit both normal disks and these U/LIRGs depends sensitively on the sizes of the U/LIRGs (which we measured), on the assumption that the star formation, traced by 33 GHz, and the molecular gas, as traced by CO, have matched structure, and the prescription for the CO-to-H$_2$ conversion factor (which is highly uncertain). We show results for three common approaches to the conversion factor, and the power law index relating normal disk galaxies to the U/LIRGs studied in this paper varies from $\sim 1.4$ to $\sim 1.9$. 7. The high column densities that we infer imply high opacities outside the $\sim$cm and mm-wave regime. By adopting a “starburst" conversion factor, the average extinction at optical wavelengths is $A_V \sim 22-12,000$ mag for this sample. $13$ of the observed sources appear X-ray Compton thick, with average $\mathrm{N_{H} > 1.5\times10^{24}~cm^{-2}}$. At IR wavelengths, the opacity is less, $\tau_{100} \sim 0.02-12$, however they are still affected by dust with those same 13 sources, but one, being optically thick at 100 $\mu$m. The combination of the measured sizes at 33 GHz with the $1.5$ GHz flux densities from @Condon90 also indicate that opacity must play a significant role at lower radio frequencies. 8. The targets show high infrared surface brightnesses, with 7 sources having $\mathrm{\Sigma_{IR}^{33GHz}>10^{13} L_{\odot}~kpc^{-2}}$, a characteristic value suggested by @Thompson05 for dusty, radiation-pressure supported starburst galaxies. However, if we consider feedback from supernovae and adopt a nuclear gas fraction of 1, we find 11 out of 22 systems are super-Eddington. This number decreases to 5 if we adopt a gas fraction of 0.3 instead. We note the need for both detailed observations of the inner disk structure and several observational subtleties that should be accounted for in comparing the observed $\Sigma_{\rm IR}^{\rm 33GHz}$ to models. We thank the anonymous referee for providing excellent comments that further improved this paper. We thank Jason Chu for making Herschel/PACS images available for quick visual comparison and further discussions on the mid-IR data. Support for this work was provided by the NSF through the Grote Reber Fellowship Program administered by Associated Universities, Inc./National Radio Astronomy Observatory. L.B-M. was supported by Fulbright, Becas Chile - CONICYT. The work of A.K.L. is supported by the National Science Foundation under Grants No. 1615105 and 1615109. A.S.E., G.C.P. and L.B-M. were supported by NSF grant AST 1109475. G.C.P. was supported by a FONDECYT Postdoctoral Fellowship (No. 3150361). T.D-S. acknowledges support from ALMA-CONICYT project 31130005 and FONDECYT regular project 1151239. This research made use of the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration, and NASA’s Astrophysics Data System Bibliographic Services. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. Ali-Ha[ï]{}moud, Y., Hirata, C. M., & Dickinson, C. 2009, , 395, 1055 Anantharamaiah, K. R., Viallefond, F., Mohan, N. R., Goss, W. M., & Zhao, J. H. 2000, , 537, 613 Andrews, B. H., & Thompson, T. A. 2011, , 727, 97 Armus, L., Charmandaris, V., Bernard-Salas, J., et al. 2007, , 656, 148 Armus, L., Mazzarella, J. M., Evans, A. S., et al. 2009, , 121, 559 Barcos-Muñoz, L., Leroy, A. K., Evans, A. S., et al. 2015, , 799, 10 Bohlin, R. C., Savage, B. D., & Drake, J. F. 1978, , 224, 132 Bolatto, A. D., Leroy, A. K., Rosolowsky, E., Walter, F., & Blitz, L. 2008, , 686, 948-965 Bolatto, A. D., Wolfire, M., & Leroy, A. K. 2013, , 51, 207 Bouch[é]{}, N., Cresci, G., Davies, R., et al. 2007, , 671, 303 Chu, J. K., Sanders, D. B., Larson, K. L., et al. 2017, arXiv:1702.01756 Clemens, M. S., Vega, O., Bressan, A., et al. 2008, , 477, 95 Clemens, M. S., Scaife, A., Vega, O., & Bressan, A. 2010, , 405, 887 Comastri, A. 2004, Supermassive Black Holes in the Distant Universe, 308, 245 Condon, J. J., Helou, G., Sanders, D. B., & Soifer, B. T. 1990, , 73, 359 Condon, J. J., & Yin, Q. F. 1990, , 357, 97 Condon, J. J., Huang, Z.-P., Yin, Q. F. & Thuan, T. X. 1991, , 378, 65 Condon, J. J. 1992, , 30, 575 Condon, J. J., Helou, G., Sanders, D. B., & Soifer, B. T. 1996, , 103, 81 Condon, J. J., Cotton, W. D., Greisen, E. W., et al. 1998, , 115, 1693 Daddi, E., Elbaz, D., Walter, F., et al. 2010, , 714, L118 D[í]{}az-Santos, T., Charmandaris, V., Armus, L., et al. 2010, , 723, 993 D[í]{}az-Santos, T., Armus, L., Charmandaris, V., et al. 2013, , 774, 68 Downes, D., & Solomon, P.M. 1998, , 507, 615 Draine, B. T., & Lazarian, A. 1998, , 508, 157 Evans, A. S., Mazzarella, J. M., Surace, J. A., & Sanders, D. B. 2002, , 580, 749 Faucher-Gigu[è]{}re, C.-A., Quataert, E., & Hopkins, P. F. 2013, , 433, 1970 Fischer, J., Abel, N. P., Gonz[á]{}lez-Alfonso, E., et al. 2014, , 795, 117 Fukui, Y., & Kawamura, A. 2010, , 48, 547 Genzel, R., Tacconi, L. J., Rigopoulou, D., Lutz, D., & Tecza, M. 2001, , 563, 527 Genzel, R., Tacconi, L. J., Gracia-Carpio, J., et al. 2010, , 407, 2091 Genzel, R., Tacconi, L. J., Lutz, D., et al. 2015, , 800, 20 Haan, S., Surace, J. A., Armus, L., et al. 2011, , 141, 100 Helmboldt, J. F., Taylor, G. B., Tremblay, S., et al. 2007, , 658, 203 Huang, M.-L., & Kauffmann, G. 2015, , 450, 1375 Hummer, D. G., & Storey, P. J. 1987, , 224, 801 Iono, D., Wilson, C. D., Yun, M. S., et al. 2009, , 695, 1537 Iwasawa, K., Sanders, D. B., Evans, A. S., et al. 2009, , 695, L103 Iwasawa, K., Sanders, D. B., Teng, S. H., et al. 2011, , 529, A106 Johnson, K. E., Leroy, A. K., Indebetouw, R., et al. 2015, , 806, 35 Kennicutt, R. C., Jr. 1998, , 498, 541 Keto, E., Ho, L. C., & Lo, K.-Y. 2005, , 635, 1062 Kim, C.-G., & Ostriker, E. C. 2015, , 802, 99 Krumholz, M. R., & McKee, C. F. 2005, , 630, 250 Krumholz, M. R., Dekel, A., & McKee, C. F. 2012, , 745, 69 Lacki, B. C., Thompson, T. A., & Quataert, E. 2010, , 717, 1 Larson, K. L., Sanders, D. B., Barnes, J. E., et al. 2016, , 825, 128 Leroy, A. K., Evans, A. S., Momjian, E., et al. 2011, , 739, L25 Leroy, A. K., Walter, F., Sandstrom, K., et al. 2013, , 146, 19 Leroy, A. K., Walter, F., Martini, P., et al. 2015, , 814, 83 Li, A., & Draine, B. T. 2001, , 554, 778 Liu, L., Gao, Y., & Greve, T. R. 2015, , 805, 31 Lonsdale, C. J., Lonsdale, C. J., Smith, H. E., & Diamond, P. J. 2003, , 592, 804 Luhman, M. L., Satyapal, S., Fischer, J., et al. 2003, , 594, 758 Lutz, D., Genzel, R., Sternberg, A., et al. 1996, , 315, L137 Lutz, D., Berta, S., Contursi, A., et al. 2016, , 591, A136 Malhotra, S., Kaufman, M. J., Hollenbach, D. et al. 2001, , 561, 766 Marvil, J., Owen, F., & Eilek, J. 2015, , 149, 32 McMullin, J. P., Waters, B., Schiebel, D. et al. 2007, Astronomical Data Analysis Software and Systems XVI, 376, 127 Murphy, E. J., Helou, G., Condon, J. J., et al. 2010, , 709, L108 Murphy, E. J., Bremseth, J., Mason, B. S., et al. 2012, , 761, 97 Murphy, E. J., Stierwalt, S., Armus, L., Condon, J. J., & Evans, A. S. 2013, , 768, 2 Murray, N., Quataert, E., & Thompson, T. A. 2005, , 618, 569 Narayanan, D., Krumholz, M. R., Ostriker, E. C., & Hernquist, L. 2012, , 421, 3127 Neff, S. G., Ulvestad, J. S., & Teng, S. H. 2004, , 611, 186 Ostriker, E. C., & Shetty, R. 2011, , 731, 41 Papadopoulos, P. P., van der Werf, P., Xilouris, E., Isaak, K. G., & Gao, Y. 2012, , 751, 10 Petric, A. O., Armus, L., Howell, J., et al. 2011, , 730, 28 Privon, G. C., Aalto, S., Falstad, N., et al. 2017, , 835, 213 Puxley, P. J., Brand, P. W. J. L., Moore, T. J. T., et al. 1989, , 345, 163 Rabidoux, K., Pisano, D. J., Kepley, A. A., Johnson, K. E., & Balser, D. S. 2014, , 780, 19 Rupke, D. S. N., Veilleux, S., & Baker, A. J. 2008, , 674, 172-193 Rupke, D. S. N., & Veilleux, S. 2013, , 768, 75 Saintonge, A., Kauffmann, G., Wang, J., et al. 2011, , 415, 61 Sanders, D. B., Soifer, B. T., Elias, J. H., et al. 1988, , 325, 74 Sanders, D. B., & Mirabel, I. F. 1996, , 34, 749 Sanders, D. B., Mazzarella, J. M., Kim, D.-C., et al. 2003 , 126, 1607 Sault, R.J., & Wieringa, M.H. 1994, , 108, 585 Schruba, A., Leroy, A. K., Walter, F., et al. 2011, , 142, 37 Scoville, N. 2003, Journal of Korean Astronomical Society, 36, 167 Scoville, N., Aussel, H., Sheth, K., et al. 2014, , 783, 84 Scoville, N., Sheth, K., Aussel, H., et al. 2016, , 820, 83 Scoville, N., Murchikova, L., Walter, F., et al. 2017, , 836, 66 Semenov, D., Henning, T., Helling, C., Ilgner, M., & Sedlmayr, E. 2003, , 410, 611 Semenov, V. A., Kravtsov, A. V., & Gnedin, N. Y. 2015, arXiv:1512.03101 Shetty, R., Glover, S. C., Dullemond, C. P., & Klessen, R. S. 2011, , 412, 1686 Sliwa, K., Wilson, C. D., Aalto, S., & Privon, G. C. 2017, , 840, L11 Socrates, A., Davis, S. W., & Ramirez-Ruiz, E. 2008, , 687, 202 Soifer, B. T., Neugebauer, G., Matthews, K., et al. 2000, , 119, 509 Solomon, P. M., Downes, D., Radford, S. J. E., & Barrett, J. W. 1997, , 478, 144 Stierwalt, S., Armus, L., Surace, J. A., et al. 2013, , 206, 1 Thompson, T.A., Quataert, E., & Murray, N. 2005, , 630, 167 Thompson, T. A., Quataert, E., Waxman, E., Murray, N., & Martin, C. L. 2006, , 645, 186 Ulvestad, J. S., Wrobel, J. M., & Carilli, C. L. 1999, , 516, 127 Vardoulaki, E., Charmandaris, V., Murphy, E. J., et al. 2015, , 574, A4 Veilleux, S., Rupke, D. S. N., Kim, D.-C., et al. 2009, , 182, 628-666 Wei, L. H., Keto, E., & Ho, L. C. 2012, , 750, 136 Wilson, C. D., Petitpas, G. R., Iono, D., et al. 2008, , 178, 189-224 Yun, M. S., Reddy, N. A., Scoville, N. Z., et al. 2004, , 601, 723 [lccccccc]{} CGCG 436-030&MCG +02-04-025&$\mathrm{01^{h}20^{m}02\,\fs722}$ & $\mathrm{+14^{\circ}21'42\,\farcs94}$ & 11.64 & 127 &0.601 & 1\ IRAS F01364-1042 & 2MASXJ01385289-1027113 &$\mathrm{01^{h}38^{m}52\,\fs921}$ & $\mathrm{-10^{\circ}27'11\,\farcs42}$ & 11.81 & 201 &0.942 & 2\ III Zw 035 & &$\mathrm{01^{h}44^{m}30\,\fs500}$& $\mathrm{+17^{\circ}06'05\,\farcs00}$ & 11.58 & 111 &0.526 & 3\ VII Zw 031 & & $\mathrm{05^{h}16^{m}46\,\fs096}$ & $\mathrm{+79^{\circ}40'13\,\farcs28}$ & 11.95 & 229 &1.066 & 4\ IRAS 08572+3915 & & $\mathrm{09^{h}00^{m}25\,\fs390}$ & $\mathrm{+39^{\circ}03'54\,\farcs40}$ & 12.13 & 254 &1.176 & 5\ UGC 04881 & Arp 55 & $\mathrm{09^{h}15^{m}55\,\fs100}$ & $\mathrm{+44^{\circ}19'55\,\farcs00}$ & 11.70 & 169 &0.796 & 6\ UGC 05101 & & $\mathrm{09^{h}35^{m}51\,\fs595}$ & $\mathrm{+61^{\circ}21'11\,\farcs45}$ & 11.97 & 168 &0.792 & 7\ MCG +07-23-019 & Arp 148 & $\mathrm{11^{h}03^{m}53\,\fs200}$ & $\mathrm{+40^{\circ}50'57\,\farcs00}$ & 11.61 & 149 &0.704 & 8\ NGC 3690 & Arp 299 & $\mathrm{11^{h}28^{m}32\,\fs300}$ & $\mathrm{+58^{\circ}33'43\,\farcs00}$ & 11.82 & 45.2 &0.217 & 9\ UGC 08058 & Mrk 231 & $\mathrm{12^{h}56^{m}14\,\fs234}$ & $\mathrm{+56^{\circ}52'25\,\farcs24}$ & 12.52 & 181 &0.849 & 10\ VV 250 & UGC 08335 NED02& $\mathrm{13^{h}15^{m}34\,\fs980}$ & $\mathrm{+62^{\circ}07'28\,\farcs66}$ & 11.77 & 132 &0.621 & 11\ UGC 08387 & Arp 193, IC 883 & $\mathrm{13^{h}20^{m}35\,\fs300}$ & $\mathrm{+34^{\circ}08'21\,\farcs00}$ & 11.65 & 101 &0.479 & 12\ UGC 08696 & Mrk 273 & $\mathrm{13^{h}44^{m}42\,\fs111}$ & $\mathrm{+55^{\circ}53'12\,\farcs65}$ & 12.15 & 162 &0.761 & 13\ VV 340a & UGC 09618 NED02& $\mathrm{14^{h}57^{m}00\,\fs826}$ & $\mathrm{+24^{\circ}37'04\,\farcs12}$ & 11.67 & 144 &0.665 & 14\ VV 705 & I Zw 107 & $\mathrm{15^{h}18^{m}06\,\fs344}$ & $\mathrm{+42^{\circ}44'36\,\farcs69}$ & 11.87 & 172 &0.807 & 15\ IRAS 15250+3608 & & $\mathrm{15^{h}26^{m}59\,\fs404}$ & $\mathrm{+35^{\circ}58'37\,\farcs53}$ & 12.02 & 238 &1.105 & 16\ UGC 09913 & Arp 220 & $\mathrm{15^{h}34^{m}57\,\fs116}$ & $\mathrm{+23^{\circ}30'11\,\farcs47}$ & 12.16 & 77.2 &0.369 & 17\ IRAS 17132+5313 & & $\mathrm{17^{h}14^{m}20\,\fs000}$ & $\mathrm{+53^{\circ}10'30\,\farcs00}$ & 11.90 & 217 &1.012 & 18\ IRAS 19542+1110 & &$\mathrm{19^{h}56^{m}35\,\fs440}$ & $\mathrm{+11^{\circ}19'02\,\farcs60}$ & 12.07 & 277 &1.277 & 19\ CGCG 448-020 & II Zw 096 &$\mathrm{20^{h}57^{m}23\,\fs900}$& $\mathrm{+17^{\circ}07'39\,\farcs00}$& 11.79 & 148 &0.698 & 20\ IRAS 21101+5810 & 2MASX J21112926+5823074 &$\mathrm{21^{h}11^{m}30\,\fs400}$ & $\mathrm{+58^{\circ}23'03\,\farcs20}$ & 11.74 & 162 &0.764& 21\ IRAS F23365+3604 & 2MASX J23390127+3621087 & $\mathrm{23^{h}39^{m}01\,\fs273}$ & $\mathrm{+36^{\circ}21'08\,\farcs31}$ & 12.16 & 273 &1.262 & 22 [lcc]{}\[th\] CGCG 436-030 & 3C 48 & J0117+1418\ IRAS F01364-1042 & 3C 48 & J0141-0928\ III Zw 035 & 3C 48 & J0139+1753\ VII Zw 031 & 3C 48 & J0410+7656\ IRAS 0857+3915 & 3C 286 & J0916+3854\ UGC 04881 & 3C 286 & J0920+4441\ UGC 05101 & 3C 286 & J0921+6215\ MCG +07-23-019 & 3C 286 & J1101+3904\ NGC 3690 & 3C 286 & J1128+5925\ Mrk 231 & 3C 286 & J1302+5748\ VV 250 & 3C 286 & J1302+5748\ UGC 08387 & 3C 286 & J1317+3425\ UGC 08696 & 3C 286 & J1337+5501\ VV 340a & 3C 286 & J1443+2501\ VV 705 & 3C 286 & J1521+4336\ IRAS 15250+3609 & 3C 286 & J1522+3144\ Arp 220 & 3C 286 & J1539+2744\ IRAS 17132+5313 & 3C 286 & J1740+5211\ IRAS 19542+1110 & 3C 48 & J1955+1358\ CGCG 448-020 & 3C 48 & J2051+1743\ IRAS 21101+5810 & 3C 48 & J2123+5500\ IRAS F23365+3604 & 3C 48 & J2330+3348 [lcccccrr]{} CGCG 436-030 & $\mathrm{01^{h}20^{m}02\,\fs628}$ & $\mathrm{+14^{\circ}21'42\,\farcs37}$ & $0\,\farcs109 \times 0\,\farcs090$ & 26.0 & 5.6 $\pm$ 0.7& 18.6 $\pm$ 0.04 & 49.1 $\pm$ 2.5\ IRAS F01364-1042 & $\mathrm{01^{h}38^{m}52\,\fs885}$ & $\mathrm{-10^{\circ}27'11\,\farcs54}$ & $0\,\farcs141 \times 0\,\farcs086$ & 31.4 & 4.7 $\pm$ 0.6 & 10.0 $\pm$ 0.04& 15.2 $\pm$ 0.8\ III Zw 035 & $\mathrm{01^{h}44^{m}30\,\fs536}$ & $\mathrm{+17^{\circ}06'08\,\farcs65}$ & $0\,\farcs145 \times 0\,\farcs117$ & 31.2 & 7.3 $\pm$ 0.9& 25.4 $\pm$ 0.05 & 41.2 $\pm$ 2.1\ VII Zw 031 & $\mathrm{05^{h}16^{m}46\,\fs028}$ & $\mathrm{+79^{\circ}40'12\,\farcs80}$ & $0\,\farcs797 \times 0\,\farcs566$ & 39.1 & 3.0 $\pm$ 0.5 & 12.5 $\pm$ 0.04 & 41.6 $\pm$ 4.2\ IRAS 08572+3915 & $\mathrm{09^{h}00^{m}25\,\fs353}$ & $\mathrm{+39^{\circ}03'54\,\farcs22}$ & $0\,\farcs254 \times 0\,\farcs193$ & 27.3 & 2.1 $\pm$ 0.4 & 4.44 $\pm$ 0.04 & 4.5 $\pm$ 0.2\ UGC 04881 & & & $0\,\farcs253 \times 0\,\farcs191$ & 26.6 & 1.6 $\pm$ 0.3 & 11.4 $\pm$ 0.09 & 31.6 $\pm$ 1.6\ & $\mathrm{09^{h}15^{m}55\,\fs513}$ & $\mathrm{+44^{\circ}19'57\,\farcs79}$ & “ & ” & 0.88 $\pm$ 0.15 & ... & ...\ & $\mathrm{09^{h}15^{m}54\,\fs787}$ & $\mathrm{+44^{\circ}19'49\,\farcs83}$ & “ & ” & 0.76 $\pm$ 0.23 & ... & ...\ UGC 05101 & $\mathrm{09^{h}35^{m}51\,\fs882}$ & $\mathrm{+61^{\circ}21'10\,\farcs84}$ &$0\,\farcs291 \times 0\,\farcs273$ & 25.7 & 14.0 $\pm$ 1.7 & 61.5 $\pm$ 0.08 & 150.0 $\pm$ 7.5\ MCG +07-23-019 & $\mathrm{11^{h}30^{m}54\,\fs018}$ & $\mathrm{+40^{\circ}50'59\,\farcs739}$ & $0\,\farcs228 \times 0\,\farcs202$ & 29.5 & 5.9 $\pm$ 0.8 & 16.0 $\pm$ 0.06 & 31.3 $\pm$ 1.6\ NGC 3690 & & & $0\,\farcs260 \times 0\,\farcs240$ & 25.7 & 115.2 $\pm$ 11.3 & 275.5 $\pm$ 0.34 & 658.0 $\pm$ 32.9\ & $\mathrm{11^{h}28^{m}30\,\fs851}$ & $\mathrm{+58^{\circ}33'44\,\farcs67}$ &“ & ” & 39.0 $\pm$ 6.0 & ... & ...\ & $\mathrm{11^{h}28^{m}33\,\fs596}$ & $\mathrm{+58^{\circ}33'48\,\farcs02}$ & “ & ” & 76.3 $\pm$ 9.5 & ... & ...\ UGC 08058 & $\mathrm{12^{h}56^{m}14\,\fs186}$ & $\mathrm{+56^{\circ}52'25\,\farcs29}$ &$0\,\farcs257 \times 0\,\farcs227$ & 25.1 & 92.8 $\pm$ 11.1 & 312.8 $\pm$ 0.2 & 296.0 $\pm$ 14.8\ VV 250 & & &$0\,\farcs236 \times 0\,\farcs219$ & 24.7 & 9.9 $\pm$ 1.2 & 19.6 $\pm$ 0.05\ & $\mathrm{13^{h}15^{m}34\,\fs890}$ & $\mathrm{+62^{\circ}07'27\,\farcs98}$ &“ & ” & 8.5 $\pm$ 1.1 & ... & 44.6 $\pm$ 2.2\ & $\mathrm{13^{h}15^{m}30\,\fs359}$ & $\mathrm{+62^{\circ}07'44\,\farcs51}$ &“ & ” & 1.4 $\pm$ 0.3 & ... & 8.5 $\pm$ 0.4\ UGC 08387 & $\mathrm{13^{h}20^{m}35\,\fs352}$ & $\mathrm{+34^{\circ}08'21\,\farcs11}$ &$0\,\farcs098 \times 0\,\farcs073$ & 19.2 & 17.7 $\pm$ 2.1 & 46.3 $\pm$ 0.08 & 101.0 $\pm$ 5.1\ UGC 08696 & $\mathrm{13^{h}44^{m}42\,\fs133}$ & $\mathrm{+55^{\circ}53'13\,\farcs02}$ &$0\,\farcs259 \times 0\,\farcs240$ & 26.2 & 19.9 $\pm$ 2.4 & 60.3 $\pm$ 0.08 & 143.0 $\pm$ 7.2\ VV 340a & $\mathrm{14^{h}57^{m}00\,\fs703}$ & $\mathrm{+24^{\circ}37'03\,\farcs69}$ &$0\,\farcs494 \times 0\,\farcs437$ & 22.9 & 3.8 $\pm$ 0.5 & 23.7 $\pm$ 0.12 & 74.6 $\pm$ 3.7\ VV 705 & $\mathrm{15^{h}18^{m}60\,\fs175}$ & $\mathrm{+42^{\circ}44'44\,\farcs51}$ &$0\,\farcs071 \times 0\,\farcs062$& 17.7 & 7.2 $\pm$ 0.9 & 19.6 $\pm$ 0.05 & 47.8 $\pm$ 2.4\ IRAS 15250+3609 & $\mathrm{15^{h}26^{m}59\,\fs440}$ & $\mathrm{+35^{\circ}58'37\,\farcs32}$ & $0\,\farcs075 \times 0\,\farcs067$ & 20.7 & 5.1 $\pm$ 0.6 & 12.0 $\pm$ 0.04 & 13.8 $\pm$ 0.7\ Arp 220 & $\mathrm{15^{h}34^{m}57\,\fs260}$ & $\mathrm{+23^{\circ}30'11\,\farcs04}$ &$0\,\farcs087 \times 0\,\farcs069$ & 21.9 & 65.6 $\pm$ 7.9 & 194.5 $\pm$ 0.08 & 324.0 $\pm$ 16.2\ IRAS 17132+5313 & $\mathrm{17^{h}14^{m}20\,\fs172}$ & $\mathrm{+53^{\circ}10'29\,\farcs77}$ & $0\,\farcs082 \times 0\,\farcs075$ & 18.5 & 3.1 $\pm$ 0.5 & 9.3 $\pm$ 0.04 & 25.8 $\pm$ 1.3\ IRAS 19542+1110 & $\mathrm{19^{h}56^{m}35\,\fs770}$ & $\mathrm{+11^{\circ}19'04\,\farcs98}$ & $0\,\farcs087 \times 0\,\farcs081$ & 21.3 & 2.7 $\pm$ 0.4 & 9.5 $\pm$ 0.04 & 20.3 $\pm$ 2.0\ CGCG 448-020 & $\mathrm{20^{h}57^{m}24\,\fs229}$ & $\mathrm{+17^{\circ}07'39\,\farcs04}$ & $0\,\farcs090 \times 0\,\farcs079$ & 23.1 & 5.3 $\pm$ 0.7 & 14.6 $\pm$ 0.06 & 36.3 $\pm$ 3.6\ IRAS 21101+5810 & $\mathrm{21^{h}11^{m}29\,\fs300}$ & $\mathrm{+58^{\circ}23'08\,\farcs65}$ & $0\,\farcs137 \times 0\,\farcs108$ & 27.1 & 3.9 $\pm$ 0.5 & 9.8 $\pm$ 0.04 &22.2 $\pm$ 2.2\ IRAS F23365+3604 & $\mathrm{23^{h}39^{m}01\,\fs259}$ & $\mathrm{+36^{\circ}21'08\,\farcs66}$ & $0\,\farcs098 \times 0\,\farcs091$ & 27.5 & 2.9 $\pm$ 0.4 & 10.6 $\pm$ 0.04 & 24.6 $\pm$ 2.5\ [lcccccccc]{}\[th\] CGCG 436-030 &$0\,\farcs705 \times 0\,\farcs573$ & 19.9 &-0.309 & 0.238 & no & 0.308\ IRAS F01364-1042 & $0\,\farcs141 \times 0\,\farcs086$ & 40.7 &-2.147 & 0.045 & no & 0.447\ III Zw 035 & $0\,\farcs145 \times 0\,\farcs117$ & 64.5 &-1.956 & 0.031 & no& 0.502\ VII Zw 031 &$2\,\farcs060 \times 1\,\farcs415$ & 5.7&0.922 & 1.739 & yes & 0.463\ IRAS 08572+3915 & $0\,\farcs254 \times 0\,\farcs193$& 39.7 & -1.415& 0.130 & yes & 0.535\ UGC 04881 & & & & & &\ & $0\,\farcs253 \times 0\,\farcs191$ & 23.4 & -1.420 & 0.088 & yes & 0.683\ & $5\,\farcs049 \times 3\,\farcs543$ & 3.0 & 1.148 & 1.683 & yes & 0.229\ UGC 05101 & $0\,\farcs291 \times 0\,\farcs273$ &70.1 &-0.969 & 0.146 & no & 0.255\ MCG +07-23-019 & $0\,\farcs228 \times 0\,\farcs202$&25.6& -1.230 & 0.096 & no & 0.253\ NGC 3690 &$0\,\farcs379 \times 0\,\farcs327$ & &0.785 & 0.302 & yes &\ & “ & 5.1 & 0.701 & 0.275 & yes & 0.024\ & ” & 20.2 & 0.029 & 0.127 & no & 0.030\ UGC 08058 & $0\,\farcs257 \times 0\,\farcs227$ &895.8 & -1.534 & 0.082 & no & 0.345\ VV 250 && & & & &\ & $0\,\farcs236 \times 0\,\farcs219$ & 5.0 & -0.180 & 0.285 & yes & 0.109\ & $10\,\farcs71 \times 9\,\farcs340$ & 3.2 & 1.895 & 3.106 & yes & 0.058\ UGC 08387 &$0\,\farcs098 \times 0\,\farcs073$ &7.4& -0.867 & 0.100 & yes & 0.026\ UGC 08696 &$0\,\farcs259 \times 0\,\farcs240$ &56.6& -0.820 & 0.167 & no & 0.205\ VV 340a &$2\,\farcs983 \times 2\,\farcs488$ &5.0& 1.249 & 1.581 & yes & 0.491\ VV 705 &$0\,\farcs583 \times 0\,\farcs579$& 5.0&0.276 & 0.625 & no & 0.071\ IRAS 15250+3609 & $0\,\farcs075 \times 0\,\farcs067$ & 52.5&-2.412 & 0.039 & no & 0.366\ Arp 220 &$0\,\farcs087 \times 0\,\farcs069$ &45.3 &-1.047 & 0.062 & no & 0.132\ IRAS 17132+5313 & $0\,\farcs911 \times 0\,\farcs864$ &5.8& 0.504 & 1.020 & yes & 0.259\ IRAS 19542+1110 & $0\,\farcs087 \times 0\,\farcs081$ &7.6& -1.410 & 0.142 & yes & 0.342\ CGCG 448-020 &$0\,\farcs970 \times 0\,\farcs841$ &6.9& 0.338 & 0.581 & no & 0.141\ IRAS 21101+5810 &$0\,\farcs686 \times 0\,\farcs590$ &11.1& -0.268 & 0.317 & no & 0.293\ IRAS F23365+3604 &$0\,\farcs098 \times 0\,\farcs091$ &6.8& -1.405 & 0.141 & yes & 0.259 [cccccc]{}\[th\] CGCG 436-030 & 0$\,\farcs072\times0\,\farcs061$ & 36.5 & 16.4& -1.873 & 0.039\ [**IRAS F01364-1042**]{} & 0$\,\farcs101\times0\,\farcs060$ & 44.0 & 59.3& -1.530 & 0.091\ [**III Zw 035**]{} & 0$\,\farcs073\times0\,\farcs062$ & 45.6 & 61.3& -1.476 & 0.054\ VII Zw 031 & 0$\,\farcs119\times0\,\farcs062$ & 24.8 & 0.3& -3.222 & 0.015\ & 0$\,\farcs241\times0\,\farcs180$ & 28.8 & 97.5& -0.831 & 0.255\ & 0$\,\farcs247\times0\,\farcs184$ & 27.7 & 84.3& -0.739 & 0.192\ & 0$\,\farcs259\times0\,\farcs240$ & 27.2 & 74.4& -0.112 & 0.393\ & 0$\,\farcs216\times0\,\farcs189$ & 31.5 & 67.8& -0.615 & 0.196\ NGC 3690 & 0$\,\farcs239\times0\,\farcs218$ & 26.6 & 42.3& 0.312 & 0.175\ & 0$\,\farcs227\times0\,\farcs202$ & 27.6 & 80.4& -0.299 & 0.340\ VV 250 & 0$\,\farcs219\times0\,\farcs202$ & 26.5 & 38.2& -0.305 & 0.249\ UGC 08387 & 0$\,\farcs073\times0\,\farcs051$ & 24.4 & 39.8& -1.301 & 0.060\ & 0$\,\farcs210\times0\,\farcs204$ & 28.4 & 65.5& -0.233 & 0.328\ VV 340a & 0$\,\farcs085\times0\,\farcs065$ & 17.8 & 0.9& -2.959 & 0.013\ VV 705 & 0$\,\farcs059\times0\,\farcs051$ & 20.9 & 17.2& -1.712 & 0.063\ [**IRAS 15250+3609**]{} & 0$\,\farcs058\times0\,\farcs051$ & 25.9 & 73.2& -1.767 & 0.082\ Arp 220 & 0$\,\farcs066\times0\,\farcs052$ & 24.8 & 66.5& -0.589 & 0.106\ IRAS 17132+5313 & 0$\,\farcs060\times0\,\farcs053$ & 24.5 & 10.5& -2.137 & 0.049\ IRAS 19542+1110 & 0$\,\farcs072\times0\,\farcs063$ & 25.7 & 49.4& -1.472 & 0.132\ CGCG 448-020 & 0$\,\farcs073\times0\,\farcs063$ & 27.8 & 23.1& -1.752 & 0.052\ IRAS 21101+5810 & 0$\,\farcs075\times0\,\farcs052$ & 27.0 & 26.9& -1.785 & 0.055\ IRAS F23365+3604 & 0$\,\farcs069\times0\,\farcs062$ & 37.1 & 33.9& -1.818 & 0.088 [lcccccccc]{}\[th\] CGCG 436-030 & 7.47E+00 & 3.12E+04 & 0.73&3.85E+03\[3.51E+03\] & 4.82E+23\[4.40E+23\] & 2.89E+02\[2.64E+02\] & 1.49E+02 & 1.23E+12\ IRAS F01364-1042 & 4.35E+02 & 6.65E+05 &0.20& 1.79E+05\[4.54E+04\] & 2.24E+25\[5.69E+24\] & 7.12E+04\[1.80E+04\] & 6.19E+03 & 5.11E+13\ III Zw 035 & 4.34E+02 & 1.16E+06 &0.23& 1.28E+05\[3.64E+04\] & 1.61E+25\[4.55E+24\] & 7.32E+04\[2.08E+04\] & 7.51E+03 &6.20E+13\ VII Zw 031 & 2.37E-01 & 1.55E+03 &1.65& 3.33E+02\[6.87E+02\] & 4.17E+22\[8.60E+22\] & 3.41E+00\[7.04E+00\] & 5.68E+00 & 4.69E+10\ IRAS 08572+3915 & 3.61E+01 & 3.65E+04 &0.49& 1.24E+04\[7.66E+03\]& 1.55E+24\[9.58E+23\] & 1.70E+03\[1.05E+03\] & 1.54E+03 &1.27E+13\ UGC 04881NE & 2.82E+01 & 2.59E+05 &0.28 &6.61E+04\[2.34E+04\] & 8.27E+24\[2.93E+24\] & 1.34E+04\[4.75E+03\] & 6.76E+02 & 5.58E+12\ UGC 05101 & 8.55E+01 & 4.35E+05 & 0.35 & 3.39E+04\[1.50E+04\] & 4.24E+24\[1.87E+24\] & 4.12E+03\[1.82E+03\] & 8.38E+02 & 6.92E+12\ MCG +07-23-019 & 6.50E+01 & 1.66E+05 & 0.30 & 5.59E+04\[2.09E+04\] & 7.00E+24\[2.62E+24\] & 1.03E+04\[3.86E+03\] & 8.45E+02 & 6.98E+12\ NGC 3690 & 1.24E+01 & 3.37E+04 & 0.86 &2.33E+03\[2.51E+03\] & 2.91E+23\[3.14E+23\] & 1.37E+02\[1.48E+02\] & 1.39E+02 &1.15E+12\ UGC 08058 & 2.08E+03 & 3.16E+06 & 0.30 & 5.44E+04\[2.05E+04\] & 6.80E+24\[2.57E+24\] & 1.18E+04\[4.47E+03\] & 9.52E+03 &7.86E+13\ VV 250a & 9.82E+00 & 2.11E+04 & 0.90 & 2.08E+03\[2.33E+03\] & 2.60E+23\[2.91E+23\] & 1.30E+02\[1.46E+02\] & 1.20E+02 & 9.94E+11\ UGC 08387 & 8.54E+01 & 2.32E+05 & 0.32 & 4.36E+04\[1.77E+04\] & 5.46E+24\[2.22E+24\] & 7.80E+03\[3.17E+03\] & 8.68E+02 &7.17E+12\ UGC 08696 & 8.63E+01 & 2.95E+05 & 0.44 &1.73E+04\[9.58E+03\] & 2.17E+24\[1.20E+24\] & 1.85E+03\[1.02E+03\] & 9.77E+02 &8.06E+12\ VV 340a & 1.41E-01 & 1.31E+03 & 1.63 &3.45E+02\[7.04E+02\] & 4.32E+22\[8.81E+22\] & 3.89E+00\[7.93E+00\] & 3.61E+00 & 2.98E+10\ VV 705 & 2.51E+00 & 7.90E+03 &1.16 & 9.60E+02\[1.39E+03\] & 1.20E+23\[1.74E+23\] & 2.74E+01\[3.97E+01\] & 3.65E+01&3.02E+11\ IRAS 15250+3609 & 8.71E+02 & 1.11E+06 & 0.22 & 1.45E+05\[3.95E+04\] & 1.81E+25\[4.94E+24\] & 6.66E+04\[1.81E+04\] & 1.34E+04 &1.11E+14\ Arp 220 & 4.79E+02 & 1.12E+06 & 0.31 & 4.89E+04\[1.91E+04\] & 6.12E+24\[2.39E+24\] & 1.40E+04\[5.46E+03\] & 7.16E+03 &5.91E+13\ IRAS 17132+5313 & 6.46E-01 & 2.52E+03 & 1.54 & 4.12E+02\[7.92E+02\] & 5.16E+22\[9.92E+22\] & 7.20E+00\[1.38E+01\] & 1.47E+01 &1.21E+11\ IRAS 19542+1110 & 4.52E+01 & 1.63E+05 & 0.41 & 2.21E+04\[1.13E+04\] & 2.76E+24\[1.41E+24\] & 2.77E+03\[1.41E+03\] & 1.12E+03 & 9.26E+12\ CGCG 448-020 & 1.59E+00 & 5.20E+03 & 1.52& 4.24E+02\[8.07E+02\] & 5.30E+22\[1.01E+23\] & 1.30E+01\[2.47E+01\] & 3.52E+01 & 2.90E+11\ IRAS 21101+5810 & 4.75E+00 & 1.28E+04 & 0.89 & 2.14E+03\[2.37E+03\] & 2.67E+23\[2.97E+23\] & 1.20E+02\[1.33E+02\] & 1.06E+02 & 8.72E+11\ IRAS F23365+3604 & 4.77E+01 & 1.95E+05 & 0.33 & 4.15E+04\[1.71E+04\] & 5.20E+24\[2.15E+24\] & 5.24E+03\[2.16E+03\] & 1.40E+03 &1.15E+13 [cccc]{}\[th\] 0.8 & U/LIRGs only (this paper) & 1.02 $\pm$ 0.10 & -1.33 $\pm$ 0.47\ 4.35 & U/LIRGs only (this paper) & 1.02 $\pm$ 0.11 & -2.08 $\pm$ 0.55\ $\propto$ $\Sigma_{gas}^{-0.5}$ & U/LIRGs only (this paper) & 1.52 $\pm$ 0.16 & -3.09 $\pm$ 0.66\ 4.35 & U/LIRGs (this paper) + kpc-size regions [@Leroy13] & 1.35 $\pm$ 0.04 & -3.85 $\pm$ 0.13\ 0.8 & U/LIRGs (this paper) + kpc-size regions [@Leroy13] & 1.63 $\pm$ 0.07 & -4.18 $\pm$ 0.22\ $\propto$ $\Sigma_{gas}^{-0.5}$ & U/LIRGs only (this paper) + kpc-size regions [@Leroy13]& 1.87 $\pm$ 0.06 & -4.63 $\pm$ 0.19\ 4.35 & kpc-size regions [@Leroy13]& 1.06 $\pm$ 0.02 & -3.48 $\pm$ 0.03 Notes on the sources ==================== [**CGCG 436-030**]{}: This system has two well separated components (east and west), however we only detected the western component at 33 GHz. [**CGCG 448-020**]{}: This is an interacting system showing a complex morphology. It is still not clear whether there are two or more systems interacting. It hosts an off-nuclear starburst (north-east component in Figure \[fig:fig1\]) which contributes $\sim$ 80% of the total infrared luminosity of the galaxy at infrared wavelengths (Stierwalt et al. in prep). In the final map (i.e., the one with the highest resolution, not shown in this work), this off-nuclear starburst is still only partially resolved, even at 0$\,\farcs$08, while the more extended component (south-west) is resolved out. [**III Zw 035**]{}: This galaxy has the most compact 33 GHz continuum emission in the sample. [**IRAS 08572+3915**]{}: We only detect the north west (NW) component of this system. The NW component is optically classified as a Seyfert 2 galaxy and is suspected to have a highly obscured AGN[e.g., @Iwasawa11; @Rupke13]. The flat spectrum observed in Figure \[fig:fig2\] suggests this is a flat spectrum AGN, which was also suggested by @Condon91 based on 1.49 and 8.44 GHz continuum observations. The high thermal fraction predicted from the IR luminosity only indicates that the IR emission is mostly dominated by an AGN instead of star formation, and that the 33 GHz emission is dominated by synchrotron instead. [**IRAS 15250+3608**]{}: This systems is one of the sources emitting at, or close to, the Eddington limit. The optical and mid-IR diagnostics classify this galaxy as a composite source. The fact that it is close to the Eddington limit, agrees with the potential coexistence of an AGN and a strong starburst. [**IRAS 17132+5313**]{}: This system has two components. The galaxy towards the north east is extended and resolved out in the highest resolution image (0$\,\farcs$08 $\times$ 0$\,\farcs$07). We had to taper the map in order to recover its emission. The galaxy towards the south west is compact and contributes $\sim$ 40% of the integrated flux density of the system. [**NGC 3690**]{}: This clearly interacting system consists of multiple components. Two of the components them are associated with the nuclei of the progenitors, NGC 3690E (east) and NGC 3690W (west), while the others are a combination of off nuclear starbursts. The strongest nucleus (NGC 3690E) has been observed with VLBI. At least 30 point sources have been found plus a potential AGN [e.g., @Neff04]. Due to the proximity of this system and its spatial extent, the 33 GHz emission is resolved at the D configuration resolution ($\sim$ 2"), clearly showing 5 components (see white crosses in map from Figure \[fig:fig1\]). In order to measure its total flux density, we tapered the D configuration map until the system showed 2 unresolved components (east and west systems). We then proceed as explained in Section \[sec:Iflux\], by fitting a Gaussian to each one. [**UGC 04881**]{}: This system has two components and its total flux density was recovered by adding the Gaussian fit results of each component separately. The error of this measurement was obtained by adding in quadrature the errors associated to each component (see Section \[sec:Iflux\]). The D configuration map of this system had low signal-to-noise ratio and the quality was not good enough to recover the total flux density. For this reason, we used the final image (with the different array configurations combined, see Section \[sec:red\]) tapered such that we recovered a point-like structure for each component. Even though both components contribute about the same to the total flux density observed at 32.5 GHz, the brightest component (north-east) is more compact. The south-west component is resolved out at the highest resolution we can achieve. We measured the size of this component from the image we used to obtain the total flux density ($\mathrm{A_{beam}=20.3~arcsec^{2}}$) and found an upper limit of $\mathrm{A_{50}=19.8~arcsec^{2}}$, i.e., it is unresolved in this coarser map. The brightest component is shown in Figure \[fig:fig1\]. The flux density of this bright component should also be treated as a lower limit. The potential calibration issues mentioned before could very well be originating the abnormally high thermal fractions observed in Figure \[fig:fig2\]. [**UGC 08058**]{}: This is the most powerful IR source in our sample. It is known to host an AGN [e.g., @Lonsdale03; @Iwasawa09] and potentially represents the stage before becoming an elliptical galaxy according to the evolutionary model proposed by @Sanders88. [**Arp 220**]{}: This is the closest ULIRG in the local universe. This galaxy shows extreme dust opacities and very compact nuclear disks. We present a detailed analysis of the 33 and 6 GHz emission from this galaxy in @BM15, where we find that the disks are better described by exponential disks, rather than Gaussian. The 33 GHz map reported in @BM15 is slightly different to the one presented here since the imaging procedures differ, however the flux density measured here and the morphology are in agreement with those shown in @BM15. [**VII Zw 031**]{}: At the highest resolution image (0$\,\farcs$8 $\times$ 0$\,\farcs$6, done with natural weighting) the emission was completely resolved out. We had to taper the image heavily in order to recover the emission. This is one of the most extended systems in our sample along with VV 340a. [**VV 250**]{}: This system has two well separated components, south-east (VV 250a) and north-west (VV 250b). In Figure \[fig:fig1\], we only show VV 250a since it contains $\sim$86% of the total flux density of the system (obtained by adding the flux density of both components). The north-west component is faint with an 11$\sigma$ peak detection. To recover A$_{50}$ for this faint component, we used the tapered D array map ($\sim$ 10" resolution) since we could not recover half of the integrated flux density of this component in higher resolution maps. Even in this low resolution map, we recover A$_{50}$ for $\mathrm{C_{50}=3.2\sigma}$, which is lower than our conservative limit of 5$\sigma$, however we inspected this contour and made sure the emission within it looked real. For the north-west component, $\mathrm{A_{50}=64.8~arcsec^{2}}$ in a map with $\mathrm{A_{beam}=113.4~arcsec^{2}}$, i.e., it is unresolved, and then A$_{50}$ is only an upper limit. [**VV 340a**]{}: In the final combined image, where we achieved an angular resolution of 0$\,\farcs$5 $\times$ 0$\,\farcs$4 (using natural weighting), the emission from this system was completely resolved out. To recover the extended emission, we had to taper the image heavily. VV 340 has two components, an edge on galaxy to the north (VV 340a), shown in Figure \[fig:fig1\], and a face on galaxy to the south (VV 340b). Inconveniently, the pointing of the VLA observation was centered on VV 340b, from which we tentatively detected an off nuclear feature at a $\sim 3 \sigma$ level in our lowest resolution image. The bright edge on galaxy, VV 340a, is clearly detected, although it was hard to perform the Gaussian fit since the source fell close to the edge of the primary beam. [**VV 705**]{}: This system shows two nuclei in Figure \[fig:fig1\], north-west and south-east. In the D configuration map they are indistinguishable. [^1]: Throughout the paper we use 33 and 32.5 GHz interchangeably, however for calculation/estimation purposes we use $\nu = 32.5$ GHz. [^2]: For UGC 04881NE, $\alpha_{6-33}\approx -1.2$ which is unusually steep. We also consider its flux density at 33 GHz as a lower limit. [^3]: $\sigma_{A~(or~B)}$ is the rms noise of the A (or B) array configuration image. [^4]: In order to obtain values that are comparable to those in the literature, we use SFR = SFR$_{\rm IR}$ to derive $\mathrm{\Sigma^{33GHz}_{SFR_{IR}}}$. [^5]: We normalized their sizes to the scale we use in this paper (see Table \[table:tbl-1\]), defined in the same way we define A$_{\rm beam}$. [^6]: Full list of Herschel images presented in @Chu17. [^7]: This is the correction to obtain the mass inside a sphere of radius R$_{50,d}$ (see Section \[sec:sizes\]) with a Gaussian mass distribution. [^8]: This turnover frequency normally occurs at MHz frequencies, when present, and it shifts to higher frequencies for high star forming, very compact systems. [^9]: This omits the upper limits obtained for the faint components in the systems UGC 04881 and VV250, for which we did not derive the physical parameters described in section \[sec:impl\_radio\]. [^10]: We used the `scipy.optimize.curve_fit` algorithm and a function of the form $\mathrm{Y = slope~X + coefficient}$ to obtain the slope and coefficient, and their standard deviation errors. We excluded sources with upper limits to their sizes.

ArXiv

--- abstract: 'Nonlinear saturation of beta induced Alfvén eigenmode, driven by slowing down energetic particles via transit resonance, is investigated by the nonlinear hybrid magnetohyrodynamic gyro-kinetic code (XHMGC). Saturation is characterized by frequency chirping and symmetry breaking between co- and counter-passing particles, which can be understood as the the evidence of resonance-detuning. The scaling of the saturation amplitude with the growth rate is also demonstrated to be consistent with radial resonance detuning due to the radial non-uniformity and mode structure.' author: - 'X. Wang$^1$, S. Briguglio$^2$, L. Chen$^{1,3}$, G. Fogaccia$^2$, C. Di Troia$^2$, G. Vlad$^2$, F. Zonca$^{2,1}$' title: Nonlinear dynamics of beta induced Alfvén eigenmode driven by energetic particles --- Apart from Toroidal Alfvén Eigenmodes [@cheng85] (TAE), the existence of Beta induced Alfvén Eigenmodes [@heidbrink93] (BAE) in the kinetic thermal ion (KTI) [@zonca96a; @chen07] gap of the shear Alfvén wave (SAW) continuous spectrum is also widely recognized to be a concern for the good confinement of energetic particles (EP) in fusion plasmas [@heidbrink93; @heidbrink95]. BAEs are particularly important in the study of low frequency fluctuations of the shear Alfvén wave (SAW) spectrum, since they can be excited by both fast ions (at long wavelengths) as well as by thermal ions (at short wavelengths) [@zonca99; @nazikian06]. In addition to Alfvén eigenmodes (AE), which are normal modes of the thermal plasma, strongly driven energetic particle continuum modes [@chen94] (EPM) may be also excited at the characteristic frequencies of the EPs, in the presence of a sufficiently intense EP free energy source. The different nature of AEs and EPMs suggests the existence of two regimes in the nonlinear dynamic evolution of a single toroidal mode number ($n$) coherent SAW driven by EPs. Near marginal stability, when the system is weakly driven and EPMs are not excited, the nonlinear dynamics is the same as in a uniform system and wave saturation can occur when wave-particle trapping flattens the particle distribution in the resonance region [@berk90]. Meanwhile, other physical mechanisms, such as Compton scattering off the thermal ions [@hahm95] and mode-mode couplings, enhancing the interaction with the SAW continuous spectrum by nonlinear frequency shift [@zonca95; @chen98], can also be important, depending on the parameter regimes. All these phenomena are local and independent of the radial mode structure. However, when the system is strongly driven and EPMs [@chen94] are excited at the EP characteristic frequencies, strong EP transport occurs in avalanches [@zonca05]. This phenomenology is strictly related with the resonant character of the modes, which tend to be radially localized where the drive is strongest, and with global readjustments in the EP radial profiles. In the strong nonlinear regime, local and global effects must be treated at the same footing in order to capture the crucial role of radial non-uniformity. In this letter, we investigate the transition regime in between the two limiting conditions described above, in order to illustrate the effect of equilibrium non-uniformity as the mode drive is increased above marginal stability. In particular, we analyze AEs resonant excitation below the EPM threshold and choose BAE as specific case because: (1) rich phenomena are observed in experiments [@heidbrink93; @heidbrink95]; (2) a single poloidal harmonic is dominant; (3) the mode is very localized, so that finite mode width effects and finite interaction length are easily illustrated. Here, we focus on EP nonlinear dynamic behaviors and their consequences on BAE saturation. Thermal ion nonlinear dynamics may also be important, but it will be analyzed in a separate paper. In our work, BAE driven by EPs are studied with the extended version of nonlinear magnetohydrodynamic (MHD)-Gyrokinetic code HMGC [@briguglio95] (XHMGC). These extensions include both thermal ion compressibility and diamagnetic effects [@wang11], in order to account for thermal ion collisionless response to low-frequency Alfvénic modes driven by EPs (e.g., kinetic BAE (KBAE) [@wang10]) and finite parallel electric field due to parallel thermal electron pressure gradient. Previous linear simulation studies show that BAE/KBAE can be destabilized by EPs via resonant wave particle interactions [@zhang10; @wang10]. Here, we ignore diamagnetic effects of kinetic thermal ions by assuming uniform thermal ion density and temperature profiles, but keep kinetic thermal ion compression effects, in order to correctly describe the KTI gap. Meanwhile, we consider nonlinear wave-particle interactions with EPs, but neglect both mode-mode coupling effects and nonlinear kinetic thermal ion response. The purpose is to isolate the nonlinear physics due to EPs, which do play dominant roles for sufficiently strong drive. We investigate a BAE mode localized around the rational surface at $(r/a)\approx 0.5$ with safety factor $q=2$, dominated by toroidal mode number $n=2$ and poloidal mode number $m=4$ harmonic. The BAE-SAW continuum accumulation point frequency can be calculated from kinetic theory [@zonca96a] and is given by $\omega_{cap}=0.127/\tau_A$, where $\tau_A=R_0/V_A$ is the Alfvén time, with $V_A$ the Alfvén speed on axis and $R_0$ the major radius. The mode is driven by EPs with an isotropic slowing-down distribution and characteristic transit frequency at the birth energy $\omega_{tmax}=(2E_0/m_H)^{1/2}/(qR_0)\simeq0.2/\tau_A$. The mode is excited at $\omega_0\simeq 0.114/\tau_A$, within the KTI gap, and the linear growth rate is $\gamma_L\simeq 0.006/\tau_A$. As the mode grows and saturates, the mode frequency is chirping up quickly, as shown in Fig. \[fig:omega\], because of EP radial redistribution and the corresponding nonlinear change in the BAE dielectric response. To show this, we use the general fishbone like dispersion relation, $i\Lambda(\omega)=\delta W_f+\delta W_k$ [@chen07; @zonca99], with the generalized inertia term $i\Lambda\simeq-\sqrt{\omega^2_{cap}-\omega^2}/\omega_A$ for BAE [@zonca96a], $\omega_A=v_A/(qR_0)$ and $v_A$ being the Alfvén frequency and speed, respectively, and $\delta W_f$ and $\delta W_k$ representing fluid and kinetic contributions to the potential energy [@chen07; @zonca99]. This allows us to write the nonlinear frequency shift as $$\begin{aligned} \label{eq:fldr} \left(\frac{\omega-\omega_0}{\omega_A}\right)\bigg [ \frac{\omega_0}{\sqrt{\omega^2_{cap}-\omega_0^2}} -\omega_A\frac{\partial}{\partial\omega_0}\mathbb{R}e\delta W_{kL}\nonumber\\ -i\omega_A\frac{\partial}{\partial\omega_0}\mathbb{I}m\delta W_{kL}\bigg ]=\mathbb{R}e\delta W_{kNL}+i\mathbb{I}m\delta W_{kNL} \ \ ,\end{aligned}$$ having separated linear (L) and nonlinear (NL) contributions to the potential energy. Equation (\[eq:fldr\]) shows that upward frequency chirping is expected, due to EP radial redistributions $(\mathbb{R}e\delta W_{kNL}>0, \mathbb{I}m\delta W_{kNL}<0$ [@chen07; @zonca99]). This result still holds when $|\mathbb{R}e\delta W_{kNL}|\ll|\mathbb{I}m\delta W_{kNL}|$, since – in the present case of slowing down EPs – $\partial_{\omega_0}\mathbb{I}m\delta W_{kL}>0$. ![\[fig:omega\] Upper panel shows the amplitude evolution (black line) and frequency evolution of the mode (blue line), given by the maximum in the wavelet transform spectrum. Lower panel shows the wavelet transform spectrum for the frequency evolution.](frequency_iso.eps){width="90mm" height="65mm"} ![\[fig:power\] Wave-particle power exchange in the EP $(\mu, v_\parallel)$ plane during linear phase (upper panel) and saturation phase (lower panel), where $\mu$ is the magnetic moment ($v_\parallel$ and $\mu$ are normalized w.r.t. $\sqrt{E_0/m_H}$ and $E_0/\Omega_0$ with $\Omega_0=eB_0/(m_Hc)$, respectively). The positive sign in the color bar corresponds to ions pumping energy into the wave. Solid and dashed lines are passing-trapped boundaries, calculated for the inner and outer limits of the mode radial width.](power.eps){width="75mm" height="70mm"} Due to the preferred direction in mode frequency chirping, the symmetry between co-passing ($v_\parallel>0$) and counter-passing ($v_\parallel<0$) particles is also broken, as visible in Fig. \[fig:power\], showing the wave-particle power exchange in both linear (upper panel) and saturation phase (lower panel). Results show that the interaction is predominantly driven by transit resonance. Both co-passing and counter-passing particles are giving energies to the wave in the linear phase, with almost symmetric response in $v_\parallel$ space. Different fine structures in the resonant region are due to different orbits of co- and counter-passing particles. However, in the saturation phase, co-passing particles are still driving the mode, while counter-passing particles do not. This is evidence of resonance detuning, explained as follows. The passing particle resonance condition is $$\label{eq:rc} \omega=\ell\omega_t+(n\bar{q}-m)\sigma\omega_t\ \ ,$$ where $\sigma=sgn(v_\parallel)$, $\bar{q}$ is the orbit average of $q$, $\bar{q}=(2\pi)^{-1}\oint qd\theta$ and the transit frequency is defined as $\omega_t=2\pi/\oint d\theta/\dot{\theta}$. Here we have used Clebsch toroidal flux coordinates $(\psi, \theta, \xi)$, with the equilibrium magnetic field ${\bf B}=\nabla\psi\times\nabla\xi$, $\psi$ the magnetic flux and the poloidal angle $\theta$ describing the position along ${\bf B}$. The resonance condition, Eq. (\[eq:rc\]), can be expressed as $\dot{\Theta}=0$; i.e., stationarity in the wave-particle phase. For the transit resonance under investigation, we can define the resonance detuning as $$\label{eq:detuning} \Delta\dot{\Theta}\simeq(n\bar{q}_r\sigma+\frac{\omega_0}{\omega^2_t}\partial_r\omega_t)\omega_t\Delta r-(\omega-\omega_0)\ \ ,$$ where $\bar{q}_r=d\bar{q}/dr$ and $\omega_0=\ell\omega_t+(n\bar{q}-m)\sigma\omega_t$. In the present condition, the resonance-detuning, due to radial particle displacement, is dominated by the first term in parenthesis. Thus, for $\dot{\omega}>0$ (upward frequency chirping), particles that are transported out, while transferring energy to the wave, more easily maintain the resonance condition for $v_\parallel>0$ (co-passing) than $v_\parallel<0$ (counter-passing). This explains the results of Fig. \[fig:power\], as well as of Fig. \[fig:density\], where the relative magnetic flux averaged EP fluctuating distribution function, $\delta n/n_0$, is shown vs $(r/a)$, after integration in $\mu$ and over a narrow interval $\Delta v_\parallel$ near resonance. From these results, it is clear that the wave interaction with co-passing particles is lasting longer and covering a broader radial domain, of the order of the radial mode width. ![\[fig:density\] Radial dependence of the magnetic flux averaged EP fluctuating distribution, integrated over $\mu$ and on a narrow interval $\Delta v_\parallel$ near resonance. Upper panel shows the result for co-passing particles. Lower panel shows the result for counter-passing particles. The radial mode structure with $n=2$ and $m=4$ is shown in the background for reference.](density.eps){width="70mm" height="54mm"} Equation (\[eq:detuning\]) also provides useful insights for the analysis of phase-space structures shown in Fig. \[fig:phase\]. There, test particle motion in the phase-space is illustrated for both co- and counter-passing particles. For $v_\parallel>0$, the wave-particle phase $\Theta$ is essentially constant, i.e. the $\dot{\Theta}\simeq0$ resonance condition is effectively maintained through saturation. This occurs because of Eq. (\[eq:fldr\]), showing that frequency chirps up due to the effect of EP radial transport. Frequency chirping, in turn, is connected with phase-locking $(\dot{\Theta}\simeq0)$, for this is the condition that allows particles to be most efficiently transported outward, while driving the mode. This condition is similar to that, which underlies EP transports in avalanches [@zonca05], due to EPM [@chen94]. In the present case, however, the finite BAE radial structure determines the finite interaction length and mode saturation is reached when particles are pumped out of the effective mode radial domain. For counter-passing particles, meanwhile, Fig. \[fig:phase\] shows rapid wave-particle phase mixing, as expected from Eq. (\[eq:detuning\]), and the corresponding formation of phase-space island structures near resonance, which explain the reduced drive for $v_\parallel<0$ in Fig. \[fig:power\] [@oneil65; @mazitov65]. This result, which crucially depends on radial non-uniformity of the system and applies for $|\omega-\omega_0|\gtrsim\gamma_L$, is different from the well-known case of nonlinear saturation due to wave-particle trapping [@onishchenko70; @oneil71; @shapiro71; @berk92]. It is also different from the adiabatic frequency chirping of hole-clump structures in phase-space [@berk97], since, in the present case, $|\dot{\omega}|\gtrsim\gamma^2_L$ (c.f. Eq. (\[eq:detuning\])) and the frequency sweeping is non-adiabatic [@chen07; @zonca99]. This is a crucial element for maximizing EP transport, for a broader region of the phase-space can be affected by radial redistribution, as shown in Fig. \[fig:density\] and \[fig:phase\] by the $v_\parallel$ symmetry breaking. Meanwhile, considering Eq. (\[eq:fldr\]), this explains why frequency chirping is connected with the nonlinear evolution of the mode amplitude, as shown in Fig. \[fig:omega\] and often observed in experiments [@podesta11]. ![\[fig:phase\] Phase space structures in $(\Theta, -P_\phi)$ ($\Theta$ and $P_\phi$ are normalized w.r.t. $2\pi$ and $a\sqrt{mE_0}$ with $a$ the minor radius) phase-space for co-passing particles (upper panel) and counter-passing particles (lower panel) at saturation, where $\Theta$ is sampled at $\theta=0$.](1729co.eps "fig:"){width="75mm" height="35mm"} ![\[fig:phase\] Phase space structures in $(\Theta, -P_\phi)$ ($\Theta$ and $P_\phi$ are normalized w.r.t. $2\pi$ and $a\sqrt{mE_0}$ with $a$ the minor radius) phase-space for co-passing particles (upper panel) and counter-passing particles (lower panel) at saturation, where $\Theta$ is sampled at $\theta=0$.](1729_counter.eps "fig:"){width="75mm" height="35mm"} This interpretation of simulation results is further supported by the scaling of the saturation amplitude with the growth rate, which is shown in Fig. \[fig:amp\] and can be understood theoretically by analyzing the nonlinear pendulum about the stable fixed point by radially localized perturbation. The corresponding wave-particle resonance is then described by $$\begin{aligned} \label{eq:pendulum1} dY/d\tau &=& X \nonumber \\ dX/d\tau &=& -A_0Y/(1+X^2) \ \ .\end{aligned}$$ Here, $X=|k_r|x/(nq')$ and $Y=y|k_r|/(nq')$ are the rescaled variables for $x = nq'\Delta r$, $y = \Theta - \Theta_0$, $\Theta_0$ begin the fixed point and $\tau=\omega_t t$. We assume $A=A_0/(1+X^2)$ to investigate the effect of finite characteristic radial mode width, i.e. $k^{-1}_r$, where $A_0$ is in general proportional to the mode saturation amplitude. Equation (\[eq:pendulum1\]) is separable and can be trivially integrated by quadratures. As a result, the period is given by ${\cal T} = 2\pi A_0^{-1/2}/\Gamma$, with $$\label{eq:int} \sqrt{A_0} \Gamma = \left( \int^1_0\frac{(2/\pi)(1+z^2X_0^2)dz}{\sqrt{(1+X^2_0/2)-z^2(1+z^2X^2_0/2)}} \right)^{-1} .$$ The integral in Eq. (\[eq:int\]) is a function of $X_0$, the normalized maximum radial particle excursion, and is easily computed by asymptotic expansion, yielding $1$ for $X_0\rightarrow 0$ and $\propto X_0$ for $X_0\gg1$. The characteristic rate for a particle to undergo a radial excursion $X_0=|k_r|\Delta r_0$, is then given by $\gamma_{NL} =A_0^{1/2} \omega_t \Gamma(X_0)$. If $X_0=|k_r|\Delta r_0\ll1$, as assumed in the nonlinear saturation of TAE modes [@berk90], the nonlinear dynamics is the same as in a uniform system and wave saturation can occur only when the particle distribution function in the resonance region is flattened, so that the nonlinear drive is significantly reduced and brought back to threshold. This occurs for $\gamma_{NL}/\omega_t \sim \gamma_L/\omega_t\sim A_0^{1/2}$, yielding the well known estimate $A_0\propto(\gamma_L/\omega_t)^2$ [@onishchenko70; @oneil71; @shapiro71]. When $X_0 = |k_r|\Delta r_0 \gtrsim 1$, saturation occurs because particle get off resonance by radial detuning, i.e., after a radial displacement of the order of the mode width, as noted above and observed in [@briguglio98] for the first time. Similar to the uniform case, saturation is expected when $\gamma_{NL}/\omega_t = A_0^{1/2} \Gamma(X_0) \propto A^{1/2}_0/X_0\approx\gamma_L/\omega_t$. Meanwhile, noting that $Y \sim |k_r|/(nq')$, Eqs. (\[eq:pendulum1\]) also give $X_0 \propto (A_0^{1/2}/X_0)|k_r|/(nq')$. Thus, the saturation condition becomes $A_0\propto(\gamma_L/\omega)^4(k_r)^2/(nq')^2$. By further increasing the drive, the radial scale $1/(nq')$ of the BAE mode will become increasingly more important, till the saturation amplitude is independent of $(\gamma_L/\omega)$. For BAE modes, the transition between these two regimes is expected to occur for $X_0\simeq 1$, i.e. for $A_0\approx(nq')^2/k^2_r$, where the radial wave number estimate for the BAE short scale radial structure is $|k_r|/(nq')\propto(\delta\omega/\omega)^{-1/2}$ [@zonca96a], i.e. $|k_r|/(nq')\propto(\gamma_L/\omega)^{-1/2}$ for $\gamma_L\gtrsim|\mathbb{R}e\delta\omega|$. The expected scaling is then given by $A_0\propto(\gamma_L/\omega)^3$. Figure \[fig:amp\] illustrates the transition between these various regimes as $\gamma_L/\omega$ is increased, further demonstrating the important role of plasma non-uniformity and radial mode structures for mode saturation and EP transport processes. ![\[fig:amp\] The saturation amplitude of the fluctuating scalar potential vs. the linear growth rate at different values of EP density. The “$\circ$” are simulation results. The black, blue and red lines are, respectively, denoting $\propto(\gamma/\omega)^2, (\gamma/\omega)^3, (\gamma/\omega)^4$ for references. ](amp.eps){width="80mm" height="49mm"} In summary, we have found that nonlinear saturation of BAEs, driven by slowing down EPs via transit resonance, is characterized by upward frequency chirping and $v_\parallel$ symmetry breaking between co- and counter-passing particles, which can be understood as the the evidence of resonance-detuning. Upward frequency chirping is the preferred condition for maximizing EP radial transport while they drive the mode, and saturation eventually occurs when the particle radial displacement is of the order of the radial mode width. The scaling of the saturation amplitude with the growth rate further demonstrates the role of radial resonance detuning, due to the radial non-uniformity and mode structure. Both EP phase space behaviors as well as nonlinear dynamics and mode structures demonstrate the crucial roles of non-uniformities and geometry of the system. This work was supported by the National ITER Program of China and the Euratom Communities under the contract of Association between EURATOM/ENEA. One of the authors, Xin Wang, would like to thank R.B. White and H. Zhang for the valuable suggestions in the analysis of phase space structures; Liu Chen also acknowledges support of U.S. DoE Grant. Useful discussions with Z. Lin are also kindly acknowledged. [99]{} C.Z. Cheng, L.Chen and M.S. Chance, Ann. Phys. [**161**]{} 21 (1985). W.W. Heidbrink [*et al.*]{}, Phys. Rev. Lett. [**71**]{} 855 (1993). F. Zonca [*et al.*]{} Plasma Phys. Control. Fusion [**38**]{} 2011 (1996). L. Chen and F. Zonca Nucl. Fusion [**47**]{} S727 (2007). W.W. Heidbrink and E.M. Carolipio, Nucl. Fusion [**35**]{} 1481 (1995). F. Zonca [*et al.*]{}, Phys. Plasmas [**6**]{} 1917 (1999). R. Nazikian [*et al.*]{}, Phys. Rev. Lett. [**96**]{} 105006 (2006). L. Chen, Phys. Plasmas [**1**]{} 1519 (1994). H.L. Berk and B.N. Breizman, Phys. Fluids B [**2**]{} 2246 (1990). T.S. Hahm and L. Chen, Phys, Rev. Lett. [**74**]{} 266 (1995). F. Zonca [*et al.*]{} Phys. Rev. Lett. [**74**]{} 698 (1995). L. Chen [*et al.*]{} Plasmas Phys. Control. Fusion [**40**]{} 1823 (1998). F. Zonca [*et al.*]{}, Nucl. Fusion [**45**]{} 477 (2005). S. Briguglio [*et al.*]{}, Phys. Plasmas [**2**]{} 3711 (1995). X. Wang [*et al.*]{}, Phys. Plasmas [**18**]{} 052504 (2011). X. Wang [*et al.*]{}, Plasma Phys. Control. Fusion [**52**]{} 115005 (2010). H.S. Zhang [*et al.*]{}, Phys. Plasmas [**17**]{} 112505 (2010). T.M. O’Neil, Phys. Fluids [**8**]{} 2255 (1965). R.K. Mazitov, Zh. Prikl. Mekh. Tekh. Fiz. [**1**]{} 27 (1965). I.N. Onishchenko [*et al.*]{}, Zh. Eksp. Teor. Fiz. Pisma Red. [**12**]{} 407 (1970) \[JETP Lett. [**12**]{} 281 (1970)\]. T.M. O’Neil [*et al.*]{}, Phys. Fluids [**14**]{} 1204 (1971). V.D. Shapiro and V.I. Shevchenko, Zh. Eksp. Teor. Fiz. [**60**]{} 1023 (1971) \[Sov. Phys. JETP [**33**]{} 555 (1971)\]. H.L. Berk, B.N. Breizman and H. Ye, Phys. Rev. Lett. [**68**]{} 3563 (1992). H.L. Berk, B.N. Breizman and N.V. Petiashivili, Phys. Lett. A [**234**]{} 213 (1997). M. Podestà [*et al.*]{}, Nucl. Fusion [**51**]{} 063035 (2011). S. Briguglio [*et al.*]{} Phys. Plasmas [**5**]{} 3287 (1998).

ArXiv

--- abstract: | We developed the generalized tight-binding model to study the magneto-electronic properties of AAB-stacked trilayer graphene. Three groups of Landau levels (LLs) are characterized by the dominating subenvelope function on distinct sublattices. Each LL group could be further divided into two sub-groups in which the wavefunctions are, respectively, localized at 2/6 (5/6) and 4/6 (1/6) of the total length of the enlarged unit cell. The unoccupied conduction and the occupied valence LLs in each sub-group behave similarly. For the first group, there exist certain important differences between the two sub-groups, including the LL energy spacings, quantum numbers, spatial distributions of the LL wavefunctions, and the field-dependent energy spectra. The LL crossings and anticrossings occur frequently in each sub-group during the variation of field strengths, which thus leads to the very complex energy spectra and the seriously distorted wavefunctions. Also, the density of states (DOS) exhibits rich symmetric peak structures. The predicted results could be directly examined by experimental measurements. The magnetic quantization is quite different among the AAB-, AAA-, ABA-, and ABC-stacked configurations. 1.0 truecm *Keywords*: trilayer graphene; Landau level; magnetic field; energy spectra; anticrossings 1.0 truecm Corresponding author. [  Tel:  +886-6-275-7575.]{}\  [[*E-mail addresses*]{}: [email protected] (C.Y. Lin), [email protected] (M.F. Lin)]{} author: - | Thi-Nga Do $^{a}$, Chiun-Yan Lin$^{a,*}$,Yi-Ping Lin $^{a,*}$, Po-Hsin Shih $^{a,*}$, Ming-Fa Lin$^{a,*}$ $$\ $^a$Department of Physics, National Cheng Kung University, Tainan, Taiwan 701\ title: 'Configuration-enriched magnetoelectronic spectra of AAB-stacked trilayer graphene\' --- [**1. Introduction**]{} 0.3 truecm Graphene, being a one-atom-thick layer of carbon atoms densely packed in a two-dimensional honeycomb lattice, has attracted a lot of theoretical and experimental research [@MKoshino; @ADMartino; @JSari; @YHHo; @MMKruczynski; @YPLin; @YKHuang; @SHRSena; @AKumar; @YBZhang; @CLKane; @KSNovoselov; @KSNovoselov2; @VPGusynin; @FGuinea; @XDu; @JWJiang; @MMShokrieh; @JULee; @QLu; @CLee]. It exhibits many unusual physical properties, e.g., a rich magnetic quantization [@MKoshino; @ADMartino; @JSari; @YHHo; @MMKruczynski; @YPLin; @YKHuang; @SHRSena], half-integer Hall effect [@AKumar; @YBZhang; @CLKane; @KSNovoselov; @KSNovoselov2; @VPGusynin; @FGuinea; @XDu], high Young’s modulus [@JWJiang; @MMShokrieh; @JULee; @QLu; @CLee], high Fermi velocity ($10^6$ m/s), and others. Graphene could play an important role in technological applications such as electric circuits [@LHuang; @YMLin], field-effect transistors [@IMeric; @FNXia], light-emitting diodes [@JBWu; @THHan; @GHJo], solar cells [@JBWu2; @JDRoyMayhew; @NLYang; @WJHong], and durable touch screens [@SKBae; @JWang]. Due to the hexagonal symmetry with a rotational angle of 60$^\circ$, a non-doped graphene is a zero-gap semiconductor with a vanishing density of state at the Fermi level. The essential electronic properties can be drastically changed by the layer number [@HHibino; @EHMFerreira; @AHCNeto], stacking configuration [@AHCNeto; @CYLin; @JSLee; @MAoki; @FGuinea2; @KFMak], magnetic field [@YZhang; @MOGoerbig], electric field [@CLLu; @EVCastro; @KFMak2], dopping [@DMBasko; @CCasiraghi], mechanical strain [@JELee; @SMChoi; @JHWong], and temperature variation [@YWTan; @VVCheianov]. Few- and multi-layer graphenes have been successfully produced by experimental methods such as exfoliation of highly orientated pyrolytic graphite [@MJWebb; @ZLiu; @ZYRong; @JMCampanera], metalorganic chemical vapour deposition (MOCVD) [@CTEllis; @JHHwang; @ZYJuang; @DRLenski; @BJayasena; @CHLui], chemical and electrochemical reduction of graphene oxide [@KFMak3; @LYZhang; @LBBiedermann], and arc discharge [@YPWu; @ZSWu]. There exist important stacking configurations, including AAB [@ZYRong; @JMCampanera; @LBBiedermann], ABC [@KSNovoselov3; @BLalmi; @CHLui; @KFMak3; @LYZhang; @LBBiedermann], AAA [@JHHwang; @ZYJuang], ABA [@BLalmi; @CTEllis; @CHLui; @LBBiedermann], and twisted [@JHHwang; @LBBiedermann] and turbostratic ones [@DRLenski]. The interlayer atomic interactions and stacking configurations induce the rich electronic properties of graphene. In this work, we investigate the complex relationship between the magnetic quantization and the interlayer atomic interactions in AAB-stacked trilayer graphene by using the generalized tight-binding model. The low-lying energy dispersions of monolayer graphene possess a pair of linear bands intersecting at $E_F$. These energy bands are further quantized by the application a uniform perpendicular magnetic field $B=B_0\hat{z}$ [@SWu]. The magneto-electronic spectrum of the isotropic Dirac cones satisfies a simple $E^{c,v}\propto \pm\sqrt{nB_0}$ relationship, where $n$ is the quantum number, and $c$ and $v$ denote the conduction and valence Landau levels (LLs), respectively. The band structure and the LLs of monolayer graphene have been verified by a number of experimental methods [@VWBrar; @ADeshpande; @KSNovoselov3]. Bilayer graphene, being held together by Van der Waals interactions, can exhibit the highly symmetric AA and AB configurations. The former and the latter have, respectively, two pairs of linear and parabolic bands. The two isotropic Dirac-cones in the AA-stacked system are magnetically quantized into two groups of LLs with monolayer-like wavefunctions and energy spectrum [@YHHo2]. Also, the AB-stacked system presents two groups of LLs, with each LL having a single-mode wavefunction in the absence of LL anticrossing [@YHLai]. The field-dependent LL energy spectrum of AB-stacked bilayer graphene is different from that of monolayer graphene. It includes a few intergroup LL anticrossings at large field strengths ($B_0>$ 100T). The band structures [@VWBrar; @KSNovoselov4; @TOhta] and the first LL groups [@KSNovoselov4] have been experimentally verified for bilayer graphene. The magneto-electronic properties of AAA- and ABA- stacked trilayer graphenes can be regarded as combinations of those of monolayer graphene [@CYLin], and AB-stacked bilayer and monolayer graphenes [@CYLin], respectively. Distinctly, the zero-field band structure of ABC-stacked trilayer graphene exhibits three pairs of energy dispersions: linear cones, sombrero-shaped bands, and parabolic bands [@YPLin; @CYLin]. Accordingly, the $B_0$-dependent energy spectrum has both intergroup and intragroup LL anticrossings, in which the latter are arised from the sombrero-shaped energy bands. The band structures and the first LL groups of these stacking systems have been examined by numerous experimental measurements [@LYZhang; @CHLui2; @FZhang; @TTaychatanapat]. As for the stacking symmetry, the AAB-stacked system is in sharp contrast with the AAA- and ABA-stacked ones, but is also different from the BAC-stacked one. Clearly, the AAB- and ABC-stacked trilayer graphenes can not be regarded as a superposition of two or three sub-systems. We develop the generalized tight-binding model, based on the subenvelope functions on the distinct sublattices, to study the rich electronic properties of AAB-stacked trilayer graphene. This work shows that the AAB-stacked configuration presents an abnormal band structure, in which there are pairs of oscillatory, sombrero-shaped and parabolic bands, which are different from those of the other stacking systems mentioned above. The special band structure could be verified by angle-resolved photoemission spectroscopy (ARPES) [@MSprinkle]. The LL evolution under a magnetic field reveals a complex pattern of LL anticrossings and splittings, as a result of the specific interlayer atomic interactions derived from the full tight-binding model. The magnetoelectronic spectra are directly reflected in DOS which can be examined by experimental measurements using scanning tunneling spectroscopy (STS). The $B_0$-dependent energy spectra exhibit intragroup and intergroup LL anticrossings. Especially, the low-lying LLs anticross frequently, leading to the existences of non-well-behaved and perturbed LLs. The spatially dramatic changes of the LL wavefunctions in the anticrossings can be verified by scanning tunneling microscope (STM) measurements [@ZYRong; @JMCampanera], as done for 2D electron gas [@MFCrommie] and topological insulators [@ZHYang]. In this paper, the important differences among trilayer graphenes such as their energy band structures, LL splittings, LL ordering, and LL anticrossings, are investigated. 0.6 truecm [**2. The generalized tight-binding model** ]{} 0.3 truecm The low-energy $\pi$-electronic structure of AAB-stacked trilayer graphene, mainly coming from the 2$p_z $ orbitals, is calculated with the generalized tight-binding model. The two sublattices in the $l$th $(l=1,2,3)$ layer are denoted as $A^{l}$ and $B^{l}$. The first two layers, shown in Fig. 1, are arranged in the AA-stacking configuration; that is, all carbon atoms have the same (x,y) projections. The third layer can be obtained by shifting the first (or the second) layer by a distance of $b$ along the armchair direction. In this system, the A atoms (black) have the same (x,y) coordinates, while the B atoms (red) on the third layer are projected at the hexagonal centers of the other two layers. The interlayer distance and the C-C bond length are, respectively, $d=3.37 \dot{A}$ and $b=1.42 \dot{A}$. There are six carbon atoms in a primitive unit cell. The low-energy electronic properties are characterized by the carbon $2p_{z}$ orbitals. The zero-field Hamiltonian, which is built from the six tight-binding functions of the $2p_{z}$ orbitals, is dominated by the intralayer and the interlayer atomic interactions $\gamma_i^{'}s$. There exist 10 kinds of atom-atom interactions corresponding to the 10 atomic hopping integrals which appear in the Hamiltonian matrix. $\gamma_{0}=-2.569 $ eV represents the nearest-neighbor intralayer atomic interaction; $ \gamma_{1}=-0.263$ eV, $\gamma_{2}=0.32 $ eV present the interlayer atomic interactions between the first and second layer; $ \gamma_{3}=-0.413$ eV, $\gamma_{4}=-0.177 $ eV, $ \gamma_{5}=-0.319$ eV are associated with the interlayer atomic interactions between the second and third layer; $ \gamma_{6}=-0.013 $ eV, $ \gamma_{7}=-0.0177 $ eV, and $ \gamma_{8}=-0.0319 $ eV relate to the interlayer atomic interactions between the first and third layer; and $ \gamma_{9}=-0.012 $ eV accounts for the difference in the chemical environment of A and B atoms. The hopping integrals $\gamma_1$, $\gamma_3$, and $\gamma_5$ belong to the vertical interlayer atomic interactions, while the others are non-vertical ones. When applying a uniform perpendicular magnetic field, the unit cell becomes enlarged as indicated in Fig. 1. There appears an extra Peierls phase $G_{R}$ in the tight-binding functions. $G_{R}$ has the form of $ \frac{2\pi}{\phi_{0}}\int_R^r \vec{A}. d\vec{l} $, where $\vec{A}$ is the vector potential, and $\phi_{0}=hc/e$ is the flux quantum. The vector potential in the Landau gauge is chosen as $\vec{A}=(0,Bx,0)$. The Peierls phase has a period of $2\phi_{0}/ \phi=2R_{B}$ along the x-axis. Under the effect of the magnetic field, the unit cell becomes enlarged as a rectangle with $12R_{B}$ atoms included. There are $12R_B$ tight-binding functions, which are arranged in the sequence $\{|A_{1}^{1}\, |B_{1}^{1}\rangle, |A_{1}^{2}\rangle, |B_{1}^{2}\rangle, |A_{1}^{3}\rangle, |B_{1}^{3}\rangle, ..., |A_{2R_{B}}^{1}\, |B_{2R_{B}}^{1}\rangle, |A_{2R_{B}}^{2}\rangle, |B_{2R_{B}}^{2}\rangle, |A_{2R_{B}}^{3}\rangle, |B_{2R_{B}}^{3}\rangle\}$. The superscript Hamiltonian is a $12R_{B} \times 12R_{B} $ matrix, in which the non-zero matrix elements can be presented by the equations below. $$\langle B_{j}^{1} |H|A_{i}^{1}\rangle = \gamma_{0} \sum { \frac{1}{N} exp [ i\vec{k}.(\vec{R}_{A_{i}^{'}}- \vec{R}_{{B}_{j}^{'}} )] exp[\frac{2i\pi}{\phi_0} (G_{ \vec{R}_{{B}_{j}^{'}}} - G_{ \vec{R}_{{B}_{j}^{'}}})] }\\$$ $$=\gamma_{0}t_{1,i}\sigma_{i,j} + \gamma_{0}q\sigma_{i,j+1}$$ $$\langle B_{j}^{2} |H|A_{i}^{2}\rangle =\gamma_{0}t_{1,i}\sigma_{i,j} + \gamma_{0}q\sigma_{i,j+1}$$ $$\langle B_{j}^{3} |H|A_{i}^{3}\rangle =\gamma_{0}t_{3,i}\sigma_{i,j-1} + \gamma_{0}q\sigma_{i,j}$$ $$\langle A_{j}^{2} |H|A_{i}^{2}\rangle =\langle A_{j}^{3} |H|A_{i}^{3}\rangle =\gamma_{9}\sigma_{i,j}$$ $$\langle A_{j}^{2} |H|A_{i}^{1}\rangle =\langle B_{j}^{2} |H|B_{i}^{1}\rangle =\gamma_{1}\sigma_{i,j}$$ $$\langle B_{j}^{1} |H|A_{i}^{2}\rangle =\langle A_{j}^{1} |H|B_{i}^{2}\rangle =\gamma_{2} t_{1,i}\sigma_{i,j}$$ $$\langle A_{j}^{3} |H|A_{i}^{2}\rangle =\gamma_{3}\sigma_{i,j}$$ $$\langle B_{j}^{3} |H|B_{i}^{2}\rangle =\gamma_{5}t_{2,i}\sigma_{i,j}$$ $$\langle B_{j}^{2} |H|A_{i}^{3}\rangle =\gamma_{4}t_{1,i}\sigma_{i,j} + \gamma_{4}q\sigma_{i,j+1}$$ $$\langle A_{j}^{2} |H|B_{i}^{3}\rangle =\gamma_{4}t_{3,i}\sigma_{i,j-1} + \gamma_{4}q\sigma_{i,j}$$ $$\langle A_{j}^{3} |H|A_{i}^{1}\rangle =\gamma_{6}\sigma_{i,j}$$ $$\langle B_{j}^{3} |H|B_{i}^{1}\rangle =\gamma_{8}t_{2,i}\sigma_{i,j} + \gamma_{8}q\sigma_{i,j+1}$$ $$\langle B_{j}^{1} |H|A_{i}^{3}\rangle =\gamma_{7}t_{1,i}\sigma_{i,j} + \gamma_{7}q\sigma_{i,j+1}$$ $$\langle A_{j}^{1} |H|B_{i}^{3}\rangle =\gamma_{7}t_{3,i}\sigma_{i,j-1} + \gamma_{7}q\sigma_{i,j}$$ The four independent phase terms are:\ $t_{1,j} = exp \{ i[ -(k_{x}b/2) - (\sqrt {3}k_{y}b/2) +\pi\phi (j-1+ 1/6) ] \} \\ + exp \{ i[ -(k_{x}b/2) + (\sqrt {3}k_{y}b/2) -\pi\phi (j-1+ 1/6) ] \} $\ $t_{2,j} = exp \{ i[ -(k_{x}b/2) - (\sqrt {3}k_{y}b/2) +\pi\phi (j-1+ 3/6) ] \} \\ + exp \{ i[ -(k_{x}b/2) + (\sqrt {3}k_{y}b/2) -\pi\phi (j-1+ 3/6) ] \} $\ $t_{1,j} = exp \{ i[ -(k_{x}b/2) - (\sqrt {3}k_{y}b/2) +\pi\phi (j-1+ 5/6) ] \} \\ + exp \{ i[ -(k_{x}b/2) + (\sqrt {3}k_{y}b/2) -\pi\phi (j-1+ 5/6) ] \} $\ $q=exp\{ ik_{x}b\}$.\ In order to deal with the small values of magnetic field strength and a huge $R_B$, we arrange the Hamiltonian in a band-like symmetric matrix. The Landau wave functions, which are investigated to identify the spatial distributions of the Landau levels, can be expressed as $$|\psi\rangle = \sum_{l=1}^{3}\sum_{m=1}^{2R_{B}} (A_m^l | A_m^l \rangle + B_m^l | B_m^l \rangle ),$$ where $A_m^l (B_m^l)$ are the subenvelope functions presenting the amplitude of the tight-binding functions based on the $m$th A (B) atom at the $l$th layer in the unit cell. 0.6 truecm [**3. The zero-field band structure and quantized Landau levels**]{} 0.3 truecm The zero-field band structure of AAB-stacked trilayer graphene consists of three pairs of conduction and valence subbands labeled $S_1^{c,v}$, $S_2^{c,v}$, and $S_3^{c,v}$, as shown by the solid curves in Fig. 2(a). Near the Fermi energy, the two subbands which belong to the first pair, $S_1^{c,v}$, have strong oscillatory energy dispersions. The conduction subband starts to increase from a local minimum value of about 4 meV at the K point (the corner of the first Brillouin zone), along the KM and K$\Gamma$ directions. After reaching a local maximum value of about 58 meV, it decreases until reaching a local minimum energy again (4 meV), and then grows steadily. The curvature of the valence subband is in the opposite direction, which is almost symmetric to the conduction subband about $E_F$. The first pair of subbands, with three constant energy contours within $\pm$ 58 meV and a narrow gap $ E_g\sim 8$ meV in between them (Fig. 2(b)), is special in that it has never appeared in any other stacking configuration. The triple-degenerate states are not suitable for low-energy expansion, indicating that the effective-mass model can not further deal with the magnetic quantization of these energy bands. The second pair of subbands, $S_2^{c,v}$, has a sombrero-shaped and a local energy minimum (maximum) and maximum (minimum), situated at around 0.24 $(-0.24)$ eV and 0.26 $(-0.26)$ eV, respectively. The energy difference between the two extreme points is quite narrow, being only about 20 meV. Located away from $E_F$, the third pair of subbands, $S_3^{c,v}$, consists of monotonic parabolic bands with a minimum (maximum) value of about 0.49 $(-0.49)$ eV. The above-mentioned features of the low-lying energy bands are consistent with those by the first-principle calculations (dashed curves), clearly indicating that the complex interlayer amomic interactions used in the generalized tight-binding model are suitable. As for the low-lying band structure, there are important differences among trilayer graphenes with distinct stacking configurations. The AAA-, ABA-, and ABC-stacked trilayer graphenes possess special band structures with three pairs of linear bands, a pair of linear bands and two pairs of monotonic parabolic bands, and a pair of sombrero-shaped bands and two pairs of parabolic subbands, respectively. The band structures of the AAA and ABA stackings can be, respectively, regarded as the superposition of monolayer graphene, and monolayer and bilayer graphenes. In contrast, ABC- and AAB-stacking configurations exhibit special energy band structures, which are not similar to those of monolayer and bilayer systems. It should be noted that the energy band structure of AAB-stacking system exhibits an extremely small gap, as indicated in Fig. 2(b). That is to say, while the three stacking trilayer graphenes are gapless 2D semimetals with a slight overlap between the valence and conduction bands, the AAB-stacked system is a narrow-gap semiconductor. A perpendicular magnetic field can quantize electronic states into dispersionless LLs with high degeneracy. AAB-stacked trilayer graphene presents rich magnetoelectronic properties. For the $(k_x=0, k_y=0)$ state, each LL is two-fold degenerate in the absence of spin degeneracy; that is, the wavefunctions localized near 2/6 and 5/6 (4/6 and 1/6) are identical. The total carrier density, in which each LL can be occupied, is $D$=4$eB_0/hc$ per unit area. There exist three groups of LLs, each group can furthermore be divided into two sub-groups corresponding to the two localizations 2/6 and 4/6. Each group consists of unoccupied conduction and occupied valence LLs. The former and the latter are almost symmetric about $E_F$ and present similar behavior. We first discuss the LLs localized near 2/6 at $B_0=40$ T, being characterized by the subenvelope functions on the six sublattices shown in Figs. 3-5. The first, second and third groups are, respectively, initiated at 0, 0.24 and 0.55 eV for the conduction states, and at 0, -0.23 and -0.58 eV for the valence states. Near $E_F$, the first group starts to show up with the first three LLs (blue lines), which are very close to each other. The quantum numbers are determined by the $A^1$ sublattice with the dominating subenvelop function (Fig. 3). Starting from the middle LL, $n_1^{c,v}=0$, which is located almost right at $E_F$, the ordering then increases for the higher conduction states, and the lower valence states. In particular, the quantum number of the next conduction (valence) LL, which is placed at about $E^c=$10 meV ($E^v=-$5.2 meV), is assigned $n_1^c = 1$ ($n_1^v =1$), and so on. For the first group, the quantum number ordering is similar to that of monolayer graphene, while the LL spacing is irregular. The largest spacing between LLs is that of $n_1^{c,v}=1$ and $n_1^{c,v}=2$, which is about $ \Delta E^{c,v}=$ 40 meV; it then decreases with increasing quantum numbers. Apparently, there is no simple relationship between the LL energy and quantum number. This is in contrast to the relationships found in monolayer graphene and 2D electron gas, which can be described by $E^{c,v} \propto \sqrt {n^{c,v}}$ and $E(n) \propto n^{c,v}$, respectively. The destruction of the inversion symmetry in AAB-stacked trilayer graphene results in certain important differences for the LL wavefunctions near 4/6 and 2/6, including differences in the LL spacing, quantum number and spatial distribution of the wavefunctions. At $B_0=40$ T, the first, second and third LL groups near the 4/6 center are, initiated at 0, (0.26 eV, $-$0.24 eV) and $\pm$0.52 eV, as indicated in Fig. 4(a)-4(c), respectively. For the low-lying LLs, it is not easy to define the quantum numbers due to the non-well-behaved spatial distributions of the LL wavefunctions. Particularly, the subenvelope functions oscillate abnormally as a result of the complex LL anticrossings (details in Fig. 6(c)). The eleven LLs nearest to $E_F$, except for the $n=0$ one, are arranged in pairs with very small spacings, in which the LL crossings or anticrossings are clearly revealed (Fig. 4(a)). At low-lying energy levels, the LL wavefunctions are well-behaved only when the applied field strength is sufficiently large (the available magnetic field in experiment is nowadays up to 80 T [@BFauque]), as shown in Fig. 5(a)-5(c) for $B_0=80$ T. The dominating subenvelope functions of the sublattice $B^3$ are used to define the quantum numbers. Similarly to those near 2/6 center, the conduction and valence LLs are also almost symmetric about the $n=0$ LL, which is located right at $E_F$. However, the quantum number ordering is slightly different to that of monolayer graphene. Particularly, the next unoccupied (occupied) LLs are assigned, respectively, $n_1^c=2$ ($n_1^v=2$), $n_1^c=3$ $n_1^v=3$, $n_1^c=1$ ($n_1^v=1$), and so on. Concerning the second group, all the LL wavefunctions are single-mode at $B_0=40$ T. Also, the LL quantum numbers are determined by the dominating $ B^1 $ sublattice for the wavefuntions distributed around 2/6 and 4/6 (green color in Figs. 3(b) and 4(b)). The assigned quantum numbers of the former are $n_2^{c,v}=1$, $n_2^{c,v}=0$, $n_2^{c,v}=2$, $n_2^{c,v}=3$, ... in the order of increasing (decreasing) energy for the conduction (valence) LLs, similarly to those of the latter ($n_2^{c,v}=1$, $n_2^{c,v}=0$, $n_2^{c,v}=2$, $n_2^{c,v}=3$). It should be noted that when the field strength is sufficiently large, the quantum number ordering becomes equivalent to that of monolayer graphene, as shown in Fig. 5 (b) near the 4/6 center for $B_0=80$ T (the available magnetic field in experiment is nowadays up to 80 T [@BFauque]). On the other hand, the third group exhibits a normal quantum number sequence for two localization centers, similarly to that of monolayer graphene. That is to say, the ordering is $n_3^{c,v}=0$, $n_3^{c,v}=1$, $n_3^{c,v}=2$, $n_3^{c,v}=3$, and so on. The LL wavefuntions based on the dominating subenvelope function $A^2$, are almost identical to those of monolayer graphene, and independent of the field strength. The $B_0$-dependent energy spectra are very useful in understanding the rich magnetic quantization in AAB-trilayer graphene. The LL energies exhibit monotonic and entangled spectra, in which the multi-crossings and -anticrossings occur frequently (Fig. 6). For all the LLs of the first group localized at 2/6, the quantum numbers have a normal sequence when the field strength is very large ($B_0> 40$ T). Moreover, the conduction (valence) LL energies monotonously decline (grow) when the field strength is reduced to $B_0=40$ T. With a further decrease of field strength, the $B_0$-dependent energy spectrum becomes oscillatory, which leads to the frequent and pronounced multi-crossings and -anticrossings. If two different multi-mode LLs simultaneously possess the same mode, they are forbidden to cross each other. For example, the three conduction LLs, $n_1^{c}=1$, $n_1^{c}=4$ and $n_1^{c}=7$ continuously anticross, as indicated in Fig. 6(b). In particular, the anticrossing between the former two occurs in the range of $B_0\sim\,19-21$ T, while in the latter two they occur between $B_0\sim\,9-11$ T; the anticrossing between the first and the third conduction LLs appears at $B_0\sim\,11-13$ T. The LL anticrossings clearly indicate the unusual sequence of quantum numbers. For $B_0 >$ 21 T, the main $n_1^c=1$ mode has a normal distribution on the dominating sublattice, in which the side modes of $n_1^c=4$ and $n_1^c=7$ are faint. With the decreasing of $B_0$, the $n_1^c=1$ main mode declines and the $n_1^c=4$ side mode quickly reaches the maximum at $B_0 = 20$ T in the range of $E^c\sim\, 20-40$ meV. At the center of the anticrossing regions, their subenvelop functions have the same $n=1$ and $n=4$ modes, which forbids them to have the same energy. Similarly, the $n_1^c=4$ and $n_1^c=7$ LLs anticross in the range of $E^c\sim\, 15-35$ meV. The anticrossing center is at $B_0\sim\,10$ T where the comparable $n=4$ and $n=7$ modes exist in two subenvelope functions. Moreover, the other conduction LLs exhibit similar $B_0$-dependent energy spectra. Particularly, there are other pairs of LLs, such as $n_1^c=2$ and $n_1^c=5$, $n_1^c=3$ and $n_1^c=6$, ... , for which the LL anticrossings appear, respectively, in the ranges of 16 T $<B_0<$ 18 T ($E^c\sim\,18-38$ meV), 13 T $<B_0<$ 15 T ($E^c\sim\,17-37$ meV), and so on. The LL anticrossings form a wide stateless region which is indicated in Fig. 6(b). Generally, the main modes of anticrossing LLs are different by 3 at large field strengths and 6 at small ones. Furthermore, there exists another region of LL anticrossings which is indicated in Fig. 7(a). These anticrossings happen at $B_0<$ 19 T and form very narrow in-between gaps. For example, the $n_1^c=1$ and $n_1^c=7$ LLs are forbidden to cross in the energy range of $E^c\sim\, 34-36$ meV (Fig. 7(a)). Likewise, the other pairs of LLs, such as ($n_1^c=2$ and $n_1^c=8$ ) ($E^c\sim\, 39-40$ meV), ($n_1^c=3$ and $n_1^c=9$ ) ($E^c\sim\, 41-41.5$ meV), anticross at about $B_0=$ 11 T and $B_0=$ 10.6 T, respectively. With a further decrease of $B_0$, the $n_1^c=4$ and $n_1^c=10$ ($n_1^c=5$ and $n_1^c=11$) LLs anticross each other at around $B_0=$ 6 T ($B_0=$ 5 T) in the range of $E^c\sim\, 36-38$ meV ($E^c\sim\, 40-42$ meV), and so on. In short, for each pair of anticrossing LLs, the main modes differ by 6. In addition to the intragroup anticrossings of the LLs in the first group, the second group LLs also avoid crossing each other at smaller field strengths ($B_0 < 5$ T) due to the sombrero-shaped energy dispersions [@YPLin; @CYLin] shown in Fig. 7(b). The third group consists of monotonic LLs without intragroup anticrossings. The intergroup LL anticrossings occur among all three groups at sufficiently large field strengths. For instance, the $n_1^{v}= 3$ LL anticrosses the $n_2^{v}=0$ and $n_2^{v}=3$ ones continuously at $B_0\sim\,75-97$ T, in the range of, respectively, $E^v\sim\,$-23$-$18 meV and $E^{v}\sim\,$-27$-$-23 meV, as shown in Fig. 7(c). In general, the main modes of anticrossing LLs are different by 3m, with m being an integer. It should be noted that the interlayer atomic interaction $\gamma 5$ between the non-vertical sites in the second and third layers induces both intergroup and intragroup LL anticrossings. This is similar to ABC-stacked trilayer graphene [@CYLin]. Significantly, near $E_F$, the $n=0$ LL presents the unusual field-dependent LL wavefunctions, as indicated in Fig. 6(b). When the magnetic field is sufficiently large, e.g., $B_0>22$ T, both $n=0$ and $n^{c,v}=3$ LLs present the well-behaved wavefunctions. However, the $n=0$ LL is forbidden to cross the $n^{c}=3$ one when the field strength is reduced to $B_0\sim 22$ T. Accordingly, there appear some $n=0$ mode wavefunctions near the turning point of the $n^c=3$ LL, which is located at $B_0\sim 20$ T and $E^c\sim 13$ meV. Besides, the $n^c=3$ and $n^v=3$ LLs also weakly anticross right at their extreme points, forming an exceedingly narrow gap in the anticrossing center. As the field strength is gradually decreased, the $n=0$ and $n^c=3$ LLs anticross each other again at $B_0\sim 16$ T. Therefore, three LLs in the range of $E^c\sim\,$-10$-15$ meV and $B_0\sim\,16-22$ T exhibit an entangled energy spectrum including the special anticrossings. On the other hand, there exist certain critical differences between the field-dependent LL spectra at the 4/6 and 2/6 localization centers. For the first group, the intragroup LL anticrossings of the former appear in a wider magnetic-field range, 0 T $<B_0<80$ T (Fig. 6(c)). Specifically, the complex LL anticrossings of the $n=0$ and $n^{c,v}=3$ LLs are revealed in the range of larger field strengths ($B_0\sim\,32-45$ T) compared to that of the latter ($B_0\sim\,16-22$ T), as indicated in Figs. 6(b)-(c). The $n=0$ LL anticrosses the $n^v=3$ LL twice, leading to the fact that the zero-mode LLs only exist above $E_F$, which is in contrast to those near the 2/6 center. Moreover, the stateless gap in the anticrossing center of $n^c=3$ and $n^v=3$ LLs is about 5 meV, clearly larger than that at 2/6 localization. The $B_0$-dependent LL wavefunctions of AAB-stacked trilayer graphene exhibit diverse real-space distributions. They can be verified by the spectroscopic-imaging STM measurement, as done for 2D electron gas [@MFCrommie] and topological insulators [@ZHYang]. From the measurements on the variations of the local density of states (DOS) in graphene planes, STM reveals a nodal structure corresponding to the well-behaved LL wavefunctions. For the hybridized wavefunctions due to the LL anticrossings in AAB-trilayer graphene, the results from the STM measurements are expected to be helpful in distinguishing the main mode and the side modes of the LLs. Moreover, the energy spectra and the internal structure of the wavefunctions could be directly examined by STM and STS measurements (discussed later), respectively, giving a useful identifiable picture of the LLs. Reflecting the main characteristics of LL spectra, the DOS is defined as $D(\omega )=\underset{n^{c},n^{v}}{\sum} \int_{1stBZ}\delta (\omega - E^{c,v}(n,k))dk.$ The discrete LLs lead to many symmetric delta-function-like peaks in the DOS, where the peak intensities are proportional to the LL degeneracy [@DLMiller; @GMRutter]. Such peaks correspond to the sharp structures in the differential conductance map of dI/dV$-$V from STS measurements. For the DOS of the AAB-stacked trilayer graphene at B=40 T, the first, second and third groups, respectively, have the onset conduction-states about 0, 0.26 and 0.52 eV, as shown in Fig. 8(a). Each group consists of two sub-groups, which are separated by a symmetry-breaking induced energy difference of about 10 meV. In the vicinity of the Fermi level, the peaks of the first groups are neither regularly sequenced nor do they follow a simple relationship as for monolayer graphene. This is due to the fact that the special quantization effects on the lowest subband leads to the multi-crossings and anticrossings within the band width. On the other hand, the second and third groups exhibit a normal sequence in the order of increasing energies. A crossover of LLs in different groups induces higher DOS, which is thus expected to exhibit stronger tunneling currents in experimental measurements. The DOS exhibits distinct characteristics among different stacking configurations, such as the peak sequence, intensity, energy and splitting. In AB-stacked trilayer graphene, peaks that follow the sequence in monolayer and bilayer graphenes are observed, as presented in Fig. 8(b). Similarly, for the AAB stacking, the half-intensity peaks near the Fermi level are attributed to the symmetry-broken structure, which leads to a considerable energy splitting of about 10 meV. However, this splitting is hardly observable at higher energies. The AA-stacked trilayer graphene exhibits three groups of monolayer-like sequence of peaks [@DLMiller; @SJung; @YJSong] located at energies described by the simple relationship $E^{c,v}\propto\sqrt{n^{c,v}B}$, as indicated in Fig. 9(a). On the contrary, the DOS of ABC-stacked trilayer graphene is distinct from that of monolayer and bilayer graphenes, as shown in Fig. 9(b). The exceptionally high peak at the Fermi level is a superposition of three peaks corresponding to the Dirac points; its intensity is proportional to the number of graphene layers. The essential differences of the DOS can be verifed by STS [@DLMiller; @GMRutter; @SJung; @YJSong]; those profiles can then be used as a tool to identify the stacking configuration of graphene sheets. 0.6 truecm [**4. Conclusion** ]{} 0.3 truecm The generalized tight-binding model is developed to investigate the rich magneto-electronic properties of AAB-stacked trilayer graphene. This system exhibits special energy bands and thus a rich magnetic quantization. The complex interlayer atomic interactions indicate that the low-energy expansion about the K point is not suitable in obtaining energy dispersions; therefore, the effective-mass model can not be used to achieve further magnetic quantization. The three zero-field pairs of energy bands contain oscillatory, sombrero-shaped and parabolic dispersions, which could be examined by ARPES [@MSprinkle]. They are magnetically quantized into three groups of LLs, defined by the dominating subenveloped functions. The field-dependent energy spectra, in which the frequent intragroup and intergroup LL anticrossings happen simultaneously, are very complex. There exist important differences between the LLs near 4/6 and 2/6 in terms of the LL splitting, quantum number, spatial distribution of wavefunctions, and anticrossing LL energy spectra. Moreover, the electronic properties of AAB-stacked trilayer graphene sharply contrast to those of the AAA-, ABA- and ABC-stacked ones, such as state degeneracy, LL anticrossing behavior, initial energies of each group, quantum number ordering and wavefunction distributions. This principally demonstrates the unique characteristics of the special AAB configuration. The first and the second LL groups exhibit unusual energy spectra with continuous intragroup LL anticrossings. Especially for the first group, each low-lying LL is forbidden to cross with the others more than once. The LL anticrossing region near $E_F$ has a wide stateless gap, while that of the other region is quite narrow. The main modes of LLs during anticrossings are different by 3 and 6 for the former, and only 6 for the latter. Furthermore, the intergroup LL anticrossings among three groups are revealed only at sufficiently large field strengths. Remarkably, the DOS exhibits many symmetric delta-function-like peaks with an irregular sequence near $E_F$ due to the LL multi-anticrossings. The web-like energy spectra and the seriously distorted wavefunctions could be examined by STS and STM experimental measurements, respectively. The complex LL spectra are expected to induce the rich mechanical, excitonic, thermal and optical properties of the AAB-stacking system. [**Acknowledgments**]{} This work was supported in part by the National Science Council of Taiwan, the Republic of China, under Grant Nos. NSC 98-2112-M-006-013-MY4 and NSC 99-2112-M-165-001-MY3. [99]{} M. Koshino, T. Nakanishi, T. Ando. Phys. Rev. B 82 (2012) 205436. A. D. Martino, A. Hutten, R. Egger. Phys. Rev. B 84 (2011) 155420. J. Sari, C. Toke. Phys. Rev. B 87 (2013) 085432. Y. H. Ho, S. J. Tsai, M. F. Lin, W. P. Su. Phys. Rev. B 87 (2013) 075417. M. M. Kruczynski, I. L. Aleiner, V. I. Falko. Phys. Rev. B 84 (2011) 041404. Y. P. Lin, J. Wang, J. M Lu, C. Y. Lin, M. F. Lin. RSC Advances 4 (2014) 56552. Y. K. Huang, S. C. Chen, Y. H. Ho, C. Y. Lin, M. F. Lin. Scientific Reports 4 (2014) 7509. S. H. R. Sena, J. M. Pereira Jr., F. M. Peeters, G. A. Farias. Phys. Rev. B 84 (2011) 205448. A. Kumar, W. Escoffier, J. M. Poumirol, C. Faugeras, D. P. Arovas, M. M. Fogler, et al. Phys. Rev. Lett. 107 (2011) 126806. Y. B. Zhang, Y. W. Tan, H. L. Stormer, P. Kim. Nature 438 (2005) 201. C. L. Kane, E. J. Mele. Phys. Rev. Lett. 95 (2005) 226801. K. S. Novoselov, E. Mccann, S. V. Morozov, V. I. Falko, M. I. Katsnelson, U. Zeitler, et al. Nature Physics 2 (2006) 177. K. S. Novoselov, Z. Jiang, Y. Zhang, S. V. Morozov, H. L. Stormer, U. Zeitler, et al. Science 315 (2007) 1379. V. P. Gusynin, S. G. Sharapov. Phys. Rev. Lett. 95 (2005) 146801. F. Guinea, M. I. Katsnelson, A. K. Geim. Nature Physics 6 (2009) 30. X. Du, I. Skachko, F. Duerr, A. Luican, E. Y. Andrei. Nature 462 (2009) 192. J. W. Jiang, J. S. Wang, B. W. Li. Phys. Rev. B 80 (2009) 113405. M. M. Shokrieh, R. Rafiee. Materials and Design 31 (2010) 790. J. U. Lee, D. Yoon, H. Cheong. Nano Lett. 12 (2012) 4444. Q. Lu, M. Arroyo, R. Huang. Journal of Physics D 42 (2009) 102002. C. Lee, X. D. Wei, J. W. Kysar, J. Hone. Science 321 (2008) 1157996. L. Huang, Y. Huang, J. J. Liang, X. J. Wan, Y. S. Chen. Nano Reseach 4 (2011) 675. Y. M. Lin, A. V. Garcia, S. J. Han, D. B. Farmer, I. Meric, Y. N. Sun, et al. Science 332 (2011) 1294. I. Meric, M. Y. Han, A. F. Young, B. Ozyilmaz, P. Kim, K. L. Shepard. Nature Nanotechnology 3 (2008) 654. F. N. Xia, D. B. Farmer, Y. M. Lin, P. Avouris. Nano Lett. 10 (2010) 715. J. B. Wu, M. Agrawal, H. A. Becerrll, Z. N. Bao, Z. F. Liu, Y. S. Chen, et al. ACS Nano 4 (2010) 43. T. H. Han, Y. B. Lee, M. R. Choi, S. H. Woo, S. H. Bae, B. H. Hong, et al. Nature Photonics 6 (2012) 105. G. H. Jo, M. Y. Choe, C. Y. Cho, J. H. Kim, W. J. Park, S. C. Lee, et al. Nanotechnology 21 (2010) 175201. J. B. Wu, H. A. Becerril, Z. N. Bao, Z. F. Liu, Y. S. Chen, P. Peumans. Appl. Phys. Lett. 92 (2008) 263302. J. D. Roy-Mayhew, D. J. Bozym, C. Punckt, L. A. Aksay. ACS Nano 4 (2010) 6203. N. L. Yang, J. Zhai, D. Wang, Y. S. Chen, L. Jiang. ACS Nano 4 (2010) 887. W. J. Hong, Y. X. Xu, G. W. Lu, C. Li, G. Q. Shi. Electrochemistry Communications 10 (2008) 1555. S. K. Bae, H. K. Kim, Y. B. Lee, X. F. Xu, J. S. Park, Y. Zheng, et al. Nature Nanotechnology 5 (2010) 574. J. Wang, M. H. Liang, Y. Fang, T. F. Qiu, J. Zhang, L. J. Zhi. R. Coating. Advanced Materials 24 (2012) 2874. H. Hibino, H. Kageshima, M. Kotsugi, F. Maeda, F.-Z. Guo, Y. Watanabe. Phys. Rev. B 79 (2009) 125437. E. H. M. Ferreira, M. V. O. Moutinho, F. Stavale, M. M. Lucchese, R. B. Capaz, C. A. Achete, et al. Phys. Rev. B 82 (2010) 125429. A. H. C. Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov, A. K. Geim. Rev. of Modern Phys. 81 (2009) 109. C. Y. Lin, J. Y. Wu, Y. H. Chiu, C. P. Chang, M. F. Lin. Phys. Rev. B 90 (2014) 205434. J. S. Lee, K. S. Novoselv, H. S. Shin. ACS Nano 5 (2010) 608. M. Aoki, H. Amawashi. Solid State Communications 142 (2007) 123. F. Guinea, A. H. Castro Neto, N. M. R. Peres. Phys. Rev. B 73 (2006) 245426. K. F. Mak, J. Shan, T. F. Heinz. Phys. Rev. Lett. 104 (2010) 176404. Y. Zhang, Z. Jiang, J. P. Small, M. S. Purewal, Y.-W. Tan, M. Jazlollahi, et al. Phys. Rev. Lett. 96 (2006) 136806. M. O. Goerbig. Rev. of Modern Phys. 83 (2011) 1193. C. L. Lu, C. P. Chang, Y. C. Huang, R. B. Chen, M. F. Lin. Phys. Rev. B 73 (2006) 144427. E. V. Castro, K. S. Novoselov, S. V. Morozov, N. M. R. Peres, J. M. B. Lopes dos Santos, J. Nilsson, et al. Phys. Rev. Lett. 99 (2007) 216802. K. F. Mak, C. H. Lui, J. Shan, T. F. Heinz. Phys. Rev. Lett. 102 (2009) 256405. D. M. Basko, S. Piscanec, A. C. Ferrari. Phys. Rev. B 80 (2009) 165413. C. Casiraghi. Phys. Rev. B 80 (2009) 233407. J. E. Lee, G. H. Ahn, J. Y. Shim, Y. S. Lee, S. M. Ryu. Nature Communications 3 (2012) 2022. S. M. Choi, S. H. Jhi, Y. W. Son. Phys. Rev. B 81 (2010) 081407. J. H. Wong, B. R. Wu, M. F. Lin. J. Phys. Chem. C 116 (2012) 8271. Y. W. Tan, Y. Zhang, H. L. Stormer, P. Kim. Eur. Phys. J. Special Topics 148 (2007) 15. V. V. Cheianov, V. I. Fal’ko. Phys. Rev. Lett. 97 (2006) 226801. M. J. Webb, P. Palmgren, P. Pal, O. Karis, H. Grennberg. Carbon 49 (2011) 3242. Z. Liu, Q. S. Zheng, J. Z. Liu. Appl. Phys. Lett. 96 (2010) 201909. Z. Y. Rong, P. Kuiper. Phys. Rev. B 48 (1993) 17427. J. M. Campanera, G. Savini, I. Suarez-Martinez, M. I. Heggie. Phys. Rev. B 75 (2007) 235449. K. S. Novoselov, A. K. Geim, S. V. Morozov, D. Jiang, Y. Zhang, S. V. Dubonos, et al. Science 306 (2004) 666. B. Lalmi, J. C. Girard, E. Pallecchi, M. Silly, C. David, S. Latil, et al. Scientific Reports 4 (2014) 04066. C. T. Ellis, A. V. Stier, M. H. Kim, J. G. Tischler, E. R. Glaser, R. L. M. Ward, et al. Scientific Reports 3 (2013)03143. J. H. Hwang, V. B. Shields, C. I. Thomas, S. Shivaraman, D. Hao, M. K. Kim, et al. Journal of Crystal Growth 312 (2010) 3219. Z. Y. Juang, C. Y. Wu, A. Y. Lu, C. Y. Su, K. C. Leou, F. R. Chen, et al. Carbon 48 (2010) 3169. D. R. Lenski, M. S. Fuhrer. Journal of Appl. Phys. 110 (2011) 013720. B. Jayasena, S. Subbiah. Nanoscale Res. Lett. 6 (2011) 95. C. H Lui, L. M. Malard, S. H. Kim, G. Lantz, F. E. Laverge, R. Saito, et al. . Nano Lett. 12 (2012) 5539. K. F. Mak, M. Y. Sfeir, J. A. Misewich, T. F. Heinz. PNAS 107 (2010) 14999. L. Y. Zhang, Y. Zhang, J. Camacho, M. Khodas, I. Zaliznyak. Nature Phys. 7 (2011) 953. L. B. Biedermann, M. L. Bolen, M. A. Capano, D. Zemlyanov, R. G. Reifenberger. Phys. Rev. B 79 (2009) 125411. Y. P. Wu, B. Wang, Y. F. Ma, Y. Huang, N. Li, F. Zhang, et al. Nano Reseach 3 (2010) 611. Z. S. Wu, W. Ren, L. B. Gao, J. P. Zhao, Z. P. Chen, B. L. Liu, et al. ACS Nano 3 (2009) 411. S. Wu, M. Killi, A. Paramekanti. Phys. Rev. B 85 (2012) 195404. V. W. Brar, Y. B. Zhang, Y. Yayon. Appl. Phys. Lett. 91 (2007) 122102. A. Deshpande, W. Bao, F. Miao, C. N. Lau, B. J. LeRoy. Phys. Rev. B 79 (2009) 205411. K. S. Novoselov, A. K. Geim, S. V. Morozov, D. Jiang, M. I. Katsnelson, I. V. Grigorieva, et. al. Nature 438(2005) 197. Y. H. Ho, J. Y. Wu, R. B. Chen, Y. H. Chiu, M. F. Lin. Appl. Phys. Lett. 97 (2010) 101905. Y. H. Lai, J. H. Ho, C. P. Chang, M. F. Lin. Phys. Rev. B 77 (2008) 085426. T. Ohta, A. Bostwick, T. Seyller, K. Horn, E. Rotenberg. Science 313 (2006) 951. C. H. Lui, Z. Q. Li, K. F. Mak, E. Cappelluti, T. F. Heinz. Nat. Phys. 7 (2011) 944. F. Zhang, B. Sahu, H. K. Min, A. H. MacDonald. Phys. Rev. B 82 (2010) 035409. T. Taychatanapat, K. Watanabe, T. Taniguchi, P. Jarillo-Herrero. Nat. Phys. 7 (2011) 621. M. Sprinkle, D. Siegel, Y. Hu, J. Hicks, A. Tejeda, A. Taleb-Ibrahimi, et. al. Phys. Rev. Lett. 103 (2009) 226803. M. F. Crommie, C. P. Lutz, D. M. Eigler. Nature 363 (1993) 524. Z. H. Yang, J. H. Han. Phys. Rev. B 83 (2011) 045415. B. Fauque, D. LeBoeuf, B. Vignolle, M. Nardone, C. Proust, K. Behnia. Phys. Rev. Lett. 110 (2013) 266601. D. L. Miller, K. D. Kubista, G. M. Rutter, M. Ruan, W. A. de Heer, P. N. First, et. al. Science 324 (2009) 924. G. M. Rutter, S. Jung, N. N. Klimov, D. B. Newell, N. B. Zhitenev, and J. A. Stroscio. Nat. Phys. 7 (2011) 649. S. Jung, G. M. Rutter, N. N. Klimov, D. B. Newell, I. Calizo, A. R. Hight-Walker, et. al. Nat. Phys. 7 (2011) 245. Y. J. Song, A. F. Otte, Y. Kuk, Y. Hu, D. B. Torrance, P. N. First, et. al. Nature 467 (2010) 185. **FIGURE CAPTIONS** Fig. 1 - The interlayer atomic interactions and the geometric structure under a uniform magnetic field $B_0\hat{z}$. The shaded region corresponds to a rectangular unit cell. The first and second layers have the same (x, y) projections. Fig. 2 - The (a) energy band structures of AAB-stacked trilayer graphene with (b) a narrow energy gap. Fig. 3 - The distributions of the LL wavefunctions based on the dominating sublattices centered at the (a)-(c) 2/6 localization under $B_0=40$ T. Fig. 4 - The distributions of the LL wavefunctions based on the dominating sublattices centered at the (a)-(c) 4/6 localization under $B_0=40$ T. Fig. 5 - The distributions of the LL wavefunctions based on the dominating sublattices centered at the (a)-(c) 4/6 localization under $B_0=40$ T. Fig. 6 - Three groups of field-dependent LL energy spectrum for (a) the 2/6 localization center; the low-lying LL spectra for (b) the 2/6 and (c) 4/6 centers. Fig. 7 - The (a) second region of intragroup LL anticrossings in the first group, the (b) frequent intragroup LL anticrossings of the second group at low field strength, and the (c) intergroup LL anticrossings of the first and second groups. Fig. 8 - The DOS of (a) AAB- and (b) ABA-stacked trilayer graphene at $B_0=40$ T. Fig. 9 - The DOS of (a) AAA- and (b) ABC-stacked trilayer graphene at $B_0=40$ T. 0.5 truecm {width="0.9\linewidth"} {width="0.6\linewidth"} {width="0.9\linewidth"} {width="0.9\linewidth"} {width="0.9\linewidth"} {width="0.9\linewidth"} {width="0.9\linewidth"} {width="0.9\linewidth"} {width="0.9\linewidth"}

ArXiv

--- abstract: 'We prove a result on the representation of squares by second degree polynomials in the field of $p$-adic meromorphic functions in order to solve positively Büchi’s $n$ squares problem in this field (that is, the problem of the existence of a constant $M$ such that any sequence $(x_n^2)$ of $M$ - not all constant - squares whose second difference is the constant sequence $(2)$ satisfies $x_n^2=(x+n)^2$ for some $x$). We prove (based on works by Vojta) an analogous result for function fields of characteristic zero, and under a Conjecture by Bombieri, an analogous result for number fields. Using an argument by Büchi, we show how the obtained results improve some theorems about undecidability for the field of $p$-adic meromorphic functions and the ring of $p$-adic entire functions.' author: - | Hector Pasten\ University of Concepción date: - - title: 'Representation of squares by monic second degree polynomials in the field of $p$-adic meromorphic functions' --- Introduction ============ In 1970, after the work developed by M. Davis, H. Putnam and J. Robinson, Hilbert’s Tenth Problem was answered negatively by Y. Matiyasevich. In logical terms, it was shown that the positive existential theory of ${\mathbb{Z}}$ in the language of rings ${\mathcal{L}}_R=\{0,1,+,\cdot\}$ is undecidable, which means that there exists no algorithm to decide whether a system of diophantine equations (or equivalently, a single diophantine equation) has integer solutions or not. Soon after the problem was solved, J. R. Büchi proved in an unpublished work (see [@Lipshitz]) that a positive answer to a certain problem in Number Theory (which we write here ${\mathbf{BP}}({\mathbb{Z}})$) would allow to show that there exists no algorithm to decide whether a system of diagonal quadratic diophantine equations has integer solutions or not. The number-theoretical problem ${\mathbf{BP}}({\mathbb{Z}})$ is based on the following observation. If we consider the first difference of a sequence of consecutive integer squares (for example $1,4,9,16$), we obtain a sequence of odd integers (in our example $3,5,7$). Hence, the second difference is constant and equal to two. One may ask whether a sequence of squares having second difference equal to two must be a sequence of consecutive squares. The sequence $6^2,23^2,32^2,39^2$ shows that in general such a reciprocal is not true. \[buchi\] Does there exist an integer $M$ such that the following happens:\ If the second difference of a sequence $(x_i^2)_{i=1}^M$ of integer squares is constant and equal to $2$, then there exists an integer $\nu$ such that $x_i^2=(\nu + i)^2$ for $i=1,\ldots,M$ (that is, the squares must be consecutive). This problem became known as the *$n$ Squares Problem* or *Büchi’s Problem*. Numerical evidence suggests that $M=5$ should work, but ${\mathbf{BP}}({\mathbb{Z}})$ still is an open problem. Assuming a positive answer to ${\mathbf{BP}}({\mathbb{Z}})$, Büchi was able to construct an algorithm to do the following: given a diophantine equation, to construct a system of quadratic diagonal equations such that the former has a solution if and only if the latter has. Therefore, using the negative answer given to Hilbert’s tenth problem and assuming a positive answer to ${\mathbf{BP}}({\mathbb{Z}})$, we get the non-existence of an algorithm to decide whether a system of diophantine diagonal quadratic equations has an integer solution. The non-existence of such an algorithm can be shown to be equivalent to the undecidability of the positive existential theory of ${\mathbb{Z}}$ over the language ${\mathcal{L}}_2=\{0,1,+,P_2\}$ where $P_2(x)$ is interpreted as ‘$x$ is a square’. In order to get similar consequences in Logic for other rings of interest, and motivated by the arithmetical interest of the problem, several authors have studied variants of ${\mathbf{BP}}({\mathbb{Z}})$. A natural thing to do is to replace the ring ${\mathbb{Z}}$ by another commutative ring $A$ with unit. Depending on the ring, we sometimes need to make additional hypothesis in the statement of ${\mathbf{BP}}(A)$: - If $A$ is a ring of functions of characteristic zero in the variable $z$, then we ask for at least one $x_i$ to be non-constant. - If $A$ is a ring of positive characteristic, then we ask $M$ to be at most the characteristic of $A$. The positive existential ${\mathcal{L}}_2$-theory of a ring is usually much weaker than its positive existential ${\mathcal{L}}_R$-theory. But when Büchi’s problem has a positive answer for a ring $A$ then those theories for $A$ are (in general) equivalent. This is what happens for example for $p$-adic analytic functions and for $p$-adic meromorphic functions (see Section \[ResLogic\]). In this work, we will solve ${\mathbf{BP}}(A)$ for some rings of functions (namely, the field of p-adic meromorphic functions and function fields of curves in characteristic zero) by showing in each case a somewhat stronger result on representation of squares by polynomials, in the spirit of the following: *Given an algebra $A\subseteq B$, there exists a constant $M$ satisfying the following condition:\ For any set $\{a_1,\ldots,a_M\}$ of $M$ elements in $A$, there exists a ‘small’ set $E\subseteq B[X]$ such that, if a monic polynomial of degree two $P\in B[X]$ has the property that each $P(a_i)$ is a square in $B$, then $P\in E$ or $P$ is a square in $B[X]$.* Also, we will show that such a result on representation of squares should hold in number fields (and hence Büchi’s problem should also be true there) under the hypothesis that a conjecture by Bombieri holds for surfaces: \[ConjBombieri\] If $X$ is a smooth variety of general type defined over a number field $K/{\mathbb{Q}}$, then $X(K)$ is contained in a proper Zariski closed set of $X$. Finally, we will use the positive answer obtained for Büchi’s problem for $p$-adic meromorphic functions in order to improve some undecidability results for $p$-adic analytic and meromorphic functions. In the next section, we present in details the main results that are proven in this paper. Main results ============ Representation of squares in the field of $p$-adic meromorphic functions ------------------------------------------------------------------------ Let $p$ be a prime number and let ${\mathbb{C}}_p$ be the field of $p$-adic complex numbers (the completion of the algebraic closure of the field ${\mathbb{Q}}_p$ of $p$-adic numbers). Throughout the paper, one can replace ${\mathbb{C}}_p$ by any algebraically closed field of characteristic zero, complete with respect to a non-trivial non-Archimedean valuation. Let ${\mathcal{A}}_p$ be the ring of entire functions over ${\mathbb{C}}_p$ and let ${\mathcal{M}}_p$ be the field of meromorphic functions over ${\mathbb{C}}_p$. We prove the following theorem on representation of squares by polynomials. \[RepMer\] Let $P\in {\mathcal{M}}_p[X]$ be a monic polynomial of degree two. If $P(a)$ is a square in ${\mathcal{M}}_p$ for at least $35$ values of $a\in{\mathbb{C}}_p$, then either $P$ has constant coefficients or $P$ is a square in ${\mathcal{M}}_p[X]$. \[BuchiMer\] The problems ${\mathbf{BP}}({\mathcal{A}}_p)$ and ${\mathbf{BP}}({\mathcal{M}}_p)$ have a positive answer. We will prove these results in Section \[SecMer\]. Representation of squares in number fields ------------------------------------------ We will show in Sections \[geometric\] and \[SecNum\] the following results on representation of squares in number fields. \[RepNum\] Assume Conjecture \[ConjBombieri\] holds for surfaces. Let $K$ be a number field and $\{a_1,\ldots,a_8\}$ a set of eight elements in $K$. There exists a finite (possibly empty) set $E=E(K,(a_i)_i)$ of polynomials in $K[x]$ such that the following holds: for each polynomial $f$ of the form $x^2+ax+b\in K[x]$, if $f(a_i)$ are squares in $K$ for each $i$ then either $f\in E$, or $f=(x+c)^2$ for some $c\in K$. It is an obvious but remarkable fact that if one could find a number field $K$ and a sequence $(a_i)$ such that the set $E(K,(a_i))$ is infinite, then one would automatically obtain a counterexample to Bombieri’s Conjecture. From the finiteness of the set $E(K,(a_i))$ one can easily derive the following (see Section \[SecNum\]). \[buchistrong\] Assume that Bombieri’s conjecture holds for surfaces defined over ${\mathbb{Q}}$. Let $a_1,a_2,\ldots$ be a sequence of integers without repeated terms. There exists a constant $M$ (depending on the sequence $(a_i)_i$) such that: if a polynomial $f=x^2+ax+b\in {\mathbb{Q}}[x]$ satisfies the property ‘$f(a_i)$ is a square in ${\mathbb{Z}}$ for $i=1,\ldots,M$’, then $f$ is of the form $f=(x+c)^2$, for some $c\in{\mathbb{Z}}$. Observe that the dependence of $M$ on the sequence cannot be dropped. Consider for example the polynomial $f_N=x^2-4(2N)!$, and define $$a_i=i! +\frac{(2N)!}{i!}.$$ Then it is obvious that $(a_i)_{i=1}^N$ is a strictly decreasing sequence in ${\mathbb{Z}}$ and each $f_N(a_i)$ is a square in ${\mathbb{Z}}$. Note that, if in Corollary \[buchistrong\] we set $a_n=n$ for each $n$, then we obtain a positive answer to Büchi’s Problem for ${\mathbb{Z}}$ (under Bombieri’s Conjecture). For ${\mathbb{Z}}$, we can actually state a stronger result, which will be proved in Section \[SecNum\]. In order to state it, we need first to give a definition. Observe that if a sequence $(x_n)$ satisfies the equation $$x_{n+2}^2-2x_{n+1}^2+x_n^2=2$$ for $n=1,\dots,M-2$ then solving the induction gives $$x_n^2=(n-1)(n-2)-(n-2)x_1^2+(n-1)x_2^2$$ for each $n$. This observation motivates the following: For $n\ge 3$, we say that a complex projective surface $X\subseteq{\mathbb{P}}^n$ is an *$n$-Büchi surface*, if there exists a sequence of distinct nonzero *integers* $\delta_2,\delta_3,\ldots,\delta_n$ such that $X$ is defined by the system $$\delta_2 x^2_i=\delta_i\delta_2(\delta_i-\delta_2)x^2_0-(\delta_i-\delta_2)x^2_1+\delta_ix^2_2$$ where $i$ ranges from $3$ to $n$ (the $x_i$ being the independent variables). One can verify that all Büchi surfaces are smooth, and that $n$-Büchi surfaces for $n\ge 6$ are of general type (we will not prove this fact, because the proof is essentially the same as the one we give for Lemma \[Xn\]). \[LogicNum\] If Bombieri’s conjecture for $K={\mathbb{Q}}$ holds for *some* $n$-Büchi surface for $n\ge 8$, then multiplication is existentially definable in ${\mathbb{Z}}$ over ${\mathcal{L}}_2$. Therefore, we need quite less than a positive answer to Büchi’s problem in order to define multiplication (because Büchi’s problem corresponds to the particular case where $\delta_k=k-1$ for each $k$). Representation of squares in function fields -------------------------------------------- The geometric results in Section \[geometric\] will be used in Section \[SecFun\] to prove the following theorem, which is the analogue of Theorem \[RepMer\]. \[RepFun\] Let $F$ be a field of characteristic zero and $C$ a non-singular projective curve defined over $F$. Define the integer $M=\max\{8,4(g+1)\}$ where $g$ is the genus of $C$. Write $K(C)$ for the function field of $C$ and let $X$ be transcendental over $K(C)$. Let $P\in K(C)[X]$ be a monic polynomial of degree two. If $P(a)$ is a square in $K(C)$ for at least $M$ values of $a\in F$, then either $P$ has constant coefficients or $P$ is a square in $K(C)[X]$. Theorem \[RepFun\] gives as a direct consequence a positive answer to Büchi’s problem, but such a positive answer is not new since it was (implicit) in [@Vojta2] and recently was solved by a new method in characteristic zero and (large enough) positive characteristic in [@ShlapentokhVidaux]. Undecidability for $p$-adic entire and meromorphic functions in Büchi’s language {#ResLogic} -------------------------------------------------------------------------------- Corollary \[BuchiMer\] allows us to obtain very strong undecidability results for $p$-adic analytic and meromorphic functions, improving results by Lipshitz, Pheidas and Vidaux. In order to state the theorems, we need to introduce some notation. Recall that ${\mathcal{A}}_p$ stands for the ring of entire functions over ${\mathbb{C}}_p$ and ${\mathcal{M}}_p$ stands for the field of meromorphic functions over ${\mathbb{C}}_p$, with variable $z$. Consider the language ${\mathcal{L}}_2'=\{0,1,+,f_z,P_2\}$ where $P_2(x)$ is interpreted as ‘$x$ is a square’ and $f_z(x,y)$ is interpreted as ‘$y=zx$’. \[LogicMer\] Multiplication is positive existentially definable in ${\mathcal{M}}_p$ and in ${\mathcal{A}}_p$ over the language ${\mathcal{L}}_2'$. See Section \[SecLogicMer\] for a proof. Define the languages ${\mathcal{L}}_R^z=\{0,1,+,\cdot,z\}$, ${\mathcal{L}}_R^*=\{0,1,+,\cdot,z,{\mbox{\upshape{ord}}}\}$, ${\mathcal{L}}_2^z=\{0,1,+,P_2,f_z\}$ and ${\mathcal{L}}_2^*=\{0,1,+,P_2,f_z,{\mbox{\upshape{ord}}}\}$. In the field ${\mathcal{M}}_p$ and the ring ${\mathcal{A}}_p$, we interpret ${\mbox{\upshape{ord}}}(x)$ as ‘$x(0)=0$’. We recall that the following two theories are undecidable: the positive existential theory of ${\mathcal{A}}_p$ in the language ${\mathcal{L}}_R^z$ (see [@LipshitzPheidas]) and the positive existential theory of ${\mathcal{M}}_p$ in the language ${\mathcal{L}}_R^*$ (see [@Vidaux3]). From this and Theorem \[LogicMer\] we conclude The positive existential theory of ${\mathcal{A}}_p$ in the language ${\mathcal{L}}_2^z$ and the positive existential theory of ${\mathcal{M}}_p$ in the language ${\mathcal{L}}_2^*$ are undecidable. Some results in $p$-adic Nevanlinna Theory ========================================== First we present the notation we use for the usual functions in Nevanlinna Theory. We will work over the field ${\mathbb{C}}_p$ with absolute value $|\cdot|_p$. Write ${\mathcal{A}}_p$ for the ring of entire functions over ${\mathbb{C}}_p$ and ${\mathcal{M}}_p$ for the field of meromorphic functions over ${\mathbb{C}}_p$. We denote by $F^+$ the positive part of a function $F$ in to ${\mathbb{R}}$, that is $F^+=\max\{F,0\}$. We adopt the following notation for the standard functions in $p$-adic Nevanlinna theory, where $f=\frac{h}{g}\in{\mathcal{M}}_p$ is non-zero, and where $g,h\in{\mathcal{A}}_p$ are coprime: $$\begin{aligned} B[r]&=&\{z\in{\mathbb{C}}_p\colon |z|_p\leq r\}\\ n(r,h,0)&=&\mbox{number of zeroes of }h\mbox{ in }B[r]\mbox{ counting multiplicity}\\ n(r,f,0)&=&n(r,h,0)\\ n(r,f,\infty)&=&n(r,g,0)\\ N(r,h,0)&=&\int_0^r\frac{n(t,h,0)-n(0,h,0)}{t}dt+n(0,h,0)\log r\\ N(r,f,0)&=&N(r,h,0)\\ N(r,f,a)&=&N(r,f-a,0)\\ N(r,f,\infty)&=& N(r,g,0)\\ |h|_r&=&\max_{n\ge 0}|a_n|_pr^n\\ |f|_r&=&\frac{|h|_r}{|g|_r}\\ m(r,f,a)&=&\log^+\frac{1}{|f-a|_r}\\ m(r,f,\infty)&=&\log^+|f|_r\end{aligned}$$ We recall to the reader that for each $r>0$, the function $|\cdot|_r:{\mathcal{M}}\to {\mathbb{R}}$ is a non-archimedean absolute value satisfying $|a|_r=|a|_p$ when $a$ is constant. We will need the following results from $p$-adic Nevanlinna Theory. For a general presentation of $p$-adic complex analysis, see for example [@Robert]. First we have the *Logarithmic Derivative Lemma* (see [@CherryYe], Lemma 4.1): \[LDL\] If $n>0$ is a positive integer and $f\in{\mathcal{M}}_p$ then $$\left|\frac{f^{(n)}}{f}\right|_r\le \frac{1}{r^n}$$ where $f^{(n)}$ stands for the $n$-th derivative. We will also need the *Poisson-Jensen Formula* (see [@CherryYe], Lemma 3.1): \[PJF\] Given $f\in {\mathcal{M}}_p$, there exists a constant $C$ depending only on $f$ such that $$\log|f|_r=N(r,f,0)-N(r,f,\infty)+C.$$ As a consequence of the Poisson-Jensen Formula, we get the *First Main Theorem*: \[FMT\] Let $f\in{\mathcal{M}}_p$ be a non-constant meromorphic function and $a\in{\mathbb{C}}_p$. As $r\to\infty$ we have $$m(r,f,a)+N(r,f,a)=m(r,f,\infty)+N(r,f,\infty)+{\mathcal{O}}(1).$$ Finally, we state the *Second Main Theorem* (see [@Ru], Theorem 2.1): \[SMT\] Let $f\in{\mathcal{M}}_p$ be a non-constant meromorphic function and let $a_1,\ldots,a_q\in{\mathbb{C}}_p$ be distinct. Then, as $r\to\infty$ we have $$\sum_{i=1}^q m(r,f,a_i)\le N(r,f,\infty)+{\mathcal{O}}(1).$$ Proof of Theorem \[RepMer\] (Meromorphic Functions) {#SecMer} =================================================== The following equality will be used many times without mention within this section: $$\label{nN} N(r,f,x)=K+\int_1^r\frac{n(t,f,x)}{t}dt,\quad \mbox{for large }r.$$ It will be used systematically in order to deduce inequalities (for large $r$) about $N$ when we know inequalities about $n$ (the point is that the integral is a linear and monotone operator).\ In order to simplify the proof of Theorem \[RepMer\], we actually will prove the following equivalent result. \[mer2\] Let $h_1,\ldots,h_M$ be elements of ${\mathcal{M}}_p$ such that at least one $h_i$ is non-constant. Let $a_1,\ldots,a_M$ be $M$ distinct elements of ${\mathbb{C}}_p$. If there exist $f,g\in{\mathcal{M}}_p$, with $g$ non-zero, such that $$\label{Pizarra} h_j^2=(a_j+f)^2-g \qquad j=1,\ldots,M$$ then $M\le 34$. For the rest of this section, we will assume that we are under the hypothesis of Theorem \[mer2\]. Assuming $M\ge 35$ we will obtain a contradiction.\ First, we observe that $$\begin{aligned} \label{diferencia} h_i^2-h_j^2=(a_i-a_j)(2f+ a_i+a_j).\end{aligned}$$ \[fnocte\] The function $f$ is not constant. If $f$ is constant then so is $c_i=(a_i+f)^2$. Note that since some $h_i$ is non-constant, $g$ is non-constant. Taking $i$, $j$ and $k$ such that $c_i$, $c_j$, and $c_k$ are pairwise distinct constants, the following equality $$(h_ih_jh_k)^2=(c_i-g)(c_j-g)(c_k-g)$$ gives a non-constant meromorphic parametrization of an elliptic curve over ${\mathbb{C}}_p$, which is impossible by a theorem of Berkovich (see [@Berkovich]). \[poles\] Let $x\in {\mathbb{C}}_p$ be a pole of some $h_i$. There exists an index $k$ depending on $x$ such that for each $i\ne k$ we have (simultaneously) 1. ${\mbox{\upshape{ord}}}_x h_k\ge {\mbox{\upshape{ord}}}_x h_i$; 2. ${\mbox{\upshape{ord}}}_x f\ge 2{\mbox{\upshape{ord}}}_x h_i$; 3. ${\mbox{\upshape{ord}}}_x g\ge 4{\mbox{\upshape{ord}}}_x h_i$; 4. ${\mbox{\upshape{ord}}}_x h_i={\mbox{\upshape{ord}}}_x h_j$ for all $j\ne k$; and 5. ${\mbox{\upshape{ord}}}_x h_i\le -1$. Moreover, for each $i$ we have $$\label{polesord} \min\{{\mbox{\upshape{ord}}}_x h_i,0\}\geq\frac{1}{M-1}\sum_l \min\{{\mbox{\upshape{ord}}}_x h_l,0\}$$ and, there exists a positive constant $K$ such that for large enough $r$ and for each $i$ we have $$\label{polesN} N(r,h_i,\infty)\le \frac{1}{M-1}\sum_l N(r,h_l,\infty).$$ Let $i_0$ be an index such that $h_{i_0}$ has a pole at $x$. First suppose that *all* $h_i$ have the same order at $x$ (hence negative). In this case, Items (1), (4) and (5) hold trivially, Item (2) comes from Equation , and Item (3) comes from Equation . Indeed for Item (3) we have $$\begin{aligned} {\mbox{\upshape{ord}}}_x(g)&\geq2\min\{{\mbox{\upshape{ord}}}_x(h_i),{\mbox{\upshape{ord}}}_x(f+a_i)\}\\ &=2\min\{{\mbox{\upshape{ord}}}_x(h_i),{\mbox{\upshape{ord}}}_x(f)\}\\ &=2\min\{{\mbox{\upshape{ord}}}_x(h_i),2{\mbox{\upshape{ord}}}_x (h_i)\}\\ &\geq4{\mbox{\upshape{ord}}}_x h_i, \end{aligned}$$ where the last inequality comes from Item (2). The other case is when not all $h_i$ have the same order at $x$. Choose $k$ such that item (1) holds true. By Equation for indices $k$ and any $i\ne k$, Item (4) holds true. If $i_0=k$ then all $h_i$ have a pole at $x$ (by maximality of $k$), and if $i_0\ne k$ then by Item (4), for all $i\ne k$, $h_i$ has a pole at $x$. Hence Item (5) holds true. Items (2) and (3) for $i\ne k$ follow as in the previous case. Finally, by Items (1), (4) and (5), and observing that ${\mbox{\upshape{ord}}}_x h_k$ could be positive, we have for each $i$ $$(M-1)\min\{{\mbox{\upshape{ord}}}_x h_i,0\}=\sum_{l\ne k}\min\{{\mbox{\upshape{ord}}}_x h_l,0\}\geq\sum_l \min\{{\mbox{\upshape{ord}}}_x h_l,0\}.$$ Summing for $x\in B[r]$ we obtain $$(M-1)n(r,h_i,\infty)\leq \sum_l n(r,h_l,\infty).$$ which gives the inequality by Equation . \[asterisco\] The following inequality holds $$\sum_{n=1}^{M} \log|h_n|_r +\frac{1}{M-1}\sum_{n=1}^{M} N(r,h_n,\infty) \ge -\frac{1}{2}N(r,f,\infty)+{\mathcal{O}}(1).$$ By the Second Main Theorem \[SMT\], we have for each $i\in\{1,\ldots,M\}$ $$\begin{aligned} -N(r,f,\infty)+{\mathcal{O}}(1) \le -\sum_{j\ne i}\log^+\left|\frac{1}{f+\frac{a_i+a_j}{2}}\right|_r \le \sum_{j\ne i}\log\left|f+\frac{a_i+a_j}{2}\right|_r.\end{aligned}$$ Since by Equation we have $$h_i^2-h_j^2=2(a_i-a_j)\left(f+\frac{a_i+a_j}{2}\right),$$ we deduce $$-N(r,f,\infty)+{\mathcal{O}}(1) \le \sum_{j\ne i}\log\left|h_i^2-h_j^2\right|_r.$$ If for a given $r$, $i_r$ is an index such that $|h_i|_r$ is minimal, then $$\begin{aligned} \frac{1}{2}\sum_{j\ne i_r}\log\left|h_{i_r}^2-h_j^2\right|_r&\le & \sum_{j\ne i_r}\log\left|h_j\right|_r \\ &=&C+ \sum_{j\ne i_r} \left(N(r,h_j,0)-N(r,h_j,\infty)\right)\\ &\le& C+N(r,h_{i_r},\infty)+ \sum_n \left(N(r,h_n,0)-N(r,h_n,\infty)\right)\\ &=&C'+N(r,h_{i_r},\infty)+ \sum_n \log\left|h_n\right|_r\\ &\le& C''+\frac{1}{M-1}\sum_n N(r,h_n,\infty)+ \sum_n \log\left|h_n\right|_r\end{aligned}$$ where the first and second equalities are given by the Poisson-Jensen Formula \[PJF\], the third inequality is given by Lemma \[poles\] (see Equation ), and $C$, $C'$, $C''$ are fixed constants (not depending on $r$ nor on $i_r$). Finally we have $$-\frac{1}{2}N(r,f,\infty)+{\mathcal{O}}(1) \le \frac{1}{2}\sum_{j\ne i_r}\log\left|h_{i_r}^2-h_j^2\right|_r\le \sum \log\left|h_n\right|_r + \frac{1}{M-1}\sum N(r,h_n,\infty) + C''$$ for each $r$ large enough, and the lemma is proven. \[L\] The following inequalities hold: $$n(r,f,\infty)\le \frac{2}{M-1}\sum_n n(r,h_n,\infty)$$ and $$\sum_n N(r,h_n,0)\ge \frac{M-3}{M-1}\sum_n N(r,h_n,\infty)+{\mathcal{O}}(1).$$ Observe that by Lemma \[poles\] (Item (2) and Equation ) we have $$(M-1)n(r,f,\infty)\le 2\sum n(r,h_j,\infty),$$ hence $$(M-1)N(r,f,\infty)\le 2\sum N(r,h_n,\infty)+{\mathcal{O}}(1).$$ The second formula comes immediately by Lemma \[asterisco\] and the Poisson-Jensen Formula \[PJF\]. \ The equations $$\begin{aligned} h_n^2+g&=&(a_n+f)^2\\ 2h'_nh_n+g'&=&2f'(a_n+f)\end{aligned}$$ are directly deduced by reordering and differentiating the one given in the hypothesis. From this we deduce $$\begin{aligned} (2h'_nh_n+g')^2=4f'^2(h_n^2+g)\end{aligned}$$ hence $$\begin{aligned} g'^2-4f'^2g=4h_n(h_nf'^2-h'^2_nh_n-h_n'g').\end{aligned}$$ Writing $$\begin{aligned} \Delta&=&g'^2-4f'^2g\\ \Delta_n&=&h_nf'^2-h'^2_nh_n-h'_ng'\end{aligned}$$ we have $$\label{Deltas} \Delta=4h_n\Delta_n.$$ \[C1\] If $\Delta$ is not identically zero, then $$N(r,\Delta,0)\ge \frac{1}{2}\sum N(r,h_n,0)-\frac{8}{M-1}\sum N(r,h_n,\infty)+{\mathcal{O}}(1).$$ On the one hand, for a given $x\in{\mathbb{C}}_p$ suppose $f$ has a pole at $x$ and $h_j(x)=0$ for some index $j$. Set $l={\mbox{\upshape{ord}}}_x (h_j)$ and $m={\mbox{\upshape{ord}}}_x(f)$. Note that ${\mbox{\upshape{ord}}}_x(g)=2m$ because $h_j(x)=0$ (see Equation ). Write $$h_j=u_l(z-x)^l+u_{l+1}(z-x)^{l+1}+\cdots,$$ $$f=v_m(z-x)^m+v_{m+1}(z-x)^{m+1}+\cdots$$ and $$g=w_{2m}(z-x)^{2m}+w_{2m+1}(z-x)^{2m+1}+\cdots$$ for the Laurent series of $h_j$, $f$ and $g$ at $x$. Observe that $w_{2m}=v^2_m$. The first term of the Laurent series at $x$ for respectively $h_jf'^2$, $h'^2_jh_j$ and $h'_jg'$ is, respectively, $$\begin{aligned} m^2u_lv^2_m(z-x)^{l+2m-2}\\ l^2u_l^3(z-x)^{3l-2}\\ 2lm u_lv^2_{m}(z-x)^{l+2m-2}\end{aligned}$$ hence ${\mbox{\upshape{ord}}}_x\Delta_j=l+2m-2$ since $2l\ne m$. Therefore, we have $${\mbox{\upshape{ord}}}_x\Delta=2(l+m-1).$$ On the other hand, if $x\in{\mathbb{C}}_p$ is not a pole of $f$ and is a zero of some $h_j$, then we have $${\mbox{\upshape{ord}}}_x\Delta\geq{\mbox{\upshape{ord}}}_x(h_j)$$ because by Equation , $g$ does not have a pole, hence $\Delta_j$ does not have a pole and we conclude by Equation . Let $A_r$ be the set of $x\in B[r]$ such that $f$ has not a pole at $x$ and $h_j(x)=0$ for some index $j$, and let $B_r$ be the set of $x\in B[r]$ such that $f$ has a pole at $x$ and $h_j(x)=0$ for some index $j$. Observe that, by Equation , no three of the $h_n$ can share a zero (we use it for the fifth inequality below). We have then $$\begin{aligned} n(r,\Delta,0)&\ge& \sum_{x\in A_r}{\mbox{\upshape{ord}}}_x\Delta + \sum_{x\in B_r}{\mbox{\upshape{ord}}}_x\Delta\\ &\ge& \sum_{x\in A_r}\max_{h_i(x)=0} {\mbox{\upshape{ord}}}_x(h_i) + \sum_{x\in B_r}\max_{h_i(x)=0} 2({\mbox{\upshape{ord}}}_x(h_i)+{\mbox{\upshape{ord}}}_x(f)-1)\\ &\ge& \sum_{x\in A_r\cup B_r}\max_{h_i(x)=0} {\mbox{\upshape{ord}}}_x(h_i) + 2\sum_{x\in B_r} \max_{h_i(x)=0}({\mbox{\upshape{ord}}}_x(f)-1)\\ &=& \sum_{x\in A_r\cup B_r}\max_{h_i(x)=0} {\mbox{\upshape{ord}}}_x(h_i) + 2\sum_{x\in B_r} ({\mbox{\upshape{ord}}}_x(f)-1)\\ &\ge& \sum_{x\in A_r\cup B_r}\max_{h_i(x)=0} {\mbox{\upshape{ord}}}_x(h_i) + 4\sum_{x\in B_r} {\mbox{\upshape{ord}}}_x(f)\\ &\ge& \frac{1}{2}\sum_i n(r,h_i,0) - 4n(r,f,\infty)\\ &\ge& \frac{1}{2}\sum_i n(r,h_i,0) - \frac{8}{M-1}\sum n(r,h_i,\infty)\end{aligned}$$ where the last inequality comes from Lemma \[L\]. The result follows. \[C2\] If $\Delta$ is not identically zero, then $$N(r,\Delta,\infty)\le \frac{8}{M-1}\sum N(r,h_n,\infty)+{\mathcal{O}}(1).$$ Suppose that some $x\in{\mathbb{C}}_p$ is a pole of $\Delta$. Then, by definition of $\Delta$, it is a pole of $f$ or of $g$. If none of the $h_i$ has a pole at $x$ then by Equation $f$ does not have a pole, and by Equation , $g$ does not have a pole, which contradicts our hypothesis. Therefore, some $h_i$ has a pole at $x$. Take $k$ as in Lemma \[poles\]. For each index $i\ne k$ we have (observing that ${\mbox{\upshape{ord}}}_x(h_i)\leq -1$ and that if $g'=0$ then ${\mbox{\upshape{ord}}}_x h_i'g'$ is infinite) $$\begin{aligned} {\mbox{\upshape{ord}}}_x\Delta &\ge& {\mbox{\upshape{ord}}}_x h_i + \min\{{\mbox{\upshape{ord}}}_x h_if'^2,{\mbox{\upshape{ord}}}_x h'^2_ih_i,{\mbox{\upshape{ord}}}_x h'_ig'\}\\ &\ge& {\mbox{\upshape{ord}}}_x h_i +\min\{7{\mbox{\upshape{ord}}}_x h_i,5{\mbox{\upshape{ord}}}_x h_i,7{\mbox{\upshape{ord}}}_x h_i\}\\ &=&8{\mbox{\upshape{ord}}}_x h_i.\end{aligned}$$ Hence, using the Lemma \[poles\] (Equation ) we have $${\mbox{\upshape{ord}}}_x\Delta\ge \frac{8}{M-1}\sum_l \min\{{\mbox{\upshape{ord}}}_x h_l,0\}.$$ Write $D_r$ for the set of poles of $\Delta$ in $B[r]$. We have $$\begin{aligned} n(r,\Delta,\infty)&=&\sum_{x\in D_r}-{\mbox{\upshape{ord}}}_x\Delta \\ &\le& \frac{8}{M-1}\sum_{x\in D_r}\sum_l \max\{-{\mbox{\upshape{ord}}}_x h_l,0\}\\ &\le& \frac{8}{M-1}\sum_l \sum_{x\in B[r]} \max\{-{\mbox{\upshape{ord}}}_x h_l,0\}\\ &=&\frac{8}{M-1}\sum_l n(r,h_l,\infty).\end{aligned}$$ and the result follows. \[PeloCorto\] 1. For each $r>0$, there exists an index $k_r$ such that $|h_{k_r}|_r$ is minimal. 2. There exists a positive constant $K_f$ such that, for any $r>0$ and for all $i\ne k_r$, we have $$\log |f|_r\le \max\{0,2\log |h_i|_r\} + K_f.$$ 3. There exists a positive constant $K_g$ such that, for any $r>0$ and for all $i\ne k_r$, we have $$\log |f|_r\le \max\{0,4\log |h_i|_r\} + K_g.$$ Item (1) is immediate since for each $r$, the set $\{|h_i|_r\colon i=1,\dots,M\}$ is finite. Let us prove Item (2). There exists a positive constant $K'>1$ such that for each $r>0$, $i$ and $j$, we have $$\label{Pelado} |2f|_r\le |2f+a_i+a_j|_r+|a_i+a_j|_r\leq K'+ |2f+a_i+a_j|_r.$$ On the other hand, by Equation there exists a constant $K''>1$ such that, for any $r>0$, $i\ne k_r$ and $j$, we have $$\begin{aligned} |2f+a_i+a_j|_r&=\left|\frac{h_i^2-h_{k_r}^2}{a_i-a_{k_r}}\right|_r&\qquad\\ &\leq \left|\frac{h_i^2}{a_i-a_{k_r}}\right|_r &\qquad \textrm{(by Item (1))}\\ &\leq K''|h_i^2|_r& \end{aligned}$$ hence by Equation $$|2f|_r\leq K''|h_i^2|_r+K'\leq K'' \max\{|h_i^2|_r,1\}+ K'.$$ Therefore, we have $$\begin{aligned} \log|f|_r&\leq \log(K'' \max\{|h_i^2|_r,1\}+ K')-\log|2|_r\\ &\leq \log(K'' \max\{|h_i^2|_r,1\})+\log K'+\log2-\log|2|_r\\ &\leq \max\{2\log |h_i|_r,0\} + K_f \end{aligned}$$ with $K_f$ is a positive constant greater than $\log K''+\log K'+\log2-\log|2|_r$, and where the second inequality comes from the fact that for all real numbers $x,y\geq 1$, we have $\log(x+y)\leq\log x+\log y+\log2$ (just write $(1-x)(y-1)\le 0$). Finally, we prove Item (3). By Equation and Item (2), for each $i\ne k_r$ we have $$\begin{aligned} \log |g|_r&=\log|(f+a_i)^2-h_i^2|_r \\ &\le \log\left(\max\{|h_i^2|_r, |f^2|_r,|2a_if|_r,|a_i^2|_r\}\right)\\ &\le \log\left(\max\{|h_i^2|_r, |f^2|_r,|a_i^2|_r\}\right)\\ &\le \max\{2\log |h_i|_r, 0, 2\log |f|_r\} + \max\{|a_i^2|_r\} \\ &\le \max\{2\log |h_i|_r, 0, 2\max\{0,2\log |h_i|_r\} + 2K_f\} + \max\{|a_i^2|_r\} \\ &\le \max\{2\log |h_i|_r+2K_f,2K_f, 4\log |h_i|_r+2K_f\} + \max\{|a_i^2|_r\} \\ &\le \max\{0,4\log |h_i|_r\} + K_g \end{aligned}$$ where $K_g$ is a fixed positive constant bigger than $2K_f+\max\{|a_i^2|_r\}$, and where the second inequality comes from the following: $$|2a_if|_r\le|a_i|_r|f|_r\leq\frac{|a_i^2|_r+|f^2|_r}{2}\leq\max\{|a_i^2|_r,|f^2|_r\}.$$ \[C3\] If $\Delta$ is not identically zero, then $$\log |\Delta|_r\le \frac{6M-2}{M(M-1)}\sum \log |h_n|_r+\frac{8}{(M-1)^2}\sum N(r,h_n,\infty)-2\log r+{\mathcal{O}}(1).$$ By the Poisson-Jensen Formula \[PJF\] and Lemma \[poles\] (Equation ) we have for $r$ large enough and for each $i$ $$\begin{aligned} \log|h_{i}|_r&= N(r,h_{i},0)-N(r,h_{i},\infty)+C\\ &\ge -N(r,h_{i},\infty)+C\\ &\ge -\frac{1}{M-1}\sum_n N(r,h_n,\infty)+C' \end{aligned}$$ for some constants $C,C'$. Given $r>0$ take $k_r$ as in Lemma \[PeloCorto\]. Choose $i_r$ such that $|h_{i_r}|_r$ is minimal in $\{|h_i|_r\colon i\ne k_r\}$, and note that $$\log |h_{i_r}|_r\le \frac{1}{M-1}\sum_{i\ne k_r}\log |h_i|_r.$$ By Item (2) in Lemma \[PeloCorto\], we have for each $r$ large enough $$\begin{aligned} \log |f|_r&\le \max\{0,2\log |h_{i_r}|_r\} + K_f\\ &\le \max\left\{0,\frac{2}{M-1}\sum_{i\ne k_r}\log |h_i|_r\right\} + K_f\\ &\le \max\left\{0,\frac{2}{M-1}\sum \log\left|h_n\right|_r + \frac{2}{(M-1)^2}\sum N(r,h_n,\infty)\right\} - \frac{2C'}{M-1} + K_f \end{aligned}$$ and by Item (3) in Lemma \[PeloCorto\] we have for each $r$ large enough $$\begin{aligned} \log |g|_r&\le \max\{0,4\log |h_{i_r}|_r\} + K_g\\ &\le \max\left\{0,\frac{4}{M-1}\sum_n \log\left|h_n\right|_r + \frac{4}{(M-1)^2}\sum_n N(r,h_n,\infty)\right\} - \frac{4C'}{M-1}+K_g. \end{aligned}$$ Hence, for large $r$ we get $$\begin{aligned} \label{cotaf}\log |f|_r &\le& \max\left\{0,\frac{2}{M-1}\sum_n \log\left|h_n\right|_r + \frac{2}{(M-1)^2}\sum_n N(r,h_n,\infty)\right\} + {\mathcal{O}}(1)\\ \label{cotag}\log |g|_r &\le& \max\left\{0,\frac{4}{M-1}\sum_n \log\left|h_n\right|_r + \frac{4}{(M-1)^2}\sum_n N(r,h_n,\infty)\right\} + {\mathcal{O}}(1).\\\end{aligned}$$ Since $\Delta$ is not the zero function, from Lemma \[LDL\] (Logarithmic Derivative Lemma) we have for each index $n$ $$\begin{aligned} |\Delta|_r\le |h_n|_r\max\{|h_nf'^2|_r,|h'^2_nh_n|_r,|h'_ng'|_r\}\le \frac{1}{r^2}|h_n|^2_r\max\{|f|^2_r,|h_n|^2_r,|g|_r\}\end{aligned}$$ and by Equation for each $n$ holds $$2\log |h_n|_r\le \max\{2\log |f|_r, 0 , \log |g|_r\}+{\mathcal{O}}(1)$$ therefore we have for each $n$ $$\begin{aligned} \log |\Delta|_r\le \log\left(\frac{1}{r^2}|h_n|^2_r\right)+ \max\{2\log |f|_r, 0 , \log |g|_r\}+{\mathcal{O}}(1).\end{aligned}$$ Since this last expression is true *for each* $n$, we have $$\begin{aligned} \log |\Delta|_r\le \frac{2}{M}\sum\log |h_n|_r - 2\log r+ \max\{2\log |f|_r, 0 , \log |g|_r\}+{\mathcal{O}}(1).\end{aligned}$$ Note that by equations and $$\max\{2\log |f|_r, 0 , \log |g|_r\}\le \max\left\{0,\frac{4}{M-1}\sum \log\left|h_n\right|_r + \frac{4}{(M-1)^2}\sum N(r,h_n,\infty)\right\} + {\mathcal{O}}(1)$$ and by Lemma \[asterisco\] we have that this last expression is less than or equal to $$\frac{4}{M-1}\sum \log\left|h_n\right|_r + \frac{4}{(M-1)^2}\sum N(r,h_n,\infty)+\frac{2}{M-1}N(r,f,\infty)+{\mathcal{O}}(1).$$ Therefore $$\begin{aligned} \log |\Delta|_r\le \left(\frac{2}{M}+\frac{4}{M-1}\right)\sum\log |h_n|_r - 2\log r + \frac{4}{(M-1)^2}\sum N(r,h_n,\infty)\\ + \frac{2}{M-1}N(r,f,\infty)+{\mathcal{O}}(1)\end{aligned}$$ Finally, we bound $N(r,f,\infty)$ using Lemma \[L\] and the result follows. \[eqdif\] $\Delta=0$ for $M\ge 35$. The proof goes by contradiction, so we assume $\Delta$ is not identically zero. Consider the equation $$\log|\Delta|_r=N(r,\Delta,0)-N(r,\Delta,\infty) +{\mathcal{O}}(1).$$ Lemmas \[C1\], \[C2\] and \[C3\] allow us to bound $\log |\Delta|_r$ above and below, obtaining $$\begin{aligned} \frac{6M-2}{M(M-1)}\sum \log |h_n|_r+\frac{8}{(M-1)^2}I-2\log r \ge \frac{1}{2}Z-\frac{8}{M-1}I - \frac{8}{M-1}I +{\mathcal{O}}(1) \end{aligned}$$ where we write $Z=\sum N(r,h_n,0)$ and $I=\sum N(r,h_n,\infty)$. Observe that $$\sum \log |h_n|_r= Z-I+{\mathcal{O}}(1)$$ by Poisson-Jensen Formula \[PJF\]. This and Lemma \[L\] give $$\begin{aligned} -2\log r &\ge& \left(\frac{1}{2}-\frac{6M-2}{M(M-1)}\right)Z +\left(\frac{6M-2}{M(M-1)} -\frac{16}{M-1} -\frac{8}{(M-1)^2}\right)I+{\mathcal{O}}(1)\\ &\ge& \left(\left(\frac{1}{2}-\frac{6M-2}{M(M-1)}\right)\frac{M-3}{M-1} -\frac{10M^2-2}{M(M-1)^2} \right)I+{\mathcal{O}}(1)\\ &=&\frac{M^2-35M+8}{2M(M-1)}I+{\mathcal{O}}(1).\end{aligned}$$ A contradiction for $M\ge 35$. Finally, we have that $\Delta$ is the zero function, $f$ is not a constant and $g$ is non-zero. We get the equation $g'^2=4f'^2g$, which implies that exists a meromorphic function $u$ such that $g=u^2$ and $u'^2=f'^2$. Hence $u= \alpha f + b$ for certain $\alpha\in\{-1,1\}$ and $b\in{\mathbb{C}}_p$, and we obtain $$\begin{aligned} h_n^2&=&(a_n+f)^2-(\alpha f+b)^2\\ &=&(a_n+f)^2-(f+\alpha b)^2\\ &=&(a_n-\alpha b)(a_n+\alpha b+2f).\end{aligned}$$ From this we get $$\begin{aligned} \left(\frac{h_ih_jh_k}{\sqrt{(a_i-\alpha b)(a_j-\alpha b)(a_k-\alpha b)}}\right)^2=(a_i+\alpha b +2f)(a_j+\alpha b +2f)(a_k+\alpha b +2f).\end{aligned}$$ This is a contradiction because $f$ is not a constant. Therefore the Theorem \[mer2\] is proven. Some geometric results {#geometric} ====================== This section contains most of the geometric results that we will use in the next two sections. The arguments given here essentially are adaptations of the arguments given by Vojta in [@Vojta2]. Anyway, we prefer to perform most of the computations in order to give a clear presentation. During the whole section, we assume that the base field is ${\mathbb{C}}$, and we write $g(X)$ for the genus of the curve $X$. Let $S=(\delta_2,\delta_3,\ldots)$ be a sequence in ${\mathbb{C}}^*$ with pairwise distinct terms. Set $X_2={\mathbb{P}}^2({\mathbb{C}})$ and for $n>2$ let $X_n\subset{\mathbb{P}}^n({\mathbb{C}})$ be the algebraic set defined by the equations $$\begin{aligned} \label{defining} \delta_2 x^2_i=\delta_i\delta_2(\delta_i-\delta_2)x^2_0-(\delta_i-\delta_2)x^2_1+\delta_ix^2_2\end{aligned}$$ as $i$ ranges from $3$ to $n$. If $[x_0:\cdots:x_n]\in X_n$ is easy to see that at most $2$ of the $x_i$ can be zero, hence $X_n\subset U_0\cup U_1\cup U_2$ where $U_i$ is the open set $\{x_i\ne 0\}$. \[Xn\] $X_n$ is a smooth surface in ${\mathbb{P}}^n$, contains the lines $$\begin{aligned} \label{triviallines} \pm x_1=\pm x_2 -\delta_2 x_0=\cdots=\pm x_n - \delta_n x_0\end{aligned}$$ and has canonical sheaf ${\mathcal{O}}_{X_n}(n-5)$. In particular, $X_n$ is of general type for $n\ge 6$. Observe that, for $[x_0:\cdots,x_n]\in X_n\cap U_0$ the matrix $$\begin{aligned} \label{matrix} \left[\begin{array}{ccccccc} (\delta_3-\delta_2)x_1 &-\delta_3 x_2 &\delta_2 x_3 &0 &\cdots &0 \\ (\delta_4-\delta_2)x_1 &-\delta_4 x_2 &0 &\delta_2 x_4 &\ddots &0 \\ \vdots &\vdots &\vdots &\ddots &\ddots &\vdots \\ (\delta_n-\delta_2)x_1 &-\delta_n x_2 &0 &0 &\cdots &\delta_2 x_n \end{array}\right]\end{aligned}$$ has rank $n-2$; indeed, we have $3$ cases depending on the number of zeroes between $x_3,\ldots, x_n$: 1. No zero: trivial. 2. One zero: at least one of the first two columns has no zero. 3. Two zeroes: suppose that $x_i=x_j=0$ where $3\le i<j\le n$, then no entry in the first two columns is zero. Therefore $$\begin{aligned} \left|\begin{array}{cc} (\delta_i-\delta_2)x_1 &-\delta_i x_2 \\ (\delta_j-\delta_2)x_1 &-\delta_j x_2 \end{array}\right|=\delta_2x_1x_2(\delta_j-\delta_i)\ne 0.\end{aligned}$$ hence, $X_n$ is nonsingular at each point in $X_n\cap U_0$. The verification that $X_n$ is nonsingular at each point in $X_n\cap U_1$ and $X_n\cap U_2$ is quite similar, but the determinants at case (3) are $$\begin{aligned} \left|\begin{array}{cc} -\delta_i\delta_2(\delta_i-\delta_2)x_0 &-\delta_i x_2 \\ -\delta_j\delta_2(\delta_j-\delta_2)x_0 &-\delta_j x_2 \end{array}\right|=\delta_2\delta_i\delta_jx_0x_2(\delta_j-\delta_i)\ne 0\end{aligned}$$ and $$\begin{aligned} \left|\begin{array}{cc} -\delta_i\delta_2(\delta_i-\delta_2)x_0 &(\delta_i-\delta_2)x_1 \\ -\delta_j\delta_2(\delta_j-\delta_2)x_0 &(\delta_j-\delta_2)x_1 \end{array}\right|=\delta_2x_0x_1(\delta_j-\delta_i)(\delta_j-\delta_2)(\delta_i-\delta_2)\ne 0\end{aligned}$$ respectively. Therefore $X_n$ is a smooth surface in ${\mathbb{P}}^n$.\ The claim about the lines is an easy computation (look at $U_0\cap X_n$).\ Finally, since $X_n$ is a complete intersection surface in ${\mathbb{P}}^n$ defined as the intersection of $n-2$ smooth hypersurfaces of degree 2, its canonical sheaf is ${\mathcal{O}}(2(n-2)-n-1)={\mathcal{O}}(n-5)$. Define the trivial lines of $X_n$ as the lines . Observe that for $n\ge 3$ the rational map $[x_0:\cdots:x_{n}]\mapsto [x_0:\cdots:x_{n-1}]$ induces a finite morphism $\pi_n:X_n\rightarrow X_{n-1}$ of degree $2$ ramified along the curve $C_n\subset X_n$ defined by $x_n=0$. This curve is nonsingular; indeed, if $[x_0:\cdots:x_n]\in C_n=X_n\cap\{x_n=0\}$ then at most one of the $x_0,\ldots,x_{n-1}$ can be zero and the remaining verification can be performed as in the proof of Lemma \[Xn\] for cases (2) and (3) since $x_n=0$, but adding the extra row $(0,\ldots,0,1)$ to each matrix.\ Define $\phi_n=\pi_3\circ\cdots\circ\pi_n$, we note that the image of $C_n$ in $X_2$ via $\phi_n$ is $$\begin{aligned} \label{CninP2} \delta_n\delta_2(\delta_n-\delta_2)x^2_0-(\delta_n-\delta_2)x^2_1+\delta_n x^2_2=0\end{aligned}$$ Let $X$ be a smooth surface over ${\mathbb{C}}$ and let ${\mathcal{L}}$ be an invertible sheaf on $X$. Take a section $\omega\in H^0(X,{\mathcal{L}}\otimes S^2(\Omega^1_X))$. Let $Y\subset X$ be a curve with normalization $i:\tilde{Y}\rightarrow Y$. We say that $Y$ is *$\omega-$integral* if $i^*\omega\in H^0(\tilde{Y},i^*({\mathcal{L}})\otimes S^2(\Omega^1_{\tilde{Y}}))$ vanishes identically on $\tilde{Y}$. On $U_0\subset{\mathbb{P}}^2=X_2$ define $$\begin{aligned} \omega=x_1x_2dx_1\otimes dx_1+(\delta^2_2-x_1^2-x_2^2)dx_1\otimes dx_2+ x_1x_2dx_2\otimes dx_2\end{aligned}$$ Note that, after the change of variables $y_0=x_0/x_1,y_2=x_2/x_1$, on $U_0\cap U_1$ we have $$\begin{aligned} \omega=\frac{1}{y_0^5}\left(\delta_2^2y_0y_2dy_0\otimes dy_0+(1-\delta^2_2y_0^2-y_2^2)dy_0\otimes dy_2+ y_0y_2dy_2\otimes dy_2\right)\end{aligned}$$ hence $\omega$ extends to a section $$\omega_2\in H^0(X_2,{\mathcal{O}}_{X_{2}}(5)\otimes S^2(\Omega^1_{X_2})).$$ \[2integral\] Write $[x_0:x_1:x_2]$ for homogeneous coordinates on ${\mathbb{P}}^2=X_2$. The only $\omega_2-$integral curves on $X_2$ are 1. $x_0=0$, $x_1=0$, and $x_2=0$ 2. the four trivial lines 3. the conics $\delta_2c(c-\delta_2)x^2_0 - (c-\delta_2)x^2_1 + cx^2_2=0$ for $c\ne 0,\delta_2$. It is easy to see that curves of type (1) and (2) are $\omega_2-$integral. Let’s show that curves of type (3) are $\omega_2$-integral. If we look at the affine chart $U_0$, on a curve of type (3) we have $$(c-\delta_2)x_1dx_1=cx_2dx_2$$ hence $$\begin{aligned} \omega_2&=&\left( \frac{c^2x^3_2}{(c-\delta_2)^2x_1} + \frac{cx_2}{(c-\delta_2)x_1}(\delta^2_2-x^2_1-x^2_2) + x_1x_2\right)dx_2\otimes dx_2\\ &=&\left( c^2x^2_2 + c(c-\delta_2)(\delta^2_2-x^2_1-x^2_2) + (c-\delta_2)^2x_1^2\right)\frac{x_2dx_2\otimes dx_2}{(c-\delta_2)^2x_1}\\ &=&\left( \delta^2_2c(c-\delta_2) - \delta_2(c-\delta_2)x^2_1 + \delta_2cx^2_2 \right)\frac{x_2dx_2\otimes dx_2}{(c-\delta_2)^2x_1}\\ &=&\delta_2\left( \delta_2c(c-\delta_2) - (c-\delta_2)x^2_1 + cx^2_2 \right)\frac{x_2dx_2\otimes dx_2}{(c-\delta_2)^2x_1}=0.\end{aligned}$$ Conversely, let $Y$ a $\omega_2-$integral curve on $X_2$ not of type (1) or (2), we will show that $Y$ is of type (3). Let $P\in Y$ be a regular point of $Y$ not in a line of type (1) nor (2). As $Y$ is regular at $P$, in some neighborhood of $P$ one can assume that one affine coordinate is function of the other, say $x_1=x_1(x_2)$. Since $Y$ is $\omega_2-$integral, we get a quadratic ordinary differential equation for $x_1$, hence there are $2$ local solutions at $P$. But exactly $2$ curves of type (3) pass through $P$. Therefore $Y$ is locally of type (3) on a dense set of points, so $Y$ is of type (3). Observe that the image of $C_n$ in $X_2$ is $\omega_2-$integral (see Equation ).\ Write $\omega'_n=\phi_n^*\omega_2$ and note that $$\omega'_n\in H^0(X_n,{\mathcal{O}}_{X_{n}}(5)\otimes S^2(\Omega^1_{X_n}))$$ because $\pi_n^*{\mathcal{O}}_{X_{n-1}}(1)={\mathcal{O}}_{X_n}(1)$ for each $n\ge 3$. \[nintegral\] Let $n\ge 6$ be an integer. The only $\omega'_n-$integral curves on $X_n$ are 1. the pull-backs via $\phi_n$ of the coordinate axes on $X_2$ to $X_n$ 2. the trivial lines 3. the pull-backs via $\phi_n$ of the conics $\delta_2c(c-\delta_2)x^2_0 - (c-\delta_2)x^2_1 + cx^2_2=0$ for $c\ne 0,\delta_2$. Moreover, these curves are nonsingular and the only of them with genus $\le 2^{n-3}$ are the trivial lines, with genus $0$. Let $Y\subset X_n$ be a $\omega'_n-$integral curve. Write $Z=\phi_n(Y)$ and $Y'=\phi^*_n(Z)$. Note that $Z$ is $\omega_2-$integral, hence we have 3 cases by Lemma \[2integral\].\ Suppose $Z=\{x_j=0\}\subset X_2$ is a coordinate axe. $Y'=X_n\cap\{x_j=0\}$ is nonsingular by a verification similar to the one done for $C_n$. Since that $Z$ meets all the curves $\phi(C_i)$ for $i=3,\ldots,n$ and they for the branch divisor, we have that $Y'$ is connected. Hence $Y'=Y$ and $Y$ is nonsingular. Note that $\phi_n|_Y:Y\rightarrow Z$ has degree $2^{n-2}$ and is ramified at $2^{n-2}(n-2)$ points, hence $g(Y)=2^{n-3}(n-4)+1$ by the Hurwitz formula.\ Now suppose $Z$ is a trivial line in $X_2$. Replacing the value of $x_2$ in terms of $x_1$ in the defining equations of $X_n$ we obtain that $Y$ is a trivial line, with genus $0$.\ Finally suppose $Z$ is a curve of type (3) in Lemma \[2integral\]. By the same argument as in the first case, $Y'$ is connected. One can show that $Y'$ is nonsingular by a direct computation (on the affine chart $U_0$ we add the row $((c-\delta_2)x_1,-cx_2,0,\ldots,0)$ in \[matrix\], and for $U_1,U_2$ the computation is similar) therefore $Y=Y'$. Consider the map $\psi_n=\phi_n|_Y:Y\rightarrow Z$. If $Y$ lies above one of the curves $C_i$ then $\deg(\psi_n)=2^{n-3}$ and if $Y$ does not lie above any $C_i$ then $\deg(\psi_n)=2^{n-2}$. Anyway, $\phi_n$ is ramified at least in $(n-3)\cdot 4\cdot 2^{n-4}=2^{n-2}(n-3)$ points and $g(Z)=0$, thus for $n\ge 6$ by the Hurwitz formula we have $$\begin{aligned} g(Y)>-2^{n-2}+2^{n-3}(n-3)=2^{n-3}(n-5)\ge 2^{n-3}.\end{aligned}$$ \[vojta\] Let $\pi:X'\rightarrow X$ be a finite morphism of smooth projective surfaces over ${\mathbb{C}}$, ramified along a curve $Y\subset X'$. Let ${\mathcal{L}}$ be a invertible sheaf on $X$, and take a section $\omega\in H^0(X,{\mathcal{L}}\otimes S^2(\Omega^1_X))$. If $\pi(Y)$ is $\omega-$integral, then $\pi^*\omega\in H^0(X',\pi^*{\mathcal{L}}\otimes S^2(\Omega^1_{X'}))$ vanishes identically on $Y$. This is a particular case of [@Vojta2] Lemma 2.10. We recall to the reader that $\omega'_n=\phi_n^*\omega_2$. \[omegan\] Define $\omega'_2=\omega_2$. The sections $\omega'_n$ determine sections $$\omega_n\in H^0(X_n,{\mathcal{O}}_{X_{n}}(7-n)\otimes S^2(\Omega^1_{X_n}))$$ such that each $\omega_n-$integral curve is a $\omega'_n-$integral curve. Moreover, the $\omega_n-$integral curves are the same as the $\omega'_n-$integral curves, with the only possible exception of $\omega'-$integral curves lying over $C_3,\ldots,C_n$. By induction. The case $n=2$ is clear. Assume it for $n=m-1$ with $m>2$. Note that $\pi_m(C_m)$ does not lie over any of the curves $C_3,\ldots,C_{m-1}$ because they have different images in $X_2$, hence $\pi_m(C_m)$ is $\omega_{m-1}-$integral by Lemma \[nintegral\] and induction hypothesis. Consider the section $$\pi^*_m\omega_{m-1}\in H^0(X_m,\pi_m^*{\mathcal{O}}_{X_{m-1}}(7-(m-1))\otimes S^2(\Omega^1_{X_{m}}))=H^0(X_m,{\mathcal{O}}_{X_{m}}(7-(m-1))\otimes S^2(\Omega^1_{X_{m}}))$$ (recall that $\pi_n^*{\mathcal{O}}_{X_{n-1}}(1)={\mathcal{O}}_{X_n}(1)$). By Lemma \[vojta\] we have that $\pi^*_m\omega_{m-1}$ vanishes identically on $C_m$, thus $\pi^*_m\omega_{m-1}$ determines a global section $\omega_m$ in ${\mathcal{O}}_{X_{m}}(7-m)\otimes S^2(\Omega^1_{X_{m}})$ by taking $$\omega_m=\frac{1}{x_m}\pi^*_m\omega_{m-1}.$$ Call $U_m$ the open set of $X_m$ obtained by deleting the curves lying over any of the $C_3,\ldots,C_m$. The sections $\omega'_m$ and $\omega_m$ agree on $U_m$ up to a non-vanishing factor, therefore the $\omega'_m-$integral curves and the $\omega_m-$integral curves are the same on $U_m$. A curve lying over some $C_i$ is of type (3) in Lemma \[nintegral\] (see Equation \[CninP2\]), hence it is $\omega'_m$-integral, and we are done. \[lastcurves\] For $n\ge 6$, the only $\omega_n-$integral curves with genus $\le 2^{n-3}$ on $X_n$ are the trivial lines, with genus $0$. This follows from Lemma \[nintegral\] and Lemma \[omegan\]. \[spset\] For $n\ge 8$, the only curves of genus $0$ or $1$ on $X_n$ are the trivial lines. Let $Y\subset X_n$ be a curve of genus $0$ or $1$ and write $i:\tilde{Y}\rightarrow Y$ for its normalization. On the one hand, the curve $\tilde{Y}$ has genus $0$ or $1$, hence ${\mathcal{K}}_{\tilde{Y}}$ has non-positive degree. On the other hand, the sheaf $i^*{\mathcal{O}}_{X_{n}}(7-n)$ has negative degree because $n\ge 8$. Therefore, $i^*{\mathcal{O}}_{X_{n}}(7-n)\otimes{\mathcal{K}}_{\tilde{Y}}^{\otimes 2}$ has no nonzero global section on $\tilde{Y}$, hence $i^*\omega_n$ vanishes identically on $\tilde{Y}$. From this we deduce that $Y$ is a $\omega_n-$integral curve with genus $\le 1$ on $X_n$, and we are done by Corollary \[lastcurves\]. Proofs of results related to number fields {#SecNum} ========================================== We understand that, given a sequence $\delta_2,\delta_3,\ldots$ of distinct non-zero elements in $K/{\mathbb{Q}}$, the surfaces $X_n$ are defined by Equation . \[correspondence\] Fix a sequence $(a_1,a_2,\ldots a_n)$ in $K/{\mathbb{Q}}$, with $n\ge 3$ and pairwise distinct $a_i$. Set $\delta_i=a_i-a_1$ for $i\ge 2$. There is a bijective correspondence between the set of monic polynomials $f\in K[x]$ of degree two satisfying that $f(a_i)$ is a square for $i=1,\ldots, n$, and $X_n(K)\cap\{x_0\ne 0\}$. This correspondence is given by the map $j(f)=[1:\sqrt{f(a_1)}:\cdots :\sqrt{f(a_n)}]$ and has the property that $f$ is a square in $K[x]$ if and only if $j(f)$ lies in a trivial line of $X_n$. Take a polynomial $f=x^2+ax+b\in K[x]$ with the property that $f(a_1)=b_1^2,\ldots,f(a_n)=b_n^2$ are squares in $K$, then $$\begin{aligned} \label{verificacion} \delta_2 b_i^2&=&(a_2-a_1)f(a_i)=(a_2-a_1)(a_i^2+ua_i+v)\\ &=&(a_i-a_1)(a_2-a_1)(a_i-a_2)\cdot 1 - (a_i-a_2)(a_1^2+ua_1+v)+(a_i-a_1)(a_2^2+ua_2+v)\\ &=&\delta_i\delta_2(\delta_i-\delta_2)1^2-(\delta_i-\delta_2)b_1^2+\delta_ib_2^2\end{aligned}$$ Therefore, for each polynomial $f=x^2+ux+v\in K[x]$ with the property that $f(a_1),\ldots,f(a_n)$ are squares in $K$, we have that $j(f)\in X_n(K)\cap\{x_0\ne 0\}$.\ Conversely, given a point $p=[1:b_1:\cdots:b_n]\in X_n(K)\cap\{x_0\ne 0\}$, define $$f_p=x^2+ \frac{b_2^2-b_1^2-a_2^2+a_1^2}{a_2-a_1}x+\frac{a_1a_2(a_2-a_1)-a_1b_2^2+a_2b_1^2}{a_2-a_1}\in K[x]$$ The polynomial $f_p$ is the only monic polynomial of degree two satisfying $f_p(a_1)=b_1^2$ and $f_p(a_2)=b_2^2$. Moreover, after a standard computation we get $$\delta_2 f_p(a_1+\delta_i)= \delta_i\delta_2(\delta_i-\delta_2)-(\delta_i-\delta_2)b_1^2+\delta_ib_2^2$$ and, since $p\in X_n(K)\cap\{x_0\ne 0\}$, we obtain $\delta_2f_p(a_1+\delta_i)=\delta_2 b_i^2$. Therefore $f_p(a_i)=b_i^2$ for each $i$.\ Clearly $j$ and $p\mapsto f_p$ are inverses, hence $j$ is bijective. Assume that $j(f)=[1:b_1:\cdots:b_n]$ lies in a trivial line for some $f=x^2+ux+v\in K[x]$. Thus we have an equation of the kind $\pm b_2-\delta_2=\pm b_1$, say $\epsilon' b_2=\epsilon b_1 +a_2-a_1$. Therefore $b_2^2=b_1^2+2\epsilon(a_2-a_1)b_1+(a_2-a_1)^2$ and we get $$\begin{aligned} \left(\frac{b_2^2-b_1^2-a_2^2+a_1^2}{a_2-a_1}\right)^2-4\left(\frac{a_1a_2(a_2-a_1)-a_1b_2^2+a_2b_1^2}{a_2-a_1}\right)=4b_1^2(\epsilon^2-1)=0\end{aligned}$$ So we have $f=f_{j(f)}=\left(x+\frac{u}{2}\right)^2$. \ First we prove Theorem \[RepNum\].\ We follow the notation of Section \[geometric\]. For $i=2,\ldots,8$ set $\delta_i=a_i-a_1$ and note that $X_2,\ldots,X_8$ are defined over $K$. If Conjecture \[ConjBombieri\] holds then there exists a proper Zariski closed subset $Z\subset X_8$ such that all the $K-$rational points of $X_8$ belong to $Z$. Given an irreducible curve $Y\subset X_n$, if $Y(K)$ is dense in $Y({\mathbb{C}})$ then $Y$ is defined over $K$ and, by Faltings’ Theorem, $Y$ has genus at most $1$. Therefore we can take $Z$ as the union of a finite number of curves on $X_8$ with genus $0$ or $1$, up to a finite number of $K$-rational points. By Theorem \[spset\] and Lemma \[correspondence\] we can conclude. \ Now we prove Corollary \[buchistrong\].\ Since the set $E({\mathbb{Q}},(a_i)_i)$ is finite, it is enough to show that a monic polynomial $f\in{\mathbb{Z}}[z]$ which is not a square, satisfies that $f(n)$ is a square at most for a finite number of $n\in {\mathbb{Z}}$. Indeed, the graph of $y=\sqrt{f(x)}$ is asymptotic to the graph of $y=|x|$ hence for large enough $|x|$ it has no integer point. \ Finally, here we have the proof of Theorem \[LogicNum\].\ Let $X$ be a Büchi surface of length $n\ge 8$ such that Bombieri’s conjecture holds for $X$. Complete the sequence $\delta_2,\ldots,\delta_n$ to an infinite sequence $(\delta_i)_{i\ge 1}$ of non-zero distinct integers, and consider the corresponding surfaces $X_i$, where $X=X_n$. By an obvious modification of Corollary \[buchistrong\], Lemma \[correspondence\] allows us to show that there exists an integer $M\ge n$ such that any point $[1:b_1:\cdots:b_M]\in X_M$ with $b_i\in{\mathbb{Z}}$ must lie in a trivial line. Therefore we can write a ${\mathcal{L}}_2$-formula $\psi$ with the property that, ${\mathbb{Z}}\vDash \psi(c_1,\ldots,c_n)$ if and only if the $c_i$ are integer squares and $p=[1:\sqrt{c_1}:\cdots:\sqrt{c_n}]\in X_M$ lies in a trivial line. By Lemma \[correspondence\] there exists $\nu$ such that $c_2=\nu^2$ and $c_i=(\nu+\delta_i)^2$ for $i\ge 3$. This proves the non-trivial implication of the fact that the ${\mathcal{L}}_2$-formula $\Psi(x,y)$ $$\exists c_1,\ldots,c_n (\psi(c_1,\ldots,c_n)\wedge c_2-c_1=2\delta_2 x+\delta_2^2\wedge y=c_1$$ defines the relation $y=x^2$ in ${\mathbb{Z}}$. From here it is clear that multiplication is positive existentially definable in ${\mathbb{Z}}$ over ${\mathcal{L}}_2$. Proof of Theorem \[RepFun\] (Function Fields) {#SecFun} ============================================= We use the same notation as in Section \[geometric\]. Let $n\ge 8$. If $Y\subseteq X_n$ is a curve, its normalization is $i:\tilde{Y}\to Y$ and $g(\tilde{Y})< \frac{n-3}{4}$, then $Y$ is a $\omega_n$-integral curve. Let $i:\tilde{Y}\to Y$ be the normalization map. We have $$i^*\omega_n\in H^0(X_n,i^*{\mathcal{O}}(7-n)\otimes{\mathcal{K}}_{\tilde{Y}}^{\otimes 2}).$$ As $\deg i^*{\mathcal{O}}_{X_n}(1)\ge 1$, for $n\ge 8$ we get $$\begin{aligned} \deg \left(i^*{\mathcal{O}}_{X_n}(7-n)\otimes{\mathcal{K}}_{\tilde{Y}}^{\otimes 2}\right)=(7-n)\deg i^*{\mathcal{O}}_{X_n}(1) + 4g(\tilde{y})-4\\ \le 7-n+4g(\tilde{Y})-4=4g(\tilde{Y}) +3 -n<0.\end{aligned}$$ therefore $i^*\omega_n$ is zero in $\tilde{Y}$. \ Now we present the proof of Theorem \[RepFun\].\ We can assume $F={\mathbb{C}}$. Suppose $P$ has some non-constant coefficient and $P(a_i)=h_i^2, i=1,\ldots,M$ for some $a_i\in {\mathbb{C}}$ and $h_i\in K(C)$. Using Lemma \[correspondence\] twice, with $K=K(C)$ and $K={\mathbb{C}}$, one can check that $h=[1:h_1:\ldots:h_M]$ defines a non-constant morphism $h:C\to X_M$, where we consider $\delta_i=a_i-a_1$ in the definition of $X_M$. Since $C$ is a complete variety we obtain that ${\mbox{\upshape{im}}}(h)$ is algebraic. Let $Y$ be an irreducible curve containing ${\mbox{\upshape{im}}}(h)$, since $h$ is dominant on $Y$, we conclude that $h$ factors through $\tilde{Y}$. By Riemann-Hurwitz Formula $g(\tilde{Y})\le g(C)\le \frac{M}{4}-1<\frac{M-3}{4}$ hence $Y$ is a $\omega_M$ integral curve by the previous lemma. Therefore $Y$ is nonsingular and $g(Y)\le \frac{M}{4}-1< 2^{M-3}$ with $M\ge 8$, thus $Y$ is a trivial line by Lemma \[spset\]. This implies that ${\mbox{\upshape{im}}}(h)$ is contained in a trivial line, and the conclusion follows from Lemma \[correspondence\]. Proof of Theorem \[LogicMer\] {#SecLogicMer} ============================= We will use the positive answer of ${\mathbf{BP}}({\mathcal{M}}_p)$ and ${\mathbf{BP}}({\mathcal{A}}_p)$. Let $R$ be the ring ${\mathcal{A}}_p$ or the field ${\mathcal{M}}_p$. The following formula $$F[x,y]: \exists u_1\cdots\exists u_{35}\left(\wedge_{i=1}^{35}P_2(u_i)\right)\wedge \left(\wedge_{i=2}^{34} u_{i-1}+u_{i+1}=2u_i+2\right) \wedge x=u_1 \wedge 2y+1=u_2-u_1$$ is satisfied in $R$ if and only if $y=x^2$ or $x,y\in{\mathbb{C}}_p$. Then the ${\mathcal{L}}_2'$-formula (actually we should use $f_z$) $$G[x,y]: F[x,y]\wedge F[zx,z^2y]$$ is satisfied in $R$ if and only if $y=x^2$. Therefore, the ${\mathcal{L}}_2'$-formula $$H[x,y,w]: \exists u\exists v \left(G[x+y,u]\wedge G[x-y,v]\wedge u=v+4w\right)$$ is satisfied in $R$ if and only if $w=xy$. This proves Theorem \[LogicMer\] [99]{} W. Berkovich, *Spectral theory and analytic geometry over non-Archimedean fields*, Math. Surveys and Monographs, Coll. Amer. Math Soc. (1990). A. Bremner, *On square values of quadratics*, Acta Arith. [**108**]{}, no. 2, 95-111 (2003). W. Cherry and Z. Ye, *Non-Archimedean Nevanlinna theory in several variables and non-Archimedean Nevanlinna inverse problem*, Transactions of the American Mathematical Society [**349**]{}, 5047-5071, (1997). J. Denef, *The Diophantine Problem for polynomial rings and fields of rational functions*, Transactions of the American Mathematical Society [**242**]{}, 391-399 (1978). D. Hensley, *Sequences of squares with second difference of two and a problem of logic*, unpublished (1980-1983). —– *Sequences of squares with second difference of two and a conjecture of Büchi*, unpublished (1980-1983). L. Lipshitz, *Quadratic forms, the five square problem, and diophantine equations*, The collected works of J. Richard Büchi (S. MacLane and Dirk Siefkes, eds.) Springer, 677-680, (1990). L. Lipshitz and T. Pheidas, *An analogue of Hilbert’s tenth problem for $p$-adic entire functions*, Jour. Symb. Logic [**60**]{}, no. 4 (1995). Y. Matiyasevic, *Enumerable sets are diophantine*, Dokladii Akademii Nauk SSSR, [**191**]{} (1970), 279-282; English translation. Soviet Mathematics Doklady [**11**]{}, 354-358 (1970). T. Pheidas and X. Vidaux, *Extensions of Büchi’s problem: Questions of decidability for addition and $n$-th powers*, Fundamenta Mathmaticae [**185**]{}, 171-194 (2005). —– *The analogue of Büchi’s problem for rational functions*, Journal of The London Mathematical Society [**74-3**]{}, 545-565 (2006). —– *Corrigendum: The analogue of Büchi’s problem for rational functions*, submitted to the Journal of the London Mathematical Society (2009). —– *The analogue of Büchi’s problem for cubes in rings of polynomials*, Pacific Journal of Mathematics [**238**]{} (2), 349-366 (2008). A. M. Robert, *A course in $p$-adic analysis*, Springer, Graduate Texts in Mathematics [**198**]{}. M. Ru, *A note on $p$-adic Nevanlinna Theory*, Proceedings of the American Mathematical Society, [**129(5)**]{}, 1263-1269 (2000). A. Shlapentokh and X. Vidaux *The analogue of Büchi’s problem for function fields*, preprint. X. Vidaux, *An analogue of Hilbert’s tenth problem for fields of meromorphic functions over non-Archimedean valued fields*, Journal of Number Theory [**101**]{}, Issue 1, 48-73 (2003). P. Vojta *Diagonal quadratic forms and Hilbert’s Tenth Problem*, Contemporary Mathematics [**270**]{}, 261-274 (2000).

ArXiv

--- abstract: 'In precision medicine, machine learning techniques have been commonly proposed to aid physicians in early screening of chronic diseases such as Parkinson’s Disease. These automated screening procedures should be interpretable by a clinician who must explain the decision-making process to patients for informed consent. However, the methods which typically achieve the highest level of accuracy given early screening data are complex black box models. In this paper, we provide a novel approach for explaining black box model predictions of Parkinson’s Disease progression that can give high fidelity explanations with lower model complexity. Specifically, we use the Parkinson’s Progression Marker Initiative (PPMI) data set to cluster patients based on the trajectory of their disease progression. This can be used to predict how a patient’s symptoms are likely to develop based on initial screening data. We then develop a black box (random forest) model for predicting which cluster a patient belongs in, along with a method for generating local explainers for these predictions. Our local explainer methodology uses a computationally efficient information filter to include only the most relevant features. We also develop a global explainer methodology and empirically validate its performance on the PPMI data set, showing that our approach may Pareto-dominate existing techniques on the trade-off between fidelity and coverage. Such tools should prove useful for implementing medical screening tools in practice by providing explainer models with high fidelity and significantly less functional complexity.' author: - Qiaomei Li - Rachel Cummings - Yonatan Mintz bibliography: - 'pdinterp.bib' title: 'Locally Interpretable Predictions of Parkinson’s Disease Progression' --- Introduction ============ In precision medicine, machine learning techniques have been commonly proposed to aid physicians in early screening of chronic diseases. Many of these diseases become more difficult to treat as they progress, so accurate early screening is critical to ensure resources are directed towards the most effective treatment plan [@pagan2012improving]. Since the final treatment decision must inevitably be made by a doctor, these screening procedures should be interpretable such that a clinician can explain the decision-making process to patients for informed consent. However, the types of models that achieve the highest level of accuracy given early screening data tend to be extremely complex, meaning that even machine learning experts have difficulties explaining why certain predictions are made, leading many to describe them as “black box” [@breiman2001statistical]. In this paper, we bridge this gap by providing a novel approach for explaining black box model predictions which can give high fidelity explanations with lower model complexity. In particular we will focus on early screening of Parkinson’s Disease (PD). PD is a complicated neurodegenerative disorder that affects the central nervous system and specifically the motor control of individuals [@mjf2019]. This disorder is estimated to affect 930,000 individuals in the US by 2020, and is more prevalent in the geriatric population affecting more then 1% of the population over the age of 60 and 5% of the population over age 85 [@findley2007economic; @kowal2013current; @rossi2018projection]. These statistics and other recent studies on Parkinson’s epidemiology indicate that as the population ages, the prevalence of PD is expected to grow to over 1.2 million by 2030 in the US alone, increasing the total economic burden of the disorder to approximately \$26 billion USD [@kowal2013current; @rossi2018projection]. One of the most challenging aspects of PD research is that there is still no clear consensus on the root cause, and whether it is a single disease or a group of diseases characterized by similar symptoms known as Parkinsonism [@rao2006parkinson]. Since the disorder manifests differently between individuals (with different primary symptoms expressed across different patients) [@rao2006parkinson; @fereshtehnejad2017clinical; @fereshtehnejad2017subtypes], studying sub-categorization of PD disease progression has been of great interest in the medical community, particularly using novel advances in data-driven and statistical methods. In this paper, we will use data from the Parkinson’s Progression Marker Initiative (PPMI) [@PPMI] to develop and analyze a method for classifying patients based on their disease progression, and to provide data-driven PD sub-types. We develop a model that can be used with screening measurements to predict how a potential patient’s symptoms are likely to develop over the course of the following two years. Our resulting five sub-types correlate well with known primary PD symptoms and have clear medical interpretations. We then develop a random forest model which can accurately predict which of these sub-types a patient should be classified into, given common screening data. Since this model is a black box, we additionally develop a method for generating *local explainers* for each prediction. Our local explainer methodology uses a computationally efficient information filter to include only the most relevant features to explain a given prediction, resulting in a methodology we believe useful for implementing such screening tools in practice by providing explainer models with high fidelity and significantly less functional complexity. We then develop a global explainer methodology by aggregating local explainers. We use an Integer Programming based approach to determine which local explainers to include in our global explainer. Our global explainer must trade off between coverage, fidelity of predictions, and interpretability. We show that it is on the Pareto frontier of this trade-off space, relative to existing methods. Additionally, many other global explainers are constrained to have perfect coverage, while our method has an additional degree of freedom, which allows for improvements in fidelity and interpretability. The remainder of the paper will proceed as follows. A discussion of related literature and previous work connected to interpretable machine learning and PD diagnosis is given next in Section \[sec:lit\_rev\]. We will discuss our data driven cluster analysis for determining PD progression sub-types in Section \[sec:cluster\]. Then in Section \[sec:loc\_exp\] we will discuss our local explainer methodology and provide with numerical validation of this methodology in Section \[sec:exp\_res\]. In Section \[s.global\] we show how to extend this local explainer framework into a global explainer using an Integer Programming (IP) framework, and in Section \[s.globalexp\], we provide empirical validation and compare the performance of our IP-based approach with other local and global explainer methods. Related Work {#sec:lit_rev} ------------ Due to the prevalence and complexity of PD, there has been a significant amount of literature focused on using data-driven methods and machine learning to assist with diagnoses. Several diagnosis methods have been proposed including those that use classical ML models such as kernel SVMs [@prashanth2016high], ensemble models [@latourelle2017large; @castillo2018robust], and both supervised and unsupervised deep learning methods [@hirschauer2015computer; @adeli2016joint; @peng2017multilevel; @singh2018determination]. The classical and ensemble methods have typically focused on lab tests and genetic markers, while the deep learning methods were developed to incorporate MRI imaging into these predictions. The majority of this work focuses on binary diagnoses, labeling individuals as either healthy or having PD, but do not give information on disease progression or disease sub-types. Also, most of the proposed methods—particularly the ensemble and deep learning methods—are difficult to interpret. For example, they may identify a region of interest in an MRI image or highlight certain genetic markers, but it is difficult to explain to clinicians or patients why these regions are important for the model’s final decision. In contrast, our model is meant to predict the disease progression of individuals based on early screening data. To ensure interpretability, we introduce a local and global explainer techniques so that proper and clear rationale can be given to classifications. In addition to the work on diagnoses, there has also been significant research into the use of data-driven methods for PD sub-type identification [@graham1999data; @erro2013heterogeneity; @fereshtehnejad2015new; @fereshtehnejad2017subtypes]. The majority of analyses that fall in this stream of literature focuses on using unsupervised methods such as $k$-means clustering to create patient sub-types based on screening data, and then track the importance of the clustering based on longitudinal progression observations. In contrast, our model will first cluster patient types based on the dynamic behavior of disease progression, and then attempt to predict these clusters using screening data. We believe this approach will be useful in identifying the most effective course of treatment by directly treating the primary symptoms that develop in each progression cluster. Our paper also draws on previous work in the broader field of interpretable machine learning. The two primary types of interpretable learning include models that are interpretable by design [@aswani2019behavioral], and black box models that can be explained using global explainer [@wang2015falling; @lakkaraju2016interpretable; @ustun2016supersparse; @bastani2018interpreting] and local explainer [@ribeiro2016should; @ribeiro2018anchors] methods. Models that are interpretable by design are perhaps the gold standard for interpretable ML; however, these models often require significant domain knowledge to formulate and train, and are therefore not suited for exploratory tasks such as PD diagnosis. Global explainer methodology attempts to train an explainable model (such as a decision tree with minimal branching) in order to match the predictions of a black box model across the entirety of its feature space. While these models may provide some understanding on the general behavior of the black box model, if the relationship between features and black-box predictions is too complex, then the global explainer may remove many subtleties that are vital for validation and explanation. In contrast, local explainer methods attempt to train simpler models centered around a the prediction of a single data point. The most commonly used local explainer methods are Local Interpretable Model-Agnostic Explanations (LIME) [@ribeiro2016should] and anchors [@ribeiro2018anchors]. While local methods cannot validate the full black box model, they are useful for understanding the subtleties and justification for particular predictions. The method we propose in this paper follows from the idea of local explanations. We then aggregate these local explainers into a global explainer, trading off between coverage of the global model and fidelity of the local explainers that comprise our global model. We believe this method is most appropriate for the problem of PD diagnosis, where the relationship between different screening measures and the diagnosis is quite complex, and the model should incorporate the richness of this relationship in its predictions. Clustering Methodology and PPMI Dataset {#sec:cluster} ======================================= PD is a complex disorder, and is often expressed differently by different patients, which has motivated the need to create PD sub-types to better direct treatment. While many existing data-driven methods focus on clustering patients based on their baseline measurements [@fereshtehnejad2017subtypes], we propose clustering patients using the trajectory of how their symptoms progress. We will use data collected in the PPMI study [@PPMI], which is a long run observational clinical study designed to verify progression markers for PD. To achieve this aim, the study collected data from multiple sites and includes lab test data, imaging data, genetic data, among other potentially relevant features for tracking PD progression. The study includes measurements of all these various values for the participants across 8 years at regularly scheduled follow up appointments. The complete data set contains information on 779 patients, and included 548 patients diagnosed with PD or some other kind of Parkinsonism and 231 healthy individuals as a control group. Determination of Criterion and Cluster Analysis {#s.crit} ----------------------------------------------- Since there is significant heterogeneity in how PD symptoms are expressed, there also is no agreement on a single severity score or measurement that can be used as a surrogate for PD progression. Thus instead of considering a single score, we will model the severity of the disease as a multivariate vector, and the disease progression as the trajectory of this vector through a multidimensional space. Using the PPMI data [@PPMI] and other previous literature on PD progression [@rao2006parkinson; @martinez2017rating; @bhat2018parkinson], we considered the following measures of severity to model disease progression: - Unified Parkinson’s Disease Rating Scale (UPDRS) II & III [@martinez1994unified]: The UPDRS is a questionnaire assessment that is commonly used to track symptoms of PD by an observer. It consists of four major sections, each meant to measure a different aspect of the disease. These sections are: (I) Mentation Behavior and Mood, which includes questions related to depression and cognitive impairment; (II) Activities for Daily Living, which includes questions related to simple daily actions such as hygiene and using tools; (III) Motor Examination, which includes questions related to tremors and other physical ticks; and (IV) Complications of Therapy, which attempts to assess any adverse affects of receiving treatment. For our analysis we focused on the aggregate scores of sections II and III of the UPDRS to track physical symptoms of the disease. - Montreal Cognitive Assessment (MoCA) [@nasreddine2005montreal]: Although not exclusively used for PD, the MoCA is a commonly used assessment for determining cognitive impairment and includes sections related to attention, executive functions, visual reasoning, and language. For our analysis, we used the MoCA scores of the individual patients as surrogates for their cognitive symptoms. - Modified Schwab and England Activities of Daily Living Scale (MSES) [@siderowf2010schwab]: The MSES is a metric used to measure the difficulties that individuals face when trying to complete daily chores due to motor deficiencies. This assessment is generally administered at the same time as the UPDRS and is often appended as a section V or VI. We used this score as a measure of how much autonomy the patients experience based on their symptoms. We formed the empirical trajectory of these scores for each patient using the values measured during the patients’ participation in the PPMI study [@PPMI]. For our cluster analysis we used longitudinal measurements that were taken across the first seven visits of the study corresponding to a period of 21 months, where the first measurement formed the patient’s baseline, and the next five measurements were taken at follow up visits at regular three month intervals; the final measurements were taken after six months. We chose this timeline for our analysis because participation was high among all participants in the study during this period, so we did not have to exclude any patients, and visits were more frequent to better capture disease progression over time. After these seven measurements, follow-up visits were scheduled too infrequently to provide useful trajectory modeling information. ![Mean trajectory progression for given score by cluster. Blue corresponds to Group 0, orange corresponds to Group 1, green corresponds to Group 2, and red corresponds to Group 3. The y-axis of each plot the is numerical value of the corresponding disease severity measure.[]{data-label="fig:mean_traj"}](mean_progression_trajectory.pdf) We used these trajectories to cluster the patients together into progression sub-types. The main motivation for this approach is that if patients’ severity scores progress in a similar way, then it may identify a useful sub-type for treatment design. Only patients diagnosed with PD were included in the cluster analysis, since we are interested in finding useful sub-types of disease progression. Each trajectory was then flattened out as a 28 dimensional vector, with the first four entries corresponding the measurements at baseline, the next four for the 3 month follow up, and so on. Using scikit-learn and Python 3.7, we performed $k$-means clustering on these trajectories to define our sub-types [@pedregosa2011scikit; @friedman2001elements]. Using cross validation and the elbow method (as seen in Figure \[fig:elbow\_plotl\] in the appendix), we determined that there are four potential sub-types of disease progressions for the PPMI participants. We label these as: moderate physical symptoms cognitive decline cluster (Group 0), stagnant motor symptoms autonomy decline cluster (Group 1), motor symptom dominant cluster (Group 2), and moderate symptoms cluster (Group 3). The names we assigned to each individual cluster were given by the observed mean trajectories of the relevant scores for individuals that were classified into a particular cluster as shown in Figure \[fig:mean\_traj\]. In Figure \[fig:projections\] we show two 2-dimensional projections of the different cluster groups. Figure \[fig:pca\_cluster\] shows the projection onto the first two principal components of the data using PCA [@friedman2001elements]; this projection method is meant to preserve linear relationships among data points as well as distances between data points that are far apart. The projection shown in Figure \[fig:tsne\_cluster\] corresponds to the tSNE projection of the data onto a two-dimensional space [@maaten2008visualizing], this projection method was designed with manifolds in mind and is meant to preserve close distances (i.e., data points close in the tSNE projection should be also close in the higher dimensional space). Note that in both projections our resulting clusters are distinct and do not significantly overlap. [0.49]{} ![Two different 2-dimensional projections for visualizing trajectory clusters. Purple corresponds to Group 0, blue corresponds to Group 1, green corresponds to Group 2, and yellow corresponds to Group 3. []{data-label="fig:projections"}](alt_pca_cluster.pdf "fig:"){width="\textwidth"} [0.49]{} ![Two different 2-dimensional projections for visualizing trajectory clusters. Purple corresponds to Group 0, blue corresponds to Group 1, green corresponds to Group 2, and yellow corresponds to Group 3. []{data-label="fig:projections"}](alt_tsne_cluster.pdf "fig:"){width="\textwidth"} Validation of Clusters ---------------------- To test whether these clustered sub-types provide additional insight into the health of the patients, we performed several statistical comparisons of each patients’ characteristics at baseline across all four sub-types plus healthy patients, to determine if there were any statistically significant differences. The results and values of these comparisons are presented in Table \[table:stat\_comp\] below. Group 0 Group 1 Group 2 Group 3 Healthy p value -- -------------- ---------------- ---------------- --------------- ---------------- ----------- ----------- 1.643$^m$ 1.749 1.642$^n$ 1.704$^p$ 1.850$^{mnp}$ 0.01 5.549$^{de}$ 1.892$^{dfgh}$ 5.969$^{fij}$ 5.087$^{gik}$ 3.247$^{ehjk}$ &lt;0.001 6.594 6.482 7.981 3.272 N/A &lt;0.001 23.654 21.277 26.503 15.382 N/A &lt;0.001 92.256 91.506 91.321 96.214 N/A &lt;0.001 58.925$^a$ 60.446 62.912$^{abc}$ 58.387$^b$ 59.571$^c$ 0.02 Anosmia 46 10 57 41 6 &lt;0.001 Hyposmia 68 11 91 98 68 Normosmia 19 5 11 34 122 95.49% 93.98% 94.34% 94.22% 94.37% 0.99 67.67% 57.83% 65.41% 63.01% 65.80% 0.63 5.391 5.069 5.270 5.231 5.168 0.68 : Comparison of baseline and screening measurements between clusters. p-values labeled in the table represent difference between all groups, and significant pairwise comparisons using a two sample T-test are marked by superscripts with p-values a-0.008; b-0.001; c-0.02;d,e,f,g,h,i,j,k-$<$0.001, m-0.003;n-0.004;p-0.04 []{data-label="table:stat_comp"} As seen in Table \[table:stat\_comp\], many of the key screening measurements of the populations from the different clusters are significantly different, implying our clusters are informative about the health of individuals. In particular, we note that Group 0—which corresponds to moderate physical symptoms with cognitive decline—tends to be younger on average then the other groups, indicating this group may contain many more individuals with early onset PD. Moreover, the sub-types vary substantially in their sleep score and olfactory evaluation, which are both measures that have previously been shown to be strong indicators of PD [@rao2006parkinson] indicating that these progression sub-types are sensitive to these important predictors. Overall, the comparisons shown in Table \[table:stat\_comp\] show that our data driven clusters are not only informative when comparing different forms of disease progression, but also correspond to variations in screening measurements. Based on this analysis, we believe that using screening data to predict these clusters could lead to clinically significant insights that can help with treatment. Local Explainer Algorithm {#sec:loc_exp} ========================= After identifying the four disease progression sub-types, we would like to predict which kind of disease progression an individual might experience, given measurements collected during a screening visit. As we will show in our experiments in Section \[sec:exp\_res\], this task is best performed by complex black box models such as artificial neural networks (ANN) and bagged forests. This means that while the prediction may be accurate, it will not be easily explained, which make such models difficult to use for diagnosis recommendations. Our goal is to instead develop a method that trains simple auxiliary explainer models, and can still accurately describe the relationship between the data and the model output within a small region of a given prediction. This methodology is known as *training local explainer models* and has been shown to be useful in understanding black box predictions [@ribeiro2016should; @ribeiro2018anchors]. One of the key tradeoffs in generating model explanations is that of *fidelity*—how well the explainer approximates the black box model—and *interpretability*—how easy it is for a practitioner to trace the predictions of the model. In contrast to previous literature which has proposed the use of regularization to achieve this goal, we propose directly computing locally significant features using an information filter. Generally, computing such filters can be computationally expensive and requires the use of numerical integration; however, one of our main contributions in this paper is to introduce an efficient algorithm for filtering out less significant features. This methodology will allow us to train local explainers that are significantly less complex then those that use regularization, with better fidelity. Notation {#sec:notation} -------- Before proceeding to our discussion on the local explainer method, we will first establish some technical notation. We assume that for each patient $i = 1,...,n$ we have an ordered pair $(x_i,y_i)$, where $x_i \in \mathcal{X} \subseteq \mathbb{R}^m$ are the features values of the patient and $y_i \in \mathcal{L} \subseteq \mathbb{Z}$ is the corresponding class label generated by a black box model $f$. Through our analysis we will also refer to this set of points through matrix notation where $X \in \mathcal{X}^n \subseteq \mathbb{R}^{m\times n}$ is the feature value matrix and $y \in \mathcal{L}^n \subseteq \mathbb{Z}^n$ is the vector of class labels, where each row in these matrices corresponds to a single patient’s data. For our analysis we assume that $\mathcal{X}$ is a compact set. Let $\Phi= \{1,...,m\}$ be the set of features, and it may also be used to denote the index set of the features. This set can be partitioned into two sets $\Phi_c,\Phi_b \subseteq \Phi$ that represent the set of continuous and binary features respectively. Furthermore we define the set-valued function $\Phi^*: \mathcal{X} \rightarrow \Phi$ as the function which extracts the minimum set of necessary features to accurately predict the class of a point $x$. Namely, $$\Phi^*(x) = \operatorname*{arg\,min}_{\varphi \subseteq \Phi}\{ |\varphi| : p(y|x) = p(y|x[\varphi])\},$$ where $x[\varphi]$ is an indexing operation that maintains the values of $x$ but only for the features in $\varphi$, and $p$ is the conditional probability mass function of the labels $y$ given the observation of some features. Specifically, if a feature index is not included $\Phi^*(x)$, then it is not required to understand the particular label of $x$. In addition, we will denote the ball around a point $x$ of radius $r$ with respect to a metric $d$ as $\mathcal{B}(x,r,d)$. Finally, a key feature of the explainer training method we propose includes the use of *mutual information*. In information theory, mutual information is a quantity that measures how correlated two random variables are with one another. If $X,Y$ are two random variables with joint density $p$ and marginal densities $p_x,p_y$, then the mutual information between $X$ and $Y$ is denoted $I(X;Y)$ and calculated as: $$I(x;y) = \mathbb{E}\log\frac{p(X,Y)}{p_x(X),p_y(Y)} = \int_{x}\int_{y} p(x,y) \log\frac{p(x,y)}{p_x(x),p_y(y)}dx dy.$$ If $X$ and $Y$ are independent then $I(X;Y) = 0$; otherwise $I(X;Y) > 0$, meaning that $X$ contains some information about $Y$. A similar quantity can be computed using a conditional distribution on another random variable $Z$, known as the *conditional mutual information* and denoted $I(X;Y|Z)$. Local Explainer Algorithm Description ------------------------------------- Our main local explainer algorithm extends previous local explainer methods such as LIME [@ribeiro2016should] by restricting the sampling region around the prediction, and including an information filter to ensure that fewer features are included in the final explainer mode. Our general local explainer is formally presented in Algorithm \[alg:loc\_exp\], but we will give a brief overview of its operations here. The algorithm takes in hyper-parameters including number of points $N$ to be sampled for training the explainer, a distance metric $d$, and a radius $r$ around the point $\bar{x}$ being explained. First the algorithm samples $N$ points uniformly from within a $r$ radius of $\bar{x}$; we call this set of points $T(\bar{x})$. Depending on the distance metric being used this can often be done quite efficiently, especially if the features are binary valued or an $\ell^p$ metric is used [@barthe2005probabilistic]. Then using the sampled points, the algorithm uses the Fast Forward Feature Selection (FFFS) algorithm as a subroutine (formally presented in Section \[sec.fffs\] and Appendix \[app.fffs\]), which uses an information filter to remove unnecessary features and reduce the complexity of the explainer model. The FFFS algorithm uses an estimate of the joint empirical distribution of $(T(\bar{x}),f(T(\bar{x}))$ to select the most important features for explaining the model’s predictions in the given neighborhood. We denote this set of features $\hat{\Phi}$. Then, using these features and the selected points, the explainer model $g$ is trained by minimizing an appropriate loss function that attempts to match its predictions to those of the black box model. In principle a regularization term can be added to the training loss of explainer $g$. However, through our empirical experiments in Section \[sec:exp\_res\] we found that FFFS typically selected at most five features, so even the unregularized models where not overly complex. sampling radius $r$, number of sample points $N$, black box model $f$, data point to be explained $\bar{x}$, and loss function $L$ for the explainer model $(\bar{x},\bar{y})$ Initialize $T(\bar{x}) = \emptyset$ Sample $x \sim U(\mathcal{B}(\bar{x},r,d))$ $T(\bar{x}) \leftarrow T(\bar{x}) \cup x$ Obtain $\hat{\Phi}(\bar{x}) = \text{FFFS}(T(\bar{x}), \Phi, f)$ Train $g = \operatorname*{arg\,min}_{\hat{g} \in \mathcal{G}}\{\sum_{x\in T(\bar{x})}L(f(x) - \hat{g}(x[\hat{\Phi}])) \}$ g Fast Forward Selection Information Filter {#sec.fffs} ----------------------------------------- A key step in our algorithm is the use of a mutual information filter to reduce the number of features that will be included in the training of the local explainer. Mutual information filters are commonly used in various signal processing and machine learning applications to assist in feature selection [@brown2012conditional]. However, these filters can be quite challenging to compute depending on the structure of the joint density function of the features and labels, and can require the use of (computationally expensive) numerical integration. We counteract this by considering an approximation of the density function, using histograms to calculate continuous features. When multiple combinations of features need to be considered as in our setting, the problem of finding the maximum-information minimum-sized feature set is known to be computationally infeasible [@brown2012conditional]. As such, our proposed method for computing the filter includes a common heuristic known as *forward selection*, which essentially chooses the next best feature to be included in the selected feature pool in a greedy manner. Using this method alone would still require recomputing the conditional distribution of the data based on previously selected features, which can result in long run times for large $N$. However, using some prepossessing techniques, we can show that these quantities can be stored efficiently using a tree structure, which allows quick computation of the filter. The general idea of the FFFS algorithm is to consider the feature selection process as a tree construction. Part of this construction relies on an estimate of the empirical density of the features as a histogram with at most $B$ bins and preprocessed summary tensor $M \in \{0,1\}^{B\times |\Phi| \times N}$ which indicates which bin of the histogram a feature value for a particular data point lays in. For each entry, $M[b,\varphi,x] = 1$ if the value of feature $\varphi$ at point $x$ falls in the bin $b$. Otherwise, $M[b,\varphi,x] = 0$. The depth of the tree represents the number of selected features and each node of the tree is a subset of $T(\bar{x})$. For instance, at the beginning of the selection process, we have a tree with exactly one node $R$ where $R=T(\bar{x})$. Assume binary feature $\varphi_1$ is selected in the first round. Then two nodes $a,b$ are added under $R$, where $a = \{x_j: M[1,\varphi_1,j]=1\}$ and $b = \{x_j: M(2,\varphi_1,j)=1\}$. In the second round, we use the partition sets $a,b$ to compute the mutual information instead of the complete set $R$. The set $a$ is used for computing $\hat{p}(\varphi|\varphi_1=1),\; \hat{p}(y|\varphi_1=1) \text{, and }\hat{p}(\varphi;y|\varphi_1=1)$, while $b$ is used when the condition is $\varphi_1=2$. In each round the leaves $\mathcal{L}$ of the current tree represent the set of partition sets corresponding to all random permutation of selected features information. Therefore, $\mathcal{L}$ provides us sufficient information for calculating the desired mutual information. As shown in Algorithm \[alg:sf\], the algorithm only outputs the leaves $\mathcal{L}$, not the entire tree. The main algorithmic challenge is to efficiently calculate the marginal distributions $(\hat{p}(\varphi|S), \hat{p}(y|S)$ and joint distribution $\hat{p}(\varphi;y|S)$, which we are able to do using the tree structure. The detailed structure of the FFFS algorithm used to compute the filtered feature set $\hat{\Phi}$ requires several subroutines, and the formal algorithmic construction for computing the filter is presented across Algorithms \[alg:fffs\], \[alg:Recur\], \[alg:sf\], and \[alg:bin\]. The main FFFS algorithm is Algorithm \[alg:fffs\], and it calls the subroutines for recursion (Algorithm \[alg:Recur\]), selecting features (Algorithm \[alg:sf\]), and partitions (Algorithms \[alg:bin\]). Formal presentation of these algorithms, as well as detailed descriptions, are given in Appendix \[app.fffs\]. Experimental Validation of Local Explainer {#sec:exp_res} ========================================== In this section we empirically evaluate the quality of our local explainer methodology by first showing that accurate sub-type predictions of our PD sub-type clusters (as described in Section \[sec:cluster\]) can be achieved using black-box methods applied to the data of individuals measured during the screening visit. We then apply our local explainer methodology developed in Section \[sec:loc\_exp\] to explain the predictions given by these black-box models. Our clusters were derived from longitudinal measurements of the four metrics of disease severity described in Section \[s.crit\], measured across the first seven visits in the study over a period of 21 months. Treating these cluster (and the healthy patients) as our ground truth class labels, we first train black box machine learning models to predict which of these progression sub-types an individual will most likely experience given her screening data. This is meant to model the data available to a physician when she must make treatment decisions for a new patient. From screening data in the PPMI data set, we included the following 31 features: PTT, Lymphocytes, Hematocrit, Eyes, Psychiatric, Head-Neck-Lymphatic, Musculoskeletal, Sleep Score, Education Years, Geriatric Depression Score, Left Handed, Right Handed, Gender Male, Female Childbearing, Race White, Race Hispanic, Race American Indian, Race Asian, Race Black, Race PI, Anosmia, Hyponosmia, Normosmia, MRI Normal, MRI Abnormal Insignificant, MRI Abnormal Significant, BL/SC UPDRSII, BL/SC UPDRSIII, BL/SC MOCA, BL/SC MSES, and BL/SC Age. Among these 31 features, 20 features are binary variables and 11 features are continuous variables. For accurate sub-type predictions using this data, in Section \[s.mlcluster\] we trained three machine learning prediction models: one interpretable model (logistic regression) and two complex black box models (a feed forward ANN and a bagged forest). Our results indicate that the black box models outperform the simpler model, which necessitates the use of a local explainer method for this application to achieve both accurate classification and explainability. In Section \[s.localval\] we computed local explanations based on the random forest model predictions (which was the model with the highest accuracy) using our proposed FFFS method with the information filter and a local explainer method. This is analogous to LIME [@ribeiro2016should] which does not contain an information filter. Our results show that given a requirement of high explainer fidelity, the use of the information filter will result in less complex explainer models. All experiments described in this section were run on a laptop computer with a 1.2GHz Intel Core m3-7Y32 processor and MATLAB version R2019a with the machine learning and deep learning tool kits [@MATLAB:2010]. Machine Learning Models for Cluster Prediction {#s.mlcluster} ---------------------------------------------- We considered three different kinds of machine learning models for the task of predicting the progression cluster: logistic regression, feed forward ANN, and a bagged forest model. The patient data was split into training, validation, and testing sets with $70\%$ of the data used for training, $15\%$ for validation, and $15\%$ for testing. Among 779 patients, 545 patients were selected for training, and 117 patients were selected for validation and testing. Since bagged forests and ANNs are sensitive to hyperparamter settings, we used cross-validation to set their respective hyperparamters. Using cross validation and MATLAB’s hyperparemeter optimization methods we found that the most effective ANN architecture for our task was with a single hidden layer containing one hundred hidden ReLu units. For the random forest model, we found that an ensemble of 50 bagged trees gave the best results compared to other forest sizes. [0.33]{} ![Confusion Matrices[]{data-label="fig:cm"}](logistics_cm.jpg "fig:"){width="\textwidth"} [0.33]{} ![Confusion Matrices[]{data-label="fig:cm"}](neural_cm.jpg "fig:"){width="\textwidth"} [0.33]{} ![Confusion Matrices[]{data-label="fig:cm"}](random_cm.jpg "fig:"){width="\textwidth"} [0.33]{} ![ROC Curves[]{data-label="fig:roc"}](logistics_roc.jpg "fig:"){width="\textwidth"} [0.33]{} ![ROC Curves[]{data-label="fig:roc"}](neural_roc.jpg "fig:"){width="\textwidth"} [0.33]{} ![ROC Curves[]{data-label="fig:roc"}](random_roc.jpg "fig:"){width="\textwidth"} Figures \[fig:cm\] and \[fig:roc\] show the performance of the models on the same training, validation, and testing sets. In both figures, the classes 1-4 correspond to Groups 0-3, and class 5 corresponds to healthy patients (which we will also call Group 4). Figure \[fig:cm\] contains the confusion matrix for each model. The rows of the matrix are the *output class*, which represents the predicted class, and the columns of the matrix are the *target class*, which is the true class. The cells on the diagonal of the matrix count accurate predictions. Each cell in the rightmost column has two values: the top number is the percentage of patients that are correctly predicted to each class, and the bottom number is the percentage of patients that are incorrectly predicted to each class. For each cell on the bottom row, the top number is the percentage of patients that belong to each class and is correctly predicted, and the bottom number is the percentage of patients that are incorrectly predicted. For the rest of cells in the matrix, the number in each cell counts for the number of patients that fall in this observation. The cell at the bottom right corner of each matrix shows the total percentage of patients that were correctly and incorrectly predicted. As shown in Figures \[fig:cm\] and \[fig:roc\], the logistic regression model under-performs relative to the ANN and bagged forest models. Even though the bagged forest model has a lower prediction rate for Group 0 compared to the ANN, it has equal or higher rates of accurate prediction for the other classes. Additionally, the bagged forest model consistently performed better than the ANN and logistic regression models in our experiments. We concluded from these results that the bagged forest classification model is the most effective for our prediction task, and we chose to consider its predictions when evaluating our local explainer method. Local Explainer Validation {#s.localval} -------------------------- Since the main difference between our local explainer training algorithm and those in the literature is our use of the FFFS information filter, our experiments on the local explainer are focused on validating the effectiveness of using this information filter. We compare the performance of our local explainer training algorithm to a similar algorithm without a filtering step. We then compare the performance of these methods in terms of explainer complexity and fidelity, across different sampling radii and across all patients. For the sampling parameters of our algorithm, we sampled $N=10,000$ points centered around each patient within a radius $r$ of either 3, 7, 11, or 15. The distance metric for computing this radius was a combination of the $\ell_\infty$ norm for the continuous features and the $\ell_1$ norm for the binary features. The continuous value feature of each of the points was sampled uniformly using standard techniques ([@barthe2005probabilistic]). For binary valued features, we randomly chose at most $r$ binary features and flipped their values. We first randomly generated an integer $k$ between $0$ and $r$, and randomly selected $k$ binary features which we then flipped from their current value (that is, values of 1 were set to 0 and vice versa). To compute probability density estimates, we found that the method performed well with histograms with only three bins for continuous features and two bins for binary features. Intuitively three bins allows us to categorize feature values as low, medium, or high relative to their range. For both training methods, we chose to train decision trees as our the local explainer class because these have been shown to be ergonomically suitable for explaining black box models in healthcare contexts [@bastani2018interpreting]. Then we computed the corresponding *fidelity score*, defined as the percentage of data where the prediction of the decision tree matched the prediction of the random forest model. We used the number of leaves on the decision tree as a measure of the explainer complexity. ![Comparison of local explainer algorithm with the information filter (solid line) and without the the information filter (dashed line) for various different radius settings for the algorithms. The x-axis corresponds to the given fidelity score of the model and the y-axis measures the complexity of the decision tree explainer by the number of leaves. For a small radius ($r=3$) and large radius ($r=15$), the addition of an information filter does not lead to a significant difference in model complexity across all levels of fidelity. However, using the information filter in explainer training for moderate sized radii ($r=7$ and $r=11$) results in less complex models at higher levels of fidelity ($>0.6$).[]{data-label="fig:filter_comparison"}](filter_v_no_filter_leaf_count.pdf) In Figure \[fig:filter\_comparison\], we compare the explainer complexity and fidelity level of the explainers generated by the two different training methodology across the four different tested sampling radii. Unsurprisingly, when the sampling radius is small (i.e., $r=3$), there is not much advantage to using the information filter in terms of reducing model complexity for a given fidelity level. Since all points are sampled so closely together, the relevant features are easily learned in explainer training. Conversely, when the sampling radius is large ($r=15$), the addition of the information filter only helps slightly. With such a large radius, sampling feature values that are far from the point that is meant to be explained may not give useful information for that prediction. However, when considering the medium radius ranges, for high levels of fidelity, the inclusion of the information filter provides simpler models across the board. In particular, consider the plots corresponding for local explainer radius of $r=7$ and $r=11$ in Figure \[fig:filter\_comparison\]. Note that in both of these figures, when considering high fidelity explainers generated by both methods (fidelity $\geq$ 0.6), the explainers generated by the information filter method are less complex then those generated without the filter. This would indicate that using our information filter, we can obtain high fidelity local explainers that are on average less complex then those generated without this filter. When considering low fidelity explainers, the no filter method creates less complex models then the filter method. This is because our filter method is better equipped to find relevant features even in more complex regions of the black box model, while the no filter method is unable to learn these regions effectively with a fixed sample size. This is significant since this would indicate that our proposed methodology is able to explain a larger portion of the feature space using less complex models while still finding meaningful features for explanations, relative to existing methodologies. Overall, the plots in Figure \[fig:filter\_comparison\] show that incorporating an information filter into local explainer training can be more effective in extracting relevant features then using regularization, and can generate less complex models with high fidelity. In addition, these results indicate that using an information filter allows for local explainers with information filters to obtain higher fidelity over a larger radius with relatively less complex models. This is particularly significant since less complex models can be me more easily interpreted by domain experts, making it easier for them to translate the clinical significance of the black box model outputs. while larger explanation radii are useful for model validation and generalization of explanations. Moreover, even in complex decision regions generated by the black box model, using an information filter in conjunction with local explainers is better at extracting relevant features for predictions which again can be useful for model validation and providing clinical insights. Global Explainer Methodology {#s.global} ============================ While our proposed local explainer method is useful for providing insight into the behaviour of the black box model in a restricted region of the feature space it cannot be used for total model validation. For this task we require a global explainer that could provide insights into the behavior of the model across the entirety of the feature space. However, an explainer that is constrained to explain all of the feature space is likely have low fidelity since the explainer model is less complex then the black box. This introduces a trade-off between two qualities of an explainer model: its *coverage* and *fidelity*. One way to address this challenge is to create a global explainer model by aggregating several local explainer models. Several existing approaches use this idea [@ribeiro2016should] by formulating the construction of the global explainer as an optimization problem. Generally, this problem is framed as maximizing the total coverage of the explainer subject to a constraint on the total number of local explainers included in the aggregate. Solving this problem is conjectured to computationally intractable [@ribeiro2016should], and therefore in existing work it is solved using heuristics. In this section, we formulate the problem of constructing the aggregate explainer explicitly as an integer linear program that allows us to directly trade off coverage and fidelity. Mathematical Programming Formulation of Aggregation Problem {#sec:math_prog_form} ----------------------------------------------------------- In order to formulate the optimization problem for the global explainer, we first need to formally define the concepts of coverage and fidelity. Building upon the notation from Section \[sec:notation\], let $g_{i,r}:\mathcal{X} \rightarrow \mathcal{L}$ be the local explainer generated by using Algorithm \[alg:loc\_exp\] on patient $i$ with radius $r$. Furthermore let $\mathcal{X}_{i,r} \subset \mathcal{X}$ be defined as the set of points explained by explainer $g_{i,r}$; that is $\mathcal{X}_{i,r}:=\{x \in \mathcal{X}: \|x - x_i\| \leq r \}$. Define $\gamma$ as the aggregate set of all explainers generated in this process: $\gamma = \{g_{1,r_1},g_{2,r_2},...,g_{n,r_n}\}$ for some potential local explainers $g_{1,r_1},g_{2,r_2},...,g_{n,r_n}$. Using these quantities we define the *coverage of aggregate exmplainer $\gamma$* on the data set $\mathcal{X}$ as the total number of points that are covered by the explanation radius of at least one explainer contained in $\gamma$. We denote the coverage as: $$\label{eq.cov} \text{Cov}(\gamma,\mathcal{X}) = \sum_{x \in \mathcal{X}} \max_{i\in \{i: g_{i,r} \in \gamma \}} \mathbbm{1}[x \in \mathcal{X}_{i,r}].$$ Next recall that the fidelity of a single local explainer can be defined as the accuracy of that explainer using the predicted labels of the black box model as the ground truth. We will define the *fidelity of aggregate explainer $\gamma$* on the data set $\mathcal{X}$ as the minimum of the fidelity obtained by each individual local explainer. We denote this as: $$\label{eq.fid} \text{Fidelity}(\gamma,\mathcal{X}) = \min_{\{i: g_{i,r_i} \in \gamma\}} \frac{1}{|\mathcal{X}_{i,r}| }\sum \mathbbm{1}[g_{i,r}(x) = f(x)]\mathbbm{1}[x \in \mathcal{X}_{i,r}].$$ Although the fidelity of $\gamma$ can be defined as the average of the fidelities of its component explainers, we believe using the minimum fidelity gives a stricter measure on how well the global explainer captures the behavior of the black box model. Let $K$ denote the budget of the maximum number of explainers that can be contained in $\gamma$, and let $\varphi$ be the minimum fidelity we would like the aggregate explainer to obtain. Then the optimization problem to be solved can be written as: $$\label{eq:global_problem} \begin{aligned} \max_{\gamma}&\ \{\text{Cov}(\gamma,\mathcal{X}) : \text{Fidelity}(\gamma,\mathcal{X}) \geq \varphi, |\gamma| \leq K\}. \end{aligned}$$ Reformulation as Integer Program (IP) {#sec:reform} ------------------------------------- Note that as written, optimization problem is not trivial to solve, and in particular could require enumerating all possible subsets $\gamma$ of local explainers. To address this challenge, we propose reformulating problem as an Integer Program (IP) that can be solved using current commercial software. To do this, we first define the three different sets of binary variables that we will call $w_i,y_j,z_{ij}$. Let $w_i$ be a binary variable that is equal to 1 if explainer $g_{i,r_i} \in \gamma$. That is, $w_i = \mathbbm{1}[g_{i,r_i} \in \gamma]$. Let $y_j$ be defined as a binary variable that is equal to 1 if point $j$ is covered by the global explainer $\gamma$. That is $y_j = \mathbbm{1}[x_j \in \cup_{\{i: g_{i,r_i} \in \gamma\}} \mathcal{X}_{i,r_i}]$. Finally, let $z_{ij}$ be a binary variable that is equal to 1 if explainer $g_{i,r_i} \in \gamma$ covers point $x_j$. That is, $z_{ij} = \mathbbm{1}[x_j \in \mathcal{X}_{i,r_i}]$. We can now define the coverage and fidelity of aggregate explainer $\gamma$ in terms of these three sets of variables. \[prop:cov\] $\text{Cov}(\gamma,\mathcal{X})$, the coverage of local explainer set $\gamma$ on data set $\mathcal{X}$, can be expressed with the following set of integer variables and constraints: $$\begin{aligned} \text{Cov}(\gamma,\mathcal{X}) &=\sum_{j\in \mathcal{X}} y_{j}, \\ \text{s.t.} \quad &z_{ij} \leq w_i, \quad i,j \in \mathcal{X},\\ & y_j \geq z_{ij}, \quad i,j \in \mathcal{X},\\ & y_j \leq \sum_{i\in \mathcal{X}} z_{ij}, \quad j\in \mathcal{X},\\ &\|x_i - x_j\|z_{ij} \leq r_i, \quad i,j \in \mathcal{X}. \end{aligned}$$ Recall the definition of $\text{Cov}(\gamma,\mathcal{X})$ as given in Equation . We will directly reconstruct this definition using the binary variables defined above. First note that through a simple direct substitution we obtain $\text{Cov}(\gamma,\mathcal{X}) = \sum_{x \in \mathcal{X}} \max_{i\in \{i: g_{i,r} \in \gamma \}} z_{ij}$. Since taking the maximum of binary variables is equivalent to the Boolean OR operator, we see that $y_j = \max_{i\in \{i: g_{i,r} \in \gamma \}} z_{ij}$, which provides us with the first equality. The next two inequalities directly models that explainer $g_{i,r_i}$ can only explain point $x_j$ if it is included in $\gamma$, which is a standard way of modeling conditional logic in IP [@wolsey1999integer]. The next two constraints come from modeling the Boolean OR operator using integer constraints [@wolsey1999integer]. The final constraint ensures that a point $x_j$ can be covered by an explainer $g_{i,r_i}$ if $x_j \in \mathcal{X}_{i,r_i}$, thus preserving the local region of the local explainer. Next we consider the average fidelity constraint. \[prop:fide\] The constraint $\text{Fidelity}(\gamma,\mathcal{X}) \geq \varphi$ can be modeled using the following set of integer linear constraints: $$\begin{aligned} &\|x_i - x_j\|z_{ij} \leq r_i, \quad i,j \in \mathcal{X},\\ &z_{ij} \leq w_i, \quad i,j \in \mathcal{X},\\ &\sum_{j \in \mathcal{X}}\big( \mathbbm{1}_{\{f(x_j)=g_{i,r_i}(x_j)\}} - \varphi \big)z_{ij} \geq 0, &i\in \mathcal{X}. \end{aligned}$$ The first two constraints ensure proper local behavior of the local explainer as in Proposition \[prop:cov\]. Thus we will focus the derivation of the final constraint. Using the definition of $\text{Fidelity}(\gamma,\mathcal{X})$ in Equation and directly substituting variables for indicators, we can express the lower bound constraint as, $$\min_{\{i: g_{i,r_i} \in \gamma\}} \frac{1}{|\mathcal{X}_{i,r}| }\sum_{j \in \mathcal{X}} \mathbbm{1}[g_{i,r_i}(x_j) = f(x_j)]z_{ij} \geq \varphi.$$ Note that if the minimum over all explainers $g_{i,r_i}$ must have fidelity of at least $\varphi$, then every local explainer must also have fidelity at least $\varphi$. This allows us to disaggregate this constraint across all $i \in \mathcal{X}$. Let us consider the constraint for a single local explainer $g_{i,r_i}$. Using the definition of $z_{ij}$, we note that $|\mathcal{X}_{i,r_i}| = \sum_{j \in \mathcal{X}} z_{ij} $. Thus the new lower bound fidelity constraint for a single explainer can be written as: $$\frac{\sum_{j \in \mathcal{X}} \mathbbm{1}[g_{i,r_i}(x_j) = f(x_j)]z_{ij}}{\sum_{j \in \mathcal{X}} z_{ij} } \geq \varphi.$$ Note that the denominator of the left hand side can only be zero when the numerator is also zero because $\sum_{j \in \mathcal{X}} z_{ij} \geq \sum_{j \in \mathcal{X}} \mathbbm{1}[g_{i,r_i}(x_j) = f(x_j)]z_{ij}$. This means that we can multiply both sides of the inequality by the sum $\sum_{j \in \mathcal{X}} z_{ij}$ while still maintining its validity. Distributing $\varphi$ and combining like terms gives us with the form of the constraint presented in the proposition statement. We can then use these expressions to for coverage and fidelity to re-write our optimization problem as an integer program that can then be solved using commercial solvers. \[prop:reform\] The optimization problem in can be written as the following integer program: $$\begin{aligned} \max &\sum_{j\in \mathcal{X}} y_{j}, \\ \text{s.t.} \quad &z_{ij} \leq w_i, \quad i,j \in \mathcal{X},\\ & y_j \ge z_{ij}, \quad i,j \in \mathcal{X}, \\ & y_j \leq \sum_{i\in \mathcal{X}} z_{ij}, \quad j\in \mathcal{X},\\ &\|x_i - x_j\|z_{ij} \leq r_i, \quad i,j \in \mathcal{X}, \\ &\sum_{j \in \mathcal{X}}\big( \mathbbm{1}_{\{f(x_j)=g_{i,r_i}(x_j)\}} - \varphi \big)z_{ij} \geq 0, \quad i\in \mathcal{X}, \\ & \sum_{i \in \mathcal{X}} w_i \leq K, \\ & y_j,w_i,z_{ij} \in \{0,1\} \quad i,j \in \mathcal{X}. \end{aligned}$$ The objective function and first five constraints come directly from Propositions \[prop:cov\] and \[prop:fide\]. The next constraint comes using the definition of $w_i$ and direct substitution to obtain that $|\gamma| = \sum_{i \in \mathcal{X}} w_i$, which is then used to rewrite the budget constraint from . The final constraint ensures that our new variables are binary integers. Experimental Validation of Global Explainer {#s.globalexp} =========================================== In this section we empirically validate the quality of our global explainer methodology using the PPMI dataset described in Sections \[sec:cluster\] and \[sec:exp\_res\]. In our experiments, we first use the clustering algorithm of Section \[sec:cluster\] to assign labels and then apply our local explainer as described in Section \[sec:loc\_exp\]. We then use the IP-based global explainer described in Section \[s.global\] to choose the local explainers that will comprise our global explainer. [0.45]{} ![Performance of IP and other local and global explainer methods on PPMI dataset.[]{data-label="fig.global"}](2Class_Cov.png "fig:"){height="6cm"} [0.45]{} ![Performance of IP and other local and global explainer methods on PPMI dataset.[]{data-label="fig.global"}](2Class_FS.png "fig:"){height="6cm"} \ [0.45]{} ![Performance of IP and other local and global explainer methods on PPMI dataset.[]{data-label="fig.global"}](5Class_Cov.png "fig:"){height="6cm"} [0.45]{} ![Performance of IP and other local and global explainer methods on PPMI dataset.[]{data-label="fig.global"}](5Class_FS.png "fig:"){height="6cm"} We ran the optimization algorithm for binary classification (2 class) between healthy individuals and patients diagnosed with PD, and for multi-class (5 class) classification among healthy individuals and the four PD sub-types. Figure \[fig.global\] shows the coverage and average fidelity score of the global explainer generated by our IP-based approach, compared with other existing local and global explainers methods as a function of the budgeted number of local explainers that are allowed in the global explainer. We tested performance for budgets $K=5,10,15,20,30,40,50,60,70$. The lines labeled lb=0.5,0.7 or 0.9 in Figure \[fig.global\] represent the result of the IP with the lower bound of the fidelity score set to $50\%,70\% \text{ or }90 \%$, respectively. Based on testing different sampling radii (shown above in Figure \[fig:filter\_comparison\]), we found that when the sampling radius was in a medium range (i.e., $r=7$ or $11$) our method produced a simpler model with same precision than those produced by existing methodologies. Therefore we used sampling radius $r=11$ in our experiments for all local explainers. We compared our IP with the following prior methods: the submodular pick algorithm from [@ribeiro2016should], a simple decision tree trained on the labels of the black box model [@friedman2001elements], an extracted decision tree method method [@bastani2018interpreting], an interpretable decision set [@lakkaraju2016interpretable], and an anchor points method [@ribeiro2018anchors]. The interpretable decision set, decision tree, and extracted approaches are all global explainer methods, so they are always guaranteed have perfect coverage. The submodular pick and anchor point methods, along with our IP method must trade off between coverage and fidelity. For the simpler classification problem with only two classes, our IP approach, extracted decision tree, and regular decision tree methods all produced a model with significantly higher fidelity score and higher coverage than other methods. Since a large percentage of the patients are healthy individuals, it is reasonable that the decision tree and extracted tree methods achieve higher precision than our IP. However, our IP outperforms the other existing methodologies when the classification problem is more complex (5 class). Even though the coverage of IP is lower than the other global explainer methods, the model produced by IP has much higher coverage and precision than the existing local explainers methods in both problems. [0.45]{} ![Trade-off between fidelity and coverage for IP and other local and global explainer methods on PPMI dataset. For our IP-based approach, the lower bound on fidelity is varied from $0\%$ to $90\%$, while the other methods do not exhibit this degree of freedom. []{data-label="fig.pareto"}](2Class_Interp.png "fig:"){height="6cm"} [0.45]{} ![Trade-off between fidelity and coverage for IP and other local and global explainer methods on PPMI dataset. For our IP-based approach, the lower bound on fidelity is varied from $0\%$ to $90\%$, while the other methods do not exhibit this degree of freedom. []{data-label="fig.pareto"}](5Class_Interp.png "fig:"){height="6cm"} Figure \[fig.pareto\] shows the relationship between the coverage and fidelity of each method. The lines labeled MIP 5, MIP 10, and MIP 15 respectively show the smooth transition of coverage and fidelity score of our IP approach with fixed budgets of 5, 10, and 15 as the lower bound of the fidelity score varies from $0\%$ to $90\%$ in increments of $10\%$. Since our IP approach is the only one that has a lower bound constraint on fidelity, we are able to observe this trade-off between the coverage and fidelity for MIP, while the other existing methods lack such freedom. FFFS Algorithmic Details {#app.fffs} ======================== In this appendix, we present and discuss the FFFS algorithm used in our local explainer method. The main algorithm is presented in Algorithm \[alg:fffs\], and the required subroutines are presented in Algorithms \[alg:Recur\], \[alg:sf\], and \[alg:bin\]. Since the main structure of the algorithm requires a recursive tree traversal, Algorithm \[alg:fffs\] includes a general prepossessing wrapper algorithm that starts the recursion. In this part of the algorithm, the sampled data points are used to compute the empirical densities of their feature values. These densities are approximated using histograms which can vary in the number of bins. For simplicity of presentation, we assume each histogram has the same bin size, but of course this detail can be modified in implementation. The key addition here is the computation of tensor $M$, which tracks the inclusion of each data point’s features into their respective histogram bin. Algorithm \[alg:Recur\] contains the main recursion of the filter computation, and it outputs the selected features when it terminates. The recursion of Algorithm \[alg:Recur\] requires a set of selected features $S$, a set of unselected features $U$, the binary tensor $M$, the black box model predictions $Y$, and $\mathcal{L}$, which is a set of partition sets of points in $T(\bar{x})$. Since no features are selected prior to the first call to Algorithm \[alg:Recur\], we initialize the inputs $S =\emptyset$, $U=\Phi$, $Y = f(T(\bar{x}))$ and $\mathcal{L} = T(x_i)$ when it is first called in Algorithm \[alg:fffs\]. The recursion terminates and outputs the current set of selected features when either all features are selected or $\mathcal{L}$ becomes empty. If the termination condition is not met, Algorithm \[alg:Recur\] calls Algorithm \[alg:sf\], which updates $S,U, \text{ and }\mathcal{L}$ using a bin expansion. Then Algorithm \[alg:Recur\] makes a recursive call with updated inputs and repeat the previous steps. Algorithm \[alg:sf\] is used to select one feature from the set of unselected features that maximizes the mutual information $I(\varphi;Y|U)$, and to update $\mathcal{L}$ given the current selected feature. We apply forward selection in Algorithm \[alg:sf\]. In order to find $\varphi^*=\operatorname*{arg\,max}_{\varphi \in U} I(\varphi;Y|U)$, we compute $I(\varphi;y|S)$ for each unselected feature $\varphi$. The approximated mutual information $I(\varphi;y|S)$ is computed using the following equation [@brown2012conditional]: $$I(\varphi;y|S) \approx {\hat{I}(\varphi;y|S)}=\frac{1}{|T(\bar{x})|}\sum_{i=1}^{N}\log\frac{\hat{p}(\varphi;y|S)}{\hat{p}(\varphi|S)\hat{p}(\varphi|S)}.$$ If $I(\varphi^*;y|S)$ is not positive, then we do not select any new features. If no new feature is selected, we terminate the process by setting $U=\emptyset$, which satisfies the termination condition of Algorithm \[alg:Recur\], and the feature selection process will be complete. If $I(\varphi^*;y|S)>0$, then we can obtain additional information on the prediction by adding $\varphi^*$ to the set of selected features $S$ and removing it from the set of unselected features $U$. Algorithm \[alg:sf\] then calls Algorithm \[alg:bin\] to update $\mathcal{L}$ to $\mathcal{L}'$. Algorithm \[alg:bin\] is used to partition each set in $\mathcal{L}$ given current selected feature $\varphi^*$. Using the binary tensor $M$, we can collect the set of bins for $\varphi^*$. As an illustrative example of this process, let $B_{\varphi^*}=\{b_1,b_2\}$ and $\mathcal{L}=T(\bar{x})=\{x_1,x_2,....,x_p\}$. Assume $x_i^{\varphi^*}\in b_1$ for $i < 5$ and $x_i^{\varphi^*}\in b_2$ otherwise. Then we can partition the set $\{x_1,x_2,....,x_p\}$ into 2 sets $\ell_1, \ell_2$ s.t. $\ell_1=\{x_1,...,x_4\}$ and $\ell_2=\{x_5,...,x_p\}$. Next we add sets $\ell_1,\ell_2$ to $\mathcal{L'}$. Since $\mathcal{L}$ contains exactly one set, we finish the partition process, and Algorithm \[alg:bin\] outputs $\mathcal{L'}=\{\{x_1,...,x_4\}, \{x_5,...,x_p\}\}$. $T(\bar{x}),\Phi, f$ Form histogram with bin set $B_\varphi$ and frequencies $\hat{p}_\varphi$ set $M\in |B_\varphi| \times |\Phi| \times N$ as a zero tensor Set $M[b,\varphi,x] = 1$ RecursionFFS($\emptyset,\Phi,M,f(T(\bar{x})),T(\bar{x})$) $S,U, M,Y,\mathcal{L}$ $S$ $[S',U',\mathcal{L}']= \text{ SelectFeature}(S,U,M,Y,\mathcal{L})$ RecursionFFS$(S',U',M,Y,\mathcal{L}')$ $S,U, M,Y,\mathcal{L}$ $f^* = \operatorname*{arg\,max}_{f \in U} I(f;Y|U)$ $U = U \setminus f^*$ $S = S \cup f^*$ $\mathcal{L}' $= BinPartition $(M, \mathcal{L},f^*)$ $S,U,\mathcal{L}'$ $U=\emptyset$ $S,U,\mathcal{L}$ $M,\mathcal{L},f^*$ Use $M$ to find $B_{f^*}$ s.t. $B_{f^*}=\{b_1,b_2,...,b_k\}$ is the set of bins for feature $f^*$ $\mathcal{L'} = \emptyset$ Partition $\ell$ into smaller sets $\{\ell_1,\ell_2,...\ell_k\}$ w.r.t $B_{f^*}$: $\ell_i=\{t \in l : t^{f^*} \in b_i\} \forall i \in \{1,...,k\}$ $\mathcal{L}'= \mathcal{L}' \cup \{l_1,...,l_k\}$ $\mathcal{L}'$ The time complexity of the FFFS algorithm for a fixed maximum discretization bin size is $\mathcal{O}(N|\Phi|)$. Note that the size of the generated points is given by the input parameter $N$, and the set of all features is denoted by $\Phi$. First, since the bin sized is fixed as a constant, and the preprocessing step requires a nested **for** loop, the total time complexity of the preprocessing is $\mathcal{O}(N|\Phi|)$. The FFFS algorithm operates as a tree traversal, where the depth of the tree at the final stage corresponds to the number of selected features. In each level of the tree, the mutual information of all points is evaluated using Algorithm \[alg:sf\] and the sets of generated points are partitioned into smaller sets using Algorithm \[alg:bin\], which combined require $\mathcal{O}(N)$ operations. Next, since in the worst case, all features contain positive mutual information on the prediction value of the black box model, the maximum possible tree depth is given by $|\Phi|$. Combining these two facts gives the desired result. Additional figures {#app.fig} ================== ![Elbow plot for determining number of clusters to use for $k$-means clustering. Red marked value is located at 4 clusters and roughly corresponds to the bend in the elbow. The $x$-axis describes the total number of clusters used in $k$-means clustering, and the $y$-axis represents the MSE loss associated with the resulting clusters.[]{data-label="fig:elbow_plotl"}](alt_elbow.pdf){width="50.00000%"}

ArXiv

--- abstract: 'In over forty presentations on experiments at the 2012 conference on Flavor Physics and CP Violation (FPCP 2012), there was an abundance of beautiful and significant results. This summary of these experiment presentations begins with a reminder of the context in which the measurements have been made and the motivations for making the specific measurements reported at the symposium. Given the number and breadth of physics topics covered at the meeting, this review covers only a limited set of highlights, sort of a traveler’s set of souvenir postcards of favorite slides. The selected slides are grouped into eight overlapping categories as an aid to flipping through the postcards and being reminded of the high points of the conference. Finally, there are some summarizing comments about how the experiment results presented here compare to expectations and what we may hope for the future.' author: - 'Jeffrey A. Appel' title: FPCP 2012 Summary Talk on Experiments --- Introduction ============ Before turning to the experiment presentations at FPCP 2012, it may be useful to review the broader context of the FPCP conference and of the measurements. We often say that particle physics is the study of matter, energy, space, and time. What do we really want to know? For matter, we want to know: - Why is the universe so dominantly matter; why is there so little antimatter around? - Why is matter made of quarks and leptons, antiquarks and antileptons? - Why are these constituents spin $\frac{1}{2}$ particles? - Why do quarks and leptons come in three generations? (And, are there only three?) - Why do the quark generations have such different masses? - Why are neutrino masses so small; and why different by generation? Are neutrinos Dirac or Majorana particles? - Why do quarks come in three colors? For energy: - Why is so little of the energy density in the universe composed of the mass that I just listed? - What is the dark matter we claim is the rest of the matter? - Why are the force carriers spin 1; not spin 0 for example? Why are gravitons spin 2? - Why are the strong and electroweak forces flavor independent? - Why aren’t all interactions flavor independent? And, for space and time: - Why are there three obvious spatial dimensions? Are there more? - Why is the expansion of the universe accelerating? Or, do we not understand gravity/space? As an aside, note the preponderance of the number three in my list of questions: three dimensions, three generations, and three colors. Are all these trinities related? Are any of them related? Sometimes we have ideas about possible keys to unlocking the answers to some of these questions: - Some kind of substructure to explain the pattern of quarks and leptons and generations we see. - The Higgs mechanism as what nature has chosen for ElectroWeak Symmetry breaking? Do the masses of the $W$ and $Z$ come from the same mechanism as that for quark masses. - A seesaw mechanism involving very massive right-handed partners as the source of the very light neutrino masses. - The neutrino sector as the source of the matter/antimatter asymmetry of the visible universe. We know now that the asymmetry is not from the CKM matrix in the quark sector (inadequate by 8-10 orders of magnitude). Not all these ideas are directly testable. Those that are, we are working hard to test. We are doing it by going to higher energies at the LHC and by making more and more precise measurements in the flavor sector. The latter is the focus of our FPCP conferences, of course. I am reminded of the story of the drunk and the lamp post. Maybe this is only well known in the West, so indulge me if I tell it here. A drunk has lost his keys and is spending a long time looking for them under a lamp post. When asked why he is still looking there, he says that that is where there is enough light to see them! We need to be careful not to make the same mistake. Certainly, look where we think we may find the answers because we have good models to test. However, in parallel, probe as deeply as we can where we don’t have such light to guide us. We have vast new data sets, and we need to check for the unexpected, too. As a trivial example, in studying decays of particles with heavy flavor to $h^{\pm} \ell^+ \ell^-$, look for $h^- \ell^+ \ell^+$, etc. as well. No matter what our theorist friends tell us about where the answers lie, we have already seen the preferred space of minimal SUSY disappear. And, the space for the Higgs to hide is also closing down. Sorry, I am supposed to say that we are closing in on the Higgs at about 125 $GeV/c^2$! (I note in this post conference write-up the announcements from the LHC and Tevatron after the FPCP 2012 conference indicating the possible discovery of a Standard-Model Higgs boson [@ATLAS-Higgs; @CMS-Higgs]!) Results shown at FPCP 2012 ========================== This brings us to the beautiful results shown at FPCP 2012, many results new since the last meeting, some shown here publicly for the first time. Since I cannot include all my favorite slides in the write-up of the talk, I can only refer the reader to the slides for this summary available from the Proceedings link on the conference web site [@FPCP] as an accompaniment to reading this article. As you look over the slides I have selected from the meeting, you will see a personal selection. Afterall, choosing post cards when you are on travel is a personal matter. So, I show some of my favorite postcards from FPCP 2012, those I found especially pretty or revealing. I won’t repeat all the excellent explanations of the results. I could not do as well as we have heard from the presenters themselves. These explanations are available in slides from the other talks, also available at the conference web site, or in the individual write-ups from the presenters, also available on links from the conference web site. I would certainly choose a postcard each to remind me of the conference site, the Chinese opera we saw, and the Bao Gong Memorial Park. However, there are also lots of postcards to select with pretty plots and physics content. My souvenir postcards are organized in eight categories: - Standard-Model confirmations - Significant reductions in uncertainty - New-Physics space ruled out - Tension with the Standard Model - New signals and structures - Hints of new physics - New techniques and looking beyond the lamp post - Postcards of the future Remember, these are just souvenir post cards, visually impressive views of the various physics topics presented, not part of a guide book. See the individual talks for details at the level of a guide book. It should also be obvious that most measurements that I reference could be listed for more than one category. Standard-Model confirmations ---------------------------- New measurements of the speed of neutrinos have confirmed that the earlier measurement of neutrino speed by the OPERA experiment (now revised also by OPERA itself) was wrong. Neutrinos are not superluminal. They do not travel faster than the speed of light. In his talk, Andrew Cohen noted [@cohen] that we should have realized this from known physics, in particular the lack of radiation by high-energy neutrinos from very far away, for example. The OPERA result did motivate serious thinking about the issue, and generated calculations of new, very sensitive limits on the violation of Lorentz invariance by neutrinos. Andrzej Bozek showed Belle’s result on the rate of $D_S \rightarrow \tau \nu$, consistent with lepton universality [@bozek]. There is also continuing progress in measurements of $CP$ violation in the quark sector. Giovanni Marchiori showed BaBar’s first three-sigma evidence of $CP$ violation in the decay of the $B$ to three $K_S^0$’s from an analysis of the time-dependence of the decays [@marchiori]; and we saw the first three-sigma evidence of $CP$ violation in $B_S$ decays from LHCb as shown by Irina Nasteva [@nasteva]. Also, note the evidence shown by Yuehong Xie that the heavier $B_S$ lives longer than its lighter sister (also from LHCb) [@xie]. Finally, the window for a Standard-Model Higgs continues to be better defined with the more precise measurements of the $W$ mass by CDF and DZero at the Tevatron [@charlton]. The continued reduction in the “oval of uncertainty" in the Higgs mass from ever more precise top-quark and $W$ mass measurements at the Tevatron (consistent with a light Higgs as predicted from electroweak measurements and possibly observed as announced soon after the FPCP meeting [@ATLAS-Higgs; @CMS-Higgs]). Significant reductions in uncertainty ------------------------------------- Perhaps the most startling reduction in a measurement uncertainty has come with the surprisingly-quick measurement of $\theta_{13}$ of the neutrino-mixing matrix parameterization as shown in the talks of Werner Rodejohann, Jianglai Liu, and Phillip Litchfield [@rodejohann; @liu; @litchfield]. The value measured, first and best so far by the Daya Bay experiment, is near the previous upper limit on this parameter. The optimists were right in this case. There is also the improvement in the uncertainties in the parameters of the so-called unitarity triangles of the CKM quark-mixing matrix as shown in individual presentations and summarized in the talk by Sebastien Descotes-Genon [@genon]. Also impressive are the measurements of the properties of the $h_C$ by BESIII shown by Guangshun Huang [@huang] and improvements in the now-lower value of $y_{CP}$ in $D^0$ decays, including the new measurements from BaBar and Belle, reported by Chunhul Chen [@c-chen]. Finally, I note separately the improvements in $B_S$ mixing parameters, including the value of $\phi_S$ reported by Sebastien, Fabrizio Ruffini, and Yuehong Xie [@genon; @ruffini; @xie], progress in reducing the semileptonic decay uncertainties, e.g., in bottom decays as reported by Vera Luth [@luth], and in charm form factors from BESIII as reported by Jonas Rademacker [@rademacker]. New physics space ruled out --------------------------- The parameter space for physics beyond the Standard Model is multidimensional. We have become used to presentations of limits when models are reduced to two relevant parameters, whether we are talking about dark matter in terms of cross sections and particle mass or SUSY models at the selected internal-parameter level. Some of the slides presented show measurement limits for physical parameters with a range of model-possibility predictions of the parameters as generated by Monte Carlo techniques to give a sense of how effective the measurements are in restricting the range of model parameters. Other limits come directly from two-dimensional plots of possible model parameters with sections of the space ruled out by the measurements. Some of the plots are quite colorful and artistic in appearance, as well as providing physics insight. Some of my favorite plots come from the presentation on dark-matter search limits by Xinchou Lou [@lou] both for generic dark matter and for a possible dark Higgs. Improved limits on new physics come from lepton-flavor violation searches as in the slide on $\mu \rightarrow e \gamma$ shown by Francesco Renga [@renga] and from the ratio of leptonic two-body $K$ decays, $R_K = \Gamma(K \rightarrow e \nu)$ over $\Gamma(K \rightarrow \mu \nu)$ shown by Evgueni Goudzovski [@goudzovski]. Two of the most colorful plots are those showed by Vincenzo Chiochia where constraints on new physics come from top-quark-production asymmetries (forward-backward and charge asymmetries) measured at the Tevatron and LHC [@chiochia] and by Nicola Serra showing limits from the rare processes $B \rightarrow \mu \mu$ and $B_S \rightarrow \mu \mu$ [@serra] taken from the presentation by David Straub at the EW Rencontres de Moriond this year. Finally, I would include in my postcard collection, the distributions of observed events in the search for $\tau \rightarrow \mu \mu \mu$ at LHCb shown by Paul Seyfert [@seyfert]. Tension with the Standard Model ------------------------------- Various fits to measurements of CKM unitarity triangles have been shown to highlight possible discrepancies in the single-phase paradigm of the Standard Model. Tensions with the Standard-Model overall fits have been observed, mention being made by Sebastien Descotes-Genon [@genon] and Koji Hara [@hara] of issues between the value of $\sin (2\beta)$ and the rate of $B$ decay to $\tau \nu$. Also mentioned by Sebastien were the semileptonic asymmetry in $B$-decay and $B_s$-decay parameters and the same-sign dimuon charge asymmetry measured by CDF and DZero at the Tevatron. Amarjit Soni showed a nice plot from a fit without inputs from semileptonic decays to address concerns over sensitivity to $V_{cb}$ [@soni]. Vera Luth noted that the “tension" between inclusive and exclusive analyses of semileptonic B decays remains [@luth], while stated uncertainties on the branching fractions and on $|V_{ub}|$ and $|V_{cb}|$ are being reduced. The search for discrepancies between measurements and Standard-Model predictions of $B$-decay rates has revealed a significant excess of events in $B \rightarrow D \tau \nu$ and in $B \rightarrow D^* \tau \nu$ (3.4 $\sigma$ when the two BaBar excesses are combined). This feature cannot be easily explained. Finally, I would point to the nice plots in the presentation by Rick Van Kooten [@vankooten] showing some tension in $B_S$ decay parameters. New signals and structures -------------------------- Clearly, the observation of electron neutrinos coming from the oscillation of muon neutrinos, giving the large observed value of $\theta_{13}$, is a new signal. I show two more plots on this, which include the individual measurements from Daya Bay, Reno, Double Chooz, T2K, MINOS, KamLAND, and solar-neutrino measurements [@liu; @litchfield]. We also saw unexpected structures in baryonic $B$ decays from BaBar as shown by Irina Nasteva [@nasteva] and, at 6.3 $\sigma$, the first observation of $B_S \rightarrow \phi \mu \mu$ from CDF shown by Rick Van Kooten [@vankooten]. The evidence for new $Z$-onium states is becoming more and more compelling, given the nice plots of $Z_{1,2} \rightarrow h_b(1,2) \pi$ from Belle shown by Jin Li [@li]. Another visually compelling set of plots demonstrated the suppression of the production the heavier upsilon mesons relative to the ground-state upsilon in heavy-ion collisions at CMS as shown by Zebo Tang [@tang]. Another signal that is becoming clear with the recent increase of data at LHCb is the zero-crossing point in the forward-backward asymmetry in the decay of $B \rightarrow K^* \mu \mu$ shown by Nicola Serra [@serra]. I have also selected three slides from the presentation of BESIII results shown by Guangsung Huang [@huang]: the newly observed isospin-breaking decay of $\eta(1405) \rightarrow f_0(980) \pi^0$, an anomalous lineshape of the $f_0$ in the decay $J/\psi \rightarrow \gamma f_0 \pi^0$, and the first evidence of $\psi(2S) \rightarrow \gamma \gamma J/\psi$. Perhaps the most unusual new signal was the first direct observation of time-reversal-symmetry violation in any system. The direct observation comes from using entangled $\Upsilon(4S)$ decays to tagged $B$ mesons. The results shown by Pablo Villanueva Perez [@perez] come from BaBar. As already noted, from a time-dependent Dalitz-plot analysis, we saw the first evidence of CPV in $B_S \rightarrow K_S K_S K_S$ in the talk by Giovanni Marchiori. Hints of new physics -------------------- Among the hints of new physics is the forward-back asymmetry in $t \bar{t}$ production at CDF and DZero, increasing with the mass of the $t \bar{t}$ system at CDF as shown by Marc Besancon [@besancon]. There are also hints of the Higgs at ATLAS and CMS shown by David Charlton [@charlton], preceding the already-mentioned announcements after the FPCP conference. I continue to think of this as new physics, though most of you may already consider this a part of the Standard Model. The surprisingly-large $D^0$ mixing observed has led to suggestions that there might be $CP$-symmetry violation in the charm sector. Combining LHCb and CDF results on the difference of $CP$ asymmetries in $D^0$ decays to $K^-K^+$ and $\pi^-\pi^+$ led Vincenzo Vagnoni to say that the data “is consistent with no $CP$ violation at 0.006% CL" [@vagoni]. Even stated in this way, of course, at this point, we can only say that the data is inconsistent with no asymmetry at the given confidence level. The issue of $CP$ violation is still being debated among theorists! Can the asymmetry be due to Standard-Model long-distance effects? New techniques and looking beyond the lamp post ----------------------------------------------- New techniques include both the application of new analysis methods and improvements made to previous analyses. Manuel Tobias Schiller showed LHCb multibody-decay results using combined Gronau-London-Wyler (GLW) and Atwood-Dunietz-Soni (ADS) methods and also taking advantage of the strong variation of hadronic parameters over the Daliz-plot phase space [@schiller]. David Charlton showed how analysis improvements allowed CDF to obtain more precise results in their search for the Higgs boson than what one would expect from the simple increase in data as the integrated luminosity grew over time [@charlton]. Similar improvements have been made at DZero, and we may expect similar things from the LHC experiments too, as they accumulate more data and experience. As mentioned earlier, searching for the unexpected is looking beyond the lamp post. Liang Sun showed the event distributions for searches for an unexpected $B^+ \rightarrow D^- \ell^+ \ell^+$ by Belle [@sun]. Postcards of the future ----------------------- I was going to reserve the last slide for postcard images of the new facilities shown in the first session of the final day. However, I decided to focus on the physics we expect most. The fact is, I am less certain what to expect than I have been for many years. So, I have left space in my postcard collection for next year’s souvenirs! Summary ======= As we flipped through the souvenir picture post cards of experiment results selected from all those presented at FPCP 2012, we have seen a wonderfully-rich abundance of new results. Yet, the selection presented here is necessarily incomplete. At best, it gives a sense of the impressive range of activity in flavor-physics and $CP$-violation experiments, and of the very high quality of the data and analyses being generated. There were over 40 experiment talks! In these talks, there were presentations of results which included Standard-Model confirmations, significant reductions in uncertainty, New Physics spaces ruled out, new signals and structures, hints of New Physics, and still some tension with the Standard Model. And, there were also places where results were presented which explored the possibility of unexpected signals, looking beyond the lamp post. There were, perhaps, fewer outstanding experimental issues relative to what was presented at FPCP 2011, less tension with the Standard Model this year. Nevertheless, there is growing disquiet over not seeing directions for the answers to the questions about matter, energy, space, and time that were listed in my introduction. There is no certainty about the direction of the needed New Physics. \[The signals around 125 $GeV/c^2$ shown by ATLAS and CMS after FPCP 2012 also seems consistent with the Standard-Model Higgs boson, though this is by no means yet proven.\] We may hope for one or more bright new ideas that could provide additional lamp posts to light our way. But this hope is not something that we can count on soon. Thus, it is hard to predict what the post cards will look like next year at FPCP 2013 in Buzios, Brazil. Nevertheless, given the huge data sets collected and anticipated, there is excellent reason for hope! As has happened in the past, data may be the key to future progress in our understanding. I close this last talk of the Flavor Physics and CP Violation 2012 meeting with a big and sincere thank you to all the presenters. Obviously, I have taken freely from their presentations for this summary. I also want to thank especially the organizers of this very enjoyable, interesting, and informative meeting. My work is supported, in part, by the US Department of Energy through Fermilab, which is operated by Fermi Research Alliance, LLC under Contract No. DE-AC02-07CH11359 with the Department of Energy. [99]{} http://www.atlas.ch/news/2012/latest-results- from-higgs-search.html http://cms.web.cern.ch/news/observation-new- particle-mass-125-gev http://hepg-work.ustc.edu.cn/fpcp2012/ A. Cohen,“Theoretical aspects of neutrino speed anomaly", FPCP2012, Hefei, China, May, 2012. A. Bozek, “Leptonic $D$ decays", FPCP2012, Hefei, China, May, 2012. G. Marchiori, “$CP$ violation and rare $B$ decays at the $B$ factories", FPCP2012, Hefei, China, May, 2012. I. Nasteva, “Hadronic $B$ decays", FPCP2012, Hefei, China, May, 2012. Y. Xie, “$CP$ violation in $J/\psi \phi$", FPCP2012, Hefei, China, May, 2012. D. Charlton, “Higgs searches at LHC and Tevatron", FPCP2012, Hefei, China, May, 2012. W. Rodejohann, “Status of the PMNS matrix", FPCP2012, Hefei, China, May, 2012. J. Liu, “Status report from Daya Bay", FPCP2012, Hefei, China, May, 2012. P. Litchfield, “$Theta_13$", FPCP2012, Hefei, China, May, 2012. S. Descotes-Genon, “CKM matrix and constraints on new physics from quark flavour observables", FPCP2012, Hefei, China, May, 2012. G. Huang, “Recent results from BESIII", FPCP2012, Hefei, China, May, 2012. C. Chen, “$D$-$D$bar mixing, direct and indirect $CP$ violation", FPCP2012, Hefei, China, May, 2012. F. Ruffini, “Hadronic $B_s$ decays", FPCP2012, Hefei, China, May, 2012. V. Luth, “Semileptonic $B$ decays", FPCP2012, Hefei, China, May, 2012. J. Rademacker, “Semileptonic $D$ decays", FPCP2012, Hefei, China, May, 2012. X. Lou, “Searches for low mass new physics", FPCP2012, Hefei, China, May, 2012. F. Renga, “Searches for $LFV$ and $LNV$ in charged lepton decays", FPCP2012, Hefei, China, May, 2012. E. Goudzovski, “Recent kaon physics results and prospects", FPCP2012, Hefei, China, May, 2012. V. Chiochia, “Top quark properties and top decays", FPCP2012, Hefei, China, May, 2012. N. Serra, “$B_s \rightarrow \mu \mu, K^* \mu \mu$", FPCP2012, Hefei, China, May, 2012. P. Seyfert, “Searches for $LFV$ and $LNV$ in hadron decays", FPCP2012, Hefei, China, May, 2012. K. Hara, “$B$ leptonic decays", FPCP2012, Hefei, China, May, 2012. A. Soni, “New ideas and directions in flavor physics/$CP$ violation", FPCP2012, Hefei, China, May, 2012. R. Van Kooten, “Semileptonic $B_s$ decays", FPCP2012, Hefei, China, May, 2012. J. Li, “New and conventional bottomonium states", FPCP2012, Hefei, China, May, 2012. Z. Tang, “Heavy flavor production in heavy ion collisions", FPCP2012, Hefei, China, May, 2012. P. Villanueva Perez, “First direct measurement of time-reversal violation $B$ decays", FPCP2012, Hefei, China, May, 2012. M. Besancon, “Top quark production at the Tevatron", FPCP2012, Hefei, China, May, 2012. V. Vagnoni, “$CP$ violation in $D$ decays", FPCP2012, Hefei, China, May, 2012. M. Tobias Schiller, “Measurements of CKM angle gamma from tree-dominated decays", FPCP2012, Hefei, China, May, 2012. L. Sun, “$b \rightarrow s \gamma$", FPCP2012, Hefei, China, May, 2012.

ArXiv

--- abstract: 'Hue modification is the adjustment of hue property on color images. Conducting hue modification on an image is trivial, and it can be abused to falsify opinions of viewers. Since shapes, edges or textural information remains unchanged after hue modification, this type of manipulation is relatively hard to be detected and localized. Since small patches inherit the same Color Filter Array (CFA) configuration and demosaicing, any distortion made by local hue modification can be detected by patch matching within the same image. In this paper, we propose to localize hue modification by means of a Siamese neural network specifically designed for matching two inputs. By crafting the network outputs, we are able to form a heatmap which potentially highlights malicious regions. Our proposed method deals well not only with uncompressed images but also with the presence of JPEG compression, an operation usually hindering the exploitation of CFA and demosaicing artifacts. Experimental evidences corroborate the effectiveness of the proposed method.' author: - - - bibliography: - 'references.bib' title: Hue Modification Localization By Pair Matching --- Hue modification, patch matching, Siamese network Introduction {#sec:intro} ============ Modern photography is losing its innocency due to the diversed use of image manipulation software, which allows even unexperienced users to modify digital images in different ways. Image contents are characterized mainly by geometric information like texture, edges and shapes, and by color information. Color modifications, even if does not effect geometric details, deceive human perception. They are very easy to be performed, and hard to be detected if implemented carefully. In this paper, we address the problem of local hue modification, which is defined as the adjustment of angular position on the color circle (or color wheel) within an image area. Figure \[fig:hue\_mod\] illustrates hue modification by different angles [^1]. ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ [ \[+(.5ex)\][![Hue modification by different angles. $0^\circ$ means no modification. Better viewed in color.[]{data-label="fig:hue_mod"}](mod_hue_0.png "fig:"){width="17mm"}]{} ]{} [ \[+(.5ex)\][![Hue modification by different angles. $0^\circ$ means no modification. Better viewed in color.[]{data-label="fig:hue_mod"}](mod_hue_60.png "fig:"){width="17mm"}]{} ]{} [ \[+(.5ex)\][![Hue modification by different angles. $0^\circ$ means no modification. Better viewed in color.[]{data-label="fig:hue_mod"}](mod_hue_120.png "fig:"){width="17mm"}]{} ]{} [ \[+(.5ex)\][![Hue modification by different angles. $0^\circ$ means no modification. Better viewed in color.[]{data-label="fig:hue_mod"}](mod_hue_180.png "fig:"){width="17mm"}]{} ]{} $0^\circ$ $60^\circ$ $120^\circ$ $180^\circ$ ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ To cope with local image manipulations, previous works seek for artifacts of Color Filter Array (CFA) and camera sensor pattern noise. Vast camera sensors employ a CFA, where each sensor element captures the light at a certain wavelength corresponding to a color component. The remaining color components at blind positions are interpolated from surrounding pixels. This interpolation is referred to as *demosaicing*. Image manipulations will likely generate some local or global disturbances which are inconsistent to our ordinary demosaicing artifacts [@Ferrara2012]. More blind way to detect local disturbances is the extraction of statistical features of rich models capturing different types of neighboring dependencies [@Fridrich2012]. These features have been proved to be effective in manipulation detection and localization, see for instance [@Cozzolino2015; @Li2017ForgeryLocalization]. Besides demosaicing, the imperfections of camera sensors also create sort of camera fingerprint, the so-called Photo-Response Nonuniformity (PRNU) noise, which is supposed to be present in every image [@Chen2008]. In the presence of manipulation, this pattern noise is distorted and this distortion can be exploited as a useful clue, provided that the reference PRNU can be reliably estimated and the forged region is sufficiently large. The specific local image manipulation considered in this paper, hue modification, distorts artifacts of demosaicing and neighboring dependencies. Based on this fact, the pioneering work in [@Choi2013] analyzes demosaicing artifacts and then estimates hue modification. Based on the observation that an interpolated value is bigger than the minimum and smaller than the maximum of its neighborhood, on the green channel the number of pixel values unsatisfying this condition should be the majority of pixel which are originally captured in this channel and a minority of interpolated pixels, resulting a big ratio between two quantities. The estimation of hue modification is done via searching over a set of modification angles until the aforementioned ratio is maximized. We want to point out that CFA analysis requires the knowledge of CFA configuration, at least the positions of green component. Such information is not always available, especially for online images. Moreover, when the image undergoes JPEG compression, demosaicing artifacts are significantly distorted. Differently, the method proposed in [@Hou2014] and [@Hou2017] recovers the modification angle, by modifying the questioned image with a set of angles and matching its residual with the reference PRNU. In real scenarios this technique is very difficult to be exploited since the assumption to know the reference PRNU (or have access to images to estimate it) is very strong and cannot be easily satisfied. In this work, we propose a novel method for detecting hue modification. Our methodology exploits the fact that two patches on the same image have the same inherent CFA configuration and demosaicing. Hue modification on a local region creates inconsistencies with the rest of the image, and thus pair-wise patch matching can reveal the forged region. To achieve such purpose, we propose a solution based on Siamese neural networks [@Bromley1993], trained on positive pairs (two pristine patches) and negative pairs (a pristine and a modified patch). JPEG compression before and after hue modification is included during training, granting the network the capability to deal with real-world conditions. Finally, we fuse multiple outputs of patch matching to obtain a unique decision map (heatmap), on which a postprocessing is applied to precisely localize the forged region (Section II). Experiments demonstrate the effectiveness of the proposed approach (Section III). Proposed Method {#sec:method} =============== Hue modification is performed in the HSV space by adding an angle $\alpha$ to the value of $H$. Since defined on a circle, hue modification is periodic with a period $360^\circ$, i.e. a modification of $\alpha$ is identical to $\alpha \pm 360^\circ$. Besides hue, other attributes of a color in HSV space are saturation and value (brightness), whose changes are different from hue modification. Here, we investigate the detection of hue modification on: i) uncompressed images, ii) JPEG images where the modification is carried out before and after compression. Given two rectangular patches $\mathcal{P}_i, \mathcal{P}_j$ of size $h \times w$ from the same image, we desire to estimate the logistic prediction $p_{ij}$ that two patches are inconsistent with respect to two corresponding modification angles $\alpha_i, \alpha_j$. The two patches are *consistent* if $\alpha_i = \alpha_j = 0$ and *inconsistent* if $\alpha_i \neq \alpha_j = 0$. We propose to verify the inconsistency of $\mathcal{P}_i$ and $\mathcal{P}_j$ by means of a Siamese neural network [@Bromley1993]. Siamese neural networks have been recently exploited for applications in multimedia forensics [@Cozzolino2018NoisePrint; @Mayer2018; @Huh2018]. This network architecture consists of two identical sub-networks $f_\theta$, followed by a non-linear classifier $g_\gamma$ that outputs an inconsistency score $z_{ij}$ whose standard logistic activation is defined as: $$\begin{aligned} p_{ij} = \frac{1}{1+e^{-z_{ij}}} \text{, and } z_{ij} = g_\gamma \left ( \left [ f_\theta \left (\mathcal{P}_i \right ) - f_\theta \left (\mathcal{P}_j \right ) \right ]^2_{\text{pointwise}} \right) \nonumber \text{.}\end{aligned}$$ The network parameters $\theta, \gamma$ are jointly optimized to minimize the binary cross-entropy of network logistic predictions $p_{ij}$ and patch inconsistencies $y_{ij}$, written in terms of a loss function over $N$ training patches: $$\mathcal{L} = \frac{1}{N^2-N}\sum_{i=1}^{N} \sum_{j=1,j \neq i}^{N} - y_{ij} \log(p_{ij}) - (1-y_{ij})\log(1-p_{ij}). \nonumber$$ In Figure \[fig:net\_arch\], we provide the sketch of the network architecture. We use the 50-layer Residual Network (ResNet50) [@He2016] as the feature extractor $f_\theta$, which outputs a $256$-dimensional feature vector. The inconsistency of features extracted from two patches are evaluated by a pointwise squared difference operator. The classifier $g_\gamma$ is a multilayer perceptron network composed by one hidden layer of $16$ units and one single-unit output layer with sigmoid activation outputting $p_{ij}$. ![Proposed Siamese network architecture.[]{data-label="fig:net_arch"}](net_arch.pdf){width="\linewidth"} We train two separate Siamese networks end-to-end on large-scale synthetic training sets. The first model is trained on $400,000$ $64 \times 64$ patches extracted from uncompressed images from RAISE [@Dang-Nguyen2015] and Dresden [@Gloe2010]. To train the second model, we use the same $400,000$ patches and perform hue modification before or after JPEG compression with random quality factors in $[55,100]$. Parameters $\theta$ are initialized using ResNet50 pretrained on ImageNet [@Deng2009]. On each training iteration, we optimize the loss function with respect to $(\theta, \gamma)$ on a mini-batch of $64$ pairs, half of which is labeled as positive, i.e. both two patches are unmodified, and another haft is labeled as negative, i.e. one patch is modified by an angle randomly selected in $[30,330]$ with step 30, and its counterpart is unmodified. Hue modification and JPEG compression are carried out during training. We use Adam optimizer with the starting learning rate $10^{-4}$, and schedule to halve it every $5$ epochs after the first $30$ epochs until convergence. Detection and Localization -------------------------- ### Heatmap creation The described architecture outputs the logistic patches inconsistency. Given a test image, we collect all inconsistency scores and generate a unique localization heatmap which potentially indicates malicious regions. Let $H,W$ be height and width of the image, and $h,w$ be height and width of the small patch. By using a sliding window with stride $s$, the total number of patches will be $N = N_H \times N_W$, where $N_H = \floor*{\frac{H-h}{s}} + 1$ and $N_W= \floor*{\frac{W-w}{s}} + 1$ are number of patches along each dimension. Generally, computing inconsistency scores on all possible pairs is expensive because the number of pairs grows quaratically w.r.t. $N$. Nevertheless, almost computational burden is attributed to operations of feature extraction network $f_\theta$ which composes convolutional layers. In pairwise manner, one patch is paired with other $N-1$ patches and passed through $f_\theta$ about $N-1$ times. This redundancy can be reduced. We first pre-extract low-dimensional features of all patches by evaluating $f_\theta (\mathcal{P}_i)$, $1 \leq i \leq N$, and proceed to compute $p_{ij}$ for all possible pairs using all computed features. For each patch $\mathcal{P}_k$ within the image, an inconsistency map $\mathcal{I}^{k} \in \mathbb{R}^ {N_H \times N_W}$ is built. If we consider all patches according to their spatial location on the image, $\mathcal{I}^k_{ij}$ is the inconsistency of $(i,j)$-th patch and $\mathcal{P}_k$, where $1 \leq i \leq N_H$ and $1 \leq j \leq N_W$. It is typical to assume that the forged region is relatively small compared to the background, thus majority of $\mathcal{I}^{k}$ ($k$ refers to patches on the pristine region) exposes inconsistencies with the forged region, while remaining maps expose inconsistencies with the pristine region, as shown in Figure \[fig:hm\]. In order to fuse inconsistency maps of majority patches belonging to the pristine region to obtain a unique map $\mathcal{\bar{I}} \in \mathbb{R}^ {N_H \times N_W}$, we follow the approach in [@Huh2018], computing $\mathcal{\bar{I}}$ by mean shift algorithm[@Cheng1995], which iteratively finds mean of majority (mode). ![Fusing patch-level inconsistency maps into an image-level map.[]{data-label="fig:hm"}](inconsistency_map.pdf){width="\linewidth"} Eventually, $\mathcal{\bar{I}}$ is a subsampled heatmap which potentially highlights malicious region. The full-size heatmap can be obtained by resizing $\mathcal{\bar{I}}$ with bilinear interpolation. If the forged region is larger than the background, we obtain the inverted heatmap since the background is the smaller area. ### Postprocessing {#sec:pp} The standard logistic output $p_{ij}$ can be interpreted as the posterior probability that two patches $\mathcal{P}_i$ and $\mathcal{P}_j$ are inconsistent. After mean shifting, each element $\bar{p}_{ij} = \mathcal{\bar{I}}_{ij}$ tells us how probable $(i,j)$-th patch is forged because $\mathcal{\bar{I}}$ is the representative inconsistency map of pristine patches to all patches. While the threshold $0.5$ may be a reasonable choice for deciding if two patches are inconsistent, it is not straightfoward to apply this rule to pixel-level predictions. Moreover, as keeping False Alarm Rate (FAR) low is critial in forensic applications, a postprocessing step is important for pixel-level prediction. With this respect, postprocessing on each image is cast to finding a statistical threshold $\tau$ based on which a pixel is masked as forged or pristine. We apply a simple postprocessing based on the assumption that $\mathcal{\bar{I}}_{ij}$ (to avoid adding new notation, we mean $\mathcal{\bar{I}}$ after resized) follows a Gaussian distribution, $\mathcal{\bar{I}}_{ij} \sim N(\mu, \sigma^2)$. We fix $\tau$ such that $5\%$ of the right tail are decided as being forged. $\tau$ is lower bounded by $0.5$ to maintain acceptable FAR, namely $\tau = \max \left( 0.5, t \right)$. $t$ is the solution of: $ 0.95 = \frac{1}{\sigma \sqrt{2\pi}} \int_{-\infty}^{t} e^{\frac{-(x-\mu)^2}{2\sigma^2}} dx \text{.} \nonumber$ Compared to the threshold $0.5$, $\tau$ results in better or equal FAR. An example of postprocessing is shown in Figure \[fig:pp\]. ![The heatmap before (left) and after (right) being postprocesed. The middle plot presents CDF of $\mathcal{\bar{I}}_{ij}$ and the threshold $\tau=0.61$.[]{data-label="fig:pp"}](postprocess.pdf){width="0.8\linewidth"} Experiments {#sec:exp} =========== Towards experimental evidences, we evaluate our approach under different configurations and test sets. Test set -------- To the best of our knowledge, there is no publicly available dataset on the problem of hue modification. Thus, to evaluate our method, we generated the test set from $120$ raw images of an external Canon 600D camera (never appeared in training phase) having CFA pattern $GBRG$. Raw images are decoded by version $9.27$. For each image, a top-left region is cropped out such that $H=768$ and $W=1024$. The forged area follows random convex shape fixed within a $256 \times 256$ bounding box, which is positioned at random location on the image. Next, we perform hue modification on pixels inside the polygons and generate multiple test sets: - $\mathcal{D}^\alpha_{png}$: Uncompressed images are demosaiced from raw images by and subject to local hue modification. For each modification angle $\alpha \in [30,330]$ step $30$, hue modification is carried out on $120$ uncompressed images. - $\mathcal{D}^{QF}_{b-jpg}$: Hue modification by different angles ($10$ images for each modification angle $\alpha \in [30,330]$ step $30$) are carried out on $110$ images, and $10$ images are unmodified. Afterwards, all images are compressed using quality factors $QF \in [55,100]$, step $5$. - $\mathcal{D}^{QF}_{a-jpg}$: $120$ images are first compressed using $QF \in [55,100]$, step $5$. Afterwards, hue modification by different angles (one angle for $10$ images) are carried out on $110$ JPEG images, while the remaining $10$ images are unmodified. All of them are compressed again using the default quality factor $75$. By the second JPEG compression, $\mathcal{D}^{QF}_{a-jpg}$ is more challenging since the training images are only subject to single JPEG compression. Setups ------ The performance of our method is compared with the following state of the art methods: [[Choi *et al.*]{}]{} [@Choi2013], based on CFA-based artifacts and explicitly designed for the estimation of hue modification, and SpliceBuster [@Cozzolino2015], based on statistical features of rich models [@Fridrich2012] and selected for comparison since those features potentially capture local disturbances caused by local hue modification. We do not compare with [@Hou2014; @Hou2017] given their strong assumption about the availability of the reference PRNU which is unrealistic in practical scenarios. This work particularly focuses on the localization of hue modification rather than its estimation. Choi *et al.* [@Choi2013] is an estimator which potentially returns the modification angle by searching over a feasible range. To convert Choi *et al.* into a localization method, we use a sliding window $35$ similarly to our method, and search the angle over $[0,359]$, step $8$. If the angle found is $0$ or $352$, the patch is marked as pristine. Choi *et al.* therefore outputs a binary map. The other method, SpliceBuster [@Cozzolino2015], returns the negative log-likelihood that a pixel is pristine. It means, a large value indicates high probability that a pixel is forged. We linearly scale the returned map into $[0,1]$ and apply the same postprocessing described in Section \[sec:pp\] to get the binary map. In order to demonstrate the advantage of our postprocessing, we also report performance of the proposed method when a simple thresholding is applied to binarize the heatmap. We empirically found that the threshold $0.8$ yields most acceptable results. We aggregate True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN) over all images and report average True Positive Rate (TPR), True Negative Rate (TNR) and F1 score. Quatitative Evaluation ---------------------- ### Detection on uncompressed images In this section, we evaluate the first model trained on uncompressed images, and compare with [[Choi *et al.*]{}]{} and SpliceBuster on $\mathcal{D}^\alpha_{png}$. Figure \[fig:png\_tpr\_tnr\] presents TPR and TNR obtained by all methods. ![TPR and TNR of all methods on $\mathcal{D}^{\alpha}_{png}$.[]{data-label="fig:png_tpr_tnr"}](png_tpr_tnr.pdf){width="\linewidth"} [[Choi *et al.*]{}]{} is guaranteed to detect hue modification on uncompressed images since this type of manipulation distorts demosaicing artifacts. It achieves high TPR which implies that almost forged pixels are correctly detected. This comes at a cost of slightly worse TNR. SpliceBuster, on the other hand, detects correctly only about $20\%$ of forged pixels, and as a consequence, yields very high TNR. We assume that the features used in [@Fridrich2012] are ineffective for hue modification detection. Our Siamese network with heatmaps thresholded simply by $0.8$ is denoted by [[Siamese-T-0.8]{}]{}. The other alternative is denoted by [[Siamese-G-0.95]{}]{}, where heatmaps are postprocessed by threshold $\tau$ as designed in Section \[sec:pp\]. We can clearly see that [[Siamese-G-0.95]{}]{} outperforms [[Siamese-T-0.8]{}]{} in all cases. In fact, a fixed threshold over all heatmaps cannot deal with high variability of predictive scores on each heatmap, and thus an adaptive threshold is more effective. Interestingly, the TPR reveals the fact that the middle range of modification angles are easier to detect by our methodologies. This is explainable since the strength of hue modification is periodic with the period of $360^\circ$. Very small or very large positive angles correspond to little modifications. We summarize the overall performance for some selective modification angles in Table \[tab:png\_f1\]. In terms of F1 score, [[Siamese-G-0.95]{}]{} outperforms all other methods. [ C[1.85cm]{} | C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} ]{} Angle $\alpha \rightarrow$ Method $\downarrow$ & 30 & 90 & 150 & 210 & 270 & 330\ & 66.16 & 67.15 & 65.05 & 68.34 & 67.92 & 68.21\ SpliceBuster & 18.12 & 29.57 & 32.99 & 26.80 & 21.97 & 12.29\ [[Siamese-T-0.8]{}]{} & 66.44 & 65.28 & 66.22 & 69.28 & 70.00 & 63.23\ [[Siamese-G-0.95]{}]{} & **71.79** & **73.82** & **73.04** & **74.41** & **74.29** & **69.86**\ \[tab:png\_f1\] ### Detection in the presence of JPEG compression We target more practical scenarios where hue modification is done with the presence of JPEG compression. It has been acknowledged that JPEG compression has strong impact on demosaicing artifacts [@Gallagher2008; @Cao2009; @Kirchner2010]. [[Choi *et al.*]{}]{} is also very sensitive to JPEG compression since the count of interpolated and recorded pixels is less accurate [@Choi2013]. We assess the second model trained on JPEG images under two testing circumstances: i) hue modification is performed on uncompressed images followed by JPEG compression, and ii) hue modification is performed on JPEG compressed images, and those are subsequently compressed again using quality factor $75$. Note that during training, we do not perform second JPEG compression. ![TPR and TNR of all methods on $\mathcal{D}^{QF}_{b-jpg}$ and $\mathcal{D}^{QF}_{a-jpg}$. $QF$ is the first compression quality factor.[]{data-label="fig:jpg_tpr_tnr"}](jpg_tpr_tnr.pdf){width="\linewidth"} The TPR and TNR of all methods, where hue modification is performed before JPEG compression, i.e. dataset $\mathcal{D}^{QF}_{b-jpg}$, is shown in the first column of Figure \[fig:jpg\_tpr\_tnr\]. [[Choi *et al.*]{}]{} fails to spot forged area unless the image is compressed with highest $QF$. At $QF=100$, [[Choi *et al.*]{}]{} achieves TPR $82.88\%$ and TNR $88.15\%$. SpliceBuster, on the other hand, can only detect about $10\%$ of forged pixels. Our proposed methods perform far better than the other two competitors on $\mathcal{D}^{QF}_{b-jpg}$, by keeping TPR at acceptable level and retaining always high TNR, i.e., [[Siamese-G-0.95]{}]{} achieves an average $64.67\%$ of TPR ($\approx 65\%$ of forged pixels are correctly detected) and $97.72\%$ of TNR, while [[Siamese-T-0.8]{}]{} attains $54.35\%$ of TPR and $97.39\%$ of TNR. In the right column of Figure \[fig:jpg\_tpr\_tnr\], i.e. dataset $\mathcal{D}^{QF}_{a-jpg}$, the overall TPR and TNR of our methods are slightly degraded compared to the performance on $\mathcal{D}^{QF}_{b-jpg}$. This degradation can be attributed to the second JPEG compression. In fact, we can generally observe the correlation of performance degradation and compression rate: the higher the first $QF$, the lower the performance. While [[Choi *et al.*]{}]{} behaves positively on $\mathcal{D}^{100}_{b-jpg}$, it loses that capability on $\mathcal{D}^{100}_{a-jpg}$. [ C[1.85cm]{} | C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} ]{} $QF$ $\rightarrow\;\;\;$ Method $\downarrow$ & 75 & 80 & 85 & 90 & 95 & 100\ & 9.80 & 9.74 & 9.69 & 9.85 & 10.05 & 39.47\ SpliceBuster & 12.80 & 12.47 & 14.26 & 14.66 & 15.57 & 21.44\ [[Siamese-T-0.8]{}]{} & 51.96 & 53.43 & 54.77 & 54.56 & 56.81 & 61.60\ [[Siamese-G-0.95]{}]{} & **61.41** & **60.83** & **63.03** & **64.65** & **66.73** & **69.11**\ \[tab:b\_jpg\_f1\] The overall F1 scores for several selective $QF$ are shown in Table \[tab:b\_jpg\_f1\] and \[tab:a\_jpg\_f1\]. Our two methodologies, in particular [[Siamese-G-0.95]{}]{}, outperform the other two methods to a large margin. We might notice that [[Choi *et al.*]{}]{} achieves F1 score $39.47\%$ on $\mathcal{D}^{100}_{b-jpg}$ while TPR and TNR in the same configuration are over $80\%$, see left column in Figure \[fig:jpg\_tpr\_tnr\]. This phenomenon is due to the high FP which penalizes precision, and as a consequence, F1 score. However, since TN dominates FP (due to the large pristine area compared to the forged area), TNR is not effectively penalized. [ C[1.85cm]{} | C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} C[0.5cm]{} ]{} $QF$ $\rightarrow\;\;\;$ Method $\downarrow$ & 75 & 80 & 85 & 90 & 95 & 100\ & 9.68 & 9.62 & 9.55 & 9.70 & 9.65 & 9.65\ SpliceBuster & 5.07 & 4.63 & 3.08 & 6.35 & 4.74 & 4.82\ [[Siamese-T-0.8]{}]{} & 52.85 & 46.76 & 41.29 & 49.93 & 46.13 & 46.09\ [[Siamese-G-0.95]{}]{} & **63.51** & **57.52** & **49.83** & **58.36** & **56.32** & **55.75**\ \[tab:a\_jpg\_f1\] Qualitative Inspection ---------------------- In Figure \[fig:realistic\_examples\], we provide detection results on realistic examples manually created using GIMP. Hue modification is carried out on uncompressed images (the first $5$ lines) and the modified images are JPEG compressed using highest quality $QF=100$ (the last $5$ lines). [[Siamese-G-0.95]{}]{} (the last column) clearly results in better detection maps compared with [[Choi *et al.*]{}]{} and SpliceBuster. Conclusion {#sec:con} ========== We have proposed a data-driven countermeasure for hue modification on color images based on patch matching. This task is done by means of a Siamese architecture which receives the two inputs and outputs the likelihood that the two inputs are inconsistent. A unique localization map is generated from inconsistency scores of multiple patches. Our models perform well on uncompressed and JPEG compressed images even though JPEG compression distorts CFA and demosaicing artifacts. Our future investigations will focus on the estimation of hue modification angles, based on which the original image can be recovered. ![Detection results on realistic examples.[]{data-label="fig:realistic_examples"}](examples.pdf){width="\linewidth"} [^1]: https://www.imagemagick.org/Usage/color\_mods/, last access: 15/02/2019

ArXiv

--- abstract: 'We introduce stochastic normalizing flows, an extension of continuous normalizing flows for maximum likelihood estimation and variational inference (VI) using stochastic differential equations (SDEs). Using the theory of rough paths, the underlying Brownian motion is treated as a latent variable and approximated, enabling efficient training of neural SDEs as random neural ordinary differential equations. These SDEs can be used for constructing efficient Markov chains to sample from the underlying distribution of a given dataset. Furthermore, by considering families of targeted SDEs with prescribed stationary distribution, we can apply VI to the optimization of hyperparameters in stochastic MCMC.' author: - 'Liam Hodgkinson[^1]' - 'Chris van der Heide[^2]' - 'Fred Roosta[^3]' - 'Michael W. Mahoney[^4]' title: '<span style="font-variant:small-caps;">Stochastic Normalizing Flows</span>' --- Introduction ============ *Normalizing flows* [@rezende2015variational] are probabilistic models constructed as a sequence of successive transformations applied to some initial distribution. A key strength of normalizing flows is their expressive power as generative models, while enjoying an explicitly computable form of the likelihood function evaluated on the transformed space. This makes them especially well-equipped for variational inference (VI). Neural networks are often used as inspiration for finding effective transformations [@dinh2014nice; @berg2018sylvester]. *Continuous normalizing flows* were later developed in [@chen2018neural] as a means to perform maximum likelihood estimation and VI for large-scale probabilistic models derived from ordinary differential equations (ODEs). The framework stems from the computation of the evolving density of an ODE with random initial value, as the solution to another ODE. The jump to continuous-time dynamics affords a few computational benefits over its discrete-time counterpart, namely the presence of a trace in place of a determinant in the evolution formulae for the density, as well as the adjoint method for memory-efficient backpropagation. Motivated by deep learning, a family of ODEs, called *neural ordinary differential equations* were constructed, whose Euler discretizations resembled layer-wise transformations of residual neural networks. Further algorithmic improvements to the framework were presented by @grathwohl2018ffjord, enabling virtually arbitrary choices of parameterized classes of ODEs. Doing all this involves some technical subtlety, and effective neural ODE architectures remain the subject of ongoing research — see for example [@dupont2019augmented; @gholami2019anode; @zhang2019anodev2]. There has also been recent interest in extending these frameworks to a stochastic scenario, that is, training probabilistic models derived from *stochastic* differential equations (SDEs). For physical models, where the evolution of a dynamical system is no longer deterministic, or microscopic fluctuations are dependent on components changing too rapidly to quantify, an SDE can be more appropriate. Stochastic extensions of neural ODEs have been considered in [@tzen2019neural; @liu2019neural; @jia2019neural; @peluchetti2019infinitely] as limits of deep latent Gaussian models, where they have been suggested to show increased robustness to noisy / adversarial data. Furthermore, unlike deterministic flows, there is a foolproof recipe for constructing a family of SDEs that are ergodic with respect to some target distribution [@ma2015complete]. This particular property guarantees the convergence of the solution of an SDE to a prescribed target distribution. Such SDEs are prime candidates for the construction of stochastic MCMC algorithms, by generating sample paths via approximate stochastic integration methods. However, developing an analogue of the continuous normalizing flows framework for flows constructed from SDEs—in particular, one that comes with simple and rigorous mathematical theory and that does not rely on ad hoc or problem-specific assumptions—is far from trivial. A common approach for conducting VI with SDEs is to rely on Girsanov’s theorem. This allows one to estimate the Kullback-Leibler divergence between densities of solutions to two SDEs (for the prior and posterior distributions) with differing drift coefficients [@beskos2006exact; @tzen2019neural]. Following this approach, @li2020scalable developed a stochastic adjoint method which scales well to high dimensions, and enables SDEs as latent models in variational autoencoders. Theoretical justification of the method proved challenging, as stochastic calculus is ill-suited for analyzing backward (approximate) solutions to SDEs. Notable deficiencies with these previous approaches include difficulties with non-diagonal diffusion, incompatibility with higher-order adaptive SDE solvers, and a complex means of reconstructing Brownian motion paths from random number generator seeds. Furthermore, the method cannot be justifiably combined with existing approaches of density estimation for SDEs (see @hurn2007seeing). On the other hand, recent efforts have made significant strides in applying variational and MCMC methods for idealized Bayesian computation. One of the most significant contributions in this direction is @salimans2015markov, who performed VI with respect to distributions formed from steps of a reversible Markov chain. For example, the setting of Hamiltonian Monte Carlo was examined in @wolf2016variational. More recently, @liu2016two considered optimizing step size in stochastic gradient Langevin dynamics using methods derived from kernelized Stein discrepancy. Langevin flows [@rezende2015variational] have been discussed as a potential VI framework that takes inspiration from the SDEs underlying stochastic MCMC [@ma2015complete]. Once again, implementation of Langevin flows relies on the approximation of the log-likelihood for a general class of SDEs. Contributions {#contributions .unnumbered} ------------- We provide a general theoretical framework (which we refer to as *stochastic normalizing flows*) for approximating generative models constructed from SDEs using continuous normalizing flows. These approximations can then be trained using existing techniques. By this process, we find that theoretical and practical developments concerning continuous normalizing flows and neural ODEs extend readily to the stochastic setting, without the need of an independent framework. The key theoretical enabler underlying our strong results and simple analysis is the *theory of rough paths* [@friz2014course], an alternative stochastic calculus that enables approximation and pathwise treatment of SDEs. Our approach 1. enables (i) density estimation, (ii) maximum likelihood estimation, and (iii) variational approximations beyond autoencoders, for *arbitrary* SDE models; and 2. is easily implemented using any general continuous normalizing flows implementation, such as that of @grathwohl2018ffjord. Our framework recovers the stochastic adjoint method of @li2020scalable, but our approach is sufficiently flexible to overcome its deficiencies. Moreover, using our approach, any existing neural ODE framework (such as @zhang2019anodev2) can be extended to SDEs, simply by the addition of a few extra terms. Following a review of background material in §\[sec:Review\], the stochastic normalizing flows framework is introduced and discussed in §\[sec:Stochastic\], with our main approximation result presented in Theorem \[thm:RoughNormFlow\]. Some numerical investigations are conducted in §\[sec:Numerics\], including an application to hyperparameter optimization in stochastic MCMC. Background Review {#sec:Review} ================= Continuous Normalizing Flows ---------------------------- We shall begin by reviewing the continuous normalizing flow framework for training ODE models, as our development of random and stochastic normalizing flows will build upon it. Consider a parameterized class of models $\{Z_\theta\}_{\theta \in \mathbb{R}^m}$ of the following form: for $f:\mathbb{R}^d \times [0,T] \times \mathbb{R}^m \to \mathbb{R}^d$, let $Z = Z_\theta \in \mathbb{R}^d$ satisfy the ODE with random initial condition (often called a *random ordinary differential equation*) $$\label{eq:ODEModel} \frac{{\mathrm{d}}}{{\mathrm{d}}t} Z(t) = f(Z(t), t, \theta),\quad Z(0) \sim p_0(\theta).$$ In a general machine learning context, one might choose $f$ such that the Euler discretization of (\[eq:ODEModel\]) resembles layer-wise updates of a residual neural network [@lu2017beyond; @chen2018neural], or one may parameterize $f$ as a neural network itself [@grathwohl2018ffjord]. The resulting ODEs constitute the class of so-called *neural ordinary differential equations*. The following theorem is a consequence of the Liouville equation (equivalently, Fokker-Planck equation) applied to the solution $Z(t)$ of the random ODE (\[eq:ODEModel\]), and it yields an ODE for the log density of $Z(t)$ evaluated at $Z(t)$. \[thm:ContNormFlow\] Suppose that $Z(t)$ satisfies (\[eq:ODEModel\]). The distribution of $Z(t)$ is absolutely continuous with respect to Lebesgue measure, with probability density $p_t$ satisfying $$\label{eq:ContNormFlow} \frac{{\mathrm{d}}}{{\mathrm{d}}t} \log p_t(Z(t)) = -\nabla_z \cdot f(Z(t), t, \theta) $$ Naively computing the divergence in (\[eq:ContNormFlow\]) with automatic differentiation is of quadratic complexity in the dimension $d$. As pointed out by @grathwohl2018ffjord, this can be improved to linear complexity using a trace estimator [@roosta2015improved]: $$\label{eq:TraceEst} \nabla_z \cdot f(z) = {\mathrm{tr}}\left(\frac{\partial f}{\partial z}\right) \approx \frac1n \sum_{k=1}^{n} \epsilon_k^\top \frac{\partial f}{\partial z} \epsilon_k,$$ where each $\epsilon_k$ is an independent and identically distributed copy of a random vector $\epsilon \in \mathbb{R}^d$ with zero mean and $\mathbb{E}[\epsilon \epsilon^\top] = I$. Common choices for $\epsilon_k$ include standard normal and Rademacher random vectors. The Adjoint Method ------------------ Training continuous normalizing flows often involves minimizing a scalar loss function involving $Z$ and/or the log-density computed via Theorem \[thm:ContNormFlow\] with respect to the parameters $\theta$. For this, we require gradients of $Z(t)$ with respect to $\theta$ for $t \in [0,T]$. The most obvious approach is to directly backpropagate through a numerical integration scheme such as in @ryder2018black, but this does not scale well in $T$. The superior alternative is the *adjoint method*, which computes derivatives of a scalar loss function by solving another differential equation in reverse time. Letting $L$ denote a scalar loss depending on $Z(T)$, the *adjoint* given by $a(t) = \frac{\partial L}{\partial Z(t)}$, as well as the gradient of $L$ in $\theta$, satisfy [@pontryagin2018mathematical §12] \[eq:AdjointMethod\] $$\begin{aligned} \frac{{\mathrm{d}}}{{\mathrm{d}}t} a(t) &= -\nabla_z f(Z(t),t,\theta) a(t), \\ \nabla_\theta L &= \int_0^T \nabla_\theta f(Z(t), t, \theta)a(t) {\mathrm{d}}t.\end{aligned}$$ Together with (\[eq:ODEModel\]), the equations (\[eq:AdjointMethod\]) are solved in reverse time, starting from the terminal values $Z(T)$ and $\nabla L(Z(T))$. By augmenting $Z(t)$ together with (\[eq:ContNormFlow\]), this method also allows for loss functions depending on $p_T(Z(T))$. Solving (\[eq:ODEModel\]), (\[eq:ContNormFlow\]), and (\[eq:AdjointMethod\]) can be achieved using off-the-shelf numerical integrators. Adaptive solvers prove particularly effective, although, as pointed out in @gholami2019anode, the backward solve (\[eq:AdjointMethod\]) can often run into stability issues, suggesting a Rosenbrock or other implicit approach [@hairer1996solving]. We point out that the same is also true in the stochastic setting; see @hodgkinson2019implicit, for example. For further implementation details concerning continuous normalizing flows, we refer to @grathwohl2018ffjord. Rough Path Theory ----------------- The theory of rough paths was first introduced in [@lyons1998differential] to provide a supporting pathwise theory for SDEs. It has since flourished into a coherent pathwise alternative to stochastic calculus, facilitating direct stochastic generalizations of results from the theory of ODEs — we refer to @friz2014course for a gentle introduction, and @friz2010multidimensional for a thorough treatment of the topic. Suppose that we would like to prescribe meaning to the infinitesimal limit of the sequence of iterates $$\label{eq:RDELimit} Z_{t+h} = Z_t + f(Z_t)(X_{t+h} - X_t),\quad \mbox{as }h \to 0^+.$$ In the case of SDEs, $X_t$ is a sample path of Brownian motion, so that each $X_{t+h} - X_t$ is a realization of a normal random vector with zero mean and covariance $h I$. Unfortunately, a strong limit of (\[eq:RDELimit\]) fails to exist if $X_t$ is too “rough”. In particular, suppose that $X_t$ is $\alpha$-Hölder continuous for $\alpha \in (0,1)$, that is, there exists some $C > 0$ such that $\|X_s - X_t\| \leq C|s - t|^{\alpha}$ for any $s,t \geq 0$. Since the limit (\[eq:RDELimit\]) is only well-defined if $\alpha \geq 1/2$ [@young1936inequality], a function on $[0,T]$ is *rough* if it is Hölder-continuous only for $\alpha < 1/2$. Sample paths of Brownian motion constitute rough paths under this definition. The problem is that the discretization (\[eq:RDELimit\]) invokes the *zeroth-order* approximation $f(Z_{t+s}) \approx f(Z_t)$ for $0 \leq s \leq h$, which proves too poor. By instead taking a *first-order* approximation $$\begin{aligned} f(Z_{t+s}) &\approx f(Z_t) + \nabla_z f(Z_t)(Z_{t+s}-Z_t) \\&\approx f(Z_t) + \nabla_z f(Z_t)f(Z_t)(X_{t+s}-X_t),\end{aligned}$$ we arrive at the *Davie scheme* [@davie2008differential] $$\label{eq:DavieApprox} Z_{t+h} = Z_t + f(Z_t) (X_{t+h} - X_t) + \nabla_z f(Z_t)f(Z_t)\mathbb{X}_{t,t+h},$$ where $\mathbb{X}_{s,t}$ represents the “integral” $\int_s^t X_r {\mathrm{d}}X_r^\top$. Once again, we cannot uniquely define $\mathbb{X}$ from the path $X$ itself, so instead we prescribe it. In fact, each choice of $\mathbb{X}$ satisfying Chen’s relations $$\mathbb{X}_{s,t} - \mathbb{X}_{s,u} - \mathbb{X}_{u,t} = (X_s - X_u)(X_t - X_u)^\top,$$ for any $s,u,t \geq 0$, will reveal a *different* limit for (\[eq:DavieApprox\]) as $h \to 0^+$, provided $\alpha \geq 1/3$ (for smaller $\alpha$, higher-order approximations are necessary). The pair $\boldsymbol{X} = (X,\mathbb{X})$ is referred to as a *rough path*, and the limit of (\[eq:DavieApprox\]) as $h \to 0^+$ is the solution to the *rough differential equation* (RDE) $$\label{eq:RoughDiffEq} {\mathrm{d}}\boldsymbol{Z}_t = f(Z_t) {\mathrm{d}}\boldsymbol{X}_t.$$ Hölder continuity is critical to rough path theory — in the sequel, we equip the space of $\alpha$-Hölder functions with the $\alpha$-Hölder norm, defined by $$\|X\|_\alpha \coloneqq \sup_{t \in [0,T]}\|X_t\| + \sup_{\substack{s,t\in [0,T]\\s \neq t}}\frac{\|X_t - X_s\|}{|t - s|^{\alpha}}.$$ This definition extends to the iterated integral $\mathbb{X}$ by replacing $X_t$ and $X_t - X_s$ with $\mathbb{X}_{0,t}$ and $\mathbb{X}_{s,t}$, respectively. It is useful to identify a calculus which satisfies the usual chain and product rules. This occurs when the rough path $\boldsymbol{X}$ is *geometric*, that is, $$\label{eq:GeoRoughPath} \mathbb{X}_{s,t} - \mathbb{X}_{t,s} = \tfrac12 (X_t - X_s) (X_t - X_s)^\top,\quad \forall s,t \geq 0.$$ Every continuous and piecewise differentiable function $X$ is canonically lifted to a geometric rough path by taking $\mathbb{X}_{s,t} = \int_s^t X_r \frac{{\mathrm{d}}}{{\mathrm{d}}r} X_r^\top {\mathrm{d}}r$, where the derivative is interpreted in the weak sense. In these cases, (\[eq:RoughDiffEq\]) equates to the ODE $\frac{{\mathrm{d}}}{{\mathrm{d}}t}Z_t = f(Z_t) \frac{{\mathrm{d}}}{{\mathrm{d}}t} X_t$. Geometric rough paths have two key properties of interest: 1. The canonical lifts of any sequence of smooth approximations $X^{(n)}$ which converge to $X$ as $n \to \infty$ in the $\alpha$-Hölder norm, also converge in the $\alpha$-Hölder rough path metric $$\varrho_{\alpha}((X,\mathbb{X}),(Y,\mathbb{Y})) = \|X - Y\|_{\alpha} + \|\mathbb{X} - \mathbb{Y}\|_{2\alpha},$$ to a geometric rough path $(X,\mathbb{X})$. Conversely, any geometric rough path can be approximated by some sequence of smooth paths [@friz2014course Proposition 2.5]. 2. The reverse-time process $\tilde{Z}_t = Z_{T-t}$ of a solution $Z_t$ to any rough differential equation (\[eq:RoughDiffEq\]) with Lipschitz $f$, itself satisfies the reversed rough differential equation ${\mathrm{d}}\tilde{\bm{Z}}_t = -f(T-t,\tilde{Z}_t){\mathrm{d}}\bm{X}_{T-t}$ *if and only if* $\bm{X}$ is geometric. By property I, any solution to RDEs driven by a geometric rough path can be approximated by solutions to ODEs. Property II, which follows readily from the definition (\[eq:GeoRoughPath\]) in the limit (\[eq:DavieApprox\]), enables the adjoint method for rough differential equations driven by a geometric rough path. Stochastic Normalizing Flows {#sec:Stochastic} ============================ Let $Z_t$ satisfy the Itô SDE $$\label{eq:MainSDE} {\mathrm{d}}Z_t = \mu_t(Z_t, \theta) {\mathrm{d}}t + \sigma_t(Z_t, \theta) {\mathrm{d}}B_t, \quad Z_0 \sim p_0(\theta),$$ where $B_t$ is an $m$-dimensional Brownian motion, and $\mu_t:\mathbb{R}^d \to \mathbb{R}^d$, $\sigma_t:\mathbb{R}^d \to \mathbb{R}^{d \times m}$ are the drift, and diffusion coefficients, respectively. Analogous to neural ODEs, neural SDEs choose $\mu_t$ to resemble a single layer of a neural network [@tzen2019neural]. The dropout-inspired construction of @liu2019neural suggests taking $\sigma_t \propto \mathrm{diag}(\mu_t)$. Alternatively, one can parameterize both $\mu_t$ and $\sigma_t$ by multi-layer neural networks. The reliance of stochastic calculus on non-anticipating processes as well as the lack of continuity for solution maps of Itô SDEs necessitates complicated and delicate arguments for extending each piece of the continuous normalizing flow framework from §\[sec:Review\] to SDEs. We bypass the intricacies of existing theoretical treatments of neural SDEs by an approximation argument: for a smooth approximation $\tilde{B}_t$ of Brownian motion $B_t$, we estimate solutions of an SDE by a random ODE involving $\tilde{B}_t$. One must take great care with such approximations. For example, geometric Brownian motion, that is, the solution to ${\mathrm{d}}Z_t = \sigma Z_t {\mathrm{d}}B_t$, has the explicit expression $Z_t = Z_0 \exp(-\frac{\sigma^2}{2} t + \sigma B_t)$, which is not well-approximated by the solution $\tilde{Z}_t = Z_0 \exp(\sigma \tilde{B}_t)$ to $\frac{{\mathrm{d}}}{{\mathrm{d}}t} \tilde{Z}_t = \sigma \tilde{Z}_t \frac{{\mathrm{d}}\tilde{B}_t}{{\mathrm{d}}t}$. Theoretical verification of this approach is challenging using traditional stochastic calculus due to the irregularity of solution maps. Instead, we rely on rough path theory — particularly properties I and II of geometric rough paths. In the rough path framework, one can reconstruct the Itô stochastic calculus via the rough path $\bm{B}^{{\text{It\^{o}}}} = (B, \mathbb{B}^{{\text{It\^{o}}}})$, where $\mathbb{B}^{{\text{It\^{o}}}}_{s,t} = B_t (B_t - B_s)^\top - \frac{t-s}2 I$. Indeed, by @friz2014course [Theorem 9.1], letting $\bm{B}^{{\text{It\^{o}}}}(\omega)$ denote a realization of the Itô Brownian motion rough path, the solution to the rough differential equation $$\label{eq:ItoRDE} {\mathrm{d}}\bm{Z}_t = \mu_t(Z_t, \theta) {\mathrm{d}}t + \sigma_t(Z_t, \theta) {\mathrm{d}}\bm{B}^{{\text{It\^{o}}}}_t(\omega)$$ is a realization of the strong solution to (\[eq:MainSDE\]). Likewise, the Davie scheme (\[eq:DavieApprox\]) corresponds to the Milstein integrator for SDEs [@kloeden2013numerical §10.3]. Unfortunately, $\bm{B}^{{\text{It\^{o}}}}(\omega)$ is not a geometric rough path, and so Theorem \[thm:RoughNormFlow\] cannot be directly applied. Instead, we shall proceed according to the following steps: 1. Convert the Itô SDE to a Stratonovich SDE (§\[sec:Strat\]). 2. Interpret the Stratonovich SDE pathwise as an RDE driven by a geometric rough path $\bm{B}^{{\text{Strat}}}$ (\[eq:StratRDE\]). 3. Approximate the pathwise Stratonovich RDE by a random ODE (§\[sec:WongZakai\]). 4. Train the random ODE as a continuous normalizing flow with added latent variables (§\[sec:RandCNF\]). Consequently, the RDE (\[eq:ItoRDE\]) is estimated by the ODE $\frac{{\mathrm{d}}Z_t(\omega)}{{\mathrm{d}}t} = F_\omega(Z_t(\omega),t,\theta)$ where $$F_\omega(z, t, \theta) = \underset{\text{Stratonovich drift}}{\underbrace{\tilde{\mu}_t(z, \theta)}} + \;\;\sigma_t(z,\theta)\underset{\text{approximation}}{\underbrace{\frac{{\mathrm{d}}B_t(\omega)}{{\mathrm{d}}t}}}.$$ Stratonovich calculus {#sec:Strat} --------------------- The unique geometric rough path formed from Brownian motion yields the Stratonovich calculus: $\mathbb{B}^{{\text{Strat}}}_{s,t} = B_t(B_t - B_s)^\top$. A Stratonovich differential equation is commonly written in the form ${\mathrm{d}}Z_t = \mu_t(Z_t) {\mathrm{d}}t + \sigma_t(Z_t) \circ {\mathrm{d}}B_t$, where $\circ$ denotes Stratonovich integration: for a process $Y_t$ adapted to the filtration generated by $B_t$, $$\int_s^t Y_t \circ {\mathrm{d}}B_t = \lim_{|\mathcal{P}|\to 0}\sum_{k=1}^N \frac12(Y_{t_k} + Y_{t_{k-1}}) (B_{t_k} - B_{t_{k-1}}),$$ where $\mathcal{P} = \{0 = t_0 < \cdots < t_N = T\}$ is a partition with mesh size $|\mathcal{P}| = \max_k |t_k - t_{k-1}|$, and the limit is in $L^2$. This is to be compared with Itô integration which is defined instead by $$\int_s^t Y_t \,{\mathrm{d}}B_t = \lim_{|\mathcal{P}|\to 0}\sum_{k=1}^N Y_{t_{k-1}} (B_{t_k} - B_{t_{k-1}}).$$ Stratonovich differential equations were recognized in @li2020scalable to be the correct setting for extending the adjoint method to SDEs. However, the adherence to classical stochastic calculus, which relies on adaptedness, somewhat complicates the argument. In our setting, the advantages of Stratonovich differential equations are clear. Because Stratonovich differential equations can be arbitrarily well-approximated by random ODEs, *all* methods of training continuous normalizing flows extend to them, including the adjoint method. Any Itô SDE can be converted into a Stratonovich SDE by adjusting the drift [@evans2012introduction p. 123], a fact readily seen by comparing limits of (\[eq:DavieApprox\]) with $\mathbb{B}^{{\text{It\^{o}}}}$ and $\mathbb{B}^{{\text{Strat}}}$. The following formula is particularly amenable to implementation with automatic differentiation: the Itô SDE ${\mathrm{d}}Z_t = \mu_t(Z_t) {\mathrm{d}}t + \sigma_t(Z_t) {\mathrm{d}}B_t$ is equivalent to the Stratonovich SDE ${\mathrm{d}}Z_t = \tilde{\mu}_t(Z_t) {\mathrm{d}}t + \sigma_t(Z_t) \circ {\mathrm{d}}B_t$ provided that for each $i=1,\dots,d$, $$\label{eq:ItoCorrection} \tilde{\mu}_t^i(x) = \mu_t^i(x) - \tfrac12\nabla_x \cdot (\sigma_t(x)\sigma_t^\top(x^{\ast}))_i,$$ where $x^\ast$ is an independent copy of $x$, and the subscript denotes the $i$-th row. Once again, we can make use of the trace estimator (\[eq:TraceEst\]) to increase performance in higher dimensions. In the rough path theory, Stratonovich SDEs are interpreted pathwise according to the RDE $$\label{eq:StratRDE} {\mathrm{d}}\bm{Z}_t = \tilde{\mu}_t(Z_t, \theta) {\mathrm{d}}t + \sigma_t(Z_t,\theta) {\mathrm{d}}\bm{B}_t^{{\text{Strat}}}(\omega),$$ which is equivalent to (\[eq:ItoRDE\]). Wong–Zakai approximations {#sec:WongZakai} ------------------------- A random ODE $\frac{{\mathrm{d}}}{{\mathrm{d}}t} Z_t^{(n)} = \mu_t(Z_t^{(n)}) + \sigma_t(Z_t^{(n)}) \frac{{\mathrm{d}}B_t^{(n)}}{{\mathrm{d}}t}$ estimating a Stratonovich SDE ${\mathrm{d}}Z_t = \mu_t(Z_t) {\mathrm{d}}t + \sigma_t(Z_t) \circ {\mathrm{d}}B_t$ is commonly referred to as a *Wong–Zakai approximation* [@twardowska1996wong], after the authors of the seminal paper [@wong1965convergence], who first illustrated this concept for one-dimensional Brownian motion. We shall consider two types of Wong–Zakai approximation: a Karhunen-Loève expansion, and a piecewise linear function. These approximations are compared in Figure \[fig:WienerApprox\]. In practice, we have found that the Karhunen-Loève expansion with $4 \leq n \leq 10$ terms works well for training, while the piecewise linear approximation is preferable for testing. ![Karhunen-Loeve (left) and piecewise linear (right) approximations of a Brownian motion sample path with $n = 6$ and $\Delta t = \frac{1}{6}$ respectively.[]{data-label="fig:WienerApprox"}](wiener_approximations.pdf){width="\columnwidth"} ### Piecewise linear Easily the most common approximation of Brownian motion involves exact simulation on a discrete set of times $\{0,t_1,t_2,\dots,t_n\}$, followed by linear interpolation. More precisely, letting $\Delta t_k = t_{k+1}-t_k$, for each $k=0,\dots,n-1$, we let $$B_{t_{k+1}}^{(n)} = B_{t_k}^{(n)} + \sqrt{\Delta t_k}\omega_k,\quad \omega_k \sim \mathcal{N}(0,1),$$ and consider the approximation $$B_t^{(n)} = B_{t_k}^{(n)} + \frac{t - t_k}{t_{k+1}-t_k} (B_{t_{k+1}}^{(n)} - B_{t_k}^{(n)}),\quad t \in [t_k,t_{k+1}].$$ Integrating the resulting Wong–Zakai approximation using Euler’s method on the same set of time points is equivalent to performing the Euler–Maruyama method for solving the Stratonovich SDE. By @friz2010multidimensional [Theorem 15.45], as the mesh size $\max_k \Delta t_k \to 0$, the piecewise linear approximation converges almost surely to Brownian motion in the $\alpha$-Hölder norm for any $\alpha < 1/2$. ### Karhunen-Loève expansion For any zero-mean Gaussian process $X_t$ on $\mathbb{R}^d$ with $t\in [0,T]$, the covariance function $K(s,t) = \mathbb{E}[X_s X_t^\top]$ is a positive-definite kernel. If $K$ is also continuous, Mercer’s theorem guarantees the existence of an orthonormal basis on $L^2([0,T],\mathbb{R}^d)$ of eigenfunctions $\{e_k\}_{k=1}^{\infty}$ with corresponding positive eigenvalues $\{\lambda_k\}_{k=1}^{\infty}$ such that $K(s,t)=\sum_{j=1}^{\infty} \lambda_j e_j(s) e_j(t)$. The process $X_t$ can be expanded in terms of these eigenfunctions as $$X_t = \sum_{k=1}^{\infty} \sqrt{\lambda_k} \omega_k e_k(t), \quad \omega_k \sim \mathcal{N}(0,1),$$ where each $\omega_k$ is independent. This is called the *Karhunen-Loève expansion* of $X$. Truncating the series after $n$ terms yields the $n$-th order Karhunen-Loève approximation, and has the smallest mean squared error over all expansions with $n$ orthogonal basis elements. Recalling that we are primarily interested in the endpoints of the solution, instead of expanding Brownian motion itself, we consider an approximation $B_t^{(n)}$ derived from the Karhunen-Loève expansion of the Brownian bridge $B_t - B_T \frac{t}{T}$: $$B_t^{(n)} = \omega_0 \frac{t}{\sqrt{T}} + \sum_{k=1}^{n - 1} \omega_k \frac{\sqrt{2T} \sin(k\pi t / T)}{k\pi},\quad n = 1,2,\dots.$$ Using this approximation ensures that the terminal density for SDEs with constant drift and diffusion coefficients is computed exactly. By @friz2010multidimensional [Theorem 15.51], $(B_t^{(n)})_{t \in [0,T]}$ converges almost surely as $n \to \infty$ to Brownian motion in the $\alpha$-Hölder norm for any $\alpha < 1/2$. Furthermore, since $B_t^{(n)}$ is smooth, Wong–Zakai approximations involving $\frac{{\mathrm{d}}B_t^{(n)}}{{\mathrm{d}}t}$ may be readily solved using adaptive ODE solvers. Main result ----------- Using Wong–Zakai approximations, a Stratonovich SDE can be uniformly approximated in Hölder norm by random ODEs. In Theorem \[thm:RoughNormFlow\], we show that the log-densities and loss function gradients for these random ODEs also converge appropriately. More generally, geometric rough paths (including the Stratonovich paths (\[eq:StratRDE\])) with random initial conditions can be approximately trained as random ODEs. \[thm:RoughNormFlow\] Let $\bm{X} = (X, \mathbb{X})$ be an $\alpha$-Hölder geometric rough path, and $\{X^{(n)}\}_{n=1}^{\infty}$ a sequence of piecewise differentiable functions on $[0,T]$ that approximate $X$ under the $\beta$-Hölder norm for $\beta \in (\frac13, \frac12)$, that is, $\|X^{(n)} - X\|_\beta \to 0$ as $n \to \infty$. Let $\bm{Z},Z^1,Z^2,\dots$ be solutions to the differential equations \[eq:RoughFlow\] $$\begin{aligned} {\mathrm{d}}\bm Z_t &= f(Z_t, t, \theta) {\mathrm{d}}\bm{X}_t, &&& Z_0 &\sim p_0,\\ \frac{{\mathrm{d}}Z_t^{(n)}}{{\mathrm{d}}t} &= f(Z_t^{(n)}, t, \theta) \frac{{\mathrm{d}}X_t^{(n)}}{{\mathrm{d}}t} &&& Z_0^{(n)} &= Z_0,\end{aligned}$$ where $f \in \mathcal{C}_b^4(\mathbb{R}^d \times [0,T] \times \mathbb{R}^m)$ and $p_0$ is a density on $\mathbb{R}^d$ such that $\log p_0$ is continuous. Let $p_t^{(n)}$ denote the probability density of $Z_t^{(n)}$ at time $t$, given by (\[eq:ContNormFlow\]). The distribution of $Z_t$ is absolutely continuous with respect to Lebesgue measure with corresponding continuous density $p_t$ satisfying: 1. For any $x \in \mathbb{R}^d$, $\displaystyle{\sup_{t \in [0,T]}}|\log p_t^{(n)}(x) - \log p_t(x)| \to 0$ as $n \to \infty$. 2. The path $t \mapsto \log p_t(Z_t)$ is the unique solution to the rough differential equation $$\label{eq:RoughNF} {\mathrm{d}}\log p_t(Z_t) = -\nabla_z \cdot (f(Z_t, t, \theta) {\mathrm{d}}\bm X_t). $$ 3. For any smooth loss function $L:\mathbb{R}^{d+1}\to\mathbb{R}$ and $t \geq 0$, as $n \to \infty$, $$\label{eq:ApproxAdjoint} \nabla_\theta L(Z_t^{(n)},\log p_t^{(n)}(Z_t^{(n)})) \to \nabla_\theta L(Z_t,\log p_t(Z_t)).$$ Recall that each $X^{(n)}$ can be lifted canonically to a rough path $\bm{X}^{(n)}$ such that $\rho_{\beta}(\bm{X}^{(n)},\bm{X}) \to 0$ as $n \to \infty$. For an arbitrary rough path $\bm{Y}$, we let $\Phi_t(\boldsymbol{Y}, \xi)$ and $\Psi_t(\bm{Y},\ell)$ denote the solution maps for the rough differential equations ${\mathrm{d}}\bm{Z}_t = f(Z_t,t,\theta){\mathrm{d}}\bm{Y}_t$, $Z_0 = \xi$ and ${\mathrm{d}}\bm{L}_t = -\nabla_z\cdot f(Z_t,t,\theta) {\mathrm{d}}\bm{Y}_t$, $L_0 = \ell$, respectively. By @friz2014course [Theorem 8.10], $\Phi_t(\boldsymbol{Y},\cdot)$ is a $\mathcal{C}^1$-diffeomorphism, and hence, for $Z_0 \sim p_0(\theta)$ and any $t \in [0,T]$, $Z_t = \Phi_t(\boldsymbol{X}, Z_0)$ is an absolutely continuous random variable, whose corresponding density we denote by $p_t$. In fact, denoting by $\Phi_{-t}(\boldsymbol{Y},\cdot)$ the inverse of $\Phi_t(\bm{Y},\cdot)$, $$\begin{aligned} \label{eq:DensityNCoM} p_t^{(n)}(x) &= p_0(\Phi_{-t}(\bm{X}^{(n)},x))\left|\det \frac{\partial \Phi_{-t}(\bm{X}^{(n)},x)}{\partial x}\right|,\\ \label{eq:DensityCoM} p_t(x) &= p_0(\Phi_{-t}(\bm{X},x))\left|\det \frac{\partial \Phi_{-t}(\bm{X},x)}{\partial x}\right|,\end{aligned}$$ and so both $p_t^{(n)}$ and $p_t$ are continuous. Furthermore, by @friz2014course [Theorem 8.5], for any $\frac13 < \gamma < \beta$, there exist constants $C_\gamma^{\Phi}$ and $C_\gamma^{\Psi}$ such that for any $\beta$-Hölder continuous rough paths $\boldsymbol{X}$, $\boldsymbol{Y}$ and $\xi, \tilde{\xi} \in \mathbb{R}^d$, $\ell, \tilde{\ell} \in \mathbb{R}_+$, $$\begin{aligned} \label{eq:PhiRegularity} \|\Phi(\boldsymbol{X}, \xi) - \Phi(\boldsymbol{Y}, \tilde{\xi})\|_{\gamma} &\leq C_\gamma^\Phi (\|\xi - \tilde{\xi}\| + \varrho_{\beta}(\boldsymbol{X}, \boldsymbol{Y}))\\ \label{eq:PsiRegularity} \|\Psi(\bm{X}, \ell) - \Psi(\bm{Y}, \tilde{\ell})\|_{\gamma} &\leq C_\gamma^\Psi (|\ell - \tilde{\ell}| + \varrho_{\beta}(\bm{X},\bm{Y})).\end{aligned}$$ We deduce the following for any $t \in [0,T]$ and $x \in \mathbb{R}^d$: **(i)** $\|Z^{(n)} - Z\|_\gamma \to 0$ by (\[eq:PhiRegularity\]); **(ii)** using (i) and continuity of $p_t$, $\log p_t(Z_t^{(n)}) \to \log p_t(Z_t)$; **(iii)** as a consequence of (\[eq:DensityNCoM\]), (\[eq:DensityCoM\]), (\[eq:PhiRegularity\]), and @friz2014course [Theorem 8.10], $p_t^{(n)}(x)\to p_t(x)$; **(iv)** combining (ii) and (iii), $\log p_t^{(n)}(Z_t^{(n)}) \to \log p_t(Z_t)$. Since $\Psi_t(\bm{X}^{(n)},\log p_0(Z_0)) = \log p_t^{(n)}(Z_t^{(n)})$ by Theorem \[thm:ContNormFlow\], (iv) and (\[eq:PsiRegularity\]) imply $\log p_t(Z_t) = \Psi(\bm{X}, \log p_0(Z_0))$ and hence (\[eq:RoughNF\]). Let $x \in \mathbb{R}^d$ be arbitrary. To show that $\log p_t^{(n)}(x)$ converges uniformly in $t$, observe that $$\log p_t(x) = \Psi(\bm{X}, \log p_0(\Phi_{-t}(\bm{X},x))),$$ and similarly for $\log p_t^{(n)}(x)$. Together with property II of geometric rough paths, inequality (\[eq:PhiRegularity\]) with $\bm{Y} \equiv \bm{0}$ reveals that $\Phi_{-t}(\bm{X}^{(n)},x)$ and $\Phi_{-t}(\bm{X},x)$ are uniformly bounded in $t \in [0,T]$. Since $\log p_0$ is continuous, $\log p_0(\Phi_{-t}(\bm{X}^{(n)},x))$ converges to $\log p_0(\Phi_{-t}(\bm{X},x))$ uniformly in $t \in [0,T]$. Applying (\[eq:PsiRegularity\]), $$\begin{gathered} \sup_{t\in[0,T]}|\log p_t^{(n)}(x) - \log p_t(x)| \leq C_{\gamma}^\Psi(\varrho_\beta(\bm{X}^{(n)}, \bm{X}) + |\log p_0(\Phi_{-t}(\bm{X}^{(n)}, x)) - \log p_0(\Phi_{-t}(\bm{X}, x))| \to 0.\end{gathered}$$ Finally, to prove (\[eq:ApproxAdjoint\]), by @friz2014course [Proposition 5.6], we can write $(\theta,L)$ as the solution to the rough differential equation \[eq:LossSDE\] $$\begin{aligned} &{\mathrm{d}}\theta = 0 \\ &{\mathrm{d}}L(Z_t,\log p_t(Z_t)) = \nabla_z L(Z_t, \log p_t(Z_t)) \cdot {\mathrm{d}}Z_t + \nabla_\ell L(Z_t, \log p_t(Z_t)) {\mathrm{d}}\log p_t(Z_t)\end{aligned}$$ and similarly for $Z_t^{(n)}$ and $\log p_t^{(n)}(Z_t^{(n)})$, where $\nabla_z$ and $\nabla_\ell$ denote the gradients with respect to $Z_t$ and $\log p_t(Z_t)$, respectively. The derivative of $L$ with respect to $\theta$ is a derivative of (\[eq:LossSDE\]) with respect to its initial condition, and hence (\[eq:ApproxAdjoint\]) follows from @friz2014course [Theorem 8.10]. Random continuous normalizing flows {#sec:RandCNF} ----------------------------------- By a conditioning argument, any random ODE, such as a Wong-Zakai approximation, may be treated as a continuous normalizing flow. Let $Z_t$ be the solution to a random ODE of the form $$\label{eq:RandomODE} \frac{{\mathrm{d}}}{{\mathrm{d}}t}Z_t = f(Z_t, \omega, t, \theta),$$ where $\omega = (\omega_1,\dots,\omega_n) \sim q(\omega)$ is a random vector independent of $Z_t$, $t$, and $\theta$. The reduction of a random ODE to this form is in keeping with the reparameterization trick [@xu2019variance]. In particular, for the piecewise linear and Karhunen-Loève approximations, each $\omega_i \sim \mathcal{N}(0,1)$. After conditioning on $\omega$, Theorem \[thm:ContNormFlow\] applied to (\[eq:RandomODE\]) provides a means of computing $\log p_t(Z_t\vert\omega)$, after sampling $Z_0 \sim p_0$. The density $p_t(Z_t)$ can be computed using a naive Monte Carlo estimator $$\label{eq:DensityEstNaive} p_t^\theta(Z_t) = \int p_t^\theta(Z_t \vert \omega)q(\omega) {\mathrm{d}}\omega \approx \frac1N \sum_{i=1}^{n} p_t^\theta(Z_t \vert \omega_i),$$ where the dependence on $\theta$ has been made explicit, and can be optimized over using the adjoint method. Analogously to @chen2018neural [@grathwohl2018ffjord], the density of data ${\boldsymbol{x}}$ may be estimated along a single sample path $B_t(\omega)$ (denoted $p({\boldsymbol{x}}\vert \omega)$) in the following way: letting $\Delta \log p_t^\omega = \log p_t(Z_t\vert \omega) - \log p({\boldsymbol{x}} \vert \omega)$, we see that $\Delta \log p_t^\omega$ also satisfies (\[eq:ContNormFlow\]). By solving (\[eq:RandomODE\]) and the corresponding (\[eq:ContNormFlow\]) in reverse time from the initial conditions $Z_T = {\boldsymbol{x}}$ and $\Delta \log p_T^\omega = 0$, we obtain $Z_0$ and $\Delta \log p_0^\omega$, and compute $\log p({\boldsymbol{x}} \vert \omega) = \log p_0(Z_0) - \Delta \log p_0^\omega$. This is shown in Algorithm \[alg:DensityEst\], which depends on an ODE solver <span style="font-variant:small-caps;">odesolve</span>, and yields a density estimation procedure for stochastic normalizing flows when paired with (\[eq:DensityEstNaive\]). Note that by comparison to @grathwohl2018ffjord [Algorithm 1] which encompasses steps 6–12 of our Algorithm \[alg:DensityEst\], we see much of the density estimation procedure can be accomplished using an existing continuous normalizing flow implementation. In variational settings where $\log p_T(Z_T)$ is required, the same procedure applies, where $\mathbf{x}$ becomes $Z_T$ and is generated by the SDE as well. drift function $\mu$, diffusion function $\sigma$, an initial distribution $p_0$, final time $T$, minibatch of samples ${\boldsymbol{x}}$, sample path $\tilde{B}_t(\omega)$ of Brownian motion approximation. an estimate of $\log p({\boldsymbol{x}} \vert \omega)$ Generate $\epsilon = (\epsilon_1,\dots,\epsilon_d)$ for (\[eq:TraceEst\]). <span style="font-variant:small-caps;">odefunc</span>$(z,t)$ Compute $\tilde{\mu}(z,t)$ via (\[eq:ItoCorrection\]).$\triangleright$ Itô correction $\tilde{\mu}(z,t) + \sigma(z,t) \frac{{\mathrm{d}}\tilde{B}_t(\omega)}{{\mathrm{d}}t}$. <span style="font-variant:small-caps;">aug</span>$((z,\log p_t),t)$ $f_t \gets $ <span style="font-variant:small-caps;">odefunc</span>($z$, $t$, $\omega$) $J_t \gets -\nabla_z(\epsilon \cdot f_t) \cdot\epsilon$$\triangleright$ Trace estimator (\[eq:TraceEst\]); $n = 1$ $(f_t, J_t)$ $(z,\Delta\log p_t^\omega) \gets $ <span style="font-variant:small-caps;">odesolve</span>(<span style="font-variant:small-caps;">aug</span>,$(x,0)$,$0$,$T$) $\log p_0(z) - \Delta \log p_t^\omega$ \[alg:DensityEst\] A number of techniques exist for debiasing the logarithm of (\[eq:DensityEstNaive\]) — see @rhee2015unbiased and @rischard2018unbiased, for example. Alternatively, we lie in the setting of *semi-implicit variational inference* seen in @yin2018semi and @titsias2018unbiased, and those techniques directly extend to our case as well. Naturally, it would be easiest to instead optimize the upper bound $$-\log p_t^\theta(Z_t) \leq -\mathbb{E}_{\omega} \log p_t^\theta(Z_t\vert \omega),$$ and in many cases we have found this to be effective. Observing that $$\label{eq:ELBO} \log p_t^\theta(Z_t) - {D_{\mathrm{KL}}}(q \Vert p_{\omega \vert Z_t}^{\theta}) = \mathbb{E}_\omega \log p_t^\theta(Z_t \vert \omega),$$ minimizing $-\mathbb{E}_{\omega}\log p_t^{\theta}(Z_t\vert \omega)$ maximizes the true log-likelihood regularized by the KL-divergence between the prior and posterior distributions for $\omega$, which reduces the effect of noise on the model. At the same time, parameterizations of the diffusion coefficient that allow $\|\sigma\|$ to shrink to zero will often do so, and should be avoided to remain distinct from a continuous normalizing flow. Numerical experiments {#sec:Numerics} ===================== Samplers and density estimation from data ----------------------------------------- For our first experiments, we train a stochastic normalizing flow (\[eq:MainSDE\]) — using Algorithm \[alg:DensityEst\] with the upper bound (\[eq:ELBO\]) — to data generated from a specified target distribution. For our drift function, we adopt the same architecture used in the toy examples of @grathwohl2018ffjord; a four-layer fully-connected neural network with 64 hidden units in each layer. Dependence on time is removed to ensure a time-homogeneous, and hence, potentially ergodic SDE after training. All networks were trained using Adagrad [@duchi2011adaptive], with $p_0 \sim \mathcal{N}(0,I)$ and a batch size of 1000 samples. ### A two-dimensional toy example {#sec:Toy} In our first example, our data is generated from the banana-shaped distribution $$p(x,y) \propto \exp(-\tfrac12 (x^2 + \tfrac12 (x^2 + y)^2)).$$ Two choices of diffusion coefficient are considered: the first, where $\sigma = I$, yields a neural SDE that can be trained using the techniques of @li2020scalable. For the second, we choose $$\label{eq:SigmaChoice} \sigma(x) = \lambda \left(\begin{matrix} 1 & \sigma_1(x) \\ \sigma_2(x) & 1 \end{matrix}\right),$$ with $\lambda = 1$, and parameterize $(\sigma_1,\sigma_2)$ by a two-layered neural network with 64 hidden units. This SDE can only be trained using our method. After training, to emulate the application of these SDEs as approximate samplers, a single sample path with 10,000 steps was simulated for each model using the Euler–Maruyama method. The resulting paths are compared in Figure \[fig:FixedVsVariable\]. From data alone, both models constructed recurrent processes. The addition of a trainable diffusion coefficient led to improved adaptation of the sampler to the underlying curvature. ![Sample paths from SDEs trained as stochastic normalizing flows to a banana-shaped distribution.[]{data-label="fig:FixedVsVariable"}](fixed_vs_variable_sigma.png){width="0.6\columnwidth"} ### Visualizing regularization As discussed in @liu2019neural, the stochastic noise injection in SDEs is a natural form of regularization, that can potentially improve robustness to noisy or adversarial data. We visualize this effect by considering the same stochastic normalizing flows treated in §\[sec:Toy\] with diffusion coefficient (\[eq:SigmaChoice\]), and adjusting the parameter $\lambda > 0$. Our data is generated in polar coordinates from a ten-pointed star-shaped distribution by $$\theta \sim \mathrm{Unif}(-\pi,\pi),\quad r\vert \theta \sim \mathcal{N}(\tfrac{2}{\sqrt{1 + \frac12 \sin(10 \theta)}}, \tfrac{9}{400}).$$ In Figure \[fig:RegDensity\], we plot the densities for $\lambda \in \{0, \tfrac{1}{10}, \tfrac12, 1\}$ computed using Algorithm \[alg:DensityEst\], noting that the $\lambda = 0$ case corresponds to a continuous normalizing flow. Increasing $\lambda$ reveals generative models with expectedly higher variance, but with improved capacity to smooth out minor (potentially, unwanted) details. ![Density plots of stochastic normalizing flows trained to a star-shaped distribution with varying diffusion coefficients.[]{data-label="fig:RegDensity"}](reg_density.png){width="0.5\columnwidth"} Optimizing stochastic MCMC -------------------------- An interesting class of SDE models for approximating a target distribution $p$ are *targeted diffusions*, solutions to SDEs that are $p$-ergodic. A convenient representation of such diffusions are known [@ma2015complete]. Because these diffusions are frequently used in MCMC algorithms, in a sense, conducting VI with respect to targeted diffusions is analogous to optimizing the convergence rate of stochastic MCMC algorithms. To illustrate the potential applications of stochastic normalizing flows for finding and examining optimal stochastic MCMC algorithms for a particular target distribution, we consider a basic setup, where $p$ is the one-dimensional Cauchy distribution $p(x) \propto (1 + x^2)^{-1}$. All $p$-ergodic SDEs are of the form $$\label{eq:CauchySampler} {\mathrm{d}}Z_t = (-2 \sigma(Z_t)^2 Z_t / (1 + Z_t^2) + \tfrac12 \sigma'(Z_t)) {\mathrm{d}}t + \sigma(Z_t) {\mathrm{d}}B_t,$$ and we may choose $\sigma$ arbitrarily. *A priori*, an optimal choice of $\sigma$ (up to constants) to ensure rapid mixing of (\[eq:CauchySampler\]) does not appear obvious. The present rule of thumb from second-order methods takes $\sigma \approx (\log p)''$ [@girolami2011riemann]. We train a stochastic normalizing flow for (\[eq:CauchySampler\]) with $\sigma$ parameterized by a four-layer neural network with 32 hidden units in each layer. The corresponding loss function is taken to be the Kullback-Leibler divergence $\log p_T(Z_T) - \log p(Z_T)$, estimated using Algorithm \[alg:DensityEst\], with an $L^1$ penalty term $10^{-4}\|{\boldsymbol{w}}\|_1$ over the weights ${\boldsymbol{w}}$ of the neural network, to prevent taking $|\sigma| \to +\infty$. The results are presented in Figure \[fig:CauchyPlot\]. ![Stochastic normalizing flow targeting Cauchy distribution. *Left*: trained diffusion coefficient (blue) compared to $\sqrt{1 + x^2}$ (orange dashed). *Right*: Histogram of 20,000 generated samples (blue) and Cauchy density (orange). []{data-label="fig:CauchyPlot"}](cauchy.pdf){width="0.7\columnwidth"} Curiously, after training, we found the apparent “optimal” choice is approximately $\sigma(x) \propto \sqrt{1 + x^2} \propto \sqrt{\frac{(\log \varphi)'}{(\log p)'}}$, where $\varphi$ is the density of the standard normal distribution. Conclusion {#sec:Conclusion} ========== We have extended the continuous normalizing flows framework to generative models involving SDEs. Justified by rough path theory, our framework enables practitioners of neural ODEs to apply their existing implementation for training neural SDEs. This is advantageous, as neural SDEs have been suggested to be more robust than neural ODEs in high-dimensional real-world examples [@liu2019neural; @li2020scalable]. Stochastic normalizing flows can be implemented as a device for investigating “optimal” hyperparameters in stochastic MCMC, which could prove useful for informing future research, but they may require implementational improvements for high-dimensional cases, e.g., variance reduction techniques and improved loss estimators. #### Acknowledgements. We would like to acknowledge DARPA, NSF, and ONR for providing partial support of this work. [41]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{} Beskos, A., Papaspiliopoulos, O., Roberts, G. O., and Fearnhead, P. Exact and computationally efficient likelihood-based estimation for discretely observed diffusion processes. *Journal of the Royal Statistical Society: Series B (Statistical Methodology)*, 680 (3):0 333–382, 2006. Chen, T. Q., Rubanova, Y., Bettencourt, J., and Duvenaud, D. K. Neural ordinary differential equations. In *Advances in Neural Information Processing Systems*, pp. 6571–6583, 2018. Davie, A. M. Differential equations driven by rough paths: an approach via discrete approximation. *Applied Mathematics Research eXpress*, 2008, 2008. Dinh, L., Krueger, D., and Bengio, Y. . In *Proceedings of the 3rd International Conference on Learning Representations (ICLR)*, 2015. Duchi, J., Hazan, E., and Singer, Y. Adaptive subgradient methods for online learning and stochastic optimization. *Journal of Machine Learning Research*, 12:0 2121–2159, 2011. Dupont, E., Doucet, A., and Teh, Y. W. Augmented neural [ODEs]{}. In *Advances in Neural Information Processing Systems*, 2019. Evans, L. C. *An introduction to stochastic differential equations*, volume 82. American Mathematical Society, 2012. Friz, P. and Hairer, M. *A Course on Rough Paths*. Springer International Publishing, 2014. Friz, P. K. and Victoir, N. B. *Multidimensional stochastic processes as rough paths: theory and applications*, volume 120. Cambridge University Press, 2010. Gholami, A., Keutzer, K., and Biros, G. . In *Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19)*, 2019. Girolami, M. and Calderhead, B. . *Journal of the Royal Statistical Society: Series B (Statistical Methodology)*, 730 (2):0 123–214, 2011. Grathwohl, W., Chen, R. T., Betterncourt, J., Sutskever, I., and Duvenaud, D. . *arXiv preprint arXiv:1810.01367*, 2018. Hairer, E. and Wanner, G. *[Solving Ordinary Differential Equations II]{}*, volume 14 of *Springer Series in Computational Mathematics*. Springer-Verlag Berlin Heidelberg, 1996. Hodgkinson, L., Salomone, R., and Roosta, F. . *arXiv preprint arXiv:1903.12322*, 2019. Hurn, A. S., Jeisman, J., and Lindsay, K. A. . *Journal of Financial Econometrics*, 50 (3):0 390–455, 2007. Jia, J. and Benson, A. R. . In *Proceedings of the 33rd Conference on Neural Information Processing Systems (NeurIPS 2019)*, 2019. Kloeden, P. E. and Platen, E. *Numerical solution of stochastic differential equations*, volume 23. Springer Science & Business Media, 2013. Li, X., Wong, T.-K. L., Chen, R. T., and Duvenaud, D. Scalable gradients for stochastic differential equations. *arXiv preprint arXiv:2001.01328*, 2020. Liu, Q. and Feng, Y. Two methods for wild variational inference. *arXiv preprint arXiv:1612.00081*, 2016. Liu, X., Si, S., Cao, Q., Kumar, S., and Hsieh, C.-J. . *arXiv preprint arXiv:1906.02355*, 2019. Lu, Y., Zhong, A., Li, Q., and Dong, B. . *arXiv preprint arXiv:1710.10121*, 2017. Lyons, T. J. Differential equations driven by rough signals. *Revista Matem[á]{}tica Iberoamericana*, 140 (2):0 215–310, 1998. Ma, Y.-A., Chen, T., and Fox, E. . In *Advances in Neural Information Processing Systems*, pp. 2917–2925, 2015. Peluchetti, S. and Favaro, S. Infinitely deep neural networks as diffusion processes. *arXiv preprint arXiv:1905.11065*, 2019. Pontryagin, L. S. *Mathematical theory of optimal processes*. Routledge, 2018. Rezende, D. J. and Mohamed, S. Variational inference with normalizing flows. In *Proceedings of the 32nd International Conference on Machine Learning*, 2015. Rhee, C.-h. and Glynn, P. W. . *Operations Research*, 630 (5):0 1026–1043, 2015. Rischard, M., Jacob, P. E., and Pillai, N. . *arXiv preprint arXiv:1810.01382*, 2018. Roosta, F. and Ascher, U. Improved bounds on sample size for implicit matrix trace estimators. *Foundations of Computational Mathematics*, 150 (5):0 1187–1212, 2015. Ryder, T., Golightly, A., McGough, A. S., and Prangle, D. Black-box variational inference for stochastic differential equations. *arXiv preprint arXiv:1802.03335*, 2018. Salimans, T., Kingma, D., and Welling, M. . In *International Conference on Machine Learning*, pp. 1218–1226, 2015. Titsias, M. K. and Ruiz, F. J. Unbiased implicit variational inference. In *Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics (AISTATS)*, 2019. Twardowska, K. . *Acta Applicandae Mathematica*, 430 (3):0 317–359, 1996. Tzen, B. and Raginsky, M. . *arXiv preprint arXiv:1905.09883*, 2019. van den Berg, R., Hasenclever, L., Tomczak, J., and Welling, M. Sylvester normalizing flows for variational inference. In *Proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI)*, 2018. Wolf, C., Karl, M., and van der Smagt, P. . *arXiv preprint arXiv:1609.08203*, 2016. Wong, E. and Zakai, M. On the convergence of ordinary integrals to stochastic integrals. *The Annals of Mathematical Statistics*, 360 (5):0 1560–1564, 1965. Xu, M., Quiroz, M., Kohn, R., and Sisson, S. A. Variance reduction properties of the reparameterization trick. In *The 22nd International Conference on Artificial Intelligence and Statistics*, pp.  2711–2720, 2019. Yin, M. and Zhou, M. Semi-implicit variational inference. In *Proceedings of the 35th International Conference on Machine Learning*, 2018. Young, L. C. . *Acta Mathematica*, 67:0 251–282, 1936. Zhang, T., Yao, Z., Gholami, A., Keutzer, K., Gonzalez, J., Biros, G., and Mahoney, M. W. . *Advances in Neural Information Processing Systems*, 2019. [^1]: Department of Statistics, University of California at Berkeley, USA, and International Computer Science Institute, Berkeley, CA, USA. Email: `[email protected]` [^2]: School of Mathematics and Physics, University of Queensland, Australia. Email: [^3]: School of Mathematics and Physics, University of Queensland, Australia, and International Computer Science Institute, Berkeley, CA, USA. Email: [^4]: Department of Statistics, University of California at Berkeley, USA, and International Computer Science Institute, Berkeley, CA, USA. Email: `[email protected]`

ArXiv

--- abstract: 'New observations are reported of $J$-band spectra (1.04 $\mu$m – 1.4 $\mu$m) of three Seyfert 2 galaxies, Mkn 34, Mkn 78 and NGC 5929. In each case the spectral range includes the near-infrared lines of [\[Fe[ii]{}\]]{}, [\[P[ii]{}\]]{}, [He[i]{}]{} and [Pa$\beta$]{}. Each Seyfert galaxy has a known radio jet, and we investigate the infrared line ratios of the nuclear and extended regions of each galaxy compared to the radio structure. In Mkn 34 there is a clear indication of an extranuclear region, probably coincident with a shock induced by the radio jet, in which [\[Fe[ii]{}\]]{} is considerably enhanced, although the nuclear emission is almost certainly the result of photoionization by the continuum of the active nucleus. Similar effects in extranuclear regions are seen in the other objects, in the case of Mkn 78 confirming recent studies by Ramos Almeida et al. A possible detection of extranuclear [\[P[ii]{}\]]{} emission suggests, if real, that photoionization by the active nucleus is the dominant line excitation mechanism over the whole source, including the regions coincident with the radio jet.' author: - | N. Jackson and R. J. Beswick\ The University of Manchester, Jodrell Bank Observatory, Macclesfield, Cheshire, SK11 9DL U.K.\ title: 'Near-infrared spectra of Seyfert galaxies and line production mechanisms\' --- \[firstpage\] galaxies:active – galaxies:individual:Mkn34 – galaxies:individual:Mkn78 – galaxies:individual:NGC5929 – line formation – infrared:galaxies Introduction ============ Optical and near-infrared studies of nearby active galaxies can give us a lot of information about the physical processes taking place in these objects. Seyfert galaxies, as relatively nearby low-luminosity active galaxies, can be studied in particular detail. These objects contain emission lines within their spectrum, including in general broad lines produced within the inner parsec, close to the central black hole and accretion disk, and narrow lines whose region of production may extend over a kiloparsec. In addition, many Seyfert galaxies have weak radio emission, often in a linear structure corresponding to an outflow from the central active nucleus. The physics of emission lines is of particular interest. These are thought mostly to arise through photoionization, and many studies have been done which use sophisticated photoionization models to reproduce emission lines in great detail. However, additional physics may be needed to reproduce some lines. In particular, the near-infrared [\[Fe[ii]{}\]]{}line at 1.257$\mu$m has been suggested as a diagnostic of shock excitation, which may arise as dust grains are dissociated in fast shocks (e.g. Forbes & Ward 1993). The idea that shocks may have general applicability to AGN spectra was argued by Sutherland et al. (2003) and Dopita & Sutherland (1995) who succeeded in reproducing many characteristic photoionized spectra using shock models. This arises because shocks generate UV and X-ray photons which can then ionize the gas – indeed, the difference between the observed properties of such “autoionizing” shocks and regions photoionized by a central active nucleus can be quite subtle and depend on the details of the spectra of the incident photons. More recently, Simpson et al. (1996) showed that the observed [\[Fe[ii]{}\]]{} line is also consistent with predominantly being produced by photoionization, and Mouri, Kawara & Taniguchi (2000) conducted more sophisticated calculations suggesting that electron collisions in a partially-ionized zone associated with a shock may be the production mechanism rather than grain dissociation. Based on such studies, Alonso-Herrero et al. (1997) propose the use of the ratio of infrared [\[Fe[ii]{}\]]{} lines to the hydrogen recombination lines to separate active galaxies from starbursts. Oliva et al. (2001) studied the problem of shock ionization versus photoionization. They suggested that the [\[P[ii]{}\]]{} line is potentially useful in disentangling this problem, because similar physical conditions are needed to produce it to those which produce [\[Fe[ii]{}\]]{}, but unlike iron, phosphorus is not produced in destruction of dust grains. A high Fe/P ratio is thus a prediction of models in which dust grains are dissociated by shocks. Oliva et al. find that pure photoionization better fits their spectrum of the very bright Seyfert galaxy NGC 1068. In fainter galaxies where [\[P[ii]{}\]]{} is not detected, even lower limits on the Fe/P ratio may be useful in ruling out shock models if the spectra are sensitive enough. Although Seyfert galaxies have been studied extensively in the optical, high signal-to-noise observations in the near-infrared are rarer. A compilation has recently been published by Riffel, Rodríguez-Ardila & Pastoriza (2006) including many of their own $JHK$ spectra (Rodríguez-Ardila, Riffel & Pastoriza 2005). Many of these spectra suggest that the line ratio of active galaxies have a [\[Fe[ii]{}\]]{}/[H$\beta$]{} ratio suggestive of X-ray heating or regions of star formation. Recently, Ramos Almeida et al. (2006) have published infrared spectra of one of the objects discussed here, Mkn 78, and deduce from observations and simulations that the predominant line-excitation mechanism in the nuclear region is photoionization by the hard UV continuum from the active nucleus, although radio-induced shocks may contribute within one of the radio lobes. In this work we choose objects in which linear radio structure is clearly seen in order to investigate the emission-line structure along the jet axis and look for evidence for shock excitation. If shocks are indeed an important part of the physics, the area where the radio jet is depositing energy into the interstellar medium is an obvious place to look. The spectra presented here are from deep exposures (1–2 hours with an 8-m class telescope) and therefore allow us to investigate both nuclear and off-nuclear emission. Observations and analysis ========================= Infrared and optical data ------------------------- All infrared observations were conducted using the 8-m Gemini-North Telescope, Mauna Kea, on the nights of 2005 February 17, February 18 and May 22 using the NIRI imaging spectrograph in f/6 spectroscopy mode. The J grism was used with a J-band order-sorting filter, giving a wavelength range of approximately 1.05$\mu$m – 1.42$\mu$m, together with a 4-pixel slit which corresponds to 046 on the sky and a resolving power of about 600 (Hodapp et al. 2003). Mkn78 was observed for 2 hours on 2005 February 17, Mkn34 for 2 hours on 2005 February 18 and NGC 5929 for 70 minutes on 2005 May 22. Each observation was divided into 3$\times$30 second exposures, and each successive exposure was chopped up and down the slit to make sky subtraction easier and minimize effects of bad pixels. Lamp flatfield exposures and argon arc exposures were also obtained for Mkn34 and Mkn78. In the case of NGC5929, no argon lamp spectrum is available and in this case the atmospheric OH lines have been used for wavelength calibration. Data reduction was performed using the Starlink [figaro]{} package and began with combination of the flatfields and division of the data frames by a flatfield normalised in the spectral direction. A second-order polynomial fit to at least six arc lines was obtained which gave a maximum error of 0.1 nm across the wavelength range. The spectra were corrected for S-distortion using three spectra of the Hipparcos stars taken at different positions on the chip. The sky was then subtracted using the Figaro task [polysky]{} and the spectra extracted using the optimal extraction algorithm of Horne (1986) using the Figaro task [optextract]{}. Relative flux calibration was performed assuming a blackbody spectrum of 7200K for the Hipparcos F0V star HIP36366 (Perryman 1997) and temperatures of 7200K and 7000K for the stars HIP56601 and HIP75411 respectively. The Pa5$\rightarrow$3, 6$\rightarrow$3 and higher order absorption bands were interactively edited out of the stellar spectrum before this process in order to avoid them appearing as emission lines in the divided spectrum. For deriving velocity profiles, successive rows of the sky-subtracted image have been extracted, continuum-subtracted using line-free regions immediately around each line, and fitted with Gaussian profiles. Profile fitting has been done by software written by us, using the Levenberg-Marquardt algorithm as implemented by Press et al. (1992). In addition to the infra-red data, some data from the Isaac Newton Group (ING) archive was extracted for the Seyfert galaxy Markarian 34. These data were taken using the William Herschel 4.2-m telescope on La Palma on 1998 December 17 using the ISIS two-arm spectrograph. On the red arm of this spectrograph the TEK-2 chip was used together with an R600R grating, using a 123 slit and a resulting spectral resolution of just over 0.2nm. The data were co-added and fluxes of [\[O[i]{}\]]{} and [H$\alpha$]{} derived assuming a flat spectral response over this wavelength range, because both the TEK-2 chip and the R600R grating have a flat spectral response over this range to within a few percent. Radio data ---------- Complementary, mainly previously-published (Falcke, Wilson & Simpson 1998; Whittle & Wilson 2004) public domain radio data for each of these three sources were obtained and re-mapped for comparison purposes from the NRAO’s Very Large Array (VLA) archive. In each case these data were observed at a wavelength 3.6cm and in the VLA’s highest resolution configuration, providing an angular resolution of $\sim$0.2-0.3 arcsec which is directly comparable to the spectroscopic observations presented. A brief summary of the observing parameters is given in Table\[tab1\]. Each data-set was calibrated using standard data reduction techniques within NRAO’s [aips]{} package where these data were edited before phase solutions from a nearby reference source were applied. In each case the flux density scale was calibrated with respect to 3C286 using the Baars [*et al.*]{} (1977) scale. [lcccl]{} Source&Date&Phase ref. & Time on source & Original publication\ &&&&\ Mkn34&1996 Nov 4&1030+611&166 min & Falcke, Wilson & Simpson 1998\ NGC5929&2003 Jun 20&1506+426&8 min&\ Mkn78&1990 Apr 16&0716+714&340 min&Whittle & Wilson 2004\ Results and line diagnostics ============================ Mkn 34 ($z$=0.0505) ------------------- Markarian 34 is a Seyfert 2 galaxy with a 24 extended radio structure whose extended optical line emission has been studied in detail by Whittle et al. (1988) and Falcke, Wilson & Simpson (1998). Its off-nuclear line emission shows a velocity profile with relatively broad [\[O[iii]{}\]]{} emission extending slightly beyond the radio lobes. There is a distinct difference between the velocity of the line emission near to the nucleus and that beyond the radio lobes. The new infra-red spectrum was taken with the slit at a position angle 158$^{\circ}$ East of North, coincident with the radio axis. Figure 1 shows the image of the spectrum along the slit, showing a clear rotation-type velocity profile in all securely detected lines ([He[i]{}]{}, [\[Fe[ii]{}\]]{}, [Pa$\beta$]{} and [Pa$\gamma$]{}). [\[P[ii]{}\]]{} is detected at a marginal level in the nuclear spectrum. In the nuclear region, the ratio [\[Fe[ii]{}\]]{}$\lambda$1.257/[\[P[ii]{}\]]{} is 3.2$\pm$1.0, which is consistent with the range of values typical of nuclear photoionization (Oliva et al. 2001). The radio map of Falcke et al. (1998) shows structure similar to that seen earlier by Ulvestad & Wilson (1984) at 6 and 20cm, but with higher angular resolution and sensitivity shows the definition of the radio jet structure more clearly. In particular the jet is resolved into several discrete knots and a significant change in the projected position angle of the tip of the southeastern jet is seen. The most striking feature of the spectrum is the different distribution between [\[Fe[ii]{}\]]{} and the other lines. Figure 2 illustrates the rotation curve and strength of the lines, and it is clear that [\[Fe[ii]{}\]]{} has a bright knot about 12 to the SE. This is just behind the bright radio knot in the southeastern jet imaged here and by Whittle et al. (1988). Although the Paschen recombination lines are slightly enhanced in the off-nuclear regions compared to the nucleus, the ratio of [\[Fe[ii]{}\]]{}/[Pa$\beta$]{} becomes a factor of 2–3 stronger away from the nucleus, suggesting that different ionization mechanisms may be operating here. The [He[i]{}]{} line displays a different behaviour to the hydrogen recombination lines, showing a smooth decrease in each direction away from the nucleus with little sign of any increase in strength in the radio lobes. NGC5929 ($z$=0.00831) --------------------- NGC5929 is a Seyfert galaxy with a radio jet extending about 07 in PA $\sim 60^{\circ}$ either side of a compact nucleus (Su et al. 1996; Cole et al. 1998) and extended optical emission partly coincident with the western radio lobe (Bower et al. 1994), strongly suggesting an interaction. Nuclear infrared spectroscopy was performed by Simpson et al. (1996) who clearly detected the [\[Fe[ii]{}\]]{} line and found it to be marginally resolved. The current observations have a smaller pixel scale, and we clearly see the resolution in this object, using a slit at the position angle of the extended optical emission previously seen (65$^{\circ}$). The nuclear spectrum is shown in Figure 3. In the nuclear emission the [\[P[ii]{}\]]{} line is detected at just over the 3$\sigma$ level; a previous spectrum reported by Rodrígues-Ardila, Riffel & Pastoriza (2005) and Riffel et al. (2006) did not detect this line. A Gaussian fit to the [\[Fe[ii]{}\]]{} and [\[P[ii]{}\]]{} lines was carried out, constraining the [\[P[ii]{}\]]{} line width to be equal to that of [\[Fe[ii]{}\]]{}. The [\[Fe[ii]{}\]]{}/[\[P[ii]{}\]]{} ratio is 4.3$\pm$1.1 in the nuclear region, which again is typical of photoionized regions and considerably less than the value of $\sim$20 expected for shock ionization (Oliva et al. 2001). The [\[Fe[ii]{}\]]{}/[Pa$\beta$]{} ratio is approximately 1, in agreement with the previous spectra. Figure 4 shows the radio map; similar structure is seen to the earlier radio map of Su et al. (1996) and Cole et al. (1998) with a compact core and two lobe structures detected. The most remarkable feature of the spectrum is the wholesale displacement of the line emission from the centre of the galaxy and of the velocity profile (Figure 5). This effect is extreme around the southwestern radio lobe, where the [\[Fe[ii]{}\]]{} line strength peaks. The [He[i]{}]{} line peak is also offset, although by less than the [\[Fe[ii]{}\]]{}. Optical studies with the [*Hubble Space Telescope*]{} (Bower et al. 1994) have also revealed [\[O[iii]{}\]]{} emission which peaks around the southwestern radio lobe. Around the northeastern radio lobe, there is some evidence of disturbed kinematics in the [\[Fe[ii]{}\]]{} line. Mkn78 ($z$=0.03715) ------------------- Mkn 78 has recently been the subject of an extensive optical study with the Hubble Space Telescope by Whittle & Wilson (2004), Whittle et al. (2004) and Whittle et al. (2005) specifically directed at disentangling shock ionization and photoionization. This Seyfert galaxy has extended radio emission, in a linear structure about 2$^{\prime\prime}$ long in an E-W direction, and Whittle & Wilson (2004) have imaged a large and complex extended structure in [\[O[iii]{}\]]{} extending 3$^{\prime\prime}$ from the nucleus in each direction. In these observations the slit was aligned in PA 90$^{\circ}$ in the hope of detecting significant extended structure in the infrared emission lines. The nuclear spectrum of this object is shown in Figure 6, and the measurements of the extended emission and velocity profile are displayed in Figure 7. In the optical observations of Whittle et al. (2005), very complex extended optical line emission was detected in all major recombination and forbidden lines. In these infrared observations, the structure is not well resolved either spatially or spectrally. However, the overall rotation curve is detected in extended emission in three major emission lines, [He[i]{}]{}, [\[Fe[ii]{}\]]{} and [Pa$\beta$]{}, covering about 2$^{\prime\prime}$ to the East of the nucleus. Relatively little line emission is seen to the West, which is also the direction of lower surface brightness in [\[O[iii]{}\]]{} from the maps of Whittle & Wilson (2004). A recent detailed infrared study has been carried out with the LIRIS spectrograph on the William Herschel Telescope by Ramos Almeida et al. (2006). They find a nuclear ratio of [\[Fe[ii]{}\]]{}/[Pa$\alpha$]{} which is consistent with photoionization from the active nucleus. Moreover, they detect [\[P[ii]{}\]]{} in the nucleus at a strength that is a factor of 1.82 lower than the [\[Fe[ii]{}\]]{}$\lambda$1.644 line. This ratio is also consistent with photoionization, being similar to the ratio found in NGC1068 by Oliva et al. (2001), and is inconsistent with shock models. However, in the extended emission associated with the western radio lobe, they find that the [\[Fe[ii]{}\]]{} line strength increases by a factor $\sim$2 relative to the hydrogen recombination lines, suggesting a contribution from interaction with the radio bow shock. We find a similar effect in our data, with an enhancement in [\[Fe[ii]{}\]]{} which peaks at $\sim 0\farcs6$ from the nucleus. However, the association of increased [\[Fe[ii]{}\]]{} emission with the radio structure is [*less*]{} obvious in this object than in the other two in this study. In the western lobe the peak of the [\[Fe[ii]{}\]]{} does coincide with the W-knot of radio emission seen at higher resolution by Whittle et al. (2004) and shown in the superposed uniformly weighted image in the top panel of figure 7. The stronger, eastern radio jet is co-spatial with a region in which the [\[Fe[ii]{}\]]{} (and [Pa$\gamma$]{}) emission becomes extremely faint and difficult to measure accurately. The [He[i]{}]{} line displays different behaviour; as in Mkn34, its strength decreases smoothly away from the nucleus. The [Pa$\gamma$]{} line, although it can be followed for about 2$^{\prime\prime}$ either side of the nucleus, is not detected at high enough signal-to-noise to describe its structure in detail. Ionization mechanisms and diagnostics ===================================== Nuclear emission: the [\[Fe[ii]{}\]]{}/[\[P[ii]{}\]]{} ratio ------------------------------------------------------------ This paper adds to the total number of detections of [\[P[ii]{}\]]{}$\lambda$1.188$\mu$m in the literature. Since its original detection in NGC 1068 by Oliva et al. (2001) the line has been detected in the Seyfert 1 galaxy Akn 564 by Rodríguez-Ardila, Riffel & Pastoriza (2005), in Mkn 78 by Ramos Almeida et al. (2006), and in a total of 11 galaxies in the compilation of Riffel et al. (2006). Although very marginal, the detections of [\[P[ii]{}\]]{} in Mkn 34 and NGC 5929 suggest [\[Fe[ii]{}\]]{}/[\[P[ii]{}\]]{} ratios of 3–4, in agreement with ratios determined in other objects. Comparison with the models presented by Oliva et al. (2001) suggests that the dominant ionization mechanism is photoionization by the active nucleus rather than shock excitation; fast shocks would be expected to disrupt dust grains and produce line ratios a factor of 10 higher. Extended emission ----------------- ### Line diagnostic diagrams There have been a number of investigations which have attempted to compare near-infrared line ratios with ionization codes. Many have used the [cloudy]{} code (Ferland et al. 1998) for a range of physical parameters, namely: density of hydrogen, ionization parameter, ionizing power-law slope, metal abundance and presence or absence of dust grains (Simpson et al. 1996; Ramos Almeida et al. 2006). Others have used the [mappings]{} code or modifications to it (Sutherland et al. 1993; Mouri et al. 2000). Of particular interest is the ratio of [\[Fe[ii]{}\]]{} to the other lines available in the near-infrared spectrum, in particular the hydrogen recombination lines. Also interesting is the optical [\[O[i]{}\]]{}$\lambda$6300 line, as this has a similar ionization potential to [\[Fe[ii]{}\]]{} and is produced in a partially-ionized zone with similar physical properties to the region producing [\[Fe[ii]{}\]]{}. It can therefore serve as an alternative diagnostic of shock excitation to [\[P[ii]{}\]]{}. Although much stronger than [\[P[ii]{}\]]{}, [\[O[i]{}\]]{} suffers from the disadvantage that the [\[O[i]{}\]]{}/[\[Fe[ii]{}\]]{}ratio is subject to substantial reddening and is difficult to measure with the same instrument. The effect of different physical parameters on the relative strengths of the [\[Fe[ii]{}\]]{} and hydrogen recombination lines (in particular [Pa$\beta$]{} ) can be summarized as follows (Simpson et al. 1996, Mouri et al. 2000; Ramos Almeida et al. 2006): - Metal abundance changes the [\[Fe[ii]{}\]]{}/[Pa$\beta$]{} ratio by approximately the scaling of the abundance, for example by a factor of $\sim$6 from solar to Orion abundance. - Neutral hydrogen density has relatively little effect, except at very high densities when the forbidden lines are in any case close to collisional de-excitation. For most values of ionization parameter, increasing densities decreases the relative strength of [\[Fe[ii]{}\]]{} very slightly. - Increasing ionization parameter causes a slow decrease in the [\[Fe[ii]{}\]]{}/[Pa$\beta$]{} ratio below $U=10^{-1.5}$, and a rapid increase above this. - The slope of the power-law has a drastic effect on the [\[Fe[ii]{}\]]{}/[Pa$\beta$]{} ratio; a steepening by 0.4 (from $-$1.0 to $-$1.4) typically halves the ratio, and the ratio becomes even smaller if a 40000-K blackbody model is adopted instead. This was considered in detail by Simpson et al. (1996), although Ramos Almeida et al. (2006) use an index of $-$2.0 throughout their simulations. In order to compare results with physical models and extract physical information from the line ratios, we have used the [cloudy]{} program (Ferland et al. 1998), version c06.02b. Models have been calculated for ionization parameters ranging from $10^{-3.5}$ to $10^{-1.5}$ in steps of $10^{0.5}$ and electron densities between $10^{3.5}$ and 10$^6$, also in steps of $10^{0.5}$. For each combination of ionization parameter and density, two different metallicities ($0.4Z_\odot$ and $Z_\odot$) and two values of the spectral index of the power-law photoionizing continuum ($-1$ and $-2$) have been used. Grains were added to the simulation and were adjusted to be similar to Orion grains. The diagnostic diagram of [\[Fe[ii]{}\]]{}$\lambda$1.257/[Pa$\beta$]{} vs. [\[O[i]{}\]]{}/[H$\alpha$]{} is shown for Mkn 34 in Figure 8. This is the same diagram as presented in Figure 3 of Mouri et al. (2000). The trends in line ratio with ionization parameter, hydrogen density and slope of the ionizing continuum agree with previous determinations. In particular, an increase in ionization parameter causes decreases in both the [\[Fe[ii]{}\]]{}$\lambda$1.257/[Pa$\beta$]{} and the [\[O[i]{}\]]{}/[H$\alpha$]{} ratio of about the same magnitude, provided that the slope of the ionizing continuum, is relatively steep ($<-1.5$); in practice, relatively steep continua are needed to fit these data. We have already noted that the ratio between [\[Fe[ii]{}\]]{}$\lambda$1.257 and the hydrogen recombination lines increases in the objects studied as one moves from the nuclear regions to the areas associated with radio lobe emission. In principle, if the sole mechanism operating is ionization by hard-UV radiation from an active nucleus, this change in line ratio could be caused by a decrease in ionization parameter resulting from a geometrical dilution in the ionizing photon flux or an increase in the electron density associated with compression by the radio jet. This is indeed the conclusion of Falcke et al. (1998) who study the variation of the [\[O[iii]{}\]]{}/[H$\alpha$]{} ratio and find that it decreases away from the nucleus. In this case, however, one would also expect the [\[O[i]{}\]]{}/[H$\alpha$]{}  ratio to increase, and this does not seem to be the case (Figure 8). In fact, this ratio is actually smaller in the southeastern lobe of Mkn 34, where the strongest interaction may be taking place, than in the nucleus of Mkn 34. Since the [\[O[i]{}\]]{} and [\[Fe[ii]{}\]]{} lines are produced in the same type of partially ionized zone and in similar physical conditions, the lack of correlation between their ratios to hydrogen recombination lines is puzzling. The same, however, does not appear to be the case in Mkn 78. Whittle et al. (2005)’s extensive study of the extended optical emission lines in this object suggest that all lines, including [\[O[i]{}\]]{}, follow standard tracks from photoionization diagnostics, with a variation of about a factor of 10 in $U$; this is sufficient to produce a difference of about a factor of 2 in [\[Fe[ii]{}\]]{}$\lambda$1.257/[Pa$\beta$]{}. Images of some optical emission lines in the extended regions of NGC 5929 are presented by Bower et al. (1994) and discussed further by Su et al. (1996); however, we do not have enough information on this object to present detailed diagnostic diagrams comparable with previous work. ### Tentative detections of extended [\[P[ii]{}\]]{} A definitive test would be the detection of the [\[P[ii]{}\]]{} line in the extended emission line region. In Figure 9 an extraction is shown of 07 of the slit in the region where the [\[Fe[ii]{}\]]{} line is strongest, just behind the southeastern radio lobe of Mkn 34. The [\[Fe[ii]{}\]]{} line in this region has been fitted with a Gaussian, and is shown shifted to the [\[P[ii]{}\]]{} wavelength and divided in intensity by factors of 4 and 20. There is a spike at exactly the expected wavelength, although it is extremely close to the noise level. If this is real, it clearly indicates that the mechanism operating in this extended region is photoionization, with little or no shock contribution. Unfortunately the signal-to-noise in this spectrum is just too low to make a definite statement. However, repeating this exercise in the extended region of NGC5929 associated with the southwestern radio lobe (Figure 5) and in which the [\[Fe[ii]{}\]]{} is also enhanced over its value in the nuclear region, reveals a similar [*very*]{} marginal [\[P[ii]{}\]]{} detection in the extended emission. Adding the two together, with the appropriate scaling in wavelength (Figure 9) gives a more secure detection of the extended [\[P[ii]{}\]]{} at a level of about 0.2 times the strength of the extended [\[Fe[ii]{}\]]{} line. This suggests, somewhat surprisingly, and despite the association with the radio jet, that the dominant mechanism of line excitation in these regions is still photoionization rather than shock excitation. Conclusions =========== A number of near-IR studies of Seyfert galaxies with extended optical emission have now been carried out, in particular Mkn78 (Ramos Almeida et al. 2006 and this work), Mkn34 and NGC1068 (Oliva et al. 2001). The picture which emerges with reasonable consistency is that the nuclear regions are almost certainly dominated by photoionization by UV photons, consistent with a mechanism involving the active nucleus. This is indicated both by the ability of photoionization to reproduce observed emission line ratios, and specifically by the inconsistency of the [\[Fe[ii]{}\]]{}/[\[P[ii]{}\]]{} diagnostic with the enhanced ratios expected from shock models. In the extended emission regions which coincide with radio jets, there is a consistent pattern of changes in emission line ratio consistent with compression of the gas by the bowshock associated with radio emission. Tentative detections of [\[P[ii]{}\]]{} in these regions, however, suggest that photoionization is still the dominant mechanism. The fact that the peak line emission is slightly behind the head of the radio jet in Mkn34 may suggest that the gas has been compressed by the shocks but ionized by the active nucleus after the [\[Fe[ii]{}\]]{} has been returned to the gas phase, rather than by UV and X-ray photons from the shocks themselves. In this case the fact that NGC5929’s peak of [\[Fe[ii]{}\]]{} is coincident with the radio lobe could indicate that the shock compression of the gas has occurred more recently. It is important to confirm these determinations in other objects, using longer integrations on large (8-m class) telescopes. Acknowledgments {#acknowledgments .unnumbered} =============== Based on observations obtained at the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the Particle Physics and Astronomy Research Council (United Kingdom), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), CNPq (Brazil) and CONICET (Argentina). RJB acknowledges financial support from the European Commission’s I3 Programme “RADIONET” under contract No. 505818. The VLA is operated by the National Radio Astronomy Observatory which is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. We thank the observers, I. Song and S. Knights and the operators, B. Walls and G. Trancho, for carrying out the observations. References {#references .unnumbered} ========== Alonso-Herrero A., Rieke M.J., Rieke G.H., Ruiz M., 1997, ApJ, 482, 747 Baars, J. W. M., Genzel, R., Pauliny-Toth, I. I. K., Witzel, A., 1977, A&A, 61, 99 Bower G.A., Wilson A.S., Mulchaey J.S., Miley G.K., Heckman T.M., Krolik J.H., 1994, AJ, 107, 1686 Cole G.H.C., Pedlar A., Mundell C.G., Gallimore J.F, Holloway A. J., 1998, MNRAS, 301, 782 Dopita M.A., Sutherland R.S. 1995, ApJ, 455, 468. Falcke H., Wilson A.S., Simpson C., 1998. ApJ, 502, 199. Ferland G., Korista K.T., Verner D.A., Ferguson J.W., Kingdon J.B., Verner E.M., 1998. PASP, 110, 761 Forbes D.A., Ward M.J. 1993, ApJ, 416, 150. Hodapp K.W., Jensen J.B., Irwin E.M., Yamada H. Chung R. Fletcher K. Robertson L. Hora J.L., Simons D.A., Mays W. 2003, PASP, 115, 1388. Horne K., 1986. PASP 98, 609 Mouri H., Kawara K., Taniguchi Y., 2000, ApJ, 528, 186. Oliva E., Marconi A., Maiolino R., Testi L., Mannucci F., Ghinassi F., Licandro J., Origlia, L., Baffa, C., Checcucci, A., 2001, A&A, 369L, 5. Pedlar A., Harrison B., Unger S.W., Graham D.A., Preuss E., Saikia D.J., Yates G.J. 1988, LNP, 307, 310. Perryman M.A.C., 1997. The Hipparcos and Tycho catalogues. Astrometric and photometric star catalogues derived from the ESA Hipparcos Space Astrometry Mission, Noordwijk, Netherlands, ESA Publications Division, 1997, ESA SP Series vol. 1200. Press W.H., Teukolsky S.A., Vetterling W.T., Flannery B.P., 1992. Numerical Recipes in C, Cambridge University Press, Cambridge. Ramos Almeida C., Pérez-Garcia A.M., Acosta-Pulido J.A., Rodríguez-Espinosa J.M., Barrena R., Manchado A., 2006. ApJ 645, 148 Riffel R., Rodríguez-Ardila A., Pastoriza M.G., 2006, A&A 457, 61 Rodrígues-Ardila A., Pastoriza M.G., Viegas S., Sigut T.A.A., Pradhan A.K., 2004, A&A, 425, 457 Rodrígues-Ardila A., Riffel R., Pastoriza M.G., 2005, MNRAS, 364, 1041 Simpson, C. Forbes, D.A., Baker, A.C., Ward, M.J. 1996, MNRAS, 283, 777. Su B.M., Muxlow T.W.B., Pedlar A., Holloway A.J., Steffen W., Kukula M.J., Mutel R.L. 1996, MNRAS, 279, 1111. Sutherland R.S., Bicknell G.V., Dopita M.A. 2003, ApJ, 591, 238. Ulvestad J.S., Wilson A.S. 1984, ApJ, 278, 544. Whittle M., Wilson A.S. 2004, AJ, 127, 606. Whittle M., Pedlar A., Meurs E.J.A., Unger S.W., Axon D.J., Ward M.J. 1988, ApJ, 326, 125. Whittle M., Silverman J.D., Rosario D.J., Wilson A.S., Nelson C.H., 2004, I.A.U. Symp. 222 “The interplay among black holes, stars and ISM in galactic nuclei”, Gramado, Brazil. March 2004, Eds. T. Storchi-Bergmann, L. C. Ho and H. R. Schmitt. p299-302 Whittle M., Rosario D.J., Silverman J.D., Nelson C.H., Wilson A.S. 2005, AJ, 129, 104.

ArXiv

--- abstract: 'We study the onset of a classical order parameter after a second-order phase transition in quantum field theory. We consider a quantum scalar field theory in which the system-field (long-wavelength modes), interacts with its environment, represented both by a set of scalar fields and by its own short-wavelength modes. We compute the decoherence times for the system-field modes and compare them with the other time scales of the model. We analyze different couplings between the system and the environment for slow quenches. Within our approximations decoherence is in general a short time event.' address: | [*$^1$ Departamento de Física, Facultad de Ciencias Exactas y Naturales\ Universidad de Buenos Aires - Ciudad Universitaria, Pabell' on I\ 1428 Buenos Aires, Argentina\ $^2$ Theoretical Physics Group, The Blackett Laboratory,\ Imperial College, London SW7 2BZ, U.K.*]{} author: - 'F. C. Lombardo$^{1}$ [^1], F. D. Mazzitelli $^1$ [^2], and R. J. Rivers $^2$ [^3]' title: 'Decoherence in Field Theory: General Couplings and Slow Quenches' --- 0.5cm PACS numbers: 03.70.+k, 05.70.Fh, 03.65.Yz Introduction ============ The emergence of classical physics from quantum behaviour is important for several physical phenomena in the early Universe. This is beyond the fundamental requirement that only after the Planck time can the metric of the Universe be assumed to be classical. For example, - The inflationary era is assumed to have been induced by scalar inflaton fields, with simple potentials [@linde]. Such fields are typically assumed to have classical behaviour, although in principle a full quantum description should be used. - The origin of large scale structure in the Universe can be traced back to quantum fluctuations that, after crossing the horizon, were frozen and became classical, stochastic, inhomogeneities [@todosLSS]. - It is generally assumed that several phase transitions have occurred during the expansion of the Universe [@old]. As in the case for the inflaton fields, the (scalar) order parameter fields that describe these transitions are described classically. However, the description of early universe phase transitions from first principles is intrinsically quantum mechanical[@cormier]. - As a specific application [@Kibble] of the previous point, the very notion of topological defects (e.g. strings and monopoles) that characterize the domain structure after a finite-time transition, and whose presence has consequences for the early universe, is based on this assumption of classical behaviour for the order parameter[@vilen], as it distributes itself between the several degenerate ground states of the ordered system. In the present paper we are concerned with the third point above, the quantum to classical transition of the order parameters in second order phase transitions. Any approach must take into account both the quantum nature of the order parameter and the non-equilibrium aspects of the process. The problem of the quantum to classical transition in the context of inflationary models was first addressed by Guth and Pi [@guthpi]. In that work, the authors used an inverted harmonic oscillator as a toy model to describe the early time evolution of the inflaton, starting from a Gaussian quantum state centered on the maximum of the potential. They subsequently showed that, according to Schrödinger’s equation, the initial wave packet maintains its Gaussian shape (due to the linearity of the model). Since the wave function is Gaussian, the Wigner function is positive for all times. Moreover, it peaks on the classical trajectories in phase space as the wave function spreads. The Wigner function can then be interpreted as a classical probability distribution for coordinates and momenta, showing sharp classical correlations at long times. In other words, the initial Gaussian state becomes highly squeezed and indistinguishable from a classical stochastic process. In this sense, one recovers a classical evolution of the inflaton rolling down the hill. A similar approach has been used by many authors to describe the appearance of classical inhomogeneities from quantum fluctuations in the inflationary era [@staro]. Indeed, the Fourier modes of a massless free field in an expanding universe satisfy the linear equation $$\phi_k''+(k^2-{a''\over a})\phi_k = 0.$$ For sufficiently long-wavelengths ($k^2\ll a''/a$), this equation describes an unstable oscillator. If one considers an initial Gaussian wave function, it will remain Gaussian for all times, and it will spread with time. As with the toy model of Guth and Pi, one can show that classical correlations do appear, and that the Wigner function can again be interpreted as a classical probability distribution in phase space. \[It is interesting to note that a similar mechanism can be invoked to explain the origin of a classical, cosmological magnetic field from amplification of quantum fluctuations\]. However, classical correlations are only one aspect of classical behaviour. It was subsequently recognized that, in order to have a complete classical limit, the role of the environment is crucial, since its interaction with the system distinguishes the field basis as the pointer basis [@kiefer]. \[We are reminded that, even for the fundamental problem of the space-time metric becoming classical, simple arguments based on minisuperspace models suggest that the classical treatment is only correct because of the interaction of the metric with other quantum degrees of freedom [@halli].\] While these linear instabilities cited above characterise [*free*]{} fields, the approach fails when interactions are taken into account. Indeed, as shown again in simple quantum mechanical models (e.g. the anharmonic inverted oscillator), an initially Gaussian wave function becomes non-Gaussian when evolved numerically with the Schrödinger equation. The Wigner function now develops negative parts, and its interpretation as a classical probability breaks down [@diana]. One can always force the Gaussianity of the wave function by using a Gaussian variational wave function as an approximate solution of the Schrödinger equation, but this approximation deviates significantly from the exact solution as the wave function probes the non-linearities of the potential [@diana; @stancioff]. When interactions are taken into account, classical behaviour is recovered only for “open systems”, in which the relevant degrees of freedom interact with their environment. When this interaction produces [*both*]{} a diagonalization of the reduced density matrix and a positive Wigner function, the quantum to classical transition is completed [@giulinibook]. Going from quantum mechanical toy models to quantum field theory is, of course, extremely difficult. For this reason, several authors [@mihaila] have considered different approximations in quantum mechanics, and compared them to the exact results. If successful in quantum mechanics, they could be implemented with greater confidence in field theory. As already mentioned, when this procedure is applied in the context of [*closed systems*]{}, the conclusion is that, in general, mean-field approximations do not reproduce the evolution of the system at late time[@diana; @stancioff; @mihaila]. Therefore, in principle, there is not reason to believe they will do so in field theory. In spite of this, computational necessity has lead many authors to perform calculations in closed field theories with Hartree, mean field, or $1/N$ approximations, since they are non-perturbative, well-defined and suitable for numerical calculations [@salman; @mottola; @devega; @Boya]. In such calculations, classical correlations do appear in some field theory models [@mottola; @devega]. However, since such decoherence (in a time-averaged sense) takes place at long times after the transition has been achieved initially, when the mean field approximation has broken down, this may be an artifact of the Gaussian-like approximations [@diana]. In a previous paper [@diana] we considered similar arguments for quantum mechanics, but for [*open systems*]{}. When an anharmonic inverted oscillator is coupled to a high temperature environment, it becomes classical very quickly, even before the wave function probes the non-linearities of the potential. Being an early time event, the quantum to classical transition can now be studied perturbatively. In general, recoherence effects are not expected [@nuno]. Taking these facts into account, we have extended the approach to field theory models [@lomplb]. In field theory, one is usually interested in the long-wavelengths of the order parameter. Moreover, the early universe is replete with fields of all sorts which comprise a rich environment. For this reason, we considered a model in which the order parameter interacts with a large number of environmental fields, including its own short-wavelengths. Assuming weak coupling and high critical temperature, we have shown that decoherence is a short time event, shorter than the spinodal time $t_{\rm sp}$, which is essentially that time by which the order parameter field has sampled the degenerate ground states. As a result, perturbative calculations are justified[@lomplb]. Subsequent dynamics can be described by a stochastic Langevin equation, the details of which are only known for early times. Our approach in Ref.[@lomplb] has some connections with well-established classical behaviour of thermal scalar field theory [@htl] at high temperature. In many articles it has been shown that, at high temperatures, the behaviour of long-wavelength modes is determined by classical statistical field theory. The effective classical theory is obtained after integrating out the hard modes with $k\geq gT$. Although similar, this approach has some important differences from ours: the “classical behaviour” in this soft thermal mode analysis is defined through the coincidence of the quantum and the statistical correlation functions. In particular, thermal equilibrium is assumed to hold at all times. Finally, the cutoff that divides system and environment depends on the temperature, which is externally fixed. In our approach, the quantum to classical transition is defined by the diagonalization of the reduced density matrix, which is not assumed to be thermal. In phase transitions the separation between long and short-wavelengths is determined by their stability, which depends on the parameters of the potential. In Ref.[@lomplb] we only considered the case of an instantaneous quench, and bi-quadratic coupling between the system and the environment. In Ref.[@lomplb2] we began to extend those results to the case of a slow quench. In this paper we provide the details of that analysis and extend it to other couplings between system and environment. The consideration of slow quenches is very important since the Kibble-Zurek mechanism predicts the relation between the subsequent domain structure and the quench time [@kzm; @Rivers2; @laguna; @zurek3] (by indirectly counting defects). The paper is organized as follows. In Section 2 we introduce our models. These are theories containing a real system field $\phi$, which undergoes a transition, coupled to other scalar fields $\chi_{\rm a}$ (${\rm a} = 1,2,...,N$), which constitute the external part of the environment. Gauge fields bring their own specific difficulties and we shall not discuss them here. We compute the influence functional by integrating out the environmental fields for different couplings. Section 3 is dedicated to reviewing the evaluation of the master equation and the diffusion coefficients which are relevant in order to study decoherence. In Section 4 we evaluate upper bounds on the decoherence time for slow quenches. As we will see, provided quenches are not to slow, decoherence takes place before the field samples the minimum of the potential, i.e. decoherence time is typically shorter than the spinodal time. However, since all relevant timescales depend only logarithmically upon the parameters of the theory, it is necessary to keep track of $O(1)$ prefactors carefully, something that we rather took for granted in Ref.[@lomplb] and in Ref.[@lomplb2]. We will also show that the bi-quadratic coupling is the most relevant for the quantum to classical transition. Section 5 contains the conclusions of our work. Two short appendices fill in some of the detail. The Influence Action for an External Environment {#sec:model} ================================================ For the infinite degree of freedom quantum field $\phi$ undergoing a continuous transition, the field ordering after the transition begins is due to the growth in amplitude of its unstable long-wavelength modes. For these modes the environment consists of the short-wavelength modes of the field, together with all the other fields with which $\phi$ inevitably interacts[@huzhang; @lombmazz; @greiner] in the absence of selection rules. The inclusion of explicit environment fields is both a reflection of the fact that a scalar field in isolation is physically unrealistic, as well as providing us with a systematic approximation scheme. The $\phi$-field describes the scalar order parameter, whose ${\cal Z}_2$ symmetry is broken by a double-well potential. Specifically, we take the simplest classical action with scalar and environmental fields $\chi_{\rm a}$ $$S[\phi , \chi ] = S_{\rm syst}[\phi ] + S_{\rm env}[\chi ] + S_{\rm int}[\chi_{\rm a},\phi ], \label{action0}$$ where (with $\mu^2$, $m^2 >0$ ) $$S_{\rm syst}[\phi ] = \int d^4x\left\{ {1\over{2}}\partial_{\mu} \phi\partial^{\mu} \phi + {1\over{2}}\mu^2 \phi^2 - {\lambda\over{4!}}\phi^4\right\}, \nonumber$$ $$S_{\rm env}[\chi_{\rm a} ] = \sum_{\rm a=1}^N\int d^4x\left\{ {1\over{2}}\partial_{\mu}\chi_{\rm a} \partial^{\mu} \chi_{\rm a} - {1\over{2}} m_{\rm a}^2 \chi^2_{\rm a}\right\}. \nonumber$$ The most important interactions will turn out to be of the biquadratic form $$S_{\rm int}[\chi_{\rm a},\phi ] = S_{\rm qu}[\phi ,\chi ] = - \sum_{\rm a=1}^N\frac{g_{\rm a}}{8} \int d^4x ~ \phi^{\rm 2} (x) \chi^{\rm 2}_{\rm a} (x). \label{Sint}$$ Even if there were no external $\chi$ fields with a quadratic interaction of kind (\[Sint\]), the interaction between short and long-wavelength modes of the $\phi$-field can be recast, in part, in this form (see later), showing that such a term is obligatory. The generalization to a complex field $\phi$ is straightforward, and has been considered elsewhere [@lomplb2]. Later, we shall consider additional interactions to the biquadratic interaction of a bilinear form $$S_{\rm bilin}[\phi ,\chi ] = - \sum_{\rm a=1}^N\frac{g'_{\rm a}\mu}{4} \int d^4x ~ \phi^{\rm 2} (x) \chi_{\rm a} (x), \label{Sint2}$$ in which the environment couples linearly. We could also have included fermionic Yukawa interactions but these provide a smaller contribution to the diffusion constant because of Fermi statistics (which gives a diffusion coefficient relatively $O(\mu^2/T^{2})$). However, because the biquadratic self-interaction is the overwhelming term in the diffusion coefficient, in general we are not interested in effects that are overshadowed by it, and we shall not consider Yukawa interactions here. Our exception to this rule of ignoring small contributions is in the the inclusion of linear couplings of the form $$S_{\rm lin}[\phi ,\chi_{\rm a} ] = - \sum_{\rm a=1}^N\frac{g''_{\rm a}\mu^2}{4} \int d^4x ~ \phi (x) \chi_{\rm a} (x), \label{Sint3}$$ in which the order parameter field couples linearly to a linear environment. Much if not most of the work on decoherence has been for linear coupling to the environment of this type (e.g. see [@unruh; @kim]). While it is sensible in quantum mechanics, in the context of quantum field theory such linear couplings signal an inappropriate field diagonalisation, although they are exactly solvable in some circumstances [@kim]. Although the system field $\phi$ can never avoid the decohering environment of its own short-wavelength modes, to demonstrate the effect of an environment we first consider the case in which the environment is taken to be composed only of the fields $\chi_{\rm a}$. We are helped in this by the fact that environments have a cumulative effect on the onset of classical behaviour. That is, the inclusion of a further component of the environment [*reduces*]{} the time $t_D$ it takes for the system to behave classically (for the definition of $t_D$ see later). Thus it makes sense to include the environment one part after another, since we can derive an [*upper*]{} bound on $t_D$ at each step. The short-wavelength modes of the $\phi$ field will be considered last. To keep our calculations tractable, we need a significant part of the environment to have a strong impact upon the system-field, but not vice-versa, from which we can bound $t_D$. The simplest way to implement this is to take a large number $N\gg 1$ of scalar $\chi_{\rm a}$ fields with comparable masses $m_{\rm a}\simeq \mu$ weakly coupled to the $\phi$, with $\lambda$, $g_{\rm a},g'_{\rm a},g''_{\rm a}\ll 1$. Thus, at any step, there are $N$ weakly coupled environmental fields influencing the system field, but only one weakly self-coupled system field to back-react upon the explicit environment. We first consider the case in which the fields $\chi_{\rm a}$ interact through the biquadratic interaction (\[Sint\]) alone. For one-loop consistency at second order in our calculation of the diffusion coefficient (that enforces classicality) it is sufficient, at order of magnitude level, to take identical $g_{\rm a} = g/\sqrt{N}$. Further, at the same order of magnitude level, we take $g\simeq\lambda$. This is very different from the more usual large-N $O(N+1)$-invariant theory with one $\phi$-field and $N$ $\chi_{\rm a}$ fields, dominated by the $O(1/N)$ $(\chi^2)^2$ interactions, that has been the standard way to proceed for a [*closed*]{} system. With our choice there are no direct $\chi^4$ interactions, and the indirect ones, mediated by $\phi$ loops, are depressed by a factor $g/\sqrt{N}$. In this way the effect of the external environment qualitatively matches the effect of the internal environment provided by the short-wavelength modes of the $\phi$-field, but in a more calculable way. For small $g$ the model has a continuous transition at a temperature $T_{\rm c}$. The environmental fields $\chi_{\rm a}$ reduce $T_{\rm c}$ and, in order that $T_{\rm c}^2/\mu^2=24/(\lambda + \sum g_{\rm a})\gg 1$, we must take $\lambda + \sum g_{\rm a} \ll 1$, whereby $1\gg 1/\sqrt N\gg g$. Further, with this choice the dominant hard loop contribution of the $\phi$-field to the $\chi_{\rm a}$ thermal masses is $$\delta m^2_T = O (g T^2_{\rm c}/\sqrt{ N}) = O(\mu^2/N)\ll \mu^2.$$ Similarly, the two-loop (setting sun) diagram which is the first to contribute to the discontinuity of the $\chi$-field propagator is of magnitude $$g^2 T_{\rm c}^2/N = O(g\mu^2/N^{3/2})\ll\delta m^2_T,$$ in turn. That is, the effect of the thermal bath on the propagation of the environmental $\chi$-fields is ignorable. This was our intention; to construct an environment that reacted on the system field, but was not reacted upon by it to any significant extent. In particular, the infinite $N$ limit does not exist. Dependence on $N$ is implicit through $T_{\rm c}$ as well as through the couplings, for initial temperatures $T_0 = O(T_{\rm c})$. With $\eta = \sqrt{6\mu^2/\lambda}$ determining the position of the minima of the potential and the final value of the order parameter, this choice of coupling and environments gives the hierarchy of scales necessary for establishing a reliable approximation scheme, as has been shown in [@lomplb]. We shall assume that the initial states of the system and environment are both thermal, at a high temperature $T_{0}>T_{\rm c}$. We then imagine a change in the global environment (e.g. expansion in the early universe) that can be characterised by a change in temperature from $T_{0}$ to $T_{\rm f}<T_{\rm c}$. That is, we do not attribute the transition to the effects of the environment-fields. Given our thermal initial conditions it is not the case that the full density matrix has $\phi$ and $\chi$ fields uncorrelated initially, since it is the interactions between them that leads to the restoration of symmetry at high temperatures. Rather, on incorporating the hard thermal loop ’tadpole’ diagrams of the $\chi$ (and $\phi$) fields in the $\phi$ mass term leads to the effective action for $\phi$ quasiparticles, $$S^{\rm eff}_{\rm syst}[\phi ] = \int d^4x\left\{ {1\over{2}}\partial_{\mu} \phi\partial^{\mu} \phi - {1\over{2}} m_{\phi}^2(T_0) \phi^2 - {\lambda\over{4!}}\phi^4\right\} \label{Stherm}$$ where $m_{\phi}^2(T_0)\propto (1-T_0/T_{\rm c})$ for $T\approx T_{\rm c}$. As a result, we can take an initial factorised density matrix at temperature $T_0$ of the form ${\hat\rho}[T_0] = {\hat\rho}_{\phi}[T_0] \times {\hat\rho}_{\chi}[T_0]$, where ${\hat\rho}_{\phi}[T_0]$ is determined by the quadratic part of $S^{\rm eff}_{\rm syst}[\phi ]$ and ${\hat\rho}_{\chi}[T_0]$ by $S_{\rm env}[\chi_{\rm a} ]$. That is, the many $\chi_{\rm a}$ fields have a large effect on $\phi$, but the $\phi$-field has negligible effect on the $\chi_{\rm a}$. Provided the change in temperature is not too slow the exponential instabilities of the $\phi$-field grow so fast that the field has populated the degenerate vacua well before the temperature has dropped significantly below $T_{\rm c}$. Since the temperature $T_{\rm c}$ has no particular significance for the environment fields, for these early times we can keep the temperature of the environment fixed at $T_{\chi}\approx T_{\rm c}$ (our calculations are only at the level of orders of magnitude). Meanwhile, for simplicity the $\chi_{\rm a}$ masses are fixed at the common value $ m\simeq\mu$. Our interests in the transitions in the early universe (domain structure, defect formation and, indirectly, structure formation) and the subsequent field evolution [@vilen] are naturally couched in terms of fields, rather than particles. Since we need to be able to distinguish between different classical system-field configurations evolving after the transition, we will only be interested in the field-configuration basis for this reduced density matrix (in analogy with the usual quantum Brownian motion model, e.g. [@zurekTA]). The resulting reduced density matrix $\rho_{{\rm r}}[\phi^+,\phi^-,t]=\langle\phi^+ \vert {\hat\rho}_{\rm r} (t)\vert \phi^- \rangle$ describes the evolution of the system under the influence of the environment, and is defined by $$\rho_{{\rm r}}[\phi^+,\phi^-,t] = \int \prod_{{\rm a}=1}^N {\cal D} \chi_{\rm a} ~ \rho[\phi^+,\chi_{\rm a} ,\phi^-,\chi_{\rm a} ,t],$$ where $\rho[\phi^+,\chi^+_{\rm a} ,\phi^-,\chi^-_{\rm a},t]= \langle\phi^+ \chi^+_{\rm a}\vert {\hat\rho}(t) \vert \phi^- \chi^-_{\rm a}\rangle$ is the full density matrix. For the reasons given above, the environment will have had the effect of making the system effectively classical once $\rho_{\rm r}(t)$ is, approximately, diagonal. Quantum interference can then be ignored and we obtain a classical probability distribution from the diagonal part of $\rho_{\rm r}(t)$, or equivalently, by means of the reduced Wigner functional, which is positive definite after $t_D$ [@diana]. For weak coupling there will be no ’recoherence’ at later times in which the sense of classical probability will be lost [@nuno]. The temporal evolution of $\rho_{\rm r}$ is given by $$\rho_{\rm r}[\phi_{{\rm f}}^+,\phi_{{\rm f}}^-,t]= \int d\phi_{{\rm i}}^+ \int d\phi_{{\rm i}}^- J_{\rm r} [\phi_{{\rm f}}^+,\phi_{{\rm f}}^-,t\vert \phi_{{\rm i}}^+,\phi_{{\rm i}}^-,t_0] ~~ \rho_{\rm r}[\phi_{{\rm i}}^+ \phi_{{\rm i}}^-,t_0],$$ where $J_{\rm r}$ is the reduced evolution operator $$J_{\rm r}[\phi_{{\rm f}}^+,\phi_{{\rm f}}^-,t\vert \phi_{{\rm i}}^+, \phi_{{\rm i}}^-,t_0] = \int_{\phi_{{\rm i}}^+}^{\phi_{{\rm f}}^+} {\cal D}\phi^+ \int_{\phi_{{\rm i}}^-}^{\phi_{{\rm f}}^-} {\cal D}\phi^- ~e^{i\{S[\phi^+] - S[\phi^-]\}} F[\phi^+,\phi^-]. \label{evolred}$$ The Feynman-Vernon [@feynver] influence functional (IF) $F[\phi^+, \phi^-]$ in (\[evolred\]) can be written in terms of influence action $\delta A[\phi^+,\phi^-]$ and the coarse grained effective action (CGEA) $A[\phi^+,\phi^-]$ by (see [@lombmazz] for formal definitions) $$F[\phi^+,\phi^-] = \exp {i \delta A[\phi^+,\phi^-]},\label{IA}$$ $$A[\phi^+,\phi^-] = S[\phi^+] - S[\phi^-] + \delta A[\phi^+,\phi^-]. \label{CTPEA}$$ Beginning from the initial distribution, peaked around $\phi = 0$, we follow the evolution of the system under the influence of the environment fields. We will calculate the influence action to lowest non-trivial order (two vertices) for large $N$. Similarly to the biquadratic interaction, we assume weak coupling $\lambda \simeq g'\simeq g''\ll 1$, where we have defined $g', g''$ by the order of magnitude relations $g'_{\rm a}= g'/\sqrt{N}$ and $g''_{\rm a}= g''/\sqrt{N}$ respectively. As we are considering weak coupling with the environment fields, we may expand the influence functional $F[\phi^+, \phi^-]$ up to second non-trivial order in coupling strengths. The general form of the influence action is then [@lombmazz; @calhumaz] $$\begin{aligned} \delta A[\phi^+,\phi^-] = &&\{\langle S_{\rm int}[\phi^+,\chi^+_{\rm a}]\rangle_0 - \langle S_{\rm int}[\phi^-,\chi^-_{\rm a}] \rangle_0\} \nonumber \\ &&+{i\over{2}}\{\langle S_{\rm int}^2[\phi^+,\chi^+_{\rm a}]\rangle_0 - \big[\langle S_{\rm int}[\phi^+,\chi^+_{\rm a}]\rangle_0\big]^2\}\nonumber \\ &&- i\{\langle S_{\rm int}[\phi^+,\chi^+_{\rm a}] S_{\rm int}[\phi^-,\chi^- _{\rm a}] \rangle_0 - \langle S_{\rm int}[\phi^+,\chi^+_{\rm a}]\rangle_0\langle S_{\rm int}[\phi^-, \chi^-_{\rm a}]\rangle_0\} \label{inflac} \\ &&+{i\over{2}}\{\langle S^2_{\rm int}[\phi^-,\chi^-_{\rm a}]\rangle_0 - \big[\langle S_{\rm int}[\phi^-,\chi^-_{\rm a}]\rangle_0\big]^2\}.\nonumber \end{aligned}$$ For weak couplings it is relatively easy to compute the upper bound on the decoherence time due to the interactions $S_{\rm int}[\phi ,\chi ]$ of (\[Sint\]), (\[Sint2\]) and (\[Sint3\]). We shall find that, in our particular model, it is in general shorter than the spinodal time $t_{\rm sp}$, defined as the time for which $$\langle \phi^2\rangle_{t_{\rm sp}} \sim \eta^2= 6\mu^2/\lambda\,\,\, .\label{deftsp}$$ \[More exactly, $t_{\rm sp}$ should be defined as the time that the r.m.s of the field takes to reach the spinodal point, the point of inflection in the potential, but the difference is only logarithmically small, so we stay with (\[deftsp\]).\] It is for this reason that, with the qualifications below, we can use perturbation theory. In consequence, by the time that the field is ordered it can be taken to be classical. This has implications[@lomplb2] for the formation of the defects that are a necessary byproduct of transitions. Master Equations and the Diffusion Coefficients {#sec:diff} =============================================== In this Section we will obtain the evolution equation for the reduced density matrix (master equation), paying particular attention to the diffusion term, which is responsible for decoherence. We will closely follow the quantum Brownian motion (QBM) example [@unruh; @qbm], translated into quantum field theory [@lomplb; @lombmazz]. This Section contains a generalization of the result already given in [@lomplb], adapted to the slow quenches that are physically relevant, in which $m^{2}_{\phi}(T)$ varies linearly in time. The formal deduction of the master equation is not altered by the duration of the quench. For the details we will follow Ref.[@lombmazz]. The first step in the evaluation of the master equation is the calculation of the density matrix propagator $J_{\rm r}$ from Eq.(\[evolred\]). In order to solve the functional integration which defines the reduced propagator, we perform a saddle point approximation $$J_{\rm r}[\phi^+_{\rm f},\phi^-_{\rm f},t\vert\phi^+_{\rm i},\phi^-_{\rm i}, t_0] \approx \exp{ i A[\phi^+_{\rm cl},\phi^-_{\rm cl}]}, \label{prosadle}$$ where $\phi^\pm_{\rm cl}$ is the solution of the equation of motion ${\delta Re A\over\delta\phi^+}\vert_{\phi^+=\phi^-}=0$ with boundary conditions $\phi^\pm_{\rm cl}(t_0)=\phi^\pm_{\rm i}$ and $\phi^\pm_{\rm cl}(t)=\phi^\pm_{\rm f}$. Even then, it is very difficult to solve this equation analytically. We are helped by the observation that the ordering of the field is due to the growth of the long-wavelength unstable modes. Unstable long-wavelength modes start growing exponentially as soon as the quench is performed, whereas short-wavelength modes will oscillate. As a result, the field correlation function rapidly develops a peak (Bragg peak) at wavenumber $k = {\bar k}\ll \mu$. Specifically [@Karra], initially as ${\bar k}^2 = {\cal O}(\mu/\sqrt{t\tau_{\rm q}})$, where $\tau^{-1}_{\rm q}$ is the quench rate. Assuming that a classical description can be justified post-hoc, a domain structure forms quickly with a characteristic domain size $O({\bar k}^{-1})$, determined from the position of this peak. \[As an example, see the numerical results of [@laguna], where this classical behaviour has been assumed through the use of stochastic equations (see later)\]. With this in mind, we adopt an approximation in which the system-field contains only one Fourier mode with $\vec k = \vec k_0 = O({\bar k}^{-1})$, characteristic of the domain size. As “trial” classical solutions, we take $$\phi_{\rm cl}(\vec x, s) = f(s,t)\cos(k_0 x)\cos(k_0 y)\cos(k_0 z), \label{classsol}$$ where $f(s,t)$ satisfies $f(0,t)= \phi_{\rm i}$ and $f(t,t) = \phi_{\rm f}$. Eq. (\[classsol\]) is an exemplary configuration that mimics domain formation. As intended, it represents domains of finite size $k_0^{-1}$. Although such a regular domain structure is idealised, numerical results (as in [@laguna]) suggest that it is a reasonable first step. In practice, after this detailed motivation, we shall find that the decoherence time $t_D$ is insensitive to wavelength for $k_0$ small. For the purpose of calculation, it is sufficient to set $k_0 = 0$, even though the physical situation requires a finite $k_0$. An exact matching of $k_0$ to ${\bar k}$ is unnecessary. \[We note that, in [@lomplb] and [@lomplb2], we adopted a simpler form $\phi_{\rm cl}(\vec x, s) = f(s,t)\cos({\vec k}_0.{\vec x}) $. Although showing a domain structure in one dimension is less physical, our conclusions are unchanged although details differ.\] We write $f(s,t)$ as $$f(s,t) = \phi_{\rm i} u_1(s,t) + \phi_{\rm f} u_2(s,t),\,\, \label{us}$$ where, during the quench, $u_i(s,t)$ are solutions of the mode equation with boundary conditions $u_1(0,t) = 1$, $u_1(t,t) = 0$ and $u_2(0,t) = 0$, $u_2(t,t) = 1$. In order to obtain the master equation we must compute the final time derivative of the propagator $J_{\rm r}$. After that, all the dependence on the initial field configurations $\phi^\pm_{\rm i}$ (coming from the classical solutions $\phi^\pm_{\rm cl}$) must be eliminated. The free propagator, defined as $$J_0[\phi^+_{\rm f}, \phi^-_{\rm f}, t\vert \phi^+_{\rm i}, \phi^-_{\rm i}, 0] = \int_{\phi^+_{\rm i}}^{\phi^+_{\rm f}}{\cal D}\phi^+ \int_{\phi^-_{\rm i}} ^{\phi^-_{\rm f}} {\cal D}\phi^- \exp\{i [ S_0(\phi^+) - S_0(\phi^-)]\}; \label{propdeJ0}$$ satisfies the general identities [@lombmazz; @qbm] (which is valid for instantaneous and for slow quenches) $$\phi^\pm_{\rm cl}(s) J_0 = \Big[ \phi^\pm_{\rm f} [u_2(s,t) - \frac{{\dot u}_2(t,t)}{{\dot u}_1(t,t)}u_1(s,t)] \mp i {u_1(s,t)\over{{\dot u}_1(t,t)}} \partial_{\phi^\pm_{\rm f}}\Big]J_0. \label{rel1}$$ These identities allow us to remove the initial field configurations $\phi^\pm_{\rm i}$, by expressing them in terms of the final fields $\phi^\pm_{\rm f}$ and the derivatives $\partial_{\phi^\pm_{\rm f}}$, and obtain the master equation. The full equation is very complicated and, as for quantum Brownian motion, it depends on the system-environment coupling. In what follows we will compute the diffusion coefficients for the different couplings described in the previous Section. As we are solely interested in decoherence, it is sufficient to calculate the correction to the usual unitary evolution coming from the noise kernels (imaginary part of the influence action). The result reads $$i {\dot \rho}_{\rm r} = \langle \phi^+_{\rm f}\vert [H,\rho_{\rm r}] \vert \phi^-_{\rm f}\rangle - i V \sum_j \Gamma_j D_{j}(\omega_0, t) \rho_{\rm r}+ ... \label{master}$$ where $j$ labels the different couplings. $D_j$ are the diffusion coefficients. The ellipsis denotes other terms coming from the time derivative that do not contribute to the diffusive effects. $V$ is understood as the minimal volume inside which there are no coherent superpositions of macroscopically distinguishable states for the field. The $\Gamma_j$, which depend on the sums and differences $\phi^+_{\rm f}\pm\phi^-_{\rm f}$ of the field amplitudes, have to be calculated case by case. The effect of the diffusion coefficient on the decoherence process can be seen considering the following approximate solution to the master equation $$\rho_{\rm r}[\phi^+, \phi^-; t]\approx \rho^{\rm u}_{\rm r}[\phi^+, \phi^-; t] ~ \exp \left[-V\sum_j \Gamma_{\rm j} \int_0^t ds ~D_{\rm j}(k_0, s) \right],$$ where $\rho^{\rm u}_{\rm r}$ is the solution of the unitary part of the master equation (i.e. without environment). The system will decohere when the non-diagonal elements of the reduced density matrix are much smaller than the diagonal ones. This decoherence time $t_{D}$ sets the scale after which we have a classical system-field configuration, and depends strongly on the properties of the environment. It is constrained by $$1 \approx V \sum_j\Gamma_{\rm j} \int_{0}^{t_{\rm D}} ds ~D_{\rm j}(k_0,s) \gtrsim V \Gamma_{\rm l} \int_{0}^{t_{\rm D}} ds ~D_{\rm l}(k_0,s), \label{Dsum}$$ for any particular $j=l$, and corresponds to the time after which we are able to distinguish between two different field amplitudes, inside a given volume $V$. Our conservative choice is that this volume factor is ${\cal O}(\mu^{-3})$ since $\mu^{-1}$ (the Compton wavelength) is the smallest scale at which we need to look. In the next Section we will compute the corresponding diffusion coefficients, from which we will estimate the decoherence time bounds. Decoherence Times for Slow Quenches {#sec:tdec} =================================== Our earliest results from Ref.[@lomplb] were for instantaneous quenches, for which the quench time is effectively $\mu^{-1}\ll t_{\rm sp}$. [*A priori*]{}, it was not obvious that this slow quench preserves these results about the decoherence time, that $t_D\lesssim t_{\rm sp}$. This is assumed in the Kibble scenario [@Kibble] for domain formation as determined by the appearance of defects, in which classical defects are assumed to have appeared as soon as the transition is effected. This was first addressed by us in a rather schematic way in [@lomplb2], to which what follows provides a more sophisticated analysis and extensions to other couplings. To tackle this problem, we assume that the quench begins at $t=0$ and ends at time $t = 2\tau_{\rm q}$, with $\tau_{\rm q}\gg t_{\rm r}\sim \mu^{-1}$. At the qualitative level at which we are working it is sufficient to take $m_{\phi}^2(T_0) = \mu^2$ exactly. Most simply, we consider a quench linear in time, with temperature $T(t)$, for which the mass function is of the following form [@bowick] $$m^2(t) = m_{\phi}^2(T(t)) = \left\{ \matrix{ \mu^2&\mbox{for} ~ t \le 0\cr \mu^2 - {t\mu^2\over{\tau_{\rm q}}}& ~~~~~~~~\mbox{for} ~ 0 < t \le 2\tau_{\rm q}\cr - \mu^2&~\mbox{for} ~ t \ge 2\tau_{\rm q}\cr} \right.$$ Note that our $\tau_{\rm q}$ is the inverse quench rate $T_{\rm c}^{-1}dT/dt|_{T=T_{\rm c}}$, and so differs from that of [@bowick] by a factor of 2. The field behaves as a free field in an inverted parabolic potential for an interval of approximately $t_{\rm sp}$ [@Karra], where $$\langle \phi^2\rangle_{t_{\rm sp}} \sim \eta^2\,\,\, .\label{deftsp2}$$ We have rederived $t_{\rm sp}$, using the exact results of [@bowick] (see Appendix). The end result is that $t_{\rm sp}$ is the solution to $${\eta^2\over C}\simeq {T_{\rm c}\over \mu{\bar t}^2}\exp\bigg[{4\over 3}\bigg({\Delta_0(t_{\rm sp})\over{\bar t}}\bigg)^{3/2}\bigg], \label{soltsp}$$ where $\Delta_0(t) = t - \tau_{\rm q}$, $C = (64 \sqrt{2}\pi^{3/2})^{-1}$, and ${\bar t}= (\tau_{\rm q}/\mu^2)^{1/3}$ (see details in the Appendix A). In order to find the master equation (or more strictly the diffusion coefficient) we follow the same procedure as in Section III, assuming a dominant wavenumber $k_0$. The classical solution is given in Eq. (\[us\]), where $u_i$ satisfy the mode equation $$\left[{d^2\over{ds^2}} + k_0^2 + \mu^2 - {\mu^2 s\over{\tau_{\rm q}}}\right] u_i(s,t) = 0\label{modeeqn}.$$ Since we are neglecting the self-interaction term, our conclusions are only believable for $t\lesssim t_{\rm sp}$. The solution of Eq.(\[modeeqn\]) is given by $$u_1(s,t) = {-Ai[{\Delta_k (t)\over{{\bar t}}}] Bi[{\Delta_k (s) \over{{\bar t}}}] + Ai[{\Delta_k (s)\over{{\bar t}}}] Bi[{\Delta_k (t) \over{{\bar t}}}]\over{Ai[-\omega^2 {\bar t}^2] Bi[{\Delta_k (t) \over{{\bar t}}}] - Ai[{\Delta_k (t)\over{{\bar t}}}] Bi[-\omega_0^2 {\bar t}^2]}},$$ $$u_2(s,t)= {-Ai[{\Delta_k (s)\over{{\bar t}}}] Bi[-\omega_0^2 {\bar t}^2] + Ai[-\omega_0^2 {\bar t}^2] Bi[{\Delta_k (s) \over{{\bar t}}}] \over{Ai[-\omega_0^2 {\bar t}^2] Bi[{\Delta_k (t) \over{{\bar t}}}] - Ai[{\Delta_k (t)\over{{\bar t}}}] Bi[-\omega_0^2 {\bar t}^2]}} ,$$ where $Ai[s]$ and $Bi[s]$ are the Airy functions, with $\Delta_k (s) = s - \omega_0^2{\bar t}^3 $, $\Delta_k (t) = t - \omega_0^2 {\bar t}^3 $, and $\omega_0^2 = \mu^2 + k_0^2$. We note that $\Delta_0 (t)= t - \tau_{\rm q}$. In the causal analysis of Kibble[@Kibble] $\bar t$ ($\mu^{-1}\ll \bar t \ll \tau_{\rm q}$) is the time at which the adiabatic field correlation length collapses at the speed of light, the earliest time in which domains could have formed. Our analysis suggests that this earliest time is not ${\bar t}$, but $t_{\rm sp}$. Quadratic coupling ------------------ We start by considering the biquadratic coupling first, for which the IF is given by $${\rm Re} \delta A = \frac{g^2}{8} \int d^4 x\int d^4y ~ \Delta_2 (x) K_{\rm q}(x-y) \Sigma_2 (y),$$ $${\rm Im} \delta A = - \frac{g^2}{16} \int d^4x\int d^4y ~ \Delta_2 (x) N_{\rm q} (x,y) \Delta_2 (y),$$ where $K_{\rm q} (x-y) = {\rm Im} G_{++}^2(x,y) \theta (y^0-x^0)$ is the dissipation kernel and $ N_{\rm q} (x-y) = {\rm Re} G_{++}^2(x,y)$ is the noise (diffusion) kernel. $G_{++}$ is the relevant closed-time-path correlator of the $\chi$-field at temperature $T_0$. We have defined $$\Delta_2 ={1\over{2}}(\phi^{+2} - \phi^{-2}) ~~~;~~~ \Sigma_2 ={1\over{2}} (\phi^{+2} + \phi^{-2}).$$ We can formally find the master equation in the same way that we did in Section III. It is given by Eq.(\[master\]), with a diffusion coefficient of the form (details on how to get this coefficient can be found in Appendix B) $$D_{\rm qu}(k_0,t) = \int_0^t ~ ds ~ u(s,t)~ F(k_0,s,t),\label{dqsq}$$ with $$u(s,t) =\bigg[u_2(s,t) - \frac{{\dot u}_2(t,t)}{{\dot u}_1(t,t)}u_1(s,t)\bigg]^{2},$$ and $$F(k_0,s,t) = {1\over{64}}\left\{{\rm Re}G_{++}^2(0; t-s) + {3\over{2}}{\rm Re} G_{++}^2(2k_0; t-s) + {3\over{4}}{\rm Re}G_{++}^2(2\sqrt{2}k_0; t-s) + {1\over{8}}{\rm Re}G_{++}^2(2\sqrt{3}k_0; t-s) \right\}.$$ It is only in $u(s,t)$ that the slow quench is apparent. $G^2_{++}(k, t-s)$ is the Fourier transform of the square of the Feynman propagator ($\chi$ propagator). In the high temperature limit ($T \gg \mu$), the explicit expression for the kernels can be shown, with a little labour, to be $${\rm Re}G_{++}^2(k; t-s)={T_{\rm c}^2\over{64\pi^2}} {1\over{k}} \int_0^\infty dp~{p\over{p^2 + \mu^2}} \int_{\sqrt{\vert p - k\vert^2 + \mu^2}}^{\sqrt {\vert p + k\vert^2 + \mu^2}}{du\over{u}} ~\cos[(\sqrt{p^2 + \mu^2}+ u) (t-s)], \label{D1}$$ and $${\rm Re}G_{++}^2(0; t-s)={T_{\rm c}^2\over{64\pi^2}} ~ \int_0^\infty dp~{p^2\over{(p^2 + \mu^2)^2}} \cos[2\sqrt{p^2 + \mu^2}(t-s)], \label{D}$$ where $\mu$ is the thermal $\chi$-field mass at temperature $T\sim T_{\rm c}$. In our scheme, this is approximately the cold $\chi$ mass. It is because the $\chi$-field propagator is unaffected by the $\phi$-field interactions that the detail of the expressions (\[D1\])-(\[D\]) are possible. Fortunately, it is inessential. We see that, for times $\mu t\gtrsim 1$, the behaviour of $D_{\rm qu}(k_0, t)$ is dominated by the exponential growth of $u(s,t)$, and the integral in Eq.(\[dqsq\]) by the interval $s\approx 0$. Indeed, it is easy to prove that $u(0,t) = ({{\dot u}_2(t,t)\over{{\dot u}_1(t,t)}})^2 \gg 1$, while $u(t,t) = 1$. Therefore, although the expessions in Eqs.(\[dqsq\]), (\[D1\]), and (\[D\]) are complicated, we can approximate the whole diffusion coefficient by $$D_{\rm qu}(k_0, t) \approx F(k_0, 0,t) ~ u(0,t) ~ \int_0^{\omega_0^{-1}} ds ~ {u(s,t)\over{u(0,t)}} \label{D0}$$ where we used the fact that $F(k_0, s,t)$ is bounded at $s=0$. We will assume large $\Delta_k (t)$ (and $\Delta_k (s)$), which means $\Delta_k (t), \Delta_k (s) \gg {\bar t}$. This condition is satisfied provided $s$ is larger and not to close to $\omega_0^2 \tau_{\rm q}/\mu^2$, and allow us to use the asymptotic expansions of the Airy functions and their derivatives for the evaluation of $u_i(s,t)$. This will be justified posthoc. In particular we obtain $$\begin{aligned} u(0,t)&\simeq&{1\over 4 \omega_0 {\bar t}}\sqrt{\Delta_k (t)\over {\bar t}} \exp{\left\{ {4\over{3}}\left({\Delta_k (t)\over{{\bar t}}}\right)^{3\over{2}} \right\}}\nonumber\\ \dot u(0,t)&\simeq&{1\over 2 \omega_0^2{\bar t}^3}\sqrt{\Delta_k (t) \over {\bar t}} \exp{\left\{ {4\over{3}}\left({\Delta_k (t)\over{{\bar t}}}\right)^{3\over{2}} \right\}}\label{uupunto}.\end{aligned}$$ Therefore, it is straightforward to check that in these cases the integral in Eq.(\[D0\]) is given by $$\omega_0 ~\int_0^{\omega_0^{-1}}ds \left[1 + s {{\dot u}(0,t) \over{u(0,t)}} + ...~ \right] \approx \omega_0 ~\int_0^{\omega_0^{-1}}ds \left[ 1 + s {2\over{\omega_0 {\bar t}^2}} + ... \right] ,$$ and it is ${\cal O}(1)$, due to the fact that $\mu {\bar t} \gg 1$. We can estimate the decoherence time for the quadratic coupling as $$1 \gtrsim V\Gamma\int_0^{t_{D}} dt ~ D_{\rm qu}(k_0,t),$$ where $V$ is the decoherence volume, as before. In terms of the dimensionless fields $\bar\phi = (\phi^+ + \phi^-)/2\mu,$ and $ \delta = (\phi^+ - \phi^-)/2\mu$ it follows that $\Gamma_{\rm qu} = g^2\mu^4 \bar\phi^2 \delta^2$. In order to quantify the decoherence time we have to fix the values of $\delta$, and $\bar\phi$. Inside the volume $V$ we do not discriminate between field amplitudes which differ by $ {\cal O}(\mu) $, and therefore, for the sake of argument, we take $\delta = 1$. For $\bar\phi$ we set $\bar\phi^2 =\alpha /\lambda$, where $\lambda\leq\alpha\leq 1$ is to be determined self-consistently. It is necessary to be as precise as this in the first instance, since $t_{\rm sp}$ and $t_D$ only differ by logarithms. It is important that the prefactors in the arguments of these logarithms are determined carefully, given that terms nominally ${\cal O}(1)$ can be small. Once we have established that the difference is large enough, this artificial precision can be dropped. Posthoc it is sufficient to take $\delta \sim {\cal O}(1)$ and $\bar\phi^2\sim {\cal O}(\alpha /\lambda)$, as we did in [@lomplb] and [@lomplb2]. Then, for quadratic interactions alone the decoherence time reads $$\exp{\left\{ {4\over{3}}\left({\Delta_k (t_D)\over{\bar t }}\right)^{3\over{2}} \right\}} \approx 4 . 10^4{\omega_0^2\over{\lambda T_{\rm c}^2 \alpha}}.$$ Using $\langle\phi^2\rangle$ of (\[phi2\]) we obtain $\alpha^2 \approx 7 (\mu/T_{\rm c})(\mu \tau_{\rm q})^{-2/3}$. Thus the decoherence time satisfies $$\exp{\left\{ {4\over{3}}\left({\Delta_k (t_D)\over{\bar t }}\right)^{3\over{2}} \right\}} \approx 2 . 10^3 {\omega_0^2\over{\mu^2}}{\eta^2\over{T^{{3\over{2}}}_{\rm c}}} {(\mu\tau_{\rm q})^{{1\over{3}}}\over{\mu^{1\over{2}}}},$$ for long-wavelength modes. If we compare it with the spinodal time of (\[soltsp\]), we get, $$(\mu \Delta_0 (t_{\rm sp}))^{{3\over{2}}} - (\mu \Delta_k (t_{D}))^{{3\over{2}}} \approx {3\over{4}}~ \sqrt{\mu \tau_{\rm q}}~ \left\{2 \ln {\mu\over{\omega_0}} + \ln \left[{1\over{2}}\left({T_{\rm c}\over{\mu}}\right)^{1\over{2}} (\mu\tau_{\rm q})^{1\over{3}}\right]\right\} > 0. \label{ts2}$$ We see that the numerical prefactor in the argument of the logarithm is, indeed, ${\cal O}(1)$, and our concern for precision was unnecessary, in retrospect. Our approximation scheme depends, as for the instantaneous quench, on the peaking of the power in the field fluctuations at long-wavelength $k_0\ll\mu$ by time $t_{\rm sp}$, and it is sufficient to take $k_0\approx 0$. In this case (\[ts2\]) becomes $$(\mu \Delta_0(t_{\rm sp}))^{{3\over{2}}} - (\mu \Delta_0 (t_{D}))^{{3\over{2}}} \approx {3\over{4}}~ \sqrt{\mu \tau_{\rm q}}~ \ln \left[{1\over{2}} \left({T_{\rm c}\over{\mu}}\right)^{1\over{2}} (\mu\tau_{\rm q})^{1\over{3}}\right] > 0, \label{ts3}$$ from which $t_D < t_{\rm sp}$ follows, as in the instantaneous case. The inclusion of further interactions, including the self-interaction with short-wavelength ($k > \mu$) modes can only reduce $t_D$ further. We should point out that, whereas peaking is inevitable for the instantaneous quench for weak enough coupling, it is not the case for very slow quenches. In such cases our approximations break down and a different analysis is required. The details are rather messy, but a sufficient condition for our approximation to be valid is that $\mu\tau_{\rm q} \lesssim\eta/\mu$[@Karra]. Tighter, but less transparent bounds can be given[@Karra]. When these bounds are satisfied the minimum wavelength for which the modes decohere by time $t_{\rm sp}$ can be shown[@lomplb2] to be shorter than that which characterises domain size at that time. Although we can talk loosely, but sensibly, about a classical domain structure at time $t_{\rm sp}$ we cannot yet talk about classical defects on their boundaries, as the naive picture might suggest. Defects (in this case, walls) are described by shorter wavelength modes ($k\lesssim\mu$). Nonetheless, the classical domain structure is sufficient to determine their density [@lomplb2]. Bilinear and linear couplings ----------------------------- Another possibility that needs to be considered is that of a bilinear coupling of the $\phi$-field to the environment. This interaction preserves the $\phi\rightarrow -\phi$ symmetry of $ S_{\rm syst}[\phi ]$ in which, for simplicity we continue to take bilinear couplings equal, as $g_{\rm a}'=g'/\sqrt{N}$. We treat this interactions as an [*additional*]{} set of interactions to the biquadratic interactions, whereby $T_{\rm c}$ is qualitatively unchanged. The influence action is still obtained from Eq.(\[inflac\]), as $${\rm Re} \delta A_{\rm bilin} = {g'^{2}\mu^2\over{8}} \int d^4 x\int d^4y ~ \Delta_2 (x) K_{\rm b}(x-y) \Sigma_2 (y),$$ and $${\rm Im} \delta A_{\rm bilin} = - {g'^{2}\mu^2\over{16}} \int d^4x\int d^4y ~ \Delta_2 (x) N_{\rm b}(x-y) \Delta_2 (y),$$ where $$K_{\rm b} = {\rm Re}G_{++}(x,y)\theta (y^{0}-x^{0}) ,\,\,\,\,\,\,N_{\rm b} = {\rm Im}G_{++}(x,y).$$ The temporal diffusion coefficient is now given by $$D_{\rm bilin}(k_0, t) = \int_0^t ds ~ u(s,t) ~ F_{\rm bilin}(k_0, s, t), \label{diff2}$$ where $$F_{\rm bilin}(k_0, s, t) = {1\over{64}}\left\{{\rm Im}G_{++}(0; t-s) + {3\over{2}}{\rm Im} G_{++}(2k_0; t-s) + {3\over{4}}{\rm Im}G_{++}(2\sqrt{2}k_0; t-s) + {1\over{8}}{\rm Im}G_{++}(2\sqrt{3}k_0; t-s) \right\}.$$ with $${\rm Im} G_{++}(k; s) = \frac{T_{\rm c}}{4}\frac{\cos[\sqrt{k^{2} + \mu^2 }s]}{k^{2} + \mu^2}.$$ Following the same arguments as in the previous section, we can evaluate the diffusion coefficient by $$D_{\rm bilin}(k_0, t) \approx F_{\rm bilin}(k_0, 0, t){u(0, t) \over{\omega_0}},$$ where $u(0,t)$ is given in Eq.(\[uupunto\]). Thus, $$D_{\rm bilin}(k_0, t) \approx 10^{-2} \frac{T_{\rm c}}{\mu^2\omega_0} ~u(0, t).$$ For the unstable long-wavelengths ($k_0^2 < \mu^2/2$ approximately) we find, for times $\mu t\gtrsim 1$, that $D_{\rm bilin}(k_0, t)$ again shows the exponential growth $$D_{\rm bilin}(k_0, t) \approx \frac{3 . 10^{-3} T_{\rm c}}{{\bar t}\mu^2 \omega_0^2} \sqrt{{\Delta_k (t)\over{{\bar t}}}} \exp\left\{ {4\over{3}}\left( {\Delta_k (t)\over{{\bar t}}}\right)^{3\over{2}} \right\}. \label{Dyu2}$$ It follows that $\Gamma_{\rm bilin} = g'^2\mu^4 \bar\phi^2 \delta^2$. As a result, the contribution to the decoherence time from bilinear interaction (for long-wavelength modes) can be evaluated as $$\exp\left\{ {4\over{3}}\left( {\Delta_k (t_D)\over{{\bar t}}} \right)^{3\over{2}} \right\} \approx 6 . 10^2 {\omega_0^2\over{\lambda T_{\rm c}\mu \alpha}}.$$ The value of $\alpha$ is again determined from the condition that, at time $t_D$, $\langle \phi^2\rangle\sim\alpha\eta^2$. That is, $\alpha^2 = 0.1 (\mu \tau_{\rm q})^{-2/3}$. Thus, decoherence time is given by $$\exp\left\{ {4\over{3}}\left( {\Delta_k (t_D)\over{{\bar t}}}\right)^{3\over{2}} \right\} \approx 3 . 10^2 {\omega_0^2\over{\mu^2}} {\eta^2\over{\mu T_{\rm c}}}(\mu \tau_{\rm q})^{1\over{3}}.$$ From (\[soltsp\]), $t_D$ and $t_{\rm sp}$ are related by $$(\mu\Delta_0 (t_{\rm sp}))^{{3\over{2}}} - (\mu\Delta_0 (t_{D}))^{{3\over{2}}} \approx {3\over{4}}~ \sqrt{\mu \tau_{\rm q}}~ \ln \left[3 (\mu \tau_{\rm q})^{1\over{3}}\right],$$ which is positive for a sufficiently slow quench. As we have already observed, early studies of decoherence were confined largely to quantum mechanical systems, for which the environment was typically a collection of harmonic oscillators, to which the system coupled linearly. Such systems have the virtue of exact solvability (or closed equations) and have been very instructive. However, in the context of quantum field theory linear terms are usually a signal of an inappropriate choice of field basis. Further, as we are considering models with spontaneous symmetry breaking, linear couplings with external fields are not a natural choice since they break the vacuum degeneracy. We include isolated linear couplings for completeness. Again, choosing couplings equal, as $g''_{\rm a}=g''/\sqrt{N}$, and defining $\Delta_{\rm 1} = (\phi^{+} - \phi^{-})/2$ and $\Sigma_{\rm 1}= (\phi^{+} + \phi^{-})/2$, we are able to write the real and imaginary parts of the influence functional as $${\rm Re} \delta A_{\rm lin} = {g''^{2}\mu^4\over{8}} \int d^4 x\int d^4y ~ \Delta_{\rm 1} (x) K_{\rm b}(x-y) \Sigma_{\rm 1} (y),$$ and $${\rm Im} \delta A_{\rm lin} = - {g''^{2}\mu^4\over{16}} \int d^4x\int d^4y ~ \Delta_{\rm 1}(x) N_{\rm b}(x-y) \Delta_{\rm 1}(y).$$ For times $\mu t\gtrsim 1$ the diffusion contribution to the master equation is $$D_{\rm lin}(k_0, t) \approx \frac{T_{\rm c}}{8 \omega_0^3}~\sqrt{u(0, t)}. \label{DL}$$ Note that the exponent is only [*half*]{} that of the quadratic and bilinear interactions. In this case, it can be shown that the decoherence time comes from $$1 \approx {{V \Gamma_{\rm lin} T_{\rm c}}\over{16 \omega_0^3}} \sqrt{{{\bar t}\over{\omega_0}}} \left[- \Gamma [{5\over{6}}] + \Gamma [{5\over{6}}, -{2\over{3}}({\Delta_k(t_D)\over{{\bar t}}})^{3\over{2}}]\right], \label{incomplete}$$ where $ \Gamma [a,z]$ is the incomplete Gamma function and, in this case $\Gamma_{\rm lin} = {1\over{4}} g''^2 \mu^2 \delta^2$. As $t_D \gg {\bar t}$, we can approximate Eq.(\[incomplete\]) and obtain, $$\exp\left\{ {2\over{3}}\left( {\Delta_k(t_D)\over{{\bar t}}}\right)^{3\over{2}} \right\} \approx {16\over{9}}{\omega_0^3\over{\mu^3}}~ {\eta^4\over{\mu^4}} ~{1\over{T_{\rm c}}}{\omega_0^{1\over{2}}\mu^{1\over{3}}\over{\tau_{\rm q}^{1\over{6}}}} .$$ The decoherence time associated to the linear interaction term, can be written as $$\Delta_k(t_D)^{3\over{2}} \approx {3\over{2}} {\bar t}^{3\over{2}} ~ \left\{3\ln {\omega_0\over{\mu}} + \ln {16\over{9}} {\eta^4\over{\mu^4}} ~{1\over{T_{\rm c}}}{\omega_0^{1\over{2}}\mu^{1\over{3}}\over{\tau_{\rm q}^{1\over{6}}}}\right\}.$$ $t_D$ and $t_{\rm sp}$ are related by $$(\mu\Delta_0(t_{\rm sp}))^{{3\over{2}}} - (\mu\Delta_0(t_{D}))^{{3\over{2}}} \sim {3\over{4}}~ \sqrt{\mu \tau_{\rm q}}~~ \ln \left[10^3 {\mu^2\over{\eta^2}} (\mu\tau_{\rm q})^{5\over{6}}\right].$$ In this case, due to the bound on the quench time ($\tau_{\rm q}\leq \eta /\mu^2$), we find that it looks as if $t_{D} > t_{\rm sp}$. However this result is not believable as it stands, since the diffusion coefficient has been computed assuming $t\lesssim t_{\rm sp}$. Whatever, the rapid decoherence of the biquadratic and other couplings is not present. This shows how adopting linear coupling to an environment in mimicry of quantum mechanics can be misleading. Comparison between different couplings {#sec:weak} -------------------------------------- In our present model the environment fields $\chi_{\rm a}$ are not the only decohering agents. The environment is also constituted by the short-wavelength modes of the self-interacting field $\phi$. Therefore, we split the field as $\phi = \phi_< + \phi_>$, and define the system ($\phi_<$) by those modes with wavelengths longer than the critical value $\Lambda^{-1}$, while the bath or environment-field ($\phi_>$) contains wavelengths shorter than $\Lambda^{-1}$. In order to consider only the unstable modes inside our system, we will set this critical scale $\Lambda$ of the order of $\mu$. In practice, where the separation is made exactly is immaterial[@Karra] by time $t_D$, when the power of the $\phi$-field fluctuations is peaked at $k_0\ll\mu$. The effect is to give a separation of system from environment through the decomposition of $S[\phi , \chi ]$ of (\[action0\]) introducing an interaction terms of the form $\phi_< \phi_>^3$, $\phi_<^2 \phi_>^2$, and $\phi_<^3 \phi_>$. The relevant interaction term is $$S_{\rm couple}[\phi_{<} ,\chi ] = S_{\rm int}[\phi_{<} ,\chi ]-{\lambda\over{4}}\int d^4x \,(\phi_{<}(x) \phi_{>} (x))^2. \label{action}$$ All terms omitted in the expansion [@lombmazz; @greiner] do not contribute to the one-loop calculations for the long-wavelength modes that we shall now consider. [^4] The net consequence of the separation of the long-wavelength system modes $\phi_<$ of $\phi$ from the short-wavelength modes $\phi_>$ of the environment through the interaction action $S_{\rm couple}$ of (\[action\]) gives an additional one-loop contribution $D_{\phi}(k_0,t)$ to the diffusion function with the same $u(s,t)$ and the same form as in (\[dqsq\]). However, $G_{\pm\pm}$ is now constructed from the short-wavelength modes $\phi_>$ of the $\phi$-field as it evolves from the top of the potential hill. Without the additional powers of $N^{-1}$ to order contributions the one-loop calculation is unreliable. In fact, we would not expect the inclusion of the $\phi$-field to give a qualitative change at one-loop level. In the first instance, the approximation (\[D0\]) remains valid, and the 1-loop term is driven by the exponential growth of $u(0,t)$. Secondly, the effect is that the short-wavelength modes have been kept at the initial temperature $T_0$, on the grounds that passing through the transition quickly has no effect on them. That is, with $g\simeq\lambda$ and no $1/N$ factor, the short-wavelength $\phi$ modes give a contribution comparable, qualitatively, to [*all*]{} the explicit environmental fields put together. At an order of magnitude level there is no change, since the effect is to replace $g^2$ by $g^2 + O(\lambda^2) = O(g^2)$. However, since the contribution of $D_{\phi}(k_0,t)$ to the overall diffusion function is positive we can derive an [*upper*]{} bound on the decoherence time $t_D$ from the reliable diffusion functions $D_{\rm qu}(k_0,t)$, $D_{\rm bilin}(k_0,t)$ (and $D_{\rm lin}(k_0,t)$). Let us suppose we have a theory where the three considered couplings with the environment are present. Eq.(\[Dsum\]) will be satisfied because one of its terms will grow faster than the others, rather than because many terms will each give a small fraction of unity. Specifically, we have (up to numerical factors $O(1)$) $$\begin{aligned} && \Gamma_{\rm qu} \int_{0}^{t_{D}} ds ~D_{\rm qu}(k_0,s): \Gamma_{\rm bilin} \int_{0}^{t_{D}} ds ~D_{\rm bilin}(k_0,s): \Gamma_{\rm lin} \int_{0}^{t_{D}} ds ~D_{\rm lin}(k_0,s) \nonumber \\ &\sim& g\,\,:\,\, g'\bigg(\frac{\mu}{T_{\rm c}}\bigg)\,\,:\,\, g''\frac{1}{\delta^2}\bigg(\frac{\mu} {T_{\rm c}}\bigg)\,\exp\left\{{2\over{3}}\left({\Delta_0(t_D) \over{\bar t}}\right)^{3\over{2}}\right\}.\end{aligned}$$ Since $T_{\rm c}\gg\mu$ and $\delta \approx O(1)$ then, for $\mu t_D\gg 1$, we have $$\Gamma_{\rm qu} \int_{0}^{t_{D}} ds ~D_{\rm qu}(k_0,s)\,\, \gg\,\, \Gamma_{\rm bilin} \int_{0}^{t_{D}} ds ~D_{\rm bilin}(k_0,s)\,\,\gg\,\, \Gamma_{\rm lin} \int_{0}^{t_{D}} ds ~ D_{\rm lin}(k_0,s),$$ if $g,g'$ and $g''$ are comparable. Thus, it is sufficient to evaluate the constraint on the time $t_{D}$ from this biquadratic interaction alone. Final remarks {#sec:final} ============= We first summarize the results contained in this paper, in which we have shown how the environment leads to the decoherence of the order parameter after a transition. After the integration over the scalar environment-fields $\chi_{\rm a}$ in Section II, we have obtained the coarse-grained effective action (CGEA) for the system (field $\phi$). From the imaginary part of the CGEA we obtained the diffusion coefficient of the master equation, at 1-loop and in the high temperature environment limit. Terms omitted are relatively $O(N^{-1/2})$ for $N$ weakly coupled environmental fields. Subsequently, we evaluated the decoherence time for the long-wavelength modes of the system-field (for each of the different couplings with the environment considered) for slow quenches. This decoherence time depends on the coupling between system and bath, the self-coupling of the system (through the environment temperature $T$), and the mass. In our model, we have shown that the decoherence time is in general [*smaller*]{} than the spinodal time. We stress that the inequality $t_D < t_{\rm sp}$ is insensitive to the strength of the couplings, for weak coupling. The reason is twofold. Firstly, there is the effect that $\Gamma\propto T_0^2$, and $T_0^2 \propto\lambda^{-1}$ is non-perturbatively large for a phase transition. Secondly, because of the non-linear coupling to the environment, obligatory for quantum field theory, $\Gamma\propto{\bar\phi}^2$. The completion of the transition finds ${\bar\phi}^2\simeq\eta^2\propto\lambda^{-1}$ also non-perturbatively large. This suggests that $\Gamma$, and hence $t_D$, can be independent of $\lambda$. In fact, the situation is a little more complicated, but the inequality holds. This arises because the diffusion coefficients, which trigger classical behaviour when they become large enough, are controlled by the exponential growth of the unstable modes. It is this same exponential growth that determines $t_{\rm sp}$. This result provides a post-hoc justification of our assertion [@Karra; @Rivers1] that the spinodal time sets the scale for the onset of classical behaviour (in open systems). Our emphasis has been on the many weak environments because of the control that this gives us on establishing a robust upper bound on $t_D$. However, we noted earlier that their total contribution was qualitatively that of the short-wavelength modes of the $\phi$ field alone. Environmental fields are an important feature of the early universe, but even had we not included them we would have expected a similar result from the one-loop couplings of short-wavelength to long-wavelength $\phi$ modes. Although, in this case, we have no way to control the higher loop terms, it is quite probable that our prescription of the environments is unnecessarily detailed, and early decoherence is a general feature. Ideally, we could extend our ideas to gauge theories. This is beyond the scope of this paper, but we would like to emphasize an important point. As diffusion is additive (each extra term in the interaction action gives additional diffusion coefficients to the master equation), the inclusion of further couplings and fields can only reduce $t_D$. Further, the coupling of short to long-wavelength modes of the global theory is omnipresent. However, theories with derivative couplings (scalar QED, for example) tend to produce small additional diffusive effects at low $\vec k$ (scaled by $k$). It is for this reason that, if scalar fields are present, there is no need include gauge field interactions in order to estimate an upper bound to the decoherence time. The results obtained in this paper also justify in part the use of phenomenological stochastic equations to describe the dynamical evolution of the system field, as we will now discuss. As it is well known [@lombmazz; @greiner], for $\phi^2\chi^m$ interactions one can regard the imaginary part of $\delta A$ as coming from a noise source $\xi_{\rm m}(x)$, with a Gaussian functional probability distribution. Taking the biquadratic coupling to the external environment $\chi$-field first leads to a noise, termed $\xi_2$, say, with distribution $$P[\xi_{2}(x)]= N_{\xi_{2}} \exp\bigg\{-{1\over{2}}\int d^4x\int d^4y ~\xi_{2} (x) \Big[ g^2 N_{\rm q}\Big]^{-1}\xi_{2} (y)\bigg\},$$ where $N_{\xi_{2}}$ is a normalization factor. This enables us to write the imaginary part of the influence action as a functional integral over the Gaussian field $\xi_{2} (x)$ $$\begin{aligned} \int {\cal D}\xi_{2} (x) P[\xi_{2} ]&&\exp{\left[ -i \bigg\{\Delta_{2} (x) \xi_{2} (x) \bigg\}\right]}\nonumber \\ &=& \exp{\bigg\{-i\int d^4x\int d^4y \ \Big[\Delta_{2}(x) ~g^2 N_{\rm q}(x,y)~ \Delta_{2}(y)\Big] \bigg\}}.\end{aligned}$$ In consequence, the CGEA can be rewritten as $$A[\phi^+,\phi^-]=-{1\over{i}} \ln \int {\cal D} \xi_{2} P[\xi_{2}] \exp\bigg\{i S_{\rm eff}[\phi^+,\phi^-, \xi_{2}]\bigg\},$$ where $$S_{\rm eff}[\phi^+,\phi^-,\xi_{2}]= {\rm Re} A[\phi^+,\phi^-]- \int d^4x\Big[\Delta_{2} (x) \xi_{2}(x) \Big].$$ Therefore, taking the functional variation $$\left.{\delta S_{\rm eff}[\phi^+,\phi^-, \xi_{2}]\over{\delta \phi^+}}\right\vert_{\phi^+=\phi^-}=0, \label{statphase}$$ we are able to obtain the “semiclassical-Langevin” equation for the system-field [@lombmazz; @greiner] $$\begin{aligned} \Box \phi (x) - {\tilde\mu}^2 \phi + {{\tilde\lambda}\over{6}}\phi^3(x) &+& g^2 \phi (x) \int d^4y ~ K_{\rm q}(x-y)~ \phi^2(y) = \phi (x) \xi_{2}(x)\label{lange2} ,\end{aligned}$$ where ${\tilde\mu}$ and $ {\tilde\lambda}$ are constants “renormalised” because of the coupling with the environment. Since (\[lange2\]) is, from (\[statphase\]), the stationary phase approximation, it is only valid once the system has become classical. It can be used to establish domain formation only because $t_D < t_{\rm sp}$. Each part of the environment that we include leads to a further ’dissipative’ term on the left hand side of (\[lange2\]) with a countervailing noise term on the right hand side. Once we include the interactions between $\phi_<$ and $\phi_>$ at one-loop level we get a similar equation to (\[lange2\]) from the term $S_{\rm couple}$ alone (see Eq.(\[action\])). However, although the $\phi_<\phi^3_>$ and $\phi^3_<\phi_>$ terms were ignorable in the bounding of $t_D$, in the Langevin equations they give further terms, with quadratic $\phi^2\xi_3$ noise and linear noise $\xi_1$ respectively. The inclusion of bilinear interactions leads to the inclusion of further terms in (\[lange2\]) of the same form. For the linear interaction with the environment (to the exclusion of self-interaction) we do recover the [*additive*]{} noise that has been the basis for stochastic equations in relativistic field theory that confirm the scaling behaviour of Kibble’s and Zurek’s analysis [@laguna]. For times later than $t_{\rm sp}$, neither perturbation theory nor more general non-Gaussian methods are valid. It is difficult to imagine an [*ab initio*]{} derivation of the dissipative and noise terms from the full quantum field theory. In this sense, a reasonable alternative is to analyze phenomenological stochastic equations numerically and check the robustness of the predictions against different choices of the dissipative kernels and of the type of noise. We stress that this is only possible because $t_D < t_{\rm sp}$. This will be considered further elsewhere [@nunopedroray]. Finally, we see that the role initially attributed by Kibble [@Kibble; @kzm] (and subsequently by others e.g.[@joao]) to the Ginzburg regime is just not present. F.C.L. and F.D.M. were supported by Universidad de Buenos Aires, CONICET, Fundación Antorchas and ANPCyT. R.J.R. was supported, in part, by the COSLAB programme of the European Science Foundation. The spinodal time {#rw-appendix} ================= In this Section, we show the estimation of the spinodal time $t_{\rm sp}$ defined from $\langle \phi^2 \rangle_{t=t_{\rm sp}} \sim \eta^2$. The equation of motion for the mode ${\cal U}_k(t)$, with wavenumber $k$ is, in the quench period, $$\left[{d^2\over{ds^2}} + k^2 + \mu^2 - {\mu^2 s\over{\tau_{\rm q}}}\right] {\cal U}_k(t) = 0, \label{airy1}$$ subject to the boundary condition ${\cal U}_k(t) = e^{-i\omega t}$ for $t\leq 0$, where $\omega^2 = \mu^2 + k^2$. Instead of the simple exponentials of the instantaneous quench, ${\cal U}_k(t)$ has solution $${\cal U}_k(t) = a_k Ai[{\Delta_k (t)\over{{\bar t}}}] + b_k Bi[{\Delta_k (t) \over{{\bar t}}}], \label{airy1sol}$$ with $Ai[s]$, $Bi[s]$ the Airy functions; $\Delta_k (t) = t - \omega^2 {\bar t}^3 $ and ${\bar t} = (\tau_{\rm q}/\mu^2)^{1/3}$. Note that $\Delta_0 (t) = t - \tau_Q$, the time since the onset of the transition. In the causal analysis of Kibble[@Kibble] ${\bar t}$ ($\mu^{-1}\ll {\bar t}\ll \tau_{\rm q}$) is the time at which the adiabatic field correlation length collapses at the speed of light, the earliest time in which domains could have formed. Our analysis suggests that this earliest time is not ${\bar t}$, but $t_{\rm sp}$. It is straightforward to establish a relationship between ${\bar t}$ and $t_{\rm sp} > {\bar t}$. The constants of integration in (\[airy1sol\]) are $$\begin{aligned} a_k &=& \pi [Bi'(-\omega^2{\bar t}^2) + i\omega{\bar t}Bi (-\omega^2{\bar t}^2)],\\ \nonumber b_k &=& - \pi [Ai'(-\omega^2{\bar t}^2) + i\omega{\bar t}Ai (-\omega^2{\bar t}^2)].\end{aligned}$$ It follows that, when $\Delta_k (t)/{\bar t}$ is large, then $$\begin{aligned} |{\cal U}_k(t)|^2&\approx& \omega{\bar t} \bigg({{\bar t}\over\Delta_k(t)} \bigg)^{1/2}\exp\bigg[{4\over 3}\bigg({\Delta_k(t)\over{\bar t}}\bigg)^{3/2}\bigg] \nonumber \\ &\approx& \mu{\bar t} \bigg({{\bar t}\over\Delta_0(t)} \bigg)^{1/2}\exp\bigg[{4\over 3}\bigg({\Delta_0(t)\over{\bar t}}\bigg)^{3/2}\bigg]e^{-k^2/{\bar k}^2},\end{aligned}$$ where ${\bar k}^2 = {\bar t}^{-3/2}(\Delta_0(t))^{-1/2}/2$. For large initial temperature $T_0 = O(T_c)$, we find the power spectrum for field fluctuations peaked around ${\bar k}$, and $$\langle \phi^2\rangle_{t}\approx {T_0\over 2\pi^2\mu^2}\int k^2\,dk\,|{\cal U}_k(t)|^2\approx {CT_0\over\mu{\bar t}^2}\bigg({\Delta_0(t)\over{\bar t}} \bigg)^{-5/4}\exp\bigg[{4\over 3}\bigg({\Delta_0(t)\over{\bar t}}\bigg)^{3/2}\bigg]. \label{phi2}$$ We have intentionally included the prefactor $C$ to show that terms, nominally $O(1)$, can in fact be large or small (in this case $C= (64\sqrt{2}\pi^{3/2})^{-1}=O(10^{-3}))$. Note that, although the unstable modes have a limited range of $k$-values, increasing in time, this is effectively no restriction when $\Delta_0(t)/{\bar t}$ is significantly larger than unity. Finally, we obtain $${\eta^2\over C'}\simeq {T_{\rm c}\over \mu{\bar t}^2}\exp\bigg[{4\over 3}\bigg({\Delta_0(t_{\rm sp})\over{\bar t}}\bigg)^{3/2}\bigg],$$ where $C' = C[\ln(\mu{\bar t}^2\eta^2/CT_{\rm c})^{-5/6}]$. Since the effect on $t_{\rm sp}$ only arises at the level of ‘$\ln\ln$’ terms, $C'\approx C$ is a good estimation in all that follows. Since this choice underestimates $t_{\rm sp}$ it only strengthens our results that $t_{\rm sp}>t_D$. The diffusion coefficient {#dc-appendix} ========================= In this Appendix we show how to obtain the diffussion coefficient Eq.(\[dqsq\]) (for the quadratic coupling) from the reduced propagator. Following the same techniques used for the quantum Brownian motion to obtain the master equation we must compute the time derivative of the propagator $J_{\rm r}$ (Eq.(\[prosadle\])), and eliminate the dependence on the initial field configurations $\phi^\pm_{\rm i}$ that enters through the classical solutions $\phi_{\rm cl}^\pm$. This can be easily done using the propagator identities of Eq.(\[rel1\]). Thus, using Eq.(\[us\]) and identities (\[rel1\]), we can write the initial field configuration $\phi_{\rm i}^\pm$, in terms of the final configurations ($\phi_{\rm f}^\pm$), final field derivatives ($\partial_{\phi^\pm_{\rm f}}$), and the functions of time $u_{\rm i}(s,t)$, $$(\phi_{\rm i}^{+} - \phi_{\rm i}^{-}) ~ u_1(s,t)~ J_0 = - {{\dot u}_2(t,t)\over{{\dot u}_1(t,t)}} ~ u_1(s,t) ~ ( \phi_{\rm f}^{+} - \phi_{\rm f}^{-}) ~ J_0 + ..., \label{cond}$$ neglecting terms proportional to derivatives respect to the final field configuration which do not contribute to normal diffusion. The temporal derivative is given by $$\begin{aligned} i\hbar \partial_t J_r[&&\phi_{\rm f}^+,\phi_{\rm f}^-,t\vert \phi_{\rm i}^+,\phi_{\rm i}^-,0] = \bigg\{h_{\rm ren}[\phi^+] - h_{\rm ren}[\phi^-] - i g^2 {(\phi_{\rm f}^{+2} - \phi_{\rm f}^{-2})V\over{16}} \int_0^t ds ~ \Delta_2^{\rm cl} (s)~ F(k_0;s,t)\nonumber \\ && + g^2 {(\phi_{\rm f}^{+2} + \phi_{\rm f}^{-2})V\over{16}} \int_0^t ds ~ \Delta_2^{\rm cl} (s) ~ {\tilde K}_{\rm q}(k_0; s, t) + ... \bigg\}J_r[\phi_{\rm f}^+,\phi_{\rm f}^-,t\vert \phi_{\rm i}^+,\phi_{\rm i}^-,0],\label{timeder} \end{aligned}$$ where $ {\tilde K}_{\rm q}(k_0; s, t)$ is the Fourier transform of the dissipation kernel. The ellipsis denotes other terms coming from the time derivative which do not contribute to diffusion. Diffusive effects are associated with terms proportional to $(\phi_{\rm f}^{+2} - \phi_{\rm f}^{-2})^2$ in the master equation. Using Eq.(\[cond\]) we remove initial conditions from Eq.(\[timeder\]), and looking only those terms proportional to ${\Delta_2^{\rm f}}^2$, we get the master equation Eq.(\[master\]). The diffusion coefficient $D(k_0, t)$ comes from the noise contribution to the influence functional and it is the time dependent coefficient that multiplies $\Delta_2^{\rm f2}$. Thus, for the quadratic coupling example, we find $$D_{\rm qu}(k_0, t) = \int_0^t ds ~ u(s,t) ~ F(k_0;s,t),$$ where $$u(s,t) = \left[u_2(s,t) - {{\dot u}_2(t,t)\over{{\dot u}_1(t,t)}} u_1(s,t)\right]^2.$$ For different couplings between system and environment one should follow the same procedure shown in this Appendix. A.D. Linde, [*Particle Physics and Inflationary Cosmology*]{}, Harwood, Chur, Switzerland (1990) S.W. Hawking, Phys. Lett. [**B115**]{}, 295 (1982); A.A. Starobinsky, Phys. Lett. [**B117**]{}, 175 (1982); A.H. Guth and S.Y. Pi, Phys. Rev. Lett. [**49**]{}, 1110 (1982) D.A. Kirzhnits and A.D. Linde, Phys. Lett. [**B42**]{}, 471 (1972) S.A. Ramsey and B.L. Hu, Phys. Rev. [**D56**]{}, 678 (1997); S.A. Ramsey, B.L. Hu, and A.M. Stylianopoulos, Phys. Rev. [**D57**]{}, 6003 (1998); D. Cormier and R. Holman, Phys. Rev. [**D62**]{}, 023520 (2000) and references therein T.W.B. Kibble, Phys. Rep. [**67**]{}, 183 (1980) A. Vilenkin, Phys. Rep. [**121**]{}, 263 (1985); A. Rajantie, Int. J. Mod. Phys. [**A17**]{}, 1 (2002) A. Guth and S.Y. Pi, Phys. Rev. [**D32**]{}, 1899 (1991) D. Polarski and A.A. Starobinsky, Class. Quantum Grav. [**13**]{}, 377 (1996); J. Lesgourgues, D. Polarski and A.A. Starobinsky, Nucl. Phys. [**B497**]{}, 479 (1997) C. Kiefer, D. Polarski, and A.A. Starobinsky, Int. J. Mod. Phys. [**D7**]{}, 455 (1998) J.J. Halliwell, Phys. Rev.[**D36**]{}, 3626 (1987) F.C. Lombardo, F.D. Mazzitelli, and D. Monteoliva, Phys. Rev. [**D62**]{}, 045016 (2000) F. Cooper, S.Y. Pi, and P. Stancioff, Phys. Rev. [**D34**]{}, 383 (1986) D. Giulini, C. Kiefer, E. Joos, J. Kupsch, I.O. Stamatescu, and H.D. Zeh, [*Decoherence and the apperance of a classical world in quantum theory*]{}, Springer, Berlin, Germany (1996) B. Mihaila, T. Athan, F. Cooper, J. Dawson, and S. Habib, Phys. Rev. [**D62**]{}, 125015 (2000); B. Mihaila, J. F. Dawson, and F. Cooper, Phys. Rev. [**D63**]{}, 096003 (2001) S. Habib, Y. Kluger, E. Mottola and J.P. Paz, Phys. Rev. Lett. [**76**]{}, 4660 (1996) F. Cooper, S. Habib, Y. Kluger, and E. Mottola, Phys. Rev. [**D55**]{}, 6471 (1997) D. Boyanovsky, H.J. de Vega, and R. Holman, ”Non-equilibrium phase transitions in condensed matter and cosmology: spinodal decomposition, condensates and defects”. Lectures delivered at the NATO Advanced Study Institute: Topological Defects and the Non-Equilibrium Dynamics of Symmetry Breaking Phase Transitions, hep-ph/9903534 D. Boyanovsky, D. Lee, and A. Singh, Phys. Rev. [**D48**]{}, 800 (1993); D. Boyanovsky and H.J. de Vega, Phys. Rev. [**D47**]{}, 2343 (1993) N.D. Antunes, F.C. Lombardo and D. Monteoliva, Phys. Rev. [**E64**]{}, 066118 (2001) F.C. Lombardo, F.D. Mazzitelli, and R.J. Rivers, Phys. Lett. [**B523**]{}, 317 (2001) G. Aarts and J. Berges, Phys.Rev.Lett. [**88**]{}, 041603 (2002); W. Buchmüller and A. Jakovac, Nucl. Phys. [**B521**]{}, 219 (1998); G. Aarts and J. Smit, Nucl. Phys. [**B511**]{}, 451 (1998) R.J. Rivers, F.C. Lombardo, and F.D. Mazzitelli, Phys. Lett. [**B539**]{}, 1 (2002) T.W.B. Kibble, J. Phys. [**A9**]{}, 1387 (1988); W.H. Zurek, Nature [**317**]{}, 505 (1985); Phys. Rep. [**276**]{}, 177 (1996) R.J. Rivers, Journal of Low Temperature Physics [**124**]{}, 41-84, (2001) cond-mat/0105171; R.J. Rivers, E. Kavoussanaki, and G. Karra, Condensed Matter Physics vol. 3, 133 (2000) P. Laguna and W.H. Zurek, Phys. Rev. Lett. [**78**]{}, 2519 (1997); Phys. Rev. [**D58**]{}, 085021 (1998); A. Yates and W.H. Zurek, Phys. Rev. Lett. [**80**]{}, 5477 (1998) N.D. Antunes, L.M.A. Bettencourt, W.H. Zurek, Phys. Rev. Lett. [**82**]{}, 2824 (1999) B.L. Hu, in [*Relativity and Gravitation: Classical and Quantum*]{}, J.C.D. Olivo et al (Eds.), World Scientific, Singapore (1991) F.C. Lombardo and F.D. Mazzitelli, Phys. Rev. [**D53**]{}, 2001 (1996) C. Greiner, B. Muller, Phys. Rev. [**D55**]{}, 1026 (1997) W. G. Unruh and W. H. Zurek, Phys. Rev. [**D40**]{}, 1071 (1989); A. Caldeira and A. Leggett, Phys. Rev. [**A31**]{}, 1059 (1985) S.P. Kim and C. H. Lee, Phys. Rev. [**D65**]{}, 045013 (2002) W.H. Zurek, “Prefered Sets of States, Predictability, Classicality, and Environment-Induced Decoherence”; in [*The Physical Origin of Time Asymmetry*]{}, ed. by J.J. Halliwell, J. Perez Mercader, and W.H. Zurek (Cambridge University Press, Cambridge, UK, 1994) R. Feynman and F. Vernon, Ann. Phys. (N. Y.) 24, 118 (1963) E.A. Calzetta, B.L. Hu, and F.D. Mazzitelli, Phys. Rep. [**352**]{}, 459 (2001) B.L. Hu, J.P. Paz, and Y. Zhang, Phys. Rev. [**D45**]{}, 2843 (1993); [**D47**]{}, 1576 (1993); J.P. Paz, S. Habib, and W.H. Zurek, Phys. Rev. [**D47**]{}, 488 (1993) G. Karra and R.J. Rivers, Phys. Lett. [**B414**]{}, 28 (1997) M. J. Bowick and A. Momen, Phys. Rev. [**D58**]{}, 085014 (1998) A.J. Gill and R.J. Rivers, Phys. Rev. [**D51**]{}, 6949 (1995) N.D. Antunes, P. Gandra, and R.J. Rivers, in preparation. R.H. Brandenberger and J. Maguiejo, e-print astro-ph/0002030. Lecture notes of the International School on Cosmology, Kish Island, Iran, Jan. 22 - Feb. 4 1999 [^1]: Electronic address: [email protected] [^2]: Electronic address: [email protected] [^3]: Electronic address: [email protected] [^4]: Strictly speaking, for $\mu/3<k_0<\mu$ one should include an additional term proportional to $\phi_<^3\phi_>$ in the interaction Lagrangian. See Ref. [@lombmazz] for details.

ArXiv

--- abstract: '[Double Weyl nodes are topologically protected band crossing points which carry chiral charge $\pm2$. They are stabilized by $C_{4}$ point group symmetry and are predicted to occur in $\mathrm{SrSi_{2}}$ or $\mathrm{HgCr_{2}Se_{4}}$. We study their stability and physical properties in the presence of a disorder potential. We investigate the density of states and the quantum transport properties at the nodal point. We find that, in contrast to their counterparts with unit chiral charge, double Weyl nodes are unstable to any finite amount of disorder and give rise to a diffusive phase, in agreement with predictions of Goswami and Nevidomskyy \[Phys. Rev. B 92, 214504 (2015)\] and Bera, Sau, and Roy \[Phys. Rev. B 93, 201302(R) (2016)\]. However, for finite system sizes a crossover between pseudodiffusive and diffusive quantum transport can be observed.]{}' author: - Björn Sbierski - Maximilian Trescher - 'Emil J. Bergholtz' - 'Piet W. Brouwer' bibliography: - 'library.bib' title: 'Disordered double Weyl node: Comparison of transport and density-of-states calculations' --- Introduction ============ Topological metals and semimetals are among the driving themes in contemporary condensed matter physics. Their most prominent three-dimensional realizations are Weyl (semi)metals, which have recently been experimentally confirmed in a number of different material systems [@Bernevig2015; @Xu2015; @Lv2015b; @Xu2015d; @Borisenko2015; @Liu2015a]. Pioneering experimental studies used spectroscopic measurements to study surface Fermi arcs and characteristic Weyl node bulk dispersions. Recently, also (magneto-)transport properties received growing interest in experiment [@Zhang2015; @Zhang2015a; @Shekhar2015]. While sample quality matures continuously, more controlled engineering of the chemical potential comes into reach [@Ruan2016; @Ruan2016a]. Weyl nodes with chemical potential $\mu$ sufficiently close to the nodal point ($\mu=0$) are predicted to show unusual transport characteristics for sample length $L\lesssim\hbar v/\mu$, with $v$ the Fermi velocity [@Baireuther2014; @Sbierski2014a]. Without disorder, the conductance scales with system size as $G\propto W^{2}/L^{2}$, where $W$ is the sample width. The inclusion of weak disorder is irrelevant in the renormalization-group (RG) sense [@Fradkin1986; @Goswami2011; @Kobayashi2014; @Ominato2014; @Syzranov2015a; @Syzranov2015cc] and consequently does not change the size dependence of the conductance. This so-called “pseudoballistic” regime is further characterized by an unusual Fano factor (the ratio of shot-noise power and average current) $F \approx 0.57$ [@Baireuther2014; @Sbierski2014a]. Only if disorder increases above a critical strength, the conductivity and density of states at the nodal point attains a non-zero value and transport becomes diffusive, $G\propto W^{2}/L$ and $F=1/3$. The simple Weyl node (SWN) band structure discussed above carries a topological charge of $\pm1$. Beyond the SWN, the existence of topological band touching points with higher topological charge is tied to the presence of point-group symmetries [@Fang2012a]. In this paper, we consider double Weyl nodes (DWN) with chiral charge of magnitude two, stabilized by $C_{4}$ rotation symmetry. The Hamiltonian reads $$H=\hbar v[\sigma_{x}\eta_{x}\,(k_{x}^{2}-k_{y}^{2})/2+\sigma_{y}\eta_{y}\, k_{x}k_{y}+k_{z}\sigma_{z}],\label{eq:H}$$ with $\eta_{x,y}$ internal length scales. The fourfold rotational symmetry around the $z$-axis is realized as $H\!\left(k_{x},k_{y},k_{z}\right)=\sigma_{z}H\left(k_{y},-k_{x},k_{z}\right)\sigma_{z}$. Time-reversal symmetry is present, $\sigma_{x}H^{*}\left(-\mathbf{k}\right)\sigma_{x}=H\left(\mathbf{k}\right)$ with the time reversal operator $T=\sigma_{x}K$ squaring to $+1$. To simplify the subsequent analysis, we specialize to the case $\eta_{x}=\eta_{y}\equiv\eta$ where the discrete rotation symmetry is extended to a continuous rotation symmetry $C_{\infty}$, in cylindrical coordinates $H\left(k_{\perp},\phi,k_{z}\right)=e^{-i\theta\sigma_{z}}H\left(k_{\perp},\phi-\theta,k_{z}\right)e^{i\theta\sigma_{z}}$. The corresponding energy dispersion $\varepsilon_{\pm}^{2}/(\hbar v)^{2}=(k_{\perp}^{2}\eta/2)^{2}+k_{z}^{2}$ is quadratic in the momentum $\mathbf{k}_{\perp}=(k_{x},k_{y})$ transverse to the rotation axis and linear in $k_{z}$, see Fig. \[fig:summary\](a). A photonic crystal realization of DWNs is reported in Ref. [@Chen2015] and fermionic candidate materials have been identified from first-principle calculations, such as $\mathrm{HgCr_{2}Se_{4}}$[@Xu2011] or $\mathrm{SrSi_{2}}$[@Huang2015]. The latter material might be experimentally more feasible since no magnetic ordering is required. An interesting proposal to detect the monopole charge in electronic Weyl materials using transport measurements has recently been formulated in Ref. .  Dispersion for DWN Hamiltonian $H$ with $\eta_{x,y}=\eta$ in Eq. (\[eq:H\]) at $k_{z}=0$. (b) The potential disorder profile is characterized by its Gaussian correlations decaying on a length scale $\xi$, here a slice at $z=z_{0}$ is shown.]{}](dWN_disordered_sketch) In view of the requirement of point-group symmetries, the stability of a DWN to disorder, which typically breaks such symmetry [\[]{}see Fig. \[fig:summary\](b)[\]]{}, is a relevant question. Several groups have addressed this question theoretically, with partially diverging results. Using a simplified version of the self-consistent Born approximation, Goswami and Nevidomskyy [@Goswami2014] argued that the DWN is unstable to disorder, and that inclusion of even a small amount of disorder drives the system to a diffusive phase with zero-energy scattering rate $\hbar/\tau\sim e^{-A/K}$, where $K$ is a dimensionless measure of the disorder strength and $A$ a material-dependent parameter. The same conclusion was drawn by Bera, Sau and Roy [@Bera2015], based both on an RG analysis [\[]{}which found disorder a marginally relevant perturbation to Eq. (\[eq:H\])[\]]{} and a numerical calculation of the density of states at zero energy, which was claimed to be compatible with the exponential form proposed above. Recently, Shapourian and Hughes [@Shapourian2015] revisited the same problem, conducting a finite-size scaling analysis of the decay length in the $z$ direction using a transfer-matrix method. Their data indicates the presence of a critical point at a *finite* disorder strength (but below the Anderson transition), leading them to conclude the stability of the DWN phase against weak disorder. A possible scenario for such an observation would be the splitting of the DWN into two equally charged SWNs under the influence of disorder, where the latter individually would indeed feature a critical point. This interesting scenario and the apparent contradiction between results in the literature motivated us to revisit the problem of a disordered DWN. We first investigate the density of states using the Kernel Polynomial method and the self-consistent Born approximation (Sec. \[sec:DOS\]). We discuss the shortcomings of either method and move on to a scattering matrix-based transport calculation, much better suited to study the physics right at the nodal point (Sec. \[sec:Quantum-transport\]). These combined numerical efforts allow us to put forward the following interpretation: In the presence of any finite amount of disorder, the clean DWN fixed point is unstable and gives rise to a diffusive phase. We find no evidence in support of a critical point at finite disorder strength and, accordingly, of the DWN splitting scenario. However, due to exponentially small scattering rate, a crossover behavior can be observed in the quantum transport properties of weakly disordered mesoscopic samples. \[sec:DOS\]Density of states ============================ Kernel-Polynomial Method ------------------------ We start by calculating the density of states in a disordered DWN which we regularize on a cubic lattice $$\begin{aligned} H_{\mathrm{L}}(\mathbf{k}) & \!= & \!\varepsilon_{0}\frac{\eta}{a}\left[\sigma_{x}(\cos ak_{x}-\cos ak_{y})+\sigma_{y}\sin ak_{x}\sin ak_{y}\right]\nonumber \\ & & -\varepsilon_{0}\sigma_{z}\cos ak_{z}\end{aligned}$$ where $\varepsilon_{0}=\hbar v/a$ and $a$ is the lattice constant. The effective low energy approximation of $H_{\mathrm{L}}$ around $\varepsilon=0$ consists of four DWNs centered at $k_{z}=\pm\frac{\pi}{2a}$ and $(k_{x},k_{y})=(0,0)$ or $(\frac{\pi}{a},\frac{\pi}{a})$ with minimal distance $\Delta k=\pi/a$. We include a Gaussian disorder potential $U(\mathbf{r})$ characterized by zero mean and real space correlations given by $$\left\langle U(\mathbf{r})U(\mathbf{r}^{\prime})\right\rangle _{\rm dis}=\frac{K\left(\hbar v\right)^{2}}{\sqrt{2\pi}^{3}\xi^{2}}e^{-|\mathbf{r}-\mathbf{r}^{\prime}|^{2}/2\xi^{2}},\label{eq:dis_correlator}$$ where $\xi$ is the correlation length and $K$ the dimensionless disorder strength. In the following, we use $\xi=\eta/2$ but different choices do not qualitatively change our conclusions. To smoothly represent $U(\mathbf{r})$ on the lattice scale, we take $\xi=2.9a$ which suppresses the inter-node scattering rate by a factor $e^{-(\Delta k)^{2}\xi^{2}/2}<10^{-18}$ compared to the intra-node rate, so single node physics (i.e. $H+U$) is realized to a very good approximation. We numerically calculate the density of states of $H_{\mathrm{L}}+U(\mathbf{r})$ using the Kernel Polynomial method (KPM) (see Ref. [@Weisse2006] for a description of the method). The resulting density of states normalized to a single DWN is shown as solid lines in Fig. \[fig:DOS\](a). Further simulation parameters are given in the figure caption. The analytical result for an infinite clean system, $$\nu_{0}(\varepsilon)=\frac{\varepsilon}{4\pi\left(\hbar v\right)^{2}\eta},\label{eq:cleanDOS}$$ is shown as a dotted line in Fig. \[fig:DOS\] and compares well with the $K=0$ KPM results except at $\varepsilon=0$. At the nodal point, the KPM method has intrinsic difficulties to simulate the vanishing (or very small) density of states, which is due to the finite expansion order of $\nu(\varepsilon)$ in Chebyshev polynomials and the discrete nature of eigenstates in a finite tight-binding model. In Fig. \[fig:DOS\](b), we plot $\nu(\varepsilon=0)$ vs. $K$. Our findings are in qualitative agreement with similar numerical results in Ref. : The presence of disorder scattering fills the dip in the density of states for any finite disorder strength. ![[\[fig:DOS\]Density of states $\nu$ as a function of energy $\varepsilon$ (top) and at the nodal point $\varepsilon=0$ (bottom) normalized to a]{} ** [single DWN as computed from the Kernel polynomial method (KPM) applied to the lattice Hamiltonian $H_{\mathrm{L}}$ (solid lines). The results of the self-consistent Born approximation based on $H$ (dashed lines) are in good agreement with the KPM data except in the vicinity of the nodal point and for large disorder strengths $K>6$. We take the disorder correlation length $\xi=\eta/2$. The system size of the tight-binding model underlying the KPM calculation is $L_{x,y}=100a$, $L_{z}=260a$ and we apply periodic boundary conditions. The expansion order in Chebyshev Polynomials $N$ is taken in between 1000 and 6000 depending on the energy $\varepsilon$ so that $\nu$ is minimized but oscillations due to the underlying discrete Eigenenergies of the finite system are sufficiently smoothed out. An average over 10 disorder realizations is taken and 20 random vectors were used to calculate the trace in the KPM.]{}](KPM_SCBA.pdf) Self-consistent Born approximation ---------------------------------- A frequently employed analytical approach to disordered electronic systems is the self-consistent Born approximation (SCBA). Although a simplified SCBA calculation has been performed in Ref. , in the following we compute the SCBA self-energy for $H+U$ and the associated density of states without any further approximations. The results are shown as dashed lines in Fig. \[fig:DOS\], comparison to the KPM results confirm that SCBA is accurate only at large energies $\varepsilon$ and weak disorder values, when $\varepsilon\tau\gg\hbar$ with $\tau$ being the quasiparticle scattering time. We start from Hamiltonian $H$ with $\eta_{x}=\eta_{y}\equiv\eta$ and seek to describe the disorder averaged retarded Green function $\left\langle G^{R}\right\rangle _{\rm dis}=1/(\varepsilon-H-\Sigma^{R})$ in terms of a translationally invariant self energy term $\Sigma^{R}$ that fulfills the SCBA equation $$\Sigma^{R}(\mathbf{k})=\int\frac{d\mathbf{k}^{\prime}}{(2\pi)^{3}}\left\langle G^{R}(\mathbf{k}^{\prime})\right\rangle _{\rm dis}\left|U(\mathbf{k}^{\prime}-\mathbf{k})\right|^{2}\label{eq:SCBA equation}$$ where $\left|U(\mathbf{k}^{\prime}-\mathbf{k})\right|^{2}$ is the Fourier transform of the disorder correlator in Eq. . Since a disorder average restores the $C_{\infty}$-symmetry of the system around the $k_{z}$ axis, the projection of $\Sigma^{R}(k,\phi,k_{z})$ to the $\sigma_{x}-\sigma_{y}$ plane in Pauli-matrix space should point into $\phi$ direction, the angle between this plane and the $\sigma_{z}$ projection of $\Sigma^{R}(k,\phi,k_{z})$ is not dictated by symmetry and can be different from the angle in $H\left(k,\phi,k_{z}\right)$. With these considerations, a natural ansatz for the self energy is $$\begin{aligned} \Sigma^{R}(k,\phi,k_{z})/\hbar v & \!= & m(k,k_{z})\!\left(\cos\left[2\phi\right]\sigma_{x}+\sin\left[2\phi\right]\sigma_{y}\right)\nonumber \\ & & +\sigma_{z}m_{z}\left(k,k_{z}\right)-im_{0}\left(k,k_{z}\right)\label{eq:SigmaAnsatz}\end{aligned}$$ with $m$, $m_{z}$ and $m_{0}$ complex and $Re[m_{0}]>0$. At $\varepsilon=0$, in order to avoid an unphysical spontaneous generation of a chemical potential from disorder with $\left\langle U(\mathbf{r}))\right\rangle _{\rm dis}=0$, $m_{0}$ has to be chosen purely real which enforces also $m$ and $m_{z}$ to be real quantities. The resulting self-consistency equations for $m$, $m_{z}$ and $m_{0}$ are given in the appendix and can be solved numerically by iteration. The density of states follows from $$\nu\left(\varepsilon\right)=-\frac{1}{\pi}\mathrm{Im}\int\frac{d\mathbf{k}}{(2\pi)^{3}}\mathrm{Tr}\left\langle G^{R}(\mathbf{k})\right\rangle _{\rm dis}.\label{eq:SCBA_DOS}$$ Results of this calculation are shown as dashed lines in Fig. \[fig:DOS\] for various representative disorder strengths $K$. Discussion ---------- The SCBA calculation in Sec. \[sec:DOS\] can be simplified by taking the disorder correlation length to zero and choosing a finite (half-)bandwidth $\Lambda\gg\hbar v/\eta$. Then we could define $K$ such that $\left|U(\mathbf{k}^{\prime}-\mathbf{k})\right|^{2}=K\Lambda^{2}\eta^{3}$ and insert this in Eq. at $\varepsilon=0$, where $\Sigma^{R}\equiv-i\Gamma$ becomes independent of $\mathbf{k}$. Transforming to an energy integral and using the density of states along with the assumption $\Gamma\ll\Lambda$, one finds $$\Gamma=\Lambda e^{-A/K}\label{eq:Goswami}$$ with $A=2\pi\left(\hbar v/\Lambda\eta\right)^{2}$. This was first observed by Goswami and Nevidomskyy in Ref. and states that any finite disorder strength gives rise to a finite lifetime $1/\Gamma$ of quasiparticles and a finite density of states $\nu(\varepsilon=0)\propto\Gamma$ at the nodal point. Our SCBA analysis which takes into account a more realistic disorder model and infinite bandwidth confirms the simplified result in Eq. qualitatively, see dashed line in Fig. \[fig:DOS\] bottom panel. However, it is well known that the SCBA is not reliable around gapless points, where the smallness of the parameter $k_{F}l$ spoils the suppression of crossed diagram contributions to the self energy (see, [*e.g.*]{} Ref. [@Sbierski2014a] for a discussion in the context of simple Weyl nodes). Indeed, comparing the non-perturbative KPM results for $\nu(\varepsilon)$ to the SCBA in Fig. \[fig:DOS\], good agreement is achieved away from the nodal point only. At the nodal point, it is difficult to judge the qualitative validity of Eq. based on the KPM results. The reason is that, for the latter method, finite size and smoothing effects tend to overestimate $\nu(\varepsilon=0)$. (For example, the KPM method returns a finite value of $\nu(\varepsilon=0)$ even for $K=0$, see Fig. \[fig:DOS\], bottom panel.) In summary, neither numerical nor analytical calculations of the density of states as presented above are conclusive in gauging the qualitative validity of Eq. against the alternative scenario of a finite critical disorder strenght below which the bulk density of states vanishes. In this situation, we switch to a quantum transport framework which is ideally suited to study the disordered DWN at the nodal point. \[sec:Quantum-transport\]Quantum transport ========================================== Clean case ---------- We start this section by calculating the conductance and shot noise power of a clean mesoscopic DWN sample of length $L$ and width $W$ coupled to ideal leads, building on earlier work by Tworzydlo [*et al.*]{}on two-dimensional Dirac nodes [@Tworzydlo2006]. We choose the transport direction as the $z$ direction and place the chemical potential at the nodal point. We model the leads as highly doped DWNs, $H_{\mathrm{lead}}=H+V$ with $V\rightarrow\infty$. By matching wavefunctions at the sample-lead interfaces we calculate the transmission amplitudes $t_{0}(\mathbf{k}_{\perp})$ and $t_{0}^{\prime}(\mathbf{k}_{\perp})$ and reflection amplitudes $r_{0}(\mathbf{k}_{\perp})$ and $r_{0}^{\prime}(\mathbf{k}_{\perp})$, where the primed (unprimed) amplitudes refer to electrons incident from the positive (negative) $z$ direction, $$\begin{aligned} t_{0}=t_{0}^{\prime} & = & 1/\cosh(\eta Lk_{\perp}^{2}/2),\nonumber \\ r_{0}=-r_{0}^{\prime*} & = & ie^{-2i\varphi}\tanh(\eta Lk_{\perp}^{2}/2),\label{eq:S0-1}\end{aligned}$$ $\mathbf{k}_{\perp} = (k_x,k_y)$ the transverse component of the wavevector, and $\varphi=\arctan(k_{y}/k_{x})$ is the azimuthal angle of incidence. The associated basis spinors for propagating states in the lead are $(0,1)^{\mathrm{T}}$ and $(1,0)^{\mathrm{T}}$ for left- and right-moving modes, respectively. From the transmission amplitude $t_{0}(\mathbf{k}_{\perp})$ we compute the clean-limit conductance and Fano factor as $G_{0}=\frac{e^{2}}{h}\mathrm{tr}[t_{0}^{\dagger}t_{0}]$ and $F_{0}=\mathrm{tr}[t_{0}t_{0}^{\dagger}(1-t_{0}t_{0}^{\dagger})]/\mathrm{tr}[t_{0}t_{0}^{\dagger}]$ [@NazarovBlanterBook]. Modes with $k_{\perp}\gg(L\eta)^{-1/2}\equiv k_{\perp}^{\star}\left(L\right)$ are strongly suppressed in transmission and the spacing of the quantized transversal wave vectors in a finite sample is $\Delta k_{\perp}=2\pi/W$. If $\Delta k_{\perp}\ll k_{\perp}^{*}$, we can compute conductance and Fano factor analytically by replacing the sum over discrete modes $\mathbf{k}_{\perp}$ by an integral and find $$\begin{aligned} G_{0}\left(W,L\right) & = & \frac{e^{2}}{h}\frac{1}{2\pi\eta}\frac{W^{2}}{L},\label{eq:G0-1}\\ F_{0}\left(W,L\right) & = & 1/3,\label{eq:F0-1}\end{aligned}$$ which resembles transport in a diffusive conductor with conductivity $\sigma_{0}=e^{2}/(2\pi h\eta)$. Thus, the clean DWN has pseudodiffusive transport characteristics — similar to Dirac electrons in two dimensions [@Tworzydlo2006]. Disordered case --------------- We extend the scattering matrix approach to include a Gaussian disorder potential $U(\mathbf{r})$ with correlations as in Eq. and $\xi=\eta/2$ like in the density of states calculation. We compute the transmission matrix of the disordered DWN $H+U(\mathbf{r})$ by concatenating the reflection and transmission amplitudes of a thin slice of DWN without disorder, see Eq. (\[eq:S0-1\]), with reflection and transmission matrices of a thin slice with disorder, which can be calculated using the first-order Born approximation, and repeating this procedure for many slices. We apply periodic or antiperiodic boundary conditions in the $x$ and $y$ directions, cutting off the number of transverse modes to keep the dimensions of the transmission and reflection matrices finite. We take the mode cutoff large enough and the slice length thin enough so that the results do not depend on either, and we take $\Delta k_{\perp}/k_{\perp}^{*}$ small enough that the results do not depend on the choice of the boundary conditions. A similar method has been previously applied to study disordered Dirac materials in two [@Bardarson2007; @Adam2009] and in three dimensions [@Sbierski2015; @Sbierski2014a], and we refer to those references for more details on the numerical method. Figure \[fig:Transport-properties\] shows our results for the resistance $R=1/\langle{G}\rangle_{\mathrm{dis}}$ as a function of sample length $L$, where $\langle {...}\rangle_{\mathrm{dis}}$ denotes an average over 60 disorder realizations as well as the two choices for the boundary conditions, to further suppress statistical uncertainty. Compared to the clean pseudo-resistance $R_{0}=L/(\sigma_{0}W^{2})$, the resistance of the disordered samples is slightly decreased by up to about 10 percent, see top panel. The difference $\Delta R=R_{0}-R$ is shown in the bottom panel. For the smallest disorder strength considered, $K=1$, $\Delta R$ scales linearly with $L$ for the system lengths considered, for intermediate $K=2,4,6$, $\Delta R$ is not a linear function of $L$ but instead has an “S”-like dependence, which prevents any meaningful assignment of a (change of the) bulk resistivity. The resistance at the largest system size $R(L_{\rm max}=72\eta)$, shows a non-monotonous behavior with increasing disorder strength. For larger $K=8,10,14$, the $\Delta R$ traces are purely convex and tend to be linear for large $L$. We have also investigated the Fano factor which stays around $F=1/3$ (not shown) for all values of $K$. ![[\[fig:Transport-properties\]Resistance for disordered DWN for disorder strengths $K=1$, $2$, $4$, $6$, $8$, $10$, and $14$ and disorder correlation length $\xi=\eta/2$. Results are averaged over periodic and antiperiodic transverse boundary conditions and over 60 disorder realizations per boundary condition. The dashed line indicates the clean-limit pseudodiffusive result of Eq. (\[eq:G0-1\]). We chose $W=72\eta$, and keep transverse modes with $|k_{x,y}|<2\pi M/W$ with $M=34$. ]{}](transport_disordered_DWN.pdf) Discussion ---------- A finite lifetime $1/\Gamma$ implies diffusive transport with resistance scaling $R\propto L$. While this is (approximately) observed in our transport simulations for $K>0$, see Fig. \[fig:Transport-properties\] top panel, the difficulty lies in the discrimination to transport behavior associated to the clean fixed point $K=0$: Being pseudodiffusive, the same resistance scaling holds, albeit for the very different reason of evanescent wave physics and not due to scattering between transport channels as in diffusive transport. To discriminate between the pseudodiffusive and diffusive regimes, in Fig. \[fig:CrossoverLength\] (top) we show the probability $P_{\rm t}(L)$ that an electron is transmitted in the same transverse mode as it enters — for which we take the mode with $\mathbf{k}_{\perp} = 0$ —, conditional on the probability that it is transmitted, $$P_{\rm t}(L) = \frac{|t(0,0)|^2}{\sum_{\mathbf{k}_{\perp}} |t(\mathbf{k}_{\perp},0)|^2},$$ where $t(\mathbf{k}_{\perp}^{\rm{out}},\mathbf{k}_{\perp}^{\rm{in}})$ is the transmission amplitude of the disordered system at length $L$, $\mathbf{k}_{\perp}^{\rm{out}}$ and $\mathbf{k}_{\perp}^{\rm{in}}$ referring to the incoming and outgoing transverse modes. The conditional probability $P_{\rm{t}}$ is an indicator of the transition between the pseudodiffusive and diffusive regimes: At the pseudodiffusive fixed point $K=0$ one has $P_{\rm{t}}(L) = 1$, as translational translational invariance ensures that $t(\mathbf{k}_{\perp}^{\rm out},\mathbf{k}_{\perp}^{\rm in})$ is diagonal in the transversal mode indices $\mathbf{k}_{\perp}^{\rm in}$ and $\mathbf{k}_{\perp}^{\rm out}$, see . In contrast, diffusive transport is characterized by scattering between transverse modes. For sufficiently long diffusive samples with many transverse modes one therefore expects $P_{\rm{t}}(L) \to 1/N_{\perp}$, where $N_{\perp}$ is the total number of transverse modes. For finite-length samples $P_{\rm{t}}(L)$ is expected to approach this asymptotic value from above, starting from $P_{\rm{t}}(0) = 1$ in the limit of zero sample length. For the disordered DWN system our data in Fig. \[fig:CrossoverLength\] (top) indeed indicates a monotonous decrease of $P_{\rm{t}}(L)$ with $L$ and a saturation at large $L$ for disorder strengths $K > 4$. Although no saturation could be observed for weaker disorder strength at the system sizes we could access in our numerical calculations, we found no sign that $P_{\rm{t}}(L)$ behaves differently for $K < 4$, consistent with with a flow to a diffusive fixed point even for weak disorder. On the other hand, if weak disorder is an irrelevant perturbation (as it is in the case of a single Weyl node) and the pseudodiffusive fixed point would be stable, we would expect that an initial decrease of $P_{\rm{t}}(L)$ with $L$ is compensated by increase of $P_{\rm{t}}(L)$ at larger lenghts, a behavior that we confirmed for the weakly disordered SWN (data not shown). As long as $P_{\rm t} \gg 1/N_{\rm t}$, where $N_{\rm t} \sim h G/e^2 \sum_{\mathbf{k}_{\perp}} |t(\mathbf{k}_{\perp},0)|^2$ is the (effective) number of transverse modes participating in the transmission, a condition that is met for the entire parameter range we consider, we expect that $P_{\rm t}(L)$ has the functional form $$P_{\rm t}(L) = e^{-(L-L_{c})/L^{\star}},$$ where the characteristic length scale $L^{\star}$ can be identified with the mean free path and $L_{c}$ a length scale that accounts for transient effects at the sample-lead boundary, leading to a quick initial decrease of $P_{\rm t}$ for short lengths, in particular visible for $K\lesssim4$. The lower panel of Fig. \[fig:CrossoverLength\] shows fits of $L^{\star}$ based on the large-$L$ asymptotics of $P_{\rm t}(L)$. The $K$ dependence of $L^{\star}$ is consistent with the expectation based on Eq. , $L^{\star} \sim \hbar v/\Gamma\sim \hbar v/\Lambda e^{A/K}$. We disregard the data points at $K=1$ and $K=2$, for which no reliable asymptotic large-$L$ fit could be made. ![[\[fig:CrossoverLength\] Top: Conditional same-mode transmission probability $P_{\rm t}(L)$ for disorder strengths $K=0, 1$, $2$, $3$, $4$, $5$, $6$, $8$, $10$, $12$, and $14$ (top to bottom curve). The dashed lines denote fits to the form $\exp(-(L-L_{c})/L^{\star})$. Bottom: Mean free path $L^{\star}$, obtained from the fits to $P_{\rm t}(L)$, versus $1/K$.]{}](transport_disordered_DWN_T0_L72.pdf) The curves for the difference $\Delta R(L)$ of the resistances in the clean and disordered cases in Fig. \[fig:Transport-properties\] can be understood in terms of a crossover from pseudodiffusive to diffusive transport as well. The length scale $L^{\star}(K)$ roughly coincides with the length scale where the second derivative of the resistance vs. sample length curve vanishes. For the weakest disorder strength we consider the maximum sample length $L_{\rm max}$ is still much smaller than the characteristic length $L^*$ of the pseudodiffusive-to-diffusive crossover. For this disorder strength, pseudodiffusive behavior prevails for all system sizes we consider, albeit with a resistance that is slighlty smaller than $R_{0}$. A decrease of the resistivity has also been observed as a finite-size effect for a SWN at weak disorder strengths [@Sbierski2014a]. A systematic decrease of the resistivity could in principle arise as a consequence of a disorder-induced renormalization of the parameters $v$ and $\eta$ in the Hamiltonian (\[eq:H\]). For a bulk system, the renormalized parameters $v_{\rm eff}$ and $\eta_{\rm eff}$ can be calculated in the Born approximation, which yields an increased effective length scale $\eta_{\rm eff}>\eta$. Replacing $\eta$ by $\eta_{\rm eff}$ in the expression for clean conductivity of a finite system, $\sigma_{0,{\rm eff}}=e^{2}/(2\pi h\eta_{\rm eff})$ predicts an increase of the resistance, in conflict with our numerical observation. We conclude that a disorder-induced renormalization of the parameters $v$ and $\eta$ is not the explanation of the observed decrease of the resistivity. A more careful analysis of the finite-size effects could be attempted along the lines of Ref. [@Schuessler2009]. For strong disorder, $K\gtrsim4$, the characteristic length scale $L^{\star}\left(K\right)$ drops below $L_{\rm max}$ and diffusive behavior can be observed, see, [*e.g.*]{}, the resistance data for $L/\eta \gtrsim 40$ and $K=14$). Such a diffusive regime is also commonly found in other topological semimetals, such as a two-dimensional Dirac- or a three-dimensional simple Weyl node: Although disorder tends to decrease the mean free path, the conductance is still increased by the disorder-induced increase of the density of states, while band topology and, in three dimensions, standard single-parameter scaling arguments, prohibit Anderson localization [@Bardarson2007; @Nomura2007; @Sbierski2014a]. Conclusion ========== We have investigated the effects of potential disorder for a double Weyl node, using numerically exact quantum transport simulations in a mesoscopic setup for chemical potential at the nodal point as well as density of states calculations based on the self-consistent Born approximation and the Kernel Polynomial method for a range of energies. Our findings indicate that disorder physics in a double Weyl node is more conventional than in its linearly dispersing counterpart with unit chiral charge, which features a disorder induced quantum phase transition with the density of states at zero energy as an order parameter. In the double Weyl node, any finite disorder strength induces a finite quasiparticle lifetime $\tau$ at the nodal point. Our numerical and analytical calculations are consistent with previous predictions by Goswami and Nevidomskyy, indicating that the lifetime $\tau$ is exponentially large in the inverse disorder strength [@Goswami2014]. Unfortunately, a quantitative comparison of our calculations for the density of states and our transport simulations is hindered by the fact that only the SCBA can give an estimate for the quasiparticle lifetime $\tau$. However, since the SCBA density of states does not agree quantitatively with the data from KPM at $\varepsilon=0$, we must also discard its predicted value of $\tau$ for quantitative checks. The density of states, as simulated by the KPM is however a quantity integrated over k-space [\[]{}see Eq. and cannot be translated into a value for $\tau$ without further assumptions. In Ref. [@Sbierski2015], the disorder-induced phase transition point in a SWN was identified using the condition of scale invariance of the (median) conductance. We repeated a similar analysis with conductance data obtained for the disordered DWN from Sec. \[sec:Quantum-transport\] but could not find a scale invariant point (data not shown). This is consistent with the absence of a disorder induced phase transition in a DWN bandstructure. For technical convenience, we have used a model for the double Weyl node with continuous rotational symmetry [\[]{}$\eta_{x}=\eta_{y}=\eta$ in Eq. . In additional numerical calculations we checked that our conclusions do not qualitatively change when $\eta_{x}\neq\eta_{y}$ and the rotational symmetry is reduced to be fourfold. Acknowledgments {#acknowledgments .unnumbered} =============== We gratefully acknowledge discussions with Johannes Reuther and Achim Rosch as well as financial support by the Helmholtz Virtual Institute “New states of matter and their excitations” and the CRC/Transregio 183 (Project A02) of the Deutsche Forschungsgemeinschaft. We thank Jens Dreger for support on the computations done on the HPC cluster of Fachbereich Physik at FU Berlin. Appendix: SCBA equations {#appendix-scba-equations .unnumbered} ======================== Using the identities $\int_{0}^{2\pi}d\theta\cos\left(2\theta\right)\exp\left[x\,\text{cos}[\phi-\theta]\right]=2\pi\cos\left(2\phi\right)I_{2}(x)$ and $\int_{0}^{2\pi}d\theta\exp\left[x\,\text{cos}[\phi-\theta]\right]=2\pi I_{0}(x)$ where $I_{k}(x)$ is the modified Bessel function of the $k$-th kind, we find the following self-consistency equations from Eq. with the Ansatz : $$\begin{aligned} M(P,P_{z}) & = & \frac{-Kr}{\left(2\pi\right)^{2}}\int_{0}^{\infty}dQ\int_{-\infty}^{\infty}dQ_{z}\left[Q^{2}/2+M(Q,Q_{z})\right]I_{2}(QPr^{2})\label{eq:SCBAI}\\ M_{z}(P,P_{z}) & = & \frac{-Kr}{\left(2\pi\right)^{2}}\int_{0}^{\infty}dQ\int_{-\infty}^{\infty}dQ_{z}\left[Q_{z}+M_{z}(Q,Q_{z})\right]I_{0}(QPr^{2})\label{eq:SCBAII}\\ M_{0}(P,P_{z}) & = & \frac{Kr}{\left(2\pi\right)^{2}}\int_{0}^{\infty}dQ\int_{-\infty}^{\infty}dQ_{z}\left[M_{0}(Q,Q_{z})-iE\right]I_{0}(QPr^{2})\label{eq:SCBAIII}\end{aligned}$$ where $$\begin{aligned} U(Q,Q_{z}) & = & Q\frac{\exp\left[-r^{2}\left(P^{2}+Q^{2}+(Q_{z}-P_{z})^{2}\right)/2\right]}{[Q^{2}/2+M(Q,Q_{z})^{2}]+[Q_{z}+M_{z}(Q,Q_{z})]^{2}-[E+iM_{0}(Q,Q_{z})]^{2}}\end{aligned}$$ and $r=\xi/\eta$, $E/(\hbar v/\eta)=\varepsilon$, $M(Q=q\eta,Q_{z}=q_{z}\eta)\equiv m(q,q_{z})\eta$, and analogous for $M_{z}$ and $M_{0}$. Eqns. to can be solved numerically by iteration.

ArXiv

--- author: - | [^1]\ MTA Wigner RCP\ E-mail: title: Angular correlation results from ALICE --- Introduction ============ In high energy collisions, quark-antiquark pairs are produced, which fragment and hadronize into a spray of collimated hadrons, called jets. These jets interact with the hot and dense medium and their interactions can be studied in angular correlation measurements. In these measurements, the distribution of the azimuthal angle ($\Delta\varphi$) and pseudorapidity ($\Delta\eta$) difference of a trigger and an associated object is extracted. These objects can be identified or unidentified hadrons or even the jets themselves. Depending on the type of the object, different properties of the plasma and of the interactions can be studied. The data used for the presented measurements were collected by the ALICE detector, of which a detailed description can be found in Ref. [@ALICE]. The ALICE detector has good tracking capabilities down to low transverse momentum ($p_{\rm T}$) and excellent particle identification, making it suitable for studying heavy-ion collisions as well as small collision systems (e.g., pp or p–Pb). Hadron-hadron correlations in Pb–Pb =================================== In angular correlation measurements of two hadrons, i.e., when the trigger and associated objects are both hadrons, jets appear as a peak around $(\Delta\varphi,\Delta\eta) = (0,0)$ (near side) and an elongated structure in $\Delta\eta$ at $\Delta\varphi=\pi$ (away side). By studying the centrality (the overlap of the two nuclei) and $p_{\rm T}$ dependence of these structures in Pb–Pb collisions, the modifications of the jets while traversing the hot and dense medium can be explored. In \[fig:yieldPbPb\], yields from Pb–Pb collisions of the jet peak and of the away-side structure divided by the same from pp collisions (denoted by $I_{\rm AA}$) are shown for unidentified hadron-hadron and $\pi^0$–hadron correlations as a function of the $p_{\rm T}$ of the associated particle [@Adam:2016xbp]. A small enhancement on the near side and a small suppression on the away side can be seen at large $p_{\rm T}$, while at low $p_{\rm T}$, there is a large enhancement in both. The results are compatible in the unidentified and the identified cases in the region where both are measured. The away-side suppression can be explained by parton energy loss, while the enhancement at low $p_{\rm T}$ can originate from several sources (e.g., $k_{\rm T}$ broadening, medium-excitation or fragments from radiated gluons [@Vitev; @Ma; @Wang; @Kopeliovich; @Gyulassy]). The enhancement on the near side might arise from a change of the fragmentation function or the quark-to-gluon jet ratio [@Aamodt]. Apart from the yield of the jet peak, the shape of the peak can also be modified by the interaction of the partons of the jets with the medium. In \[fig:widthTwoPanelPRC\], the width of the peak from Pb–Pb collisions can be seen as a function of the centrality, and the results from pp collisions are shown as a reference [@Adam:2016tsv; @Adam:2016ckp]. The peak at low $p_{\rm T}$ gets broader towards central collisions in Pb–Pb collisions in both the $\Delta\varphi$ and the $\Delta\eta$ directions. In peripheral collisions, it converges to the value in pp collisions in the $\Delta\varphi$ direction, while in the $\Delta\eta$ direction even the peripheral Pb–Pb results are above the pp results. At low $p_{\rm T}$, the width is larger in the $\Delta\eta$ direction at the same $p_{\rm T}$ and centrality than in the $\Delta\varphi$ direction, resulting in an asymmetric peak. The multiplicity dependence of the yield and the width was also measured in pp collisions and only a very small dependence was found, therefore the use of minimum bias pp data as a reference was validated. [2016-Sep-19-widthTwoPanelPRC.pdf]{} In Pb–Pb collisions, at the lowest measured $p_{\rm T}$ and at the highest centrality, an unexpected depletion around $(\Delta\varphi,\Delta\eta) = (0,0)$ is seen [@Adam:2016tsv; @Adam:2016ckp]. In \[subfig:results1c\], this depletion is illustrated, while in \[subfig:depletion\_AMPT\_comparison\], the magnitude of the missing yield is presented. The results from hadron-hadron correlations in Pb–Pb collisions were compared to A Multi-Phase Transport (AMPT) model [@AMPT1; @AMPT2], which exhibits both the broadening and the depletion with the settings used, and also show large radial flow. This comparison suggests therefore that these features are caused by an interplay of the jets with the flowing medium. Path-length dependence in Pb–Pb =============================== The dependence on the length of the path the partons of a jet traveled in the medium can be characterized with different methods. Firstly, two-plus-one correlations are presented. In these studies, the angular distributions are calculated for two back-to-back trigger hadrons. The first trigger hadron has a higher $p_{\rm T}$, therefore it is assumed that this has traveled a shorter length in the plasma and has lost less energy than its back-to-back partner. The associated particles with this trigger particle form a peak, which corresponds to the leading jet. The associated particles with the second trigger hadron also form a peak, which corresponds to the back-to-back jet, the partons of which have traveled a longer path in the medium. The yield in central Pb–Pb collisions divided by the yield from peripheral collisions ($I_{\rm CP}$) from the peak for both trigger particles is shown in \[fig:two-plus-one\]. In the case when the two trigger particles have similar $p_{\rm T}$ (\[subfig:2017-Feb-03-Icp\_8\_12\]), the $I_{\rm CP}$ corresponding to the two trigger particles are the same, while in the case when the second trigger particle has a much lower $p_{\rm T}$ than the first (\[subfig:2017-Feb-03-Icp\_16\_20\]), the $I_{\rm CP}$ belonging to the second trigger particle is enhanced at low $p_{\rm T}$. This can be explained by a path length dependent suppression, but other explanations (e.g., a bias on the trigger selection) are also possible. Secondly, the path length dependence of jets can be studied by taking jets as the trigger objects and restricting their direction with respect to the direction of the event plane (the plane defined by the direction of the colliding nuclei and the vector connecting the centers of the colliding nuclei). In this way, jets that are parallel with the event plane travel a smaller path in the medium than jets that go perpendicular to the event plane. In \[fig:jet-hadron\], the yield of hadrons associated with jets is presented for three categories of jets depending on their angle with respect to the event plane. The yields agree within uncertainties for the three cases, therefore no dependence on the path length is found. This suggests that the modifications seen from the two-plus-one correlations are connected to another effect. Small systems ============= Angular correlations of identified particles in small systems can be used to study the production mechanisms and different conservation laws of the studied particles. In \[fig:pp\_identified\], identified hadron-hadron correlations are presented from pp collisions [@Adam:2016iwf]. When mesons (e.g., pions or kaons) are considered, a jet peak is visible, as expected, in both the like-sign and the unlike-sign case (not shown here). However, when baryons are considered, the jet peak is only visible in the unlike-sign case (\[subfig:BaryonsToPP\_UnLikeSign\]). In the like-sign case, a depletion is present instead of the peak (\[subfig:BaryonsToPP\_LikeSign\]). The size of this depletion is the same if protons or $\Lambda$ particles are considered, and it is not described by models, therefore its origin is currently not understood. Summary ======= Angular correlation measurements are powerful tools to study the interaction of the partons of jets with the medium in heavy-ion correlations, and to study the production mechanisms and conservation laws in small systems. In Pb–Pb collisions, an enhancement of the jet-peak yield and a broadening of the jet peak at low transverse momentum in central collisions were presented together with an unexpected depletion around $(\Delta\varphi,\Delta\eta) = (0,0)$. Furthermore, studies on the path length dependence of the measured quantities were shown, with no clear path length dependence observed. Finally, measurements from pp collisions were presented, where an unexpected depletion of the jet peak is visible when baryon-baryon like-sign correlations are studied. [99]{} Collaboration, *JINST* [**3**]{} (2008) S08002 Collaboration, *Phys. Lett.* [**B763**]{} (2016) 238-250 I. Vitev, *Phys. Lett.* [**B630**]{} (2005) 78-84 G.-L. Ma and X.-N. Wang, *Phys. Rev. Lett.* [**106**]{} (2011) 162301 X.-N. Wang, *Phys. Rev.* C61 (2000) 064910 B. Z. Kopeliovich, J. Nemchik, A. Schafer, and A. V. Tarasov, *Phys. Rev. Lett.* [**88**]{} (2002) 232303 I. Vitev and M. Gyulassy, *Phys. Rev. Lett.* [**89**]{} (2002) 252301 Collaboration, *Phys. Rev. Lett.* [**108**]{} (2012) 092301 Collaboration, *Phys. Rev. Lett.* [**119**]{}, (2017) 102301 Collaboration, *Phys. Rev.* [**C96**]{}, (2017) 034904 Z.-W. Lin, C. M. Ko, B.-A. Li, B. Zhang, and S. Pal, *Phys. Rev.* [**C72**]{} (2005) 064901 J. Xu and C. M. Ko, *Phys. Rev.* [**C83**]{} (2011) 034904 Collaboration, *Eur.Phys.J.* [**C77**]{} (2017) no.8, 569 [^1]: on behalf of the ALICE collaboration.

ArXiv

--- abstract: 'We present the results of an extragalactic point source search using the five-year WMAP 41, 61 and 94 GHz (Q-, V- and W-band) temperature maps. This work is an extension of our designing and applying a CMB-free technique to extract point sources in the WMAP maps. Specifically, we have formed an internal linear combination (ILC) map of the three-band maps, with the weights chosen to remove the CMB anisotropy signal as well as to favor the selection of flat-spectrum sources. We have also constructed a filter to recover the true point source flux distribution on the sky. A total of 381 sources are found in our study at the $> 5\sigma$ level outside the WMAP point source detection mask, among which 89 are “new” (i.e., not present in the WMAP catalogs). Source fluxes have been calculated and corrected for the Eddington bias. We have solidly identified 367 ($96.3\%$) of our sources, the 1$\sigma$ positional uncertainty of which is 2$^\prime$. The 14 unidentified sources could be either extended radio structure or obscured by Galactic emission. We have also applied the same detection approach to simulated maps, which yielded 364$\pm$21 detections on average. The recovered source distribution $N(>S)$ agrees well with the simulation input, which proves the reliability of this method.' author: - 'X. Chen & E. L. Wright' title: 'Extragalactic Point Source Search in Five-year WMAP 41, 61 and 94 GHz Maps' --- Introduction ============ The Wilkinson Microwave Anisotropy Probe (WMAP) is designed to advance observational cosmology by making precise measurements of the cosmic microwave background (CMB) anisotropy (@2003ApJ...583....1B). The characterizing and removing of foreground emission play a vital role in the interpretation of CMB temperature maps. At small angular scales, extragalactic radio sources, especially those with flat spectra, constitute the most important foreground. Therefore, it is essential to have efficient source detection techniques developed and applied to the WMAP maps. Meanwhile, since WMAP provides the only all-sky millimeter survey, source extraction in its maps also opens a window to study extragalactic radio sources in the millimeter wavelength region. With every WMAP data release, a catalog is provided of the brightest point sources in the WMAP maps. The detection procedure used by the WMAP science team features a filtering of the integration-time-weighted maps with $b_{\ell}/(b_{\ell}^{2}C_{\ell}^{CMB}+C_{\ell}^{noise})$ in harmonic space and a search for $> 5\sigma$ peaks, where $b_{\ell}$ is the transfer function of the WMAP beam response (@2003ApJS..148...39P, @2007ApJS..170..263J, @2009ApJS..180..246H), $C_{\ell}^{CMB}$ is the CMB angular power spectrum and $C_{\ell}^{noise}$ is the noise power. This procedure successfully generated 208, 323 and 390 point sources in the WMAP first year, three-year and five-year maps, respectively (@2003ApJS..148...97B, @2007ApJS..170..288H, @2008arXiv0803.0577W; hereafter, WMAP1, WMAP3 and WMAP5 catalogs). However, because of the limited angular resolution of WMAP, positive CMB excursions can be confused with point sources. In addition, the $C_{\ell}^{CMB}$ term in the filter counts as a systematic noise that does not integrate down with observing time. This largely limits the increasing of the number of detections with increased sensitivity of the maps. To circumvent the CMB “noise”, @2008ApJ...681..747C introduced a CMB-free technique, which involves forming internal linear combination (ILC) of multi-frequency maps to suppress the CMB signal. The number of sources N found by applying this technique to the WMAP V- and W-band maps alone varied as $t^{0.72}$ from one year to five years, in comparison to $N \sim t^{0.39}$ between the WMAP1 and WMAP5 catalogs (@2009ApJS..180..283W). In this paper, we extend this CMB-free technique and employ it to the WMAP five-year Q-, V- and W-band temperature maps. The main goal is to utilize the high signal-to-noise ratio Q-band data to find more flat-spectrum sources, which are known to dominate the 30 - 100 GHz frequency regime. Methodology =========== Point Source Detection ---------------------- The WMAP temperature maps at HEALPix[^1] Res 9 are used in this study. To minimize the noise difference between maps, we first smooth the high resolution V- and W-band maps to match the Q-band resolution. We construct the smoothing functions using a polynomial of an infinitely smooth function with compact support, i.e., $$S(\theta)~=~\sum_n~a_nS_n~(\theta),$$ where $$S_n~(\theta) = \left\{ \begin{array}{ll} \exp(- \frac {n \theta^2}{\theta_{Rc}^{2} - \theta^2}) & \theta < \theta_{Rc} \\ 0 & \theta > \theta_{Rc}. \end{array}\right. \label{eq:smooth}$$ The coefficients $a_{n}$’s are fitted by requiring the Legendre transform of the smoothing functions to match the ratios of Q-band beam transfer function to V-band and W-band beam transfer functions separately. And a cutoff radius $\theta_{R_{c}}$ of 1.25$\degr$ is chosen for the solid angle integrations. We then form the ILC map out of the Q-band map, V-band smoothed map and W-band smoothed map, $$T^{ILC} ~=~ \omega_Q T_Q + \omega_V T_V^{SM} + \omega_W T_W^{SM}.$$ The weights are determined by canceling out the CMB anisotropy signal $$\omega_Q + \omega_V + \omega_W = 0,$$ normalizing the flat-spectrum source flux to the Q-band $$\omega_Q + \omega_V (\frac {\partial B_Q}{\partial T} / \frac{\partial B_V}{\partial T})|_{T_0} + \omega_W (\frac {\partial B_Q}{\partial T} / \frac{\partial B_W}{\partial T})|_{T_0} = 1,$$ and minimizing the variance of the combination map, which can be approached by minimizing ($\omega_Q^2 + \omega_V^2 + \omega_W^2$) since the variance at each pixel is approximately the same for Q, V and W bands. Here $B_{\nu}$ is the Planck function and $T_{0}$ = 2.725$\pm$0.002 K is the CMB temperature (@1999ApJ...512..511M). The three-band ILC map is shown in Figure \[ILC\]a. For any pixel $i$ within the Q-band beam to a point source, assuming negligible contribution from the overlap of point sources, its temperature $T_i^{ILC}$ in the ILC map can be fitted into the point source intensity $f$, multiplied by the beam response at that pixel location $b^Q_i$, plus a local baseline $c$, i.e., $$T_i^{ILC} ~=~ f b^Q_i ~+~ c. \label{eq:srcfit}$$ Deviation from the fit is evaluated using the $\chi^2$, $$\chi^2 ~=~ \sum_{ij} \epsilon_i ( N^{-1}_{ij} ) \epsilon_j,$$ where $$\epsilon_i ~=~ T_i^{ILC} ~-~ (f b^Q_i + c)$$ and $$N_{ij} ~=~\delta_{ij} \sigma_Q^2 \omega_Q^2 ~+~ \sum_k S^{V2Q}_{ik} S^{V2Q}_{jk} \sigma_V^2 \omega_V^2 ~+~ \sum_k S^{W2Q}_{ik} S^{W2Q}_{jk} \sigma_W^2 \omega_W^2.$$ Here $S^{V2Q}$ and $S^{W2Q}$ are the smoothing functions used to smooth the V-band and W-band maps previously; the summation over $k$ is a sum over all the pixels within $\theta_{R_{c}}$ of pixel $i$ and $j$; $\sigma_Q$, $\sigma_V$ and $\sigma_W$ are the noise at pixel $k$ in the Q-, V- and W-band maps, respectively. The best fit is defined when the $\chi^2$ reaches a minimum, which gives $$\sum_{i j} N^{-1}_{i j} \left( \begin{array}{c} b_i^Q \\ 1 \end{array} \right) \left( \begin{array}{cc} b_j^Q & 1 \end{array} \right) \left( \begin{array}{c} f \\ c \end{array} \right) = \sum_{i j} N^{-1}_{i j} T_j^{ILC} \left( \begin{array}{c} b_i^Q \\ 1 \end{array} \right).$$ Solving the above equation, we find that the intensity of the point source can be written as a weighted sum of the intensities of its surrounding pixels, i.e., $$f = \sum_{j} \left[\sum_{i} (M^{-1})_{00} b_i^Q N^{-1}_{i j} + \sum_{i} (M^{-1})_{01} N^{-1}_{i j} \right] T_j^{ILC}, \label{eq:kernel}$$ where $$M = \sum_{i j} N^{-1}_{i j} \left( \begin{array}{c} b_i^Q \\ 1 \end{array} \right) \left( \begin{array}{cc} b_j^Q & 1 \end{array} \right).$$ Since the weights inside the square brackets of Equation (\[eq:kernel\]) are essentially a function of the angular distance to the considered source, this suggests that we can probe the point source flux in the ILC map by fitting the function and filter the map locally. As a first attempt, we generate 32 random source positions in the sky and for each position calculate the weights to a distance of 1.25$\degr$. We plot the weights against the angular distances at each selected random spot, and note that they all share the same profile. This leads us to fit a single function using the collective weights from the 32 random positions. Convolving the all-sky ILC map with this global filter, we are able to recover the point source flux distribution on the entire sky, shown in Figure \[ILC\]b. The extreme negative regions in the map are regions surrounding the brightest sources, where fluxes in the pixels are depleted and restored into their central point sources by filtering. We adopt the same Kp0+LMC+SMC detection mask used by the WMAP team (@2009ApJS..180..283W, shown as contours in Figure \[ILC\]b) to exclude extended foreground emission. Following the WMAP scan pattern that gives non-uniform observation numbers among pixels: greatest at the ecliptic poles, high at latitude $\pm45\degr$, and least in the ecliptic plane, we divide the masked sky into rings that cover 1 degree in ecliptic latitude. We then compute $\sigma$ individually within each ring and look for peaks greater than $5\sigma$. When more than one $> 5\sigma$ pixel lay within a 3 $\times$ 3 pixels area, the brightest one is chosen as a source detection. The accurate source position is approached by fitting a point source profile to the 9-pixel grid centered on the brightest pixel and defined at where the peak is. We find 381 sources in our study. The distribution of these sources on the sky is given in Figure \[overview\]; Blue, green and red dots indicate, respectively, sources in the WMAP catalogs, newly detected and identified, and newly detected but unidentified (See §\[id\] for a detailed discussion of source identification). Flux Estimation --------------- We estimate the Q-, V- and W-band flux densities of our sources by calculating $$F_{\nu} = \int I_{\nu}~\psi(\theta) \cos \theta d\Omega~=~\frac {\partial B_{\nu}}{\partial T} |_{T_{0}} \int T ~\psi(\theta) \cos \theta d\Omega,$$ where $B_{\nu}$ is again the Planck function and $T_{0}$ is the CMB temperature of 2.725$\pm$0.002 K. $T$ is the temperature measurement in each band; the solid angle is integrated to a radius of 1.25 degree. We have introduced a weighting function $\psi(\theta)$ in our flux estimator to help enhance the contrast of the point source flux to the background. This function is constructed individually in each band with the corresponding beam profile in the center and a negative Gaussian ring in the outer field. We require $$\int \psi(\theta) d\Omega = 0,$$ in order to ignore any flux that spreads uniformly across the whole integration field. The mean and variance of the Gaussian functions are optimized to preserve the source flux while minimizing the fluctuation in the integration field: $$\sigma^2 = \sum_i \sum_j \psi_i \psi_j (C(\theta_{ij}) + \delta_{ij} N_i),$$ where $$C(\theta) = \sum_{\ell} \frac {2\ell + 1}{4 \pi} C_{\ell} \omega_{\ell} P_{\ell} (cos \theta).$$ Here i and j denote pixels within the integration field; N is the radiometer noise in the five-year maps; $C_{\ell}$ is the best fit power spectrum of WMAP5 data assuming a $\Lambda$CDM model; $\omega_{\ell}$ is the window function which encodes the beam smoothing (@2003ApJS..148...39P); $P_{\ell}$ is the Legendre function. We see negative fluxes for some sources, which is likely caused by our source sitting on the top of a negative CMB fluctuation. The error on the flux density is calculated as the flux rms value of the background pixels that are close to the target source but are not “contaminated” by it. Specifically, we select a ring of pixels around the source with an inner radius equal to 1$\degr$ to exclude pixels affected by the source fluxes, and an outer radius of 2$\degr$ to include sufficient number of pixels to estimate the background fluctuation level. Since flux estimates of sources in noisy fields are on average overestimated (Eddington Bias, @1940MNRAS.100..354E), we have applied the Bayesian approach described in the Appendix B of @2007ApJS..170..108L to correct for the bias. The estimated slopes of the differential number counts are 2.67, 2.68 and 2.30 for the Q, V, and W bands separately. We compare the uncorrected fluxes of each band with the corrected ones in Figure \[eddington\]. The agreement is generally good, and the correction makes almost no difference for fluxes $\ga$ 1.5 Jy in Q and V bands. Identification {#id} -------------- We search in the NASA/IPAC Extragalactic Database (NED[^2]) and the source catalogs from two recent all/large sky radio surveys: CRATES (8.4 GHz, @2007ApJS..171...61H) and the Australia Telescope 20 GHz Survey (AT20G, @2008MNRAS.384..775M), for radio sources within 15$^\prime$ ($\theta^{Q}_{FWHM} \sim 30.6^\prime$) to our sources. We retrieve the available photometric data and make the spectral energy distribution (SED) plots for these radio sources. If the flux densities of our source can be reasonably fitted into the SED of a known radio source, we associate them. This SED approach identifies 367 (i.e. 96.3$\%$) sources on our list, the majority of which are found to have a flat spectrum, i.e., ${\alpha} \sim 0 $ in $F_{\nu} \sim \nu^{\alpha}$. This is consistent with our knowledge of the dominance of flat-spectrum sources in this wavelength regime, and with the weights we chosen for the ILC map that favors flat-spectrum sources. Nevertheless, we do detect some steep-spectrum sources as well as sources with spectra peaked at $\sim$10 GHz (see Figure \[sedexample\] for some SED examples). We have also searched NED for the optical identifications of our sources and found 269 QSOs, 61 galaxies, 1 galaxy pair, 1 galaxy triple, 1 planetary nebula and 34 unclassified radio sources among the identified ones. Additionally, NED provides redshifts for 306 identified sources. A plot of the redshift distribution of these sources is given in Figure \[redshift\], the median of z is 0.85. We then cross-correlate our list of identified sources with the WMAP catalogs. If a source is within 15$^\prime$ of a WMAP cataloged source, we tag it. We find 287 of our sources present in the WMAP5 catalog, and five more sources (J0734+5026, J1513-1013, J1553-7914, J1644-7713, J1648+4109) are in the WMAP3 catalog but missing from the WMAP5 catalog. The high associate rate clearly demonstrates that our method can generate results consistent with the method adopted by the WMAP team. In Figure \[fluxcompare\], we compare our flux estimates with the ones given in the WMAP catalogs. These two sets of fluxes agree very well with each other at $F_{\nu} \ga$ 1 Jy. Our estimates are in general lower than the WMAP ones for faint sources, which is partially due to the flux correction we made. The most deviated point in both Q and V bands is Fornax A, known as an extended source with a relatively weak core and two large extended lobes. A complete list of all the identified sources is given in Table \[src\_id\], sorted in ascending order of right ascension. For each source, we give its WMAP5 ID (if available), optical identification, flux densities along with 1$\sigma$ errors at the WMAP Q-, V- and W-band, its 5 GHz ID as given in GB6 (@1996ApJS..103..427G), PMN (@1994ApJS...90..179G [@1995ApJS...97..347G], @1994ApJS...91..111W [@1996ApJS..103..145W]), S5 (@1981AJ.....86..854K) or catalogs, and the angular distance to its 5 GHz counterpart. For four of our sources (J0422+0212, J0721+0403, J0805+6143, J1015+2258), we have assigned different 5GHz IDs than given in the WMAP5 catalog based on the SEDs of the 5 GHz sources. The position errors of our detections are evaluated with respect to the positions of their 5 GHz counterparts since the 5 GHz surveys have in general higher angular resolutions. Assuming the source deviations from their 5 GHz counterparts in both the galactic longitude and latitude axes are normally distributed, the angular distance $r$ to the 5 GHz IDs should satisfy a Rayleigh distribution and the radial positional uncertainty of our sources can then be estimated as $\sigma_r = median(r)/ \sqrt {\ln 4} = 2^\prime$. We conclude that our 5 GHz IDs are more reliable since we obtain a better positional accuracy than the WMAP5 survey (4$^\prime$ in both longitudes and latitudes). There are 14 sources left without any solid identification, as listed in Table \[src\_uid\] and noted in red in Figure \[overview\]. Looking back into the filtered ILC map (Figure \[ILC\]b), we do see bright emission features at the corresponding locations. We suggest that these sources could be either extended radio structure or obscured by Galactic emission. Considering the low or negative flux densities obtained for some of these sources, it is likely that some of them are located on the top of negative CMB fluctuations so that they do not show as strong sources in each band but show up in the CMB-free ILC map. We cross-correlate this list with the GB6, PMN, S5 and catalogs. If a 5 GHz source is within 15$^\prime$, we suggest it as the possible counterpart of our source. In cases when no 5 GHz source is found nearby, we use the closest 1.4 GHz NVSS (@1998AJ....115.1693C) source or 843 MHz SUMSS (@1999AJ....117.1578B, @2003MNRAS.342.1117M) source instead. Discussion ========== Source Number Counts -------------------- We have calculated the source number counts in bins of $\Delta$log S = 0.2 at Q-, V- and W-band[^3] and compared them with the predictions of the cosmological evolution model by . As illustrated in Figure \[numcnt\], the agreement is generally good above $\sim$ 2 Jy (i.e., log S = 0.3) in the Q and V bands. The low data points at lower fluxes are due to the incompleteness of our sources at these flux ranges. In the W-band, our data points agree more with a model rescaled by 0.86. The source counts derived from WMAP5 catalog in the corresponding bands are also plotted for comparison. They are mostly consistent with the source counts in this work. Chance-Coincidence Rate ----------------------- To estimate the influence of false identifications from random coincidences, we have generated an equal number (i.e., 381) of random positions outside the WMAP source detection mask. We repeat the same near-position search in NED. If a 5 GHz source is found to have $F_{\nu} \ga$ 100 mJy within 15$^\prime$ of a fake source position, we calculate the spectral index $\alpha$ using the 5 GHz flux estimate along with the 1.4 GHz flux estimate from the NVSS catalog or 843 MHz flux estimate from the SUMSS catalog (all such 5 GHz sources have a NVSS or SUMSS counterpart). We then use the spectral index to extrapolate this source’s 5 GHz flux to the WMAP 41 GHz channel (Q-band). We consider a source as identified when its Q-band flux is above 100 mJy. Following this approach, we associate 14 random sources with known radio sources, corresponding to a chance-coincidence rate of 3.7$\%$. In Figure \[posierr\], we plot the distribution of position offsets of real detections to their 5 GHz counterparts, in comparison with the distribution of the identified random sources. It is evident that the false identifications with random sources generally have a larger position error than the real peak positions. Based on this plot, we suggest that our source identifications within 5$^\prime$ are probably real. Analysis of Simulated Maps -------------------------- We repeat the same point source analysis on simulated maps constructed with point sources, CMB fluctuations, and radiometer noise. $10^6$ sources are sampled from a power law distribution $N(>S)$ at the WMAP Q-band (centered at 40.7 GHz), derived from WMAP5 sources with $F_Q > 1$ Jy as they appear to be complete to $\sim 1$ Jy (Figure \[intcnt\]). Spectral indices are chosen from a Gaussian with mean -0.09 and standard deviation 0.176 (@2009ApJS..180..283W), and the fluxes are scaled to the V- and W-band centers (60.8 GHz and 93.5 GHz). In each band, we calculate the appropriate temperature of every source and assign it to a random HEALPix pixel at Res 11 (a total of $12\times4^{11}$ pixels). These point source maps are then smoothed with the beam window function at each band and converted to Res 9 maps. Finally, we add in the Res 9 maps of CMB fluctuations produced using the best fit $C_{\ell}^\prime$s from WMAP5 data assuming a $\Lambda$CDM model, and of radiometer noise generated from the WMAP5 noise variance in each pixel. The point source detection process is then applied to these simulated maps, yielding on average 364$\pm$21 point sources for the 10 simulations we completed. When compared with the input point source maps, we find only 2 spurious sources out of the total detections from our simulations. The recovered $N(>S)$ agree remarkably well with the simulation input for fluxes $> 1$ Jy, which infers that our method is robust and the source counts we get from the WMAP5 ILC map are reliable for fluxes $> 1$ Jy. Conclusions =========== We present catalogs of point sources found in an ILC map constructed from the WMAP five-year Q-, V- and W-band maps. We have recovered 287 WMAP5 sources and five WMAP3 sources in our survey. Bypassing of CMB “noise” in the detection process brings us 89 new sources (23.4$\%$ of the total). We have obtained flux density estimates for our sources in each of the three bands. Most of the new sources are found to have low or even negative flux estimates in at least one band, which is probably due to the coincidence of their locations with negative CMB fluctuations. This will generally lead to missing of these sources in single frequency band searches. We identify our sources by searching in NED for nearby low frequency radio sources and fitting the fluxes of our sources to their SEDs. Of all the sources we detected, we lack of solid identifications for 14. The flux estimates and number count distribution of our identified sources are comparable with those from the WMAP5 catalog, which proves our method to be complimentary to as well as consistent with the WMAP approach. Since this CMB-free technique responds rapidly as observing time increases (@2009ApJS..180..283W), we expect a fast increase of new detections in more years of WMAP maps. We acknowledge the use of the Legacy Archive for Microwave Background Data Analysis (LAMBDA) to retrieve the WMAP data set. Support for LAMBDA is provided by the NASA Office of Space Science. This research has also made use of the VizieR catalog service and the HEALPix (@2005ApJ...622..759G) package. Bennett, C. L., et al. 2003a, , 583, 1 Bennett, C. L., et al. 2003c, , 148, 97 Bock, D. C.-J., Large, M. I., & Sadler, E. M. 1999, , 117, 1578 Chen, X., & Wright, E. L. 2008, , 681, 747 Condon, J. J., Cotton, W. D., Greisen, E. W., Yin, Q. F., Perley, R. A., Taylor, G. B., & Broderick, J. J. 1998, , 115, 1693 de Zotti, G., Ricci, R., Mesa, D., Silva, L., Mazzotta, P., Toffolatti, L., & Gonz[á]{}lez-Nuevo, J. 2005, , 431, 893 Eddington, A. S., Sir 1940, , 100, 354 G[ó]{}rski, K. M., Hivon, E., Banday, A. J., Wandelt, B. D., Hansen, F. K., Reinecke, M., & Bartelmann, M. 2005, , 622, 759 Gregory, P. C., Scott, W. K., Douglas, K., & Condon, J. J. 1996, , 103, 427 Griffith, M. R., Wright, A. E., Burke, B. F., & Ekers, R. D. 1994, , 90, 179 Griffith, M. R., Wright, A. E., Burke, B. F., & Ekers, R. D. 1995, , 97, 347 Healey, S. E., Romani, R. W., Taylor, G. B., Sadler, E. M., Ricci, R., Murphy, T., Ulvestad, J. S., & Winn, J. N. 2007, , 171, 61 Hill, R. S., et al. 2009, , 180, 246 Hinshaw, G., et al.  2007, , 170, 288 Jarosik, N., et al.  2007, , 170, 263 Kuehr, H., Pauliny-Toth, I. I. K., Witzel, A., & Schmidt, J. 1981a, , 86, 854 Kuehr, H., Witzel, A., Pauliny-Toth, I. I. K., & Nauber, U. 1981b, , 45, 367 L[ó]{}pez-Caniego, M., Gonz[á]{}lez-Nuevo, J., Herranz, D., Massardi, M., Sanz, J. L., de Zotti, G., Toffolatti, L., & Arg[ü]{}eso, F. 2007, , 170, 108 Massardi, M., et al.  2008, , 384, 775 Mather, J. C., Fixsen, D. J., Shafer, R. A., Mosier, C., & Wilkinson, D. T. 1999, , 512, 511 Mauch, T., Murphy, T., Buttery, H. J., Curran, J., Hunstead, R. W., Piestrzynski, B., Robertson, J. G., & Sadler, E. M. 2003, , 342, 1117 Page, L., et al. 2003, , 148, 39 Wright, A. E., Griffith, M. R., Burke, B. F., & Ekers, R. D. 1994, , 91, 111 Wright, A. E., Griffith, M. R., Hunt, A. J., Troup, E., Burke, B. F., & Ekers, R. D. 1996, , 103, 145 Wright, E. L., et al.  2009, , 180, 283 [^1]: The Hierarchical Equal Area isoLatitude Pixelization (HEALPix) of the sphere is used to define WMAP map pixels on the sky in Galactic coordinates. Res 9 corresponds to 3,145,728 pixels over the full sky with a pixel resolution of $0.115\degr$. See more of HEALPix at http://healpix.jpl.nasa.gov/ . [^2]: See <http://nedwww.ipac.caltech.edu/>. [^3]: By convention, S is often used to represent the flux density in the discussion of number counts distribution; it is equivalent to $F_{\nu}$ that is used throughout this paper.

ArXiv

--- abstract: 'Results of a new two-particle correlation analysis of central Pb+Au collision data at 158 GeV per nucleon are presented. The emphasis is put on pion-proton correlations and on the dependence of the two-pion correlation radii on the azimuthal emission angle with respect to the reaction plane.' author: - Dariusz Antończyk and Dariusz Miśkowiec for the CERES Collaboration title: 'Pion-pion and pion-proton correlations - new results from CERES[^1]' --- -1.35cm EXPERIMENT AND DATA ANALYSIS {#sec:intro} ============================ CERES is a dilepton experiment at the CERN SPS, known for its observation of enhanced production of low mass  pairs in collisions between heavy nuclei [@ceres-sau-pbau]. The upgrade of CERES in 1997-1998 by a radial Time Projection Chamber (TPC) allowed to improve the momentum resolution and the particle identification capability while retaining the cylindrical symmetry (Fig. \[fig:setup\]). The TPC also opened the possibility of measuring hadrons. The upgraded experiment has been described in detail elsewhere [@ceres]. The measurement of central Pb+Au collisions at the maximum SPS energy of 158 GeV per nucleon in the fall of 2000 was the first run of the fully upgraded CERES and at the same time the last run of this experiment. About 30 million Pb+Au collision events at 158 GeV per nucleon were collected, most of them with centrality within the top 7% of the geometrical cross section  = 6.94 b. Small samples of the 20% and the minimum bias collisions, as well as a short run at 80 AGeV, were recorded in addition. The first two-particle correlation analysis performed on a subset of these data resulted, among others, in an improved procedure to account for the Coulomb interaction [@heinz-pap] and a new postulate of an universal freeze-out criterion [@heinz-let]. The current analysis features a better momentum resolution $$\frac{\Delta p}{p} = 2\% \oplus 1\% \cdot p/{\rm (GeV\!\!/\!c)} ,$$ a better understanding of the two-track resolution, and was performed on the full data set. The $n$(,$y$) distribution of the analyzed pairs is shown in Fig. \[fig:acceptance\]. The two-track resolution cuts applied to the true pairs and to the pairs from event mixing were different for the two possible track pair topologies (Fig. \[fig:sailor-cowboy\]). It should be noted that the required two-track cuts depended somewhat on the quality cuts applied to single tracks: the higher number of hits required for single tracks, the more pairs were lost because of the finite two-track resolution. The two-pion analysis was performed in the longitudinally co-moving frame (LCMS) defined by the vanishing $z$-component of the pair momentum. The momentum difference in this frame, [**q**]{} = [**p**]{}$_2$-[**p**]{}$_1$, was decomposed into the “out”, “side”, and “long” components following the Bertsch-Pratt convention, with $q_{\rm out}$ pointing along the pair transverse momentum and $q_{\rm long}$ along the beam. The non-identical correlations were analyzed in the pair c.m.s., the frame in which the pair momentum is zero. The two components there,  and , were defined such that  was along the pair transverse momentum, i.e.  was equal to $q_{\rm out}$ if the latter is calculated in the pair c.m.s. Two-pion correlations {#sec:two-pion} ===================== The  and  correlation functions, defined as the three-dimensional distributions of pion pairs from the same event $n($[**q**]{}$)$, normalized to the analogous distributions of pairs constructed from different events (event mixing), were fitted by $$C_2\left({\bf q}\right)= N \cdot\left\{\left(1-\lambda\right) + \lambda \cdot F_{c}\left(q_{inv}\right)\left[1+ \exp\left(-\displaystyle\sum_{i,j = 1}^{3} R^{2}_{ij} q_{i}q_{j}\right)\right]\right\}\,\,. \label{eq:fitfunction}$$ The normalization factor $N$ is needed because the number of pairs from event mixing is arbitrary. The correlation strength $\lambda<$1 reflects the tails of the source distribution caused by the pions from long-lived resonances, the finite ${\bf q}$-resolution, and the contamination of the pion sample by other particle species. The $R^{2}_{ij}$ fit parameters, with the indices $i$,$j$ being {out, side, long}, are related to the size of the source emitting pions of given momentum [@mike-overview] and will be called here HBT radii. The $F_{c}\left(q_{inv}\right)$ factor, $q_{inv}=\sqrt{-(p_2^\mu-p_1^\mu)^2}$, accounts for the mutual Coulomb interaction between the pions and was calculated by averaging the nonrelativistic Coulomb wave function squared over a realistic source size. The Coulomb factor was attenuated by $\lambda$ similarly as the rest of the correlation function peak; the importance of this approach was demonstrated in [@heinz-pap]. The fits were performed by the minimum negative loglikelihood method with the Poissonian number of true pairs and were done separately for each pair (,y) bin. The HBT radii obtained from the fit were corrected for the finite momentum resolution. The correction was determined by Monte Carlo and was rather insignificant for  and ; for  it gets as large as $\approx$ 20% for the highest bin of the pair . The obtained HBT radii show a strong  dependence (Fig. \[fig:rpt\]). The  and  radii were fitted with [@rside; @rlong] $$\begin{aligned} R_{side}(p_t) &=& \sqrt{\frac{R_G^{2}}{1+\Mt \, \, \eta_{f}^{2}/T}} \label{eq:rsidept}\\ R_{long}(p_t) &=& \tau_{f}\sqrt{\frac{T}{\Mt} \frac{K_{2}\left(\Mt/T\right)} {K_{1}\left(\Mt/T\right)}} \label{eq:rlongpt}\end{aligned}$$ with the freeze-out temperature $T$ fixed to be 120 MeV. The results of the fit are shown in Fig. \[fig:hbtcent\]. The obtained expansion time $\tau$ of 5.5-6.5 fm, geometrical source size $R_G$ of 6-8 fm, and transverse expansion rapidity $\eta_f$ of 0.7-0.8, in reasonable agreement with the results of the previous analysis of a subset of data [@heinz-pap], indicate a long-living, longitudinally and transversally expanding pion source. Azimuthal angle dependent HBT {#sec:reaction-plane} ============================= A fireball created in a collision with a finite impact parameter is elongated in the direction perpendicular to the reaction plane. In the course of expansion, with the pressure gradient larger in-plane than out-of-plane, the initial asymmetry should get reduced or even reversed. A dependence of the two-pion correlations on the pair emission angle with respect to the reaction plane would be a signature of the source eccentricity at the decoupling time. The azimuthal angle of the reaction plane was estimated by the preferred direction of the particle emission aka elliptic flow. The particles were weighted with their transverse momentum: $$\begin{aligned} Q^X_2 &=& \sum_i p_t \, \cos(2\phi_i)\\ Q^Y_2 &=& \sum_i p_t \, \sin(2\phi_i) \, .\end{aligned}$$ The raw distribution $n(Q^X_2,Q^Y_2)$ is shown in Fig. \[fig:qdist\]. The reaction plane angle was calculated (modulo $\pi$) from the calibrated $Q_2$ components via $$\Phi_{\rm RP} = \frac{1}{2} \arctan \left[\frac{Q^Y_2}{Q^X_2}\right] .$$ The resolution of the so determined reaction plane angle, estimated via the subevevent method, was 31-38. With the event plane known (within the resolution) event-by-event the pion pairs can be sorted into 8 bins covering $(-\pi/2,\pi/2)$ according to their azimuthal angle with respect to the reaction plane $\Phi^*=\Phi_{\rm pair}-\Psi_{\rm RP}$. During event mixing it was required that the two events had a similar reaction plane angle $\Psi_{\rm RP}$. The eight correlation functions were fitted as described in Section \[sec:two-pion\], and the resulting , , , and the cross-terms are plotted versus $\Phi^*$. The squared source radii were then fitted with $$R_i^2 = R_{i,0}^2+2 R_{i,2}^2 \cos(2 \Phi^*) , \label{rfourier}$$ with $i$ denoting {out,side,long}. While the $R_{i,0}$’s obtained coincide with the results of the standard HBT analysis presented in Section \[sec:two-pion\] the second Fourier components $R_{i,2}$’s represent the eccentricity of the observed pion source. The normalized second Fourier component of  is shown in Fig. \[fig:r2\]. As far as the limited centrality range allows to judge the measured   anisotropy is consistent with zero, in contrast to the out-of-plane elongated pion source observed both at the AGS [@ags] and at RHIC [@rhic]. The source eccentricity, thus, seems to be joining the exclusive club of heavy ion observables which behave non-monotonically with the collision energy, the other members being the flow and the strangeness-to-entropy ratio. ![ Azimuthal source eccentricity, represented by the normalized second Fourier component of $(\Phi^*)$, vs. centrality. The CERES result (black full dots) is much closer to zero than the analogous measurement at the AGS (open red squares) and RHIC (open blue triangles). []{data-label="fig:r2"}](fig_R2_vs_centrality.eps){width="9cm"} Pion-proton correlations {#sec:pion-proton} ======================== For the top central 7% of the geometrical cross section the high event statistics allows to perform the proton-pion correlation analysis. The shape of non-identical particle correlation functions $C({\bf q})$ reflects the shape of the relative source distributions $S(r_2^\mu-r_1^\mu)$. Particularly, a difference between the average freeze-out position or time of two particle species reveals itself as an asymmetry of the correlation function at small $q$ [@lednicky]. A two-dimensional $\pi^-$p correlation function $C$(,) and its slice $C$() for $<$50 MeV/c are shown in Fig. \[fig:nonident\]. ![ Two-dimensional $\pi^-$p correlation function $C$(,) (left) and its slice $C$() for $<$50 MeV/c (right). The peak asymmetry in the right hand plot indicates that the proton source is located at a larger radius than the pion source, or that protons are emitted earlier than pions. The asymmetry can be parametrized using Eq. (\[eq:AsymmetryFit\]). []{data-label="fig:nonident"}](nonident.eps){width="9cm"} The peak at low $q$ comes from the attractive Coulomb interaction. The peak asymmetry indicates that the proton source is located at a larger radius than the pion source, or that protons are emitted earlier than pions. The asymmetry can be conveniently parametrized by fitting a Lorentz curve, which happens to match the shape, separately to the left and to the right half of the peak, and taking the ratio of the two widths: $$C_{2}(q_{\parallel}) = \left\{ \begin{array}{ll} N \cdot \left(1+\frac{a}{\left(q_{\parallel}/\sigma_{-}\right)^{2} + 1} \right)\,, & q_{\parallel} < 0 \\ N \cdot \left(1+\frac{a}{\left(q_{\parallel}/\sigma_{+}\right)^{2} + 1} \right)\,, & q_{\parallel} > 0 \end{array} \right. \label{eq:AsymmetryFit}$$ where $N$ is a normalization factor and $a$ is the peak amplitude. The asymmetry is then defined as the ratio between the two widths $\mathcal{A}=\sigma_{-} / \sigma_{+}$. The asymmetry was translated to a spatial displacement between the proton and pion sources using a Monte Carlo pair generator with realistic source sizes and the Coulomb wave function squared as a weighting factor for each pair. The relation between the two quantities is shown in Fig. \[fig:asymmetrycalibration\]. The resulting displacement $\Delta x$ as a function of the transverse pair momentum is shown in the right panel of Fig. \[fig:asymmetryresults\]. In the left panel the raw asymmetry $\mathcal{A}$ is shown. The asymmetry and the displacement derived from it vanish in the limit of small pair  as is expected for symmetry reasons. The displacement is rather similar to the spatial displacement seen in UrQMD v.1.3 [@urqmd] (green line). The red and blue lines represent a simultaneous fit to () and $\Delta x$() using, respectively, Eq.(\[eq:rsidept\]) and the formula derived in [@deltax]: $$\langle \Delta x \rangle = \frac{R_G \, \beta_\perp \, \beta_0}{\beta_0^2+\frac{T}{\Mt}}$$ with the pair transverse mass $$\Mt = \sqrt{ \sqrt{m_{\pi}^{2} + \frac{m_{\pi}}{m_{p}+ m_{\pi}} \cdot \left(\frac{P_{\perp}}{2}\right)^{2}} \cdot \sqrt{m_{p}^{2} + \frac{m_{p}}{m_{\pi}+m_{p}} \cdot \left(\frac{P_{\perp}}{2}\right)^{2}}}\,,$$ and the pair transverse velocity {width="18cm"} $$\beta_\perp = \frac{1}{\sqrt{1+\left(\frac{m_\pi+m_p}{P_\perp}\right)^2}} \, .$$ The relation between the rapidity $\eta_f$ and the velocity $\beta_0$ of transverse flow, which occur in the  and $\Delta x$ fit formulas, respectively, is $$\eta_f = \frac{1}{2} \ln\frac{1+\beta_0}{1-\beta_0} .$$ The fit yields a $\beta_0$ of 0.65-0.70 and a $R_G$ of about 7.5 fm. The fit is dominated by the  data with their small error bars but it reproduces the $\Delta x$ values quite well. This indicates that the finite displacement between the sources of pions and protons has a similar origin as the  dependence of , namely the transverse flow. Other, possibly more interesting, contributions to it cannot be addressed with the present statistics of the data. Blast-wave parametrization ========================== {width="18cm"} While it is notoriously difficult to describe many observables of heavy ion collisions within the same theoretical approach, in recent years simple hydrodynamics-inpired parametrizations turn out to be quite successful in this aspect. The blast-wave model of [@mike-blast], with its 8 parameters adjusted accordingly, nicely reproduces the transverse momentum spectra, the elliptic flow, and the two-particle correlations of CERES, including the emission angle dependence (or, rather, the lack of it). The transverse spectra and the elliptic flow coefficients from the top 7% of  for several hadron species are shown along with the blast-wave model lines in Fig. \[fig:blast0\]. In Fig. \[fig:blast1\] the two-particle correlation results are compared to the calculations performed with the blast-wave model with the same parameter values. The agreement is rather good. Two of the eight parameters of the model were fixed: the freeze-out temperature $T=100$ MeV and the sharpness of the emission volume $a=0.01$ (sharp). The other six parameters, adjusted to the data by the simplex method, include the transverse flow rapidity $\rho$ of 0.88 and 0.85 in-plane and out-of-plane, respectively; the source radius $(R_x,R_y)$ of (11.26, 11.42) fm; the emission time (i.e. the duration of the expansion) $\tau$ of 7.4 fm, and the duration of emission $\Delta \tau = 1.55$ fm/c. The nearly-circular source $R_x\approx R_y$ was enforced by the smallness of the second Fourier component of the HBT radii (bottom part of Fig. \[fig:blast1\]). Summary ======= In summary, the new high-statistics CERES hadron data allow for state of the art analysis of two-particle correlation data. The pion source anisotropy, accessible via the dependence of the two-pion HBT radii on the emission angle with respect to the reaction plane, is unexpectedly small compared to the analogous results obtained at lower and higher energies. The pion-proton correlations indicate that the proton source is located at a larger radius than the pion source, or that protons are emitted earlier than pions. The amount of the displacement, and its functional dependence on the transverse momentum of the two involved particles, corroborate the transverse expansion picture deduced from the behavior of the   HBT radius. Finally, the blast-wave parametrization is able to describe simultaneously the transverse spectra and the elliptic flow coefficients of several hadron species, and the two-particle correlations. This work was supported by the German BMBF, the U.S. DoE, the Israeli Science Foundation, and the MINERVA foundation. The data analysis was performed by D. Antończyk as a part of his Ph.D. thesis work. The author (DM) thanks the workshop organizers for the invitation and for the good and inspiring atmosphere of the meeting. [99]{} G. Agakichiev et al. (CERES), Phys. Rev. Lett. 75 (1995) 1272 and Eur. Phys. J. C41 (2005) 475. D. Miskowiec, Nucl.Phys.A774, 43-50 (2006); A. Marin, J. Phys. G30, S709 (2004); Nucl. Instr. Meth., in preparation. D. Adamova et al., Nucl. Phys. A714, 124 (2003). D. Adamova et al., Phys. Rev. Lett. 90, 022301 (2003). M.A.Lisa, S.Pratt, R.Soltz, and U.Wiedemann, Ann. Rev. Nucl. Part. Sci. 55, 357 (2005). S.Chapman, J.R.Nix, and U.Heinz, Phys. Rev. C52, 2694 (1995); T.Csörgő and B.Lörstad, Phys. Rev. C54, 1390 (1996); R.Scheibl and U. Heinz, Phys. Rev. C59, 1585 (1999). A.N.Makhlin and Yu.M.Sinyukov, Z. Phys. C39, 69 (1988); S.V.Akkelin and Yu.M.Sinyukov, Phys. Lett. B356, 525 (1995); M.Herrmann and G.F.Bertsch, Phys. Rev. C51, 328 (1995); U.A.Wiedemann, P.Scotto, and U.Heinz, Phys. Rev. C53, 918 (1996). M.A. Lisa et al., Phys. Lett. B496, 1 (2000). J.Adams et al., Phys. Rev. C72, 014904 (2005). R.Lednicky, V.L.Lyuboshitz, B.Erazmus, and D.Nouais, Phys. Lett. B373, 30 (1996). S.A.Bass et al., Prog. Part. Nucl. Phys. 41, 225 (1998); M. Bleicher et al., J. Phys. G: Nucl. Part. Phys. 25, 1859 (1999). S.V.Akkelin and Yu.M.Sinyukov, Z. Phys. C72, 501 (1996); R.Lednicky, Phys. Atom. Nucl. 67, 72 (2004). F.Retiere and M.A.Lisa, Phys. Rev. C70, 044907 (2004). [^1]: Based on the Ph.D. thesis work of Dariusz Antończyk, Technical University Darmstadt, 2006

ArXiv

--- abstract: 'LECTURE GIVEN AT TH2002: Given a set of Boolean variables, and some constraints between them, is it possible to find a configuration of the variables which satisfies all constraints? This problem, which is at the heart of combinatorial optimization and computational complexity theory, is used as a guide to show the convergence between these fields and the statistical physics of disordered systems. New results on satisfiability, both on the theoretical and practical side, can be obtained thanks to the use of physics concepts and methods.' author: - Marc Mézard title: 'Optimization and Physics: On the satisfiability of random Boolean formulae ' --- Combinatorial optimization aims at finding, in a finite set of possible configurations, the one which minimizes a certain cost function. The famous example of the traveling salesman problem (TSP) can serve as an illustration: A salesman must make a tour through $N$ cities, visiting each city only once, and coming back at its starting point. The cost function is a symmetric matrix $c_{ij}$, where $c_{ij}$ is the cost for the travel between cities $i$ and $j$. A permutation $\pi$ of the $N$ cities gives a tour $\pi(1)\to \pi(2) \to \pi(3)\to...$. Taking into account the equivalence between various starting points and the direction of the tour, one sees that the number of distinct tour is $(N-1)!/2$. For each tour $\pi$, the total cost is $C=c_{\pi(N)\pi(1)}+\sum_{r=1}^{N-1} c_{\pi(r)\pi(r+1)}$, which can be computed in $N$ operations. The problem is to find the tour $\pi$ with lowest cost. As can be seen on this example, the basic ingredients of the optimization problems which will interest us are the following: - An integer $N$ giving the size of the problem (in the TSP, it is the number of cities). - A set of configurations, which typically scales like $\exp(N^\alpha)$. - A cost function, which one can compute in polynomial time $O(N^\gamma)$. Let me mention a few examples, beside the TSP [@papad]. In the assignment problem, one is given a set of $N$ persons $i=1,...,N$, a set of $N$ tasks $a=1,...,N$, a $N \times N$ cost matrix $c$ where $c_{ia}$ is the cost for having task $ a$ performed by person $i$. An assignment is a permutation $\pi \in S_N$ assigning task $a=\pi(i)$ to person $i$, and the problem is to find the lowest cost assignment, i.e. the permutation which minimizes $C=\sum_{i=1}^N c_{i\pi(i)}$. In the spin glass problem[@MPV], one is given a set of $N$ spins, which could be for instance Ising variables $\sigma_i\in\{ \pm 1 \}$, the total energy of the configuration is a sum of exchange interaction energies between all pairs of spins, $E(\{\sigma\})=-\sum_{1\le i<j \le N}J_{ij} \sigma_i \sigma_j$, and one seeks the ground state of the problem, i.e. the configuration (among the $2^N$ possible ones) which minimizes the energy. In physical terms, optimization problems consist in finding ground states. This task can be non trivial if a system is frustrated, which means that one cannot get the ground state simply by minimizing the energy locally. This is typically what happens in a spin glass. In some sense, statistical physics addresses a more general question. Assuming that the system is at thermal equilibrium at temperature $T$, every configuration ${{\cal C}}$ is assigned a Boltzmann probability, $P({{\cal C}})=\frac{1}{Z} e^{-{ \beta} { E({{\cal C}})}}$. Beside finding the ground state, one can ask also interesting questions about which are the typical configurations at a given temperature, like counting them (which leads to the introduction of entropy), or trying to know if they are located in one single region of phase space, or if they build up well separated clusters, as often happens in situation where ergodicity is broken. The generalization introduced by using a finite temperature, beside leading to interesting questions, can also be useful for optimization, both from the algorithmic point of view (for instance this is the essence of the simulated annealing algorithm[@KGV], which is a general purpose heuristic algorithm that can be used in many optimization problems), but also from an analytic point of view [@MPV]. Conversely, smart optimization algorithms turn out to be very useful in the study of frustrated physical systems like spin glasses or random field models, and the cross-fertilization between these two fields (and also with the related domain of error correcting codes for information transmission [@codes]) has been very fruitful in the recent years [@TCS_issue]. Before proceeding with one such example, let us briefly mention a few important results in optimization which will provide the necessary background and motivation. One of the great achievements of computer science is the theory of computational complexity. It is impossible to present it in any details here and I will just sketch a few main ideas, the interested reader can study it for instance in [@papad2]. Within the general framework explained above, we can define three types of optimization problems: the optimization problem in which one wants to find the optimal configuration, the evaluation problem in which one wants to compute the optimal cost (i.e. the ground state energy), the decision problem in which one wants to know, given a threshold cost $C_0$, if there exists a configuration of cost less than $C_0$. One classification of decision problems is based on the scaling with $N$ of the computer time required to solve them in the [*worst case*]{}. There are two main complexity classes: - [ [**Class P**]{}  ]{} = polynomial problems: they can be solved in a time $ t <N^\alpha$. The assignment is an example of a polynomial problem, as is the spin glass problem in 2 dimensions. - [ [**Class NP**]{}]{} = non-deterministic polynomial: Given a ’yes’ solution to a NP problem, it can be checked in polynomial time. Roughly speaking this means that the energy is computable in polynomial time, so this class contains a wide variety of problems, including most of the ones of interest in physics. All problems mentioned above are in NP. One nice aspect of focusing on polynomiality is that it allows to forget about the details of the definition of $N$, the implementation, language, machine, etc...: any ’reasonable’ such change (for instance one could have used the number of possible links in the assignment) will change the exponent of $N$ appearing in the computer time of a problem in P, but not transform it into an exponential behavior. A problem $A$ is said to be at least as hard as a problem $B$ if there exists a polynomial algorithm which transforms any instance of $B$ into an instance of $A$. This allows to define the very important class: - [ [**NP-complete**]{}]{} A problem is NP complete if it is at least as hard as any problem in NP. So the NP-complete are the hardest problems in NP. If one such problem can be solved in polynomial time, then all the problems in NP are solved in polynomial time. Clearly P is contained in NP, but it is not yet known whether P = NP , and this is considered as a major challenge. A great result was obtained in this field by Cook in 1971 [@Cook]: The satisfiability problem, which we shall describe below, is NP-complete. Since then hundreds of problems have been shown to belong to this class, among which the decision TSP or the spin glass in dimension larger or equal to 3. The fact that 3-d spin glass is NP-complete while 2-d spin glass is P might induce one to infer that NP-completeness is equivalent with the existence of a glass transition. This reasoning is too naive and wrong; the reason is that the complexity classification relies on a worth-case analysis, while physicists study the behavior of a typical sample. The development of a typical case complexity theory has become a major goal [@average], also motivated by the experimental observations that many instances of NP-complete problems are easy to solve[@TCS_issue]. One way of addressing this issue of a typical case complexity is to define a probability measure on the instances (= the ’samples’) of the optimization problem which one is considering. Typical examples are: - TSP with independent random points uniformly distributed in the unit square - assignment with independent affinities uniformly distributed on $[0,1]$ - CuMn spin glass at one percent Mn Once this measure has been defined, one is interested in the properties of the [ generic]{} sample in the $N \to \infty$ limit. In most cases, global properties (e.g. extensive thermodynamic quantities, among which the ground state energy), turn out to be self averaging. This means for instance that the distribution of the ground state energy density becomes more and more peaked around an asymptotic value in the large $N$ limit: almost all samples have the same ground state energy density. In this situation, a statistical physics analysis is appropriate. Early examples of the use of statistical physics in such a context are the derivation of bounds for the optimal length of a TSP[@MVan], the exact prediction of the ground state energy in the random assignment problem defined above, where the result $E_0=\frac{\pi^2}{6}$, derived in 1985 through a replica analysis[@MP_match], was recently confirmed rigorously by Aldous [@Aldous], or the link between spin glasses and graph partitioning [@FuAnd]. As statistical physics can be quite powerful at understanding the generic structure of an optimization problem, one may also hope that it can help finding better optimization algorithms. A successful example which was developed recently is the satisfiability problem, to which we now turn. As we have seen, satisfiability is a core problem in optimization and complexity theory. It is defined as follows [@Hayes]: A configuration is a set of $N$ Boolean variables $x_i \in \{0,1\} \quad i=1,\ldots,N$. One is given $M$ constraints which are clauses, meaning that they are in the form of an OR function of some variables or their negations. For instance: $x_1 \vee x_{27} \vee \bar x_3$, $ \bar x_{11} \vee x_2$, are clauses (notation: $\bar x_3$ is the negation of $x_3$). So the clause $x_1 \vee x_{27} \vee \bar x_3$ is satisfied if either $x_1=1$, or $x_{27}=1$, or $x_3=0$ (these events do not exclude each other). The satisfiability problem is a decision problem. It asks whether there exists a choice of the Boolean variables such that all constraints are satisfied (we will call it a SAT configuration). This is a very generic problem, because one can see it as finding a configuration of the $N$ Boolean variables which satisfies the logical proposition built from the AND of all the clauses (in our example[ $(x_1 \vee x_{27} \vee \bar x_3) \wedge (\bar x_{11} \vee x_2) \wedge \ldots$]{}), and any logical proposition can be written in such a ’conjunctive normal form’ . Satisfiability is known to be NP complete if it contains clauses of length $\ge 3$, but common sense and experience show that the problem can often be easy; for instance if the number of constraints per variable $\alpha=\frac{M}{N}$ is small, the problem is often SAT, if it is large, the problem is often UNSAT. This behavior can be characterized quantitatively by looking at the typical complexity of the random 3-SAT problem, defined as follows. Each clause involves exactly three variables, chosen randomly in $\{ x_1,..,x_N\}$; a variable appearing in the clause is negated randomly with probability $1/2$. This defines the probability measure on instances for the random 3-SAT problem. The control parameter is the ratio Constraints/Variables $ \alpha=\frac{M}{N}$ . The properties of random 3-SAT have been first investigated numerically [@KirkSel; @Crawford], and exhibit a very interesting threshold phenomenon at $\alpha_c \sim 4.26$: a randomly chosen sample is generically SAT for $\alpha<\alpha_c$ (meaning that it is SAT with probability $1$ when $N \to \infty$), generically UNSAT for $\alpha>\alpha_c$. The time used by the best available algorithms (which have an exponential complexity) to study the problem also displays an interesting behavior: For $\alpha$ well below $\alpha_c$, it is easy to find a SAT configuration; for $\alpha$ well above $\alpha_c$, it is relatively easy to find a contradiction in the constraints, showing that the problem is UNSAT. The really difficult region is the intermediate one where $\alpha \sim \alpha_c$, where the computer time requested to solve the problem is much larger and increases very fast with system size. A lot of important work has been done on this problem, to establish the existence of a threshold phenomenon, give upper and lower bounds on $\alpha_c$, show the existence of finite size effects around $\alpha \sim \alpha_c$ with scaling exponents. We refer the reader to the literature [@KirkSel; @Crawford; @MZKST; @bounds_1; @bounds_2; @bounds_3; @bounds_4], and just here extract a few crucial aspects for our discussion. The threshold phenomenon is a phase transition, and the neighborhood of the transition is the region where the algorithmic problem becomes really hard. This relationship between phase transition and complexity makes a statistical physics analysis of this problem particularly interesting. Monasson and Zecchina have been the first ones to recognize this importance and to use statistical physics methods for an analytic study of the random 3-SAT problem [@MoZe_prl; @MoZe_pre]. Basically this problem falls into the broad category of spin glass physics. This is immediately seen through the following formulation. To make things look more familiar, physicists like to introduce for each Boolean variable $x_i\in \{0,1\}$ an Ising spin $\sigma_i \in\{-1,1\}$ through $x_i= \frac{1+\sigma_i}{2}$. A satisfiability problem like $${ (x_1 \vee x_{27} \vee \bar x_3) }{\wedge }{ (\bar x_{11} \vee x_3 \vee x_2) }{\wedge}{ \ldots }$$ can be written in terms of an energy function, where the energy is equal to one for each violated clause. Explicitly, in the above example, one would have $$E= \frac{1+\sigma_1}{2}\; \frac{1+\sigma_{27}}{2}\; \frac{1-\sigma_3}{2} \;{ +} \;\frac{1-\sigma_{11}}{2}\; \frac{1+\sigma_3}{2}\; \frac{1+\sigma_2}{2} \;{ +} \; \ldots$$ This is clearly a problem of interacting spins, with 1,2, and 3 spin interactions, disorder (in the choice of which variable interacts with which), and frustration (because some constraints are antagonist). More technically, the problem has a special type of three-spin interactions on a random hyper-graph. Using the replica method, Monasson and Zecchina first showed the existence of a phase transition within the replica symmetric approximation, at $\alpha_c=5.18$, then showed that replica symmetry must be broken in this problem. Some variational approximation to describe the replica symmetry breaking effects were developed in particular in [@BMW; @FLRZ]. Recently, in a collaboration with G.Parisi and R. Zecchina [@MEPAZE; @MZ_pre], we have developed a new approach to the statistical physics of the random 3-SAT problem using the cavity method. While the cavity method had been originally invented to deal with the SK model where the interactions are of infinite range [@MPV], it was later adapted to problems with ’finite connectivity’, in which a given variable interacts with a finite set of other variables. While this is easily done for systems which are replica symmetric (like the assignment, or the random TSP with independent links), it turned out to be considerably more difficult to develop the corresponding formalism and turn it into a practical method, in the case where replica symmetry is broken. This has been done in the last two years in joint works with G.Parisi [@MP_Bethe], and has opened the road to the study of finite connectivity optimization problems with replica symmetry breaking like random K-sat. Curiously, while the cavity method is in principle equivalent to the replica method (although it proceeds through direct probabilistic analysis instead of using an analytic continuation in the number of replicas), it turns out that it is easier to solve this problem with the cavity method. ![A pictorial description of the phase diagram for random 3-SAT obtained in the cavity method: clusters of SAT (in green/light gray) or UNSAT (in red/dark grey) configurations. One finds three qualitatively different phases, the EASY-SAT phase for $\alpha<\alpha_d$, the HARD-SAT phase for $\alpha_d<\alpha<\alpha_c$, the UNSAT phase obtained for $\alpha_c<\alpha$ (see text) []{data-label="phdi"}](phase_diag.eps){width="9.cm"} The analytic study of [@MEPAZE; @MZ_pre] for the random 3-SAT problem at zero temperature shows the following phase diagram (see fig.\[phdi\]): - For $\alpha<\alpha_d=3.921$, the problem is generically SAT; the solution can be found relatively easily, because the space of SAT configurations builds up a single big connected cluster. A $T=0$ Metropolis algorithm, in which one proposes to flip a randomly chosen variable, and accepts the change iff the number of violated constraints in the new configuration is less or equal to the old one, is able to find a SAT configuration. We call this the EASY-SAT phase - For $\alpha_d<\alpha<\alpha_c=4.267$, the problem is still generically SAT, but now it becomes very difficult to find a solution (we call this the HARD-SAT phase). The reason is that the configurations of variables which satisfy all constraints build up some clusters. Each such cluster, which we call a ’state’, is well separated from the other clusters (passing from one to the other requires flipping an extensive number of variables). But there also exist many “metastable states”: starting from a random configuration, a local descent algorithm will get trapped in some cluster of configurations with a given finite energy (they all have the same number of violated clauses, and it is impossible to get out of this cluster towards lower energy configurations through any descending sequence of one spin flip moves). The number of SAT clusters ${\cal N}$ is exponentially large in $N$, it behaves as ${\cal N}\sim \exp(N \Sigma)$; but the number of metastable clusters is also exponentially large with a larger growth rate, behaving like $ \exp(N \Sigma_{ms})$ with $\Sigma_{ms}> \Sigma$. The most numerous metastable clusters, which have an energy density $e_{th}$, will trap all local descent algorithms (zero temperature Metropolis of course, but also simulated annealing, unless it is run for an exponential time). - For $\alpha>\alpha_c$, the problem is typically UNSAT. The ground state energy density $e_0$ is positive. Finding a configuration with lowest energy is also very difficult because of the proliferation of metastable states. A more quantitative description of the thermodynamic quantities in the various phases is shown in fig.\[phdi2\]. The most striking result is the existence of an intermediate SAT phase where clustering occurs. A hint of such a behavior had been first found in a sophisticated variational one step replica symmetry breaking approximation developed in [@BMW]; however this approximation predicted a second order phase transition (clusters separating continuously), while we now think that the transition is discontinuous: an exponentially large number of macroscopically separated clusters appears discontinuously at $\alpha=\alpha_d$. Another point which should be noticed is the fact that the complexity, and the energy $e_{th}$ in the HARD-SAT phase are rather small: around $e_{th} \sim 3 \ 10^{-3}$ violated clause per variable for $\alpha=4.2$. This shows that until one reaches problems with at least a few thousands variables, one cannot feel the true complexity of the problem. This can explain why the existence of the intermediate phase went unnoticed in previous simulations. ![ Thermodynamic quantities for the random 3-SAT problem: $e_0$ is the ground state energy density (minimal number of violated clauses per variable); $e_{th}$ is the energy density of the most numerous metastable states,which trap the local descent algorithms; $\Sigma$ is the complexity of SAT states with $E=0$ []{data-label="phdi2"}](fig0.eps){width="10.cm"} The second type of results found in [@MEPAZE; @MZ_pre; @BMZ] is a new class of algorithms dealing with the many states structure. Basically this algorithm amounts to using the cavity equations on one given sample. Originally the cavity method was developed to handle a statistical distribution of samples. For instance in the random 3-SAT case its basic strategy is to add one extra variable and connect it randomly to a number of new clauses which has the correct statistical distribution. In the large $N$ limit, the statistics of the local field on the new variable should be identical to the statistics of the local fields on any other variable in the absence of the new one. It turns out that this strategy can be adapted to study a single sample: one considers a given clause $a$, which involves three variables $\sigma_1,\sigma_2,\sigma_3$. The cavity field on $\sigma_1$ is the field felt by $\sigma_1$ in the absence of $a$. In the case where there exist many states, there is one such field for each possible state of the system, and the order parameter is the survey of these fields, in all the states of fixed energy density $e$: $$P_0^e(h)= C^t \sum_\alpha \delta\left(h_0^\alpha-h\right) \delta \left(\frac{E^\alpha}{N}-e\right)$$ One can then write a recursion recursion between these surveys. Looking for instance at the structure of fig.\[iterfig\], one gets the following iteration equation: $$\begin{aligned} \nonumber P_0^e(h)&=& C^t \int P_{\sigma_1}^e(g_1)dg_1 \; P_{\tau_1}^e(h_1)dh_1 \;P_{\sigma_2}^e(g_2)dg_2 \; P_{\tau_2}^e(h_2)dh_2 \\ && \delta(h-f(g_1,h_1,g_2,h_2)) \exp\left(-\frac{d \Sigma}{d e} w(g_1,h_1,g_2,h_2)\right) \label{speq}\end{aligned}$$ The function $f$ just computes the value of the new cavity field on $\sigma_0$ in terms of the four cavity fields $g_1,h_1,g_2,h_2$. It is easily computed by considering the statistical mechanics problem of the five-spin system $\{ \sigma_0,\sigma_1,\tau_1 ,\sigma_2,\tau_2\}$ and summing over the 16 possible states of the spins $\sigma_1,\tau_1,\sigma_2,\tau_2$. The function $w$ computes the free energy shift induced by the addition of $\sigma_0$ to the system with the four spins $\sigma_1,\tau_1,\sigma_2,\tau_2$. The exponential reweighting term in (\[speq\]) is the crucial piece of survey propagation: it appears because one considers the survey at a fixed energy density $e$. As the number of states at energy $E=Ne+\delta E$ increases in $\exp(y \delta E)$, where $y=\frac{d \Sigma}{d e}$, this favors the states with a large negative energy shift. ![The basic iterative structure of survey propagation. This subgraph is a part of a 3-SAT problem, in which the variable $\sigma_0$ belongs to three 3-clauses, involving respectively the variables $\{\sigma_1,\tau_1\} ,\{\sigma_2,\tau_2\},\{\sigma_3,\tau_3\}$. When the clause 3 is switched off, the local cavity field survey $P_0^e(h)$ on spin $\sigma_0$ can be computed in terms of the cavity field surveys on each of the spins $\sigma_1,\tau_1 ,\sigma_2,\tau_2$.[]{data-label="iterfig"}](reseau3.eps){width="7.cm"} The usual cavity method for the random 3-SAT problem determines the probability, when a new variable is added at random, that its survey $P_0^e(h)$ is a given function $P(h)$: the order parameter is thus a functional, the probability of a probability. Because the fields are distributed on integers, this object is not as terrible as it may look and it is possible to solve the equation and compute the ’complexity curve’ $\Sigma(E)$, giving the phase diagram described above. The algorithm for one given sample basically iterates the survey propagation equation on one given graph. It is a message passing algorithm which can be seen as a generalization of the belief propagation algorithm familiar to computer scientists[@pearl]. The belief propagation is a propagation of local magnetic fields, which is equivalent to using a Bethe approximation [@yed]. Unfortunately, it does not converge in the hard-SAT region because various subparts of the graph tend to settle in distinct states, and there is no way to globally choose a state. In this region, the survey propagation, which propagates the information on the whole set of states (in the form of a histogram), does converge. Based on the surveys, one can detect some strongly biased spins, which are fixed to one in almost all SAT configurations. The “Survey Inspired Decimation” (SID) algorithm fixes the spin which is most biased, then it re-runs the survey propagation on the smaller sample so obtained, and then iterates... An example of the evolution of the complexity as a function of the decimation is shown in \[fig\_decim\]. This algorithm has been tested in the hard SAT phase. It easily solves the ’large’ benchmarks of random 3sat at $\alpha=4.2$ with $N=1000,2000$ available at [@SATLIB]. It turns out to be able to solve typical random 3-SAT problems with up to $N=10^7$ at $\alpha=4.2$ in a few hours on a PC, which makes it much better than available algorithms. The main point is that the set of surveys contains a lot of detailed information on a given sample and can probably be used to find many new algorithms, of which SID is one example. ![The total complexity (= the $\ln$ of the number of states) as a function of the total energy (= the number of violated clauses) for one given sample with $N=10^4$ variables and $M=4.2 \; 10^4$ constraints. The top curve is the original complexity. The next curves are the various complexities obtained for the decimated samples, plotted here every $200$ decimations. One sees a global decrease of the number of metastable states, and also a global decrease of the threshold energy. In the end the problem has no more metastable states and can be solved by simple algorithms. []{data-label="fig_decim"}](decimation_4p20_single_N10000.eps){width="7.cm"} To summarize, the recent statistical physics approaches to the random 3-SAT problem give the following results: - An analytic result for the phase diagram of the generic samples - An explanation for the slowdown of algorithms near $\alpha_c=4.267$: this is due to the existence of a [ HARD-SAT phase]{} at $\alpha \in [3.921,4.267]$, with exponentially many metastable states. - An algorithm for single sample analysis[@web]: Survey propagation converges and yields very non trivial information on the sample. It can be used for instance to decimate the problem and get an efficient SAT-solver in the hard-SAT region. This whole set of results calls for a lot of developments in many directions. On the analytical side, the cavity method results quoted above are believed to be correct, but they are not proven rigorously. It would be very interesting to turn these computations into a rigorous proof. A very interesting step in this direction was taken recently by Franz and Leone who showed that the result for the critical threshold $\alpha_c$ obtained by the cavity method on random K-SAT with $K$ even actually give a rigorous upper bound to the critical $\alpha_c$ [@FraLeo]. The whole construction of the cavity method with the clustering structure, the many states and the resulting reweighting, has actually been checked versus rigorous computations on a variant of the SAT problem, the random XORSAT problem, where rigorous computations [@xorsat] have confirmed the validity of the approach. On the numerical side, one needs to develop convergence proofs for survey propagation, and to derive the generalization of the algorithm for the case in which there exists local structures in the interaction graphs. This will amount to generalizing from a Bethe like approximation (with many states) to a cluster variational method with larger clusters (and with many states). The techniques can also be extended to other optimization problems like coloring [@colouring]. Beside the problems mentioned here, there exist many other fascinating problems of joint interests to physicists and computer scientists, like e.g the dynamical behavior of algorithms in optimization or error correcting codes, which I cannot survey here. Let me just quote a few recent references to help the readers through the corresponding bibliography [@cocmon; @global_algorithm; @FLMR; @MonZec]. [Acknowledgements]{} It is a pleasure to thank all my collaborators over the years on various topics at the interface between statistical physics and optimization. I am particularly thankful to G. Parisi and R. Zecchina for the wonderful collaborations over the last three years which lead to the works on the cavity method and the satisfiability problem. [1]{} C.H. Papadimitriou and K. Steiglitz, [*Combinatorial Optimization*]{}, Dover (Mineola, N.Y., 1998). Mézard, M., Parisi, G., $\&$ Virasoro, M.A. Spin Glass Theory and Beyond, World Scientific, Singapore (1987). Kirkpatrick, S., [Gelatt Jr.]{}, C. D. $\&$ Vecchi, M. P. Optimization by simulated annealing, [*Science*]{}[****]{} 220 671–680(1983). [Č]{}erny, V., Thermodynamical approach to the traveling salesman problem: An efficient simulation algorithm, [*J. Optimization Theory Appl.*]{}[**45**]{} 41 (1985). See e.g. A. Montanari, [*The glassy phase of Gallager codes*]{} cond-mat/0104079, and references therein. Dubois O. , Monasson R., Selman B. $\&$ Zecchina R. (Eds.), Phase Transitions in Combinatorial Problems, Theoret. Comp. Sci. [**265**]{} (2001). C.H. Papadimitriou, [*Computational complexity*]{} (Addison-Wesley, 1994) Cook, S. The complexity of theorem–proving procedures, in: Proc. 3rd Ann. ACM Symp. on Theory of Computing, Assoc. Comput. Mach., New York, 1971, p. 151. Levin L.A., Average case complete problems, SIAM J. Comput., [**14**]{} (1): 285-286 (1986); Ben-David S., Chor B., Goldreich O. $\&$ Luby M., On the theory of average case complexity, JCSS [**44**]{}, 193-219; Gurevich Y., Average Case completeness, JCSS [**42**]{}, 246-398 (1991). , M. Mézard and J. Vannimenus, [*On the statistical mechanics of optimization problems of the traveling salesman type*]{}, J. Physique Lett. [**45**]{} (1984) L1145. M. Mézard and G. Parisi, Replicas and Optimization, J.Phys.Lett. [**46**]{} (1985) L771. D. Aldous, The zeta(2) Limit in the Random Assignment Problem, Random Structures and Algorithms [**18**]{} (2001) 381-418. Fu Y. and Anderson P. W., Application of Statistical Mechanics to NP-Complete Problems in Combinatorial Optimization, J. Phys. A 19, 1605–1620 (1986). A nice pedagogical introduction is B. Hayes [*I can’t get no satisfacion*]{}, American Scientist [**85**]{} (1997) 108. S. Kirkpatrick, B. Selman, Critical Behaviour in the satisfiability of random Boolean expressions, Science 264, 1297 (1994) Crawford J.A. $\&$ Auton L.D., Experimental results on the cross-over point in random 3-SAT, Artif. Intell. [**81**]{}, 31-57 (1996). R. Monasson, R. Zecchina, S. Kirkpatrick, B. Selman and L. Troyansky, Nature (London) [**400**]{}, 133 (1999). A. Kaporis, L. Kirousis, E. Lalas, The probabilistic analysis of a greedy satisfiability algorithm, in [*Proceedings of the 4th European Symposium on Algorithms*]{} (ESA 2002), to appear in series: Lecture Notes in Computer Science, Springer; D. Achlioptas, G. Sorkin, [*41st Annu. Symp. of Foundations of Computer Science, IEEE Computer Soc. Press*]{}, 590 (Los Alamitos, CA, 2000). J. Franco, Results related to threshold phenomena research in satisfiability: lower bounds, Theoretical Computer Science [**265**]{}, 147 (2001) O. Dubois, Y. Boufkhad, J. Mandler, Typical random 3-SAT formulae and the satisfiability threshold, in [*Proc. 11th ACM-SIAM Symp. on Discrete Algorithms*]{}, 124 (San Francisco, CA, 2000). Monasson, R. $\&$ Zecchina, R. Entropy of the [K]{}-satisfiability problem, [*Phys. Rev. Lett.*]{} [**76**]{} 3881–3885(1996). Monasson, R. $\&$ Zecchina, R., Statistical mechanics of the random [K-S]{}at problem, [*Phys. Rev.* ]{} [**E 56**]{} 1357–1361 (1997). Biroli, G., Monasson, R. $\&$ Weigt, M. A Variational description of the ground state structure in random satisfiability problems, [*Euro. Phys. J.* ]{} [**B 14**]{} 551 (2000). S. Franz, M. Leone, F. Ricci-Tersenghi, R. Zecchina, Exact Solutions for Diluted Spin Glasses and Optimization Problems, [*Phys. Rev. Lett.*]{} [**87**]{}, 127209 (2001). M. M[é]{}zard, G. Parisi and R. Zecchina, Science 297, 812 (2002) (Sciencexpress published on-line 27-June-2002; 10.1126/science.1073287) , M. M[é]{}zard and R. Zecchina, Phys.Rev. E [**66**]{} (2002) 056126. , A. Braunstein, M. Mézard, R. Zecchina, http://fr.arXiv.org/abs/cs.CC/0212002. M[é]{}zard, M. $\&$ Parisi, G. The Bethe lattice spin glass revisited. [*Eur.Phys. J. B*]{} [**20**]{}, 217–233 (2001). M[é]{}zard, M. $\&$ Parisi, G. The cavity method at zero temperature, J. Stat. Phys. in press. J. Pearl, [*Probabilistic Reasoning in Intelligent Systems*]{}, 2nd ed. (San Francisco, MorganKaufmann,1988). J.S. Yedidia, W.T. Freeman and Y. Weiss, Generalized Belief Propagation, in [*Advances in Neural Information Processing Systems 13*]{} eds. T.K. Leen, T.G. Dietterich, and V. Tresp, MIT Press 2001, pp. 689-695. Satisfiability Library: www.satlib.org/ S. Franz and M. Leone, Replica bounds for optimization problems and diluted spin systems, preprint cond-mat/0208280, available at http://fr.arXiv.org M. Mézard, F. Ricci-Tersenghi, R. Zecchina, [*Alternative solutions to diluted p-spin models and XOR-SAT problems*]{}, cond-mat/0207140, (2002). S. Cocco, O. Dubois, J. Mandler, R. Monasson, [*Rigorous decimation-based construction of ground pure states for spin glass models on random lattices*]{}, cond-mat/0206239 (2002). , R. Mulet, A. Pagnani, M. Weigt, R. Zecchina, cond-mat/0208460. , S. Cocco and R. Monasson, cond-mat/0206242 , J. Phys. A 35, 7559 (2002), A. Braunstein, M. Leone, F. Ricci-Tersenghi, R. Zecchina. , S. Franz, M. Leone, A. Montanari, F. Ricci-Tersenghi, cond-mat/0205051. , A. Montanari, R. Zecchina, cond-mat/0112142.

ArXiv

--- author: - - bibliography: - 'bibliovrd.bib' title: Compensating Supervision Incompleteness with Prior Knowledge in Semantic Image Interpretation ---

ArXiv

--- abstract: 'The real-time flux dynamics of up to three superconducting quantum interference devices (SQUIDs) are studied by numerically solving the time-dependent Schrödinger equation. The numerical results are used to scrutinize the mapping of the flux degrees of freedom onto two-level systems (the qubits) as well as the performance of the intermediate SQUID as a tunable coupling element. It is shown that the qubit representation yields a good description of the flux dynamics during quantum annealing and the presence of the tunable coupling element does not have negative effects on the overall performance. Additionally, data obtained from a simulation of the dynamics of two-level systems during quantum annealing are compared to experimental data produced by the D-Wave 2000Q quantum annealer. The effects of finite temperature are incorporated in the simulation by coupling the qubit system to a bath of two-level systems. It is shown that an environment modeled as non-interacting two-level systems coupled to the qubits can produce data which matches the experimental data much better than the simulation data of the qubits without coupling to an environment and better than data obtained from a simulation of an environment modeled as interacting two-level systems coupling to the qubits.' author: - Madita Willsch - Dennis Willsch - Fengping Jin - Hans De Raedt - Kristel Michielsen bibliography: - '../bibliography.bib' title: 'Real-time simulation of flux qubits used for quantum annealing' --- Introduction {#sec:introduction} ============ Theoretically, an ideal quantum computer is described in terms of two-level systems, the qubits [@nielsen_chuang]. However, almost all currently popular technologies such as ion traps [@cirac95; @monroe95; @haeffner03; @Hanneke2009; @Schindler2013; @ballance16], quantum dots [@loss98; @levy02], or superconducting circuits [@wendin2017; @krantz19] employ physical devices which are only approximately described by two-level systems [@wu02]. Among these, trapped ions may be described in the two-level approximation under conditions discussed in Ref. [@leibfried03]. Single-electron quantum dots can be described as two-level systems if the orbital wave function is neglected. However, two-electron [@levy02; @petta05] and three-electron systems [@divincenzo00; @gaudreau11] confined in quantum dots are again only approximately described by two-level systems, and leakage out of the computational space may need to be taken into account [@divincenzo00; @medford13; @cerfontaine16]. For superconducting circuits, it depends on the particular circuit design how well the two relevant energy levels, which define the qubit states, are separated from the higher energy levels. For instance, the phase qubit [@martinis02; @steffen06] and the transmon [@koch2007], an extension of the charge qubit [@shnirman97; @bouchiat1998; @Nakamura1999], have rather small anharmonicities. Thus, leakage to higher energy levels is a major issue when performing gate operations [@motzoi09; @lucero10; @gambetta11; @Wallman2016; @willsch17; @wood18] and is typically alleviated by the use of pulse-shaping techniques [@motzoi09; @chow10; @lucero10; @gambetta11; @chen16; @mckay17]. Flux qubits, on the other hand, have a strong anharmonicity and are less prone to excitations to higher energy levels [@Poletto2009; @yoshihara14; @billangeon15; @chen16; @krantz19] as long as the qubit is not driven too strongly [@ferron10]. In this study we focus on the superconducting quantum interference device (SQUID)-based flux qubit [@chiarello00; @Makhlin2001; @Poletto2009], which is used in the D-Wave quantum annealer [@harris09_newjournal; @harris10]. Due to the large superconducting loop needed for the SQUID-based qubit, it is sensitive to flux noise which limits the coherence time [@wendin2017]. Other flux qubits, commonly used for gate-based quantum computing, are the three-junction qubit [@Mooij99; @orlando99; @vanderWal00; @grajcar04], the C-shunt flux qubit [@you07; @steffen10; @Yan2016], and the fluxonium qubit [@Manucharyan09; @pop2014; @nguyen18]. For the three-junction qubit and the fluxonium qubit, the large inductance is realized by using two or more Josephson junctions. In this way, the loop size and thus the sensitivity to flux noise can be reduced [@wendin07]. The C-shunt flux qubit is a capacitively shunted variant of the flux qubit with improved coherence when operated away from the degeneracy point [@you07]. There is an ongoing discussion about the role of decoherence during quantum annealing [@childs01; @sarandy05; @ashhab06; @amin09_decoherence; @boixo13; @dickson13]. In this paper we address three questions. First, we study the flux dynamics of the SQUIDs used in the D-Wave quantum annealer, addressing the issue of how well these dynamics are captured by a qubit model. Second, we investigate whether the presence of the SQUID used as a tunable coupler in the D-Wave device affects the performance of the quantum annealing process. To answer the first two questions, we study the dynamics of the flux degrees of freedom of two SQUIDs functioning as qubits and their tunable coupler, a third SQUID, by solving the time-dependent Schrödinger equation (TDSE) for the model Hamiltonian in terms of flux degrees of freedom. So far, studies including higher energy levels have been limited to four-level qudits [@johnson11; @amin13]. The approach that we adopt in this paper is to start from an idealized model of the SQUIDs which does not take into account fabrication variations of circuit elements or stray fluxes induced by the control lines. That is, in the idealized model, the two SQUIDs functioning as qubits are equal and the complete system can be regarded as a perfect device. The third question we consider is to what extent the data produced by a D-Wave quantum annealer can be described by quantum annealing of the qubit model including environment effects. In order to do so, we study the dynamics of a two-qubit system interacting with an environment of two-level systems, representing, e.g., a heat bath [@ZHAO16; @RAED17b] or a collection of defects described by non-interacting two-level systems [@shnirman05; @mueller09; @cole10], by solving the corresponding TDSE. The structure of the paper is as follows. In Sec. \[sec:quantumannealing\] we give a short introduction to quantum annealing and its relation to optimization problems. Section \[sec:squids\] contains a description of the SQUID-based model which we simulate. The mapping of the model onto the qubit model is given in Sec. \[sec:mapping\]. The flux dynamics of the SQUID model are simulated by solving the TDSE using the method described in Sec. \[sec:method\]. The results of the simulation are presented in Sec. \[sec:results\]. In Sec. \[section2\] we describe the two different models for the bath of two-level systems coupled to the qubit system and discuss the simulation and its results in comparison to data obtained from a D-Wave 2000Q quantum annealer. We conclude with a summary in Sec. \[sec:summary\]. Theoretical background {#sec:quantumannealing} ====================== In general, the Hamiltonian describing a quantum annealing process can be written as $$\begin{aligned} H(s) = A(s)H_\mathrm{init} + B(s)H_\mathrm{final},\label{eq:genH_qa}\end{aligned}$$ where $H_\mathrm{init}$ is the initial Hamiltonian whose ground state defines the state in which the system is prepared initially, $H_\mathrm{final}$ denotes the Hamiltonian at the end of the annealing process and whose ground state is the one to be determined, $s=t/t_a\in[0,1]$ is the rescaled (dimensionless) time, and $t_a$ is the total annealing time. The functions $A(s)$ and $B(s)$ determine the energy scale (in our case GHz) and the annealing scheme. They satisfy $|A(0)|\gtrsim 1$ and $B(0)\approx 0$, and $A(1)\approx 0$ and $|B(1)|\gtrsim 1$ with respect to the corresponding energy scale. From the adiabatic theorem [@born28], it follows that the system stays in the ground state of the instantaneous Hamiltonian $H(s)$ during the annealing process if $t_a\rightarrow \infty$ such that for $s=1$ the system is in the ground state of $H_\mathrm{final}$. Let $\Delta E_j(s)$ denote the difference between the energy of the ground state $\ket{GS(s)}$ and the $j$-th excited state $\ket{ES_j(s)}$ at the rescaled time $s$. A finite $t_a$ can be sufficient for the system to stay in the ground state if [@amin09_adiabatictheorem; @albash18] $$\begin{aligned} \underset{s\in[0,1]}{\mathrm{max}}\frac{|\bra{ES_j(s)}\partial_s H(s)\ket{GS(s)}|}{\Delta E_j(s)^2} \ll t_a.\end{aligned}$$ Quantum annealing can be used to solve optimization problems that can be mapped onto the Hamiltonian $H_\mathrm{final}$. The class of so-called quadratic unconstrained binary optimization (QUBO) problems can be mapped onto the Ising-spin model of the form $$\begin{aligned} H_\mathrm{QUBO}= -\sum_{k=1}^N h_k S_k - \sum_{\mathclap{1\le j<k}}J_{jk} S_j S_k,\label{eq:Hqubo}\end{aligned}$$ where $N$ is the number of binary variables $S_k\in \{-1, 1\}$, and $h_k$ and $J_{jk}$ are dimensionless real numbers defining the particular QUBO. The set of variables $\{S_k\}$ that minimizes Eq. (\[eq:Hqubo\]) gives the solution of the QUBO problem. Quantum annealing can, at least in principle, find (one of) the ground state(s) of the Ising-spin Hamiltonian Eq. (\[eq:Hqubo\]) [@kadowaki98] or, equivalently, solve the corresponding QUBO problem. For this purpose, the two-valued variables $S_k$ are replaced by the Pauli $Z$ matrices $\sigma_k^z$ with eigenvalues $\pm 1$ and eigenstates $\ket{\uparrow}$ and $\ket{\downarrow}$ such that Eq. (\[eq:Hqubo\]) transforms into $$\begin{aligned} H_\mathrm{Ising} = -\sum_{k=1}^N h_k\sigma_k^z - \sum_{\mathclap{1\le j<k}}J_{jk}\sigma_j^z\sigma_k^z.\label{eq:Hising}\end{aligned}$$ The product states of the $\sigma^z$ eigenstates define the so-called computational basis and are eigenstates of Eq. (\[eq:Hising\]). The ground state of the Hamiltonian Eq. (\[eq:Hising\]) can then be found by quantum annealing with $H_\mathrm{final}$ replaced by $H_\mathrm{Ising}$ in Eq. (\[eq:genH\_qa\]). For quantum annealing, the simplest choice for $H_\mathrm{init}$ is the Hamiltonian of spins in a transverse field [@kadowaki98] $$\begin{aligned} H_\mathrm{trans} = -\sum_{k=1}^N \sigma_k^x,\label{eq:Htrans}\end{aligned}$$ where $\sigma^x$ is the Pauli $X$ matrix. The ground state of this Hamiltonian is given by the product state $\ket{+\dots +}$, with $\ket{+} = (\ket{\uparrow} + \ket{\downarrow})/\sqrt{2}$. Equation (\[eq:Hising\]) is used to formulate optimization problems for the quantum annealer manufactured by D-Wave Systems Inc. [@harris09_newjournal]. By design, the parameters of the final Hamiltonian are restricted to $h_k\in [-2,2]$ and $J_{jk}\in [-1,1]$, and the qubit connectivity is given by the Chimera graph such that, in the notation of Eq. (\[eq:Hising\]), some $J_{jk}$ have to be set to zero [@harris10_eightqubit]. In the following sections, we discuss the SQUID model and describe the mapping of the SQUID model onto the qubit model in terms of Eqs. (\[eq:Hising\]) and (\[eq:Htrans\]). SQUID Model {#sec:squids} =========== In this section we introduce the model Hamiltonian of the three-SQUID system that is used to simulate the flux dynamics during quantum annealing. Two of the three SQUIDs serve as qubits, each qubit subspace being defined by projection onto the two lowest energy states of the individual SQUIDs. The third SQUID acts as a tunable coupler between the two other SQUIDs. Figure \[fig:squid\] shows the circuit of a SQUID with a compound Josephson junction (CJJ) loop. It is used as a building block for the flux qubits and the effective coupling between them in the D-Wave quantum annealer. Including the CJJ loop effectively leads to a tunable Josephson junction [@chiarello00]. The two qubit states correspond to the left-circulating and right-circulating persistent current in the superconducting (main) loop and the tunable Josephson junction is used to control the annealing process. For the coupler element, the tunable Josephson junction results in the tunable coupling strength [@harris09]. ![Sketch of a SQUID with a CJJ loop. The magnetic fluxes $\varphi$ and $\varphi_\mathrm{J}$ are the dynamical variables of the system. The external fluxes $\varphi^x$ and $\varphi^x_\mathrm{J}$ are used to control the operation of the device.[]{data-label="fig:squid"}](squid){width="30.00000%"} The Hamiltonian of a SQUID with a CJJ loop is given by [@han89; @harris10; @boixo16] $$\begin{aligned} H_\mathrm{SQUID} = &-E_C\partial_\varphi^2 + E_L\frac{(\varphi-\varphi^x)^2}{2}\nonumber\\ &- E_{C_{\mathrm{J}}}\partial_{\varphi_{\mathrm{J}}}^2 + E_{L_{\mathrm{J}}}\frac{(\varphi_{\mathrm{J}} -\varphi_{\mathrm{J}}^x)^2}{2}\nonumber\\ &- E_J\cos(\varphi)\cos\left( \frac{\varphi_{\mathrm{J}}}{2} \right) ,\label{eq:squid}\end{aligned}$$ where $\varphi=2\pi\Phi/\Phi_0=2e\Phi$ (we use $\hbar =1$) is the dimensionless magnetic flux in the main loop and $\varphi_\mathrm{J}$ the dimensionless magnetic flux in the CJJ loop with $\Phi_0$ denoting the magnetic flux quantum and $e$ the electron charge. In addition, $E_C$ and $E_{C_\mathrm{J}}$ are capacitive energies, $E_L$ and $E_{L_\mathrm{J}}$ are inductive energies, and $E_J$ is the Josephson energy. For an uncoupled SQUID, the inductive energy $E_L$ is given by $E_L= 1/(4e^2L)$, where $L=L_\mathrm{main}+L_\mathrm{J}/4$ [@harris10] is the total inductance. The potential of the flux variable $\varphi$, $V(\varphi)=E_L(\varphi-\varphi^x)^2/2-E_J\cos(\varphi_\mathrm{J}/2)\cos(\varphi)$, can be either a single potential well or a double-well potential depending on the value of $\varphi_\mathrm{J}$. Thus, $\varphi_\mathrm{J}$ can be used to change the shape of the potential and also the barrier height between the double wells [@han89; @Chiarello2005]. This property is used to control the annealing process via the external flux $\varphi_\mathrm{J}^x$ [@lanting14] and to set the coupling strength of the coupling SQUID [@harris09]. The external flux $\varphi^x$ can be used to tilt the potential, thereby lowering one of the wells and raising the other one [@Chiarello2005]. In terms of the qubit model (see Sec. \[sec:mapping\]), the external flux $\varphi^x$ can be used to set the parameter $h_k$ of the Ising Hamiltonian given in Eq. (\[eq:Hising\]) [@harris10_eightqubit]. Total Hamiltonian ----------------- So far, we have discussed the Hamiltonian of a single SQUID. In this section we introduce the Hamiltonian of the complete system consisting of the three SQUIDs. A tunable coupling constant $J_{jk}$ (see Eq. (\[eq:Hising\])) is realized by inserting a SQUID as a coupler element between the other two SQUIDs [@harris07; @harris09]. For the SQUID that functions as the coupler element, we denote the flux in the CJJ loop by $\varphi_{\mathrm{J},0}$ and the one in the main loop by $\varphi_0$. Accordingly, energies that correspond to the coupler main loop are labeled by an index “$0$”, and those that correspond to the coupler CJJ loop by an index “J,$0$”. The external control flux $\varphi_{\mathrm{J},0}^x$ can be used to tune the coupling strength between the SQUIDs. We label the fluxes of the SQUIDs corresponding to the qubits with indices “$1$” and “$2$”, respectively. Since in the idealized model, the SQUIDs are equal and subject to the same annealing functions $A(s)$ and $B(s)$, their energies and their external fluxes $\varphi_{\mathrm{J}}^x$ are equal. Therefore, we drop the indices “1” and “2” in these cases. Although the external fluxes $\varphi_1^x$, $\varphi_2^x$, and $\varphi_\mathrm{J}^x$ depend on time, we do not write this explicitly for reasons of readability. A sketch of the complete system is shown in Fig. \[fig:squids\]. ![Sketch of three SQUIDs to realize a tunable coupling. The magnetic fluxes $\varphi_i$ and $\varphi_{\mathrm{J},i}$ are the dynamical variables of the system. The external fluxes $\varphi_i^x$ and $\varphi^x_{\mathrm{J},i}$ are used to control the operation of the device. The parameters $\varphi^x_1$, $\varphi^x_2$, and $\varphi_{\mathrm{J},0}^x$ determine the values of the parameters $h_1$, $h_2$, and $J$ of the qubit model Eq. (\[eq:Hising\]), respectively.[]{data-label="fig:squids"}](squids){width="48.00000%"} By coupling SQUIDs, the inductive energies change such that they are given by $E_L(1+M^2/(LL_\mathrm{eff}))$ for the SQUIDs representing the qubits and by $E_{L_\mathrm{eff}}=E_{L_0}L_0/L_\mathrm{eff}$ for the coupler, where $L_\mathrm{eff}=L_0-2M^2/L$ is the effective inductance of the coupler element and $M$ is the mutual inductance between the coupler and the other SQUIDs’ main loops. In addition to the modified Hamiltonians of the three SQUIDs, the interaction terms $$\begin{aligned} H_\mathrm{int} &= \frac{M}{L_{\mathrm{eff}}}E_L\left( \varphi_1-\varphi_1^x \right)\left( \varphi_0-\varphi_0^x \right)\nonumber\\ &+ \frac{M}{L_\mathrm{eff}}E_L\left( \varphi_2-\varphi_2^x \right)\left( \varphi_0-\varphi_0^x \right) \nonumber\\ &+ \frac{M^2}{LL_\mathrm{eff}}E_L\left( \varphi_1-\varphi_1^x \right)\left( \varphi_2-\varphi_2^x \right),\label{eq:coupling}\end{aligned}$$ have to be included [@vandenbrink05]. The tunable coupler can be operated without an external flux in the coupler main loop if the junction asymmetry is negligible, i.e., the difference between the critical currents (of the two junctions of the coupler SQUID) is much smaller than the sum of these critical currents [@harris09]. Since we do not consider junction asymmetries in the idealized model, we set $\varphi_0^x=0$. Collecting all terms, the total Hamiltonian is given by $$\begin{aligned} H_\mathrm{total} &= \sum\limits_{i=1}^2 \Bigg[ - E_{J}\cos(\varphi_i)\cos\left( \frac{\varphi_{\mathrm{J},i}}{2} \right)+ E_{L_{\mathrm{J}}}\frac{(\varphi_{\mathrm{J},i}-\varphi_{\mathrm{J}}^x)^2}{2} - E_{C_{\mathrm{J}}}\partial_{\varphi_{\mathrm{J},i}}^2+E_{L}\!\left(\! 1+\frac{M^2}{LL_\mathrm{eff}} \right)\frac{(\varphi_i-\varphi_i^x)^2}{2}-E_{C}\partial_{\varphi_i}^2 \Bigg]\nonumber\\ &+E_{L_\mathrm{eff}}\frac{\varphi_0^2}{2}-E_{C_0}\partial_{\varphi_0}^2 + E_{L_{\mathrm{J},0}}\frac{(\varphi_{\mathrm{J},0}-\varphi_{\mathrm{J},0}^x)^2}{2} - E_{C_{\mathrm{J},0}}\partial_{\varphi_{\mathrm{J},0}}^2 - E_{J_0}\cos(\varphi_0)\cos\left( \frac{\varphi_{\mathrm{J},0}}{2} \right)\nonumber\\ &+\frac{M}{L_{\mathrm{eff}}}E_L\left( \varphi_1-\varphi_1^x \right)\varphi_0 + \frac{M}{L_\mathrm{eff}}E_L\left( \varphi_2-\varphi_2^x \right) \varphi_0 + \frac{M^2}{LL_\mathrm{eff}}E_L\left( \varphi_1-\varphi_1^x \right)\left( \varphi_2-\varphi_2^x \right).\label{eq:Htotal}\end{aligned}$$ This is the final Hamiltonian for which we solve the TDSE without further simplification. Effective coupling {#sec:model_coupling} ------------------ The idea of inserting the coupler element is that it leads to a tunable effective coupling between the other two SQUIDs [@vandenbrink05; @harris07; @harris09] such that the interaction Hamiltonian takes the form $$\begin{aligned} H_\mathrm{int}^\mathrm{eff} = C(\varphi_{\mathrm{J},0}^x) \varphi_1 \varphi_2 ,\label{eq:Heff_int}\end{aligned}$$ where $C(\varphi_{\mathrm{J},0}^x)$ is the effective coupling strength tunable by the external flux $\varphi_{\mathrm{J},0}^x$ of the coupler CJJ loop. To derive an approximate effective Hamiltonian that exhibits such a coupling term, we first replace the flux of the coupler CJJ loop $\varphi_{\mathrm{J},0}$ by its approximate expectation value. To obtain this expectation value, we expand the Hamiltonian of the SQUID given in Eq. (\[eq:squid\]) to second order in $\varphi_{\mathrm{J},0}-\varphi_{\mathrm{J},0}^x$ and set $\varphi_0=0$. The resulting Hamiltonian describes a shifted harmonic oscillator $$\begin{aligned} H_\mathrm{co} =\frac{E_{L_{\mathrm{J},0}}'}{2} \Bigg(\varphi_{\mathrm{J},0} &-\left(\varphi_{\mathrm{J},0}^x -\frac{2E_{J_0}\sin\left(\varphi_{\mathrm{J},0}^x/2\right)}{4E_{L_{\mathrm{J},0}}'}\right)\Bigg)^2\nonumber\\ &- E_{C_{\mathrm{J},0}}\partial_{\varphi_{\mathrm{J},0}}^2,\end{aligned}$$ where $E_{L_{\mathrm{J},0}}' = E_{L_{\mathrm{J},0}}+E_{J_0}\cos(\varphi_{\mathrm{J},0}^x/2)/4$. The expectation value of $\varphi_{\mathrm{J},0}$ in its ground state can thus be identified as $$\begin{aligned} \langle \varphi_{\mathrm{J},0} \rangle = \varphi_{\mathrm{J},0}^x -\frac{2E_{J_0}\sin\left(\varphi_{\mathrm{J},0}^x/2\right)}{4E_{L_{\mathrm{J},0}}+E_{J_0}\cos\big(\varphi_{\mathrm{J},0}^x/2\big)}.\label{eq:expv}\end{aligned}$$ With $\varphi_{\mathrm{J},0}$ replaced by $\langle \varphi_{\mathrm{J},0}\rangle$, we can find a matrix $T$ such that the transformation $$\begin{aligned} H^{\mathrm{eff}} &= e^{iT(t)}H_\mathrm{total}e^{-iT(t)} +i\left(\frac{\mathrm{d}}{\mathrm{d}t}e^{iT(t)}\right)e^{-iT(t)} \label{eq:trafo}\end{aligned}$$ of the total Hamiltonian yields an effective Hamiltonian which contains an interaction term of the form Eq. (\[eq:Heff\_int\]). Choosing $$\begin{aligned} T &= T(t) =i\frac{M}{L(1+\beta_\mathrm{eff})}\left( \varphi_1-\varphi_1^x +\varphi_2 -\varphi_2^x \right)\partial_{\varphi_0},\end{aligned}$$ where the external fluxes $\varphi_1^x$ and $\varphi_2^x$ depend on time, $$\begin{aligned} \beta_\mathrm{eff}=\frac{E_{J_0}}{E_{L_\mathrm{eff}}}\cos\left(\frac{\langle\varphi_{\mathrm{J},0}\rangle}{2}\right), \label{eq:beff}\end{aligned}$$ and expanding $$\begin{aligned} \cos\left( \varphi_0 -\frac{M}{L(1+\beta_\mathrm{eff})}\left( \varphi_1-\varphi_1^x +\varphi_2 -\varphi_2^x \right) \right)\end{aligned}$$ to second order in (the products of) $\varphi_0$, $\varphi_1-\varphi_1^x$, and $\varphi_2-\varphi_2^x$, we obtain the effective Hamiltonian $$\begin{aligned} H^\mathrm{eff} = \sum\limits_{i=1}^2\Bigg[ &E_L\left(1+\frac{M^2}{LL_\mathrm{eff}}\frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}}\right)\frac{\varphi_i^2}{2} - E_C\partial_{\varphi_i}^2 -E_J\cos(\varphi_i)\cos\left( \frac{\varphi_{\mathrm{J,}i}}{2} \right) - E_{C_{\mathrm{J}}}\partial_{\varphi_{\mathrm{J,}i}}^2 + E_{L_{\mathrm{J}}}\frac{(\varphi_{\mathrm{J},i} -\varphi_{\mathrm{J}}^x)^2}{2}\Bigg] \nonumber\\ +\sum\limits_{i=1}^2\Bigg[&-E_L\left(1+\frac{M^2}{LL_\mathrm{eff}}\frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}}\right)\varphi_i^x\varphi_i \Bigg] +\sum\limits_{i=1}^2\Bigg[-E_L\frac{M^2}{LL_\mathrm{eff}}\frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}}\varphi_{j\neq i}^x\varphi_i\Bigg] +E_L\frac{M^2}{LL_\mathrm{eff}}\frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}} \varphi_1\varphi_2 \nonumber\\ -\Big(E_{C_\mathrm{0}} &+\frac{2E_C M^2}{L^2(1+\beta_\mathrm{eff})^2} \Big)\partial_{\varphi_\mathrm{0}}^2+E_{L_\mathrm{eff}}(1+\beta_\mathrm{eff})\frac{\varphi_\mathrm{0}^2}{2} + \frac{M}{L(1+\beta_\mathrm{eff})}\left(i\frac{\mathrm{d}}{\mathrm{d}t}\left( \varphi_1^x+\varphi_2^x \right)-2E_C\left( \partial_{\varphi_1} +\partial_{\varphi_2}\right) \right)\partial_{\varphi_\mathrm{0}}.\label{H_eff}\end{aligned}$$ In the basis defined by the transformation Eq. (\[eq:trafo\]), we obtain the term $C(\varphi_{\mathrm{J},0}^x)\varphi_1\varphi_2$ where the dependence on $\varphi_{\mathrm{J},0}^x$ is given via $\langle\varphi_{\mathrm{J},0}\rangle$ in $\beta_\mathrm{eff}$ (see Eqs. (\[eq:expv\]) and (\[eq:beff\])). The only coupling term between the coupler element and the other two SQUIDs that remains is the last term in Eq. (\[H\_eff\]) which is expected to be much smaller than the previous coupling terms since $E_C\ll E_L$. Note that none of the approximations made to derive Eq. (\[H\_eff\]) affect the simulation results, as these are obtained by solving the TDSE for the Hamiltonian Eq. (\[eq:Htotal\]). However, as discussed in the next section, the approximation Eq. (\[H\_eff\]) is necessary to relate the external flux $\varphi_{\mathrm{J},0}^x$ to the coupling constant $J_{12}$ (denoted by $J$ for two qubits), which appears in the Ising Hamiltonian Eq. (\[eq:Hising\]). Mapping to the qubit model {#sec:mapping} ========================== In this section we investigate the mapping of the flux model Eq. (\[H\_eff\]) onto the qubit model Eq. (\[eq:genH\_qa\]) with $H_\mathrm{final}$ and $H_\mathrm{init}$ given by Eqs. (\[eq:Hising\]) and (\[eq:Htrans\]), respectively. The two-qubit Hamiltonian reads $$\begin{aligned} H(s)=-A(s)( \sigma_1^x +\sigma_2^x) -B(s)( h_1\sigma_1^z +h_2\sigma_2^z + J\sigma_1^z\sigma_2^z).\label{eq:H_2level}\end{aligned}$$ As we will see below, reducing Eq. (\[H\_eff\]) to the generic form of Eq. (\[eq:H\_2level\]) enforces a specific choice of the external fluxes $\varphi_i^x$ (see Eq. (\[eq:phi\_i\])) and gives a relation between $J$ and $\varphi_{\mathrm{J},0}^x$ (see Eq. (\[eq:J\_of\_phicox\])). Since we have assumed the two SQUIDs to be identical, the mapping to the qubit model is the same for both, and therefore we omit the SQUID indices in this section. The two lowest-energy states $\ket{g}$ and $\ket{e}$ of each SQUID for $\varphi^x=0$ define the computational subspace [@harris10]. We obtain them by diagonalizing the first part in square brackets of Eq. (\[H\_eff\]) in $\varphi$- and $\varphi_\mathrm{J}$-space (see Sec. \[sec:method\] for the definition of the discretized basis). Note that the first summand given in Eq. (\[H\_eff\]) contains an effective change of the inductive energy depending on the value chosen for $\varphi_{\mathrm{J},0}^x$ (because $\beta_\mathrm{eff}$ depends on it). Therefore, the definition of the computational subspace changes with the coupling strength. This leads to slightly different annealing schemes, i.e., a dependence of $A(s)$ and $B(s)$ on $\varphi_{\mathrm{J},0}^x$, as observed experimentally [@harris09; @harris10]. The computational basis states $\ket{\uparrow}$ and $\ket{\downarrow}$ are defined as the eigenstates of the operator $\varphi$ (and thus of the second part in square brackets in Eq. (\[H\_eff\])) inside the computational subspace $\mathrm{span}\{\ket{g},\ket{e}\}$. We obtain $$\begin{aligned} \ket{\uparrow} &= a\ket{g} +b\ket{e} = \int\limits_{-\infty}^\infty \int\limits_{-\infty}^\infty \mathrm{d}\varphi\mathrm{d}\varphi_\mathrm{J} \,u(\varphi,\varphi_\mathrm{J}) \ket{\varphi\,\varphi_\mathrm{J}}, \label{eq:upstate}\\ \ket{\downarrow} &= a\ket{g} -b\ket{e} =\int\limits_{-\infty}^\infty \int\limits_{-\infty}^\infty \mathrm{d}\varphi\mathrm{d}\varphi_\mathrm{J} \,d(\varphi,\varphi_\mathrm{J}) \ket{\varphi\,\varphi_\mathrm{J}}, \label{eq:downstate}\end{aligned}$$ where $|a|=|b|=1/\sqrt{2}$ [@harris10] and $u(\varphi,\,\varphi_\mathrm{J})$ and $d(\varphi,\,\varphi_\mathrm{J})$ are the resulting amplitudes of the states $\ket{\uparrow}$ and $\ket{\downarrow}$ in $\varphi$- and $\varphi_\mathrm{J}$-space. Note that $\ket{g}$ and $\ket{e}$ depend on the time-dependent external flux $\varphi_\mathrm{J}^x$, implying that the definition of the computational states changes with time. The projection of the operator $\widetilde E_L\varphi$ with $$\begin{aligned} \widetilde E_L =E_L\left(1+\frac{M^2}{LL_\mathrm{eff}} \frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}}\right)\label{eq:tilde_E_L}\end{aligned}$$ has eigenstates $\ket{\uparrow}$ and $\ket{\downarrow}$ with eigenvalues $\pm I_p(s)/2e$, respectively. Thus, in this subspace, $\widetilde E_L\varphi$ is represented by $I_p(s)\sigma^z/2e$ and the first contributions in square brackets in Eq. (\[H\_eff\]) are mapped to $-\Delta(s) \sigma^x/2$, where $\Delta(s) = E_1(s)-E_0(s)$ is the energy gap between the ground state $\ket{g}$ and the first excited state $\ket{e}$. To derive the coupling terms, we write the SQUID indices $i$ again. For the terms in $\sigma_i^z$ and $\sigma_1^z\sigma_2^z$ to scale with the same annealing function $B(s)$ [@harris10_eightqubit], $\varphi_i^x$ has to be set to $$\begin{aligned} \varphi_i^x=h_i\gamma\frac{2eI_p(s)M^2}{L_\mathrm{eff}},\label{eq:phi_i}\end{aligned}$$ where $\gamma = \mathrm{max}_{\varphi_{\mathrm{J},0}^x} \beta_\mathrm{eff}E_L^2 /(1+\beta_\mathrm{eff})\widetilde E_L^2$. Disregarding the contribution of the last term in Eq. (\[H\_eff\]), we find that the Hamiltonian for $\varphi_0$ effectively decouples from the qubit Hamiltonian and thus the effective qubit Hamiltonian can be written as $$\begin{aligned} H^{\mathrm{eff},q}\approx &-\sum\limits_{i=1}^2 \left(\frac{\Delta(s)}{2}\sigma_i^x + h_i\gamma\frac{I_p^2(s)M^2}{L_\mathrm{eff}}\sigma_i^z \right)\nonumber\\ &- \frac{E_L}{\widetilde E_L}\frac{I_p^2(s)M^4}{LL_\mathrm{eff}^2}\frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}}\gamma\left( h_1\sigma_2^z +h_2\sigma_1^z \right) \nonumber\\ &+\frac{E_L^2}{\widetilde E_L^2}\frac{I_p^2(s)M^2}{L_\mathrm{eff}}\frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}}\sigma_1^z\sigma_2^z.\label{eq:Heff}\end{aligned}$$ For all $J\in [-1,1]$, we have $$\begin{aligned} -\gamma = -\underset{\varphi_{\mathrm{J},0}^x}{\mathrm{max}} \frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}}\frac{E_L^2}{\widetilde E_L^2} \le -J\gamma \le \underset{\varphi_{\mathrm{J},0}^x}{\mathrm{max}} \frac{\beta_\mathrm{eff}}{1+\beta_\mathrm{eff}}\frac{E_L^2}{\widetilde E_L^2} = \gamma.\end{aligned}$$ Thus, and because $E_L^2\beta_\mathrm{eff}(\varphi_{\mathrm{J},0}^x)/(\widetilde E_L^2(\varphi_{\mathrm{J},0}^x)(1+\beta_\mathrm{eff}(\varphi_{\mathrm{J},0}^x)))$ is monotonic, it is possible to find $\varphi_{\mathrm{J},0}^x$ such that $$\begin{aligned} \frac{\beta_\mathrm{eff}(\varphi_{\mathrm{J},0}^x)}{1+\beta_\mathrm{eff}(\varphi_{\mathrm{J},0}^x)}\frac{E_L^2}{\widetilde E_L^2(\varphi_{\mathrm{J},0}^x)}=-J\gamma \label{eq:J_of_phicox}\end{aligned}$$ for all $J\in [-1,1]$, and Eq. (\[eq:Heff\]) becomes $$\begin{aligned} H^{\mathrm{eff},q}\approx & -\sum\limits_{i=1}^2 \frac{\Delta(s)}{2}\sigma_i^x -\gamma\frac{I_p^2(s)M^2}{L_\mathrm{eff}}\Bigg( \sum\limits_{i=1}^2 h_i\sigma_i^z + J\sigma_1^z\sigma_2^z \nonumber\\ &- \frac{\widetilde E_L}{E_L}\frac{M^2}{LL_\mathrm{eff}} J\gamma\left( h_1\sigma_2^z +h_2\sigma_1^z \right)\Bigg), \label{eq:Heff_q}\end{aligned}$$ which has the structure of an Ising model in a transverse field. Comparing Eq. (\[eq:Heff\_q\]) to Eqs. (\[eq:Hising\]) and (\[eq:Htrans\]), we can identify $A(s)=\Delta(s)/2$ and $B(s)=\gamma I_p^2(s)M^2/L_\mathrm{eff}$ and see that $$\begin{aligned} H^{\mathrm{eff},q}\approx &-A(s)( \sigma_1^x +\sigma_2^x) \nonumber\\ &-B(s)\Bigg( h_1\sigma_1^z +h_2\sigma_2^z + J\sigma_1^z\sigma_2^z \nonumber\\ &- \frac{\widetilde E_L}{E_L}\frac{M^2}{LL_\mathrm{eff}} J\gamma\left( h_1\sigma_2^z +h_2\sigma_1^z \right)\Bigg),\label{eq:H_2l_model}\end{aligned}$$ where the last term only adds a small contribution since $M^2\ll LL_\mathrm{eff}$. Simulation {#sec:method} ========== This section starts with a brief description of the numerical technique used to perform the simulation of the three-SQUID model. Then we discuss the choice of the model parameters that appear in Eq. (\[eq:Htotal\]) and explain the method by which we numerically extract the annealing scheme and the qubit-qubit coupling $J$, which appears in Eq. (\[eq:H\_2level\]), from the dynamics of the fluxes. For the simulation of the time evolution of the system defined by Eq. (\[eq:Htotal\]), we use the Suzuki-Trotter product-formula algorithm [@suzuki84; @deraedt87] to numerically solve the TDSE $$\begin{aligned} i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle.\label{eq:tdse}\end{aligned}$$ The time-dependent Hamiltonian is discretized such that the state vector $|\psi(t)\rangle$ can be updated by a time step $\tau$ to $|\psi(t+\tau)\rangle$ using the time-evolution operator $U(t,t+\tau)=\exp( -i\tau H(t+\tau/2) )$. To implement the algorithm, we fix a basis for the description of $\ket{\psi(t)}$ and a decomposition of the Hamiltonian $H(t)=\sum_k A_{k}(t)$ such that $$\begin{aligned} e^{-iH(t)\tau} \approx e^{-iA_{1}(t)\tau} e^{-iA_{2}(t)\tau} \cdots e^{-iA_{K}(t)\tau} = U_{t,1}(\tau),\end{aligned}$$ is a good approximation for sufficiently small $\tau$ and the update of the state vector can be performed with two-component updates only. For a detailed description of how to choose the $A_k$, see Ref. [@deraedt87]. In our simulation, we use the second-order approach given by $$\begin{aligned} e^{-iH(t)\tau} \approx U_{t,1}\left(\tau/2\right)U^\dagger_{t,1}\left(-\tau/2\right).\end{aligned}$$ Note that there is no need to diagonalize the Hamiltonian or to store the full matrices representing the Hamiltonian or the time-evolution operator. For the description of the state $\ket{\psi}$, the fluxes $\varphi_i$ through the main loops are discretized, i.e., the wave function is defined at $\lambda_i$ discrete points $\varphi_{i\mathrm{min}} + l_i\Delta \varphi_i$, $l_i=0,\dots ,\lambda_i-1$. By studying the convergence of the numerical results as a function of $\lambda_i$ and $\Delta\varphi_i$, we find that $\lambda_1=\lambda_2=47$ and $-2.0 \le \varphi_{1},\varphi_2\le 2.0$, and $\lambda_0=31$ and $-1.0\le \varphi_{0}\le 1.0$ provide a good compromise between accuracy and computational work to solve the TDSE. Since $E_{L_\mathrm{J}}\gg E_J$ and $E_{L_\mathrm{J,0}}\gg E_{J_0}$, the Hamiltonian for $\varphi_{\mathrm{J},i}$ resembles an oscillator with small anharmonicity. In the harmonic-oscillator basis, the evolution of $\varphi_{\mathrm{J},i}$ can be well described with the three lowest states and thus the fluxes $\varphi_{\mathrm{J},i}$ through the CJJ loops can be discretized in the harmonic-oscillator basis and labeled by $\ket{m_i}$, $m_i=0,1,2$. In summary, the state $\ket{\psi}$ is represented by $$\begin{aligned} |\psi\rangle = \sum\limits_{\mathclap{\substack{l_0,l_1,l_2,\\m_0,m_1,m_2}}} \phi_{l_0,m_0,l_1,m_1,l_2,m_2}\ket{l_0\, m_0\, l_1\, m_1\, l_2\, m_2},\label{eq:discr_psi}\end{aligned}$$ where the amplitudes $\phi_{l_0,m_0,l_1,m_1,l_2,m_2}$ are stored as an array of $\lambda_1 \lambda_2 \lambda_0\times 3^3 \approx 1.85\times 10^6$ complex double-precision numbers. To store this array, approximately $30\,\mathrm{MB}$ of memory is needed. Parallelization of the state updates is implemented using OpenMP. Testing with decreasing time steps $\tau$ and studying the convergence, we find that for $\tau=1.5\times 10^{-5}\,\mathrm{ns}$ the results are sufficiently accurate. Due to this small time step, one quantum annealing run of $5\,\mathrm{ns}$ takes about 16 hours on a 24-core node of the supercomputer JURECA [@jureca]. Parameters ---------- The parameters of the Hamiltonian Eq. (\[eq:Htotal\]) and the values of the time-dependent flux $\varphi_\mathrm{J}^x(s)$, which determines the annealing scheme, were provided to us by D-Wave Systems Inc. and are typical values of the D-Wave 2000Q processor [@dwave_private]. The device parameters used in the simulation are slightly modified and listed in Table \[tab:parameters\] and $\varphi_\mathrm{J}^x(s)$ is plotted in Fig. \[fig:phicjjx\]. The external fluxes $\varphi_i^x$ and $\varphi_{\mathrm{J},0}^x$ are computed from Eq. (\[eq:phi\_i\]) and by solving Eq. (\[eq:J\_of\_phicox\]) numerically for $\varphi_{\mathrm{J},0}^x(J)$. Using the provided parameters, we compute the annealing scheme of a single SQUID by exact diagonalization of Eq. (\[eq:squid\]) with $\varphi^x=0$ and computing $\Delta(s)$ and $I_p(s)$ as described in Sec. \[sec:mapping\]. However, the resulting annealing scheme (data not shown) does not match the data of the annealing scheme provided to us by D-Wave Systems Inc. [@dwave_private] (see Fig. \[fig:scheme\] (dashed lines)). Better agreement between the two annealing schemes was found by using $E_C=4.68\,\mathrm{GHz}$ (which was computed from the provided capacitance directly) instead of $E_C=5.85\,\mathrm{GHz}$ (value provided by D-Wave Systems Inc.), $E_{L_\mathrm{J}}=54538\,\mathrm{GHz}$ instead of $E_{L_\mathrm{J}}=73388\,\mathrm{GHz}$, and $M=15.97\,\mathrm{pH}$ instead of $M=13.7\,\mathrm{pH}$, see Fig. \[fig:scheme\] (solid lines). The disagreement in $B(s)$ for small $s$ could not be removed by slight variation of the model parameters. Changing $\varphi_i^x$ would reduce the disagreement for the single-qubit terms, but at the same time Eq. (\[eq:phi\_i\]) would be violated, effectively yielding different functions for the single-qubit and two-qubit $\sigma^z$ terms. Thus, we decided to keep $\varphi_i^x$ as given by Eq. (\[eq:phi\_i\]). Parameter Value ---------------------- --------------------------- $ E_C $ $ 4.68\,\mathrm{GHz} $ $ E_L $ $ 3.48 \,\mathrm{THz} $ $ E_{C_\mathrm{J}} $ $ 133.02 \,\mathrm{GHz} $ $ E_{L_\mathrm{J}} $ $ 54.54 \,\mathrm{THz} $ $ E_J $ $ 7.80 \,\mathrm{THz} $ $ M $ $ 15.97 \,\mathrm{pH} $ : Values of the device parameters appearing in the Hamiltonian Eq. (\[eq:Htotal\]) and used in our numerical work.[]{data-label="tab:parameters"} Parameter Value -------------------------- -------------------------- $ E_{C_0} $ $ 9.02\,\mathrm{GHz} $ $ E_{L_0} $ $ 15.67\,\mathrm{THz} $ $ E_{C_{\mathrm{J},0}} $ $ 213.50\,\mathrm{GHz} $ $ E_{L_{\mathrm{J},0}} $ $ 354.18\,\mathrm{THz} $ $ E_{J_0} $ $ 18.72\,\mathrm{THz} $ : Values of the device parameters appearing in the Hamiltonian Eq. (\[eq:Htotal\]) and used in our numerical work.[]{data-label="tab:parameters"} ![(color online) External flux $\varphi_\mathrm{J}^x=2\pi\Phi_\mathrm{J}^x/\Phi_0$ as a function of the normalized annealing time $s$, as provided to us by D-Wave Systems Inc.[]{data-label="fig:phicjjx"}](phicjjx_paper){width="35.00000%"} ![(color online) Functions $A(s)$ and $B(s)$ of the annealing scheme as provided to us by D-Wave Systems Inc.  (dashed lines) in comparison to the annealing scheme obtained from the full Hamiltonian for an uncoupled SQUID (solid lines).[]{data-label="fig:scheme"}](annealing_scheme){width="35.00000%"} Estimation of the coupling strength and the annealing scheme {#sec:method_scheme} ------------------------------------------------------------ In order to map the full state Eq. (\[eq:discr\_psi\]) to the computational space, we have to trace out the degrees of freedom of the coupler element and project the resulting reduced density matrix onto the computational subspace. To do so, we discretize Eqs. (\[eq:upstate\]) and (\[eq:downstate\]) to obtain $$\begin{aligned} \ket{\uparrow} &= \sum\limits_{l,m} u_{l,m} \ket{l\, m},\label{eq:def_up}\\ \ket{\downarrow} &= \sum\limits_{l,m} d_{l,m} \ket{l\, m}\label{eq:def_down}\end{aligned}$$ for a single qubit, and accordingly the product states for the two-qubit states, where $u_{lm}$ and $d_{lm}$ are the discretizations of $u(\varphi,\,\varphi_\mathrm{J})$ and $d(\varphi,\,\varphi_\mathrm{J})$, respectively. Since this projection is not a unitary transformation, the trace of the projected density matrix $\rho^\mathrm{comp}$ will be less than one if there is leakage to higher levels, i.e., if excitations to states outside the computational subspace occur. The deviation of the trace from one is a measure for the amount of leakage to higher levels. To obtain the effective coupling strength and the annealing scheme we proceed as follows. We start with the ideal qubit Hamiltonian (for simplicity, with $h_1=h_2=0$) $$\begin{aligned} H_2(s) = -\frac{\Delta(s)}{2}\left( \sigma_1^x + \sigma_2^x \right) + C(s)\sigma_1^z\sigma_2^z,\end{aligned}$$ where $\Delta(s)$ and $C(s)$ are to be determined by comparison with the data obtained by simulating the model Eq. (\[eq:Htotal\]). For fixed $s\in [0,1]$, the evolution determined by $H_2(s)$ of the initial state $\ket{++} = \left( \ket{\uparrow} + \ket{\downarrow}\right)\otimes\left( \ket{\uparrow} + \ket{\downarrow}\right)/2$ and the expectation values $\langle \sigma_1^\alpha\sigma_2^\beta \rangle$ for $\sigma_i^\alpha,\sigma_i^\beta\in \{\mathbb{I}_i,\sigma_i^x,\sigma_i^y,\sigma_i^z\}$ in the evolved state can be calculated analytically. On the other hand, for any time $t$, we can compute these expectation values directly from the simulation of the time evolution of the initial state $\ket{{++}}$ expressed in flux degrees of freedom using Eqs. (\[eq:def\_up\]) and (\[eq:def\_down\]) at a fixed value for $s$. The time evolution is governed by the full Hamiltonian Eq. (\[eq:Htotal\]) based on the flux degrees of freedom with fixed $s$. In this case, the expectation values are computed by $\mathrm{Tr}(\rho^\mathrm{comp}\sigma_1^\alpha\sigma_2^\beta)$. We can then estimate $\Delta (s)$ and $C(s)$, and thus the effective coupling strength and the annealing scheme, by fitting the analytical expressions to the simulation data. Results {#sec:results} ======= In this section, we present the results obtained from the simulation of the flux model described by the Hamiltonian Eq. (\[eq:Htotal\]). First, we show that the mapping between $J$ and $\varphi_\mathrm{J,0}^x$ given by Eq. (\[eq:J\_of\_phicox\]) leads to the desired effective coupling strength. Subsequently, we discuss the effective annealing scheme obtained by using the procedure described in Sec. \[sec:method\_scheme\]. We check the results of the simulation based on the flux model and the results of the qubit model against each other by comparing the probabilities during and at the end of the annealing process. Finally, we briefly discuss the data obtained from the D-Wave quantum annealer in comparison to the simulation results. Effective coupling and annealing scheme --------------------------------------- In order to assess the mapping between $J$ and $\varphi_\mathrm{J,0}^x$ using Eq. (\[eq:J\_of\_phicox\]), we first study the effective mutual inductance $M_\mathrm{eff}$ as a function of $J$. We utilize the method described in Sec. \[sec:method\_scheme\] for various values of $J$ and $s=1$ (such that $\Delta(s)\approx 0$) to determine the coupling strength. In this case, the analytical result for the expectation value $\langle \sigma_1^y\sigma_2^z \rangle=\sin(2C(1)t)$ can be used for fitting. The obtained value for $C(1)$ for each $J$ is then mapped onto the effective mutual inductance $M_\mathrm{eff}(J)=C(1)/I_p^2(1)=-J\gamma M^2/L_\mathrm{eff}$ and plotted against $J$. The result for the effective inductance $M_\mathrm{eff}(J)$ between the qubits is presented in Fig. \[fig:J\] and shows good agreement between the theoretical linear curve from the approximation and the simulation result. For $J$ in the range $[-1,1]$, we can reach all values for $M_\mathrm{eff}$ in $[-M_\mathrm{eff,max},M_\mathrm{eff,max}]$ to good precision and have thus obtained a transformation $\varphi_\mathrm{J,0}^x \leftrightarrow J$ such that the mapping $J\leftrightarrow M_\mathrm{eff}$ is linear. Therefore, we can expect that the mapping onto the qubit model and the resultant mapping $J\leftrightarrow \varphi_\mathrm{J,0}^x$ work reasonably well. ![(color online) Effective mutual inductance between the qubits as a function of the qubit-qubit coupling $J$. The solid line shows the expected behavior based on the analytical calculation presented in Sec. \[sec:mapping\]. Asterisks show the simulation data.[]{data-label="fig:J"}](J){width="35.00000%"} To assess the effective annealing scheme, we use the method described in Sec. \[sec:method\_scheme\] for different values $s'\in[0,1]$, using the analytical expression of the expectation value $$\begin{aligned} \langle \sigma^z_1\sigma_2^z\rangle = \frac{2\Delta(s')C(s')\sin^2\left(\sqrt{\Delta(s')^2+C(s')^2}t\right)}{\Delta(s')^2+C(s')^2}\end{aligned}$$ for the fitting of $\Delta(s')$ and $C(s')$. Figure \[fig:annealing\_v1-1\] shows the effective annealing scheme (data points) obtained in this way. We find that the data points in Fig. \[fig:annealing\_v1-1\] deviate from the annealing scheme for an uncoupled qubit (solid lines, obtained by using Eq. (\[eq:squid\])), but they are in better agreement with the annealing scheme obtained by using $\widetilde E_L$ (see Eq. (\[eq:tilde\_E\_L\])) instead of $E_L$ (dashed lines). Note that for the computation of the annealing scheme, the single-SQUID Hamiltonian Eq. (\[eq:squid\]) is mapped onto the effective Hamiltonian $H^{\mathrm{eff},q}\approx -A(s)\sigma_1^x-B(s)h_1\sigma_1^z$. In the qubit model (Eq. (\[eq:H\_2level\])), this gives the same function for $B(s)$ as the term proportional to $\sigma_1^z\sigma_2^z$ in the case of two coupled qubits. Because of the choice for $\varphi_1^x$ (see Eq. (\[eq:phi\_i\])), this should also be the case for the SQUID model if the mapping to the qubit model works well enough. We find, in agreement with our analytical calculation, that the effective coupling between the SQUIDs induces a shift in the inductive energy, leading to shifts in the annealing scheme. Including this shift, the effective annealing scheme can be well described by the single-SQUID annealing scheme. The influence of the coupling on the inductive energy was also observed in experiments [@harris09]. ![(color online) Annealing scheme for the uncoupled single SQUID model obtained by diagonalization of Eq. (\[eq:squid\]) with $\varphi^x=0$ (solid line) and with $E_L$ replaced by $\widetilde E_L$ (see Eq. (\[eq:tilde\_E\_L\])) where $\varphi_\mathrm{J,0}^x$ is set to correspond to $J=-1$ (dashed line) to show the effect of coupling on the annealing scheme. Blue asterisks and red squares are obtained from the simulation of the coupled model Eq. (\[eq:Htotal\]) as described in Sec. \[sec:mapping\]. Solid and dashed lines following the asterisks represent $A(s)$ and solid and dashed lines following the squares represent $B(s)$. The parameters are $J=-1$ and $h_1=h_2=0$. []{data-label="fig:annealing_v1-1"}](annealing_v-1){width="35.00000%"} Comparison to the qubit model ----------------------------- The next step is to compare the overall performance and the final probabilities between the real-time simulation with the Hamiltonian given in Eq. (\[eq:Htotal\]) and the qubit Hamiltonian given in Eq. (\[eq:H\_2level\]). As mentioned in Sec. \[sec:method\], the amount of leakage to higher excited states can be computed by projecting the density matrix onto the computational subspace. The projected density matrix can also be used to obtain the probabilities of the computational basis states. As an illustration, in Fig. \[fig:comp\_states\] we show the results for $J=-1$, $h_1=0.96$, and $h_2=0.94$. For this choice of parameters, the ground state of Eq. (\[eq:Hising\]) is $\ket{{\uparrow\downarrow}}$. The total annealing time was set to $t_a = 5\,\mathrm{ns}$ for the simulations of both the flux model and the qubit model. ![(color online) Probabilities of the four computational states $\ket{\uparrow \uparrow}$ (black triangles), $\ket{\uparrow \downarrow}$ (green asterisks), $\ket{\downarrow \uparrow}$ (blue solid circles), and $\ket{\downarrow \downarrow}$ (red squares) during the annealing process for the qubit model (dashed lines) and the full system (solid lines). Markers are used to better distinguish the lines of the different states. For the data from the simulation of the qubit model, every 120th data point is plotted with a marker and for the data from the simulation of the full system, every 6th point is plotted with a marker. For the full system, additionally the probability of leakage (cyan open circles) is shown using the right $y$ axis. The annealing time was set to $t_a=5\,\mathrm{ns}$. The parameters are $J=-1$, $h_1=0.96$, and $h_2=0.94$.[]{data-label="fig:comp_states"}](comp_states){width="35.00000%"} This annealing time is much less than typically used on the D-Wave processors (order of microseconds), but for comparison of the results of the flux simulation with the qubit description only, this difference is unimportant. Note that in this section our aim is to scrutinize the validity of the qubit model as a description of the flux dynamics governed by the Hamiltonian Eq. (\[eq:Htotal\]), not to compare simulation results with experiments performed on the D-Wave quantum annealer (see Secs. \[sec:compare\_d\_wave\] and \[section2\] below). As seen from Fig. \[fig:comp\_states\], there are small deviations from the probabilities obtained from the qubit representation. Some leakage which has its maximum at about $s=0.6$, where the change in the probabilities of the computational states is strongest, can also be observed. In general, the evolutions of both the full model and the qubit model show the same features. In the following we refer to the probability of finding the system at the end of the annealing process in the ground state of Hamiltonian Eq. (\[eq:Hising\]) as success probability. For the example case shown in Fig. \[fig:comp\_states\], the success probability for the flux model and the qubit model differ only slightly. In this example, the success probability is higher for the qubit model. However, Fig. \[fig:gap\_probs\] shows that there are also cases in which the success probability is lower for the qubit model. Note that the annealing process does not start with equal probability for all states because we start the annealing in the ground state of the system instead of in the state $\ket{++}=\ket{+}_1\otimes\ket{+}_2$, since for $B(s=0)>0$, the ground state of the qubit model Eq. (\[eq:Hising\]) is not exactly the state $\ket{++}$ but a superposition of all basis states. A simulation of the qubit model comparing the annealing processes with the two different initial states shows deviations during the annealing process, but there is no significant difference in the success probability (data not shown). ![(color online) Success probability as a function of the minimal energy gap $\delta E = \min_s E_1(s)-E_0(s)$ (computed from the qubit model) during the annealing process. Each data point represents another problem, i.e., other values for the parameters $h_1$, $h_2$, and $J$. A list with all cases is given in Appendix \[app:data\]. Closed (red) circles show the results for the qubit model and open (black) circles originate from the simulation of the flux model.[]{data-label="fig:gap_probs"}](gap_probs){width="40.00000%"} In summary, we observed an influence of the coupling on the annealing scheme and some amount of leakage to higher levels. The important question, however, is whether these effects have consequences on the final success probability. Figure \[fig:gap\_probs\] shows the success probability for many different problems (defined in Appendix \[app:data\]) as a function of the minimal energy gap between the ground state and the first excited state during the annealing process, computed from the qubit model. As can be seen in Fig. \[fig:gap\_probs\], for most of the investigated cases, the effects on the success probability of using a subspace of a larger system as the qubit instead of an ideal qubit representation are rather small. The data points generated by the simulation based on Eq. (\[eq:Htotal\]) (black open circles) show, apart from a few exceptions, only small deviations from the data points of the ideal qubit model (red closed circles). Note that the success probability can be enhanced as well as reduced compared to the ideal qubit model. Another interesting observation can be made in Fig. \[fig:gap\_probs\]. For large minimal energy gaps $\delta E = \min_s E_1(s)-E_0(s)$, the system shows Landau-Zener behavior [@landau32; @zener32]. For small minimal energy gaps $\delta E$, the success probabilities form two clusters, one cluster of probability approximately equal to $0.5$ and a second one of probability approximately equal to $ 0.3$. Considering the energy spectra of these instances, we can separate them into three classes. The first class would be the one where the energies of all three excited states come close to the ground-state energy and the success probability clusters at roughly $0.3$. The second class of problem instances has a spectrum similar to the one shown in Fig. \[fig:spec\_all\](a). Only the energy of the first excited state comes close to the ground-state energy. The energies of the second and third excited states are much higher. For this class, the success probability clusters around $0.5$. The third class of problems has a spectrum similar to the one shown in Fig. \[fig:spec\_all\](b). The energies of the first and second excited states come close to the ground-state energy. For this class, the success probability depends on the particular problem instance. The reason is that for degenerate ground states, the probabilities to find these states are not necessarily equal [@Matsuda09] and for problem instances which are close to these degenerate cases with unequal probabilities, this imbalance may have an influence when non-adiabatic transitions occur. All instances which show larger deviations between the success probabilities obtained from the qubit model and the flux model simulations (see Fig. \[fig:gap\_probs\]) belong to the third class. A possible explanation for these deviations might be that, in some cases, due to the presence of the additional states in the flux model, these unequal probabilities in the computational subspace are different during the evolution than in the qubit model. We leave a more detailed study of this phenomenon for future work. {width="90.00000%"} As noted in Ref. [@albash15; @dwave_manual] and confirmed by our analytical calculation in Sec. \[sec:mapping\], there is some crosstalk between the qubits (last term in Eq. (\[eq:H\_2l\_model\])), leading to small offsets in the parameters $h_i$. Furthermore, a dependence of the annealing scheme on the parameter $\varphi_{\mathrm{J},0}^x$ was found, also leading to small discrepancies between the ideal qubit representation and the full system. Additionally, for the mapping of $J$ to $\varphi_{\mathrm{J},0}^x$, we had to draw on an approximate analytical calculation which may be another source for the small differences between the results obtained from the two models. Nevertheless, the results fit very well. Interestingly, the coupler element, which can be viewed as part of the environment and might be the source of additional noise, does not cause significant deviations in the results compared to the results of the qubit description. Comparison to D-Wave 2000Q data {#sec:compare_d_wave} ------------------------------- Because we find good agreement between the system described by the Hamiltonian Eq. (\[eq:Htotal\]) and the qubit model, we compare the success probability for both these systems with the percentage of successful runs on the D-Wave 2000Q quantum annealer. In Appendix \[app:data\] (see Table \[tab:cases\]), we present the data obtained by at least ten repetitions of annealing simultaneously, 992 (976) copies of the two-qubit problems distributed over the D-Wave DW\_2000Q\_2 (DW\_2000Q\_2\_1) chip, for an annealing time $t_a=20\,\mu\mathrm{s}$. Postprocessing and autoscaling have been turned off for all experiments on the D-Wave 2000Q. We find that although the annealing time on the D-Wave is much larger than for our simulations ($20\mu$s instead of $5$ns), a large fraction of the D-Wave data seems to agree (approximately) with the corresponding success probabilities obtained from the simulation of the SQUID model Eq. (\[eq:Htotal\]) and its two-level approximation. This agreement is probably accidental. The annealing time of $5\,$ns was chosen to keep the real time to solve the TDSE of the SQUID model within acceptable limits as well as having some variation in the success probability at the end of the annealing process without having to use too-small values for, or differences between, the parameters $J$ and $h_i$. In spite of the large difference in annealing times, the good agreement suggests that in the D-Wave device there are physical processes at work that affect the annealing, processes which are not incorporated in the SQUID model Eq. (\[eq:Htotal\]) or the corresponding qubit model Eq. (\[eq:H\_2l\_model\]). Concrete evidence for the relevance of such processes is shown in Table \[tab:d\_wave\_data\], where we present D-Wave data for three different cases whose energy spectra are shown in Figs. \[fig:spec\_all\](a)–(c). Because the spectra of these cases differ significantly, we assume that they are a representative subset of the cases studied previously. Data characterizing the problem instances such as the minimal energy gap $\delta E$ and the problem gap $\Delta_p$ of the final Hamiltonian are listed in Table \[tab:d\_wave\_data\] as well as the frequency of runs finding the ground states ($\ket{\downarrow\uparrow},\ket{\uparrow\downarrow},$ and $\ket{\uparrow\downarrow}$ for the three cases, respectively) on the D-Wave machine for four different annealing times. The results reported in Table \[tab:d\_wave\_data\] were obtained by putting 992 copies of the two-qubit problems on the Chimera graph and performing 1000 annealing runs. ----- ------- -------- -------- --------- --------- ----------- ------------ ------------- ------------------ $1\,\mu$s $20\,\mu$s $100\,\mu$s $1\,\mathrm{ms}$ (a) $-1$ $0$ $0.05$ $1.206$ $3.519$ $63.0$ $65.6$ $67.1 $ $69.7$ (b) $-1$ $0.96$ $0.94$ $0.627$ $1.407$ $51.4$ $52.9$ $53.6 $ $55.6$ (c) $0.1$ $0.3$ $-0.3$ $5.481$ $14.07$ $92.5$ $96.2$ $97.6 $ $98.5$ ----- ------- -------- -------- --------- --------- ----------- ------------ ------------- ------------------ : Percentage for finding the ground state (GS) on D-Wave’s DW\_2000Q\_2 chip for three different problem instances and four different annealing times. The minimal and final gap (in GHz) are denoted by $\delta E$ and $\Delta_p$, respectively.[]{data-label="tab:d_wave_data"} Table \[tab:d\_wave\_data\] shows some additional interesting facts. First, recall that for the shortest annealing time considered, i.e., $1\,\mu\mathrm{s}$, simulation of the quantum annealing process in the qubit description yields the ground state with probability one for the three cases listed. Clearly, as Table \[tab:d\_wave\_data\] shows, this is not the case for the D-Wave data, not even if we increase the annealing time to $1\,\mathrm{ms}$, as is most evident for case (c). We emphasize that these differences in the observed frequencies for finding the ground state are not due to poor statistics nor can they be attributed to the differences in the minimal gaps $\delta E$. Correlating these observations with the energy level spectra shown in Fig. \[fig:spec\_all\] suggests that the observed reduction of the frequency for finding the ground state may be related to the distribution of energy levels for $s\rightarrow1$. However, the mechanism that causes the observed change of frequencies when going from case (a) to (c) cannot be found within the description of the quantum dynamics of the two-qubit system. Explaining these experimental observations requires considering additional physical processes. The first process that comes to mind is the interaction of the qubits with their environment at non-zero temperature. In the next section, we address this issue by solving the TDSE of the two-qubit model Eq. (\[eq:H\_2l\_model\]) coupled to a bath of two-level systems, complementing previous work that investigated the effects of finite temperature on quantum annealing [@amin08; @johansson09; @amin09; @amin09_decoherence; @amin09_nonmarkovian; @dickson13; @amin15]. System coupled to a bath {#section2} ======================== The annealing process of the isolated two-qubit system may be understood in terms of the adiabatic theorem. However, in the real world, the system modeling the two qubits is in contact with an environment at finite temperature. In this section we scrutinize the extent to which the coupling to a heat bath, i.e., the presence of thermal fluctuations, affects the annealing process. This, we hope, may shed light on the annealing behavior observed on the D-Wave machine in the cases studied. We do not assume the heat bath to be Markovian but instead we solve the TDSE of the system comprising the two-qubit system and the bath. In order to be able to perform such simulations, it is necessary to keep these models simple. Therefore, it would be remarkable to find good quantitative agreement between the results of the simulations and those obtained with the D-Wave machine. Thus, the goal here is limited to find out if such models can reproduce, qualitatively, the trends observed in the D-Wave data. We model the heat bath as a collection of two-level systems [@Phillips1972; @anderson72] which might represent e.g., defects in the material. Such models have been used to discuss noise and dephasing in superconducting resonators and circuits [@shnirman05; @mueller09; @cole10; @burnett14; @faoro15; @Lisenfeld15; @deGraaf18; @mueller19]. We assume that this heat bath is at thermal equilibrium, with a temperature given by the operating temperature of the D-Wave machine. We denote the inverse of this operating temperature by $\beta^*=0.588\,\mathrm{ns}$ (in units of $\hbar=k_B=1$), corresponding to a temperature of $T\approx 13\,\mathrm{mK}$. The Hamiltonian of the system (S) + bath (B) reads $$\begin{aligned} \mathcal{H}(t)&=& H_\mathrm{S}(s=t/t_a) + H_\mathrm{B} + \lambda H_\mathrm{SB} , \label{s40}\end{aligned}$$ where $\lambda$ controls the overall strength of the system-bath interaction. The time evolution during the quantum annealing process of the closed quantum system defined by the Hamiltonian Eq. (\[s40\]) is obtained by solving the TDSE given in Eq. (\[eq:tdse\]) with the initial state $$\begin{aligned} |\Psi(t=0)\rangle=\ket{++}\otimes|\Phi(\beta)\rangle.\end{aligned}$$ The method to prepare the thermal state $|\Phi(\beta)\rangle$ and other technical details are discussed in Appendix \[app:technical\_description\]. The system Hamiltonian is given by $$\begin{aligned} H_\mathrm{S}&=A(s) \left(-\sigma^x_{1}-\sigma^x_{2}\right) + B(s)\left( -J\sigma^z_{1}\sigma^z_{2} - h^z_1\sigma^z_{1} - h^z_2\sigma^z_{2}\right) \label{s41}\end{aligned}$$ and changes with time as a function of the annealing variable $s=t/t_a$. We consider two extreme cases for $H_\mathrm{B}$ and $H_\mathrm{SB}$. Model I ------- In the first case, the bath is modeled as a ring of two-level systems represented by the Pauli matrices $\bm\mu_{n}=(\mu^x_{n},\mu^y_{n},\mu^z_{n})$, described by the Hamiltonian $$\begin{aligned} H_\mathrm{B}&= -K\sum_{n=1}^{N_{\mathrm{B}}}\left( r^x_n\mu^x_{n}\mu^x_{n+1} +r^y_n\mu^y_{n}\mu^y_{n+1} +r^z_n\mu^z_{n}\mu^z_{n+1} \right) .\label{eq:H_bath}\end{aligned}$$ Here and in the following $N_{\mathrm{B}}$ denotes the number of bath particles. The couplings $r^x_n$, $r^y_n$, and $r^z_n$ are taken to be uniform random numbers in the range $[-1,+1]$ and $K$ determines the spectral range of $H_\mathrm{B}$. For random couplings it is unlikely that the model Eq. (\[eq:H\_bath\]) is integrable (in the Bethe-ansatz sense) or has any other special features such as conserved magnetization. The bath Hamiltonian Eq. (\[eq:H\_bath\]) with random couplings has the property that the distribution of nearest-neighbor energy levels is Wigner-Dyson-like [@ZHAO16]. Extensive simulation work on spin baths with very different degrees of connectivity [@JIN13a] suggests that as long as there is randomness in the system-bath coupling and randomness in the intrabath coupling, the simple model Eq. (\[eq:H\_bath\]) may be considered as a generic spin bath [@ZHAO16]. The Hamiltonian describing the interaction of the two-qubit system with the bath is taken to be $$\begin{aligned} H_\mathrm{SB}&= -r^x_{n,1}\mu^x_{n}\sigma^x_{1} -r^y_{n,1}\mu^y_{n}\sigma^y_{1} -r^z_{n,1}\mu^z_{n}\sigma^z_{1}\nonumber\\ &-r^x_{m,2}\mu^x_{m}\sigma^x_{2} -r^y_{m,2}\mu^y_{m}\sigma^y_{2} -r^z_{m,2}\mu^z_{m}\sigma^z_{2} , \label{s43b}\end{aligned}$$ where $n$ and $m$ are chosen randomly from the set $\{1,\ldots,N_{\mathrm{B}}\}$ such that $n\not=m$. The $r^\alpha_{n,1}$ and $r^\alpha_{m,2}$ are real-value random numbers in the range $[-1,+1]$. Model II -------- In this case, the bath is modeled as a collection of non-interacting two-level systems given by the Hamiltonian $$\begin{aligned} H_\mathrm{B} = -\Omega \sum\limits_{n=1}^{N_\mathrm{B}} r_n^x\mu_n^x +r_n^y\mu_n^y + r_n^z\mu_n^z,\label{eq:H_bathII}\end{aligned}$$ where the parameter $\Omega$, together with the random numbers $r_n^x,\,r_n^y,\,r_n^z\in [-1,1]$, determines the level splitting of each two-level system. The interaction between the qubits and the two-level systems of the bath is chosen such that each qubit interacts with a different half of the bath. This type of interaction is modeled by the Hamiltonian $$\begin{aligned} H_\mathrm{SB} =\!\! \sum\limits_{\alpha=x,y,z}\left[ \sigma_1^\alpha\sum\limits_{n=1}^{N_{\mathrm{B}}/2}r_{n,1}^\alpha\mu_n^\alpha + \sigma_2^\alpha\,\sum\limits_{\mathclap{n=N_\mathrm{B}/2+1}}^{N_\mathrm{B}}\,r_{n,2}^\alpha\mu_n^\alpha \right].\end{aligned}$$ Parameters ---------- Obviously, to compare the simulation results with D-Wave results it is necessary to solve the TDSE for the same time interval as used on the D-Wave machine. In practice, this requirement puts a severe constraint on the size of the problems for which one can solve the TDSE within a reasonable amount of real time. Simulating the annealing process over $1\,\mu\mathrm{s}$ (the shortest annealing time possible on the D-Wave machine) for a system comprising $N_B=16$ on a BullSequana X1000 supercomputer (JUWELS [@JUWELS]) takes about 4 hours using 40 compute cores. Performing a simulation for $N_B=28$ two-level systems in the bath takes about one week (400000 time steps of $0.0025\,\mathrm{ns}$ using 5120 compute cores). Earlier work which studied the TDSE dynamics of two spins coupled to a spin bath [@RAED17b] shows that the results for $N_B=16$ and $N_B=28$ primarily differ in the statistical fluctuations on the data (see also Appendix \[app:technical\_description\]). Therefore, we have chosen to perform most simulations with $N_B=16$, repeating runs with different random numbers to collect statistics, and use a few runs with $N_B=28$ as an additional check on the data. We use the annealing schedule of the DW\_2000Q\_2 chip (see Fig. \[fig:DW\_schedule\]) which is the machine that we used for our experiments. ![(color online) Annealing schedule of the DW\_2000Q\_2 chip which was used for the simulation with the environment. The red solid line corresponds to $A(s)$ and the blue dashed line corresponds to $B(s)$.[]{data-label="fig:DW_schedule"}](schedule){width="35.00000%"} In the case of model I, the initial state of the bath is prepared by projection with the operator $\exp(-\beta H_B)=\exp(-\beta K (H_B/K))$, as explained in Appendix \[app:technical\_description\]. From Eqs. (\[B2\]) and (\[B3\]) it is clear that baths with the same $\beta K$ (and the same $r_n^x$, $r_n^y$, and $r_n^z$; see Eq. (\[eq:H\_bath\])) will have the same thermal equilibrium properties. Therefore, we may use $K$ as an adjustable parameter to “calibrate” the temperature of the bath with respect to the operating temperature of the D-Wave machine on which we performed our experiments. For instance, for a fixed choice of $r_n^x$, $r_n^y$ and $r_n^z$, baths with $( K=5\,\mathrm{GHz},\,\beta=0.2\,\mathrm{ns}\,(T\approx 38.2\,\mathrm{mK}))$ and $(K=5/3\,\mathrm{GHz},\,\beta=0.6\,\mathrm{ns}\,(T\approx 12.7\,\mathrm{mK}) )$ have the same thermal properties. On the other hand, $K$ sets the time scale of the dynamics of the two-level systems of the bath. Simulations with $K$ in the range $[1\,\mathrm{GHz},5\,\mathrm{GHz}]$ (data not shown) reveal that the primary quantity of interest, the success probability of the two-qubit system at $s=1$, does not change significantly with $K$ (in the mentioned range and for the same value of $\lambda$). This leaves only the system-bath interaction $\lambda$ as a parameter to fit the simulation data to the D-Wave data. In the case of model II, $\Omega$ plays the role of $K$ in model I, i.e., $\beta\Omega$ determines the thermal equilibrium properties of the bath. Note that for modest values of $N_B$, model II is too simple to act as a genuine heat bath, but as a model for a few defects interacting with the SQUIDs, it can be a realistic choice [@shnirman05; @mueller09; @cole10]. Therefore, in this case, we set $\beta\approx\beta^*$ and use $\Omega$ and $\lambda$ as fitting parameters. Comparison to D-Wave 2000Q data {#comparison-to-d-wave-2000q-data} ------------------------------- {width="95.00000%"} Figures \[fig:bath\_all\](a)–(c) depict the results of the simulation with the heat bath (closed circles) averaged over ten different random initializations of the heat bath with $N_B=16$. Results for $N_B=28$ (asterisks) show that for each value of $\beta$, the averages of 10 samples of $N_\mathrm{B}=16$ data are in good agreement with the data obtained from one $N_\mathrm{B}=28$ sample. The solid line indicates the probability $p_0$ of finding the ground state of the isolated qubit system in thermal equilibrium for $H_\mathrm{S}$ at $s=1$, i.e., $p_0=\exp(-\beta E_0)/Z$, where $E_0$ is the ground-state energy and $Z=\mathrm{Tr}(\exp(-\beta H_S))$ is the partition function. Qualitatively, the simulation data obtained using model I (closed circles) nicely match the equilibrium line. The deviations from the equilibrium line may be due to the freeze-out where thermal transitions stop [@amin15] and/or too-short annealing times and/or the magnetic Foehn effect [@saito01]. However, the simulation data do not match the data generated on the D-Wave machine (crosses). Assuming that the qubit system on the D-Wave machine is in thermal equilibrium, we would infer from Figs. \[fig:bath\_all\](a)–(c) that $\beta\approx 0.2\,\mathrm{ns}$ (corresponding to $T\approx 38.2\,$mK), which is about a factor of 3 smaller than the inverse operational temperature of about $\beta^*=0.588\,\mathrm{ns}$ (corresponding to $T\approx 13\,$mK). For model II, we have searched the parameter space $1/8\,\mathrm{GHz}\le \Omega\le 2\,\mathrm{GHz}$ and $1/2\,\mathrm{GHz}\le\lambda\le 2\,\mathrm{GHz}$ for sets of ($\Omega,\lambda$) which would fit the D-Wave results best. These data are shown in Figs. \[fig:bath\_all\](a)–(c) as open circles ($\Omega=0.125\,\mathrm{GHz},\lambda=1\,\mathrm{GHz}$) and open squares ($\Omega=0.125\,\mathrm{GHz},\lambda=1.5\,\mathrm{GHz}$). For cases (b) and (c) (see Table \[tab:d\_wave\_data\]), the former fit remarkably well to the D-Wave data. However, we have not found a set ($\Omega,\lambda$) which fits all D-Wave data very well. Summary {#sec:summary} ======= We simulated the full system of three SQUIDs resembling two qubits and one tunable coupler element as used in the quantum annealer built by D-Wave Systems Inc. and found that the higher energy levels as well as the presence of the coupler element have observable effects on the annealing process which however do not have a strong influence on the final success probability compared to the ideal qubit model. In contrast to the investigation of the influence of the higher levels and resonators present in current systems for gate-based quantum computing [@willsch17], we found that, apart from a few exceptions, the effects are small for the cases of quantum annealing examined. Furthermore, we investigated in detail the derivation of the qubit representation to obtain expressions for $\varphi_i^x$ and $J$ that led to satisfying results (see Figs. \[fig:J\]–\[fig:gap\_probs\]). The simulation results are in good agreement with effects encountered in this derivation such as the change in the annealing scheme depending on the choice of $\varphi_{\mathrm{J},0}^x$, which is also supported by findings in experiments [@harris09]. Simulation data for the SQUID model Eq. (\[eq:Htotal\]) and the corresponding two-level approximation Eq. (\[eq:H\_2l\_model\]) for an annealing time $t_a=5\,\mathrm{ns}$ show remarkably good agreement with the D-Wave data obtained with an annealing time $t_a=20\,\mu\mathrm{s}$, also in those cases for which the success probability is substantially less than one. Although this agreement might be accidental, it suggests that the dynamics of the D-Wave quantum annealer are more complicated than what can be described by a closed-system model of the SQUIDs including higher levels and the tunable coupler. Therefore, as a first step, we have studied quantum annealing in the presence of a heat bath. We solve the TDSE of the two-qubit system (Eq. (\[eq:H\_2l\_model\])) plus bath (Eq. (\[eq:H\_bath\])) for three cases with qualitatively different energy spectra of the two-qubit system (see Fig. \[fig:spec\_all\]). Comparing D-Wave and simulation results for an annealing time of $1\,\mu\mathrm{s}$, we found that the simulation data for the success probabilities of the two-qubit systems quite nicely agree with the corresponding thermal equilibrium values but also that these probabilities are significantly larger than those obtained with the D-Wave annealer. We have not found a common set of parameters $(\beta,K,\lambda)$ of the two-qubit-bath model that reproduces the D-Wave results for the three different cases considered. Modeling the environment as a collection of non-interacting two-level defects (see Eq. (\[eq:H\_bathII\])) was found to yield a much more appropriate description of the D-Wave data. Although we could readily find values of the bath parameters $\Omega$ and $\lambda$ for which the solution of the TDSE yields results that are in excellent agreement with D-Wave data for two of the three different cases considered, we could not find a similar level of agreement with these data for all three cases if we impose the elementary requirement that the bath parameters $\Omega$ and $\lambda$ do not depend on the $J$ and $h$ that define the problem Hamiltonian. Unlike in the case of a time-independent problem, the exchange of energy between the two qubits in the time-dependent (annealing) field and the bath of two-level systems seems to be an intricate process which, according to our simulation data, depends on the model parameters in a complicated manner. We leave a detailed study of this challenging problem for future research. We would like to thank Mauricio Reis of D-Wave Systems Inc. for providing us with technical information and Seiji Miyashita for helpful discussions. Access and compute time on the D-Wave machine located at the headquarters of D-Wave Systems Inc. in Burnaby (Canada) were provided by D-Wave Systems Inc. The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V. for funding this project by providing computing time on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre (JSC) and the computing time granted through JARA on the supercomputer JURECA at Forschungszentrum Jülich. D.W. was supported by the Initiative and Networking Fund of the Helmholtz Association through the Strategic Future Field of Research project “Scalable solid state quantum computing” (ZT-0013). Data {#app:data} ==== Table \[tab:cases\] contains a list with the parameters $h_i$ and $J$, the minimal energy gap $\delta E$, and the success probabilities for the problems used to generate Fig. \[fig:gap\_probs\]. Numerical Solution of the TDSE {#app:technical_description} ============================== The numerical solution of the TDSE for a pure state of $N_\mathrm{B}+2$ two-level systems requires computational resources (memory and CPU time) proportional to $2^{N_\mathrm{B}+2}$. For a brute force calculation of thermal expectation values $\mathbf{Tr}( e^{-\beta \mathcal{H}}{\cal A}(t))/\mathbf{Tr} (e^{-\beta \mathcal{H}})$ this number changes to $2^{N_\mathrm{B}+2}\times 2^{N_\mathrm{B}+2}$. Fortunately, this increase in cost can be avoided by making use of random-state technology, reducing the cost to that of solving the TDSE for one pure state [@HAMS00]. If $|\Phi\rangle$ is a pure state, picked uniformly from the $D=2^{{N_\mathrm{B}+2}}$-dimensional unit hypersphere, one can show that for Hermitian matrices $X$ [@HAMS00], $$\mathbf{Tr} (X) \approx D\langle\Phi|X|\Phi\rangle , \label{s46b}$$ and that the statistical errors resulting from approximating $\mathbf{Tr} (X) $ by $D\langle\Phi| X |\Phi\rangle$ are small if $D$ is large [@HAMS00]. For large baths, this property of the random pure state renders the problem amenable to numerical simulation. In the case at hand, we proceed as follows. First, we generate a thermal random state of the bath system, meaning that we compute the pure state $$\begin{aligned} |\Phi(\beta)\rangle&=& \frac{e^{-\beta H_\mathrm{B}/2}|\Phi\rangle}{ \langle\Phi|e^{-\beta H_\mathrm{B}}|\Phi\rangle^{1/2} } , \label{B2}\end{aligned}$$ where $\beta$ denotes the inverse temperature. For any bath observable ${\cal A}(t)$ we have [@HAMS00] $$\langle {\cal A}(t)\rangle=\frac{\mathbf{Tr} (e^{-\beta H_\mathrm{B}}{\cal A}(t))}{\mathbf{Tr} (e^{-\beta H_\mathrm{B}})} \approx \langle\Phi(\beta)| {\cal A}(t)|\Phi(\beta)\rangle . \label{B3}$$ The initial state of the whole system is then a product state of the ground state of the two qubits at $s=0$ and the thermal pure state $|\Phi(\beta)\rangle$, i.e., $$\begin{aligned} |\Psi(t=0)\rangle=\ket{++}\otimes|\Phi(\beta)\rangle . \label{s4k1}\end{aligned}$$ Since the Hamiltonian Eq. (\[s40\]) depends explicitly on time, we can only solve Eq. (\[eq:tdse\]) numerically by time stepping. For this purpose, we use a Suzuki-Trotter product-formula-based algorithm [@RAED06]. This algorithm employs a decomposition in terms of unitary matrices and is unconditionally stable by construction. All our simulations for the two qubits coupled to a heat bath were carried out with the massively parallel quantum-spin dynamics simulator (in-house software), which is based on the same computational kernel as the massively parallel quantum computer simulator [@deraedt18].

ArXiv

--- abstract: 'The diffusion equation is extended by including spatial-temporal memory in such a manner that the conservation of the concentration is maintained. The additional memory term gives rise to the formation of non-trivial stationary solutions. The steady state pattern in an infinite domain is driven by a competition between conventional particle current and a feedback current. We give a general criteria for the existence of a non-trivial stationary state. The applicability of the model is tested in case of a strongly localized, time independent memory kernel. The resulting evolution equation is exactly solvable in arbitrary dimensions and the analytical solutions are compared with numerical simulations. When the memory term offers an spatially decaying behavior, we find also the exact stationary solution in form of a screened potential.' author: - 'Steffen Trimper, Knud Zabrocki' title: Memory Driven Pattern Formation --- Introduction ============ There is an increasing effort in including feedback-couplings into the underlying evolution equations. Such memory-controlled effects should be a further unifying feature of complex physical [@6a; @6b] as well as biological systems [@6c] far from equilibrium. Whereas most of the papers are addressed to a purely time dependent but homogeneous memory, the present one gives an extension to spatially dependent processes. Such systems, however without memory effects, are widely discussed in physics [@kas] and biology [@mur]. Based on our recent studies of several evolution models [@bst; @st1; @zab; @zab1], where such non-Markovian memory effects had been enclosed, and on the analysis of chemical reactions [@trizab], we focus now on the diffusive behavior with memory couplings. These effects have an essential influence on the long time behavior, but they have also an impact on the dynamics in an intermediate time regime. As the consequence of the interplay between conventional diffusive transport and feedback we demonstrate the occurrence of non-trivial stationary solutions forming patterns.\ Our analysis should be grouped into the continuing interest in feedback couplings. A memory dominated behavior is well established in analyzing the freezing processes in under-cooled liquids [@l; @goe], where the underlying mathematical representation is based on a projector formalism proposed by Mori [@mori]. As the result of the projection procedure the irrelevant degrees of freedom, fluctuating on short time and short length scales, contribute to the time evolution of the relevant degrees of freedom as well as by instantaneous and by delay-controlled terms. Since the projection formalism is not restricted to selected systems, the modification of the evolution equation due to memory effects is rather general. Here the relevant quantity may be the concentration of certain species or the probability density for a particle. The crucial influence of memory effects can be illustrated by considering a single particle moving in a disordered environment. Due to the strong disorder each member of an equally prepared ensemble makes experience by its own surrounding, which is even modified by the random walk of the single particle itself. Such a behavior can be described by a non-linear Fokker-Planck equation with an additional memory term as had been demonstrated in [@ss]. The analytical results, based on a one-loop renormalization group analysis, is in accordance with numerical simulations [@bst] and further analytical studies [@st1]. Whereas in that approach the memory effects are believed to be originated by the inherent non-linear interaction of the many particle system themselves, i. e. the time scale of the memory is determined by the relevant variable itself, there is a broad class of models which are subjected to external delay effects [@o; @g; @fe]; for a survey and applications in biology, see [@mathbio]. Even in traffic models such memory effects has been discussed [@he].\ Recently [@fk] memory effects in correlated anisotropic diffusion are studied in nanoporous crystalline solids. Likewise the effects of transport memory are discussed within the Fisher‘s equation [@ah], also applicable for bacterial population dynamics [@kk]. There appear a non-linear damping and traveling wave solutions [@abk]. The transport with memory, depending on the survivability of a population, is analyzed in [@h].\ Regarding the large variety of systems with feedback couplings it seems to be worth to study simple models, which still include the dynamical features of evolution models as conservation of the relevant quantity $p(\vec r, t)$ and moreover, a time-delayed coupling. In the present paper we emphasize in an analytical solvable model with spatio-temporal feedback coupling. The generic behavior of the system may be changed by those additional delay effects. In particular, we discuss the transport behavior which is realized after a sufficient accumulation time and after cumulating particles within a spatial region. We show that the processes are affected by additional spatial and temporal correlations leading to non-trivial steady state solutions. Model ===== In case the transport is realized after a spatial-temporal accumulation process the time evolution of the probability could also depend on the history of the sample to which it belongs, i.e. the changing rate of the probability should be influenced by the changing rate in the past. Thus the evolution of the probability $p(\vec r, t)$ has to be supplemented by memory terms. Such a term models, for example the way on which a seed probability at a certain time $t'$ had been accumulated by a delayed transport mechanism, originated by the surrounded environment of the particle. In general, the changing rate of $p$ at time $t$ is also determined by the accumulation rate at a former time $t' < t$. In between, i.e. within the interval $\tau = t - t'$, the particles are enriched while changing the probability at $t'$. Regardless that process the available amount of particles at time $t$ is governed by an instantaneous transport term as well as by the changing rate at former times $t'$. Consequently the conventional diffusion equation should be modified according to [@spie] $$\partial _{t}p(\vec r,t) = \mathcal{M}(\vec r,t; p, \nabla p) +\int\limits_{0}^{t}dt'\int\limits_{-\infty }^{\infty }d^dr^{\prime} \mathcal{K}(\vec r- \vec r^{\prime},t-t'; p, \nabla p)\mathcal{L}(\vec r^{\prime}, t'; p,\nabla p)\,. \label{2}$$ This equation is of convolution type. Whereas the operator $\mathcal{M}$ characterizes the instantaneous and local processes, both operators $\mathcal{K}$ and $\mathcal{L}$ are responsible for the delayed processes. In general the operators $\mathcal{M}, \mathcal{K}$ and $\mathcal{L}$ may be also non-linear in $p(\vec r, t)$ and $\nabla p(\vec r,t)$. They have to be specified according to the physical situation in mind. In particular we show that the operators are restricted when $p(\vec r,t)$ is assumed to be conserved. Conservation ------------ In this section we specify the model, defined by Eq. (\[2\]), under the assumption that the basic quantity $p(\vec r, t)$ is conserved. This condition is fulfilled by $$\frac{d\,P(t)}{dt} = \frac{d}{dt} \int_{-\infty}^{+\infty} d^{d}r p(\vec r, t) = 0\quad . \label{ev1}$$ To preserve $p$ the instantaneous term $\mathcal{M}$ has to be related to a current, e.g. $\mathcal{M}\propto \nabla \cdot \vec j $. Choosing natural boundary conditions, i.e. $p=0$ at the boundary, we find after Laplace transformation $P(t) \to P(z)$ $$z P(z) - P_0 = \hat{K}(z) \hat{L}(z) \quad\mbox{with}\quad \hat{K}(z) = \int d^dr\, \mathcal{K}(\vec r, z),\qquad \hat{L}(z) = \int d^dr\, \mathcal{L}(\vec r, z)\,, \label{ev2}$$ with $P_0 = P(t=0)$. For an arbitrary polynomial kernel $\hat{K}$ the conservation law is in general not fulfilled provided the operator $\mathcal{L}$ is simply defined by $\mathcal{L} \equiv - \partial_t p(\vec r, t)$ (the minus sign is only for convention). Making this ansatz we conclude from Eq. (\[ev2\]) $$[z P(z) - P_0][\hat{K}(z) + 1] = 0$$ and consequently the conservation law is guaranteed. Physically, the assumption for $\mathcal{L}$ means, that we take into account a coupling of the rates, e.g. the evolution at the observation time $t$ is directly coupled to the changing rate at $t' < t$. Processes in the past will be permanently “reevaluated” at present time. In doing so the memory kernel gives rise to a coupling of the time scales. In the vicinity of the upper limit $t^{\prime} \simeq t$ the memory term reads $\mathcal{K}[\vec r, 0;\,p(\vec r, 0)]\partial_t p(\vec r, t)$, i.e. a instantaneous change at the observation time $t$ is coupled to the value of $p$ at the initial time $t = 0$. Therefore, the very past is related to the instantaneous value of $p(\vec r, t)$. In the opposite case, at the lower limit $t^{\prime} \simeq 0$, the change of the quantity $p(\vec r, t)$ near to the initial value $\partial_{t^{\prime}}\,p(\vec r, t^{\prime} = 0)$ is directly coupled to the instantaneous value $p(\vec r, t)$ via the kernel. In such a manner the memory part represents a weighted coupling of the behavior at the initial time and the observation time. Due to that coupling of the rates the long-time behavior of the system will be modified. One reason for that could be that the moving particle is embedded into an environment of all the other particles of the system. Due to the mutual interaction, the particle is released after a certain waiting time $t - t'$. Especially in sufficiently complex diffusion-reaction systems the feedback and memory effects should be relevant. In such system additional degrees of freedom like in flexible macromolecules in melts or in concentrated solutions [@dg], nematic elastomer [@af] or in biology [@mathbio]. In that context one is interested in the description of effluent reprocessing plants in systems with closed water circulation. A special ecosystem of aerobic and anaerobic microorganisms is evolved in the clarifiers of such systems due to natural immigration or due to additional allowance. The living conditions of the microorganisms are mutually associated via the exchange of intermediate catabolic products. Each change of the concentration of one species will be stored in the food chain and effects the evolution of these species at a later time. Furthermore, the partial mixing in the clarifiers by convection and diffusion processes enlarges the effects over the whole system, so that the memory integral introduced in Eq. (\[2\]) includes both, the time and the spatial coordinates. This special example may be extended also to other complex biological, chemical or engineering problems with various hidden degrees of freedom, which are able to influence the evolution of a selected component significantly, for instance by biological interaction with other species via the food chain or via biological competition. Such effects which are partially observable, could contribute to the memory term. General properties ------------------ Starting from Eq. (\[2\]) we can make some general statements for an arbitrary kernel $\mathcal{K}(\vec r, t; p, \nabla p)$ and the specialized term $\mathcal{L}(\vec r, t; p, \nabla p) = - \partial_t p(\vec r, t)$. As stressed before the last condition guarantees the conservation of $p(\vec r, t)$. After Fourier transformation with respect to the spatial coordinate and Laplace transformation with respect to the time, we get from Eq. (\[2\]) $$p(\vec k, z) = \frac{p_0(\vec k)}{z + \hat{D}(\vec{k}, z)\,k^2}\quad\text{with}\quad \hat{D}(\vec{k}, z) = \frac{D}{1 + \mathcal{K}({\vec{k}}, z)}\,, \label{ev10}$$ where $\mathcal{K}(\vec k, z)$ is the Fourier-Laplace transformed kernel. Eq. (\[ev10\]) describes diffusion with a modified diffusion parameter $D \to \hat{D}(\vec r, t)$. Assuming a regular behavior of the kernel let us make the ansatz $$\mathcal{K}(\vec k, z) = \frac{b(\vec k)}{z} + \Lambda (\vec k, z)\quad\text{with} \quad\lim_{z \to 0}\Lambda ({\vec k}, z)= \text{finite}\quad . \label{ev11}$$ Provided the kernel reveals a finite stationary value $\lim\limits_{t \to \infty} \mathcal{K}(\vec k, t) \equiv b(\vec k)$, then $p(\vec r, t)$ yields a stationary solution, too. Inserting Eq. (\[ev11\]) in Eq. (\[ev10\]) we obtain $$\begin{aligned} p(\vec k, z) &= \frac{p_s(\vec k)}{z} + \varphi (\vec k, z)\nonumber\\ \text{with} \quad p_s(\vec k) &= \frac{p_0(\vec k) b(\vec k)}{b(\vec k) + D k^2}\quad\text{and}\quad \varphi (\vec k, z) = \frac{p_0(\vec k)\,Dk^2\,[\,1 + \Lambda (\vec k, z)]}{[\,b(\vec k) + D k^2\,] [\,z(1 + \Lambda (\vec k, z)) + b(\vec k) + Dk^2]}\,. \label{ev12}\end{aligned}$$ Summarizing the results we conclude that the model, following Eq. (\[2\]) for the conserved quantity $p(\vec r, t)$, gives rise to a non-trivial stationary solution $g(\vec k$), or equivalent $g(\vec r)$, if the Laplace transformed kernel $\mathcal{K}(\vec k, z)$ satisfies the condition $$\lim_{z \to 0} z \mathcal{K}(\vec k, z) \neq 0\,. \label{ev14a}$$ This result is a generalization of a previous one derived for a homogeneous system [@trizab]. Diffusion with time independent memory kernel ============================================== In the following let us regard the evolution equation $$\partial_t p(\vec r, t) = D\nabla^2p(\vec r, t) - \int_{-\infty}^{\infty} d^dx' \int_0^t dt'\, K(\vec r - \vec r', t - t')\, \partial_{t'}p(\vec r',t')\,. \label{ev3}$$ As discussed in the previous section the conservation of the quantity $p(\vec r,t)$ is guaranteed for an arbitrary kernel $K$. Different to a conventional evolution equation, in the present one $p(\vec r, t)$ is related to $p(\vec r, t')$ with $0< t'< t$. Moreover, Eq. (\[ev3\]) offers also a coupling between $\partial_t p(\vec r, t)$ and $\partial_{t'}p(\vec r, t')$. The mixing of time scales leads, above all, to a substantial modification in the long time limit. Spatially localized memory -------------------------- As the simplest example we analyze a strictly spatially local, but time independent kernel $$K(\vec r, t) = \mu\, \delta (\vec r)\, , \label{ev3a}$$ where $\mu > 0$ is the strength of the memory. Note that in our approach spatial and temporal variables are decoupled. In case of a coupling between space and time variable one may observe different effects discussed elsewhere [@zts]. Inserting this kernel in Eq. (\[ev3\] we get $$\partial_t p(\vec r, t) = D\nabla^2p(\vec r, t) - \mu [p(\vec r, t) - p_0(\vec r)] \quad\mbox{with}\quad p_0(\vec r) \equiv p(\vec r, t =0)\quad . \label{ev4}$$ Such a memory term leads to a very strong coupling of the instantaneous value to the initial value because all previous time steps are weighted equally. There appears a memory induced feedback to the initial distribution as a driving force. Note that without the coupling to the initial value $p_0(\vec r)$, the particle would perform a diffusive motion and the probability density $p(\vec r,t)$ would decay on the time scale $\mu ^{-1}$. Mathematically the inhomogeneous linear equation may be solved analytically for arbitrary initial conditions and in any dimension: $$p(\vec r,t)=\text{e}^{-\mu\, t} \int_{-\infty}^{\infty} G (\vec r - \vec r\,', t) p_0 (\vec r\,')\, d^d\, r' +\mu\, \text{e}^{-\mu\, t} \int_{-\infty}^{\infty}\,\int_{0}^{t} G(\vec r-\vec r\,' ,t-t')\, \text{e}^{\mu\, t'}\, p_0 (\vec r\,' ) dt' \, d^d r' \,. \label{gs}$$ From this solution one can verify that a non-negative initial distribution $p_0(\vec r\,) \geq 0$ leads always to $p(\vec r, t) \geq 0$ provided $\mu > 0$. Eq.  enables us to calculate the second moment $s(t)$. It results $$s(t)\equiv\int\limits_{-\infty}^{\infty} \vec{r}^{\,2} \, p(\vec r,t) \, d^d r =\frac{2\,d\, D\, \left(1-\text{e}^{-\mu\, t}\right)}{\mu}\, \int_{-\infty}^{\infty} p_0 (\vec r\,)\, d^d r + \int_{-\infty}^{\infty}\vec{r}^{\,2} \, p_0 (\vec r\,)\, d^d r \,.$$ Notice that for $\mu \to 0$ the last equation shows the conventional diffusive behavior.\ Let us illustrate the behavior in the one dimensional case with the initial condition $p_0(x) = p_0 \delta (x)$. The exact solution reads $$\begin{aligned} p(x,t) &= \frac{p_0}{\sqrt{4\pi Dt}}e^{-(\mu t +\frac{x^2}{4Dt})} + \frac{p_0 \kappa }{4}\left\{e^{\kappa x}\left[\rm erf\left(\frac{x}{\sqrt{4Dt}}+\sqrt{\mu t}\right)-\rm sgn(x)\right]\right\} \nonumber\\ &\quad + \frac{p_0 \kappa }{4}\left\{e^{-\kappa x}\left[\rm erf\left(-\frac{x}{\sqrt{4Dt}}+\sqrt{\mu t}\right) - \rm sgn(-x)\right]\right\}\,, \label{ev5}\end{aligned}$$ where $\rm erf (x)$ is the error function and $\kappa = \sqrt{\mu /D}$. The first part of Eq. (\[ev5\]) is the solution of the homogeneous equation showing a temporal decay with the time constant $\mu ^{-1}$. For different strength of the feedback coupling $\mu = \,0.5,\,1,\,3$ the solution $p(x,t)$ is presented in Fig. (\[Fig.1\]). The stronger the feedback strength $\mu$ the more pronounced are the deviations from the pure diffusive behavior. In particular, the system develops a non-trivial stationary solution whenever $\mu > 0$. Such a stationary solution is due the permanent coupling to the initial distribution. We find $$\lim_{t \to \infty} p(x,t) \equiv p_s(x) = \frac{p_0 \kappa }{2} e^{-\kappa \mid x \mid}\,. \label{ev6}$$ where $p_s(x)$ is also included into Fig. (\[Fig.1\]). Notice that in the limiting case $\kappa \to \infty$ the stationary solution tends to the initial distribution. Obviously our feedback offers a kind of shape memory. The stronger the memory is the stronger is the recognition of the initial state.\ The result can be generalized to arbitrary dimensions $d$. For the stationary solution $p_s(\vec r)$ we obtain with $r = \mid \vec r \mid$ $$p_s(r) = \frac{p_0}{(2\pi )^{d/2}} \kappa ^{\frac{d+2}{2}} r^{-\frac{d-2}{2}}K_{\frac{d-2}{2}}(\kappa r)\,. \label{ev7}$$ The modified Bessel function $K_{\nu}(x)$ can be expressed by standard functions for $d = 1, 3$ [@as]. In the two dimensional case the modified Bessel function offers a logarithmic behavior. The evolution equation for $p(\vec r, t)$ is integrable with $$p({\vec r} ,t) = \frac{p_0}{(4\pi D t)^{d/2}} e^{-\left(\mu t + \frac{\vec r^2}{4Dt}\right)} + \mu p_0 e^{-\mu t}\int _0^t dt' e^{\mu t'}\frac{1}{[4\pi D(t - t')]^{d/2}} e^{-\frac{\vec r^2}{4D(t-t')}}\,. \label{ev8}$$ The last integral can be calculated exactly for $d=1$ with the result given in Eq. (\[ev5\]). Asymptotically we obtain $$p({\vec r} ,t) \approx \frac{p_0}{(4\pi D t)^{d/2}} e^{-\left(\mu t + \frac{\vec r^2}{4Dt}\right)} + \frac{p_0}{(2\pi )^{d/2}} \kappa ^{\frac{d+2}{2}} r^{-\frac{d-2}{2}}K_{\frac{d-2}{2}}(\kappa \mid \vec r \mid)\,. \label{ev9}$$ The analytical results are confirmed by numerical simulations. Gaussian and exponential initial conditions ------------------------------------------- For an arbitrary initial condition $p_0(\vec r)$ the stationary solution can be directly calculated by solving equation $\nabla ^2p_s(\vec r) = \kappa ^2 [\,p_s(\vec r)-p_0(\vec r)\,]$ with $\kappa ^2 = \mu /D$. It results in $$p_s(\vec r) = \frac{\kappa^{(d+2)/2}}{(2\pi)^{d/2}} \int d^dx'\frac{p_0(\vec r')}{\mid \vec r - \vec r'\mid^{\frac{d-2}{2}}}\, K_{\frac{d-2}{2}}(\kappa \mid \vec r - \vec r'\mid)\quad . \label{ev9a}$$ In this subsection we study different realizations of the initial conditions. In case of a Gaussian initial distribution $p_0(\vec r) = p_0 e^{-\lambda x^2}$ the results for different strength of memory $\mu $ for both $p(x,t)$ as well as $p_s(x)$ are shown in Fig. (\[Fig.2\]). When the initial condition is an exponentially decaying function $p_0(\vec r) = p_0 e^{-\lambda \mid \vec r \mid}$, we find solutions presented in Fig. (\[Fig.3\]). Summarizing this section we have shown, that a simple feedback coupling lead to stable patterns which do not appear under a conventional diffusion term. Fading diffusion ================ Now we regard as a further realization of Eq. (\[2\]) with a conserved $p(\vec r, t)$ a spatially decaying memory term by choosing the kernel in the form $$\mathcal{K}(\vec r, t) = \mu \exp(-\lambda \mid \vec r \mid)\,. \label{exp}$$ Performing Laplace transformation we get $$\mathcal{K}(\vec k, z) = \frac{A \mu \lambda }{z \,(\lambda^2 + k^2)^{(d+1)/2}}\quad\mbox{with}\quad A(d)=2^d(\pi )^{(d-1)/2}\Gamma \left(\frac{d+1}{2}\right)$$ Inserting that result into Eq. (\[ev12\]) it results a non-trivial stationary solution, compare also Eq. (\[ev12\]), $$p_s(\vec k)= \frac{p_0(\vec k) A \mu \lambda }{A \mu \lambda + D k^2 [\lambda ^2 + k^2]^{\frac{d+1}{2}}}\,. \label{exst}$$ In the hydrodynamic limit $\vec k \to 0$ one finds the Fourier transformed stationary solution $$p_s(\vec r) = \frac{\kappa ^{d/2+1}}{(2\pi )^{d/2}}\int d^dx'\frac{p_0(\vec r')}{\mid \vec r -\vec r'\mid^{\frac{d-2}{2}}} K_{\frac{d-2}{2}}(\kappa \mid \vec r - \vec r'\mid)\quad\mbox{with}\quad \kappa =\sqrt{\frac{A\mu}{D \lambda ^d}}\,. \label{exst1}$$ Using the asymptotic expansion for the modified Bessel functions [@as], one concludes that the stationary solution offers a screened Coulomb potential decaying on a length scale $\kappa ^{-1}$. Note that the screening length is determined by the memory strength $\mu $ as well as by the characteristic length $\lambda ^{-1}$ of the spatial memory. The screening length increases when the memory strength $\mu $ decreases. Let us stress that the non-trivial steady state is originated by the competition between diffusion and memory and it is not driven by spatial boundary conditions. In that case there appears an additional length scale, the size of the sample $L$, which gives rise to further effects not discussed now. Conclusions =========== In this paper we have extended the conventional diffusive transport by including non-Markovian memory terms within the frame of evolution equations. The additional terms are chosen in such a manner that the relevant variable $p(\vec r, t)$ is normalized. By this requirement the form of the memory term is restricted to a class where the changing rate of $p$ at the observation time $t$ is coupled to the changing rate of $p$ at a former time $t'$. Insofar the delay effect offers a long time memory due to $0 \leq t' \leq t$. Further the model exhibits also a long range memory because the kernel $\mathcal{K}$ depends on $\vec r - \vec r^{~\prime}$. Here we have considered several realizations of the memory kernel with different behavior. As a new result there occur non-trivial stationary solutions forming patterns. Such patterns are absent for conventional diffusion in an infinite domain without boundary conditions. The reason for such a new behavior is by means of the above mentioned coupling of the rates. This time accumulation is further accompanied by an additional spatial accumulation, the effect of which is comparable to the effect a long-range interaction forces and consequently the results are basically independent of the spatial dimensions in according to scaling arguments. These many-body effects are shown to change the asymptotic behavior drastically. Due to the feedback-coupling of a particle to its environment, a subsequent particle, undergoing a diffusive motion, gains information from a modified environment. As the result there appears patterns controlled by the memory strength. This work is supported by the DFG (SFB 418). We are indebted to Peter Hänggi and Gunter Schütz for valuable discussions. [90]{} L.S.Tsimring, A.Pikovsky, Phys.Rev.Lett. [**87**]{}, 250602 (2001). C.Masoller, Phys.Rev.Lett. [**88**]{}, 034102 (2002); Phys. Rev. Lett. [**90**]{}, 020601 (2003). M. Freeman, Nature [**408**]{}, 313 (2000). K. Kassner [*Pattern formation in diffusion-limited crystal growth*]{}, (World Scientific, Singapore, 1996). J. D. Murray [*Mathematical Biology, Part II*]{} (Springer, Berlin, 2003). B. M. Schulz, S. Trimper, Phys. Lett. A [**256**]{}, 266 (1999); B. M. Schulz, S. Trimper, and M. Schulz, Eur. Phys. J [**B15**]{}, 499 (2000); B. M. Schulz, M. Schulz, and S. Trimper, Phys. Rev. E [**66**]{}, 031106 (2002). M. Schulz and S. Trimper, Phys. Rev. E [**64**]{}, 233101 (2001). S. Trimper, K. Zabrocki, M. Schulz, Phys. Rev. E [**65**]{}, 056106 (2002). S. Trimper, K. Zabrocki, M. Schulz, Phys. Rev. E [**66**]{}, 026114 (2002). S. Trimper and K. Zabrocki, Phys. Lett. [**A**]{}, 2205 (2004). E. Leutheusser, Phys.Rev.A [**29**]{}, 2765 (1984). W. Götze in [*Liquids, Freezing and the Glass Transition*]{} edited by Hansen et al. (North Holland, Amsterdam, 1991); for a recent survey, see W.Götze, J.Phys.: Condens. Matter, [**11**]{}, A1 (1999). H. Mori, Prog.Theor.Phys. [**34**]{}, 399 (1965). M. Schulz, S. Stepanow, Phys. Rev. B [**59**]{}, 13528 (1999). T. Ohira, T. Yamane, Phys. Rev. E [**61**]{}, 1247 (2000). R. Gerami, Phys. Rev. E [**65**]{}, 036102 (2002). S. Fedotov and Y. Okuda, Phys. Rev. E [**66**]{}, 021113 (2002). J. D. Murray [*Mathematical Biology, Part I*]{} (Springer, Berlin, 2000). M. Treiber, D. Helbing, Phys. Rev. E [**68**]{}, 046119 (2003). S. Fritzsche, J. Kärger, Europhys. Lett. [**63**]{}, 465 (2003). D. ben-Avraham and S. Havlin [*Diffusion and Reactions in Fractals and Disordered Systems*]{}, (Cambridge University Press, Cambridge, UK, 2000). V. M. Kenkre, M. N. Kuperman, Phys. Rev. E [**67**]{}, 051921 (2003). G. Abramson, A. R. Bishop, and V. M. Kenkre, Phys. Rev. E [**64**]{}, 066615 (2001). S. Harris, J. Phys. A: Math. Gen. [**36**]{}, 8291 (2003). S. Trimper and K. Zabrocki, Proccedings Spie [*Fluctuations and Noise* ]{} (2004). P. G. de Gennes, J. Chem. Phys. [**76**]{}, 3316 (1982). X. Xing and L. Radzihovsky, Europhys. Lett. [**61**]{}, 769 (2003). M. Schulz, S. Trimper and K. Zabrocki, in preparation. M. Abramowitz and I. A. Stegun, [*Handbook of Mathematical Functions*]{}, (Verlag Harri Deutsch, Frankfurt/Main, 1984). \[\]\[\]\[0.6\][$p$]{} \[\]\[\]\[0.6\][$p_s$]{} [ ]{}\ [ ]{}\ \[\]\[\]\[0.6\][$p$]{} \[\]\[\]\[0.6\][$p_s$]{} [ ]{}\ [ ]{}\ \[\]\[\]\[0.6\][$p$]{} \[\]\[\]\[0.6\][$p_s$]{} [ ]{}\ [ ]{}\

ArXiv

--- abstract: 'Linear response calculations based on the time-dependent density-functional theory are presented. Especially, we report results of the finite amplitude method which we have recently proposed as an alternative and feasible approach to the (quasiparticle-)random-phase approximation. Calculated properties of the giant resonances and low-energy $E1$ modes are discussed. We found a universal linear correlation between the low-energy $E1$ strength and the neutron skin thickness.' author: - Takashi Nakatsukasa - Tsunenori Inakura - Paolo Avogadro - Shuichiro Ebata - Koichi Sato - Kazuhiro Yabana bibliography: - 'myself.bib' - 'nuclear\_physics.bib' - 'current.bib' - 'chemical\_physics.bib' title: '**Linear-response calculation in the time-dependent density functional theory**' --- [ address=[RIKEN Nishina Center, Wako-shi 351-0198, Japan]{} ,altaddress=[Center for Computational Sciences, University of Tsukuba, Tsukuba 305-8571, Japan]{} ]{} [ address=[RIKEN Nishina Center, Wako-shi 351-0198, Japan]{} ]{} [ address=[Departimento di Fisica, Università degli Studi di Milano, via Celoria 16, 20133 Milan, Italy]{} ,altaddress=[RIKEN Nishina Center, Wako-shi 351-0198, Japan]{} ]{} [ address=[Center for Nuclear Study, University of Tokyo, Bunkyo-ku, 113-0033, Japan]{} ,altaddress=[RIKEN Nishina Center, Wako-shi 351-0198, Japan]{} ]{} [ address=[RIKEN Nishina Center, Wako-shi 351-0198, Japan]{} ]{} [ address=[Center for Computational Sciences, University of Tsukuba, Tsukuba 305-8571, Japan]{} ,altaddress=[RIKEN Nishina Center, Wako-shi 351-0198, Japan]{} ]{} Introduction ============ Nuclei are created in stars via nuclear reactions where the temperatures are extremely high. Understanding of astrophysical phenomena generally requires many subfields of physics. Among them, nuclear physics plays key roles in the generation of elements, the evolution of the stars, the energy production in the universe, etc. The lightest elements up to helium were produced in the Big Bang. The other heavier elements are generated by many kinds of nuclear reactions in stars. Especially, in stellar explosions, there are thousands of reactions supposed to take place producing a variety of radioactive isotopes. However, to date, their reaction rates have not been hardly determined experimentally. To overcome these experimental difficulties, reliable theoretical information is necessary. Traditional ab-initio methods start from a nucleon-nucleon potential which describes nucleon-nucleon scattering data. However, since the nuclear systems are strongly correlated because of a repulsive core in the potential, their description requires highly sophisticated many-body methods, such as the quantum monte carlo method [@PW01]. To describe the nucleus in a quantitative way, they must employ an additional three-body force. These ab-initio methods are so involved that, even at present, their investigations have been limited to very light nuclei and to the homogeneous nuclear matter [@HP00]. Under these circumstances, it is highly demanded to establish a universal theoretical approach which is able to describe properties of all species of nucleus. The nuclear density functional theory [@BHR03] is the most promising candidate among many nuclear models. It was referred to as the self-consistent mean-field model, but its concept is analogous to the density functional theory for many-electron systems [@PY89; @DG90]. The functionals have typically about ten parameters which are adjusted by extensive fits to nuclear structure data. The most prominent feature of the model is that a single energy functional enables us to quantitatively describe almost all nuclei in the nuclear chart and infinite nuclear/neutron matter as well. The nuclear density functional models have been extensively used since 1970s [@VB72]. In the beginning, there were several restrictions related to symmetries of the wave functions, which limited applications of the model. In 1990s, significant computational advances together with vast amount of new spectroscopic data obtained with large $gamma$-ray arrays changed the situation. Using the cranking prescription, the density functional models were becoming a standard tool to study rotational bands in heavy nuclei [@AFN90]. In particular, the models were very successful in studies of superdeformed rotational bands at high spin in $A=150$ [@AFN90; @NMMN95] and $A=190$ regions [@NMMS96]. It should be emphasized that the model parameters were never adjusted to these bands at all. In 2000s, systematic calculations of nuclear ground-state properties were performed, to predict properties of nuclei far from the stability line [@LPT03]. At the same time, experimental developments greatly increased our knowledge on radioactive isotopes, then, had an impact on the energy density functionals. The nuclear density functional methods has now reached a point where one need to introduce “correlations beyond the Kohn-Sham scheme” to further improve the quality of the description. The density functional theory is designed for the description of ground-state properties. Its extension to a time-dependent density functional theory is formally straightforward, which is also analogous to the time-dependent mean-field theory. The density matrix $\rho(t)$ obeys the following equation, $$i\ \hbar\frac{\partial}{\partial t} \rho(t) = \left[ h(t), \rho(t) \right] ,$$ where $h(t)$ is the Kohn-Sham (mean-field) Hamiltonian which is a functional of $\rho(t)$. In the case of open-shell superfluid nuclei with pairing, the density $\rho(t)$ and the Hamiltonian $h(t)$ should be generalized to those with double size of matrices, $R(t)$ and $H(t)$, respectively [@BR86]. Namely, $R(t)$ contains not only the normal density $\rho(t)$ but also the pairing tensor $\kappa(t)$. Similarly, $H(t)$ has the pair potential $\Delta(t)$ in addition to $h(t)$. It is not yet fully understood how we can compute excitation properties using the time-dependent density-functional theory, except for a few limiting cases. One such case is the low-energy regime of the nuclear dynamics, such as surface vibrations and shape fluctuations, which can be reached by the adiabatic limit of the time-dependent density-functional theory. It is closely related to the microscopic derivation of the Bohr model [@RS80]. Recently, we have performed the numerical calculations of the large-amplitude quadrupole dynamics in various isotopes [@HSNMM10; @HSYNMM11; @YH11; @SH11]. Another important limiting case is the small-amplitude limit which provides us with a powerful tool to study linear response in nuclei. It is known as the random-phase approximation (RPA) or the quasiparticle-random-phase approximation (QRPA) in nuclear physics, which accesses the regime of giant resonances. However, since the calculation of the residual interaction is tedious for the realistic energy functionals, it has been common to ignore some terms in practice and to sacrifice the full self-consistency. To overcome these difficulties and facilitate an implementation of the full self-consistency, we employ two different methodologies; one is based on the real-time method (RTM) [@NY05] and the other is the finite amplitude method (FAM) [@NIY07]. The RTM has an advantage that it does not require the calculation of the complex residual interactions, although it has a limitation in the achieved energy resolution inversely proportional to the time duration $T$. Our recent applications of the RTM are based on the canonical-basis framework [@Eba10] which is able to take into account dynamical pairing effects in nuclei. This is presented in another contribution to this volume [@ENI12]. In contrast, the FAM, which is complementary to the RTM, is a method of calculating the matrix elements of the residual field, $\delta h = \delta h/\delta \rho \cdot \delta \rho$, using the finite difference. This does not require excessive programming but can be done by employing the program of the static density-functional calculation. We have performed systematic and fully self-consistent RPA calculations of photoabsorption cross sections for wide mass region ($A\le 100$), for both spherical and deformed nuclei. In this report, we show results of these symmetry-unrestricted FAM calculations. Finite Amplitude Method ======================= In this section, we recapitulate the FAM. FAM for RPA ----------- First, we discuss the case without the pairing correlations. In this case, the FAM leads to residual fields appearing in the RPA. For more details, readers are referred to Ref. [@NIY07]. The linear-response RPA equation to a weak external field with a fixed frequency, $V_\mathrm{ext}(\omega)$, can be expressed in terms of the forward and backward amplitudes, ${| {X_i(\omega)} \rangle}$ and ${\langle {Y_i(\omega)} |}$. $$\begin{aligned} \omega \, | X_i(\omega) \rangle = \left( h_0 - \epsilon_i \right) | X_i(\omega) \rangle + \hat{P} \left\{ V_\mathrm{ext}(\omega) + \delta h(\omega) \right\} | \phi_i \rangle , \label{RPAeqX}\\ - \omega \, \langle Y_i(\omega) | = \langle Y_i(\omega) | \left( h_0 - \epsilon_i \right) + \langle \phi_i | \left\{ V_\mathrm{ext}(\omega) + \delta h(\omega) \right\} \hat{P} , \label{RPAeqY}\end{aligned}$$ where the subscript $i$ indicates the occupied orbitals ($i=1,2,\cdots,A$) and the operator $\hat{P}$ denotes the projector onto the particle space, $\hat{P} = 1 - \sum_i | \phi_i \rangle \langle \phi_i |$. Usually, the induced residual field $\delta h(\omega)$ is expanded to the first order with respect to $| X_i(\omega)\rangle$ and $| Y_i(\omega)\rangle$. This leads to the well-known matrix form of the linear-response equation and calculation of these matrix elements is most time-consuming in practice. Instead, we utilize the fact that the linearization is numerically achieved for $\delta h(\omega) = h[\rho_0 + \delta\rho(\omega)] - h_0$ within the linear approximation. In order to perform this numerical differentiation in the program, we use a small trick in the calculation of the single-particle Hamiltonian $h[\rho]$. First, we should notice that the $\delta h(\omega)$ depends only on the forward “ket” amplitudes ${| {X_i(\omega)} \rangle}$ and backward “bra” ones ${\langle {Y_i(\omega)} |}$. In other words, it is independent of bras ${\langle {X_i(\omega)} |}$ and kets ${| {Y_i(\omega)} \rangle}$. This is related to the fact that the transition density $\delta\rho(\omega)$ depends only on ${| {X_i(\omega)} \rangle}$ and ${\langle {Y_i(\omega)} |}$. $$\delta\rho(\omega) = \sum_i \left\{ {| {X_i(\omega)} \rangle}{\langle {\phi_i} |} + {| {\phi_i} \rangle}{\langle {Y_i(\omega)} |} \right\} .$$ Then, we can calculate the residual fields in a following manner [@NIY07]: $$\delta h(\omega) = \frac{1}{\eta} \left( h\left[ {\rho}_\eta \right] - h_0 \right) , \label{FAM}$$ where $\eta$ is a small real parameter to realize the linear approximation. $\rho_\eta$ are defined by $${\rho}_\eta \equiv \sum_i \left\{ ({| {\phi_i} \rangle}+\eta{| {X_i(\omega)} \rangle})({\langle {\phi_i} |}+\eta{\langle {Y_i(\omega)} |}) \right\} .$$ Once ${| {X_i(\omega)} \rangle}$ and ${\langle {Y_i(\omega)} |}$ are given, the calculation of $h[\rho_\eta]$ is an easy task. This does not require complicated programming, but only needs a small modification in the calculation of $h[\rho]$. Of course, eventually, we need to solve Eqs. (\[RPAeqX\]) and (\[RPAeqY\]) to determine the forward and backward amplitudes. We use an iterative algorithm to solve this problem. Namely, we start from initial amplitudes ${| {X_i^{(0)}} \rangle}$ and ${\langle {Y_i^{(0)}} |}$, then update them in every iteration, (${| {X_i^{(n)}} \rangle},{\langle {Y_i^{(n)}} |}) \rightarrow ({| {X_i^{(n+1)}} \rangle},{\langle {Y_i^{(n+1)}} |})$, until the convergence. In each step, we calculate $\delta h(\omega)$ using the FAM as Eq. (\[FAM\]). FAM for QRPA ------------ The FAM in the previous section can be extended to superfluid nuclei, namely, to the QRPA with the Bogoliubov extension of the mean fields. For more details, readers are referred to Ref. [@AN11]. A self-consistent solution of static Kohn-Sham-Bogoliubov problems determines the ground-state densities $(\rho_0,\kappa_0)$ and the ground-state Hamiltonians $(h_0,\Delta_0)$. They are given in terms of the quasiparticle wave functions, $(U_\mu, V_\mu)$. Then, following the same argument in the previous section, we can derive equations for the residual fields, $\delta h(\omega)$ and $\delta\Delta(\omega)$ as follows [@AN11]: $$\label{FAM_QRPA} \begin{split} \delta h(\omega) &= \frac{1}{\eta} \left( h[\rho_\eta,\kappa_\eta] - h_0 \right) , \\ \delta \Delta(\omega) &= \frac{1}{\eta} \left( \Delta[\rho_\eta,\kappa_\eta] - \Delta_0 \right) , \end{split}$$ where the density and pairing tensor $(\rho_\eta, \kappa_\eta)$ are defined by $$\begin{aligned} \rho_\eta &=& ( V^* + \eta U X ) (V + \eta U^* Y )^T ,\\ \kappa_\eta &=& ( V^* + \eta U X ) ( U + \eta V^* Y )^T .\end{aligned}$$ Here, the forward and backward amplitudes $(X_{\mu\nu},Y_{\mu\nu})$ have subscripts $\mu\nu$ to specify two-quasiparticles. On the other hand, the subscripts of $(U_{k\mu},V_{k\nu})$ indicate a basis of the single-particle space ($k$) and the quasiparticle ($\mu$). Again, utilizing an iterative algorithm for solution of the QRPA equation, we can solve the QRPA linear-response equation without explicitly calculating the residual interactions. Numerical results ----------------- ### Development of FAM computer programs We have developed the computer codes of the FAM in several representations. The FAM-RPA is available in the three-dimensional (3D) grid representation [@INY09; @INY11]. This provides a completely symmetry unrestricted RPA calculation. All the single-particle wave functions and RPA amplitudes are represented by these grid points: $$\left\{ \phi_i(\vec{r}_k,\sigma), X_i(\omega;\vec{r}_k,\sigma), Y_i(\omega;\vec{r}_k,\sigma)\right\}_{k=1,\cdots,N_{\rm grid}} ^{i=1,\cdots,A;\ \sigma={\rm up, down}} .$$ Since the results are not sensitive to mesh spacing in a region outside of the interacting region, the adaptive grid representation is used to reduce the number of grid points $N_{\rm grid}$ [@NY05]. In the followings, we show results obtained with this code. The code for FAM-QRPA was also developed, but still with some symmetry restrictions at present. In Ref. [@AN11], we developed a code in the radial coordinate representation based on the spherical static program [hfbrad]{} [@DFT84], which assumes the spherical symmetry in the ground state. Lately, another FAM-QRPA code has been developed in the harmonic oscillator representation [@Sto11], which is based on the program [hfbtho]{} with the axial symmetry [@Sto03]. In the present applications, we use complex energy, $\omega=E+i\Gamma/2$, with $\Gamma=1$ MeV, which introduces an artificial damping width. This smoothing is necessary in two reasons: To obtain smooth strength functions, and to speed up the convergence for the iterative procedure. However, the FAM formulae, Eqs. (\[FAM\]) and (\[FAM\_QRPA\]) themselves, are independent from the smoothing parameter $\Gamma$. In fact, we can use the FAM for explicit construction of the (Q)RPA matrix and calculate the discrete normal modes. This will be reported elsewhere [@AN12]. ![Calculated GDR peak energies as functions of mass number (top), neutron number (bottom left), and proton number (bottom right). The isotopic chains are connected by lines for the top and bottom left panels, while the isotonic chains are shown in the bottom right.[]{data-label="peak_position"}](peak_position.m2m1.Ndep.Zdep.eps){height=".3\textheight"} ### Giant dipole resonances We have carried out the systematic calculation of electric dipole response in the FAM-RPA with the 3D grid representation. We show evolution of peak energies of the giant dipole resonances (GDR) as functions of mass number, neutron number, and proton number in Fig. \[peak\_position\]. The GDR peak position is estimated by the average energy $$\label{m_k} E_\mathrm{peak} = \frac{m_1(\omega_{\rm max})}{m_0(\omega_{\rm max})} \,,\quad m_k(\omega_{\rm max}) = \int_0^{\omega_{\rm max}} \mathrm{d}\omega \, \omega^k S(\omega; D_{E1})$$ where $S(\omega;D_{E1})=\sum_n |\langle n|D_{E1}|0\rangle|^2 \delta(\omega-E_n)$. The maximum energy is $\omega_{\rm max}\approx 40$ MeV. The $E1$ operator is defined with the recoil charges for protons, $Ne/A$ and for neutrons, $-Ze/A$. In deformed nuclei, since the peak energy depends on direction of the $E1$ operator ($x$, $y$, and $z$), their averaged value is adopted in Fig. \[peak\_position\]. The upper panel shows the GDR peak energies from oxygen to nickel, as a function of mass number. In the medium-mass region, the peak energies approximately follow the empirical low, $21 A^{-1/3} + 31 A^{-1/6}$ MeV, denoted by the solid curve. However, in each isotopic chain, the peak energies in stable nuclei are the highest, while they are decreasing as leaving from the stability line. In addition, we can see some kind of shell effects. This can be more clearly seen in the lower panels of Fig. \[peak\_position\], in which the peak energies are plotted as functions of neutron and proton numbers. There are cusps at $N=$14 and 28 corresponding to the subshell closure of $1d_{5/2}$ and $1f_{7/2}$ orbitals. This may be attributed to the emergence of low-energy pygmy peaks beyond these neutron numbers (see the discussion in the next section). The proton shell effects seem not to be as significant as those of neutrons (see the bottom right panel of Fig. \[peak\_position\]). ![Calculated PDR strength fraction as a function of neutron number. []{data-label="PDR_Ndep"}](PDR.Ndep2.eps){width="\textwidth"} ### Low-energy $E1$ strength Now, let us move to the low-energy part of the $E1$ strength distribution. In some nuclei, there appear small peaks in the $E1$ strength distribution, well separated from the main GDR peak, which are often called “pygmy dipole resonance” (PDR). In contrast to the main peak of GDR, the PDR strength distribution is sensitive to nuclear properties at nuclear surface and at low density. Thus, its property may provide us with useful constraints on the energy density functional, to identify the equation of state (EOS) of the nuclear and neutron matters. For instance, the neutron skin thickness is known to be well correlated with the slope of the neutron-matter EOS [@Bro00]. Thus, if the neutron skin thickness has a strong correlation with the low-energy $E1$ strength, we may pin down the EOS property by observing the PDR in experiments. First, we define the PDR strength fraction as $$f_{\rm PDR} \equiv m_1(\omega_c)/m_1(\infty) ,$$ where $m_1(\omega)$ is given in Eq. (\[m\_k\]) and we adopt $\omega_c=10$ MeV. In Fig. \[PDR\_Ndep\], we show the neutron-number dependence of $f_{\rm PDR}$. It indicates a strong shell effect. Namely, there are clear kinks at $N=14$, 28, and 50. Let us concentrate our discussion on the kinks at $N=28$. The PDR fractions suddenly increase at $N=28\rightarrow 30$ and continue to increase till $N=34$ where the neutron $2p$ shell are filled. Beyond $N=34$, the PDR fractions are roughly constant for $34<N\leq 50$, in which the neutrons are filling high-$\ell$ orbits of $f_{5/2}$ and $g_{9/2}$. Beyond $N=50$, the neutrons start filling $2d_{5/2}$ orbits, then the $f_{\rm PDR}$ again shows a sudden increase. These behaviors strongly suggest that the spatially extended nature of the low-$\ell$ neutron orbits near the Fermi level plays a primary role for the emergence and growth of the PDR. We have also observed that the deformation tends to increase the PDR strength, especially in the region $N>56$. More detailed analysis can be found in our recent paper [@INY11]. ![Calculated PDR strength fraction as a function of the neutron skin thickness. []{data-label="PDRcorr"}](PDR.diffR.eps){width="72.00000%"} Finally, let us examine the correlation between $f_{\rm PDR}$ and the neutron skin thickness. The skin thickness is defined by the difference in radius between neutrons and protons. Plotting the PDR fraction as a function of the skin thickness, we observe a linear correlation between them, but only in specific regions of the neutron number. This is illustrated in Fig. \[PDRcorr\] for isotopes with $Z=16--22$ and with $Z=26--32$, which show the kinks at $N=28$ and 50. The PDR fraction in each isotopic chain shows a linear correlation with the skin thickness in the regions of the neutron number $N=28-34$ and $N \ge 50$. The positions of the kinks are located at different values of the skin thickness for different isotopes. However, the slope is universal for all the isotopes; $0.18\sim 0.20$ fm$^{-1}$. Despite the fact that the deformation and shell ordering are different and vary from nucleus to nucleus, the universal linear correlation remains valid for $50\leq N < 76$. It should be noted that the linear correlation can be observed only for each isotopic chain. Deleting the lines connecting isotopic chains in Fig. \[PDRcorr\], we only see scattered points showing a weak correlation. Again, for detailed analysis on this issue, readers are referred to Ref. [@INY11]. The work is supported by KAKENHI (Nos. 21340073 and 20105003) and by SPIRE, MEXT, Japan. The numerical calculations were performed on PACS-CS and T2K supercomputers in University of Tsukuba, on Hitachi SR11000 at KEK, and on the RIKEN Integrated Cluster of Clusters (RICC).

ArXiv

--- abstract: 'One of the major goals of quantum thermodynamics is the characterization of irreversibility and its consequences in quantum processes. Here, we discuss how entropy production provides a quantification of the irreversibility in open quantum systems through the quantum fluctuation theorem. We start by introducing a two-time quantum measurement scheme, in which the dynamical evolution between the measurements is described by a completely positive, trace-preserving (CPTP) quantum map (forward process). By inverting the measurement scheme and applying the time-reversed version of the quantum map, we can study how this backward process differs from the forward one. When the CPTP map is unital, we show that the stochastic quantum entropy production is a function only of the probabilities to get the initial measurement outcomes in correspondence of the forward and backward processes. For bipartite open quantum systems we also prove that the mean value of the stochastic quantum entropy production is sub-additive with respect to the bipartition (except for product states). Hence, we find a method to detect correlations between the subsystems. Our main result is the proposal of an efficient protocol to determine and reconstruct the characteristic functions of the stochastic entropy production for each subsystem. This procedure enables to reconstruct even others thermodynamical quantities, such as the work distribution of the composite system and the corresponding internal energy. Efficiency and possible extensions of the protocol are also discussed. Finally, we show how our findings might be experimentally tested by exploiting the state-of-the-art trapped-ion platforms.' address: - '$^1$ via G. Sansone 1, I-50019 Sesto Fiorentino, Italy.' - '$^2$ via S. Marta 3, I-50139 Florence, Italy & , Sezione di Firenze, Sesto Fiorentino, Italy.' - '$^3$, via Bonomea 265, I-34136 Trieste, Italy.' - '$^4$, via Bonomea 265, I-34136 Trieste, Italy & , Sezione di Trieste, I-34151 Trieste, Italy.' - '$^5$, Via Madonna del Piano 10, I-50019 Sesto Fiorentino, Italy.' author: - 'Stefano Gherardini$^{1,2}$, Matthias M. Müller$^1$, Andrea Trombettoni$^{3,4}$, Stefano Ruffo$^{4,5}$, Filippo Caruso$^1$' bibliography: - 'references.bib' nocite: '[@*]' title: Reconstructing quantum entropy production to probe irreversibility and correlations --- Introduction ============ The advent of the thermodynamics laws and their following development, from the theoretical side, and the construction of heat engines, from the technological one, drove in the 18th and 19th centuries an astonishing series of important scientific discoveries and social transformations. The crucial point was the use of heat to produce work, which corresponds to take a disordered form of energy and convert (a part of) it into a mechanical one [@Groot1984]. In the last decades, a breakthrough in non-equilibrium thermodynamics was given by the Jarzynski equality [@JarzynskiPRL1997], which relates the free-energy between two equilibrium states to an exponential average of the work done on the system, over an ideally infinite number of repeated non-equilibrium experiments. This result links together free-energy differences to work measurements along an ensemble of trajectories in the phase space of the system with same energy contribution [@Jarzynski2011]. The Jarzynski equality can be derived also from the Crooks fluctuation theorem [@CrooksPRE1999], which formalizes the existence of symmetry relations for the probability distribution of thermodynamic quantities during the forward and reverse transformations that the system undergoes due to external actions. Generalized versions of the Jarzynski equality for non-equilibrium steady states from Langevin dynamics and non-equilibrium systems subjected to feedback control have been proved, then, respectively in Refs. [@HatanoPRL2001] and [@SagawaPRL2010]. From the experimental side, the Jarzynski equality and its generalizations have been tested by a wide range of experiments, for example to determine the folding and unfolding free energies of a small RNA hairpin [@CollinNAT2005], or to prove the fundamental principle given by the information-to-heat engine, converting information into energy by means of feedback control [@ToyabeNAT2010]. Even from a purely classical point of view, the notion of thermodynamics quantities such as work, heat and entropy production have been extended to the level of individual trajectories of well-defined non-equilibrium ensembles by the stochastic thermodynamics [@Seifert2005; @TietzPRL2006], which has allowed for the introduction of a generalized fluctuation-dissipation theorem involving entropy production. At the same time, the attention moved also towards the attempts to build a thermodynamic theory for quantum systems to exploit the power and the processes of quantum physics [@Gemmer2004; @Horodecki2013; @PekolaNAT2015; @Alhambra2016]. This field of research, known as quantum thermodynamics, aims at characterizing the thermodynamical aspects behind the quantum mechanical processes, defining the role of quantum coherence and measurements for such transformations [@Lostaglio2015; @Narasimhachar2015; @LostaglioPRX2015; @Kammerlander2016]. Quantum thermodynamics, moreover, provides the theoretical tools to describe and build efficient quantum heat engines [@KimPRL2011; @AbahPRL2012; @RossnagelPRL2014]. One of the major goals of quantum thermodynamics is the definition and characterization of irreversibility in quantum processes. This could have a significant impact on technological applications for the possibility of producing work with heat engines at high efficiency using systems where quantum fluctuations are important; in this regard, a detailed analysis about the aspects that define the work done by a quantum system can be found in [@Talkner2016]. The quantum work and its distribution are generally defined by taking into account also the role of quantum measurements and, consequently, the sensitivity of the system to the interactions with the measurement apparatus [@Campisi2011; @HekkingPRL2013; @VenkateshNJP2015; @AlonsoPRL2016]. Recently a novel definition of quantum work has been proposed in [@DeffnerPRE2016], in which the work is a thermodynamic quantity depending only on the quantum system and not on the measurement apparatus. The importance of defining the concept of irreversibility in quantum thermodynamics can be hardly overestimated, as one can appreciate by considering its classical counterpart. As well known, in classical mechanics the solutions of the dynamical equations of motion are unique and the motion along the trajectories in phase space can be inverted to retrieve all the states previously occupied by the system [@Seifert2012Review]. However, the time inversion in experiments with a macroscopic number of particles cannot be practically performed. As a consequence of the information loss and of the fact that is very improbable to occupy the same state at a later time, we have to resort to a statistical description of the system. In classical thermodynamics this is the origin of the irreversibility of the system dynamics. Similarly, in quantum mechanics the dynamics of the wave function and more generally of the density matrix can be reversed in time, and it ensues the corresponding need to characterize and quantify, where possible, irreversible quantum processes [@Esposito2009; @CampisiRMP2011]. The typical instance is given by the thermalization of an open system, where the dissipative processes taking place due to the interaction of the system with its environment degrade the quantum nature of the system and the coherence of the quantum states [@Riera2012; @Gogolin2016]. Along this line, several studies have shown how to derive the quantum version of the fluctuation-dissipation theorem, both for closed [@Kurchan2001; @BuninNAT2011] and open quantum systems [@Campisi2009; @Campisi2010; @Kafri2012; @RasteginJSM2013; @Albash2013; @WatanabePRE2014; @Manzano2015]. Recently, in [@Aberg2016] a fully quantum fluctuation theorem have been formulated, explicitly including the reservoir exchanging energy with the system, and a control system driving its dynamics. In [@Huber2008; @An2015], moreover, experimental tests of the quantum version of the Jarzynski identity [@Mukamel2003; @Chernyak2004; @CrooksJSM2008] for work distributions are shown. Considerable efforts have been made in measuring irreversibility, and, consequently, the stochastic entropy production in quantum thermodynamics [@DeffnerPRL2011; @BatalhaoPRL2015; @Brunelli2016; @Frenzel2016]. The ratio between the probability to observe a given quantum trajectory and its time reversal is related to the amount of heat exchanged by the quantum system with the environment [@Sagawa2014]. Such knowledge leads then to experimental procedures for the measure of the heat backflow with the environment, even if the latter is not necessarily correlated with the information back-flow from the reservoir to the quantum system [@Schmidt2016]. Lately, it has been experimentally proved that irreversibility in quantum non-equilibrium dynamics can be partially rectified by the presence of an *intelligent observer*, identified by the well-known Maxwell’s demon [@KoskiPRL2014; @GooldJPA2016], which manages to assess additional microscopic informational degrees of freedom due to a proper feed-forward strategy [@CamatiPRL2016]. Instead, regarding the reconstruction of the fluctuation properties of general thermodynamical quantities, in Ref. [@MazzolaPRL2013; @DornerPRL2013; @GooldPRE2014; @Peterson2016] an interferometric setting for the measurement of the characteristic function of the work distribution is introduced and proposed as the key element to properly design inference strategies [@Liu2016]. This method, then, has been generalized for open quantum systems, as shown in [@CampisiNJP2014; @FuscoPRX2014]. In [@RoncagliaPRL2014], instead, a method for the sampling of the work distribution by means of a projective measurement at a single time is shown, motivating a novel quantum algorithm for the estimation of free energies in closed quantum systems. In the present work we address three issues. [*(i)*]{} We discuss how to relate the stochastic entropy production to the quantum fluctuation theorem, generalizing the Tasaki-Crooks theorem for open systems. This relation is obtained via the evaluation of the irreversibility of the quantum dynamics. [*(ii)*]{} Then, once the stochastic quantum entropy production has been defined and characterized, we introduce a protocol to reconstruct it from the measurement data, possibly with the minimum amount of resources. Here, we propose a procedure to reconstruct the stochastic entropy production of an open quantum system by performing repeated two-time measurements, at the initial and final times of the system transformation. In particular, the proposed reconstruction algorithm requires to determine the characteristic function of the probability distribution of the stochastic quantum entropy production. Indeed, by means of a parametric version of the integral quantum fluctuation theorem, we can derive the statistical moments of the entropy production. Moreover, we also prove that with this procedure the number of required measurements scales linearly with the system size. [*(iii)*]{} By assuming that the quantum system is bipartite, we apply the reconstruction procedure both for the two subsystems and for the composite system by performing measurements, respectively, on local and global observables. The comparison between the local and the global quantity allows us to probe the presence of correlations between the partitions of the system. The manuscript is organized as follows. Section 2 reviews the quantum fluctuation theorem, introducing the definition of stochastic quantum entropy production. Section 3 analyzes the physical meaning of thermodynamic irreversibility by applying a two-time measurement scheme, and shows the relation between the mean entropy production and the quantum relative entropy of the system density matrix after an arbitrary transformation. The derivation in section 3 sheds light on the importance to design protocols to effectively measure the entropy production of a quantum system. In section 4 we derive the characteristic functions of the probability distributions of the stochastic entropy production within a quantum multipartite system, while in section 5 the reconstruction algorithm is introduced. We propose an experiment implementation with trapped ions in section 6. Finally, we discuss our results and conclusions in section 7. Quantum fluctuation theorem =========================== The fluctuations of the stochastic quantum entropy production obey the quantum fluctuation theorem. The latter can be derived by evaluating the forward and backward protocols for a non-equilibrium process, according to a two-time quantum measurement scheme [@Esposito2009; @CampisiRMP2011]. In this section, we introduce this two-time quantum measurement scheme and define the stochastic quantum entropy production. Then we review the derivation of the quantum fluctuation theorem. We consider an open quantum system that undergoes a transformation in the interval $[0,\tau]$ consisting of measurement, dynamical evolution and second measurement. We call this forward process and then study also its time-reversal, which we call backward process: $$\begin{aligned} \text{FORWARD}:~\rho_{0}\underbrace{\longmapsto}_{\{\Pi^{\textrm{in}}_{m}\}}\rho_{\textrm{in}}\underbrace{\longmapsto}_{\Phi}\rho_{\textrm{fin}} \underbrace{\longmapsto}_{\{\Pi^{\textrm{fin}}_{k}\}}\rho_{\tau} \\ \text{BACKWARD}:~\widetilde{\rho}_{\tau}\underbrace{\longmapsto}_{\{\widetilde{\Pi}^{\textrm{ref}}_{k}\}}\widetilde{\rho}_{\textrm{ref}} \underbrace{\longmapsto}_{\widetilde{\Phi}}\widetilde{\rho}_{\textrm{in}'}\underbrace{\longmapsto}_{\{\widetilde{\Pi}^{\textrm{in}}_{m}\}}\widetilde{\rho}_{0'}\end{aligned}$$ At time $t=0^-$ the system is prepared in a state $\rho_{0}$ and then subjected to a measurement of the observable $\mathcal{O}_{\textrm{in}} = \sum_{m}a^{\textrm{in}}_{m}\Pi^{\textrm{in}}_{m}$, where $\Pi^{\textrm{in}}_{m}\equiv|\psi_{a_{m}}\rangle\langle\psi_{a_{m}}|$ are the projector operators given in terms of the eigenvectors $|\psi_{a_{m}}\rangle$ associated to the eigenvalues $a^{\textrm{in}}_{m}$ (the $m-$th possible outcome of the first measurement). After the first measurement (at $t = 0^{+}$), the density operator describing the ensemble average of the post-measurement states becomes $$\rho_{\textrm{in}} = \sum_{m}p(a^{\textrm{in}}_{m})|\psi_{a_{m}}\rangle\langle\psi_{a_{m}}|,$$ where $p(a^{\textrm{in}}_{m}) = \textrm{Tr}\left[\Pi^{\textrm{in}}_{m}\rho_{0}\Pi^{\textrm{in}}_{m}\right] = \langle\psi_{a_{m}}|\rho_{0}|\psi_{a_{m}}\rangle$ is the probability to obtain the measurement outcome $a^{\textrm{in}}_{m}$. Then, the system undergoes a time evolution, which we assume described by a *unital* completely positive, trace-preserving (CPTP) map $\Phi:L(\mathcal{H})\rightarrow L(\mathcal{H})$, with $L(\mathcal{H})$ denoting the sets of density operators (non-negative operators with unit trace) defined on the Hilbert space $\mathcal{H}$. Quantum maps (known also as quantum channels) represent a very effective tool to describe the effects of the noisy interaction of a quantum system with its environment [@Nielsen2000; @Petruccione2003; @Caruso14]. A CPTP map is unital if it preserves the identity operator $\mathbbm{1}$ on $\mathcal{H}$, [*i.e.*]{} $\Phi(\mathbbm{1}) = \mathbbm{1}$. The assumption of a unital map covers a large family of quantum physical transformations not increasing the purity of the initial states, including, among others, unitary evolutions and decoherence processes. We will briefly discuss later how the protocol presented in section 5 may be modified when the unital map hypothesis is relaxed. The time-evolved ensemble average is then denoted as $$\rho_{\textrm{fin}} \equiv \Phi(\rho_{\textrm{in}})$$ For example, in case of unitary evolution with Hamiltonian $H(t)$, the final quantum state at $t = \tau^{-}$ equals to $\rho_{\textrm{fin}} = \Phi(\rho_{\textrm{in}}) = \mathcal{U}\rho_{\textrm{in}}\mathcal{U}^{\dagger}$, where $\mathcal{U}$ is the unitary time evolution operator given by $\mathcal{U}=\mathbb{T}\exp\left(-\frac{i}{\hbar}\int^{\tau}_{0}H(t)dt\right)$, with $\mathbb{T}$ time-ordering operator. After the time evolution, at time $t = \tau^+$, a second measurement is performed on the quantum system according to the observable $\mathcal{O}_{\textrm{fin}} = \sum_{k}a^{\textrm{fin}}_{k}\Pi^{\textrm{fin}}_{k}$, where $\Pi^{\textrm{fin}}_{k} \equiv |\phi_{a_{k}}\rangle\langle\phi_{a_{k}}|$, and $a^{\textrm{fin}}_{k}$ is the $k-$th outcome of the second measurement (with eigenvectors $|\phi_{a_{k}}\rangle$). Consequently, the probability to obtain the measurement outcome $a_{k}^{\textrm{fin}}$ is $p(a^{\textrm{fin}}_{k}) = \textrm{Tr}\left[\Pi^{\textrm{fin}}_{k}\Phi(\rho_{\textrm{in}})\Pi^{\textrm{fin}}_{k}\right] = \langle\phi_{a_{k}}|\rho_{\textrm{fin}}|\phi_{a_{k}}\rangle$. The resulting density operator, describing the ensemble average of the post-measurement states after the second measurement, is $$\rho_{\tau} = \sum_{k}p(a^{\textrm{fin}}_{k})|\phi_{a_{k}}\rangle\langle\phi_{a_{k}}|.$$ Thus, the joint probability that the events “measure of $a^{\textrm{in}}_{m}$” and “measure of $a^{\textrm{fin}}_{k}$” both occur for the forward process, denoted by $p(a^{\textrm{fin}} = a^{\textrm{fin}}_{k}, a^{\textrm{in}} = a^{\textrm{in}}_{m})$, is given by $$p(a^{\textrm{fin}}_{k},a^{\textrm{in}}_{m}) = \textrm{Tr}\left[\Pi^{\textrm{fin}}_{k}\Phi(\Pi^{\textrm{in}}_{m}\rho_{0}\Pi^{\textrm{in}}_{m})\right].$$ To study the backward process, we first have to introduce the concept of time-reversal. Time-reversal is achieved by the time-reversal operator $\Theta$ acting on $\mathcal{H}$. The latter has to be an antiunitary operator. An antiunitary operator $\Theta$ is anti-linear, [*i.e.*]{} $$\Theta (x_1|\varphi_1\rangle + x_2|\varphi_2\rangle)= x_1^\star \Theta|\varphi_1\rangle + x_2^\star \Theta|\varphi_2\rangle$$ for arbitrary complex coefficients $x_1$, $x_2$ and $|\varphi_1\rangle$, $|\varphi_2\rangle$ $\in$ $\mathcal{H}$, and it transforms the inner product as $\langle \widetilde{\varphi}_1|\widetilde{\varphi}_2\rangle=\langle\varphi_2|\varphi_1\rangle$ for $|\widetilde{\varphi}_1\rangle=\Theta|\varphi_1\rangle$, and $|\widetilde{\varphi}_2\rangle=\Theta|\varphi_2\rangle$. Antiunitary operators satisfy the relations $\Theta^{\dagger}\Theta = \Theta\Theta^{\dagger} = \mathbbm{1}$. The antiunitarity of $\Theta$ ensures the time-reversal symmetry [@Sozzi2008]. We define the time-reversed density operator as $\widetilde{\rho}\equiv\Theta\rho\Theta^\dagger$, and we consider the time-reversal version of the quantum evolution operator, [*i.e.*]{} our unital CPTP map $\Phi$. Without loss of generality, it admits an operator-sum (or Kraus) representation: $\rho_{\textrm{fin}} = \Phi(\rho_{\textrm{in}}) = \sum_{u}E_{u}\rho_{\textrm{in}}E_{u}^{\dagger}$ with the Kraus operators $E_{u}$ being such that $\sum_{u}E_{u}^{\dagger} E_{u} = \mathbbm{1}$ (trace-preserving) [@Nielsen2000; @Petruccione2003; @Caruso14]. For each Kraus operator $E_{u}$ of the forward process we can define the corresponding time-reversed operator $\widetilde{E}_{u}$ [@CrooksPRA2008; @Manzano2015], so that the time-reversal $\widetilde{\Phi}$ for the CPTP quantum map $\Phi$ is given by $$\label{reversed_map} \widetilde{\Phi}(\rho) = \sum_{u}\widetilde{E}_{u}\rho\widetilde{E}_{u}^{\dagger},$$ where $\widetilde{E}_{u} \equiv \mathcal{A}\pi^{1/2}E^{\dagger}_{u}\pi^{-1/2}\mathcal{A}^{\dagger}$, $\pi$ is an invertible fixed point (not necessarily unique) of the quantum map, such that $\Phi(\pi) = \pi$, and $\mathcal{A}$ is an arbitrary (unitary or anti-unitary) operator. Usually, the operator $\mathcal{A}$ is chosen equal to the time-reversal operator $\Theta$. If the density operator $\pi$ is a positive definite operator, as assumed in Refs. [@CrooksPRA2008; @Horowitz2013], then also the square root $\pi^{1/2}$ is positive definite and the inverse $\pi^{-1/2}$ exists and it is unique. Since our map is unital we can choose $\pi^{1/2}=\pi^{-1/2}=\mathbbm{1}$. Thus, from Eq. (\[reversed\_map\]), we can observe that also $\widetilde{\Phi}$ is a CPTP quantum map with an operator sum-representation, such that $\sum_{u}\widetilde{E}_{u}^{\dagger}\widetilde{E}_{u} = \mathbbm{1}$. Summarizing, we have $$\widetilde{E}_{u} = \Theta E^{\dagger}_{u}\Theta^\dagger,$$ so that $$\widetilde{\Phi}(\rho) = \sum_{u}\widetilde{E}_{u}\rho\widetilde{E}_{u}^{\dagger}=\Theta\left(\sum_u E^{\dagger}_{u} \widetilde{\rho} E_{u}\right) \Theta^\dagger.$$ We are now in a position to define the backward process. We start by preparing the system (at time $t=\tau^+$) in the state $\widetilde{\rho}_{\tau} = \Theta\rho_{\tau}\Theta^\dagger$, and measure the observable $\widetilde{\mathcal{O}}_{\textrm{ref}} \equiv \sum_{k}a^{\textrm{ref}}_{k}\widetilde{\Pi}^{\textrm{ref}}_{k}$, with $\widetilde{\Pi}^{\textrm{ref}}_{k} = |\widetilde{\phi}_{a_{k}}\rangle\langle\widetilde{\phi}_{a_{k}}|$ and $|\widetilde{\phi}_{a_{k}}\rangle\equiv\Theta|\phi_{a_{k}}\rangle$, that is we choose this first measurement of the backward process to be the time-reversed version of the second measurement of the forward process. If we call the post-measurement ensemble average $\widetilde{\rho}_{\textrm{ref}}$, as a consequence $\widetilde{\rho}_{\tau}=\widetilde{\rho}_{\textrm{ref}}$, or equivalently $\rho_{\tau}=\rho_{\textrm{ref}}$, where the latter is called reference state. In particular, we recall that, although the quantum fluctuation theorem can be derived without imposing a specific operator for the reference state [@Sagawa2014], the latter has been chosen to be identically equal to the final density operator after the second measurement of the protocol. This choice appears to be the most natural among the possible ones to design a suitable measuring scheme of general thermodynamical quantities, consistently with the quantum fluctuation theorem. The spectral decomposition of the time-reversed reference state is given by $$\label{rho_ref} \widetilde{\rho}_{\textrm{ref}} = \sum_{k}p(a^{\textrm{ref}}_{k})|\widetilde{\phi}_{a_{k}}\rangle\langle\widetilde{\phi}_{a_{k}}|,$$ where $$p(a^{\textrm{ref}}_{k}) = \textrm{Tr}[\widetilde{\Pi}^{\textrm{ref}}_{k}\widetilde{\rho}_{\tau}\widetilde{\Pi}^{\textrm{ref}}_{k}] = \langle\widetilde{\phi}_{a_{k}}|\widetilde{\rho}_{\tau}|\widetilde{\phi}_{a_{k}}\rangle$$ is the probability to get the measurement outcome $a_{k}^{\textrm{ref}}$. The reference state undergoes the time-reversal dynamical evolution, mapping it onto the initial state of the backward process $\widetilde{\rho}_{\textrm{in}'}=\widetilde{\Phi}(\widetilde{\rho}_{\textrm{ref}})$. At $t = 0^{+}$ the density operator $\widetilde{\rho}_{\textrm{in}'} = \widetilde{\Phi}(\widetilde{\rho}_{\textrm{ref}})$ is subject to the second projective measurement of the backward process, whose observable is given by $\widetilde{\mathcal{O}}_{\textrm{in}} = \sum_{m}a^{\textrm{in}}_{m}\widetilde{\Pi}^{\textrm{in}}_{m}$, with $\widetilde{\Pi}^{\textrm{in}}_{m} = |\widetilde{\psi}_{a_{m}}\rangle\langle\widetilde{\psi}_{a_{m}}|$, and $|\widetilde{\psi}_{a_{m}}\rangle\equiv\Theta|\psi_{a_{m}}\rangle$. As a result, the probability to obtain the outcome $a^{\textrm{in}}_{m}$ is $p(a^{\textrm{in}}_{m}) = \textrm{Tr}[\widetilde{\Pi}^{\textrm{in}}_{m}\widetilde{\Phi}(\widetilde{\rho}_{\textrm{ref}})\widetilde{\Pi}^{\textrm{in}}_{m}] = \langle\widetilde{\psi}_{a_{m}}|\widetilde{\rho}_{\textrm{in}'}|\widetilde{\psi}_{a_{m}}\rangle$, while the joint probability $p(a^{\textrm{in}}_{m}, a^{\textrm{ref}}_{k})$ is given by $$p(a^{\textrm{in}}_{m}, a^{\textrm{ref}}_{k}) = \textrm{Tr}[\widetilde{\Pi}^{\textrm{in}}_{m}\widetilde{\Phi}(\widetilde{\Pi}^{\textrm{ref}}_{k}\widetilde{\rho}_{\tau} \widetilde{\Pi}^{\textrm{ref}}_{k})].$$ The final state of the backward process is instead $\widetilde{\rho}_{0'}=\sum_{m} p(a^{\textrm{in}}_{m}) \widetilde{\Pi}^{\textrm{in}}_{m}$. Let us observe again that the main difference of the two-time measurement protocol that we have introduced here, compared to the scheme in Ref. [@Sagawa2014], is to perform the $2$nd and $1$st measurement of the backward protocol, respectively, on the same basis of the $1$st and $2$nd measurement of the forward process after a time-reversal transformation. The irreversibility of the two-time measurement scheme can be analyzed by studying the stochastic quantum entropy production $\sigma$ defined as: $$\label{general_sigma} % \sigma(a^{\textrm{fin}}_{k},a^{\textrm{in}}_{m}) \equiv \ln\left[\frac{p(a^{\textrm{fin}}_{k}, a^{\textrm{in}}_{m})}{p(a^{\textrm{in}}_{m}, a^{\textrm{ref}}_{k})}\right] % = \ln\left[\frac{p(a^{\textrm{fin}}_{k}|a^{\textrm{in}}_{m})p(a^{\textrm{in}}_{m})} {p(a^{\textrm{in}}_{m}|a^{\textrm{ref}}_{k})p(a^{\textrm{ref}}_{k})}\right], %$$ where $p(a^{\textrm{fin}}_{k}|a^{\textrm{in}}_{m})$ and $p(a^{\textrm{in}}_{m}|a^{\textrm{ref}}_{k})$ are the conditional probabilities of measuring, respectively, the outcomes $a^{\textrm{fin}}_{k}$ and $a^{\textrm{in}}_{m}$, conditioned on having first measured $a^{\textrm{in}}_{m}$ and $a^{\textrm{ref}}_{k}$. Its mean value $$\label{general_mean_sigma} \langle\sigma\rangle = \sum_{k,m}p(a_k^{\textrm{fin}},a_m^{\textrm{in}})\ln\left[\frac{p(a_k^{\textrm{fin}},a_m^{\textrm{in}})}{p(a_k^{\textrm{in}}, a_m^{\textrm{ref}})}\right]$$ corresponds to the classical relative entropy (or Kullback-Leibler divergence) between the joint probabilities $p(a^{\textrm{fin}}, a^{\textrm{in}})$ and $p(a^{\textrm{in}}, a^{\textrm{ref}})$, respectively, of the forward and backward processes [@Cover2006; @Umegaki]. The Kullback-Leibler divergence is always non-negative and as a consequence $$\langle\sigma\rangle\geq 0.$$ As a matter of fact, $\langle\sigma\rangle$ can be considered as the amount of additional information that is required to achieve the backward process, once the quantum system has reached the final state $\rho_{\tau}$. Moreover, $\langle\sigma\rangle = 0$ if and only if $p(a^{\textrm{fin}}_{k}, a^{\textrm{in}}_{m}) = p(a^{\textrm{in}}_{m}, a^{\textrm{ref}}_{k})$, [*i.e.*]{} if and only if $\sigma = 0$. To summarize, the transformation of the system state from time $t=0^{-}$ to $t=\tau^{+}$ is then defined to be thermodynamically irreversible if $\langle\sigma\rangle > 0$. If, instead, all the fluctuations of $\sigma$ shrink around $\langle\sigma\rangle \simeq 0$ the system comes closer and closer to a reversible one. We observe that a system transformation may be thermodynamically irreversible also if the system undergoes unitary evolutions with the corresponding irreversibility contributions due to applied quantum measurements. Also the measurements back-actions, indeed, lead to energy fluctuations of the quantum system, as recently quantified in Ref. [@Elouard2016]. In case there is no evolution (identity map) and the two measurement operators are the same, then the transformation becomes reversible. We can now state the following theorem:\ \ **Theorem 1**: *Given the two-time measurement protocol described above and an open quantum system dynamics described by a unital CPTP quantum map* $\Phi$, *it can be stated that:* $$\label{equality_cond_prob} p(a^{\textrm{fin}}_{k}|a^{\textrm{in}}_{m}) = p(a^{\textrm{in}}_{m}|a^{\textrm{ref}}_{k}).$$ The proof of Theorem 1 can be found in \[appendix\_a\].\ Throughout this article we assume that $\Phi$ is unital and this property of the map guarantees the validity of Theorem 1. Note, however, that Refs. [@Horowitz2013; @Manzano2015] present a fluctuation theorem for slightly more general maps, that however violate Eq. (\[equality\_cond\_prob\]). As a consequence of Theorem 1 we obtain: $$\label{sigma} \sigma(a^{\textrm{fin}}_{k},a^{\textrm{in}}_{m}) = \ln\left[\frac{p(a^{\textrm{in}}_{m})}{p(a^{\textrm{ref}}_{k})}\right] =\ln\left[\frac{\langle\psi_{a_{m}}|\rho_{0}|\psi_{a_{m}}\rangle} {\langle\widetilde{\phi}_{a_{k}}|\widetilde{\rho}_{\tau}|\widetilde{\phi}_{a_{k}}\rangle}\right].$$ providing a general expression of the quantum fluctuation theorem for the described two-time quantum measurement scheme. Let us introduce, now, the entropy production $\widetilde{\sigma}$ for the backward processes, [*i.e.*]{} $$\widetilde{\sigma}(a^{\textrm{in}}_{m},a^{\textrm{ref}}_{k})\equiv\ln\left[\frac{p(a^{\textrm{in}}_{m}, a^{\textrm{ref}}_{k})}{p(a^{\textrm{fin}}_{k}, a^{\textrm{in}}_{m})}\right] = \ln\left[\frac{p(a^{\textrm{ref}}_{k})}{p(a^{\textrm{in}}_{m})}\right],$$ where the second identity is valid only in case we can apply the results deriving from Theorem 1. Hence, if we define $\textrm{Prob}(\sigma)$ and $\textrm{Prob}(\widetilde{\sigma})$ as the probability distributions of the stochastic entropy production, respectively, for the forward and the backward processes, then it can be shown (see e.g. Ref. [@Sagawa2014]) that $$\label{qft} \frac{\textrm{Prob}(\widetilde{\sigma} = -\Gamma)}{\textrm{Prob}(\sigma = \Gamma)} = e^{-\Gamma},$$ where $\Gamma$ belongs to the set of values that can be assumed by the stochastic quantum entropy production $\sigma$. Eq. (\[qft\]) is usually called quantum fluctuation theorem. By summing over $\Gamma$, we recover the *integral quantum fluctuation theorem*, or quantum Jarzynski equality, $\langle e^{-\sigma}\rangle = 1$, as shown e.g. in Refs. [@Kurchan2001; @Sagawa2014]. The role of the integral fluctuation theorem in deriving the probability distribution $\textrm{Prob}(\sigma)$ of the stochastic entropy production for an open quantum system is analyzed in the following sections. Mean entropy production vs quantum relative entropy {#new_sec_3} =================================================== In this section, we discuss the irreversibility of the two-time measurement scheme for an open quantum system $\mathcal{S}$ in interaction with the environment $\mathcal{E}$ (described by a unital CPTP map), deriving an inequality (Theorem 2) for the entropy growth. Following Ref. [@Sagawa2014], the essential ingredient is the non-negativity of the quantum relative entropy and its relation to the stochastic quantum entropy production. As a generalization of the Kullback-Leibler information [@Umegaki], the quantum relative entropy between two arbitrary density operators $\nu$ and $\mu$ is defined as $S(\nu\parallel\mu)\equiv\textrm{Tr}[\nu\ln\nu] - \textrm{Tr}[\nu\ln\mu]$. The Klein inequality states that the quantum relative entropy is a non-negative quantity [@Vedral2002], [*i.e.*]{} $S(\nu\parallel\mu)\geq 0$, where the equality holds if and only if $\nu = \mu$ - see [*e.g.*]{} [@Sagawa2014]. In particular, in the following theorem we will show the relation between the quantum relative entropy of the system density matrix at the final time of the transformation and the stochastic quantum entropy production for unital CPTP quantum maps:\ \ **Theorem 2**: *Given the two-time measurement protocol described above and an open quantum system dynamics described by a unital CPTP quantum map* $\Phi$, *the quantum relative entropy* $S(\rho_{\textrm{fin}}\parallel\rho_{\tau})$ *fulfills the inequality* $$\label{eq:entropy-positivity} 0\leq S(\rho_{\textrm{fin}}\parallel\rho_{\tau})\leq\langle\sigma\rangle,$$ *where the equality* $S(\rho_{\textrm{fin}}\parallel\rho_{\tau}) = 0$ *holds if and only if* $\rho_{\textrm{fin}} = \rho_{\tau}$. *Then, for* $[\mathcal{O}_{\textrm{fin}},\rho_{\textrm{fin}}]=0$ *one has* $\langle\sigma\rangle = S(\rho_{\tau}) - S(\rho_{\textrm{in}})$, *so that* $$\label{eq:entropy-of-map} 0= S(\rho_{\textrm{fin}}\parallel\rho_{\tau})\leq\langle\sigma\rangle=S(\rho_{\textrm{fin}})-S(\rho_{\textrm{in}}),$$ *where $S(\cdot)$ denotes the von Neumann entropy of $(\cdot)$. Finally,* $S(\rho_{\textrm{fin}}\parallel\rho_{\tau}) = \langle\sigma\rangle$ *if* $\mathcal{S}$ *is a closed quantum system following a unitary evolution*. A proof of Theorem 2 is in \[appendix\_b\].\ While Eq. (\[eq:entropy-positivity\]) is more general and includes the irreversibility contributions of both the map $\Phi$ and the final measurement, in Eq. (\[eq:entropy-of-map\]) due to a special choice of the observable of the second measurement we obtain $\rho_{\textrm{fin}}=\rho_{\tau}$ and, thus, the quantum relative entropy vanishes while the stochastic quantum entropy production contains the irreversibility contribution only from the map. This contribution is given by the difference between the von Neumann entropy of the final state $S(\rho_{\textrm{fin}})$ and the initial one $S(\rho_{\textrm{in}})$[^1]. Physical considerations ----------------------- To summarize the previous results, in case the environment $\mathcal{E}$ is not thermal, the stochastic quantum entropy production represents a very general measurable thermodynamic quantity, encoding information about the interaction between the system and the environment also in a fully quantum regime. Therefore, its reconstruction becomes relevant, not only for the fact that we cannot longer adopt energy measurements on $\mathcal{S}$ to infer $\sigma$ and its fluctuation properties, but also because in this way we could manage to measure the mean heat flux exchanged by the partitions of $\mathcal{S}$ in case it is a multipartite quantum system, as shown in the following sections. Recovering the second law of thermodynamics {#rec_2nd_law} ------------------------------------------- After stating the results of Theorem 2, and, accordingly, having discussed the irreversibility contributions coming from unital quantum CPTP maps, the following question naturally emerges about the connection between the Theorem 2, valid also for a quantum dynamics at $T=0$, and the second law of thermodynamics, given the fact that in Theorem 2 there is no specific reference to thermal states. Therefore we think that it is useful in this Section to address the question: How the inequality (\[eq:entropy-positivity\]) for the entropy growth can be connected to the conventional second law of thermodynamics given in terms of the energetic quantities of the quantum system? In a fully quantum regime, following [@Alicki1979; @LorenzoPRL2015], the internal energy of a quantum system $\mathcal{S}$ is given by the relation $\textrm{Tr}[\rho(t)H(t)]\equiv \textrm{Tr}[\rho H](t)$, where $H(t)$ is the (time-dependent) Hamiltonian of the system. Accordingly, an infinitesimal change of the internal energy during the infinitesimal interval $[t,t+\delta t]$ will be $\delta \textrm{Tr}[\rho H](t)\equiv\textrm{Tr}[\rho(t+\delta t)H(t+\delta t)] - \textrm{Tr}[\rho(t)H(t)]$. The latter, then, can be recast into the following equation, representing the first law of thermodynamics for the quantum system: $$\label{eq:diff_1stlaw} \delta \textrm{Tr}[\rho H](t) = \textrm{Tr}[\rho(t)\delta H(t)] + \textrm{Tr}[\delta\rho(t)H(t)],$$ where $\delta H(t)\equiv H(t+\delta t)-H(t)$ and $\delta\rho(t)\equiv\rho(t+\delta t)-\rho(t)$. The quantity $\textrm{Tr}[\rho(t)\delta H(t)]$ is the infinitesimal mean work $\delta\langle\mathrm{W}\rangle(t)$ done by the system in the time interval $[t,t+\delta t]$, while $\textrm{Tr}[\delta\rho(t)H(t)]$ denotes the infinitesimal mean heat flux $\delta\langle\mathrm{Q}\rangle(t)$, which is identically equal to zero if the quantum system dynamics is unitary. Thus, the mean work done the system is $\langle W\rangle = \textrm{Tr}[\rho_{\textrm{fin}}H(\tau)] - \textrm{Tr}[\rho_{\textrm{in}}H(0)]$, while the mean heat flux $\langle\mathrm{Q}\rangle$, for a time-independent Hamiltonian and a finite value change of the internal energy of the quantum system during the protocol, equals to $ \langle\mathrm{Q}\rangle = \textrm{Tr}[\rho_{\textrm{fin}}H] - \textrm{Tr}[\rho_{\textrm{in}}H] = \textrm{Tr}\left[(\Phi - \mathbb{I})[\rho_{\textrm{in}}]H\right], $ where $\mathbb{I}$ is the identity map acting on the sets of the density operators within the Hilbert space of $\mathcal{S}$. From the results of Theorems 1 and 2, valid under the hypothesis that the quantum CPTP map of the quantum system is unital, one has that $$\label{sigma_theorem2} \langle\sigma\rangle = - {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}] - S(\rho_{\textrm{in}}).$$ In order to recover the second law of thermodynamics as a relation between the mean work $\langle\mathrm{W}\rangle$ and the Helmholtz free-energy difference $\Delta F \equiv F(\tau) - F(0)$, we need to quantify the deviations of ${\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}]$ from the corresponding value in a thermal state $\rho_{\tau}^{{\rm th}}$, defined to be the thermal state for the Hamiltonian $H$ at time $\tau$. In formulas, $\rho_{\tau}^{{\rm th}} \equiv e^{\beta\left[F(\tau)\mathbbm{1}_\mathcal{S} - H(\tau)\right]}$ at time $t = \tau$. The derivation is done in \[appendix\_second\_law\], and the result is $$\label{new_equation_quantum_entropy} S(\rho_{\textrm{fin}}\parallel\rho_{\tau}) + {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}] = S(\rho_{\textrm{fin}}\parallel\rho_{\tau}^{{\rm th}}) + {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}^{{\rm th}}],$$ so that $${\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}] = {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}^{{\rm th}}] + S(\rho_{\textrm{fin}}\parallel\rho_{\tau}^{{\rm th}}) - S(\rho_{\textrm{fin}}\parallel\rho_{\tau}).$$ Accordingly, by substituting $\rho_{\textrm{in}} \equiv e^{\beta\left[F(0)\mathbbm{1}_\mathcal{S} - H(0)\right]}$ and $\ln\rho_{\tau}^{{\rm th}} = \beta(F(\tau) - H(\tau))$ in Eq. (\[sigma\_theorem2\]), one has that $$\langle\sigma\rangle - S(\rho_{\textrm{fin}}\parallel\rho_{\tau}) = \beta\left(\langle\mathrm{W}\rangle - \Delta F\right) - S(\rho_{\textrm{fin}}\parallel\rho_{\tau}^{{\rm th}}),$$ where $\langle\sigma\rangle - S(\rho_{\textrm{fin}}\parallel\rho_{\tau})\geq 0$, since $0\leq S(\rho_{\textrm{fin}}\parallel\rho_{\tau})\leq\langle\sigma\rangle$. Finally, observing that $S(\rho_{\textrm{fin}}\parallel\rho_{\tau}^{{\rm th}})\geq 0$ being a quantum relative entropy, we recover the conventional second law of thermodynamics $$\langle\mathrm{W}\rangle\geq\Delta F.$$ The validity of the second law of thermodynamics has been proved just by exploiting the non-negativity of the quantum relative entropy and the results from Theorems 1 and 2. However, to avoid possible misunderstadigs, let us clarify that a unital quantum process cannot in general describe the mapping between two Gibbs thermal states, and, thus, neither a thermalization process for $\mathcal{S}$. Accordingly, the density operator $\rho_{\tau}$ will not be physically equal to the corresponding thermal state $\rho_{\tau}^{{\rm th}}$, and $\langle\sigma\rangle$ is not linearly proportional (with $\beta$ as proportionality constant) to the internal energy of $\mathcal{S}$, i.e. to the mean heat flux $\langle\mathrm{Q}\rangle$. One can see this taking Eq. (\[sigma\_theorem2\]) and substituting in $\rho_{\textrm{in}}$ the thermal state $e^{\beta\left[F(0)\mathbbm{1}_\mathcal{S} - H(0)\right]}$: being $\rho_{\tau}$ a mixed but not thermal state, necessarily $\langle\sigma\rangle \neq \beta\left(\langle\mathrm{W}\rangle - \Delta F\right)$.\ Stochastic quantum entropy production for open bipartite systems ================================================================ In this section, our intent is to define and, then, reconstruct the fluctuation profile of the stochastic quantum entropy production $\sigma$ for an open multipartite system (for simplicity we will analyze in detail a bipartite system), so as to characterize the irreversibility of the system dynamics after an arbitrary transformation. At the same time, we will also study the role played by the performance of measurements both on local and global observables for the characterization of $\textrm{Prob}(\sigma)$ in a many-body context, and evaluate the efficiency of reconstruction in both cases. In particular, as shown by the numerical examples, by comparing the mean stochastic entropy productions $\langle\sigma\rangle$ obtained by local measurements on partitions of the composite system and measurements on its global observables, we are able to detect (quantum and classical) correlations between the subsystems, which have been caused by the system dynamics. To this end, let us assume that the open quantum system $\mathcal{S}$ is composed of two distinct subsystems ($A$ and $B$), which are mutually interacting, and we denote by $A-B$ the composite system $\mathcal{S}$. However, all the presented results can be in principle generalized to an arbitrary number of subsystems. As before, the initial and final density operators of the composite system are arbitrary (not necessarily equilibrium) quantum states, and the dynamics of the composite system is described by a unital CPTP quantum map. The two-time measurement scheme on $A - B$ is implemented by performing the measurements locally on $A$ and $B$ and we assume, moreover, that the measurement processes at the beginning and at the end of the protocol are independent. Since the local measurement on $A$ commutes with the local measurement on $B$, the two measurements can be performed simultaneously. This allows us to consider the stochastic entropy production for the composite system by considering the correlations between the measurement outcomes of the two local observables. Alternatively, by disregarding these correlations, we can consider separately the stochastic entropy production of each subsystem. The composite system $A - B$ is defined on the finite-dimensional Hilbert space $\mathcal{H}_{A-B}\equiv\mathcal{H}_{A}\otimes\mathcal{H}_{B}$ (with $\mathcal{H}_A$ and $\mathcal{H}_B$ the Hilbert spaces of system $A$ and $B$, respectively), and its dynamics is governed by the following time-dependent Hamiltonian $$H(t) = H_{A}(t)\otimes\mathbbm{1}_{B} + \mathbbm{1}_{A}\otimes H_{B}(t) + H_{A-B}(t).$$ $\mathbbm{1}_{A}$ and $\mathbbm{1}_{B}$ are the identity operators acting, respectively, on the Hilbert spaces of the systems $A$ and $B$, while $H_{A}$ is the Hamiltonian of $A$, $H_{B}$ the Hamiltonian of system $B$, and $H_{A-B}$ is the interaction term. We denote the initial density operator of the composite quantum system $A - B$ by $\rho_{0}$ (before the first measurement), which we assume to be a product state, then the ensemble average after the first measurement (at $t = 0^{+}$) is given by the density operator $\rho_{\textrm{in}}$, which can be written as: $$\label{rho_in} \rho_{\textrm{in}}=\rho_{A,\textrm{in}}\otimes\rho_{B,\textrm{in}},$$ where $$\begin{cases} \rho_{A,\textrm{in}}=\sum_{m}p(a_{m}^{\textrm{in}})\Pi^{\textrm{in}}_{A,m} \\ \rho_{B,\textrm{in}}=\sum_{h}p(b_{h}^{\textrm{in}})\Pi^{\textrm{in}}_{B,h} \\ \end{cases}$$ are the reduced density operators for the subsystems $A$ and $B$, respectively. The projectors $\Pi^{\textrm{in}}_{A,m} \equiv |\psi_{a_{m}}\rangle\langle\psi_{a_{m}}|$ and $\Pi^{\textrm{in}}_{B,h}\equiv|\psi_{b_{h}}\rangle\langle\psi_{b_{h}}|$ are the projectors onto the respective eigenstates of the local measurement operators for the subsystems $A$ and $B$: the observables $\mathcal{O}^{\textrm{in}}_{A}=\sum_m a_m^{\textrm{in}}\Pi^{\textrm{in}}_{A,m}$ on system $A$ and $\mathcal{O}^{\textrm{in}}_{B}=\sum_h b_h^{\textrm{in}}\Pi^{\textrm{in}}_{B,h}$ on system $B$, with possible measurement outcomes $\{a^{\textrm{in}}_{m}\}$ and $\{b^{\textrm{in}}_{h}\}$, upon measurement of $\rho_0$. After the measurement, the composite system $A - B$ undergoes a time evolution up to the time instant $t = \tau^{-}$, described by the unital CPTP quantum map $\Phi$, such that $\rho_{\textrm{fin}}=\Phi(\rho_{\textrm{in}})$. Then, a second measurement is performed on both systems, measuring the observables $\mathcal{O}^{\textrm{fin}}_{A}=\sum_k a_k^{\textrm{fin}}\Pi^{\textrm{fin}}_{A,k}$ on system $A$ and $\mathcal{O}^{\textrm{fin}}_{B}=\sum_l b_l^{\textrm{fin}}\Pi^{\textrm{fin}}_{B,l}$ on system $B$, where $\{a_k^{\textrm{fin}}\}$ and $\{b_l^{\textrm{fin}}\}$ are the eigenvalues of the observables, and the projector $\Pi^{\textrm{fin}}_{A,k}\equiv|\phi_{a_{k}}\rangle\langle\phi_{a_{k}}|$ and $\Pi^{\textrm{fin}}_{B,l}\equiv|\phi_{b_{l}}\rangle\langle\phi_{b_{l}}|$ are given by the eigenstates $|\phi_{a_{k}}\rangle$ and $|\phi_{b_{l}}\rangle$, respectively. After the second measurement, we have to make a distinction according to whether we want to take into account correlations between the subsystems or not. If we disregard the correlations, the ensemble average over all the local measurement outcomes of the state of the quantum system at $t = \tau^{+}$ is described by the following product state $\rho_{A,\tau}\otimes\rho_{B,\tau}$, where $$\begin{cases} \rho_{A,\tau}=\sum_{k}p(a_{k}^{\textrm{fin}})\Pi^{\textrm{fin}}_{A,k} \\ \rho_{B,\tau}=\sum_{l}p(b_{l}^{\textrm{fin}})\Pi^{\textrm{fin}}_{B,l} \\ \end{cases}.$$ The probabilities $p(a_{k}^{\textrm{fin}})$ to obtain outcome $a_{k}^{\textrm{fin}}$ and $p(b_{l}^{\textrm{fin}})$ to obtain the measurement outcome $b_{l}^{\textrm{fin}}$ are given by $$\begin{cases} p(a_{k}^{\textrm{fin}}) = \textrm{Tr}_{A}\left[\Pi^{\textrm{fin}}_{A,k}\textrm{Tr}_{B}\left[\rho_{\textrm{fin}}\right]\right] \\ p(b_{l}^{\textrm{fin}}) = \textrm{Tr}_{B}\left[\Pi^{\textrm{fin}}_{B,l}\textrm{Tr}_{A}\left[\rho_{\textrm{fin}}\right]\right] \\ \end{cases},$$ where $\textrm{Tr}_{A}\left[\cdot\right]$ and $\textrm{Tr}_{B}\left[\cdot\right]$ denote, respectively, the operation of partial trace with respect to the quantum systems $A$ and $B$. Conversely, in order to keep track of the correlations between the simultaneously performed local measurements, we have to take into account the following global observable of the composite system $A-B$: $$\label{eq:ovservable_A-B} \mathcal{O}^{\textrm{fin}}_{A - B} = \sum_{k,l}c_{kl}^{\textrm{fin}}\Pi^{\textrm{fin}}_{A - B,kl}~,$$ where $\Pi^{\textrm{fin}}_{A - B,kl} \equiv \Pi^{\textrm{fin}}_{A,k}\otimes \Pi^{\textrm{fin}}_{B,l}$ and $\{c^{\textrm{fin}}_{kl}\}$ are the outcomes of the final measurement of the protocol. The state of the system after the second measurement at $t = \tau^{+}$ is then described by an ensemble average over all outcomes of the joint measurements: $$\rho_{\tau} = \sum_{k,l}p(c^{\textrm{fin}}_{kl})\Pi^{\textrm{fin}}_{A - B,kl}~,$$ where $p(c^{\textrm{fin}}_{kl}) = \textrm{Tr}\left[\Pi^{\textrm{fin}}_{A - B,kl}~\rho_{\textrm{fin}}\right]$. In both cases, consistently with the previous sections, we choose $\rho_{\tau}$ as the reference state of the composite system. The measurement outcomes of the initial and final measurement for the composite system $A - B$ are, respectively, $c^{\textrm{in}}_{mh}\equiv(a_{m}^{\textrm{in}},b_{h}^{\textrm{in}})$ and $c^{\textrm{fin}}_{kl}\equiv(a_{k}^{\textrm{fin}},b_{l}^{\textrm{fin}})$. These outcomes occur with probabilities $p(c^{\textrm{in}}_{mh})$ and $p(c^{\textrm{fin}}_{kl})$, which reflect the correlation of the outcomes of the local measurements. As a result, the stochastic quantum entropy production of the composite system reads $$\sigma_{A - B}(c^{\textrm{in}}_{mh},c^{\textrm{fin}}_{kl}) = \ln\left[\frac{p(c^{\textrm{in}}_{mh})}{p(c^{\textrm{fin}}_{kl})}\right],$$ consistently with the definition in section 2. Under the same hypotheses, we can define similar contributions of the stochastic quantum entropy production separately for each subsystem, [*i.e.*]{} $\sigma_A$ for subsystem $A$ and $\sigma_B$ for subsystem $B$: $$\sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k})=\ln\left[\frac{p(a^{\textrm{in}}_{m})}{p(a^{\textrm{fin}}_{k})}\right],~~~~\text{and}~~~~ \sigma_B(b^{\textrm{in}}_{h},b^{\textrm{fin}}_{l})=\ln\left[\frac{p(b^{\textrm{in}}_{h})}{p(b^{\textrm{fin}}_{l})}\right].$$ If upon measurement the composite system is in a product state, the measurement outcomes for $A$ and $B$ are independent and the probabilities to obtain them factorize as $$\begin{cases} p(c^{\textrm{in}}_{mh}) = p(a^{\textrm{in}}_{m})p(b^{\textrm{in}}_{h}) \\ p(c^{\textrm{fin}}_{kl}) = p(a^{\textrm{fin}}_{k})p(b^{\textrm{fin}}_{l}) \end{cases}.$$ As a direct consequence, the stochastic quantum entropy production becomes an additive quantity: $$\label{eq:sigma_AB} \sigma_{A - B}(c^{\textrm{in}}_{mh},c^{\textrm{fin}}_{kl}) = \sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k}) + \sigma_B(b^{\textrm{in}}_{h},b^{\textrm{fin}}_{l}) \equiv \sigma_{A + B}(c^{\textrm{in}}_{mh},c^{\textrm{fin}}_{kl}).$$ In the more general case of correlated measurement outcomes, instead, the probabilities do not factorize anymore, and Eq. (\[eq:sigma\_AB\]) is not valid anymore. In particular, the mean value of the stochastic entropy production $\sigma_{A - B}(c^{\textrm{in}}_{mh},c^{\textrm{fin}}_{kl})$ becomes sub-additive. In other words $$\label{eq:sigma_sum} \langle\sigma_{A - B}\rangle \leq \langle\sigma_{A}\rangle + \langle\sigma_{B}\rangle \equiv \langle\sigma_{A + B}\rangle,$$ [*i.e.*]{} the mean value of the stochastic quantum entropy production $\sigma_{A - B}$ of the composite system $A - B$ is smaller than the sum of the mean values of the corresponding entropy production of its subsystems, when the latter are correlated. To see this, we recall the expression of the mean value of the stochastic entropy production in terms of the von Neumann entropies of the two post-measurement states (see \[appendix\_b\]): $$\begin{aligned} \langle\sigma_{A - B}\rangle &=& S(\rho_{\tau})-S(\rho_{\textrm{in}}) =S(\rho_{\tau})-S(\rho_{A,\textrm{in}})-S(\rho_{B,\textrm{in}})\\ &\leq& S(\rho_{A,\tau})+S(\rho_{B,\tau})-S(\rho_{A,\textrm{in}})-S(\rho_{B,\textrm{in}})\\ &=& \langle\sigma_{A}\rangle + \langle\sigma_{B}\rangle = \langle\sigma_{A + B}\rangle.\end{aligned}$$ In the following we will analyze the probability distribution of the stochastic quantum entropy productions $\sigma_{A-B}$ for the composite system and $\sigma_A$, $\sigma_B$ for the subsystems. For comparison we compute also $\sigma_{A+B}$. We will show in particular how to calculate the corresponding characteristic functions. In the next section we will then show how these characteristic functions can be measured and how they can be used to reconstruct the probability distributions. Probability Distribution ------------------------ Depending on the values assumed by the measurement outcomes $c^{\textrm{in}}\in\{c^{\textrm{in}}_{mh}\}$ and $c^{\textrm{fin}}\in \{c^{\textrm{fin}}_{kl}\}$, $\sigma_{A - B}$ is a fluctuating variable as it is true also for the single subsystem contributions $\sigma_A\in\{\sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k})\}$ and $\sigma_B\in\{\sigma_B(b^{\textrm{in}}_{h},b^{\textrm{fin}}_{l})\}$. We denote the probability distributions for the subsystems with $\textrm{Prob}(\sigma_A)$ and $\textrm{Prob}(\sigma_B)$ and $\textrm{Prob}(\sigma_{A-B})$ for the composite system. We will further compare this probability distribution for the composite system (containing the correlations of the local measurement outcomes) to the uncorrelated distribution of the sum of the single subsystems’ contributions. We introduce the probability distribution $\textrm{Prob}(\sigma_{A+B})$ of the stochastic quantum entropy production $\sigma_{A+B}$ by applying the following discrete convolution sum: $$\label{eq:convolution} \textrm{Prob}(\sigma_{A+B}) = \sum_{\{\xi_{B}\}}\textrm{Prob}((\sigma_{A+B} - \xi_{B})_{A})\textrm{Prob}(\xi_{B}),$$ where $(\sigma_{A+B} - \xi_{B})_{A}$ and $\xi_{B}$ belong, respectively, to the sample space ([*i.e.*]{} the set of all possible outcomes) of the random variables $\sigma_A$ and $\sigma_B$. The probability distribution for the single subsystem, [*e.g.*]{} the subsystem $A$, is fully determined by the knowledge of the measurement outcomes and the respective probabilities (relative frequencies). We obtain the measurement outcomes $(a^{\textrm{in}}_m,a^{\textrm{fin}}_k)$ with a certain probability $p_{a}(k,m)$, the joint probability for $a^{\textrm{in}}_m$ and $a^{\textrm{fin}}_k$, and this measurement outcome yields the stochastic entropy production $\sigma_A = \sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k})$. Likewise, for system $B$ we introduce the joint probability $p_{b}(l,h)$ to obtain $(b^{\textrm{in}}_h,b^{\textrm{fin}}_l)$, which yields $\sigma_B = \sigma_B(b^{\textrm{in}}_{h},b^{\textrm{fin}}_{l})$. Therefore, the probability distributions $\textrm{Prob}(\sigma_A)$ and $\textrm{Prob}(\sigma_B)$ are given by $$\label{prob_a} \textrm{Prob}(\sigma_A) = \left\langle\delta\left[\sigma_A - \sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k})\right]\right\rangle = \sum_{k,m}\delta\left[\sigma_A - \sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k})\right]p_{a}(k,m)$$ and $$\label{prob_b} \textrm{Prob}(\sigma_B) = \left\langle\delta\left[\sigma_B - \sigma_B(b^{\textrm{in}}_{h},b^{\textrm{fin}}_{l})\right]\right\rangle = \sum_{l,h}\delta\left[\sigma_B - \sigma_B(b^{\textrm{in}}_{h},b^{\textrm{fin}}_{l})\right]p_{b}(l,h),$$ where $\delta[\cdot]$ is the Dirac-delta distribution. In Eqs. (\[prob\_a\]) and (\[prob\_b\]), the joint probabilities $p_{a}(k,m)$ and $p_{b}(l,h)$ read $$\label{joint} \begin{cases} p_{a}(k,m) = \textrm{Tr}\left[(\Pi^{\textrm{fin}}_{A,k}\otimes\mathbbm{1}_{B})\Phi(\Pi^{\textrm{in}}_{A,m}\otimes\rho_{\textrm{B,in}})\right]p(a_{m}^{\textrm{in}}) \\ p_{b}(l,h) = \textrm{Tr}\left[(\mathbbm{1}_{A}\otimes\Pi^{\textrm{fin}}_{B,l})\Phi(\rho_{\textrm{A,in}}\otimes\Pi^{\textrm{in}}_{B,h})\right]p(b_{h}^{\textrm{in}}). \end{cases}$$ By definition, given the reconstructed probability distributions $\textrm{Prob}(\sigma_A)$ and $\textrm{Prob}(\sigma_B)$, the probability $\textrm{Prob}(\sigma_{A+B})$ can be calculated straightforwardly by calculating the convolution of $\textrm{Prob}(\sigma_A)$ and $\textrm{Prob}(\sigma_B)$ according to Eq. (\[eq:convolution\]). Equivalently, the probability distribution $\textrm{Prob}(\sigma_{A - B})$ of the stochastic quantum entropy production of the composite system (containing the correlations between the local measurement outcomes) is given by: $$\textrm{Prob}(\sigma_{A - B}) = \left\langle\delta\left[\sigma_{A - B} - \sigma_{A - B}(c^{\textrm{in}}_{mh},c^{\textrm{fin}}_{kl})\right]\right\rangle =\sum_{mh,kl}\delta\left[\sigma_{A - B} - \sigma_{A - B}(c^{\textrm{in}}_{mh},c^{\textrm{fin}}_{kl})\right]p_{c}(mh,kl),$$ where $$p_{c}(mh,kl) = \textrm{Tr}\left[\Pi^{\textrm{fin}}_{A - B,kl}\Phi\left(\Pi^{\textrm{in}}_{A,m}\otimes\Pi^{\textrm{in}}_{B,h}\right)\right]p(c^{\textrm{in}}_{mh}),$$ with $p(c^{\textrm{in}}_{mh}) = p(a^{\textrm{in}}_{m})p(b^{\textrm{in}}_{h})$. Now, the integral quantum fluctuation theorems for $\sigma_A$, $\sigma_B$ and $\sigma_{A - B}$ can be derived just by computing the characteristic functions of the corresponding probability distributions $\textrm{Prob}(\sigma_A)$, $\textrm{Prob}(\sigma_B)$ and $\textrm{Prob}(\sigma_{A - B})$, as it will be shown below. Characteristic function of the stochastic quantum entropy production and integral fluctuation theorem ----------------------------------------------------------------------------------------------------- In probability theory, the characteristic function of a real-valued random variable is its Fourier transform and completely defines the properties of the corresponding probability distribution in the frequency domain [@Kallenberg2005]. We define the characteristic function $G_C(\lambda)$ of the probability distribution $\textrm{Prob}(\sigma_C)$ (for $C\in \{A,B,A-B\}$) as $$G_C(\lambda) = \int \textrm{Prob}(\sigma_C)e^{i\lambda\sigma_C} d\sigma_C,$$ where $\lambda\in\mathbb{C}$ is a complex number. For the two subsystems, by inserting Eqs. (\[prob\_a\])-(\[joint\]) and exploiting the linearity of the CPTP quantum maps and of the trace, as shown in \[appendix\_charact\_func\], the characteristic functions for $\textrm{Prob}(\sigma_A)$ and $\textrm{Prob}(\sigma_B)$ can be written as $$\label{eq:G_A} G_A(\lambda) = \textrm{Tr}\left\{\left[(\rho_{A,\tau})^{-i\lambda}\otimes\mathbbm{1}_{B}\right]\Phi\left[ (\rho_{\textrm{A,in}})^{1 + i\lambda}\otimes\rho_{\textrm{B,in}}\right]\right\}$$ and $$\label{eq2} G_B(\lambda) = \textrm{Tr}\left\{\left[\mathbbm{1}_{A}\otimes(\rho_{B,\tau})^{-i\lambda}\right]\Phi\left[ \rho_{\textrm{A,in}}\otimes(\rho_{B,\textrm{in}})^{1 + i\lambda}\right]\right\}.$$ In a similar way, we can derive the characteristic function $G_{A - B}(\lambda)$ of the stochastic entropy production of the composite system $A - B$: $$\label{eq:G_A-B} G_{A - B}(\lambda) = \mathrm{Tr}\left[\rho_{\tau}^{-i\lambda}\Phi(\rho_{\mathrm{in}}^{1+i\lambda})\right]\,.$$ Furthermore, if we choose $\lambda = i$, the integral quantum fluctuation theorems can be straightforwardly derived, namely for $\sigma_A$ and $\sigma_B$: $$\label{eq3} \big\langle e^{-\sigma_A}\big\rangle \equiv G_A(i) = \textrm{Tr}\left\{\left[\rho_{A,\tau}\otimes\mathbbm{1}_{B}\right]\Phi\left[ \mathbbm{1}_{A}\otimes\rho_{\textrm{B,in}}\right]\right\}$$ and $$\label{eq4_bis} \big\langle e^{-\sigma_B}\big\rangle \equiv G_B(i) = \textrm{Tr}\left\{\left[\mathbbm{1}_{A}\otimes\rho_{B,\tau}\right]\Phi\left[ \rho_{\textrm{A,in}}\otimes\mathbbm{1}_{B}\right]\right\},$$ as well as $$\label{eq4} \big\langle e^{-\sigma_{A-B}}\big\rangle \equiv G_{A-B}(i) = \textrm{Tr}\left\{\rho_{\tau}\Phi\left[ \mathbbm{1}_{A-B}\right]\right\}=1$$ for $\sigma_{A-B}$ (with $\Phi$ unital). The characteristic functions of Eqs. (\[eq:G\_A\])-(\[eq:G\_A-B\]) depend exclusively on appropriate powers of the initial and final density operators of each subsystem. These density operators are diagonal in the basis of the observable eigenvectors and can be measured by means of standard state population measurements for each value of $\lambda$. As will be shown in the following, this result can lead to a significant reduction of the number of measurements that is required to reconstruct the probability distribution of the stochastic quantum entropy production, beyond the direct application of the definition of Eqs. (\[prob\_a\])-(\[joint\]). A reconstruction algorithm implementing such improvement will be discussed in the next section. Reconstruction algorithm {#VVV} ======================== In this section, we present the algorithm for the reconstruction of the probability distribution $\textrm{Prob}(\sigma)$ for the stochastic quantum entropy production $\sigma$. The procedure is based on a parametric version of the integral quantum fluctuation theorem, [*i.e.*]{} $\langle e^{-\varphi \sigma}\rangle$ ($\varphi\in\mathbb{R}$). In particular, we introduce the moment generating functions $\chi_{C}(\varphi)$ for $C\in\{A,B,A-B\}$: $$\langle e^{-\varphi\sigma_C}\rangle = G_C(i\varphi) \equiv \chi_{C}(\varphi).$$ The quantity $\langle e^{-\varphi\sigma_{C}}\rangle$ can be expanded into a Taylor series, so that $$\chi_{C}(\varphi) = \langle e^{-\varphi\sigma_{C}}\rangle = \left\langle\sum_{k}\frac{(-\varphi^{k})}{k!}\sigma_{C}^{k}\right\rangle = 1 - \varphi\langle\sigma_{C}\rangle + \frac{\varphi^{2}}{2}\langle\sigma_{C}^{2}\rangle - \ldots$$ Accordingly, the statistical moments of the stochastic quantum entropy production $\sigma_{C}$, denoted by $\{\langle\sigma_{C}^{k}\rangle\}$ with $k = 1,\ldots,N-1$, can be expressed in terms of the $\chi_{C}(\varphi)$’s defined over the parameter vector $\underline{\varphi} \equiv (\varphi_{1}, \ldots, \varphi_{N})^T$, [*i.e.*]{} $$\begin{pmatrix} \chi_{C}(\varphi_{1}) \\ \chi_{C}(\varphi_{2}) \\ \vdots \\ \chi_{C}(\varphi_{N}) \end{pmatrix} = \underbrace{\begin{pmatrix} 1 & -\varphi_{1} & +\frac{\varphi_{1}^{2}}{2} & \ldots & \frac{\varphi_{1}^{N-1}}{N-1!} \\ 1 & -\varphi_{2} & +\frac{\varphi_{2}^{2}}{2} & \ldots & \frac{\varphi_{2}^{N-1}}{N-1!} \\ \vdots & \vdots & \vdots & \vdots & \vdots \\ 1 & -\varphi_{N} & +\frac{\varphi_{N}^{2}}{2} & \ldots & \frac{\varphi_{N}^{N-1}}{N-1!} \end{pmatrix}}_{A(\underline{\varphi})} \begin{pmatrix} 1 \\ \langle\sigma_{C}\rangle \\ \langle\sigma_{C}^{2}\rangle \\ \vdots \\ \langle\sigma_{C}^{N-1}\rangle \end{pmatrix},$$ where the matrix $A(\underline{\varphi})$ can be written as a Vandermonde matrix, as detailed below. It is clear at this point that the solution to the problem to infer the set $\{\langle\sigma_{C}^{k}\rangle\}$ can be related to the resolution of a polynomial interpolation problem, where the experimental data-set is given by $N$ evaluations of the parametric integral fluctuation theorem of $\sigma_{C}$ in terms of the $\varphi$’s. Let us observe that only by choosing real values for the parameters $\varphi$ is it possible to set up the proposed reconstruction procedure via the resolution of an interpolation problem. We will explain in the next section a feasible experiment with trapped ions to directly measure the quantities $\chi_{C}(\varphi)$ by properly varying the parameter $\varphi$. By construction, the dimension of the parameters vector $\underline{\varphi}$ is equal to the number of statistical moments of $\sigma_{C}$ that we want to infer, including the trivial zero-order moment. In this regard, we define the vectors $$\widetilde{\underline{m}}\equiv\left(1, \, \, -\langle\sigma_{C}\rangle, \,\,\, \ldots, \,\, (-1)^{N-1}\frac{\langle\sigma_{C}^{N-1}\rangle}{N-1!}\right)^T,$$ with element ${\widetilde{\underline{m}}}_j=(-1)^{j}\frac{\langle\sigma_{C}^{j}\rangle}{j!}$, $j=0, \ldots, N-1$, and $$\underline{\chi}_{C}\equiv(\chi_{C}(\varphi_{1}), \ldots, \chi_{C}(\varphi_{N}))^T.$$ Then one has $$\label{vandermonde} \underline{\chi}_{C} = V(\underline{\varphi})\widetilde{\underline{m}},$$ where $$V(\underline{\varphi}) = \begin{pmatrix} 1 & \varphi_{1} & \varphi_{1}^{2} & \ldots & \varphi_{1}^{N-1} \\ 1 & \varphi_{2} & \varphi_{2}^{2} & \ldots & \varphi_{2}^{N-1} \\ \vdots & \vdots & \vdots & \vdots & \vdots \\ 1 & \varphi_{N} & \varphi_{N}^{2} & \ldots & \varphi_{N}^{N-1} \end{pmatrix}$$ is the Vandermonde matrix built on the parameters vector $\underline{\varphi}$. $V(\underline{\varphi})$ is a matrix whose rows (or columns) have elements in geometric progression, [*i.e.*]{} $v_{ij} = \varphi^{j-1}_{i}$, where $v_{ij}$ denotes the $ij-$ element of $V(\underline{\varphi})$. Eq. (\[vandermonde\]) constitutes the formula for the inference of the statistical moments $\{\langle\sigma_{C}^{k}\rangle\}$ by means of a finite number $N$ of evaluations of $\chi_{C}(\varphi)$. Moreover, given the vector $\underline{m}\equiv(1, \,\, \langle\sigma_{C}\rangle,\ldots,\langle\sigma_{C}^{N-1}\rangle)^T$ of the statistical moments of $\sigma_{C}$, the linear transformation ${\cal T}$, which relates $\widetilde{\underline{m}}$ with $\underline{m}$ such that $\underline{m} = {\cal T} \widetilde{\underline{m}}$, is ${\cal T} = \textrm{diag}\left(\{(-1)^{n}n!\}_{n=0}^{N-1}\right)$, where $\textrm{diag}(\cdot)$ denotes the diagonal matrix. The determinant of the Vandermonde matrix $V(\underline{\varphi})$ is $$\textrm{det}\left[V(\underline{\varphi})\right] = \prod_{1\leq i\leq j\leq N}(\varphi_{j} - \varphi_{i}),$$ given by the product of the differences between all the elements of the vector $\underline{\varphi}$, which are counted only once with their appropriate sign. As a result, $\textrm{det}\left[V(\underline{\varphi})\right] = 0$ if and only if $\underline{\varphi}$ has at least two identical elements. Only in that case, the inverse of $V(\underline{\varphi})$ does not exist and the polynomial interpolation problem cannot be longer solved. However, although the solution of a polynomial interpolation by means of the inversion of the Vandermonde matrix exists and is unique, $V(\underline{\varphi})$ is an ill-conditioned matrix [@Meyer2000]. This means that the matrix is highly sensitive to small variations of the set of the input data (in our case the parameters $\varphi$’s), such that the condition number of the matrix may be large and the matrix becomes singular. As a consequence, the reconstruction procedure will be computationally inefficient, especially in the case the measurements are affected by environmental noise. Numerically stable solutions of a polynomial interpolation problem usually rely on the Newton polynomials [@Trefethen2000]. The latter allow us to write the characteristic function $\chi_{C}(\varphi)$ in polynomial terms as a function of each value of $\underline{\varphi}$: $$\chi^{\textrm{pol}}_{C}(\varphi) = \sum^{N}_{k=1}\eta_{k}n_{k}(\varphi),$$ with $n_{k}(\varphi)\equiv\prod^{k-1}_{j=1}(\varphi - \varphi_{j})$ and $n_{1}(\varphi) = 1$. The coefficients $\eta_{k}$ of the basis polynomials, instead, are given by the divided differences $$\eta_{k} = \left[\chi_{C}(\varphi_{1}),\ldots,\chi_{C}(\varphi_{k})\right] \equiv \frac{\left[\chi_{C}(\varphi_{2}),\ldots,\chi_{C}(\varphi_{k})\right] - \left[\chi_{C}(\varphi_{1}),\ldots,\chi_{C}(\varphi_{k-1})\right]}{(\varphi_{k} - \varphi_{1})},$$ where $\left[\chi_{C}(\varphi_{k})\right]\equiv\chi_{C}(\varphi_{k})$, $\left[\chi_{C}(\varphi_{k-1}),\chi_{C}(\varphi_{k})\right] \equiv ([\chi_{C}(\varphi_{k})] - [\chi_{C}(\varphi_{k-1})])/(\varphi_{k} - \varphi_{k-1}) = (\chi_{C}(\varphi_{k}) - \chi_{C}(\varphi_{k-1}))/(\varphi_{k} - \varphi_{k-1})$, and all the other divided differences found recursively. Then, the natural question arises on what is an optimal choice for $\underline{\varphi}$. It is essential, indeed, to efficiently reconstruct the set $\{\langle\sigma_{C}^{k}\rangle\}$ of the statistical moments of $\sigma_{C}$. For this purpose, we can take into account the error $e_{C}(\varphi) \equiv \chi_{C}(\varphi) - \chi^{\textrm{pol}}_{C}(\varphi)$ in solving the polynomial interpolation problem in correspondence of a value of $\varphi$ different from the interpolating points within the parameter vector $\underline{\varphi}$. The error $e_{C}(\varphi)$ depends on the regularity of the function $\chi_{C}(\varphi)$, and especially on the values assumed by the parameters $\varphi$. As shown in [@Trefethen2000; @Phillips2003], the choice of the $\varphi$’s for which the interpolation error is minimized is given by the real zeros of the Chebyshev polynomial of degree $N$ in the interval $\left[\varphi_{\textrm{min}},\varphi_{\textrm{max}}\right]$, where $\varphi_{\textrm{min}}$ and $\varphi_{\textrm{max}}$ are, respectively, the lower and upper bound of the parameters $\varphi$. Accordingly, the optimal choice for $\underline{\varphi}$ is given by $$\label{cheby} \varphi_{k} = \frac{(\varphi_{\textrm{min}}+\varphi_{\textrm{max}})}{2} + \frac{\varphi_{\textrm{max}}-\varphi_{\textrm{min}}}{2}\cos\left(\frac{2k-1}{2N}\pi\right),$$ with $k = 1,\ldots,N$. Let us observe that the value of $N$, [*i.e.*]{} the number of evaluations of the characteristic function $\chi_{C}(\varphi)$, is equal to the number of statistical moments of $\sigma_{C}$ we want to infer. Therefore, in principle, if the probability distribution of the stochastic quantum entropy production is a Gaussian function, then $N$ could be taken equal to $2$. Hence, once all the evaluations of the characteristic functions $\chi_{C}(\varphi)$ have been collected, we can derive the statistical moments of the quantum entropy production $\sigma_{C}$, and consequently reconstruct the probability distribution $\textrm{Prob}(\sigma_{C})$ as $$\label{Fourier} \textrm{Prob}(\sigma_{C}) \approx \mathcal{F}^{-1}\left\{\sum_{k = 0}^{N-1}\frac{\langle\sigma_{C}^{k}\rangle}{k!}(i\mu)^{k}\right\} \equiv\frac{1}{2\pi}\int^{\infty}_{-\infty}\left[\sum_{k = 0}^{N-1}\frac{\langle\sigma_{C}^{k}\rangle}{k!}(i\mu)^{k}\right]e^{-i\mu\sigma_{C}}d\mu,$$ where $\mu\in\mathbb{R}$ and $\mathcal{F}^{-1}$ denotes the inverse Fourier transform [@Mnatsakanov], which is numerically performed [@Athanassoulis]. To do that, we fix a-priori the integration step $d\mu$ and we vary the integration limits of the integral, in order to minimize the error $\sum_{k}\left|\widetilde{\langle\sigma_{C}^{k}\rangle} - \overline{\langle\sigma_{C}^{k}\rangle}\right|^{2}$ between the statistical moments $\widetilde{\langle\sigma_{C}^{k}\rangle}$, obtained by measuring the characteristic functions $\chi_{C}(\varphi)$ (i.e. after the inversion of the Vandermonde matrix), and the ones calculated from the reconstructed probability distribution, $\overline{\langle\sigma_{C}^{k}\rangle}$, which we derive by numerically computing the inverse Fourier transform for each value of $\sigma_{C}$. This procedure has to be done separately for $C\in\{A,B,A-B\}$, while, as mentioned, the probability distribution $\textrm{Prob}(\sigma_{A+B})$ is obtained by a convolution of $\textrm{Prob}(\sigma_{A})$ and $\textrm{Prob}(\sigma_{B})$. Here, it is worth observing that Eq. (\[Fourier\]) provides an approximate expression for the probability distribution $\textrm{Prob}(\sigma_{C})$. Ideally, given a generic unital quantum CPTP map modeling the dynamics of the system, an infinite number $N$ of statistical moment of $\sigma_{C}$ is required to reconstruct $\textrm{Prob}(\sigma_{C})$ if we use the inverse Fourier transform as in Eq. (\[Fourier\]).  While we can always calculate the Fourier transform to reconstruct the probability distribution from its moments, in the case of a distribution with discrete support (as in our case), there is a different method that can lead to higher precision, especially when the moment generating function is not approximated very well by the chosen number $N$ of extracted moments. As a matter of fact, each statistical moment $\widetilde{\langle\sigma_{C}^{k}\rangle}$, with $C\in\{A,B,A-B\}$, is the best approximation of the true statistical moments of $\sigma_{C}$ from the measurement of the corresponding characteristic functions $\chi_C(\varphi)$. Hence, apart from a numerical error coming from the inversion of the Vandermonde matrix $A$ or the use of the Newton polynomials $\chi^{\textrm{pol}}_{C}$, we can state that $$\label{expansion_sigma_C} \widetilde{\langle\sigma_{C}^{k}\rangle} \simeq \sum_{i = 1}^{M_{C}}\sigma^{k}_{C,i}\textrm{Prob}(\sigma_{C,i}) = \sigma_{C,1}^{k}\textrm{Prob}(\sigma_{C,1}) + \ldots + \sigma_{C,M_{C}}^{k}\textrm{Prob}(\sigma_{C,M_{C}}),$$ with $k = 1,\ldots,N$. In Eq. (\[expansion\_sigma\_C\]), $M_{C}$ is equal to the number of values that can be assumed by $\sigma_{C}$, while $\sigma_{C,i}$ denotes the $i-$th possible value for the stochastic quantum entropy production of the (sub)system $C$. As a result, the probabilities $\textrm{Prob}(\sigma_{C,i})$, $i = 1,\ldots,M$, can be approximately expressed as a function of the statistical moments $\left\{\widetilde{\langle\sigma_{C}^{k}\rangle}\right\}$, [*i.e.*]{} $$\begin{pmatrix} \widetilde{\langle\sigma_{C}\rangle} \\ \widetilde{\langle\sigma_{C}^{2}\rangle} \\ \vdots \\ \widetilde{\langle\sigma_{C}^{N}\rangle} \end{pmatrix} = \underbrace{\begin{pmatrix} \sigma_{C,1} & \sigma_{C,2} & \ldots & \sigma_{C,M} \\ \sigma_{C,1}^{2} & \sigma_{C,2}^{2} & \ldots & \sigma_{C,M}^{2} \\ \vdots & \vdots & \vdots & \vdots \\ \sigma_{C,1}^{M} & \sigma_{C,2}^{M} & \ldots & \sigma_{C,N}^{M} \end{pmatrix}}_{\Sigma_{C}} \begin{pmatrix} \textrm{Prob}(\sigma_{C,1}) \\ \textrm{Prob}(\sigma_{C,2}) \\ \vdots \\ \textrm{Prob}(\sigma_{C,M}) \end{pmatrix},$$ where $\Sigma_{C}\in\mathbb{R}^{N\times M}$. By construction $\Sigma_{C}$ is a rectangular matrix, that is computed by starting from the knowledge of the values assumed by the stochastic quantum entropy production $\sigma_{C,i}$. Finally, in order to obtain the probabilities $\textrm{Prob}(\sigma_{C,i})$, $i = 1,\ldots,M_C$, we have to adopt the Moore-Penrose pseudo-inverse of $\Sigma_{C}$, which is defined as $$\Sigma^{+}_{C}\equiv (\Sigma^{T}_{C}\Sigma_{C})^{-1}\Sigma^{T}_{C}.$$ A pictorial representation of the reconstruction protocol is shown in Fig. \[fig:procedure\]. Let us observe, again, that the proposed algorithm is based on the expression of Eq. (\[sigma\]) for the stochastic quantum entropy production, which has been obtained by assuming unital CPTP quantum maps for the system dynamics. We expect that for a general open quantum system, not necessarily described by a unital CPTP map, one can extend the proposed reconstruction protocol, even though possibly at the price of a greater number of measurements. Notice that, since Eq. (\[equality\_cond\_prob\]) it is not longer valid in the general case, one has to use directly Eqs. (\[general\_sigma\])-(\[general\_mean\_sigma\]). However, we observe that, as shown in [@Manzano2015], the ratio between the conditional probabilities may admit for a large family of CPTP maps the form $p(a^{\textrm{fin}}_{k}|a^{\textrm{in}}_{m}) / p(a^{\textrm{in}}_{m}|a^{\textrm{ref}}_{k}) \equiv e^{-\Delta V}$, where the quantity $\Delta V$ is related to the so-called nonequilibrium potential, so that $\sigma=\sigma_{unital}+V$ and $\sigma_{unital}$ again given by Eq. (\[sigma\]). Required number of measurements ------------------------------- From an operational point of view, we need to measure (directly or indirectly) the quantities $$\begin{cases} \chi_{A}(\alpha) = \textrm{Tr}\left\{\left[(\rho_{A,\tau})^{\alpha}\otimes\mathbbm{1}_{B}\right]\Phi\left[ (\rho_{\textrm{A,in}})^{1 - \alpha}\otimes\rho_{\textrm{B,in}}\right]\right\} \\ \chi_{B}(\beta) = \textrm{Tr}\left\{\left[\mathbbm{1}_{A}\otimes(\rho_{B,\tau})^{\beta}\right]\Phi\left[ \rho_{\textrm{A,in}}\otimes(\rho_{\textrm{B,in}})^{1 - \beta}\right]\right\}\\ \chi_{A-B}(\gamma) = \textrm{Tr}\left\{(\rho_{\tau})^{\gamma}\Phi\left[ (\rho_{\textrm{in}})^{1 - \gamma}\right]\right\} \end{cases},$$ [*i.e.*]{} the moment generating functions of $\sigma_A$, $\sigma_B$ and $\sigma_{A-B}$, after a proper choice of the parameters $\alpha$, $\beta$ and $\gamma$, with $\alpha,\beta, \gamma \in\mathbb{R}$. The optimal choice for these parameters was analyzed in the previous section. For this purpose, as shown in \[appendix\_charact\_func\], it is worth mentioning that $\left(\rho_{C,\textrm{in}}\right)^{1-\varphi}\equiv\sum_{m}\Pi^{\textrm{in}}_{C,m}p(x_{m}^{\textrm{in}})^{1 - \varphi}$ and $\left(\rho_{C,\tau}\right)^{\varphi}\equiv\sum_{k}\Pi^{\tau}_{C,k}p(x^{\tau}_{k})^{\varphi}$, where $C\in\{A,B,A - B\}$, $x\in\{a,b,c\}$ and $\varphi\in\{\alpha,\beta,\gamma\}$. A direct measurement of $\chi_{C}(\varphi)$, based for example on an interferometric setting as shown in Ref. [@MazzolaPRL2013; @DornerPRL2013] for the work distribution inference, is not trivial, especially for the general fully quantum case. For this reason, we propose a procedure, suitable for experimental implementation, requiring a limited number of measurements, based on the following steps: 1. Prepare the initial product state $\rho_{\textrm{in}}=\rho_{A,\textrm{in}}\otimes\rho_{B,\textrm{in}}$, as given in Eq. (\[rho\_in\]), with fixed probabilities $p(a^{\textrm{in}}_{m})$ and $p(b^{\textrm{in}}_{h})$. Then, after the composite system $A - B$ is evolved within the time interval $[0,\tau]$, measure the occupation probabilities $p(a^{\textrm{fin}}_{k})$ and $p(b^{\textrm{fin}}_{l})$ via local measurements on $A$ and $B$. Then, compute the stochastic quantum entropy productions $\sigma_{A}(a_m^{\textrm{in}},a_k^{\textrm{fin}})$ and $\sigma_{B}(b_h^{\textrm{in}},b_l^{\textrm{fin}})$. Simultaneous measurements on $A$ and $B$ yield also the probabilities $p(c_{kl}^{\textrm{fin}})$ and thus $\sigma_{A-B}(c_{mh}^{\textrm{in}},c_{kl}^{\textrm{fin}})$. 2. For every chosen value of $\alpha$, $\beta$ and $\gamma$, prepare, for instance by quantum optimal control tools [@DoriaPRL2011], the quantum subsystems in the states $$\begin{cases} \displaystyle{\rho_{\textrm{IN}}(\alpha)\equiv\frac{\left[(\rho_{A,\textrm{in}})^{1 - \alpha}\otimes\rho_{B,\textrm{in}}\right]}{\textrm{Tr}\left[(\rho_{A,\textrm{in}})^{1 - \alpha}\otimes\rho_{B,\textrm{in}}\right]}} \\ \displaystyle{\rho_{\textrm{IN}}(\beta)\equiv\frac{\left[\rho_{\textrm{A,in}}\otimes(\rho_{B,\textrm{in}})^{1 - \beta}\right]}{\textrm{Tr}\left[\rho_{\textrm{A,in}}\otimes(\rho_{B,\textrm{in}})^{1 - \beta}\right]}}\\ \displaystyle{\rho_{\textrm{IN}}(\gamma)\equiv\frac{\left(\rho_{\textrm{A,in}}\otimes\rho_{B,\textrm{in}}\right)^{1 - \gamma}}{\textrm{Tr}\left[\left(\rho_{\textrm{A,in}}\otimes\rho_{B,\textrm{in}}\right)^{1 - \gamma}\right]}} \end{cases},$$ and let the system evolve. 3. Since the characteristic function $\chi_{C}(\varphi)$, with $C\in\{A,B,A-B\}$ and $\varphi\in\{\alpha,\beta,\gamma\}$, is given by performing a trace operation with respect to the composite system $A - B$, one can write the following simplified relation: $$\chi_{C}(\varphi) = \sum_{k}\sum_{m}\langle m|p(x^{\textrm{fin}}_{k})^{\varphi}|k\rangle\langle k|\rho_{\textrm{FIN}}(\varphi)|m\rangle = \sum_{m}p(x^{\textrm{fin}}_{m})^{\varphi}\langle m|\rho_{\textrm{FIN}}(\varphi)|m\rangle,$$ where $\{|l\rangle\}$, $l = m,k$, is the orthonormal basis of the composite system $A - B$, $x\in\{a,b,c\}$ and $\rho_{\textrm{FIN}}(\varphi)\equiv\Phi[\rho_{\textrm{IN}}(\varphi)]$ (with $p(x^{\textrm{fin}}_{m})$ measured in step 1 and $\rho_{\textrm{IN}}(\varphi)$ introduced in step 2). Thus, measure the occupation probabilities $\langle m|\rho_{\textrm{FIN}}(\varphi)|m\rangle$ in order to obtain all the characteristic functions $\chi_{C}(\varphi)$. We observe that the measure of the characteristic functions $\chi_{C}(\varphi)$ relies only on the measure of occupation probabilities. Hence, the proposed procedure does not require any tomographic measurement. Moreover, for the three steps of the protocol we can well quantify the required number of measurements to properly infer the statistics of the quantum entropy production regarding the composite quantum system. The required number of measurements, indeed, scales linearly with the number of possible measurement outcomes coming from each quantum subsystem at the initial and final stages of the protocol. Equivalently, if we define $d_{A}$ and $d_{B}$ as the dimension of the Hilbert space concerning the quantum subsystems $A$ and $B$, we can state that the number of measurements for both of the three steps scales linearly with $(d_{A} + d_{B})$, [*i.e.*]{} with the number of values $(M_{A} + M_{B})$ that can be assumed by $\sigma_{A}$ and $\sigma_{B}$, the stochastic quantum entropy production of the subsystems. It also scale linearly with $M_A M_B$ for the reconstruction of the stochastic quantum entropy production $\sigma_{A-B}$ of the composite system. The reason is that the described procedure is able to reconstruct the distribution of the stochastic quantum entropy production, without directly measuring the joint probabilities $p_a(k,m)$ and $p_{b}(l,h)$ for the two subsystems and $p_c(mh,kl)$ for the composite system. Otherwise, the number of required measurements would scale, respectively, as $M_{A}^{2}$ and $M_{B}^{2}$ for the subsystems and as $(M_AM_B)^2$ for the composite system in order to realize all the combinatorics concerning the measurement outcomes. A physical example ================== ![**Experimental implementation with trapped ions -** Pictorial representation of two trapped ions subjected to two laser fields. The internal levels of the ions allow to encode one qubit in each ion. The transition between these levels is driven by the lasers, where the driving depends on the state of the common vibrational (trap) mode of the two ions. The lasers can be focused to choose between single or global addressing. This allows to generate local gates as well as entangling gates.[]{data-label="fig:fig2"}](Fig2.pdf) In the previous section, we have introduced an algorithm for the reconstruction of thermodynamical quantities in a fully quantum regime. Here, in order to illustrate our theoretical results, we discuss in this section an experimental implementation with trapped ions. Trapped ions have been demonstrated to be a versatile tool for quantum simulation [@Friedenauer2008; @Kim2010; @Lanyon2011; @GerritsmaNAT2010], including simulation of quantum thermodynamics [@Huber2008; @An2015; @AbahPRL2012; @RossnagelPRL2014; @Rossnagel325]. The application of our protocol on a physical example relies on the availability of experimental procedures for state preparation and readout, as well as an entangling operation. We consider a system of two trapped ions, whose two internal states allow to encode the qubit states $|0\rangle$ and $|1\rangle$ of the standard computational basis. Then, the subsystems $A$ and $B$ are represented by the two qubits. The latter can interact by the common vibrational (trap) mode of the two ions, and external lasers allow to manipulate the ion states, generating arbitrary single qubit rotations through individual addressing or an entangling operation, as for example the Mølmer-Sørensen gate operation [@SoerensenPRL82; @RoosNJP2008; @MonzPRL106; @Nigg302]. Fig. \[fig:fig2\] shows a pictorial representation of the system. While usually universal state preparation for single qubits is supposed only for pure states, here we have to prepare mixed states. However, once we have prepared a pure state with the right amount of population in the two levels, we can reach the required mixed state by applying a random $Z$ rotation leading to a complete dephasing of the two levels, where $Z$ is the corresponding Pauli matrix. The two-qubit operation, that generates entanglement between $A$ and $B$, is chosen to be a partial Mølmer-Sørensen gate operation, given by the following unitary operation, depending on the phase $\phi$: $$\label{propagator} \mathcal{U}(\phi)=e^{-i\phi\left(X^{A}\otimes X^{B}\right)},$$ where $X^{A}$ and $X^{B}$ equal, respectively, to the Pauli matrix $X$ for the quantum systems $A$ and $B$, and the reduced Planck’s constant $\hbar$ is set to unity. In the following (and unless explicitly stated otherwise), we choose $\phi=\frac{\pi}{7}$, and start from the initial state $\rho_{0} = \textrm{diag}\left(\frac{6}{25},\frac{9}{25},\frac{4}{25},\frac{6}{25}\right)$ since this choice leads to a non-Gaussian probability distribution $\textrm{Prob}(\sigma_{A - B})$ of the stochastic quantum entropy production. For the sake of simplicity, we remove the label $A$ and $B$ from the computational basis $\{|0\rangle,|1\rangle\}$ considered for the two subsystems. Thus, the corresponding projectors are $\Pi_{0}\equiv|0\rangle\langle 0|$ and $\Pi_{1}\equiv|1\rangle\langle 1|$, and each ion is characterized by $4$ different values of the stochastic quantum entropy production $\sigma_{C}$, with $C\in\{A,B\}$. As a consequence, the probability distribution $\textrm{Prob}(\sigma_{A - B})$ of the stochastic quantum entropy production for the composite system $A - B$ is defined over a discrete support given by $l$ samples, with $l \leq M_{A}M_{B} = 16$. Correlated measurement outcomes and correlations witness -------------------------------------------------------- In the general case, the outcomes of the second measurement of the protocol are correlated, as in our example, and the stochastic quantum entropy production of the composite system is sub-additive, [*i.e.*]{} $\langle\sigma_{A-B}\rangle\leq\langle\sigma_{A}\rangle+\langle\sigma_{B}\rangle$. Hence, by adopting the reconstruction algorithm proposed in Fig. \[fig:procedure\] we are able to effectively derive the upper bound of $\langle\sigma_{A - B}\rangle$, which defines the thermodynamic irreversibility for the quantum process. In the simulations of this section, we compare the fluctuation profile that we have derived by performing local measurements on the subsystems $A$ and $B$ with the ones that are obtained via a global measurement on the composite system $A - B$, in order to establish the amount of information which is carried by a set of local measurements. Furthermore, we will discuss the changes of the fluctuation profile of the stochastic quantum entropy production both for unitary and noisy dynamics. The unitary operation describing the dynamics of the quantum system is given by Eq. (\[propagator\]), while the noisy dynamics can be described by the differential Lindblad (Markovian) equation $\dot{\rho}(t) = \mathcal{L}(\rho(t))$, defined as $$\label{Lindblad} \dot{\rho}(t) = -i\left[H,\rho\right] - \sum_{C\in\{A,B\}}\Gamma_{C}\left(\{\rho,L_{C}^{\dagger}L_{C}\} - 2L_{C}\rho L_{C}^{\dagger}\right).$$ In Eq. (\[Lindblad\]), $\rho(t)$ denotes the density matrix describing the composite quantum system $A - B$, $\{\cdot,\cdot\}$ is the anticommutator, $\Gamma_A$ and $\Gamma_B$ (rad/s) are dephasing rates corresponding to $L_{A}\equiv\Pi_{0}\otimes\mathbbm{1}_{B}$ and $L_{B}\equiv\mathbbm{1}_{A}\otimes\Pi_{0}$, pure-dephasing Lindblad operators, where $\mathbbm{1}_{A}$ and $\mathbbm{1}_{B}$ are the identity operators acting, respectively, on the Hilbert spaces of the ions $A$ and $B$. The Hamiltonian of the composite system $A - B$ in Eq. (\[Lindblad\]), instead, is given by $$H = \omega\left(X^{A}\otimes X^{B}\right),$$ where the interaction strength $\omega = \phi/\tau$ (rad/s) with $\tau$ kept fixed and chosen equal to $50$ s (leading to a largely relaxed system dynamics), consistently with the unitary operation (\[propagator\]).   In Figs. \[fig:fig3\] and \[fig:fig4\], we plot the first $4$ statistical moments of $\sigma_{A - B}$ and $\sigma_{A + B}$ as a function of the phase $\phi$, respectively, in case of unitary and noisy dynamics. Moreover, we show, for a given value of $\phi$, the probability distributions $\textrm{Prob}(\sigma_{A - B})$ and $\textrm{Prob}(\sigma_{A + B})$ for both unitary and noisy dynamics, compared with the corresponding reconstructed distributions obtained by applying the reconstruction algorithm, which we call [$\overline{\textrm{Prob}(\sigma_{A - B})}$]{} and [$\overline{\textrm{Prob}(\sigma_{A + B})}$]{}, respectively. Let us recall that $\textrm{Prob}(\sigma_{A + B})$ is obtained by performing the two local measurements with observables $\mathcal{O}^{\textrm{fin}}_{A}$ and $\mathcal{O}^{\textrm{fin}}_{B}$ independently (disregarding the correlations of their outcomes) on the subsystems $A$, $B$, while the distribution $\textrm{Prob}(\sigma_{A - B})$ requires to measure $\mathcal{O}^{\textrm{fin}}_{A}$ and $\mathcal{O}^{\textrm{fin}}_{B}$ simultaneously, [*i.e.*]{} measuring the observable $\mathcal{O}^{\textrm{fin}}_{A - B}$, defined by Eq. (\[eq:ovservable\_A-B\]). For unitary dynamics, the statistical moments of the stochastic quantum entropy productions $\sigma_{A - B}$ and $\sigma_{A + B}$ follow the oscillations of the dynamics induced by changing the gate phase $\phi$. Conversely, for the noisy dynamics of Eq. (\[Lindblad\]), with $\Gamma = \Gamma_{A} = \Gamma_{B}$ different from zero, when $\phi$ increases the system approaches a fixed point of the dynamics. Consequently, the statistical moments of the stochastic quantum entropy production tend to the constant values corresponding to the fixed point, and the distribution of the stochastic entropy production becomes narrower. In both Figs. \[fig:fig3\] and \[fig:fig4\], the first statistical moments (or mean values) $\langle\sigma_{A - B}\rangle$ and $\langle\sigma_{A + B}\rangle$ are almost overlapping, and the sub-additivity of $\sigma_{A - B}$ is confirmed by the numerical simulations. Furthermore, quite surprisingly, also the second statistical moments of $\sigma_{A - B}$ and $\sigma_{A + B}$ are very similar to each other.  This means that the fluctuation profile of the stochastic entropy production $\sigma_{A + B}$ is able to well reproduce the probability distribution of $\sigma_{A - B}$ in its Gaussian approximation, [*i.e.*]{} according to the corresponding first and second statistical moments. In addition, we can state that the difference of the higher order moments of $\langle\sigma_{A+B}\rangle$ and $\langle\sigma_{A-B}\rangle$ reflects the presence of correlations between $A$ and $B$ created by the map, since for a product state $\sigma_{A - B} = \sigma_{A+B}$. Therefore, the difference between the fluctuation profiles of $\sigma_{A - B}$ and $\sigma_{A + B}$ constitutes a witness for classical and/or quantum correlations in the final state of the system before the second measurement. As a consequence, if $\textrm{Prob}(\sigma_{A - B})$ and $\textrm{Prob}(\sigma_{A + B})$ are *not* identically equal, then the final density matrix $\rho_{\textrm{fin}}$ is *not* a product state, and (classical and/or quantum) correlations are surely present. Notice that the converse statement is not necessarily true because the quantum correlations can be partially or fully destroyed by the second local measurements, while the classical ones are still preserved and so detectable. In Fig. \[fig:fig5\] the first $4$ statistical moments of $\sigma_{A - B}$ and $\sigma_{A + B}$ are shown as a function of $\Gamma$ (rad/s). As before, we can observe a perfect correspondence between the two quantities when we consider only the first and second statistical moments of the stochastic quantum entropy productions, and, in addition, similar behaviour for the third and fourth statistical moments. Indeed, since the coherence terms of the density matrix describing the dynamics of the composite quantum system tend to zero for increasing $\Gamma$, the number of samples of $\sigma_{A - B}$ and $\sigma_{A + B}$ with an almost zero probability to occur is larger, and also the corresponding probability distribution approaches to a Gaussian one, with zero mean and small variance. In accordance with Figs. \[fig:fig3\] and \[fig:fig4\], this result confirms the dominance of decoherence in the quantum system dynamics, which coincides with no creation of correlations. In the following subsection, we will evaluate the performance of the proposed reconstruction algorithm for the reconstruction of $\textrm{Prob}(\sigma_{A})$ and $\textrm{Prob}(\sigma_{B})$. This choice is largely justified also by the possibility to characterize the irreversibility of an arbitrary quantum process, given by the mean value of the stochastic quantum entropy production $\sigma_{A - B}$, via the reconstruction of the corresponding upper bound in accordance with the sub-additivity property. Still, a similar behaviour was found for the probability distribution $\textrm{Prob}(\sigma_{A-B})$ of the stochastic quantum entropy production of the composite system. Reconstruction for unitary dynamics ----------------------------------- In this section, we show the performance of the reconstruction algorithm for the probability distribution of the stochastic quantum entropy production $\sigma_{A+B}$ via local measurements on the subsystems $A$ and $B$, when the dynamics of the quantum system is unitary. In particular, in the numerical simulations, we take the parameters $\alpha$ and $\beta$ of the algorithm, respectively, equal to the real zeros of the Chebyshev polynomial of degree $N$ in the intervals $\left[\alpha_{\textrm{min}},\alpha_{\textrm{max}}\right] = [0,N]$ and $\left[\beta_{\textrm{min}},\beta_{\textrm{max}}\right] = [0,N]$. This choice for the minimum and maximum values of the parameters $\alpha$ and $\beta$ ensures a very small numerical error (about $10^{-4}$) in the evaluation of each statistical moment of $\sigma_{A}$ and $\sigma_{B}$ via the inversion of the Vandermonde matrix, already for $N > 2$. Indeed, since all the elements of the vectors $\underline{\alpha}$ and $\underline{\beta}$ are different from each other, [*i.e.*]{} $\alpha_{i}\neq\alpha_{j}$ and $\beta_{i}\neq\beta_{j}$ $\forall i,j = 1,\ldots,N$, we can derive the statistical moments of $\sigma_{C}$, with $C\in\{A,B\}$, by inverting the corresponding Vandermonde matrix. The number $N$ of evaluations of the moment generating functions $\chi_{A}(\alpha)$ and $\chi_{B}(\beta)$, instead, has been taken as a free parameter in the numerics in order to analyze the performance of the reconstruction algorithm. The latter may be quantified in terms of the Root Mean Square Error (RMSE) defined as $$\label{RMSE_m} \textrm{RMSE}\left(\{\langle\sigma^{k}_{A+B}\rangle\}_{k = 1}^{N_{\textrm{max}}}\right)\equiv\sqrt{\frac{\displaystyle{\sum_{k = 1}^{N_{\textrm{max}}} \left|\langle\sigma^{k}_{A+B}\rangle - \overline{\langle\sigma^{k}_{A+B}\rangle}\right|^{2}}}{N_{\textrm{max}}}},$$ where $\{\langle\sigma^{k}_{A+B}\rangle\}$ are the true statistical moments of the stochastic quantum entropy production $\sigma_{A+B}$, which have been numerically computed by directly using Eqs. (\[eq:convolution\])-(\[prob\_b\]), while $\overline{\langle\sigma^{k}_{A+B}\rangle}$ are the reconstructed statistical moments after the application of the inverse Fourier transform or the Moore-Penrose pseudo-inverse of $\Sigma_{C}$, $C\in\{A,B\}$. $N_{\textrm{max}}$, instead, is the largest value of $N$ considered for the computation of the $\textrm{RMSE}\left(\{\langle\sigma^{k}_{A+B}\rangle\}\right)$ in the numerical simulations (in this example $N_{\textrm{max}} = 16$).  Another measure for the evaluation of the algorithm performance, which will be used hereafter, is given by the RMSE $$\label{RMSE_p} \textrm{RMSE}\left(\{\textrm{Prob}(\sigma_{A+B,i})\}_{i = 1}^{l}\right)\equiv\sqrt{\displaystyle{\frac{\sum_{i = 1}^{l}R_{i}^{2}}{l}}},$$ where $R_{i} \equiv \left|\textrm{Prob}(\sigma_{A+B,i}) - \overline{\textrm{Prob}(\sigma_{A+B,i})}\right|$ is the reconstruction deviation, [*i.e.*]{} the discrepancy between the true and the reconstructed probability distribution $\textrm{Prob}(\sigma_{A+B})$. The $\textrm{RMSE}\left(\{\textrm{Prob}(\sigma_{A+B,i})\}\right)$ is computed with respect to the reconstructed values [$\overline{\textrm{Prob}(\sigma_{A+B,i})}$]{} of the probabilities $\textrm{Prob}(\sigma_{A+B,i})$, $i = 1,\ldots,l$, for the stochastic quantum entropy production $\sigma_{A+B}$. Fig. \[fig:fig6\] shows the performance of the reconstruction algorithm as a function of $N$ for the proposed experimental implementation with trapped ions in case the system dynamics undergoes a unitary evolution. In particular, we show the first $4$ statistical moments of $\sigma_{A}$, $\sigma_{B}$ and $\sigma_{A+B}$ as a function of $N$. Let us observe that the statistical moments of the stochastic quantum entropy production of the two subsystems $A$ and $B$ are equal due to the symmetric structure of the composite system. As expected, when $N$ increases, the reconstructed statistical moments converge to the corresponding true values, and also the reconstruction deviation tends to zero. This result is encoded in the RMSEs of Eqs. (\[RMSE\_m\])-(\[RMSE\_p\]), which behave as monotonically decreasing functions. Both the $\textrm{RMSE}\left(\{\langle\sigma^{k}_{A+B}\rangle\}\right)$ and $\textrm{RMSE}\left(\{\textrm{Prob}(\sigma_{A+B,i})\}\right)$ sharply decrease for about $N \geq 6$, implying that the reconstructed probability distribution [$\overline{\textrm{Prob}(\sigma_{A+B})}$]{} overlaps with the true distribution $\textrm{Prob}(\sigma_{A+B})$ with very small reconstruction deviations $R_{i}$. Since the system of two trapped ions of this example is a small size system, we have chosen to derive the probabilities $\{\textrm{Prob}(\sigma_{A,i})\}$ and $\{\textrm{Prob}(\sigma_{B,i})\}$, $i = 1,\ldots,4$, without performing the inverse Fourier transform on the statistical moments $\{\widetilde{\langle\sigma_{C}^{k}\rangle}\}$, $C\in\{A,B\}$. Indeed, the computation of the inverse Fourier transform, which has to be performed numerically, can be a tricky step of the reconstruction procedure, because it can require the adoption of numerical methods with an adaptive step-size in order to solve the numerical integration. In this way, the only source of error in the reconstruction procedure is given by the expansion in Taylor series of the quantity $\chi_{C}(\varphi)$, with $C\in\{A,B\}$ and $\varphi\in\{\alpha,\beta\}$, around $\varphi = 0$ as a function of a *finite* number of statistical moments $\langle\sigma_{C}^{k}\rangle$, $k = 1,\ldots,N-1$. As shown in Fig. \[fig:fig6\], the choice of the value of $N$ is a degree of freedom of the algorithm, and it strictly depends on the physical implementation of the reconstruction protocol. In the experimental implementation above with two trapped ions, $N = 10$ ensures very good performance without making a larger number of measurements with respect to the number of values assumed by the stochastic quantum entropy production $\sigma_{A+B}$. In Fig. \[fig:Fig7\], moreover, we show for $N = 10$ the first $4$ true statistical moments of the stochastic quantum entropy productions $\sigma_{A}$ and $\sigma_{B}$ of the two subsystems, as well as the correlation-free convolution $\sigma_{A+B}$ as a function of $\phi\in[0,2\pi]$, along with the corresponding reconstructed counterpart [$\overline{\langle\sigma^{k}_C\rangle}$]{}, $k = 1,\ldots,4$, $C\in\{A,B,A+B\}$. As before, the reconstruction procedure yields values very close to the true statistical moments of $\sigma_{A}$, $\sigma_{B}$ and $\sigma_{A+B}$ for all values of the phase $\phi$. Reconstruction for noisy dynamics --------------------------------- Here, the performance of the reconstruction algorithm is discussed in case the system dynamics is affected by pure-dephasing contributions, described via the differential Lindblad (Markovian) equation $\dot{\rho}(t) = \mathcal{L}(\rho(t))$, given by Eq. (\[Lindblad\]). The Hamiltonian of the composite system $A - B$ is defined as $H = \omega\left(X^{A}\otimes X^{B}\right)$. Since the fixed duration $\tau$ of the transformation has been chosen as before equal to $50$ s, we choose the desired phase $\phi$ by setting the interaction strength to $\omega \equiv \phi/\tau$ (rad/s). Fig \[fig:entropy\_Lindblad\] shows the $\textrm{RMSE}\left(\{\langle\sigma^{k}_{A+B}\rangle\}\right)$ computed from the reconstructed statistical moments of the stochastic quantum entropy production $\sigma_{A+B}$. As it can be observed, apart from an initial transient, the mean error monotonically tends to zero as $N$ increases, similarly to the case of unitary dynamics (see Fig. \[fig:fig6\]), such that it can be considered sufficiently small for $N > 8$. Again, we evaluate the performance of the reconstruction algorithm also as a function of the phase $\phi = \omega\tau$, with $\tau$ fixed. As shown in Fig. \[fig:Fig9\], when $\phi$ increases (with a fixed value of $\Gamma$, set to $0.2$ rad/s) the statistical moments of $\sigma_{A+B}$ (but not necessarily the ones regarding the subsystems $A$ and $B$) increase as well, since when $\phi$ increases the system tends to a fixed point of the dynamics. Also the reconstruction procedure turns out to be more accurate for larger values of $\phi$, as shown in the two bottom panels of Fig. \[fig:Fig9\] (for this figure we use the Fourier transform). The reason is that when the dynamics approaches the fixed point, the distribution of the stochastic quantum entropy production becomes narrower and the convergence of the Fourier integral is ensured. Finally, in Fig.\[fig:Fig10\] we plot the behaviour of the first three statistical moments of $\sigma_{A}$, $\sigma_{B}$ and $\sigma_{A+B}$ as a function of the dephasing rate $\Gamma = \Gamma_{A} = \Gamma_{B}$ (rad/s), with $N = 10$ and $\phi = \pi/7$. As before, due to the symmetry of the bipartition, the statistical moments of $\sigma_{A}$ and $\sigma_{B}$ are identically equal. For increasing $\Gamma$ the dephasing becomes dominant over the interaction and all correlations between the subsystems are destroyed. As a consequence, the stochastic quantum entropy production tends to zero as is observed in the figure for all the investigated moments, both for the subsystems and the composite system. Probing irreversibility ----------------------- Once the fluctuation profile of the stochastic quantum entropy production ([*i.e.*]{} the corresponding probability distribution) is reconstructed, then the irreversibility properties of the composite system transformation can be successfully probed. The thermodynamic irreversibility, indeed, is quantified by means of the mean value $\langle\sigma_{A - B}\rangle$, with $\langle \sigma_{A - B} \rangle = 0$ corresponding to thermodynamic reversibility. As previously shown in Figs. \[fig:fig3\], \[fig:fig4\] and \[fig:fig5\], the mean value $\langle\sigma_{A - B}\rangle$ can be well approximated by $\langle\sigma_{A + B}\rangle$ and from Eqs. (\[eq:entropy-positivity\]) and (\[eq:sigma\_sum\]) we have $0\leq\langle\sigma_{A-B}\rangle \leq \langle\sigma_{A+B}\rangle$. From Figs. \[fig:fig5\] and \[fig:Fig10\], thus, we can observe that the implemented noisy transformation is more reversible with respect to the unitary one. Indeed, the statistical moments of the stochastic quantum entropy production, as well as the corresponding mean value, go to zero as $\Gamma$ increases. Since the dynamics originating from the Lindblad equation (\[Lindblad\]) admits as a fixed point the completely mixed state of the composite system $A - B$, if we increase the value of $\Gamma$ then the probability distribution of the quantum entropy production for the systems $A$, $B$ and $A - B$ tends to a Kronecker delta around zero, leading to a more-reversible system transformation with respect to the unitary case. For this reason, also the numerical simulations of Fig. \[fig:Fig10\] have been performed by using the inverse Fourier transform to reconstruct the probabilities $\{\textrm{Prob}(\sigma_{C,i})\}$, with $i = 1,\ldots,M_{C}$ and $C\in\{A,B\}$, instead of calculating the pseudo-inverse of the matrix $\Sigma_{C}$. As a matter of fact, as $\Gamma$ increases some values of $\sigma_{C}$ approach zero and $\Sigma_{C}$ becomes singular. Let us observe that when the dynamics is unitary the performance of the reconstruction algorithm adopting the inverse Fourier transform can be affected by a not-negligible error, as shown by the $\textrm{RMSE}\left(\{\langle\sigma^{k}_{A+B}\rangle\}\right)$ in the last panel of Fig.\[fig:Fig10\]. For such case, the adoption of the pseudo-inverse of $\Sigma_{C}$ is to be preferred. Moreover, we expect that increasing the number of ions the thermodynamic irreversibility becomes more and more pronounced and that this analysis may be the object of forthcoming work. In conclusion, a system transformation on a multipartite quantum system involves stochastic quantum entropy production whenever correlations between the subsystems of the multipartite system is first created by the dynamics of the composite system and then destroyed by the second measurement. This result, indeed, can be easily deduced from Figs. \[fig:Fig9\] and \[fig:Fig10\], in which, for a fixed value of $\Gamma$ ($\Gamma$ = 0.2 rad/s) and $\phi$ $(\phi = \pi/7)$ respectively, the behaviour of the statistical moments of the stochastic quantum entropy production as a function of $\phi$ ($\Gamma$) is monotonically increasing (decreasing). Indeed, the stronger is the interaction between the two ions, the larger is the corresponding production of correlations between them. On the other side, instead, the production of correlations within a multipartite system is inhibited due to the presence of strong decoherent processes. Discussion and conclusions ========================== The goal of this paper has been three-fold: We discussed how to relate the stochastic quantum entropy production to the quantum fluctuation theorem, which is the generalization of the Tasaki-Crooks theorem for open systems, and this relation is based on the evaluation and quantification of the irreversibility by means of a two-times measurement protocol. By using the definition of the stochastic quantum entropy production, then, we introduced a protocol to reconstruct its fluctuations, and characterize the dynamics of the quantum system. In particular, the proposed procedure allows us to determine the mean value $\langle\sigma\rangle$ of the stochastic quantum entropy production, quantifying and probing the amount of irreversibility, with $\langle \sigma \rangle=0$ ($\langle \sigma \rangle>0$) corresponding to thermodynamic reversibility (irreversibility). At variance, $\langle \sigma \rangle<0$ – violating the second law of the thermodynamics – do not occur due to the non-negativity of the Kullback-Leibler divergence. Furthermore, we have shown the relation between the quantum relative entropy of the system density matrix at the final time of the transformation and the stochastic quantum entropy production. In this regard, Eq. (\[eq:entropy-positivity\]) is valid also at zero temperature and without assuming thermal baths with finite temperature. It is worth noting, however, that the presented results about the stochastic quantum entropy production are based on the assumption to consider unital CPTP quantum maps for the description of the open system dynamics. Even though we limited ourself for most of the paper to unital CPTP quantum maps, in Section \[rec\_2nd\_law\] we discussed how to put in relation our results with the conventional second law of thermodynamics at finite temperature. Secondly, to reconstruct the probability distribution of $\sigma$, we proposed a suitable reconstruction protocol based on the determination of the characteristic functions of the stochastic quantum entropy production, which are evaluated $N$ times for a given set of $N$ real parameters, so as to collect an adequate amount of information for the reconstruction. In other words, the reconstruction algorithm relies on a parametric version of the integral quantum fluctuation theorem, and yields the first $N$ statistical moments of the stochastic quantum entropy production through the inversion of a Vandermonde matrix, which encodes the experimental evaluation of the characteristic function. The corresponding numerical error, however, can be reduced using the solution of a polynomial interpolation problem based on the use of (zeros of) the Chebyshev polynomials. Then, the presented algorithm in its generic form uses a numerical inverse Fourier transform to reconstruct the probability distribution of the stochastic quantum entropy production from its statistical moments. In order to optimize the minimum amount of resources needed for the protocol, we have also shown that the required number of measurements to perform the algorithm scales linearly with the number of the values that can be assumed by the stochastic quantum entropy production, and not quadratically as one would have obtained by a direct application of the definition of the corresponding probability distributions. Moreover, as third aspect, under the hypothesis that the open quantum system is composed by mutually interacting subsystems, we investigated the stochastic quantum entropy production both for the subsystems and for the composite system, showing that their mean values are sub-additive; they coincide only when the system density matrix at the end of the transformation is a product state. In this way, we have proved not only how to characterize the irreversibility of an arbitrary unital quantum process by reconstructing the corresponding upper-bound in accordance with the sub-additivity, but we have also provided a method to detect correlations between the partitions of the system, just by comparing the fluctuation profiles of $\sigma_{A+B}$ and $\sigma_{A-B}$. In this regard, let us recall that the correlation shared by two parts of a composite system (let us consider e.g. even just the entanglement, genuinely quantum type of correlation) is not directly measurable in laboratory, since there does not exist a self-adjoint operator quantifying it. To this aim, one could perform a single quantum state tomography on the final state of the system after the transformation and evaluate the quantum discord for such state, so as to measure nonclassical correlations between its partitions [@OllivierPRL2001]. However, it is worth noting that the performance of a full state tomography is not always feasible in terms of the available resources, and, along this direction, nontrivial solutions have been recently introduced (see e.g. Ref. [@GirolamiPRL2012]). Conversely, as previously shown, the measure of the probability distribution of the quantum entropy production by performing measurements both on local and global observables of the composite system does not require any quantum state tomography, since only measurements of the occupation probabilities of the final density matrix of the system are needed. Finally, to detect correlations one could also directly measure the second order dynamical correlation function of the system. However, such procedure, although it could not require a full quantum state tomography, is extremely system-dependent and usually a relevant experimental effort is necessary, though it has been recently introduced a measurement apparatus relying on weak-ancilla-system couplings, which could in part facilitate such measurements [@UhrichPRA2017]. In order to illustrate our theoretical results, we discussed an experimental implementation with trapped ions and we showed the performance of the reconstruction algorithm on a quantum system composed of two trapped ions, subjected to a unitary evolution and to a Lindbladian one. We showed that the probability distribution of the stochastic quantum entropy production can be efficiently reconstructed with a very small error already with $N = 7$ momenta. Efficiency and possible extensions of the presented procedure have been also discussed. Generalizations of our results to more general (non-unital) quantum maps have been briefly discussed and will be investigated in a future work. Our protocol, summarized in Fig. \[fig:procedure\], is experimentally oriented and it is based on the preparation of suitably prepared initial states, depending on the parameter $\varphi$ entering the characteristic functions $\chi_{C}(\varphi)$ to be measured. Such protocol appears to be within the reach of experimental realizations, given the remarkable results obtained in the last decades in the preparation of quantum states [@BergmannRMP1998]. In this regard, we observe that state preparation can be achieved in most common quantum technology platforms via optimal control procedures [@BrifNJP2010], allowing to speed up the process of coherent population transfer up to the ultimate bound imposed by quantum mechanics, [*i.e.*]{} the quantum speed limit [@CanevaPRL2009; @DoriaPRL2011]. As a final remark, we also observe that the fluctuation properties of the stochastic quantum entropy production do strongly depend on the presence of decoherent channels between an arbitrary quantum system and the environment. Hence, one could effectively determine not only the influence of the external noise sources on the system dynamics, but also characterize the environment structure and properties via quantum sensing procedures. We believe that this aspect could be the subject for future investigations, [*e.g.*]{} along the research direction proposed in [@Mueller3; @Mueller2], based on stochastic quantum Zeno phenomena [@Gherardini1; @Mueller1], and/or in Refs. [@Cosco2015; @Nokkala2016] via engineered quantum networks.\ \ [**Acknowledgments**]{}\ The authors gratefully acknowledge Mauro Paternostro, Giorgio Battistelli, Giacomo Gori, Jin Wang, Francesco Saverio Cataliotti, Duccio Fanelli, Pietro Silvi and Augusto Smerzi for useful discussions. S.G. and M.M.M. thank the Scuola Internazionale Superiore di Studi Avanzati (SISSA), Trieste (Italy) for hospitality during the completion of this work. A.T. thanks the Galileo Galilei Institute for Theoretical Physics for the hospitality in the Workshop “From Static to Dynamical Gauge Fields with Ultracold Atoms” and the INFN for partial support during the completion of this work. This work was financially supported from the Fondazione CR Firenze through the project Q-BIOSCAN. Proof of Theorem 1 {#appendix_a} ================== In this Appendix, we prove the equality between the conditional probabilities $p(a^{\textrm{fin}}_{k}|a^{\textrm{in}}_{m})$ and $p(a^{\textrm{in}}_{m}|a^{\textrm{ref}}_{k})$, respectively, of the forward and backward processes of our two-time measurement scheme. Let us recall the observables $\mathcal{O}_{\textrm{in}}\equiv\sum_{m}a^{\textrm{in}}_{m}\Pi^{\textrm{in}}_{m}$, $\mathcal{O}_{\textrm{fin}}\equiv\sum_{k}a^{\textrm{fin}}_{k}\Pi^{\textrm{fin}}_{k}$, $\widetilde{\mathcal{O}}_{\textrm{ref}}\equiv\sum_{k}a^{\textrm{ref}}_{k}\widetilde{\Pi}^{\textrm{ref}}_{k}$ and $\widetilde{\mathcal{O}}_{\textrm{in}} = \sum_{m}a^{\textrm{in}}_{m}\widetilde{\Pi}^{\textrm{in}}_{m}$, as defined in the main text. The dynamical evolution of the open quantum system between the two measurements is described by a unital CPTP map $\Phi(\cdot)$ (with $\Phi(\mathbbm{1}) = \mathbbm{1}$), whose Kraus operators $\{E_{u}\}$ are such that $\sum_{u}E_{u}^{\dagger}E_{u} = \mathbbm{1}$, where $\mathbbm{1}$ denotes the identity operator on the Hilbert space $\mathcal{H}$ of the quantum system. Accordingly, $\Phi(\rho_{\textrm{in},m}) = \sum_{u}E_{u}\rho_{\textrm{in},m}E_{u}^{\dagger}$, where $\rho_{\textrm{in},m}\equiv\Pi^{\textrm{in}}_{m}\rho_{0}\Pi^{\textrm{in}}_{m}$, and thus the conditional probability $p(a^{\textrm{fin}}_{k}|a^{\textrm{in}}_{m})$ equals to $$\begin{aligned} \label{cond_prob_forward_appA} &&p(a^{\textrm{fin}}_{k}|a^{\textrm{in}}_{m}) = \frac{\textrm{Tr}[\Pi^{\textrm{fin}}_{k}\Phi(\rho_{\textrm{in},m})]} {\textrm{Tr}[\Pi^{\textrm{in}}_{m}\rho_{0}\Pi^{\textrm{in}}_{m}]} = \frac{\textrm{Tr}[\Pi^{\textrm{fin}}_{k}\sum_{u}E_{u}\rho_{\textrm{in},m}E_{u}^{\dagger}]} {\textrm{Tr}[\Pi^{\textrm{in}}_{m}\rho_{0}\Pi^{\textrm{in}}_{m}]}\nonumber \\ &&= \sum_{u}\frac{\textrm{Tr}[\Pi^{\textrm{fin}}_{k}E_{u}\Pi^{\textrm{in}}_{m}\rho_{0}\Pi^{\textrm{in}}_{m}E_{u}^{\dagger}]} {\textrm{Tr}[\Pi^{\textrm{in}}_{m}\rho_{0}\Pi^{\textrm{in}}_{m}]} = \sum_{u}|\langle\phi_{a_{k}}|E_{u}|\psi_{a_{m}}\rangle|^{2}.\nonumber \\ &&\end{aligned}$$ Next, by inserting in Eq. (\[cond\_prob\_forward\_appA\]) the identity operator $\mathbbm{1} = \Theta\Theta^\dagger=\Theta^\dagger\Theta$, where $\Theta$ is the time-reversal operator as defined in the main text, one has: $$\begin{aligned} |\langle\phi_{a_{k}}|E_{u}|\psi_{a_{m}}\rangle|^{2} &=& |\langle\phi_{a_{k}}|\Theta^{\dagger}\left(\Theta E_{u}\Theta^\dagger \right)\Theta|\psi_{a_{m}}\rangle|^{2} = |\langle \widetilde{\phi}_{a_{k}}| \Theta E_{u} \Theta^\dagger |\widetilde{\psi}_{a_{m}}\rangle|^2 \nonumber\\ &=& |\langle\widetilde{\psi}_{a_{m}}|\Theta E_{u}^{\dagger}\Theta^\dagger|\widetilde{\phi}_{a_{k}}\rangle|^{2}.\end{aligned}$$ where we have used complex conjugation and the modulus squared to flip the order of the operators. The time-reversal of a single Kraus operator is $\widetilde{E}_{u}\equiv\mathcal{A}\pi^{1/2}E^{\dagger}_{u}\pi^{-1/2}\mathcal{A}^{\dagger}$, where we choose $\mathcal{A} = \Theta$ and $\pi = \mathbbm{1}$ (as $\Phi$ is unital, such that $\Phi(\mathbbm{1}) = \mathbbm{1}$). We can now state that $$|\langle\phi_{a_{k}}|E_{u}|\psi_{a_{m}}\rangle|^{2} = |\langle\widetilde{\psi}_{a_{m}}|\widetilde{E}_{u}|\widetilde{\phi}_{a_{k}}\rangle|^{2}.$$ Moreover, by observing that $$\sum_{u}|\langle\widetilde{\psi}_{a_{m}}|\widetilde{E}_{u}|\widetilde{\phi}_{a_{k}}\rangle|^{2} = \frac{\textrm{Tr}[\widetilde{\Pi}^{\textrm{in}}_{m}\widetilde{\Phi}(\rho_{\textrm{ref},k})]} {\textrm{Tr}[\widetilde{\Pi}^{\textrm{ref}}_{k}\widetilde{\rho}_{\tau}\widetilde{\Pi}^{\textrm{ref}}_{m}]} = p(a^{\textrm{in}}_{m}|a^{\textrm{ref}}_{k}),$$ where $\rho_{\textrm{ref},k}\equiv\widetilde{\Pi}^{\textrm{ref}}_{k}\widetilde{\rho}_{\tau}\widetilde{\Pi}^{\textrm{ref}}_{m}$, the equality $p(a^{\textrm{fin}}_{k}|a^{\textrm{in}}_{m}) = p(a^{\textrm{in}}_{m}|a^{\textrm{ref}}_{k})$, as well as the Theorem 1, follow straightforwardly. Proof of Theorem 2 {#appendix_b} ================== Here, we prove Theorem 2, [*i.e.*]{} the inequality $$0\leq S(\rho_{\textrm{fin}}\parallel\rho_{\tau})\leq\langle\sigma\rangle,$$ where $\rho_{\textrm{fin}}$ and $\rho_{\tau}$ are the density operators of the open quantum system $\mathcal{S}$ before and after the second measurement of the forward process. $S(\rho_{\textrm{fin}}\parallel\rho_{\tau})$ is called the quantum relative entropy of $\rho_{\textrm{fin}}$ and $\rho_{\tau}$ and $\langle\sigma\rangle$ is the average of the stochastic quantum entropy production. This inequality may be regarded as the quantum counterpart of the second law of thermodynamics for an open quantum system. Let us consider the stochastic entropy production $\sigma(a^{\textrm{fin}},a^{\textrm{in}})=\ln\left[\frac{p(a^{\textrm{in}})}{p(a^{\textrm{ref}})}\right]$ (as given in Eq. (\[sigma\]) in the main text) for the open quantum system $\mathcal{S}$, whose validity is subordinated to the assumptions of Theorem 1. Accordingly, the average value of $\sigma$ is $$\langle\sigma\rangle = \sum_{a^{\textrm{fin}},a^{\textrm{in}}}p(a^{\textrm{fin}},a^{\textrm{in}})\ln\left[\frac{p(a^{\textrm{in}})}{p(a^{\textrm{ref}})}\right] = \sum_{a^{\textrm{in}}}p(a^{\textrm{in}})\ln[p(a^{\textrm{in}})] - \sum_{a^{\textrm{fin}}}p(a^{\textrm{fin}})\ln[p(a^{\textrm{ref}})]\geq 0.$$ We observe that the mean quantum entropy production $\langle\sigma\rangle$ is a non-negative quantity due to the positivity of the classical relative entropy, or Kullback-Leibler divergence. Since $p(a^{\textrm{fin}})\equiv\langle\phi_{a}|\rho_{\textrm{fin}}|\phi_{a}\rangle$ and the reference state is diagonal in the basis $\{|\phi_a\rangle\}$, we have $$\sum_{a^{\textrm{fin}}}p(a^{\textrm{fin}})\ln[p(a^{\textrm{ref}})] = \sum_{a^{\textrm{fin}}}\langle\phi_{a}|\rho_{\textrm{fin}}|\phi_{a}\rangle\ln[p(a^{\textrm{ref}})] = \sum_{a^{\textrm{fin}}}\langle\phi_{a}|\rho_{\textrm{fin}}\ln\rho_{\textrm{ref}}|\phi_{a}\rangle = \textrm{Tr}\left[\rho_{\textrm{fin}}\ln\rho_{\tau}\right],$$ where the last identity is verified by assuming the equality between the reference state $\rho_{\textrm{ref}}$ and the density operator $\rho_{\tau}$ after the second measurement of the protocol. One also has: $$\sum_{a^{\textrm{in}}}p(a^{\textrm{in}})\ln[p(a^{\textrm{in}})] = \textrm{Tr}\left[\rho_{\textrm{in}}\ln\rho_{\textrm{in}}\right] = -S(\rho_{\textrm{in}}),$$ where $S(\rho_{\textrm{in}})\equiv-\textrm{Tr}\left[\rho_{\textrm{in}}\ln\rho_{\textrm{in}}\right]$ is the von Neumann entropy for the initial density operator $\rho_{\textrm{in}}$ of the quantum system $\mathcal{S}$. The mean quantum entropy production $\langle\sigma\rangle$, thus, can be written in general as $$\langle\sigma\rangle = -\textrm{Tr}\left[\rho_{\textrm{fin}}\ln\rho_{\tau}\right] - S(\rho_{\textrm{in}}).$$ The quantum relative entropy is defined as $S(\rho_{\textrm{fin}}\parallel\rho_{\tau})=- \textrm{Tr}\left[\rho_{\textrm{fin}}\ln\rho_{\tau}\right] - S(\rho_{\textrm{fin}})$ and trivially $S(\rho_{\textrm{fin}}\parallel\rho_{\tau})\geq 0$. According to our protocol, the initial and the final states are connected by the unital CPTP map $\Phi$ as $\rho_{\textrm{fin}}=\Phi(\rho_{\textrm{in}})$. As a consequence of the unitality of $\Phi$ the von Neumann entropies obey the relation $S(\rho_{\textrm{in}})\leq S(\rho_{\textrm{fin}})$ [@Sagawa2014]. Summarizing, we obtain $$0\leq S(\rho_{\textrm{fin}}\parallel\rho_{\tau}) = - \textrm{Tr}\left[\rho_{\textrm{fin}}\ln\rho_{\tau}\right] - S(\rho_{\textrm{fin}}) \leq - \textrm{Tr}\left[\rho_{\textrm{fin}}\ln\rho_{\tau}\right] - S(\rho_{\textrm{in}}) = \langle\sigma\rangle,$$ proving the original inequality. Note that if we perform the second measurement with a basis in which $\rho_{\textrm{fin}}$ is diagonal ([*i.e.*]{} vanishing commutator between measurement operator and final state, $[\mathcal{O}_{\textrm{fin}},\rho_{\textrm{fin}}]=0$), the state is unchanged by the second measurement and $\rho_{\textrm{fin}}=\rho_{\tau}$. As a consequence $$0= S(\rho_{\textrm{fin}}\parallel\rho_{\tau})\leq\langle\sigma\rangle=S(\rho_{\textrm{fin}})-S(\rho_{\textrm{in}}),$$ [*i.e.*]{} the quantum relative entropy vanishes, while the average of the stochastic entropy production equals to the difference of final and initial von Neumann entropies, $\langle\sigma\rangle=S(\rho_{\textrm{fin}})-S(\rho_{\textrm{in}})$, and thus describes the irreversibility distribution of the map $\Phi$ only (and not of the measurement, as it would be in the general case). In the general case, i.e. if the condition $[\mathcal{O}_{\textrm{fin}},\rho_{\textrm{fin}}]=0$ does not hold, still the post-measurement state $\rho_{\tau}$ is diagonal in the basis of the observable eigenstates and we obtain $$\langle\sigma\rangle = -\textrm{Tr}\left[\rho_{\textrm{fin}}\ln\rho_{\tau}\right] - S(\rho_{\textrm{in}}) = S(\rho_{\tau})- S(\rho_{\textrm{in}}).$$ Recovering the second law of thermodynamics {#appendix_second_law} =========================================== Here we show the formal derivation to connect the entropy production inequality in Theorem 2 to the second law of thermodynamics for the open quantum system $\mathcal{S}$ in term of the mean work $\langle\mathrm{W}\rangle$ and the free-energy difference $\Delta F$ as defined in the main text. The main ingredient of this proof is to express ${\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}]$ as a function of the thermal state $\rho_{\tau}^{{\rm th}} \equiv e^{\beta\left[F(\tau)\mathbbm{1}_\mathcal{S} - H(\tau)\right]}$ at time $t = \tau$. In particular, we can write that $$\begin{aligned} {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}] &=& {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}^{{\rm th}}] + {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}] - {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}^{{\rm th}}] \\ &=& {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}^{{\rm th}}] + {\rm Tr}[\rho_{\textrm{fin}}(\ln\rho_{\textrm{fin}} - \ln\rho_{\textrm{fin}} + \ln\rho_{\tau} - \ln\rho_{\tau}^{{\rm th}})] \\ &=& {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}^{{\rm th}}] + {\rm Tr}[\rho_{\textrm{fin}}(\ln\rho_{\textrm{fin}} - \ln\rho_{\tau}^{{\rm th}})] - {\rm Tr}[\rho_{\textrm{fin}}(\ln\rho_{\textrm{fin}} - \ln\rho_{\tau})], \\\end{aligned}$$ so as to obtain $${\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}] = {\rm Tr}[\rho_{\textrm{fin}}\ln\rho_{\tau}^{{\rm th}}] + S(\rho_{\textrm{fin}}\parallel\rho_{\tau}^{{\rm th}}) - S(\rho_{\textrm{fin}}\parallel\rho_{\tau}),$$ i.e. Eq. (\[new\_equation\_quantum\_entropy\]) in the main text. Therefore, by taking Eq. (\[sigma\_theorem2\]) and substituting $$S(\rho_{\textrm{in}}) = - {\rm Tr}[\rho_{\textrm{in}}\ln\rho_{\textrm{in}}] = -\beta F(0) + \beta{\rm Tr}[\rho_{\textrm{in}}H(0)]$$ with $\rho_{\textrm{in}} \equiv e^{\beta\left[F(0)\mathbbm{1}_\mathcal{S} - H(0)\right]}$, one has $$\langle\sigma\rangle = \beta F(0) - \beta{\rm Tr}[\rho_{\textrm{in}}H(0)] - \beta F(\tau) + \beta{\rm Tr}[\rho_{\textrm{fin}}H(\tau)] - S(\rho_{\textrm{fin}}\parallel\rho_{\tau}^{{\rm th}}) + S(\rho_{\textrm{fin}}\parallel\rho_{\tau}).$$ Accordingly, being $0\leq S(\rho_{\textrm{fin}}\parallel\rho_{\tau})\leq\langle\sigma\rangle$ (from the results of Theorems 1 and 2) and $S(\rho_{\textrm{fin}}\parallel\rho_{\tau}^{{\rm th}})\geq 0$ (non-negativity of the quantum relative entropy), we finally recover the conventional second law of thermodynamics, i.e. $\langle\mathrm{W}\rangle\geq\Delta F$. Derivation of the characteristic functions {#appendix_charact_func} ========================================== In this Appendix, we derive the expressions for the characteristic functions $G_A(\lambda)$ (for the probability distributions $\textrm{Prob}(\sigma_A)$) and $G_B(\lambda)$ (for the probability distributions $\textrm{Prob}(\sigma_B)$), given by Eq. (\[eq3\]) and Eq. (\[eq4\]), respectively. We start with the definition $$G_A(\lambda) = \int \textrm{Prob}_{A}(\sigma_A)e^{i\lambda\sigma_A}d\sigma_A,$$ where $$\textrm{Prob}(\sigma_A) = \sum_{k,m}\delta\left[\sigma_A - \sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k})\right]p_a(k,m),$$ as well as $$p_{a}(k,m) = \textrm{Tr}\left[(\Pi^{\tau}_{A,k}\otimes\mathbbm{1}_{B})\Phi(\Pi^{\textrm{in}}_{A,m}\otimes\rho_{\textrm{B,in}})\right]p(a_{m}^{\textrm{in}}),$$ and $$\sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k})=\ln [p(a^{\textrm{in}}_{m})] - \ln[p(a^{\textrm{fin}}_{k})]$$ Exploiting the linearity of $\Phi$ and the trace, we obtain $$\begin{aligned} \label{eq:G-appendix} &&G_A(\lambda) = \sum_{k,m}p_{a}(k,m)e^{i\lambda\sigma_A(a^{\textrm{in}}_{m},a^{\textrm{fin}}_{k})}\nonumber \\ &&= \textrm{Tr}\left[\left(\sum_{k}\Pi^{\tau}_{A,k}e^{-i\lambda \ln [p(a^{\textrm{fin}}_{k})]}\otimes\mathbbm{1}_{B}\right)\Phi\left(\sum_{m}\Pi^{\textrm{in}}_{A,m}e^{i\lambda \ln[p(a^{\textrm{in}}_{m})]}p(a_{m}^{\textrm{in}})\otimes\rho_{\textrm{B,in}}\right)\right].\nonumber \\ &&\end{aligned}$$ Recalling the spectral decompositions of the initial and final density operators, $\rho_{\textrm{A,in}}\equiv\sum_{m}\Pi^{\textrm{in}}_{A,m}p(a^{\textrm{in}}_{m})$ and $\rho_{A,\tau}\equiv\sum_{k}\Pi^{\tau}_{A,k}p(a^{\tau}_{k})$, with eigenvalues $p(a^{\textrm{in}}_{m})$ and $p(a^{\tau}_{k})=p(a^{\textrm{fin}}_{k})$, we get $$\sum_{k}\Pi^{\tau}_{A,k}e^{-i\lambda \ln[p(a^{\textrm{fin}}_{k})]} = \sum_{k}\Pi^{\tau}_{A,k}e^{-i\lambda \ln[p(a^{\tau}_{k})]} = \sum_{k}\Pi^{\tau}_{A,k}p(a^{\tau}_{k})^{-i\lambda} = \left(\rho_{A,\tau}\right)^{-i\lambda},$$ and $$\sum_{m}\Pi^{\textrm{in}}_{A,m}e^{i\lambda \ln[p(a^{\textrm{in}}_{m})]}p(a_{m}^{\textrm{in}}) = \sum_{m}\Pi^{\textrm{in}}_{A,m}p(a_{m}^{\textrm{in}})^{1 + i\lambda} = \left(\rho_{A,\textrm{in}}\right)^{1 + i\lambda}.$$ If we insert these expressions into Eq. (\[eq:G-appendix\]) we obtain the expression for the characteristic function $G_A(\lambda)$ given in Eq. (\[eq:G\_A\]). Analogously we can derive Eq. (\[eq2\]) for $G_B(\lambda)$. In a similar way we can derive the characteristic function $G_{A - B}(\lambda)$ of the stochastic entropy production of the composite system $A - B$: $$G_{A - B}(\lambda)=\mathrm{Tr}\left[\rho_{\tau}^{-i\lambda}\Phi(\rho_{\mathrm{in}}^{1+i\lambda})\right]\,.$$ References {#references .unnumbered} ========== [^1]: Let us assume that the initial density matrix $\rho_{\textrm{in}}$ is a Gibbs thermal state at inverse temperature $\beta$, [*i.e.*]{} $\rho_{\textrm{in}}\equiv e^{\beta\left[F(0)\mathbbm{1}_\mathcal{S}- H(0)\right]}$, where $F(0) \equiv -\beta^{-1}\ln\left\{\textrm{Tr}[e^{-\beta H(t = 0)}]\right\}$ and $H(0)$ are, respectively, equal to the Helmholtz free-energy and the system Hamiltonian at time $t=0$. Accordingly, the von Neumann entropy $S(\rho_{\textrm{in}})$ equals the thermodynamic entropy at $t = 0$, [*i.e.*]{} $S(\rho_{\textrm{in}}) = \beta(\langle H(0)\rangle - F(0))$, where $\langle H(0)\rangle \equiv \textrm{Tr}[\rho_{\textrm{in}}H(0)]$ is the average energy of the system in the canonical distribution. More generally, we can state that given an arbitrary initial density matrix $\rho_{\textrm{in}}$ the thermodynamic entropy $\beta(\langle H(0)\rangle - F(0))$ represents the upper-bound value for the von Neumann entropy $S(\rho_{\textrm{in}})$, whose maximum value is reached only in the canonical distribution. To prove this, it is sufficient to consider $S(\rho_{\textrm{in}}\parallel e^{\beta\left(F(0)\mathbbm{1}_\mathcal{S} - H(0)\right)}) = \beta\left(F(0) - \langle H(0)\rangle\right) - S(\rho_{\textrm{in}})$, from which, from the positivity of the quantum relative entropy, one has $S(\rho_{\textrm{in}})\leq\beta(\langle H(0)\rangle - F(0))$.

ArXiv

--- abstract: 'We present an efficient, low-cost implementation of time-hopping impulse radio that fulfills the spectral mask mandated by the FCC and is suitable for high-data-rate, short-range communications. Key features are: (i) all-baseband implementation that obviates the need for passband components, (ii) symbol-rate (not chip rate) sampling, A/D conversion, and digital signal processing, (iii) fast acquisition due to novel search algorithms, (iv) spectral shaping that can be adapted to accommodate different spectrum regulations and interference environments. Computer simulations show that this system can provide 110Mbit/s at 7-10m distance, as well as higher data rates at shorter distances under FCC emissions limits. Due to the spreading concept of time-hopping impulse radio, the system can sustain multiple simultaneous users, and can suppress narrowband interference effectively.' author: - 'Andreas F. Molisch,$^{2}$[^1] Ye Geoffrey Li, Yves-Paul Nakache, Philip Orlik,' - 'Makoto Miyake, Yunnan Wu, Sinan Gezici, Harry Sheng,' - 'S. Y. Kung, H. Kobayashi, H.Vincent Poor, Alexander Haimovich$,$and Jinyun Zhang$^{3}$[^2]' bibliography: - 'art\_ch1.bib' - 'uwb.bib' title: 'A low-cost time-hopping impulse radio system for high data rate transmission$^{1}$[^3]' --- Introduction ============ Ultrawideband (UWB) wireless systems are defined as systems that use either a large relative bandwidth (ratio of bandwidth to carrier frequency larger than 25[%]{}), or a large absolute bandwidth (larger than 500MHz). While UWB radar systems have been used for a long time, mainly in the military domain [@Taylor_1995], UWB communications systems are a fairly recent development. The first papers in the open literature are those of Win and Scholtz [@Scholtz_1993], [@Win_and_Scholtz_1998], [@Win_and_Scholtz_2000], who developed the concept of time-hopping impulse radio (TH-IR) system. This concept excited immense interest in the area of military [@Kolenchery_et_al_1998] as well as civilian [@Ho_et_al_2001] communications. Further advances of TH-IR are described, e.g., in [@LeMatret_and_Giannakis_2000], [@Conroy_et_al_1999], [@Forouzan_et_al_2000], [@Cassioli_et_al_2003], [@Fischler_and_Poor_2002]. In 2002, the Federal Communications Commission (FCC) in the US allowed *unlicensed* UWB communications [@FCC_2002]. This greatly increased commercial interest in UWB, leading to a large number of papers, see, e.g., [@UWBST_2002], [@Oulu_2003]. One of the most promising applications is data communications at rates that are higher than the currently popular 802.11b ($11$ Mbit/s) and 802.11a ($<$$54$ Mbit/s) standards. The goal, as mandated, e.g., by the standardization committee IEEE 802.15.3a, is a system that can provide multiple piconets with $110$ Mbit/s each. This data rate should be achieved for distances up to $10$ m (Personal Area Networks). Higher data rates should be feasible at shorter distances. The principle of using very large bandwidths has several generic advantages: - By spreading the information over a large bandwidth, the spectral *density* of the transmit signal can be made very low. This decreases the probability of intercept (for military communications), as well as the interference to narrowband victim receivers. - The spreading over a large bandwidth increases the immunity to narrowband interference and ensures good multiple-access capabilities [@Zhao_and_Haimovich_2002], [@Ramirez_2001]. - The fine time resolution implies high temporal diversity, which can be used to mitigate the detrimental effects of fading [@Win_and_Scholtz_1998_energy_capture]. - Propagation conditions can be different for the different frequency components. For example, a wall might be more transparent in a certain frequency range. The large bandwidth increases the chances that at least some frequency components arrive at the receiver [@Cassioli_et_al_2002]. These advantages are inherent in the use of very large bandwidths, and can thus be achieved by *any* UWB system, including the recently proposed UWB frequency-hopping OFDM system [@Batra_et_al_2003] and UWB direct-sequence spread spectrum (DS-SS) systems [@McCorkle_et_al_2003]. However, TH-IR has additional advantages: - Recent information-theoretic results indicate that higher capacities can be achieved than with DS-SS systems [@Subramanian_and_Hajek_2002], [@Molisch_et_al_2003_PACRIM]. - More important from a practical point of view, impulse radio systems operate in baseband only, thus requiring no frequency upconversion circuitry and associated RF components [@LeMatret_and_Giannakis_2000], though circuitry for accurate timing is still required. This allows low-cost implementation. A lot of progress has been made in the theoretical understanding of impulse radio, as evidenced by the papers mentioned above. However, several assumptions made in the theoretical analyses do not agree with the requirements for a practical implementation of a high-data-rate impulse radio system. Those requirements may stem from the regulations by the FCC and other frequency regulators, from the necessity of coexistence with other devices, and from cost considerations. The goal of this paper is to describe the complete physical-layer design of an IR system that is suitable for practical implementation. In this system, we combine existing and innovative aspects, giving special attention to the interplay between the different aspects. The current paper is thus more of an engineering paper, while the theoretical background of some of our innovations is described in [@Wu_et_al_2003], [@Nakache_and_Molisch_2003], [@Gezici_et_al_2003]. The remainder of the paper is organized the following way: in Section II, we present an overview of the system. Next, we discuss the transmit signal, and how its spectrum can be shaped to fit the requirements of regulators, as well as to minimize interference to nearby devices. Section IV describes the signal detection at the receiver, including the structure of the Rake receiver and the equalizer. The channel estimation procedure that is used for establishing the weights of the Rake receiver and equalizer is discussed in Section V. Finally, Section VI presents simulations of the total performance of the system in terms of coverage and resistance to interference from narrowband signals and other UWB transmitters. A summary and conclusions wrap up the paper. System overview =============== The system that we are considering is a time-hopping impulse radio (TH-IR) system. We first describe “classical” TH-IR [@Win_and_Scholtz_2000]. Each data bit is represented by several short pulses; the duration of the pulses determines essentially the bandwidth of the (spread) system. For the single-user case, it would be sufficient to transmit a single pulse per symbol. However, in order to achieve good multiple access (MA) properties, we have to transmit a whole sequence of pulses. Since the UWB transceivers are unsynchronized, so-called catastrophic collisions can occur, where pulses from several transmitters arrive at the receiver almost simultaneously. If only a single pulse would represent one symbol, this would lead to a bad signal-to-interference ratio, and thus to high bit error probability BER. These catastrophic collisions are avoided by sending a whole sequence of pulses instead of a single pulse. The transmitted pulse sequence is different for each user, according to a so-called time-hopping (TH) code. Thus, even if one pulse within a symbol collides with a signal component from another user, other pulses in the sequence will not. This achieves interference suppression gain that is equal to the number of pulses in the system. Fig. \[thir\_principle\] shows the operating principle of a generic TH-IR system. We see that the possible positions of the pulses within a symbol follow certain rules: the symbol duration is subdivided into $N_{f}$ frames of equal length. Within each frame the pulse can occupy an almost arbitrary position (determined by the time-hopping code). Typically, the frame is subdivided into chips, whose length is equal to a pulse duration. The (digital) time-hopping code now determines which of the possible positions the pulse actually occupies. \[ptb\] [THIR\_principle\_2.eps]{} The modulation of this sequence of pulses can be pulse-position modulation (PPM), as suggested in [@Win_and_Scholtz_2000], or amplitude modulation (PAM). PPM has the advantage that the detector can be simpler (an energy detector) in AWGN channels. However, reception in multipath environments requires a Rake receiver for either PPM or PAM. While this scheme shows good performance for some applications, it does have problems for high-data rate, FCC-compliant systems: 1. Due to the use of PPM, the transmit spectrum shows spectral lines. This requires the reduction of the total emission power, in order to allow the fulfillment of the FCC mask within each 1MHz band, as required by the FCC. 2. Due to the high data rate required by 802.15, and due to the high delay spread seen by indoor channels, the system works better with an equalizer. An equalizer for PPM will introduce increased complexity and cost. 3. For a full recovery of all considered multi-path components, the system requires a Rake receiver with a large number of fingers. A conventional implementation, using many digital correlators, will also introduce increased complexity and cost. 4. Due to the relatively low spreading factor of less than 40, the number of possible pulse positions within a frame is limited. This might lead to higher collision probability, and thus smaller interference suppression. The first two problems are solved by using (antipodal) pulse amplitude modulation (PAM) instead of PPM. This eliminates the spectral lines, and allows in general an easier shaping of the spectrum. Furthermore, it allows the use of simple linear equalizers. As detailed below, an innovative Rake receiver is considered to overcome the third problem; this Rake structure implements correlators by means of pulse generators and multipliers only. The problem of multiple-access interference, finally, can be addressed by interference-suppressing combining of the Rake finger signals. \[ptb\] [blockdiagram\_3.eps]{} A block diagram of the system is shown in Fig. \[blockdiagram\]. The transmit data stream is divided into blocks, and each block is encoded with a convolutional coder. We use a rate $1/2$ convolutional code with a constraint length 7. The use of turbo codes or low-density parity check codes would improve the performance by approximately 2 dB; however, decoding becomes challenging at the high data rates envisioned in this scheme. Then, a preamble is prepended that can be used for both acquisition and channel estimation. As mentioned above, the modulation and multiple access format is BPSK-modulated TH-IR. Each pulse sequence representing one symbol is multiplied by $\pm$1, depending on the bit to be transmitted. Finally, each data block (including preamble) is amplified (with power control, in order to minimize interference to other systems), and transmitted. Note that as the system is packet based and the number of packets per second can vary, it is not desirable to code across packets. In the receiver, the acquisition part of the preamble is stripped off and used to determine the timing. Once this has been established, the channel estimation part of the preamble is used to determine the coefficients for the Rake receiver and the equalizer. The main body of the data block is then received by a Rake receiver that can be interpreted as a filter that is matched to the convolution of the transmit signal and the channel impulse response. Each finger of the Rake finger is a filter that is matched to a time-delayed version of the transmit signal, encompassing both to the pulse shape and the time-hopping sequence. We use here an innovative Rake structure that requires only pulse generators and no delays, which makes an analogue implementation possible – this allows us to perform the sampling and A/D conversion only at the *symbol rate*, instead of the chip rate. Note that for chip rate sampling, A/D converters with about $20$ Gsamples/s would be required. The outputs of the Rake fingers are weighted (according to the principles of optimum combining) and summed up. The optimum location and weight of the fingers can be determined from the channel sounding sequence, which is processed before the reception of the actual data. The output of the summer is then sent through an MMSE equalizer and a decoder for the convolutional code. One important point of the system is that all the pulses are *baseband* pulses, more specifically, derivatives of Gaussian pulses. This allows a simple pulse generation, and obviates any need for passband components. This is a typical property of time-hopping impulse radio; however, it is not a trivial task within the restrictions of the FCC that the main power is emitted in the $3-10$ GHz range. We will show in Sec. 3 how this can be achieved. The goal of our design is to obtain a low-cost implementation. Thus, the design is not theoretically optimum, but rather contains a number of simplifications that reduce complexity of implementation and costs. Transmit signal and spectral shaping ==================================== Mathematical description of the transmit signal ----------------------------------------------- Throughout this paper, we use a communication system model where the transmitted signal is given by $$s_{tr}(t)=\sum\limits_{j=-\infty}^{\infty}d_{j}b_{\left\lfloor j/N_{f}\right\rfloor }w_{\text{tr}}(t-jT_{f}-c_{j}T_{c})=\sum\limits_{k=-\infty }^{\infty}b_{k}w_{\text{seq}}\left( t-kT_{s}\right)$$ where $w_{\text{tr}}(t)$ is the transmitted unit-energy pulse, $T_{f}$ is the average pulse repetition time, $N_{f}$ is the number of frames (and therefore also the number of pulses) representing one information symbol of length $T_{s}$, and $b$ is the information symbol transmitted, i.e., $\pm1;$ $w_{\text{seq}}(t)$ is the pulse sequence transmitted representing one symbol. The TH sequence provides an additional time shift of $c_{j}T_{c}$ seconds to the $j^{th}$ pulse of the signal, where $T_{c}$ is the chip interval, and $c_{j}$ are the elements of a pseudorandom sequence, taking on integer values between $0$ and $N_{c}-1$. To prevent pulses from overlapping, the chip interval is selected to satisfy $T_{c}\leq T_{f}$/$N_{c}$; in the following, we assume $T_{f}$ / $T_{c}=N_{c}$ so that $N_{c}$ is the number of chips per frame. We also allow polarity scrambling (see Sec. III.4), where each pulse is multiplied by a (pseudo) random variable $d_{j}$ that can take on the values $+1$ or $-1$ with equal probability. The sequence $d_{j}$ is assumed to be known at transmitter and receiver. An alternative representation can be obtained by defining a sequence [{]{}$s_{j}$[}]{} as follows $$s_{j}=\left\{ {\begin{array} [c]{l}d_{\left\lfloor {j/N_{c}}\right\rfloor }\text{ \ \ \ \ \ \ for \ \ \ \ \ \ \ }j-N_{f}\left\lfloor {j/N_{c}}\right\rfloor \text{=c}_{\left\lfloor {j/N_{c}}\right\rfloor }\\ 0\text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ otherwise}\ \ \ \ \ \ \ \ \text{\ \ \ \ \ }\end{array} }\right. \label{eq1}$$ Then the transmit signal can be expressed as $$s_{tr}(t)=\sum\limits_{j=-\infty}^{\infty}s_{j}b_{\left\lfloor {j/N_{f}N_{c}}\right\rfloor }w_{\text{tr}}(t-jT_{c}).\label{eq2}$$ To satisfy the spectrum masking requirement of the FCC, the transmit waveform $w_{tr}$, also known as monocycle waveform, is chosen to be the 5$^{th}$ derivative of the Gaussian pulse and it can be expressed as, $$w_{\text{tr}}(t)=p\left( t\right) =K_{2}\left( {-15\frac{t}{\sigma_{p}}+10\frac{t^{3}}{\sigma_{p}^{3}}-\frac{t^{5}}{\sigma_{p}^{5}}}\right) \exp(-\frac{t^{2}}{2\sigma_{p}^{2}}),$$ where $K_{2}$ is a normalization constant, and $\sigma_{p}$ controls the width of the pulse and it is chosen according to the spectral mask requirement of the FCC, which is [@Sheng_et_al_2003],$$\sigma_{p}=5.08\times10^{-11}\text{ s}.$$ Other signals shapes are possible; in particular, a combination of weighted pulses $p(t)$ (as explained below) can be used to improve the spectral properties. The various methods (e.g., Rake receiver, pulse polarity randomization, ....) discussed in the remainder of the paper can be applied independently of the exact shape of the transmit waveform. Spectral shaping - general aspects ---------------------------------- One of the key requirements for a UWB system is the fulfillment of the emission mask mandated by the national spectrum regulators [@Lehman_and_Haimovich_2003a]. In the USA, this mask has been prescribed by the FCC and essentially allows emissions in the $3.1-10.6$ GHz range with power spectral density of $-41.3$ dBm/MHz; in Europe and Japan, it is still under discussion. In addition, emissions in certain parts of the band (especially the $5.2-5.8$ GHz range used by wireless LANs) should be kept low, as UWB transceivers and IEEE 802.11a transceivers, which operate in the $5$ GHz range, are expected to work in close proximity. We are using two techniques in order to fulfill those requirements. - The first is a linear combination of a set of basis pulses to be used for shaping of the spectrum of a transmitted impulse radio signal. The delayed pulses are obtained from several appropriately timed programmable pulse generators. The computation of the delays and weights of those pulses is obtained in a two-step optimization procedure [@Wu_et_al_2003]. - A further improvement of the spectral properties can be obtained by exploiting different polarities of the pulses that constitute a transmit sequence $w_{\operatorname{seq}}(t)$. Using different pulse polarities does not change anything for the signal detection, as it is known at the receiver, and can thus be easily reversed. However, it does change the spectrum of the *emitted* signal, and thus allows a better matching to the desired frequency mask [@Nakache_and_Molisch_2003], [@Lehmann_and_Haimovich_2003]. The first technique (combination of pulses) leads to a shaping of the spectrum, allowing the placement of broad minima and an efficient filling out of the FCC mask. The second technique is used to reduce or eliminate the peak-to-average ratio of the spectrum, and allows the design of more efficient multiple-access codes. Note that these two aspects are interrelated, and the optimization of pulse combination and polarity randomization should be done jointly in order to achieve optimality. However, such a joint treatment is usually too complicated for adaptive modifications of the transmit spectrum. A further important aspect of the spectral shaping is that it can be used not only to reduce interference $to$ other devices, but also interference $from$ narrowband interferers. This can be immediately seen from the fact that matched filtering is used in the receiver. Placing a null in the transmit spectrum thus also means that the receiver suppresses this frequency. Furthermore, it might be advantageous in some cases to perform “mismatched filtering” at the receiver by placing minima in the receive transfer function even if there is no corresponding minimum in the transmit spectrum. This is useful especially for the suppression of narrowband interferers that could otherwise drive the A/D converter into saturation. Pulse combination ----------------- One of the key problems of conventional TH-IR radio is that it is difficult to influence its spectrum without the use of RF components. Spectral notches, e.g., are typically realized by means of bandblock filters. However, this is undesirable for extremely low cost applications; furthermore, it does not allow adaptation to specific interference situations. We have thus devised a new scheme for shaping the spectrum [@Wu_et_al_2003]. This scheme uses delaying and weighting of a set of basis pulses to influence the transmit spectrum, see Fig. \[pulsecombination\]. \[ptb\] [pulsecombination\_2.eps]{} The basic transmit waveform $w_{\text{tr}}(t)$ is a sum of delayed and weighted “basic pulse shapes” $p(t)$ that can be easily generated, e.g., Gaussian pulses and their derivatives. $$\begin{aligned} w_{tr}(t) & \equiv\sum_{i=0}^{M}u{_{i}}p(t-\xi_{i})\\ W(\text{j}\Omega) & \equiv\int_{-\infty}^{\infty}w_{tr}(t)e^{-j\Omega t}dt=\sum_{i=0}^{M}u{_{i}}P(j\Omega)e^{-j\Omega\xi_{i}},\label{eq:S_Omega}$$ where j is the imaginary unit (not to be confused with the index $j$ that denotes the considered frame), $u_{i}$ are the pulse weights, $W($j$\Omega)$ is the Fourier transform of $w_{\text{tr}}(t)$, and $\Omega$ is the transform variable. In contrast to tapped delay lines, where only certain discrete delays are feasible, we assume here that a continuum of delays can be chosen. This can be achieved by the use of programmable pulse generators. The range of allowed delays of the coefficients is determined by the pulse repetition frequency of the communication system. The number of pulse generators $M+1$ should be kept as low as possible to reduce the implementation costs. Let us introduce the following notations: $$\begin{aligned} & \underline{u}\equiv\lbrack u_{0}~u_{1}~\ldots~u_{M}]^{T}\label{eq:p}\\ & \underline{\xi}\equiv\lbrack\xi_{0}~\xi_{1}~\ldots~\xi_{M}]^{T}\\ & r(\lambda)\equiv\int_{-\infty}^{\infty}p(t-\lambda)p(t)dt=r(-\lambda),\\ & \mathbf{R}(\underline{\xi})\equiv\left( \begin{array} [c]{cccc}r(0) & r(\xi_{0}-\xi_{1}) & \cdots & r(\xi_{0}-\xi_{M})\\ r(\xi_{1}-\xi_{0}) & r(0) & \ddots & r(\xi_{1}-\xi_{M})\\ \vdots & \ddots & \ddots & \vdots\\ r(\xi_{M}-\xi_{0}) & r(\xi_{M}-\xi_{1}) & \cdots & r(0) \end{array} \right) ,\\ & \langle w_{tr}(t),w_{tr}(t)\rangle\equiv\int_{-\infty}^{\infty}w_{tr}(t)w_{tr}(t)dt=\underline{u}^{T}\mathbf{R}(\underline{\xi})\underline{u}$$ The single user spectrum shaping problem can now be formulated as follows: $$\max_{\underline{u},\underline{\xi}}\langle w_{tr}(t),w_{tr}(t)\rangle ,\;\text{subject to}\;|W(j\Omega)|^{2}\leq M(\Omega)\;,\forall\Omega\in \lbrack-\infty,\infty],$$ where $M(\Omega)$ is the upper-bound on the magnitude response regulated by FCC. This is equivalent to: $$\min_{\underline{u},\underline{\xi}}\max_{\Omega\in\lbrack-\infty,\infty ]}\frac{|W(j\Omega)|^{2}}{M(\Omega)},\qquad\text{subject to}\;\;\underline {u}^{H}R(\underline{\xi})\underline{u}=1.$$ The criteria for the optimization $M(\Omega)$ can thus stem from the FCC spectral mask, which is fixed, from the necessity to avoid interference to other users, which can be pre-defined or time-varying, or following an instantaneous or averaged determination of the emissions of users in the current environment, or other criteria. In any case, these criteria are mapped onto an instantaneous" spectral mask that has to be satisfied by the pulse. If the fulfillment of the FCC spectral mask is the only requirement, then the optimum weights can be computed a priori, and stored in the transceivers; in that case, the computation time determining the optimum weights and delays is not relevant, and exhaustive search can be used. However, in order to adjust to different interference environments, a capability to optimize the weights dynamically is desirable. This can be achieved, e.g., by an efficient two-step procedure that in the first step uses an *approximate* formulation of the optimization problem, namely 2-norm minimization that can be solved in closed form. This solution is then used as the initialization of a nonlinear optimization (e.g., by means of a neural network) to find the solution to the *exact* formulation. Details of this two -step procedure can be found in [@Wu_et_al_2003]. Note also that the spectral shaping can be refined even more by combining different basis pulses. However, this requires different pulse generators, which increases implementation complexity. Polarity randomization ---------------------- Conventional impulse radio systems use only a pseudo-random variation of the pulse position to distinguish between different users. For PAM - TH-IR, the spectrum of the transmit signal is determined by the spectrum of the transmit waveform $w_{tr}(t)$, multiplied with the spectrum of the TH sequence. Fig. \[spectrum\_ripples\] shows an example of a spectrum with a short (4 frames) time hopping sequence, in combination with a 5th-order Gaussian basis pulse. We can observe strong ripples, so that the peak-to-average ratio is about 6dB. However, the ideal case would be to find TH sequences whose spectrum is flat, so that the we can design the transmit waveform to fit the spectral mask as closely as possible. One way to achieve this goal is to use very long TH sequences (much longer than a symbol duration). However, this complicates the design of the receiver, especially the equalizer. Alternatively, we can use more degrees of freedom in the design of short sequences by allowing different amplitudes and polarities of the pulses for the design of the sequence. This helps to limit the power back-off by reducing the peak to average ratio. However, it is still true that the less pulses compose the sequence, the larger is the peak-to-average ratio. An example can be seen in Figs. \[spectrum\_ripples\] (unipolar sequence) and Fig. \[spectrum\_smoothed\] (polarity randomization); it is obvious that the ripples have been considerably reduced; specifically, we reduced the peak-to-average ratio by 1.6 dB. We also have to bear in mind that we need to generate a multitude of sequences that all should have the desired spectral properties, as well as approximate orthogonality with respect to each other for arbitrary time shifts of the sequences. This is a complex optimization problem, and has to be solved by an exhaustive search. \[ptb\] [ripplespectrum\_2.eps]{} \[ptbptb\] [smoothedspectrum\_2.eps]{} Signal detection** ** ===================== Received signal and Rake reception ---------------------------------- The Rake receiver is a key aspect of ultrawideband systems.[^4] Due to the ultra wide bandwidth, UWB systems have very fine temporal resolution, and are thus capable of resolving multi-path components that are spaced approximately at an inverse of the bandwidth. This is usually seen as a big advantage of UWB. Multipath resolution of components reduces signal fading because the multi-path components (MPCs) undergo different fading, and thus represent different diversity paths. The probability that the components are simultaneously all in a deep fade is very low. However, the fine time resolution also means that many of the MPCs have to be collected by the Rake receiver in order to obtain all of the available energy. A channel with $N_{\text{p}}$ resolvable paths requires $N_{\text{p}}$ fingers to collect all of the available energy. In a dense multi-path environment, the number of MPCs increases linearly with the bandwidth. Even a sparse environment, such as specified by the IEEE 802.15.3a standard channel model [@Molisch_et_al_2003_Mag], requires up to 80 fingers to collect 80% of the available energy. Another problem is the complexity of the Rake fingers. In the conventional Rake finger of a direct-sequence-spread spectrum (DS-SS) system, the received signal is filtered with a filter matched to the chip waveform, and then in each Rake finger, correlated to time-shifted versions of the spreading sequence. In order to do the correlation, the signal first has to be sampled and analog-to-digital (A/D) converted at the chip rate. Then, those samples have to be processed. This involves convolution with the stored reference waveform, addition, and readout. Sampling and A/D converting at the chip rate, e.g., $10$ Gsamples/s, requires expensive components.[^5] We avoid those problems by utilizing a Rake/equalizer structure as outlined in Fig. \[Rakestructure\]. Each Rake finger includes a programmable pulse generator, controlled by a pulse sequence controller. The signal from the pulse generator is multiplied with the received signal. The output of the multiplier is then sent through a low-pass filter, which generates an output proportional to a time integral of an input to the filter. The implementation is analogue, while the adjustable delay blocks have been eliminated. The hardware requirements for each Rake finger are: one pulse generator (which can be controlled by the same timing controller), one multiplier, and one sampler / AD converter. It is an important feature of this structure that the sampling occurs at the $symbol$ rate, not the chip rate. In the following, we assume the use of $10$ Rake fingers; this is a very conservative number. Obviously, a larger number of Rake fingers would give better performance; this is one of the complexity/performance trade-offs in our design [@Cassioli_et_al_2002_ICC], [@Choi_and_Stark_2002]. The weights for the combination of the fingers are determined by the channel estimation procedure described in Sec. V. \[ptb\] [Rakestructure\_2.eps]{} Next, we compute the output of the different Rake fingers. Let the impulse response of a UWB channel be $$h\left( t\right) =\sum\limits_{k}\alpha_{k}\delta\left( t-\tau_{k}\right) ,$$ where $\tau_{k}$ and $\alpha_{k}$ are the delay and (real) gain of the k-th path of the UWB channel, respectively. Then the channel output can be expressed as $$x\left( t\right) =h\left( t\right) \ast s_{tr}(t)+\overline{n}\left( t\right) =\sum\limits_{n=-\infty}^{\infty}b_{n}\widehat{h}\left( t-nT_{s}\right) +\overline{n}(t),\label{eq6}$$ where $$\widehat{h}\left( t\right) =\sum\limits_{k}\alpha_{k}w_{\text{tr}}\left( t-\tau_{k}\right) .$$ The output of the matched filter can be expressed as $$y\left( t\right) =x\left( t\right) \ast w_{tr}\left( {-t}\right) =\sum\limits_{k=-\infty}^{\infty}b_{k}\tilde{h}\left( t-kT_{s}\right) +\tilde{n}\left( t\right) ,\label{eq7}$$ where $$\tilde{h}\left( t\right) =\int\widehat{h}\left( t-\tau\right) w_{\text{tr}}\left( {-\tau}\right) d\tau=\sum\limits_{k}\alpha_{k}r\left( t-\tau_{k}\right) ,\label{eq8}$$$$r\left( t\right) =\int w_{\text{tr}}\left( t+\tau\right) w_{\text{tr}}\left( \tau\right) d{\tau,}$$ and $$\mbox{ }\tilde{n}\left( t\right) =\overline{n}\left( t\right) \ast w_{\text{tr}}\left( {-t}\right) .\label{eq9}$$ The samples of the matched filter output can be thus written as $$y[n]=y\left( {n\Delta}\right) =\sum\limits_{k=-\infty}^{\infty}{b_{k}\tilde{h}\left( {n\Delta-kp\Delta}\right) }+\tilde{n}\left( {n\Delta }\right) ,\label{eq10}$$ where $\Delta$ is the minimum time difference between Rake fingers and $p=T_{s}/\Delta.$ Combining of the Rake finger signals ------------------------------------ Let $\tilde{h}\left( n_{l}\Delta\right) $’s, for $l=1,{\ldots},L$ be the L taps with the largest absolute values, $\left\vert \tilde{h}\left( {n_{l}\Delta}\right) \right\vert $’s. The output of the Rake receiver can be expressed as $$z\left[ n,n_{o}\right] =\sum\limits_{l=1}^{L}\gamma_{l}y\left[ pn+n_{l}+n_{o}\right] ,\label{eq11}$$ where $\gamma_{l}$ is the weight for the $l$-th finger and $n_{o}$ is a time offset. It is obvious that the signal quality of the Rake receiver output depends on the weight and initial time offset. *Maximal ratio combining (MRC)* is a traditional approach to determine the weights of the Rake combiner. For the MRC Rake combiner, $\gamma _{l}=\tilde{h}\left( {n_{l}\Delta}\right) $, and $$z\left[ n,n_{o}\right] =\sum\limits_{l=1}^{L}\tilde{h}\left( n_{l}\Delta\right) y\left[ pn+n_{l}+n_{o}\right] .$$ *Minimum mean-square-error* (MMSE) Rake combining can improve the performance of the Rake receiver in the presence of interference, including intersymbol interference and multi-user interference since it automatically take the correlation of the interference into consideration. For the MMSE Rake combiner, the weights are determined to minimize $$E\left\vert z[n,n_{o}]-b_{n}\right\vert ^{2}.$$ The performance of the Rake receiver can be further improved if *adaptive timing* is used with the MMSE Rake combiner. That is, the goal is to find optimum time offset $n_{o}$ and $\gamma_{l}$ to minimize $$E\left\vert z[n,n_{o}]-b_{n}\right\vert ^{2}.$$ When there is co-channel interference, the received signal can be written as$$\bar{y}[n]=\sum\limits_{k=-\infty}^{\infty}{b_{k}}\tilde{h}\left( n\Delta-kT_{s}\right) +\underbrace{\sum\limits_{k=-\infty}^{\infty}\bar {b}_{k}\bar{h}\left( n\Delta-kT_{s}\right) +\tilde{n}\left( n\Delta\right) }_{i\left[ n\right] },$$ where $\{\bar{b}_{k}\}$ and $\bar{h}\left( n\Delta-kT_{s}\right) $ are, respectively, i.i.d. sequence and channel impulse response corresponding to the interferer, and $i[n]$ represents the interference-plus-noise. It can be shown that $i\left[ n\right] $ is not stationary but rather cyclo-stationary. Let $$P_{k}=E\left\{ \left\vert i\left[ mp+k\right] \right\vert ^{2}\right\} ,$$ for any integer m and $k=0,1,{\ldots},p-1$. Therefore, for different $k$, $h(nT_{s}+k\Delta)$ experiences different interference power. To improve the performance of the Rake receiver, we need to normalize the channel impulse response corresponding to the desired signal by $$\hat{h}\left( {n\Delta}\right) =\frac{\tilde{h}\left( {n\Delta}\right) }{\sqrt{P_{k}}},$$ and then find the $L$ taps with the largest absolute values of channel taps, $\left\vert {\hat{h}\left( {n_{l}\Delta}\right) }\right\vert $’s for the Rake receiver. Fig. \[interference\_suppression\] demonstrates the interference suppression performance for a UWB system with one interferer and 50 dB SNR. We compare the BER without normalization to the improved one that is normalized by noise power as described above. Note that this can also be interpreted as the difference between assuming the noise being stationary or cyclo-stationary. \[ptb\] [interference\_suppression\_2.eps]{} Channel equalizer ----------------- The combination of the channel and the Rake receiver constitutes an equivalent channel; however, since the symbol duration is shorter than the delay spread of the channel, intersymbol interference (ISI) does occur. We combat that by means of a MMSE (minimum mean square error) equalizer, as indicated in Fig. \[Rakestructure\]. The reasons for choosing a linear equalizer, instead of a DFE, are twofold: - the system is intended to operate at symbol error probabilities of 1-10%; strong coding is used to decrease the frame error probability. Thus, a decision feedback of the “raw symbols” (hard decision before the decoder) would result in strong error propagation. - the alternative to use the symbols after decision would require re-encoding and re-modulation before subtraction. This increases complexity considerably. As the ISI is not a dominant source of errors in our system (as determined from simulations that are not described in detail in Sec. VI), the possible gains from this improved DFE scheme do not warrant such an increase in complexity. After the Rake receiver, a linear equalizer is used to mitigate residual interference. Let the coefficients of the equalizer be $\left\{ c_{-K},c_{-K+1},....c_{-1},c_{0},c_{1},.....c_{K}\right\} $. Then the equalizer output is $$\tilde{b}[n]=\sum\limits_{k=-K}^{K}c_{k}z\left[ n-k,n_{o}\right] .$$ To optimize performance, the equalizer coefficients are chosen to minimize the MSE of its output, that is $$MSE=E\left\vert \tilde{b}[n]-b_{n}\right\vert ^{2}.$$ For the numerical simulations in Sec. VI, we will use a 5-tap equalizer. Parameter estimation ==================== A training sequence is used to determine the parameters for the Rake receivers and equalizers. It is desirable to use the correlators and A/D converters of the Rake receivers, since these components have to be available anyway. This is not straightforward, as the sampling and A/D conversion of the correlator outputs is done at the symbol rate, while the channel parameters have to be available for each possible chip sampling instant. This problem is solved by combining a “sliding correlator” approach with a training sequence that exhibits a special structure, as shown in Fig. \[training\_structure\]. \[ptb\] [training\_structure\_2.eps]{} Channel estimation ------------------ The matched filter in the Rake receiver in UWB systems is implemented using analog circuits since it needs to operate at a high speed. The output of the matched filter is sampled at symbol rate ($\mbox{1/T}_{\mbox{s}}=1/(p\Delta ))$. Therefore, during each symbol period, we can only observe $L$ outputs, each from one of $L$ fingers. On the other hand, we need to estimate channel coefficients every $\Delta$ seconds; thus we need to obtain $p$ uniform samples during each symbol period. In order to solve this seeming paradox, we use an approach that shows some similarity to the swept time delay cross correlator channel sounder proposed in [@Cox_1972]. We send the same training sequence (with guard interval) multiple times to obtain denser sampling of the matched filter output. For a Rake receiver with $10$ fingers, $10$ samples with different timings can be obtained within one symbol duration if the training sequence is sent once. Therefore, to get $32$ samples per symbol duration, the training sequence needs to be repeated $4$ times (see also Fig. \[training\_structure\]). Each training sequence consists of $511$ symbols, and $365$ ns guard interval to prevent interference caused by delay spread of UWB channels between adjacent training sequences. Consequently, the length of the whole training period for parameter estimation is $4(511\ast 5+365)=11600$ ns or $11.6$ $\mu$s. The detailed equations for the channel estimates can be found in the Appendix. Figs. \[channel\_estimation\_error1\], \[channel\_estimation\_error2\] shows the normalized MSE (NMSE) of our channel estimation, which is defined as$$NMSE=\frac{{\displaystyle\sum\limits_{n}} |\widetilde{h}(n\Delta)-h(n\Delta)|^{2}}{{\displaystyle\sum\limits_{n}} |h(n\Delta)|^{2}}\text{ \ .}$$ From Fig. \[channel\_estimation\_error1\], the channel estimation improves with the signal-to-noise ratio when it is less than 25 dB. However, when it is over 35 dB, there is an error floor. Fig. \[channel\_estimation\_error2\] shows the normalized MSE (NMSE) of the 10 largest channel taps, which is much better than the NMSE of overall channel estimation. \[ptb\] [channel\_estimation\_error1\_2.eps]{} \[ptbptb\] [channel\_estimation\_error2\_2.eps]{} After having obtained the channel estimates, we determine the optimum Rake combining weights by minimizing the mean-square error. The concatenation of channel and Rake receiver constitutes a “composite” channel that is sampled once per symbol. The equalizer is adapted such that it minimizes the mean-square error of the equalizer output compared to a special training sequence that is transmitted after the Rake weights have been adjusted. Detailed equations about the weights for Rake and equalizer can be found in the appendix. Synchronization --------------- Before any data demodulation can be done on the received UWB signal, the template signal and the received signal must be time-aligned. The aim of acquisition is to determine the relative delay of the received signal with respect to the template signal. The conventional technique to achieve this is the serial search algorithm. In this scheme, the received signal is correlated with a template signal and the output is compared to a threshold. If the output is lower than the threshold, the template signal is shifted by some amount, which usually is comparable to the resolvable path interval and the correlation with the received signal is obtained again. In this way, the search continues until an output exceeds the threshold. If the output of the correlation comes from a case where signal paths and the template signal are aligned, it is called a signal cell output. Otherwise, it is called a non-signal cell output. A false alarm occurs when a non-signal cell output exceeds the threshold. In this case, time $t_{p}$ elapses until the search recovers again. This time is called penalty time for false alarm. However, in UWB systems, such a sequential search can be very time consuming, as the number of cells is very large. This problem can be overcome by a new algorithm that we call “sequential block search”. The key idea here is to divide the possible search space, which contains the cells, into several blocks, where each of the blocks contains a number of signal cells. We then first perform a quick test to check if the whole block contains a signal cell, or not. Once we have identified the block that contains the signal, a more detailed (sequential) search is performed in that block; for details, see [@Gezici_et_al_2003]. Simulations show that acquisition can be achieved (with 90% probability) in less than $10\mu s$.[^6] This can be shortened even further if the search space is restricted, e.g., by exploiting knowledge from a beacon signal. Performance results =================== In this section, we analyze the performance of our system in multipath and interference. The performance of the system was simulated in typical UWB channels, which were developed within the IEEE 802.15.3a UWB standardization activities and are described in detail in [@Molisch_et_al_2003_Mag]. We distinguish between four different types of channels (called CM1, CM2, CM3, and CM4). CM1 describes line-of-sight (LOS) scenarios with distances between TX and RX of less than $4$ m; CM2 and CM3 describe non-LOS scenarios at distances $0-4$, and $4-10$ m, respectively. CM4 is valid for heavy multipath environments. Note that in the following, we will plot the performance in all the four different types of channels over a wide range of distances. [^7] \[ptb\] [link\_success\_100\_2.eps]{} Fig. \[link\_success\_100\] shows the probability for obtaining a successful link. A successful link means that acquisition is obtained successfully, and the packet error probability (over the ensemble of different channels) is less than 8[%]{}. For CM1, the mean coverage distance[^8] is about $10$ m. The 10[%]{} outage distance (meaning that 8[%]{} packet error rate or less is guaranteed in 90[%]{} of all channels) is $7$ m. For heavy multipath (CM4) these values decrease to $7$ and $4$ m, respectively. \[ptb\] [link\_success\_200\_2.eps]{} Fig. \[link\_success\_200\] shows the analogous curves for data rate of $200$ Mbit/s. Due to the higher rate, the original data stream is converted (demultiplexed) into two parallel data streams with $100$ Mbit/s each. The two data streams are then transmitted simultaneously, using time hopping codes that have the same hopping sequence, but are offset in delay by one chip. In an AWGN channel, those codes would remain orthogonal, and the performance should be worsened only by $3$ dB (since the E$_{b}$/N$_{0}$ is decreased). However, in a multipath channel, the temporally offset codes lose their orthogonality, which worsens the performance. One way to remedy this situation is to use different (not just offset) hopping codes. However, this decreases the number of possible simultaneous piconets. Another approach would be the use of the scheme of [@Yang_and_Giannakis_2002], which retains the orthogonality of codes even in delay-dispersive channels. \[ptb\] [SOP\_CM1.eps]{} \[ptbptb\] [SOP\_CM3.eps]{} Figs. \[SOP\_CM1\] and \[SOP\_CM4\] show the performance when two users (independent piconets) are operating simultaneously. The desired users are located at half the distance that gives the 90[%]{} outage probability (i.e., there is a $6$ dB margin[^9] with respect to the single-user case); shadowing is not considered in that graph. We find that an interfering piconet can be at a distance from the victim receiver of about $1$ m (if the desired piconet is operating in CM1 or CM2), or $1.5$ m (if the desired piconet is operating in CM3 or CM4). The performance does not depend on which channel model is used for the interfering piconet. Table 1 shows the coexistence of our system with other communications devices, obeying various narrowband standards. In the column “desired”, we list the interference power that must not be exceeded according to the specifications of the IEEE 802.15.3a technical requirements documentation (this power is derived from receiver sensitivity specifications for various systems). In the “achieved” column, we list the interference power (within the victim receiver bandwidth) received from our UWB transmitter spaced at $1 $m distance from the victim receiver. The column “FCC mask” gives the interference power created by a UWB transmitter (at $1$m distance) that transmits at all frequencies with the maximum power allowed by the FCC mask. We find that if the UWB transmitter emits with the full power allowed by the FCC, it can significantly interfere with other communications devices. A suppression of about 15dB is necessary to allow coexistence within a 1m range. We achieve this suppression with the spectral shaping as described in Sec. 3.3. Finally, we also analyzed the resistance of the UWB system to interference *from* other communications devices. We found that again, a minimum distance of 1m is sufficient to allow operation with less than $8{\%}$ PER. Summary and conclusions ======================= We have presented a UWB communications system based on time-hopping impulse radio. This system uses only baseband components, while still being compatible with FCC requirements, and providing a flexible shaping of the transmit spectrum in order to accommodate future requirements by other spectrum governing agencies, as well as not interfere with 802.11a wireless LANs and other communications receivers in the microwave range. Our system can sustain data rates of $110$ Mbit/s at $15$ m in AWGN channels, and $4-7$ m in multipath channels. It is also resistant to interference from other UWB users, as well as interference from wireless LANs, microwave ovens, and other interferers. \[c\][|p[53pt]{}|l|l|l|]{}*System* & *Desired* & *Achieved* & *FCC Mask*\ $802.11a$ & *-88dBm* & *-90dBm* & *-75dBm*\ $802.11b$ & *-82dBm* & *-85dBm* & *-70dBm*\ $802.15.1$ & *-76dBm* & *-95dBm* & *-80dBm*\ $802.15.3$ & *-81dBm* & *-85dBm* & *-70dBm*\ $802.15.4$ & *-91dBm* & *-95dBm* & *-80dBm*\ \[tab1\] Table 1: Coexistence for other systems Appendix A Parameter estimation =============================== To obtain uniform samples, the timing of the $l$-th finger corresponding to the $m$-th training sequence is adjusted as follows: $$t_{l,m}=4(l-1)\Delta+(m-1)\Delta,$$ for $l=1,\cdots,10,$ and $m=1,\cdots,4.$ Let the training sequence be $b_{k}^{t}$’s for k=0, 1, […]{}, 510, where superscript $^{t}$ denotes “training”. Then the training signal can be expressed as $$s^{t}\left( t\right) =\sum\limits_{k=0}^{510}b_{k}^{t}w\left( t-kT_{s}\right) ,$$ From Equation (\[eq10\]), the $\Delta$-spaced output of the matched filter is $$y^{t}\left( n\Delta\right) =\sum\limits_{k=0}^{510}b_{k}^{t}\tilde{h}\left( n\Delta-kp\Delta\right) +\tilde{n}\left( n\Delta\right) .$$ Consequently, the estimated channel taps can be expressed as$$\tilde{h}\left( n\Delta\right) =\frac{1}{511}\sum\limits_{k=0}^{510}b_{k}^{t}y^{t}\left( n\Delta+kp\Delta\right) .$$ It can be shown that$$\begin{aligned} \tilde{h}\left( n\Delta\right) & =h\left( n\Delta\right) +\frac{1}{511}\sum\limits_{m=0}^{510}\left( \sum\limits_{k=m}^{510}b_{k}^{t}b_{k-m}^{t}\right) h\left( n\Delta+mp\Delta\right) \\ & +\frac{1}{511}\sum\limits_{m=-510}^{0}\left( \sum\limits_{k=0}^{510-m}b_{k}^{t}b_{k-m}^{t}\right) h\left( n\Delta+mp\Delta\right) +\frac{1}{511}\sum\limits_{k=0}^{510}b_{k}^{t}\tilde{n}\left( n\Delta+kp\Delta\right) \text{ \ .}\nonumber\end{aligned}$$ The 2nd and 3rd terms in the above equation are the perturbations from other taps due to imperfect orthogonality of the training sequence and the 4th term presents the effect of channel noise. To exploit the improved approach for UWB systems with co-channel interference, interference power has to be estimated. Using the estimated channel and the training sequence, the interference can be estimated by $$i_{t}\left[ n\right] =y_{t}\left( n\Delta\right) -\sum\limits_{k=0}^{510}b_{k}^{t}\tilde{h}\left( n\Delta-kp\Delta\right) ,$$ and from it, interference-plus-noise power can be estimated by $$P_{k}=\frac{1}{511}\sum\limits_{m=0}^{510}\left\vert i_{t}\left[ mp+k\right] \right\vert ^{2},$$ for $k=0,1,{\ldots},(p-1).$ Next, we determine the Rake weights. Let $n_{1},\cdots,n_{L}$ be the indices of the $L$ largest taps. Then the weights for the MMSE Rake combiner and optimum timing can be found by minimizing $$MSE(\vec{\gamma},n_{0})=\frac{1}{511}\sum\limits_{n=0}^{510}\left\vert z_{t}(n,n_{0})-b_{n}^{t}\right\vert ^{2}=\frac{1}{511}\sum\limits_{n=0}^{510}\left\vert \sum\limits_{l=1}^{10}\gamma_{l}y_{t}(pn+n_{l}+n_{0})-b_{n}^{t}\right\vert ^{2}$$ Direct least-squares calculation yields that [@Haykin_1996] $$\vec{\gamma}=\left( {{\begin{array}{*{20}c} {\gamma _1 } \hfill \\ \vdots \hfill \\ {\gamma _{10} } \hfill \\ \end{array}}}\right) =\left( {\mathrm{\mathbf{Y}}_{t}\mathrm{\mathbf{Y}}_{t}^{H}}\right) ^{-1}\left( {\mathrm{\mathbf{Y}}_{t}\mathrm{\mathbf{b}}_{t}^{H}}\right) ,$$ where $$\mathrm{\mathbf{Y}}_{t}=\left( {{\begin{array}{*{20}c} {y_t \left[ {n_1 +n_o } \right]} \hfill & \hfill & \cdots \hfill & {y_t \left[ {510p+n_1 +n_o } \right]} \hfill \\ {y_t \left[ {n_2 +n_o } \right]} \hfill & \hfill & \cdots \hfill & {y_t \left[ {510p+n_2 +n_o } \right]} \hfill \\ \vdots \hfill & \hfill & \cdots \hfill & \vdots \hfill \\ {y_t \left[ {n_{10} +n_o } \right]} \hfill & \hfill & \cdots \hfill & {y_t \left[ {510p+n_{10} +n_o } \right]} \hfill \\ \end{array}}}\right) ,$$ and $$\mathrm{\mathbf{b}}_{t}=\left( {{\begin{array}{*{20}c} {b_0^t } \hfill & {b_1^t } \hfill & \cdots \hfill & {b_{510}^t } \hfill \\ \end{array}}}\right) .$$ From the estimated weights for the Rake receiver, its output can be calculated by $$z_{t}\left[ n,n_{o}\right] =\sum\limits_{l=1}^{10}\gamma_{l}y_{t}\left[ pn+n_{l}+n_{o}\right] .$$ The equalizer coefficients can be estimated by minimizing $$\frac{1}{511}\sum\limits_{n=0}^{511}\left\vert \sum\limits_{k=-L}^{L}c_{k}z_{t}[n-k,n_{o}]-b_{k}^{t}\right\vert ^{2}.$$ Consequently [@Haykin_1996], $$\left( {{\begin{array}{*{20}c} {c_{-2} } \hfill \\ \vdots \hfill \\ {c_2 } \hfill \\ \end{array}}}\right) =\left( {\frac{1}{511}\sum\limits_{k=0}^{510}{\mathrm{\mathbf{z}}_{k}^{t}\mathrm{\mathbf{z}}_{k}^{t}{}^{T}}}\right) ^{-1}\left( {\frac {1}{511}\sum\limits_{k=0}^{510}{\mathrm{\mathbf{z}}_{k}^{t}b_{k}^{t}}}\right) ,$$ where $$\mathrm{\mathbf{z}}_{k}^{t}=\left( {{\begin{array}{*{20}c} {z_t \left[ {k+2,n_o } \right]} \hfill \\ \vdots \hfill \\ {z_t \left[ {k-2,n_o } \right]} \hfill \\ \end{array}}}\right) .\label{eq12}$$ [^1]: $^{2}$Author for correspondence. Email: [email protected]. [^2]: $^{3}$A. F. Molisch, Y. P. Nakache, P. Orlik, and J. Zhang are with Mitsubishi Electric Research Labs, Cambridge, MA, USA. Y. G. Li is with Georgia Institute of Technology, Atlanta, GA, USA. M. Miyake is with Information Technology R&D Center, Mitsubishi Electric Corp., Ofuna, Japan..Y. Wu, S. Gezici, S. Y. Kung, H. Kobayashi, and H. V. Poor are with Princeton University, Princeton, NJ, USA. H. Sheng and A. Haimovich are with New Jersey Institute of Technology, Newark, NJ, USA. [^3]: $^{1}$Part of this work was presented at WPMC 2003, Yokosuta, Japan. [^4]: An exception is OFDM-based UWB systems, which use a different principle to collect the multipath energy [@Batra_et_al_2003]. [^5]: Note that some companies have proposed the use of *one-bit* A/D converters with $7.5-20$ Gsamples per second [@McCorkle_et_al_2003]. [^6]: Note that the treshold whether detection has taken place or not is a critical parameter of the algorithm. A discussion of how to set this threshold can be found in [@Gezici_et_al_2003]. [^7]: We also evaluate the performance at distances that the IEEE models were not originally intended for (e.g., CM1 was extracted from measurements where the distance between TX and RX is less than 4m). We do this as it gives insights into the relative importance of delay dispersion and attenuation. [^8]: The mean coverage distance is defined as the distance where the packet error rate, averaged over all channel realizations, is below the target rate. [^9]: As the channel model prescribes the received power to be proportional to $d^{-2}$, halving the distance means increasing the power by $6$dB.

ArXiv

--- abstract: 'A novel framework for consensus clustering is presented which has the ability to determine both the number of clusters and a final solution using multiple algorithms. A consensus similarity matrix is formed from an ensemble using multiple algorithms and several values for $k$. A variety of dimension reduction techniques and clustering algorithms are considered for analysis. For noisy or high-dimensional data, an iterative technique is presented to refine this consensus matrix in way that encourages algorithms to agree upon a common solution. We utilize the theory of nearly uncoupled Markov chains to determine the number, $k$ , of clusters in a dataset by considering a random walk on the graph defined by the consensus matrix. The eigenvalues of the associated transition probability matrix are used to determine the number of clusters. This method succeeds at determining the number of clusters in many datasets where previous methods fail. On every considered dataset, our consensus method provides a final result with accuracy well above the average of the individual algorithms.' author: - 'Carl Meyer, Shaina Race' title: A Flexible Iterative Framework for Consensus Clustering --- Introduction ============ Cluster analysis is an important tool used in hundreds of data mining applications like image segmentation, text-mining, genomics, and biological taxonomy. Clustering allows users to find and explore patterns and structure in data without prior knowledge or training information. Hundreds (if not thousands) of algorithms exist for this task but no single algorithm is guaranteed to work best on any given class of real-world data. This inconsistency of performance in cluster analysis is not unique to the clustering algorithms themselves. In fact, the dimension reduction techniques that are expected to aid these algorithms by revealing the cluster tendencies of data also tend to compete unpredictably, and it is difficult to know beforehand which low-dimensional approximation might provide the best separation between clusters. Having many tools and few ways to make an informed decision on which tool to use, high-dimensional cluster analysis is doomed to become an ad hoc science where analysts blindly reach for a tool and hope for the best. Cluster analysis is not the first type of data mining to encounter this problem. Data scientists were quick to develop ensemble techniques to escape the unreliability of individual algorithms for tasks like prediction and classification. Ensemble methods have become an integral part of many areas of data mining, but for cluster analysis such methods have been largely ignored. An additional problem stems from the fact that the vast majority of these algorithms require the user to specify the number of clusters for the algorithm to create. In an applied setting, it is unlikely that the user will know this information before hand. In fact, the number of distinct groups in the data may be the very question that an analyst is attempting to answer. Determining the number of clusters in data has long been considered one of the more difficult aspects of cluster analysis. This fact boils down to basics: what is a cluster? How do we define what should and should not count as two separate clusters? Our approach provides an original answer this question: A group of points should be considered a cluster when a variety of algorithms *agree* that they should be considered a cluster. If a majority of algorithms can more or less agree on how to break a dataset into two clusters, but cannot agree on how to partition the data into more than two clusters, then we determine the data has two clusters. This is the essence of the framework suggested herein. Our purpose is to address both problems: determining the number of clusters and determining a final solution from multiple algorithms. We propose a consensus method in which a number of algorithms form a voting ensemble, proceeding through several rounds of elections until a majority rule is determined. This allows the user to implement many tools at once, increasing his or her confidence in the final solution. Consensus Clustering ==================== Previous Proposals for Consensus Clustering {#oldconsensus} ------------------------------------------- In recent years, the consensus idea has been promoted by many researchers [@consensusberman; @consensusgene; @consensusspectral; @chuck; @consensusnam; @consensuszeng; @consensusLS; @consensusjain; @consensusfern; @consensusfilkov; @consensusstrehl; @amaral; @mirkin2]. The main challenge to ensemble methods using multiple algorithms is generally identified to be the wide variety in the results produced by different algorithms due to the different cluster criteria inherent in each algorithm. Thus any direct combination of results from an ensemble will not often generate a meaningful result [@consensusqian; @consensuskot]. Most often the consensus problem has been formulated as an optimization problem, where the optimal clustering, ${\mathscr{C}}^*$, minimizes some relative distance metric between ${\mathscr{C}}^*$ and all of the clusterings ${\mathscr{C}}_i$ in the ensemble. There are many ways to define the distance between two clusterings, for example one could take the minimum number of elements that need to be deleted for the two partitions to become identical [@gusfield]. Using $d({\mathscr{C}}_1,{\mathscr{C}}_2)$ to denote some measure of distance between two different clusterings, we’d write $${\mathscr{C}}^*=\operatorname*{arg\,min}_{{\mathscr{C}}} \sum_{i=1}^N d({\mathscr{C}}_i,{\mathscr{C}}). \label{medianobj}$$ This problem is known as the *median partition problem* in the literature and dates back to the 1965-‘74 work of R[é]{}gnier ([@regnier]) and Mirkin ([@mirkin]) [@consensusfilkov]. Alternatively, some authors use a relative validity metric like the normalized mutual information $NMI({\mathscr{C}}_i,{\mathscr{C}})$ in place of a distance function and attempt to maximize the objective function in [Eq. \[medianobj\]]{} [@consensusstrehl]. The median partition problem was shown by Krivanek and Moravek, and also by Wakabayashi, to be NP-complete [@consensusfilkov], but many heuristics have since been proposed to find approximate solutions [@consensusstrehl; @consensusfilkov; @consensusberman; @consensusLS]. We believe that these methods are bound to suffer because each clustering in the ensemble is given equal importance. Suppose we had 4 perfect clusterings and 1 terribly inaccurate clustering. These methods would not take into account the fact that the majority of the algorithms share 100% agreement on a perfect clustering, and instead may shift the optimal clustering away from perfection towards inaccuracy. Thus, we feel that the optimization in [Eq. \[medianobj\]]{} leads to a “middle-of-the-road” solution or a *compromise* between algorithms, rather than a solution of “agreement” or consensus. In our method, the clustering algorithms act as a voting ensemble and continually move through a series of elections until a desired level of consensus is reached. Additionally, we introduce a parameter of intolerance, which allows the user to impose a level of agreement that must be reached between algorithms in order to accept a cluster relationship between objects. The Consensus Matrix {#consensusmatrix} -------------------- To begin, we introduce some notation. Since consensus methods combine multiple solutions from multiple algorithms (or multiple runs of the same algorithm), we start with a **cluster ensemble**. A cluster ensemble, ${\mathscr{C}}=\{{\mathscr{C}}_1,{\mathscr{C}}_2,\dots,{\mathscr{C}}_N\}$, is a set of $N$ clusterings of the $n$ data objects ${\textbf{x}}_1,{\textbf{x}}_2,\dots,{\textbf{x}}_n$. That is, each clustering ${\mathscr{C}}_j$ in the ensemble is a $k_j$-way partition of the data, composed of individual clusters, $${\mathscr{C}}_j = [C_1,C_2, \dots, C_{k_j}],$$ where the number of clusters $k_j$ in each clustering may be allowed to vary. In [Figure \[ensembleex\]]{}, we illustrate a simple example with $N=3$ clusterings. ![Example of an Ensemble of $N=3$ Clusterings[]{data-label="ensembleex"}](ensembleex.pdf){width="0.7\linewidth"} The information from a cluster ensemble is then recorded in a **consensus matrix**. Given a cluster ensemble, $${\mathscr{C}}=\{{\mathscr{C}}_1,{\mathscr{C}}_2,\dots,{\mathscr{C}}_N\},$$ of $n$ data points ${\textbf{x}}_1,{\textbf{x}}_2,\dots,{\textbf{x}}_n$, the **consensus matrix** ${\textbf}{M}$ is an $n\times n$ matrix such that $${\textbf{M}}({\mathscr{C}})_{ij} = \# \mbox{ of times object } {\textbf{x}}_i \mbox{ was placed in the same cluster as } {\textbf{x}}_j \mbox{ in the ensemble }{\mathscr{C}}.$$ \[cmatrix\] One might prefer to think of the consensus matrix as the sum of individual adjacency matrices for each clustering in the ensemble. For a given clustering ${\mathscr{C}}_i$ we could define an adjacency matrix, ${\textbf{A}}_i$ as $${\textbf{A}}_{ij} = \left\{ \begin{array}{ll} 1 &\mbox{ if object } {\textbf{x}}_i \mbox{ was clustered with } {\textbf{x}}_j \\ 0 &\mbox{ otherwise } \end{array} \right.$$ Then the consensus matrix ${\textbf{M}}$ would be the sum of the adjacency matrices of each clustering in the ensemble: $${\textbf{M}}({\mathscr{C}}) = \sum_{i=1}^N {\textbf{A}}_i.$$\ As an example, the consensus matrix for the ensemble depicted in [Figure \[ensembleex\]]{} is given in [Figure \[consensusex\]]{}. ![The Consensus Matrix for the Ensemble in [Figure \[ensembleex\]]{}[]{data-label="consensusex"}](consensusex.pdf){width="0.5\linewidth"} The consensus matrix from [Figure \[consensusex\]]{} is very interesting because the ensemble that was used to create it had clusterings for various values of $k$. The most reasonable number of clusters for the colored circles in [Figure \[ensembleex\]]{} is $k^*=3$. The 3 clusterings in the ensemble depict $k_1=3, k_2=4, \mbox{ and } k_3=5$ clusters. However, the resulting consensus matrix is clearly block-diagonal with $k^*=3$ diagonal blocks! This is not an isolated phenomenon, in fact it is something we should expect from our consensus matrices if we labor under the following reasonable assumptions: - If there are truly $k$ distinct clusters in a given dataset, and a clustering algorithm is set to find $\tilde{k}>k$ clusters, then the $k$ “true” clusters will be broken apart into smaller clusters to make $\tilde{k}$ total clusters. - Further, if there is no clear “subcluster" structure, meaning the original $k$ clusters do not further break down into meaningful components, then different algorithms will break the clusters apart in different ways. This block-diagonal structure is the subject of [Section \[perron\]]{}. ### Benefits of the Consensus Matrix {#benefits} As a similarity matrix, the consensus matrix offers some benefits overs traditional approaches like the Gaussian or Cosine similarity matrices. One problem with these traditional methods is the curse of dimensionality: In high dimensional spaces, distance and similarity metrics tend to lose their meaning. The range of values for the pairwise distances tightens as the dimensionality of the space grows, and little has been done to address this fact. In [Figure \[matrixdist\]]{} we show the distribution of similarity values for the same 1 million entries in a consensus matrix compared to the cosine similarity matrix. As you can see, the consensus approach allows a user to witness some very high levels of similarity in high-dimension data, whereas the cosine similarities have a much smaller range. The dataset, which is more formally introduced in [Section \[mcc\]]{}, is the Medlars-Cranfield-CISI document collection ($\approx 4,000$ documents) [@kogan]. Such contrast is typical among high-dimensional datasets. [.49]{} ![Distribution of Similarity Values in Cosine vs. Consensus Matrix[]{data-label="matrixdist"}](MCCcoshist.pdf "fig:"){width="0.97\linewidth"} [.49]{} ![Distribution of Similarity Values in Cosine vs. Consensus Matrix[]{data-label="matrixdist"}](MCCconshist.pdf "fig:"){width="0.97\linewidth"} An additional benefit is that entries in the consensus matrix have depth. By this, we mean that they result from summing entries in adjacency matrices output by individual clustering algorithms, so more information is available about the *meaning* of each similarity value. The cosine of the angle between two data vectors ${\textbf{x}}_i$ and ${\textbf{x}}_j$ may tell us something about their correlation, but knowing, for instance that these two objects were clustered together by all algorithms with $\tilde{k} \leq 5$, by some algorithms with $6\leq\tilde{k}\leq7$, and never when $\tilde{k}\geq 7$, provides a depth of insight not previously considered. While we do not use this information explicitly in our analysis, it may be beneficial in practical research. The greatest benefit of using a consensus matrix for clustering is that it provides superior information about clustering within the data. This has been demonstrated time and again in the literature [@chuck; @phdthesis; @siamicc; @consensusfred; @consensusgene; @consensusjain; @consensusspectral; @consensusstrehl; @jain50; @mirkin2]. We add to the pile of evidence for this statement with the experiments in this paper. Iterative Consensus Clustering (ICC) {#icc} ==================================== The consensus approach outlined herein is based on the work in [@consensusspectral; @consensusjain; @thesis; @chuck; @consensusfred] where the consensus matrix is treated as similarity matrix and used as input to a clustering algorithm to reach a final solution. In [@consensusspectral] the authors suggest using many runs of the [*k*-means]{}algorithm, initialized randomly, to build the consensus matrix and then using a spectral clustering method, such as normalized cut (NCut) [@shi], to determine a final solution. In [@consensusjain; @jain50; @consensusfred], the approach is again to build a consensus matrix using many runs of the [*k*-means]{}algorithm and then to cluster the consensus matrix with one final run of [*k*-means]{}. In [@chuck] a consensus matrix is formed via [*k*-means]{}and then used as input to the stochastic clustering algorithm (SCA). While all these methods provide better results than individual algorithms, they still rely on a single algorithm to make both the consensus matrix and the final decision on cluster membership. Our method uses a variety of algorithms, rather than just [*k*-means]{}, to create the initial cluster ensemble. In addition, each algorithm is paired with different dimension reductions because it is often unclear which dimension reduction gives the best configuration of the data; each lower dimensional representation has the potential to separate different sets of clusters in the data. In this way, we essentially gather an initial round of *votes* for whether or not each pair of objects $({\textbf{x}}_i,{\textbf{x}}_j)$ belong in the same cluster. These votes are collected in a consensus matrix ${\textbf{M}}$ as defined in Definition \[cmatrix\]. Clustering Algorithms Used for Ensembles {#algs} ---------------------------------------- The experiments contained herein use a number of different clustering algorithms. The details of these popular algorithms are omitted due to space considerations but we refer the reader to the following resources: #### Data Clustering Algorithms 1. Spherical $k$-means (solution with lowest objective value from 100 runs, randomly initialized) [@datamining] 2. PDDP: Principal Direction Divisive Partitioning [@pddp] 3. PDDP-$k$-means: $k$-means with initial centroids provided by PDDP clusters [@phdthesis] 4. NMFcluster: $k$ dimensional Nonnegative Matrix Factorization for Clustering [@NMFcluster] #### Graph Clustering Algorithms (Requiring a similarity matrix) 1. PIC: Power Iteration Clustering [@poweriteration] 2. NCUT: Normalized Cuts according to Shi and Malik [@shi] 3. NJW: Normalized Cuts according to Ng, Jordan, and Weiss [@ng] Each algorithm in our ensemble is assumed to be *reasonable*, making good choices on cluster membership most of the time. It is inevitable that each of the clustering algorithms will make mistakes, particularly on noisy data, but it is assumed that rarely will the majority of algorithms make the *same* mistake. To account for this error, we introduce an *intolerance parameter*, $\tau$, for which entries in the consensus matrix ${\textbf{M}}_{ij} < \tau N$ will be set to zero. In other words, $\tau$ is the minimum proportion of algorithms that must agree upon a single cluster relationship $({\textbf{x}}_i,{\textbf{x}}_j)$ in order to keep those “votes” in the consensus matrix. After the initial consensus matrix is formed, we use it as input to each of the clustering algorithms again. Essentially we start a debate between algorithms, asking each of them to use the collective votes of the ensemble to determine a second solution. Again these solutions are collected in a consensus matrix and the process repeats until a simple majority of algorithms agree upon one solution. Once the majority of algorithms have chosen a common solution, we say the algorithms have *reached consensus* and call that resulting solution the **final consensus solution**. This process is illustrated in the flow chart in [Figure \[consensusflowchart\]]{}. ![Iterated Consensus Clustering (ICC) Process[]{data-label="consensusflowchart"}](consensusflowchart.pdf){width="0.7\linewidth"} Example with Medlars-Cranfield-CISI Text Collection {#mcc} --------------------------------------------------- To illustrate the effectiveness of this procedure, we consider a combination of 3 text datasets used frequently in the information retrieval literature. For simplicity, we assume the number of clusters is known a priori. In [Section \[perron\]]{} this information will be extracted from the data. The combined Medlars-Cranfield-CISI (MCC) collection consists of nearly 4,000 scientific abstracts from 3 different disciplines. These 3 disciplines (Medlars = medicine, Cranfield = aerodynamics, CISI = information science) form 3 natural clusters in the data [@surveytextmining; @kogan]. The document data is high-dimensional with $m \approx 11,000$ features (words). As a result, clustering algorithms tend to run slowly on the raw data. Thus, we reduce the dimensions of the data using 3 preferred algorithms: - Nonnegative Matrix Factorization (NMF) by Alternating Constrained Least Squares (ACLS) [@AppliNMF] - Singular Value Decomposition (SVD) [@drineassvd; @LSI] - Principal Components Analysis (PCA) [@PCA] One should realize that PCA is, in fact, a Singular Value Decomposition of Data under z-score normalization. However, in practice, these two decompositions generally provide quite different results, particularly for high-dimensional data. For each dimension reduction technique, the number of features is reduced from $m = 11,000$ to $r=5, 10,\mbox{ and } 20$ creating a total of 9 input data sets. On each input dataset, 3 different clustering methods were used to cluster the data: - $k$-means - PDDP - PDDP-$k$-means The accuracy (proportion of documents classified correctly [@phdthesis]) of each algorithm on each data input are given in [Table \[mccresults\]]{}. [|c | l| c| c| c|]{} \# Features & Algorithm & NMF input & SVD input & PCA input\ & [*k*-means]{}&0.8741& 0.7962&0.8260\ ($r=5$) & PDDP & 0.9599& 0.9049& 0.9026\ & PDDP-[*k*-means]{}&0.9599& 0.9049& 0.9026\ & [*k*-means]{}& 0.8628&0.8268&0.8286\ ($r=10$) & PDDP & 0.9764& 0.9774& 0.9481\ & PDDP-[*k*-means]{}&0.9764& 0.9774&0.9481\ & [*k*-means]{}& 0.8530&0.8263&0.8281\ ($r=20$) & PDDP & 0.9722&0.9802&0.9478\ & PDDP-[*k*-means]{}&0.6114&0.9802&0.9478\ \ The accuracy of the results ranges from 61% ($>$1,500 misclassified) to 98% (78 misclassified). A reasonable question one might ask is this: Why not choose the solution with the lowest [*k*-means]{}objective value? The biggest problem with this boils down to the curse of dimensionality: the distance measures used to compute such metrics lose their meaning in high-dimensional space [@phdthesis; @mirkin2]. The only comparison we could make between clusterings would be with the full dimensional data and, surprisingly, the objective function values for the minimum accuracy solution is approximately equal to the maximum accuracy solution! Other internal metrics, like the popular Silhouette coefficient [@datamining] suffer from the same problem. One must be very careful when attempting to compare high-dimensional clusterings with such metrics. Our suggestion is instead to compile the clusterings from [Table \[mccresults\]]{} into a consensus matrix, cluster that consensus matrix with multiple algorithms, and repeat that process until the majority of the algorithms agree upon a solution. This can be done with or without dimension reduction on the consensus matrix. For simplicity, we’ll proceed without reducing the dimensions of the consensus matrix, but we will include an additional clustering algorithm, NMFCluster, which was not well suited for the analysis on the low-dimensional representations in [Table \[mccresults\]]{}. [Table \[mccconsensus\]]{} provides the accuracy of these 4 clustering algorithms on the consensus matrices through iteration. Boxes are drawn around values to indicate a common solution chosen by algorithms. A final consensus solution is found in the third iteration with 3 of the 4 algorithms agreeing upon a single solution. The accuracy of this final consensus solution is much greater than the average of all the initial clustering results in [Table \[mccresults\]]{}. Such a result is typical across all datasets considered. Algorithm Consensus Iter 1 Consensus Iter 2 Consensus Iter 3 -------------------- ------------------ ------------------ ------------------ PDDP 0.939 0.969 0.969 PDDP-[*k*-means]{} 0.954 NMFcluster 0.969 0.954 [*k*-means]{} 0.966 : Medlars-Cranfield-CISI text collection: Accuracies for 4 Clustering Algorithms on Consensus Matrices through Iteration[]{data-label="mccconsensus"} Perron Cluster Analysis {#perron} ----------------------- In [Section \[consensusmatrix\]]{} an example was given that alluded to our methodology for determining the number of clusters. We approach this task using Perron-cluster methodology [@chuck; @fischer; @perroncluster; @siamicc; @phdthesis] on the consensus similarity matrix. Perron-cluster analysis involves the examination of eigenvalues of a *nearly uncoupled* or *nearly completely reducible* Markov chain. We consider the transition probability matrix ${\textbf{P}}$ of a random walker on the graph defined by the consensus matrix ${\textbf{M}}$: $${\textbf{P}}={\textbf{D}}^{-1}{\textbf{M}}$$ where ${\textbf{D}}$ is a diagonal matrix containing the row sums of ${\textbf{M}}$: ${\textbf{D}}=diag({\textbf{M}}{\textbf{e}}).$ According to our assumptions in [Section \[cmatrix\]]{}, there exists some simultaneous permutation of rows and columns of our consensus matrix such that the result is *block-diagonally dominant*. By this we essentially mean that ${\textbf{P}}$ (after row and column permutation) is a perturbation of a block-diagonal matrix ${\textbf}{B}$, such that $${\textbf{P}}={\textbf}{B}+{\textbf}{E} = \left[ \begin{array}{ccccc} {\textbf}{B}_{11} & {\textbf}{E}_{12} & {\textbf}{E}_{13}& \dots & {\textbf}{E}_{1k} \\ {\textbf}{E}_{21} & {\textbf}{B}_{22} & {\textbf}{E}_{23} & \dots & {\textbf}{E}_{2k} \\ {\textbf}{E}_{31} & {\textbf}{E}_{32} & {\textbf}{B}_{33} & \ddots & {\textbf}{E}_{3k} \\ \vdots& \vdots& \vdots & \ddots & \vdots \\ {\textbf}{E}_{k1} & {\textbf}{E}_{k2}& {\textbf}{E}_{k3} & \dots & {\textbf}{B}_{kk} \end{array} \right] \label{bdd}$$ where the off-diagonal blocks, ${\textbf}{E}_{ij}$, are much smaller in magnitude than the the diagonal blocks. In fact, the entries in the off-diagonal blocks are small enough that the diagonal blocks are *nearly stochastic*, i.e. ${\textbf}{B}_{ii} {\textbf{e}}\approx 1$ for $i=1,2,\dots,k$. A transition probability matrix taking this form describes a **nearly uncoupled** or **nearly completely reducible** Markov Chain. The degree to which a matrix is considered nearly uncoupled is dependent on one’s criteria for measuring the level of *coupling* (interconnection) between the *aggregates* (clusters of states) of the Markov chain [@fischer; @meyernumc; @chuckthesis]. In [@meyernumc], the *deviation from complete reducibility* is defined as follows: For an $m\times n$ irreducible stochastic matrix with a $k$-cluster partition $${\textbf{P}}= \left[ \begin{array}{ccccc} {\textbf}{P}_{11} & {\textbf}{P}_{12} & {\textbf}{P}_{13}& \dots & {\textbf}{P}_{1k} \\ {\textbf}{P}_{21} & {\textbf}{P}_{22} & {\textbf}{P}_{23} & \dots & {\textbf}{P}_{2k} \\ {\textbf}{P}_{31} & {\textbf}{P}_{32} & {\textbf}{P}_{33} & \ddots & {\textbf}{P}_{3k} \\ \vdots& \vdots& \vdots & \ddots & \vdots \\ {\textbf}{P}_{k1} & {\textbf}{P}_{k2}& {\textbf}{P}_{k3} & \dots & {\textbf}{P}_{kk} \end{array} \right]$$ the number $$\delta=2\max_{i} \|{\textbf{P}}_{i*}\|_{\infty},$$ where ${\textbf{P}}_{i*}$ represents a row of blocks, is called the **deviation from complete reducibility.** It is important to point out that the parameter $\delta$, or any other parameter that measures the level of coupling between clusters in a graph (like those suggested in [@fischer; @chuckthesis]) cannot be computed without prior knowledge of the clusters in the graph. Such parameters are merely tools for the perturbation analysis, used to present the following important fact regarding the spectrum of block-diagonally dominant stochastic matrices [@fischer; @kato; @chuck; @meyernumc; @perroncluster; @stewartnumc]:\ For sufficiently small $\delta \neq 0$, the eigenvalues of ${\textbf{P}}(\delta)$ are continuous in $\delta$, and can be divided into 3 parts: - The Perron root, $\lambda_1(\delta)=1$, - a cluster of $k-1$ eigenvalues $\lambda_2(\delta),\lambda_3(\delta),\dots,\lambda_k(\delta)$ that approach 1 as $\delta \to 0$ (known as the **Perron cluster**), and - the remaining eigenvalues, which are bounded away from 1 as $\delta \to 0$. In order to recover the number of blocks (clusters), we simply examine the eigenvalues of the stochastic matrix ${\textbf{P}}={\textbf{D}}^{-1}{\textbf{M}}$ and count the number of eigenvalues in the Perron cluster, which is separated from the remaining eigenvalues by the **Perron gap**, the largest difference between consecutive eigenvalues $\lambda_k$ and $\lambda_{k+1}$. The size of this gap is determined by the level of uncoupling in the graph, with larger gaps indicating more nearly uncoupled structures [@chuck]. Perron Cluster Analysis for Consensus Matrices {#perronconsensus} ----------------------------------------------- To build the consensus similarity matrix, we use one or more algorithms to cluster the data into a varying number of clusters. We set the algorithm(s) in our ensemble to find $\tilde{k}_1,\tilde{k}_2, \dots, \tilde{k}_J$ clusters in the data. The choice of the values $\tilde{k}_i$ is for the user, but we suggest choosing these values such that they might over-estimate the number of clusters but remain less than $\sqrt{n}$. We then construct the consensus similarity matrix ${\textbf{M}}$ from the resulting clusterings, examine the eigenvalues of the transition probability matrix ${\textbf{P}}={\textbf{D}}^{-1}{\textbf{M}}$, and count the number of eigenvalues near $\lambda_1 = 1$ by locating the largest gap in the eigenvalues. It is sometimes helpful in this phase of the algorithm to consider the use of an intolerance parameter $\tau$ in the construction of the consensus matrix because we may have allowed the algorithms to partition the data into fewer clusters than actually exist in the data. We return to the Medlars-Cranfield-CISI (MCC) collection discussed in [Section \[mcc\]]{} as an example where traditional SSE plots generally fail to provide a clear picture of how many clusters may be optimal [@phdthesis]. Before discussing the results of our method, we first look at the eigenvalues of the transition probability matrix that would result from using cosine as a measure of similarity (this is the most common similarity matrix used for text collections in the spectral clustering literature). The largest eigenvalues of this $3891 \times 3891$ matrix are displayed in [Figure \[MCCcosineEigs\]]{}. The plot shows only one eigenvalue in the Perron cluster and therefore, as with the other methods discussed in [@phdthesis], no information is gathered about the number of clusters in the data. ![Dataset MCC: 20 Largest Eigenvalues Found Using Cosine Similarity Matrix[]{data-label="MCCcosineEigs"}](MCCcosineEigs.pdf){width="0.5\linewidth"} Now we look at the eigenvalues of the transition probability matrix associated with a consensus similarity matrix. This consensus matrix was built from an ensemble of various algorithms paired with different dimension reductions and different levels of dimension reduction. All 3 of the authors’ preferred dimension reduction techniques (NMF, PCA, SVD) were used to reduce the dimensions of the data to $r=5,10, \mbox{ and } 20$ dimensions, creating a total of 10 data inputs (including the raw high-dimensional data) for each clustering algorithm. Three different clustering methods were used to cluster each data input: PDDP, spherical [*k*-means]{}initialized randomly, and spherical [*k*-means]{}initialized with centroids from the clusters found by PDDP. Counting every combination of dimension reduction and clustering procedure, the ensemble had 30 algorithms at work. For each of the 30 algorithms, $\tilde{k}=2,3,4,\dots,10$ clusters were determined and the resulting 270 clusterings were collected in the consensus matrix ${\textbf{M}}$. We show in [Figure \[mccicceigs\]]{} side-by-side images showing the eigenvalues of the transition probability matrix associated with the consensus similarity matrix *with* and *without* use of the intolerance parameter $\tau$. Particularly with text and other high-dimensional data, this intolerance parameter, by removing extraneous connections in the consensus graph, encourages a *nearly uncoupled* structure in the clustering results [@siamicc; @phdthesis]. This *uncoupling effect*, even for conservative values of $\tau$, is clearly identified by the widened gap after $\lambda_3$ in the eigenvalue graphs. [.47]{} ![Dataset MCC: 20 largest eigenvalues found using consensus similarity matrices with (right) and without (left) the intolerance parameter $\tau$. Ensemble of 30 algorithms, each clustering data into $\tilde{k}=2,3,\dots,10$ clusters[]{data-label="mccicceigs"}](MCCnodropEigs.pdf "fig:"){width="0.9\linewidth"} [.47]{} ![Dataset MCC: 20 largest eigenvalues found using consensus similarity matrices with (right) and without (left) the intolerance parameter $\tau$. Ensemble of 30 algorithms, each clustering data into $\tilde{k}=2,3,\dots,10$ clusters[]{data-label="mccicceigs"}](MCCdrop10Eigs.pdf "fig:"){width="0.9\linewidth"} However, both eigenvalue plots in [Figure \[mccicceigs\]]{}, with and without the intolerance parameter, reveal a Perron-cluster containing $k^*=3$ eigenvalues, as desired. ### Refining the Consensus Matrix through Iteration We have seen with the Medlars-Cranfield-CISI collection that the consensus matrix can provide better clustering information than the raw data. Therefore it seems reasonable that iterating the consensus process using multiple values of $\tilde{k}$ may refine the consensus matrix in way that minimizes or eliminates elements outside of the diagonal blocks, revealing a more identifiable Perron cluster. This is most often the case. Iterating the clustering process has an uncoupling effect on the consensus matrix [@siamicc]. In [Figure \[vismatrix0\]]{} we show a matrix heat-map of a consensus matrix formed by clustering 700 documents (100 of each in 7 clusters) into $\tilde{k}=[10, 11,\dots, 20]$ clusters with 4 different algorithms and 3 different dimension reductions. Red pixels indicate high levels of similarity while yellow pixels represent lower levels of similarity. There is a considerable amount of noise outside of the diagonal blocks. This consensus matrix was then clustered by the same 4 algorithms and 3 dimension reductions, again into $\tilde{k}=[10, 11,\dots, 20]$, and a heat map of the resulting consensus matrix (iteration 2) is shown in [Figure \[vismatrix2\]]{}. It is easy to see the refinement of the clusterings by the reduction of noise outside the diagonal blocks. The difference is also clearly shown in the eigenvalue plots displayed in [Figure \[viseigs\]]{}. For high-dimensional or noise-ridden data, we suggest this iterated procedure in determining the Perron gap, because spurious cluster relationships will often couple the clusters together. [.49]{} ![The Uncoupling Effect of Iteration: Matrix Heat Map[]{data-label="vismatrix"}](NGiter0winter.pdf "fig:") [.49]{} ![The Uncoupling Effect of Iteration: Matrix Heat Map[]{data-label="vismatrix"}](NGiter2winter.pdf "fig:") [.49]{} ![The Uncoupling Effect of Iteration: Eigenvalues[]{data-label="viseigs"}](NG7iter0.pdf "fig:") [.49]{} ![The Uncoupling Effect of Iteration: Eigenvalues[]{data-label="viseigs"}](NG7iter2.pdf "fig:") In [Section \[results\]]{} the flexibility of our approach is demonstrated using a comprehensive example on another benchmark dataset. The Iterative Consensus Clustering Framework is summarized in Algorithm \[iccframework\]. **Part I: Determining the Number of Clusters** *(If desired number of clusters is known, skip to Part II.)* - **Input:** Data Matrix ${\textbf{X}}$, intolerance parameter $\tau$ (if desired), and sequence $\tilde{k}=\tilde{k}_1,\tilde{k}_2,\dots,\tilde{k}_J$ - Using each clustering method $i=1,\dots,N$, partition the data into $\tilde{k}_j$ clusters, $j=1,\dots,J$ - Form a consensus matrix, ${\textbf}{M}$ with the $JN$ different clusterings determined in step 1. - Set ${\textbf}{M}_{ij}=0$ if ${\textbf}{M}_{ij} <\tau JN$. - Let ${\textbf}{D}=\mbox{diag} ({\textbf}{M}{\textbf{e}})$. Compute the eigenvalues of ${\textbf{P}}$ using the symmetric matrix ${\textbf}{I}-{\textbf}{D}^{-1/2}{\textbf}{M}{\textbf}{D}^{-1/2}$. - Output the number of eigenvalues in the Perron cluster, $k$. Repeat steps 1-5 using ${\textbf}{M}$ as the data input in place of ${\textbf{X}}$, if the number of eigenvalues in the Perron cluster remains the same, stop. **Part II: Determining the Consensus Solution** - **Input:** Final consensus matrix from part I, intolerance parameter $\tau$ (if desired), and the number of clusters $k$. (Or if desired number of clusters is known before hand, the raw data matrix ${\textbf{X}}$). - Using each clustering method $i=1,\dots,N$, partition the matrix into $\tilde{k}_j$ clusters, $j=1,\dots,J$ - If the majority of algorithms agree upon a single solution, stop and output this solution. - Form a consensus matrix, ${\textbf}{M}$ with the $JN$ different clusterings determined in step 1. - Set ${\textbf}{M}_{ij}=0$ if ${\textbf}{M}_{ij} <\tau JN$. - Repeat steps 1-5. Comprehensive Results on a Benchmark Dataset {#results} ============================================ NG6: Subset of 20 Newsgroups ---------------------------- The Newsgroups 6 (NG6) dataset is a subset from the infamous “Twenty Newsgroups” text corpus that has become a common benchmark for cluster analysis. The Twenty Newsgroups corpus consists of approximately 20,000 news documents (web articles) partitioned somewhat evenly across 20 different topics. The collection of these documents is attributed to Ken Lang, although it is never mentioned explicitly in his original paper [@Lang]. It is now publicly available via the web [@20newsgroups]. To create the NG6 collection, 300 documents from 6 topics were randomly selected, resulting in a term-document matrix with 11,324 terms and 1800 documents. Our initial consensus matrix (used to determine the number of clusters) was formed using *only* the [*k*-means]{}algorithm (randomly initialized), performed on the raw data and 3 dimension reductions. To determine an appropriate rank for dimension reduction we followed convention by observing the screeplot (plot of the singular values) for the NG6 data matrix. The screeplot shown in [Figure \[ng6screeplot\]]{}, indicates our decision to reduce the dimensions of the data from $m=11,324$ to $r=10$. ![Dataset NG6: Screeplot (First 100 Singular Values by Index)[]{data-label="ng6screeplot"}](ng6screeplot.pdf){width="0.5\linewidth"} The dimensionality of the data was then reduced using our 3 preferred dimension reduction algorithms: 1. Principal Components Analysis 2. Singular Value Decomposition 3. Nonnegative Matrix Factorization and 10 iterations of [*k*-means]{}clustering was performed on each dimension reduction to create $\tilde{k} = 10,11,12,\dots,20$ clusters. Since the clustering is only performed on the reduced data, this phase of the process proceeds extremely fast. The result was a total of 330 clusterings which contributed to the initial consensus matrix. An intolerance parameter was not used for this initial matrix. The next step in our analysis is to examine the eigenvalues of the transition probability matrix associated with this initial consensus matrix. ![Dataset NG6: Eigenvalues associated with Initial (unadjusted) Consensus Matrix[]{data-label="ng6iter0Eigs"}](NG6iter0Eigs.pdf){width="0.5\linewidth"} The Perron cluster in [Figure \[ng6iter0Eigs\]]{} contains 5 eigenvalues for initial consensus matrix. As discussed in [Section \[perronconsensus\]]{} there are two adjustments one might consider to further explore the cluster structure. 1. Implement an intolerance parameter $\tau$ to distinguish the Perron cluster 2. Iterate the consensus procedure using the initial consensus matrix as input (no dimension reduction was used here). In [Figure \[ng6adjust\]]{} the results of both adjustments are shown. In either scenario, a Perron cluster with $k^*=6$ eigenvalues becomes clear. When the iterative procedure is repeated once more, the number of eigenvalues in the Perron cluster does not change. [.47]{} ![Dataset NG6: Eigenvalues associated with consensus similarity matrices adjusted by intolerance (right) or adjusted by iteration (left) using $\tilde{k}=10,11,12,\dots,20$ clusters. The $k^*=6$ eigenvalues in the Perron cluster correctly identify the number of clusters.[]{data-label="ng6adjust"}](NG6Eigsp30.pdf "fig:"){width="0.9\linewidth"} [.47]{} ![Dataset NG6: Eigenvalues associated with consensus similarity matrices adjusted by intolerance (right) or adjusted by iteration (left) using $\tilde{k}=10,11,12,\dots,20$ clusters. The $k^*=6$ eigenvalues in the Perron cluster correctly identify the number of clusters.[]{data-label="ng6adjust"}](NG6iter1Eigs45iterationsEachk10to20.pdf "fig:"){width="0.9\linewidth"} For the purposes of comparison, we present in [Figure \[ng6cosineeigs\]]{} the eigenvalues of the transition probability matrix associated with the Cosine similarity matrix, which is commonly used to cluster document datasets with spectral algorithms. No information regarding the number of clusters is revealed by the Perron cluster, which contains only a single eigenvalue. ![Dataset NG6: Eigenvalues associated with Cosine Similarity Matrix[]{data-label="ng6cosineeigs"}](ng6CosEigs.pdf){width="0.5\linewidth"} ### NG6 Data: Determining a Cluster Solution #### Comparing the Consensus Matrix as Input Tables \[ng6rawaccs\], \[ng6consaccs\] and \[ng6rawaccs2\] demonstrate the superiority of the consensus matrices to traditional data inputs. The accuracies of certain algorithms increase by as much as 60% when the consensus matrix is used as input compared with the raw data. The average accuracies of all the algorithms also increases dramatically. One interesting fact to point out is that in this example, the algorithms perform quite poorly on the NMF dimension reduction. Even though these results are contained in the consensus matrices, the nature of the process is able to weed out these poor results. The authors have experimentally discovered time and again that the ICC process is not sensitive to a small number of poor results contained in the cluster ensemble. When it comes to clustering the ensemble, such results are essentially “voted out.” Algorithm Raw Data PCA $r=10$ SVD $r=10$ NMF $r=10$ -------------------- ---------- ------------ ------------ ------------ -- -- -- PDDP 0.32 0.75 0.80 0.31 PDDP-[*k*-means]{} 0.77 0.99 0.99 0.60 NMF-Basic 0.46 - - - [*k*-means]{} 0.18 0.81 0.68 0.18 Average 0.43 0.85 0.82 0.36 : Dataset NG6: Accuracies for individual algorithms on raw data and dimension reductions.[]{data-label="ng6rawaccs"} Algorithm Initial Consensus Consensus A Consensus B -------------------- ------------------- ------------- ------------- -- -- -- -- PDDP 0.85 0.93 0.98 PDDP-[*k*-means]{} 0.96 0.94 0.98 NMF-Basic 0.98 0.98 0.99 [*k*-means]{} 0.75 0.72 0.78 Average 0.88 0.89 0.93 : Dataset NG6: Accuracies for individual algorithms on 3 different consensus matrices.[]{data-label="ng6consaccs"} Algorithm Cosine Initial Consensus Consensus A Consensus B ----------- -------- ------------------- ------------- ------------- PIC 0.52 0.86 0.97 0.98 NCUT 0.98 0.98 0.98 0.99 NJW 0.87 0.82 0.80 0.99 Average 0.79 0.89 0.92 0.99 : Dataset NG6: Accuracies for spectral algorithms on different similarity matrices.[]{data-label="ng6rawaccs2"} ### NG6 Data: Determining a Final Solution Our second step in the ICC process, once the number of clusters has been determined via the eigenvalue analysis, is to iterate the consensus procedure using the determined number of clusters in an attempt to witness agreement between algorithms. Combining the clustering results of each algorithm on Consensus B (Column 3 in [Table \[ng6consaccs\]]{} and Column 4 in [Table \[ng6rawaccs2\]]{}) into another consensus matrix, we run through a second round of “voting”. The matrix Consensus B was chosen because the eigenvalue gap was larger, although using Consensus matrix A provides a similar result. The accuracies of the resulting solutions are given in [Table \[ng6consensus\]]{}. The boxed in values indicate a common solution among algorithms. We call this the *final consensus solution.* Algorithm Consensus Iter 2 -------------------- ------------------ PDDP 0.83 PDDP-[*k*-means]{} NMF-Basic [*k*-means]{} PIC NCUT 0.99 NJW 0.82 : Dataset NG6: 4 of 7 algorithms find a common solution in one iteration of the final consensus process[]{data-label="ng6consensus"} ### NG6 Data: Conclusion We began our analysis with a collection of documents and algorithms - both for dimension reduction and for clustering. Traditional tools for determining the number of clusters were not successful. An analyst, having somehow determined the number of clusters and attempting to cluster this data with the given set of tools had a chance of finding a solution with accuracy ranging from 18% to 99%. If that analyst had chosen the best dimension reduction algorithm (PCA) for this particular dataset (a task for which there are no hard and fast guidelines), the accuracy of his/her solution may have been between 77% and 99%. Internal cluster validation metrics like the [*k*-means]{}objective function would not have been much help in choosing between these solutions, as such measures are difficult to compare on high-dimensional data. However, by using *all* of the tools at our disposal in the Iterative Consensus Clustering Framework, we found that the clustering algorithms worked out their differences constructively - finally settling down on a solution with the highest level of accuracy achieved by any of the algorithms independently. Conclusion ========== Herein we have presented a flexible framework for combining results from multiple clustering algorithms and/or multiple data inputs. Not only does this framework provide the user with an above average clustering solution, it also contains a practical exploratory procedure for determining the number of clusters. We have discovered that consensus matrices built using multiple algorithms and multiple values for the number, $k$, of clusters will often allow users to estimate an appropriate number of clusters in data by determining the maximum number of clusters for which algorithms are likely to agree on a common solution. We have provided several examples to show how this approach succeeds at determining the number of clusters in datasets where other methods fail. When the initial consensus matrix does not provide this information, it can be refined through the use of an intolerance parameter or iteration to get a clearer picture of how many clusters the algorithms might be able to agree upon. While the consensus matrix itself is not a new idea, the practice of using multiple algorithms and dimension reductions together to create the matrix had not previously been explored, nor had varying the number of clusters for the purposes of approximating $k$. Our approach to building the consensus matrix is novel and improves clustering results from nearly every clustering algorithm on all datasets considered. This consensus matrix has several advantages over traditional similarity matrices as discussed in [Section \[benefits\]]{}. The ICC Framework encourages clustering algorithms to agree on a common solution to help escape the unreliability of individual algorithms. While previous consensus methods have aimed to average cluster solutions in one way or another, ours is the first to emphasize agreement between clustering algorithms. After seeing some of the results of the individual algorithms in our ensemble, it should be clear that an average solution could be very poor indeed. Rather than deciding each clustering is equally valid, we simply sum the number of times a cluster relationship was made between two points and let the algorithms decide whether this sum is considerable enough to draw those points together, or whether it might be more reasonable to dissolve the connection in favor of others. This framework iteratively encourages algorithms to agree upon a common solution because the value of the similarity metric reflects the level of algorithmic agreement at each step. Thus, through iteration, cluster relationships upon which the algorithms do not agree are abandoned in favor of relationships with higher levels of agreement. [10]{} P. Berman, B. DasGupta, M. Kao, and J. Wang. On constructing an optimal consensus clustering from multiple clusterings. , 104:137–145, 2007. M.W. Berry, editor. . Springer, 2004. D. Boley. Principal direction divisive partitioning. , 2(4):325–344, December 1998. W. Cao and W.J. Stewart. Iterative aggregation/disaggregation techniques for nearly uncoupled markov chains. , 32(2):702–719, 1985. Twenty Newsgroups Dataset. http://qwone.com/\~jason/20newsgroups/. August 5th, 2013. S. Deerwester, S.T. Dumais, G.W. Furnas, T.K. Landauer, and R. Harshman. Indexing by latent semantic analysis. , 41(6):391–407, September 1990. P. Deuflhard and M. Webber. Robust peron cluster analysis in conformation dynamics. , 398, 2004. P. Drineas, A. Frieze, R. Kanna, S. Vempala, and V. Vinay. Clustering large graphs via the singular value decomposition. , 56(1):9–33, 2004. W. W. Cohen F. Lin. Power iteration clustering. , 2010. X. Z. Fern and C. E. Brodley. Cluster ensembles for high dimensional clustering: An empirical study. Technical report, 2006. V. Filkov and S. Skiena. Integrating microarray data by consensus clustering. , 2004. A. Fred and A.K. Jain. Combining multiple clusterings using evidence accumulation. , 27(6):835–850, 2005. S. Régnier. Sur quelques aspects mathematiques des problems de classification automatique. , 4:175–191, 1965. D. Gusfield. Partition-distance: A problem and class of perfect graphs arising in clustering. , 82:159–164, 2002. A. K. Jain. Data clustering: 50 years beyond k-means. , 2009. I.T. Jolliffe. . Springer Series in Statistices. Springer, 2nd edition, 2002. T. Kato. . Springer, 1995. J. Kogan. . Cambridge University Press, Cambridge, New York, 2007. S. Kotsiantis and P. Panayiotis. Recent advances in clustering: A brief survey. , 1(1):73–81, 2004. K. Lang. Newsweeder: Learning to filter netnews. In [*Proceedings of the 12th International Conference on Machine Learning*]{}, pages 331–339, 1995. A. Langville, M.W. Berry, M. Browne, V. P. Pauca, and R. J. Plemmons. Algorithms and applications for the approximate nonnegative matrix factorization. , 2007. F. de Toledo, M. Nascimento and A. Carvalho. Consensus clustering using spectral theory. , 461-468, 2009 C. D. [Meyer]{} and C. D. [Wessell]{}. . , 33(4):1214–1236, 2012. C. D. Meyer. Stochastic complementation, uncoupling markov chains, and the theory of nearly reducible systems. , 31(2), 1989. B. Mirkin. The problems of approximation in spaces of relations and qualitative data analysis. , 35:1424–1431, 1974. S. Monti, P. Tamayo, J. Mesirov, and T. Golub. Consensus clustering: A resampling-based method for class discovery and visualization of gene expression microarray data. , 52:91–118, 2003. L. Murino, C. Angelini, I. De Feis, G. Raiconi, and R. Tagliaferri. Beyond classical consensus clustering: The least squares approach. , 2011. A. Ng, M. Jordan, and Y. Weiss. On spectral clustering: Analysis and an algorithm, 2001. N. Nguyen and R. Caruana. Consensus clusterings. In [*Seventh IEEE International Conference on Data Mining*]{}, 2007. A. Fischer, Ch. Schutte, P. Deuflhard, and W. Huisinga. Identification of almost invariant aggregates in reversible nearly uncoupled markov chains. , 315:39–59, 2000. B. G. Mirkin. . Chapman & Hall, CRC Computational Science Series, 2013. V. Kumar, P. Tan, and M. Steinbach. . Pearson, 2006. Y. Qian and C. Suen. Clustering combination method. In [*International Conference on Pattern Recognition (ICPR)*]{}, volume 2, pages 732–735, 2000. S. Race. Clustering via dimension-reduction and algorithm aggregation. Master’s thesis, North Carolina State University, 2008. S. Race. . PhD thesis, North Carolina State University, 2014. S. Race, C.D. Meyer, and K. Valakuzhy. Determining the number of clusters via iterative consensus clustering. In [*Proceedings of SIAM 13th International Conference on Data Mining*]{}, 2013. M. Sales-Pardo, R. Guimera, A.A. Moreira, and L.A.N. Amaral. Extracting the hierarchical organization of complex systems. , 104(39), 2007. F. Shahnaz, M. W. Berry, V. P. Pauca, and R. J. Plemmons. Document clustering using nonnegative matrix factorization. , 42(2):373–386, 2006. J Shi and J Malik. Normalized cuts and image segmentation. , 22(8):888–905, 2000. A. Strehl and J. Ghosh. Cluster ensembles–a knowledge reuse framework for combining multiple partitions. , 3(583-617), 2003. A. Topchy, A. K. Jain, and W. Punch. Clustering ensembles: Models of consensus and weak partitions. , 27(12):1866–1881, 2005. U. [von Luxburg]{}. A tutorial on spectral clustering. , 17:4, 2007. C. Wessell. . PhD thesis, North Carolina State University, 2011. Y. Zeng, J. Tang, J. Garcia-Frias, and G. R. Gao. An adaptive meta-clustering approach: Combining the information from different clustering algorithms. In [*Proceedings of the IEEE Computer Society Bioinformatics Conference*]{}, 2002.

ArXiv

--- abstract: 'In this work we extend the work on the recently discovered role of Cosmic Rays (CRs) in regulating the average CO/$\rm H_2$ abundance ratio in molecular clouds (and thus their CO line visibility) in starburst galaxies, and find that it can lead to a CO-poor/C[ i]{}-rich $\rm H_2 $ gas phase even in environments with Galactic or in only modestly enhanced CR backgrounds expected in ordinary star-forming galaxies. Furthermore, the same CR-driven astro-chemistry raises the possibility of a widespread phase transition of molecular gas towards a CO-poor/C[i]{}-rich phase in: a) molecular gas outflows found in star-forming galaxies, b) active galactic nuclei (AGNs), and c) near synchrotron-emitting radio jets and the radio-loud cores of powerful radio galaxies. For main sequence galaxies we find that CRs can render some of their molecular gas mass CO-invisible, compounding the effects of low metallicities. Imaging the two fine structure lines of atomic carbon with resolution high enough to search beyond the C[i]{}/CO-bright line regions associated with central starbursts can reveal such a CO-poor/C[i]{}-rich molecular gas phase, provided that relative brightness sensitivity levels of $T_b$(C[i]{} $1-0$)/$T_b$(CO $J=1-0$)$\sim $0.15 are reached. The capability to search for such gas in the Galaxy is now at hand with the new high-frequency survey telescope HEAT deployed in Antarctica and future ones to be deployed in Dome A. ALMA can search for such gas in star-forming spiral disks, galactic molecular gas outflows and the CR-intense galactic and circumgalactic gas-rich environments of radio-loud objects.' author: - | Padelis P. Papadopoulos,$^{1,2,3}$[^1] Thomas G. Bisbas,$^{4,5}$ and Zhiyu Zhang$^{6,7}$\ $^{1}$Department of Physics, Section of Astrophysics, Astronomy and Mechanics,Aristotle University of Thessaloniki,\ Thessaloniki, GR-54124, Greece\ $^{2}$Research Center for Astronomy, Academy of Athens Soranou Efesiou 4, GR-11527, Athens, Greece\ $^{3}$School of Physics and Astronomy, Cardiff University, Queen’s Buildings, The Parade, Cardiff, CF24 3AA, UK\ $^{4}$Department of Astronomy, University of Virginia, Charlottesville, VA 22904, USA\ $^{5}$Max-Planck-Institut für Extraterrestrische Physik, Giessenbachstrasse 1, D-85748 Garching, Germany\ $^{6}$Institute for Astronomy, University of Edinburgh, Royal Observatory, Edinburgh, EH9 3HJ, UK\ $^{7}$European Southern Observatory, Headquarters, Karl-Schwarzschild-Strasse 2, D-85748, Garching bei München, Germany date: 'Accepted XXX. Received YYY; in original form ZZZ' title: 'New places and phases of CO-poor/C[i]{}-rich molecular gas in the Universe' --- \[firstpage\] [(ISM:) cosmic rays, (ISM:) photodissociation regions (PDR), methods: numerical, astrochemistry, radiative transfer, galaxies: ISM]{} Introduction {#sec:intro} ============ The utility of CO and its low-$J$ rotational transitions as effective tracers of $\rm H_2$ gas mass is now well established observationally [e.g. @Youn91; @Solo92; @Solo97] and theoretically [e.g. @Dick86; @Malo88; @Brya96], provided: a) the so-called $X_{\rm CO}(=M({\rm H}_2)/L_{\rm CO})$ factor is used only for $M({\rm H}_2)\geq 10^{5}\,M_{\odot}$ (so that its statistical notion remains valid) and b) the $\rm [CO]/[H_2]$ abundance ratio does not fall much below its average Galactic value of $\sim 10^{-4}$. The latter can happen in low-metallicity interstellar medium (ISM) with strong ambient far-ultraviolet (FUV) radiation fields [@Bola99; @Pak98], such as those expected in metal-poor dwarf galaxies, leaving large amounts of $\rm H_2$ gas as CO-invisible [@Madd97; @Shi17]. In the Milky Way and other ordinary spirals, this is expected also for molecular gas at large galactocentric distances where metallicity falls to $\sim 0.2 \rm Z_{\odot}$ and much of the $\rm H_2$ gas can be rendered (CO line)-invisible by the FUV-induced destruction of CO [@Papa02; @Wolf10]. It was only recently that Cosmic Rays (CRs) have been identified as a potentially more effective CO-destruction agent in molecular clouds compared to FUV photons [@Bisb15; @Bisb17; @Bial15]. Unlike the latter whose propagation (and thus CO-dissociation capability) is blunted by the strong dust absorption of FUV light taking place in the dust-rich H[i]{} phase and in outer $\rm H_2$ cloud envelopes, CR-induced chemistry destroys CO volumetrically throughout a molecular cloud irrespective of dust column. Other observational signatures of CR-controlled versus FUV-controlled chemistry of H$_2$ clouds in galaxies, (even when CO remains abundant) have been discussed thoroughly in the literature [e.g. @Papa10; @Meij11], and will not concern us here. We should, nevertheless, mention that it is CRs that make the C[i]{} distribution concomitant with that of CO in H$_2$ gas clouds, rather existing only in a thin transition layer between C[ii]{}-rich outer and CO-rich inner H$_2$ cloud regions [as the traditional PDR view would have it, e.g. @Holl99]. This makes C[i]{} lines equally good and more straightforward H$_2$ gas mass tracers as low-J CO lines, even under conditions where CO remains abundant in H$_2$ gas clouds. A last theoretical effort to retain the PDR picture against the failure of its basic prediction of a C[ii]{}/C[i]{}/CO stratification of species on the surface of FUV-illuminated H$_2$ clouds was made by introducing density inhomogeneities on the classic PDR picture [@Meix93; @Spaa97]. There, the C[ii]{}/C[i]{}/CO species stratification remained but only on small H$_2$ clumps while C[i]{} appeared spatially extended deeper into FUV-illuminated CO-rich inhomogeneous H$_2$ clouds. However, while CO-rich H$_2$ clumps of low filling factor, each with a C[i]{} “coating", could reproduce the astonishing spatial correspondence between C[i]{} and $^{12}$CO, $^{13}$CO line emission observed in GMCs across a wide range of conditions, they could not account for the tight [*intensity correlation*]{} between $^{12}$CO, $^{13}$CO (1-0), (2-1) and C[i]{} 1-0 line intensities, unless one postulates also a very standard H$_2$ clump making up all H$_2$ clouds, with characteristics that remain invariant across the wide range of ISM conditions [see @Papa04 for details]. CRs are the simplest and most likely culprits in creating a volumetric rather than surface-like C[i]{} distribution in H$_2$ clouds, and therein lies the most notable difference between FUV-driven chemistry and gas thermal state, and a CR-driven one. Another key difference between the two mechanisms besides their spatially distinct ways to destroy CO is that FUV-induced CO destruction leaves behind C[i]{} and C[ii]{}, while CR-induced destruction yields mostly C[i]{}, provided that $T_{\rm kin}\lesssim50\,{\rm K}$, otherwise CO abundance increases via the OH channel [@Bisb17]. This makes CR-induced CO-poor $\rm H_2$ gas more accessible to observations via the two fine structures of atomic carbon at rest frequencies $\nu_{\rm rest}^{\rm CI}(1-0) \sim $492GHz $^3{\rm P}_1-^3{\rm P}_0$ (hereafter 1-0) and $\nu_{\rm rest}^{\rm CI}(2-1) \sim $809GHz $^3{\rm P}_2-^3{\rm P}_1$ (hereafter 2-1) than FUV-irradiated clouds. These \[C[i]{}\] lines can be observed over a large redshift range, starting from $z\sim $0 to $z\sim $5 for \[C[i]{}\] 1-0, and from $z\sim $0 to $z\sim $8 for \[C[i]{}\] 2-1, using ground-based telescopes, such as ALMA on the Atacama Desert Plateau, while the \[C[ii]{}\] fine structure line, even if typically much brighter than the \[C[i]{}\] lines for warm gas, will be faint for cold $\rm H_2$ gas ($T_k\sim (15-20)\,{\rm K}$, while $\rm E_{ul}($C[ii]{}$\rm )/k_B\sim $92K) away from star-forming (SF) sites. Moreover C[ii]{} has a rest frequency of $\sim $1900GHz, making it accessible to ground-based observations only once $\rm z\ga 2$, still extremely challenging until $\rm z\ga 4$. This leaves most of the star-formation history of the Universe (and its gas-fueling) outside the reach of \[C[ii]{}\]. In this paper we study the effects of a CR-regulated \[CO/C[i]{}\] average abundance in low-density molecular gas in the Galaxy, the outer regions of local spirals, and distant main sequence (MS) galaxies. We conclude this work by examining the possibility of CO-poor/C[i]{}-rich molecular gas in the CR-intense environments of molecular gas outflows from starbursts, and the environments of radio-loud objects. Low-density molecular gas in the Universe: the effects of CRs {#sec:2} ============================================================= The CR effects on the relatively low density molecular gas ($n({\rm H_2}) \sim$50-500${\rm cm}^{-3}$) have not been studied in detail, but early hints that CO can be effectively destroyed in such gas [*even at Galactic levels of CR energy densities*]{} exist [@Bisb15]. A low density molecular gas phase can be found in a variety of places in the Universe, the nearest ones being the envelopes of ordinary GMCs in the Galaxy. Should their CO-marker molecule be wiped out by CRs, it would leave the corresponding H$_2$ gas mass CO-invisible, yielding a systematic underestimate of H$_2$ gas mass even in places where CO was considered an effective $\rm H_2$ tracer. This is of particular importance since a typical log-normal distribution of $M$(H$_2$)-$n$($\rm H_2$) expected in turbulent GMCs would place most of their total mass at densities of $n({\rm H}_2)<500\,{\rm cm}^{-3}$ [@Pado02]. Moreover, CRs can act on the chemistry of H$_2$ gas as fast as the photon-driven processes driven by FUV radiation fields. The Milky Way {#ssec:MW} ------------- Studies of the H[i]{}$\rightarrow $H$_2$ phase transition in the metallicity and radiation environment of the Milky Way showed that it can commence from densities as low as $n({\rm H_2}) \sim$ 5–20${\rm cm}^{-3}$, depending on the $\rm H_2$ formation rate on grains and ambient dust shielding [@Papa02], while for density enhancements reaching above $n({\rm H_2}) \sim$50${\rm cm}^{-3}$ this transition is complete [@Jura75a; @vDis86; @Jura75b; @Shay87; @Ande93; @Shul00; @Offn13; @Bial17]. However CO (and HCN) multi-$J$ observations of Molecular Clouds (MCs) in the Galaxy, typically yield densities of $n({\rm H_2}) \sim$ 500–$10^4$${\rm cm}^{-3}$ [@Saka97; @Heye15; @Bial16]. Thus, there is a significant range of gas densities $n({\rm H_2}) \sim$ 50–500${\rm cm}^{-3}$ where Cold Neutral Medium (CNM) gas can be molecular but perhaps not (CO-line)-bright, reminiscent of the translucent clouds [@vDis86]. {width="\textwidth"} Figure \[fig:cds\] shows the column density maps of H[i]{}, $\rm H_2$ CO, C[i]{}, C[ii]{}, and $T_{\rm kin}$ distributions within inhomogeneous low-density gas clouds using the fractal rendering and thermo-chemical calculations presented by @Bisb17 and using the [3d-pdr]{} code[^2] [@Bisb12], but now subjected only to Galactic levels of interstellar radiation field and CR energy density [i.e. $\zeta'=1$, where $\zeta'=\zeta_{\rm CR}/10^{-17}\,{\rm s}^{-1}$ and $\rm G_{\circ}=1$ normalized according to @Drai78]. The GMCs were constructed using the method described in @Walc15 for a fractal dimension of ${\cal D}=2.4$ and assuming a mass of $M=7\times10^4\,{\rm M}_{\odot}$ but with different radial extent corresponding to three average number densities i.e. $\langle n\rangle\sim50,\,100,\,200\,{\rm cm}^{-3}$. We find that the maximum visual extinction, $A_V$, along the line-of-sight of these clouds is $\sim14,\,22,\,40\,{\rm mag}$, respectively. These GMCs have much smaller average number densities than the GMC studied in @Bisb17 ($\langle n\rangle\sim760\,{\rm cm}^{-3}$). Furthermore, our computations are made for metallicities of both $Z=Z_{\odot}$ and $Z=0.2Z_{\odot }$ representing the ISM for inner and the outer parts of the Milky Way. From the maps in Figure \[fig:cds\] it can be readily seen that for $\langle n\rangle\sim(50-100)\,{\rm cm}^{-3}$ [*the $\rm H_2$ gas can be rendered very CO-poor, even at Galactic levels of CR energy density*]{}, while for low metallicities this remains so up to $\langle n\rangle\sim200\,{\rm cm}^{-3}$ (the highest average density in our computations). Atomic Carbon on the other hand remains abundant throughout the H$_2$-rich parts of the cloud, except for the low-metallicity gas where its abundance drops, but nevertheless remains generally higher than that of CO. The cloud mass fractions for which the \[CO\]/\[H$_2$\] abundance drops below $10^{-5}$ (i.e. 10 times below the average \[CO\]/\[H$_2$\] abundance in the Galaxy, making hard to use CO lines as $\rm H_2$ gas mass tracers) are: 80%, 55%, 30% for $\langle n\rangle\sim50,\,100,\,200\,{\rm cm}^{-3}$, respectively, for $Z=1Z_{\odot}$, K$_{\rm vir}=1$[^3] and $\zeta'=1$. For the metal-poor case, the CO-poor cloud mass fraction becomes $\gtrsim 98\%$. Nevertheless, unlike in our past higher-density cloud models where most of C[ii]{} recombines into C[i]{} [@Bisb17], C[ii]{} now remains abundant for much of the mass of our low-density clouds (Figure \[fig:cds\]). Furthermore, classical photolectric FUV heating, and the much lower average cooling of low density gas ($\Lambda _{\rm line}\propto n^2$) allows the gas to maintain higher temperatures $T_{\rm kin}\sim(40-100)\,{\rm K}$ for $\langle n\rangle\sim(50-100)\,{\rm cm}^{-3}$, where the \[C[ii]{}\] fine structure line is expected to be luminous. Figure \[fig:cdsall\] shows how the carbon cycle abundances change as a function of the total H column density for the four different ISM models. Here, we plot the average value of all three different clouds simultaneously. It can be seen that for the Galactic conditions (panel a), the molecular gas is CO-dominated for high column densities, as expected, while C remains abundant enough ($\rm [CI/CO]\sim 0.1-0.3$) as to continue serving as a capable $\rm H_2$ gas tracer along with CO [e.g. @Papa04]. However, the molecular gas phase switches to a C[i]{}-dominated one in the lower-metallicity case (panel b) and for $Z=1\,Z_{\odot}$ and $\zeta'=30$ case (panel c). For even higher $\zeta'$, panel (d) shows a gas phase that becomes C[ii]{}-dominated even at higher column densities (i.e. $\sim6\times10^{21}\,{\rm cm}^{-2}$) more typical for inner regions of molecular clouds. {width="80.00000%"} Thus it may well be that, besides FUV photons, CRs also contribute in the making of (C[ii]{}-line)-bright gas envelopes of (CO line)-invisible gas found around (CO line)-marked GMCs in the Galaxy [e.g., @Pine13; @Lang14]. [*Then these envelopes can be bright in both C[ii]{} and C[i]{} lines,*]{} a possibility that should be investigated by sensitively imaging the latter in the C[ ii]{}-bright envelope regions of GMCs in the metal-rich inner parts of the Galaxy. The conditions for a widespread phase transition of $\rm H_2$ gas from a CO-rich to a CO-poor/CI-rich phase can exist also for outer Galactic regions. Indeed as the star formation rate density becomes lower in the outer Galaxy, along with the intervening interstellar absorption, they reduce the average FUV radiation field at large Galactocentric distances. CRs however can stream further out in the disk, and keep C[i]{} abundant in low-density molecular clouds. There is already evidence for CO-poor gas in the Galaxy at large galactocentric radii from studies of otherwise CO-bright clouds selected in the Goddard-Columbia $^{12}$CO survey. In these places an $\rm X_{CO}$ factor systematically larger by a factor of $\sim 2-3$ with respect to its standard value at the inner Galaxy is found [@Sodr91]. Other work using also CO-marked clouds finds that the diffuse H$_2$ clouds contain increasingly more $\rm H_2 $ mass at larger Galactocentric radii [@Roma16]. These are the type of clouds that could contain a significant CI-rich/CO-poor phase according to our current study. Perhaps more significantly, the only search for the so so-called CO-dark gas in our Galaxy which is not using CO-selected $\rm H_2$ clouds but dust extinction maps instead indicates that up to 55% of the gas at large Galactocentric radii may be CO-dark [@Chen15]. The same study finds that in certain regions the inferred CO-dark $\rm H_2$ gas mass can reach up to four times the CO-luminous one. We must note here that all our current arguments about CO-poor/CI-rich gas, as well as those that follow in the next sections, are phase-transition type of arguments. By this we mean that we are simply investigating the ISM conditions controlling the phase transition from a typically CO-rich to a CO-poor (and CI-rich) gas phase (with emphasis on the role of CRs, gas density, and metallicity), and the places where the conditions for such a phase transition can be fullfiled. In order to find how much $\rm H_2$ mass is actually in such a gas phase, observations are indispensible. To this purpose we conclude this section by urging sensitive imaging of the two C[i]{} lines for the molecular gas in the Galaxy, and especially of its CO-poor yet CII-luminous phase. ### The promise of ground-based high-frequency single dish telescope surveys The High Elevation Antarctic Terahertz Telescope (HEAT), with a diameter of 60 cm, is a new high-frequency telescope now operating at Ridge A of the Antarctic plateau, with capabilities to observe both \[C[i]{}\] lines[^4]. Its wide beam of $\sim 4.1^{'}$ at 492GHz (\[C[i]{}\] 1-0) and $2.5^{'}$ at 809GHz (\[C[i]{}\] 2-1) makes it appropriate for searching for low-brightness temperature (T$_{\rm b}$) C[i]{}-rich molecular clouds in the Galaxy, where $T_b$(\[C[i]{}\] 1-0)/$T_b$(CO 1-0) (hereafter $T_b$(C[ i]{})/$T_b$(CO)) drops to $\sim 0.15$ in SF-quiescent regions [@Papa04 and references therein]. An inventory of molecular gas in the Galaxy obtained using both \[C[i]{}\] lines and its comparison to the CO-rich and C[ii]{}-rich gas can reveal whether a low-density C[ i]{}-rich/CO-poor gas phase exists in the Milky Way, its spatial distribution, and temperature. Other single dish submm telescopes from excellent sites at the Atacama Plateau in North Chile (APEX, NANTEN-2, ASTE) can be used to sensitively map C[i]{} (1-0) and/or (2-1) in a more targeted fashion e.g. along a continuous strip starting from the inner parts of a given Galactic molecular cloud and continuing well beyond its (CO/CI-line)-luminous regions, searching for C[i]{}-rich/CO-poor envelopes. The upcoming 5-m Dome A Terahertz Explorer [DATE5 @Yang13] in Dome-A of the Antarctic plateau will be able to perform a systematic survey across the Galactic plane with both good angular resolution and sensitivity, given its excellent condition for high-frequency observations[@Shi16]. In Figure \[fig:heat\], we show $T_b$(C[i]{})/$T_b$(CO) ratio maps for our low-density ($\langle n\rangle$$\sim$50${\rm cm}^{-3}$) inhomogeneous cloud model. These could be used to guide such an observational campaign in the Galaxy indicating the necessary CI/CO line relative brightness sensitivity levels. {width="80.00000%"} Molecular gas outflows from galaxies {#ssec:outflows} ------------------------------------ The discovery of strong H$_2$ outflows from galaxies, induced by AGN and/or starburst activity [@Feru10; @Cico12; @Cico14; @Dasy12], shows that large amounts of molecular gas can be expelled from galaxies. It followed much earlier discoveries of huge cm-emitting synchrotron haloes around starburst galaxies [e.g. @Seaq91; @Colb96], indicating CRs outflows swept out from the star-formation (SF) galaxy at bulk speeds $\ga 1000$kms$^{-1}$. This could have significant effects since powerful molecular outflows are often discovered in those same starbursts where extended synchrotron haloes are found [e.g. @Sala13]. Thus, [*the CO-destroying CRs could be swept along the molecular gas outflow itself.*]{} Moreover, by exerting large pressures on the molecular gas of starburst galaxies (even small gas ionization fractions can allow CR-gas coupling and momentum transfer from CRs to gas), CRs could even be driving these fast molecular outflows [@Hana13; @Giri16]. Low-density and gravitationally unbound molecular gas is to be expected in such galactic outflows, a result of Kelvin-Helmholtz instabilities and shear acting on the envelopes of denser clouds in the outflow. Such a gas phase could carry a significant mass fraction of the outflow, while remaining CO-invisible, because of large scale CO destruction induced by the CRs carried within the same outflow. Large masses of low-density molecular gas could be present in galactic molecular gas outflows given the trend of progressively larger amounts of molecular gas mass discovered in them, the lower the critical density of the line tracer used to reveal them is [@Cico12]. A CR-irradiation of the outflowing molecular gas by the relativistic plasma carried along with it thus points to the possibility of CO-poor/C[i]{}-rich molecular gas in powerful galactic outflows. Using the magnetic field value found for the outflow in M82 of $\rm \langle B\rangle=25\,\mu$G [@Adeb13], and the equipartition assumption between magnetic field and CR energy densities, yields a CR energy density boost expected in such an outflow of $U_{\rm CR}=(\langle {\rm B}\rangle/\langle {\rm B}_{\rm Gal}\rangle)^2\times U_{\rm CR,Gal}\sim 17\times U_{\rm CR,Gal}$ (for $\langle \rm B_{Gal}\rangle\sim 6\,\mu$G). Stronger magnetic fields of $\rm \langle B\rangle\sim(35-40)\mu$m have been found in extended synchrotron haloes around galaxies [@Lain08], corresponding to $U_{\rm CR}\sim (34-44)\times U_{\rm CR,Gal}$ for any concomitant $\rm H_2$ gas phase outflowing along with the CRs. For our computations, we adopt $U_{\rm CR}=30\times U_{\rm CR,Gal}$ and $\rm K_{vir}=10$ corresponding to strongly unbound gas states expected in such galactic outflows. Our results are shown in Figure \[fig:highzcr\], from where it can be seen that CO is destroyed very effectively over the whole density range we consider in this work, while C[i]{} remains abundant. Thus sensitive C[i]{} 1-0, 2-1 imaging observations of galactic gas outflows could reveal significantly more $\rm H_2$ gas mass than CO lines currently find. C[ii]{} also remains abundant but starts strongly recombining to C[i]{} at the high density end of $n\sim200\,{\rm cm}^{-3}$, as expected from our previous work [@Bisb17]. {width="\textwidth"} In Figure \[fig:panels\], we correlate the H$_2$ column density with the T$_{\rm b}$(C[i]{})/T$_{\rm b}$(CO) brightness ratio for the four different ISM environments we considered. As expected, high H$_2$ column densities (i.e. $N({\rm H}_2)\gtrsim10^{22}\,{\rm cm}^{-2}$) correspond to a small T$_{\rm b}$(C[i]{})/T$_{\rm b}$(CO) brightness ratio, implying that CO (1-0) is brighter than C[i]{} (1-0). From Figures \[fig:cds\] and \[fig:highzcr\], it can be seen that low $N$(H$_2$) corresponds to gas temperatures that may exceed $\sim50\,{\rm K}$, particularly when the CR ionization rate is elevated (i.e. $\zeta'\gtrsim30$). This may result in a local formation of CO via the OH channel [see @Bisb17 for further details], which then increases T$_{\rm b}$(CO) and decreases T$_{\rm b}$(C[i]{}). This in turn, lowers their brightness ratio that we examine, resulting in a local minimum at $N$(H$_2$)$\sim0.8-1\times10^{21}\,{\rm cm}^{-2}$ where FUV radiation is also important, as it can be seen from panels b and c of Figure \[fig:panels\]. For still higher CR energy densities (i.e. $\zeta'\sim300$), all simulated GMCs become brighter in C[i]{} (1-0) than in CO (1-0). Radio galaxies: Molecular gas and AGN-injected Cosmic rays {#ssec:radio} ========================================================== The powerful jets of radio galaxies can carry CRs to great distances outside the galaxy where the radio-loud AGN resides, with the CRs also diffusing around the immediate confines of the jet to form magnetized bubbles [@Guo12]. It is possible that such powerful jets can also entrain molecular gas from the ambient ISM and drive molecular gas outflows, an effect already shown in the low-powered jets found in AGN-harbouring spirals like NGC1068 [@Lama16], and NGC4258 [@Krau07b]. This is not so difficult to imagine, given that such jets are often launched from within gas-rich galaxies where the radio-loud AGN resides. Moreover, the heavily flared star-forming $\rm H_2$ gas disks expected around AGN [@Wada09] can act as a constant source of molecular gas to be entrained by jets ‘firing’ from the AGN, given that the spins of disk and BH rotation will not necessarily be aligned. The interaction of cold gas with radio jets has already been studied theoretically and invoked to explain the H[i]{} outflows found in some powerful radio galaxies [@Morg05; @Krau07], which have been since augmented by fast molecular gas outflows observed via the traditional method of CO lines [@Morg15; @Morg16]. Finally, significant amounts of molecular gas, the fuel of SF, driven out of radio-galaxies via jet-powered outflows provides a natural explanation for the so-called ‘alignment effect’ observed in many gas-rich high-redshift radio-galaxies [e.g. @McCa87; @Pent01]. Should molecular gas be ‘caught’ in a radio-jet driven outflow, it will be subjected to its withering CR-intense environment, which could quickly render it CO-poor/invisible. Moreover, a great deal of low-density gas (i.e. the phase where CR-induced and far-UV induced CO destruction are most effective), is to be expected in such environments for the same reasons mentioned in §\[ssec:outflows\] (to which perhaps MHD-driven shear should also be added). Magnetic fields in radio-jets can be strong with $\rm \langle B\rangle \sim(35-100)\mu$G [@Staw05; @Ostr98]. Assuming equipartition between CR and magnetic field energy, yields CR energy densities within radio-jet environments of $\rm U_{CR}\sim(35-280)\times\rm \langle U_{CR,Gal}\rangle$. Thus, the CR energy density boost expected within jets and the areas near them approaches those expected in vigorously SF galaxies, even as the source of CRs is different. In Figure \[fig:highzcr\], we show images of the relative H[i]{}, $\rm H_2$, CO, C[i]{} and C[ii]{} distributions for low-density clouds subjected to CR-irradiation environments of $\rm U_{CR}=300\times\rm \langle U_{CR,Gal}\rangle$, and in a strongly unbound state ($\rm K_{vir}=10$), that are plausible for gas found in radio-jets and radio-loud AGNs. CO is very effectively destroyed while both C[i]{} and C[ii]{} remain abundant for such molecular gas (see Fig. \[fig:panels\]d). {width="80.00000%"} Nearby radio galaxies, such as Cen A and Minkowski’s object, are excellent targets for detecting CO-poor/C[i]{}-rich molecular gas in jets and their vicinity. In Cen A, such gas may have already been detected, even if its bright C[i]{} emission has been attributed to PDRs rather than CRDRs [@Isra17][^5] In the case of Minkowski’s object, sensitive CO observations detected only a small CO-marked $\rm H_2$ gas reservoir fueling the star formation observed along its jet, making this object an outlier of the Schmidt-Kennicutt (S-K) relation [@Salo15; @Lacy17]. Sensitive \[C[i]{}\] (1-0) imaging of this galaxy is particularly promising for detecting molecular gas that may have been rendered CO-invisible by the high $U_{\rm CR}$ expected in its radio jet. Similar observations of radio galaxies at high redshifts are even more promising, because in the early Universe these galaxies reside in very H$_2$-rich hosts with plenty of gas to be entrained/impacted by their powerful jets. Finally, \[C[i]{}\] (1-0) and even (2-1) line imaging observations can be conducted in the distant Universe with modest $T_{\rm sys}$ since their frequencies will be redshifted to more transparent parts of the Earth’s atmosphere. Main Sequence galaxies: CRs and low-metallicity gas {#ssec:ms} =================================================== A mostly CR-regulated average \[CO\]/\[H$_2$\] abundance in the ISM of Main Sequence galaxies and the possibility of large fractions of their molecular gas mass having their CO destroyed (with adverse effects on the calibration attempts of their X$_{\rm CO}$-factor [e.g. @Genz12; @Carl17]) have been described by @Bisb15 [@Bisb17]. Our current models of lower density gas clouds, reinforce this view by demonstrating that the conditions for a phase transition towards very CO-poor gas remain favorable for low densities [*even for Galactic or only modestly elevated levels of CR energy densities.*]{} For low-metallicity gas, the effects of CO destruction are dramatic (Figure \[fig:cds\]), and could explain the lack of total CO emission seen in some metal-poor MS galaxies [@Genz12]. Such CO-dark molecular gas may also exist in the outer regions of MS galaxies, which otherwise show CO-bright and metal-rich molecular gas in their inner SF regions. One may ask whether significant amounts of lower-density molecular gas can indeed exist in such vigorously SF disks. Given that star formation is a low-efficiency process, both on the mass scale of individual molecular clouds and that of galaxy-sized reservoirs, we can expect that large amounts of non-SF molecular gas will always be present in SF galaxies, either in their disk, or expelled out by SF feedback (e.g. the massive non-SF ${\rm H}_2 $ gas reservoir in the outer regions of M 82). With CRs able to ‘leak’ out of the SF areas along molecular gas outflows and beyond the SF disk, they can subject non-SF low density gas to significant levels of CR irradiation, destroying its CO and replacing with C[i]{} and C[ ii]{}. Sensitive C[i]{} line imaging of such systems can show whether this is so. Nevertheless at redshifts of $z\ga 4$, C[ii]{} line imaging will be the most effective in revealing such gas. C[i]{} line imaging and a critical brightness limit --------------------------------------------------- There are now several PdBI and ALMA C[i]{} (1-0), (2-1) line observations of galaxies in the distant Universe [@Walt11; @Zhan14; @Gull16; @Popp17; @Both17; @Emon17] as well as local ones using [*Herschel*]{} [@Jiao17]. In all cases, the galaxies are either unresolved or only marginally resolved in \[C[i]{}\] line emission. Such observations do not yet allow a detailed study of relative distributions of C[i]{} vs CO 1-0 emission in galaxies, since unresolved or marginally resolved emission will always be dominated by the warm and dense star-forming $\rm H_2$ gas (where both CO and C[i]{} are abundant). However, @Krip16 recently provided a 3$''$ resolution of CO(1-0) and \[C[i]{}\] (1-0) for the SF galaxy NGC 253. Intriguingly, recent high-resolution ALMA imaging of \[C[ i]{}\] (1-0) lines in a local LIRG shows significant differences between the CO and C[i]{}-bright $\rm H_2$ gas distribution (Zhang et al. [*in prep.*]{}). High-resolution imaging of CO and \[C[i]{}\] (1-0), (2-1) lines of local LIRGS is necessary to compare their emission distribution in detail and deduce the corresponding $\rm H_2$ mass distributions. The resolution however [*must be high enough to separate the more compact SF molecular gas distributions from the more extended SF-quiescent, and perhaps lower density gas reservoirs.*]{} The strongest requirement for such imaging is then placed by the conditions in latter where observations yield brightness temperature ratios T$_{\rm b}$(C[i]{})/T$_{\rm b}$(CO) $\sim0.15$ [@Papa04 and references therein]. From Figure \[fig:panels\] we can see that a C[i]{}/CO (J=1-0) relative brightness cutoff of T$_{\rm b}$(C[i]{})/T$_{\rm b}$(CO)$\sim0.10$ is adequate to encompass all the $\rm H_2$ gas in the clouds, irrespective of its thermal/chemical state and their intrinsic average \[CO\]/\[C[i]{}\] ratio. Predictably, it is the denser/colder inner regions of our low density cloud models (and where the average $N$(H$_2$) is the highest), where this ratio drops to its lowest values, as is indeed found by simulations and observations of cold/low-density non-SF molecular cloud regions in the Galaxy [@Papa04; @Offn14; @Lo14; @Glov16]. It is perhaps more beneficial to conduct such high brightness sensitivity C[i]{} (and CO) line imaging for SF disks in the more distant Universe (e.g in the disks of MS and submm galaxies, radio galaxy environments) as to take advantage of the lower $T_{\rm sys}$ ALMA values for the C[i]{} redshifted frequencies. Even at $z\ga 0.4$, the C[i]{} (1-0) frequency already shifts to $\la 350$GHz, where the atmosphere becomes much more transparent and detectors less noisy. Furthermore, it is during earlier cosmic epochs (and thus distances) when galaxies become more $\rm H_2$-rich, while the CRs generated by their elevated SFR can induce large scale CO destruction leaving behind C[i]{}-rich gas. Conclusions {#sec:3} =========== The astrochemistry that demonstrated the critical role of CRs in regulating the average $\rm [CO]/[H_2]$ abundance for the bulk of the molecular gas in SF galaxies (except in localised surface PDRs near O,B stars) indicates that CO-poor/C[i]{}-rich can exist not only in the highly CR-irradiated ISM environments of starbursts but also in environments with much lower levels of CR-irradiation, if the average molecular gas density is low. A low-density molecular gas phase with CR irradiation levels high enough to render it very CO-poor and C[ i]{}/C[ii]{}-rich can be found in a number of places in the Universe, namely: - Low-density envelopes around the CO-rich parts of ordinary molecular clouds in the Milky Way, - Molecular gas outflows from galaxies, induced by starburst and/or AGNs, - In main sequence galaxies both in their metal-rich and metal-poor regions (in the latter ones CR and far-UV irradiation penetrate deep in $\rm H_2$ destroying the CO tracer molecule). - In regions inside and around radio-jets and perhaps even near the cores of powerful radio galaxies. Sensitive C[i]{} line observations of such environments can perhaps find more molecular gas mass than the standard low-J CO line observations, provided a relative brightness temperature ratio of T$_{\rm b}$(C[i]{})/T$_{\rm b}$(CO)$\sim0.10$ (for $J=1-0$) is reached in well-resolved images. GMCs in the Galaxy, distant Main Sequence and submm galaxies, as well as areas around radio galaxies and their jets are all excellent targets for this kind of imaging. In the later case, Minkowski’s object as well as Cygnus A are some of the most prominent radio-loud objects for such observations in the local Universe. Acknowledgements {#acknowledgements .unnumbered} ================ The authors thank Ewine van Dishoeck for the useful comments and discussion on several aspects of this work. PPP would like to thank Ocean Divers at Sithonia, Halkidiki, Christos Douros and Sokratis Vagiannis for providing a most useful distraction during the last stages of this project. Z-Y.Z. and PPP acknowledge support from ERC in the form of the Advanced Investigator Programme, 321302, COSMICISM. Ackermann, M., Ajello, M., Atwood, W. B., et al. 2012, , 750, 3 Adebahr, B., Krause, M., Klein, U., et al. 2013, , 555, A23 Andersson, B.-G., & Wannier, P. G. 1993, , 402, 585 Bialy, S., & Sternberg, A. 2015, , 450, 4424 Bialy, S., & Sternberg, A. 2016, , 822, 83 Bialy, S., Burkhart, B., & Sternberg, A. 2017, , 843, 92 Bisbas, T. G., Bell, T. A., Viti, S., Yates, J., & Barlow, M. J. 2012, , 427, 2100 Bisbas, T. G., Papadopoulos, P. P., & Viti, S. 2015, , 803, 37 Bisbas, T. G., van Dishoeck, E. F., Papadopoulos, P. P., et al. 2017, , 839, 90 Bolatto, A. D., Jackson, J. M., & Ingalls, J. G. 1999, , 513, 275 Bolatto, A. D., Wolfire, M., & Leroy, A. K. 2013, , 51, 207 Bothwell, M. S., Aguirre, J. E., Aravena, M., et al. 2017, , 466, 2825 Bradford, C. M., Nikola, T., Stacey, G. J., et al. 2003, , 586, 891 Bryant, P. M., & Scoville, N. Z. 1996, , 457, 678 Dasyra, K. M., & Combes, F. 2012, , 541, L7 Dickman, R. L., Snell, R. L., & Schloerb, F. P. 1986, , 309, 326 Carleton, T., Cooper, M. C., Bolatto, A. D., et al. 2017, , 467, 4886 Chen, B.-Q., Liu, X.-W., Yuan, H.-B., Huang, Y., & Xiang, M.-S. 2015, , 448, 2187 Cicone, C., Feruglio, C., Maiolino, R., et al. 2012, , 543, A99 Cicone, C., Maiolino, R., Sturm, E., et al. 2014, , 562, A21 Colbert, E. J. M., Baum, S. A., Gallimore, J. F., et al. 1996, , 105, 75 Draine, B. T. 1978, , 36, 595 Emonts, B. 2017, Formation and Evolution of Galaxy Outskirts, 321, 348 Feruglio, C., Maiolino, R., Piconcelli, E., et al. 2010, , 518, L155 Genzel, R., Tacconi, L. J., Combes, F., et al. 2012, , 746, 69 Girichidis, P., Naab, T., Walch, S., et al. 2016, , 816, L19 Glover, S. C. O., & Clark, P. C. 2016, , 456, 3596 Gullberg, B., Lehnert, M. D., De Breuck, C., et al. 2016, , 591, A73 Guo, F., & Mathews, W. G. 2012, , 756, 181 Hanasz, M., Lesch, H., Naab, T., et al. 2013, , 777, L38 Heyer, M., & Dame, T. M. 2015, , 53, 583 Hollenbach, D. J., & Tielens, A. G. G. M. 1999, Reviews of Modern Physics, 71, 173 Israel, F. P., G[ü]{}sten, R., Meijerink, R., Requena-Torres, M. A., & Stutzki, J. 2017, , 599, A53 Jiao, Q., Zhao, Y., Zhu, M., et al. 2017, , 840, L18 Jura, M. 1975a, , 197, 575 Jura, M. 1975b, , 197, 581 Krause, M. 2007, , 51, 174 Krause, M., Fendt, C., & Neininger, N. 2007, , 467, 1037 Krips, M., Mart[í]{}n, S., Sakamoto, K., et al. 2016, , 592, L3 Lacy, M, Croft, S., Fragile, C., Wood, S., Nyland, K. 2017, , 838, 146 Laine, S., & Beck, R. 2008, , 673, 128-142 Lamastra, A., Fiore, F., Guetta, D., et al. 2016, , 596, A68 Langer, W. D., Velusamy, T., Pineda, J. L., Willacy, K., & Goldsmith, P. F. 2014, , 561, A122 Lo, N., Cunningham, M. R., Jones, P. A., et al. 2014, , 797, L17 Madden, S. C., Poglitsch, A., Geis, N., Stacey, G. J., & Townes, C. H. 1997, , 483, 200 Maloney, P., & Black, J. H. 1988, , 325, 389 McCarthy, P. J., van Breugel, W., Spinrad, H., & Djorgovski, S. 1987, , 321, L29 Meijerink, R., Spaans, M., Loenen, A. F., & van der Werf, P. P. 2011, , 525, A119 Meixner, M., & Tielens, A. G. G. M. 1993, , 405, 216 Morganti, R., Tadhunter, C. N., & Oosterloo, T. A. 2005, , 444, L9 Morganti, R., Oosterloo, T., Oonk, J. B. R., Frieswijk, W., & Tadhunter, C. 2015, , 580, A1 Morganti, R., Oosterloo, T., Oonk, J. B. R., Santoro, F., & Tadhunter, C. 2016, , 592, L9 Offner, S. S. R., Bisbas, T. G., Viti, S., & Bell, T. A. 2013, , 770, 49 Offner, S. S. R., Bisbas, T. G., Bell, T. A., & Viti, S. 2014, , 440, L81 Ostrowski, M. 1998, , 335, 134 Padoan, P., & Nordlund, [Å]{}. 2002, , 576, 870 Pak, S., Jaffe, D. T., van Dishoeck, E. F., Johansson, L. E. B., & Booth, R. S. 1998, , 498, 735 Papadopoulos, P. P., Thi, W.-F., & Viti, S. 2002, , 579, 270 Papadopoulos, P. P., Thi, W.-F., & Viti, S. 2004, , 351, 147 Papadopoulos, P. P. 2010, , 720, 226 Papadopoulos, P. P., van der Werf, P. P., Xilouris, E. M., et al. 2012, , 426, 2601 Papadopoulos, P. P., Zhang, Z.-Y., Xilouris, E. M., et al. 2014, , 788, 153 Pentericci, L., McCarthy, P. J., R[ö]{}ttgering, H. J. A., et al. 2001, , 135, 63 Pineda, J. L., Langer, W. D., Velusamy, T., & Goldsmith, P. F. 2013, , 554, A103 Popping, G., Decarli, R., Man, A. W. S., et al. 2017, , 602, A11 Remy, Q., Grenier, I. A., Marshall, D. J., & Casandjian, J. M. 2017, , 601, A78 Roman-Duval, J., Heyer, M., Brunt, C. M., et al. 2016, , 818, 144 Sakamoto, S., Hasegawa, T., Handa, T., Hayashi, M., & Oka, T. 1997, , 486, 276 Salak, D., Nakai, N., Miyamoto, Y., Yamauchi, A., & Tsuru, T. G. 2013, , 65, 66 Salom[é]{}, Q., Salom[é]{}, P., & Combes, F. 2015, , 574, A34 Sandstrom, K. M., Leroy, A. K., Walter, F., et al. 2013, , 777, 5 Seaquist, E. R., & Odegard, N. 1991, , 369, 320 Shi, Y., Wang, J., Zhang, Z.-Y., Gao, Y., Hao, C.-N., Xia, X.-Y., Gu, Q. 2017, NatCo, 713789 Shi, S.-C.., Paine, S., Yao, Q.-J. Lin, Z.-H., Duan, W.-Y., Matsou, H., Zhang Q., Yang, J., Ashley, M.C.B., Shang Z., Hu, Z.-W., 2016, NatAs,1,1 Shaya, E. J., & Federman, S. R. 1987, , 319, 76 Shull, J. M., Tumlinson, J., Jenkins, E. B., et al. 2000, , 538, L73 Sodroski, T. J. 1991, , 366, 95 Solomon, P. M., Downes, D., & Radford, S. J. E. 1992, , 387, L55 Solomon, P. M., Downes, D., Radford, S. J. E., & Barrett, J. W. 1997, , 478, 144 Spaans, M., & van Dishoeck, E. F. 1997, , 323, 953 Stawarz, [Ł]{}., Siemiginowska, A., Ostrowski, M., & Sikora, M. 2005, , 626, 120 van der Werf, P. P., Isaak, K. G., Meijerink, R., et al. 2010, , 518, L42 van Dishoeck, E. F., & Black, J. H. 1986, , 62, 109 Wada, K., Papadopoulos, P. P., & Spaans, M. 2009, , 702, 63 Walch, S., Whitworth, A. P., Bisbas, T. G., Hubber, D. A., & W[ü]{}nsch, R. 2015, , 452, 2794 Walter, F., Wei[ß]{}, A., Downes, D., Decarli, R., & Henkel, C. 2011, , 730, 18 Wolfire, M. G., Hollenbach, D., & McKee, C. F. 2010, , 716, 1191 Yang, J., Zuo, Y.-X., Lou, Z., Cheng, J.-Q.; Zhang, Q.-Z.; Shi, S.-C.; Huang, J.-S.; Yao, Q.; Wang, Z., 2013, RAA, 13, 1493 Young, J. S., & Scoville, N. Z. 1991, , 29, 581 Zhang, Z.-Y., Henkel, C., Gao, Y., et al. 2014, , 568, A122 \[lastpage\] [^1]: E-mail: [email protected] [^2]: https://uclchem.github.io/ [^3]: With $\rm K_{vir}=(dV/dR)/(dV/dR)_{vir}$ a value of $\rm K_{vir}=1$ signifies self-gravitating clouds, see [@Papa14] for details [^4]: http://soral.as.arizona.edu/HEAT/instrument/ [^5]: [*Relative*]{} molecular line ratios can always be fitted with PDRs [e.g. @vand10], yet the decisive test whether PDRs or CRDRs/XDRs are responsible for any observed extragalactic lines must use the relative gas mass fractions of warm/dense gas and warm dust [see @Brad03; @Papa14; @vand10]. No such tests have been done for the C[i]{}-bright molecular gas in Cen A.

ArXiv

--- abstract: 'The dense environment of globular clusters (GCs) can facilitate the formation of binary black holes (BBHs), some of which can merge with gravitational waves (GW) within the age of the Universe. We have performed a survey of Monte-Carlo simulations following the dynamical evolution of GCs with different masses, sizes and binary fractions and explored the impact of the host GC properties on the formation of BBH mergers. We find that the number of BBH mergers from GCs is determined by the GC’s initial mass, size and primordial binary fraction. We identify two groups of BBH mergers: a primordial group whose formation does not depend on cluster’s dynamics and a dynamical group whose formation is driven by the cluster’s dynamical evolution. We show how the BBH origin affects the BBH mergers’ main properties such as the chirp mass and merging time distributions. We provide analytic expressions for the dependence of the number of BBH mergers from individual GCs on the main cluster’s structural properties and the time evolution of the merger rates of these BBHs. These expressions provide an essential ingredient for a general framework allowing to estimate the merger rate density. Using the relations found in our study, we find a local merger rate density of 0.18–1.8 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for primordial BBH mergers and 0.6–18 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for dynamical BBH mergers, depending on the GC mass and size distributions, initial binary fraction and the number density of GCs in the Universe.' author: - | Jongsuk Hong$^{1,2}$[^1], Enrico Vesperini$^2$, Abbas Askar$^{3,4}$, Mirek Giersz$^3$, Magdalena Szkudlarek$^5$ and Tomasz Bulik$^6$\ $^1$ Kavli Institute for Astronomy and Astrophysics, Peking University, Yi He Yuan Lu 5, HaiDian District, Beijing 100871, China\ $^2$ Department of Astronomy, Indiana University, Bloomington, Swain West, 727 E. 3rd Street, IN, 47401, USA\ $^3$ Nicolaus Copernicus Astronomical Centre, Polish Academy of Sciences, ul. Bartycka 18, 00-716 Warsaw, Poland\ $^4$ Lund Observatory, Department of Astronomy, and Theoretical Physics, Lund University, Box 43, SE-221 00 Lund, Sweden\ $^5$ Janusz Gil Institute of Astronomy, University of Zielona Góra, Licealna 9, 65-407 Zielona Góra, Poland\ $^6$ Astronomical Observatory, University of Warsaw, Al. Ujazdowskie 4, 00-478 Warsaw, Poland\ date: 'Accepted 2018 August 10. Received 2018 July 11; in original form 2018 May 10' title: 'Binary Black Hole Mergers from Globular Clusters: the Impact of Globular Cluster Properties' --- \[firstpage\] globular clusters: general — stars: black holes — gravitational waves Introduction\[S1\] ================== Recently, Advanced LIGO has made the first detection of gravitational waves (GWs) and opened a new window to explore very energetic events [@2016PhRvL.116f1102A]. The event responsible for the GWs revealed by this first detection, GW150914, was the merger of black holes (BHs) in a binary system and it has been followed by four more detections of merging binary BHs [BBHs, @2016PhRvL.116x1103A; @2016PhRvX...6d1015A; @2017PhRvL.118v1101A; @2017PhRvL.119n1101A] and one merging binary neutron stars [@2017PhRvL.119p1101A]. A number of different scenarios for the formation of these merging compact binaries have been proposed so far; the different formation mechanisms proposed have invoked isolated binary evolution [e.g. @2002ApJ...572..407B; @2007ApJ...662..504B; @2012ApJ...759...52D], three-body interactions in dense stellar systems [e.g. @2000ApJ...528L..17P; @2006ApJ...637..937O; @2016ApJ...824L...8R; @2018ApJ...855..124S], the orbital evolution of hierarchical systems [e.g. @2012MNRAS.422..841A; @2014ApJ...781...45A; @2016ApJ...816...65A; @2018ApJ...856..140H; @2018arXiv180508212R], relativistic captures [e.g. @2009MNRAS.395.2127O; @2015MNRAS.448..754H; @2017PhRvD..96h4009B; @2018ApJ...860....5G]. As for the environment in which these compact binaries might form, the scenarios proposed include globular clusters , young/open clusters [e.g. @2014MNRAS.441.3703Z; @2017MNRAS.467..524B; @2018MNRAS.473..909B] and galactic nuclei [e.g. @2009MNRAS.395.2127O; @2012ApJ...757...27A; @2018MNRAS.474.5672L]. An important aspect concerning the formation of BHs is the mass fall-back after the supernova explosions. As discussed by @2002ApJ...572..407B, this fall-back (i.e., failed supernovae) mechanism can increase the remnant BH masses and reduce the natal kicks, which, in turn, can lead to a larger fraction of BHs retained inside the host stellar system [@2015ApJ...800....9M; @2016MNRAS.458.1450W; @2016MNRAS.463.2109R; @2018MNRAS.478.1844A; @2018MNRAS.479.4652A]. The retention of a large number of BHs can significantly influence not only the internal dynamics [e.g. @2013MNRAS.432.2779B] but also the observational properties of star clusters [e.g. @2008MNRAS.386...65M; @2017ApJ...834...68C; @2017arXiv171203979W; @2018ApJ...855L..15K; @2018MNRAS.476.5274L; @2018MNRAS.479.4652A; @2018MNRAS.478.1844A]. The retained BHs in a dense stellar system rapidly sink to the centre of the system due to the effects of dynamical friction and form a compact subsystem predominantly composed of BHs, on a timescale of few hundreds Myr [@2013ApJ...763L..15M]. Due to its short relaxation timescale, a BH subsystem quickly undergoes core collapse and generate energy through the formation and dynamical interactions of BBHs [@2013MNRAS.432.2779B]. Recoil velocities acquired during binary-single and binary-binary interactions can result in BHs ejection from GCs, and some numerical studies [e.g. @2015ApJ...800....9M; @2017MNRAS.469.4665P] suggested that $\sim$30% of dynamically escaping BHs are in binary systems, some of which are expected to merge within the age of the Universe. Moreover, @2018MNRAS.478.1844A suggested that some massive Galactic GCs (GGCs) are still harbouring a large number of BHs and that the formation and ejection of BBHs can still be ongoing in those GGCs. The BBHs’ properties as well as the merger and detection rates of these BBHs are significantly affected not only by the global properties of host GCs such as the initial mass, size and the metallicity [e.g. @2016PhRvD..93h4029R; @2017MNRAS.464L..36A; @2017ApJ...836L..26C] but also by the GC’s stellar initial mass function and the prescriptions for the mass fall-back and the stellar wind [see e.g. @2017ApJ...834...68C and the references therein]. In this paper, we present an analysis of the survey of Monte-Carlo simulations of GCs evolution from @2017MNRAS.464.2511H and of another set of simulations performed specifically for this paper aimed at a detailed characterization of the link between the properties of BBHs formed in GCs and the structure of the host GCs. Understanding the connection between the properties of the BBHs and those of their host GCs is an important step for more realistic estimates of the merger rate of BBHs. We extracted the information of all escaping BBHs from our GC simulations and found some empirical relations between the properties of merging BBHs and those of the host GCs. These relations provide an essential ingredient to estimate the merger and detection rates of BBHs for any assumed GC system properties and GC formation rate. We also provided examples of estimates of the local merger rate density for various assumptions concerning the properties of GC systems. The structure of this paper is as follows. We briefly describe the numerical method, the initial conditions and assumptions of our GC simulations in Section \[S2\]. The relations between the properties of merging BBHs and those of the host GC are presented in Section \[S3\]. In Section \[S4\], we then estimate the local merger rate density based on these relations. We conclude with a summary of our results in Section \[S5\]. Methods and Initial Conditions\[S2\] ==================================== The models used for this study are those of the survey of Monte-Carlo simulations presented in @2017MNRAS.464.2511H. The simulations followed the evolution of 81 cluster models with a variety of initial number of stars ($N=2\times10^5, 5\times10^5, 10^6$), half-mass radii ($r_{\rm h}=1, 2, 4$ pc), binary fractions (10, 20, 50 per cent) and galactocentric distances ($r_{\rm G}=4, 8, 16$ kpc) [see Table 1 in @2017MNRAS.464.2511H] and were run with the [mocca]{} code [@2013MNRAS.431.2184G; @2013MNRAS.429.1221H]. The initial density structure of clusters follows the @1966AJ.....71...64K density profile with the dimensionless central potential, $W_{0} = 7$. We have adopted @2001MNRAS.322..231K initial mass function with the mass range of stars from 0.1 to 100 M$_{\odot}$. The metallicity is fixed to $Z = 0.001$ for all our simulation models. All single and binary stars in the simulations evolve according to the stellar evolution recipes [SSE & BSE, @2000MNRAS.315..543H; @2002MNRAS.329..897H] implemented in the [mocca]{} code. We used the stellar wind prescription of SSE and BSE. In all of our simulations, we adopt the mass fall-back mechanism [@2002ApJ...572..407B] modifying the natal kicks for BHs. All our GC simulation models are limited by the tidal field from the host galaxies with a realistic treatment of escaping stars based on @2000MNRAS.318..753F [see also @2013MNRAS.431.2184G]. For the parameter spaces of the initial conditions considered in @2017MNRAS.464.2511H, the ratio of the half-mass radius to the tidal radius for the GC simulation models ranges from 0.005 to 0.09. In @2017MNRAS.464.2511H, we used the initial binary distribution (e.g. eccentricity, semi-major axis and the mass ratio) based on the [*initial binary population*]{} (hereafter IBP) in which the orbital parameters of short-period proto-binaries are redistributed by mutual interactions between binary components (e.g. mass transfer and tidal circularization) due to the large stellar radii during pre-main-sequence stage as suggested by @1995MNRAS.277.1507K [see also @2013pss5.book..115K].[^2] As the results of this proto-binary evolution, short period binaries with large eccentricity are preferentially depleted and tend to have similar masses. For this study, we have also run another set of 81 simulations with the [*birth binary population*]{} (hereafter BBP) from @1995MNRAS.277.1507K that follows the period distribution and the thermal eccentricity distribution to investigate the effects of the initial orbital properties of primordial binaries on the formation, dynamical evolution of BBHs and the rate of merger events among these BBHs from GCs. bbh mergers from gc simulation models\[S3\] =========================================== Correlation between the number of BBH mergers and GC properties\[S3.1\] ----------------------------------------------------------------------- ![Number of BBH mergers versus the parameter $\gamma$ defined in Eq. (\[E1\]). Different symbol types, sizes and colors represent models with different initial number of stars, half-mass radii, galactocentric distances and binary fractions [see also @2017MNRAS.464.2511H]. Dashed lines indicate the two times of Poisson errors of the locus line.[]{data-label="F1"}](figure_nmer.eps){width="1.0\columnwidth"} ![Same as Fig. \[F1\] but for only primordial BBH mergers that escape from GCs due to the natal kicks during supernova explosions.[]{data-label="F2"}](figure_npr.eps){width="1.0\columnwidth"} ![Same as Fig. \[F1\] but for only dynamical BBH mergers that form dynamically inside GCs by three-body or exchange encounters and subsequently escape.[]{data-label="F3"}](figure_ndy.eps){width="1.0\columnwidth"} We first focus on the presentation of our results for the simulations with the IBP distribution. After 12 Gyr of evolution, the 81 GC models explored for this study produced 9519 escaping BBHs and 3402 of them emit GWs and merge within 12 Gyr. To illustrate the dependence of the GW events on the properties of the host GCs in which the BBH formed, we show the correlation between the number of merging BBHs, $N_{\rm merg}$ and the initial properties of GCs in Fig. \[F1\]. We determine the number of BBH mergers that escape from GCs and subsequently merge within 12 Gyr, i.e., $t_{\rm merg}\equiv t_{\rm esc}+t_{\rm Peters} <$ 12 Gyr, where $t_{\rm esc}$ and $t_{\rm Peters}$ are, respectively, the BBH escaping time and the @1964PhRv..136.1224P timescale for GW coalescence of BBHs calculated using their semi-major axis and eccentricity at the moment of escape. We found that the number of merging BBHs is closely correlated with a parameter, $\gamma$, defined as $$\label{E1} \gamma\equiv A\cdot \frac{M_{0}}{10^{5}{\rm M}_{\odot}}\times \Big(\frac{\rho_{\rm h}}{10^{5}{\rm M}_{\odot}{\rm pc}^{-3}}\Big)^{\alpha} + B\cdot \frac{M_{0}}{10^{5}{\rm M}_{\odot}}\times f_{\rm b,0}$$ where $M_{0}$, $\rho_{\rm h}$ and $f_{\rm b,0}$ are, respectively, the initial total mass, initial half-mass density (i.e., mean density within the half-mass radius) and the initial primordial binary fraction; $A$, $B$ and $\alpha$ are the fitting parameters. Our best fitting result that minimizes the $\chi^{2}$ value is ($A$, $B$, $\alpha$) $=$ (12.53$\pm$0.22, 6.89$\pm$0.84, 0.33$\pm$0.02). The uncertainty on the best-fit parameters is determined by calculating the values for which $\chi^{2}=\chi_{\rm min}^{2}+1$ [@1976ApJ...210..642A] by assuming that the $\chi^{2}$ distribution in 1D parameter space is simply a quadratic function. Eq. (\[E1\]) implies that there are two main formation channels for BBH mergers from GCs; one, the primordial channel, is related only to the primordial binary fraction and binary stellar evolution and is described by the second term of Eq. (\[E1\]). The other channel is affected by the cluster’s internal dynamics and its contribution to the total number of merging BBHs is described by the first term in Eq. (\[E1\]) (hereafter we will refer to this as the dynamical channel). The number of BBH mergers from the primordial channel, $N_{\rm merg,p}$, depends, as was to be expected, only on the initial binary fraction and the total mass of GCs. The number of merging BBHs from the dynamical channel, $N_{\rm merg,d}$, on the other hand is the result of the combined effects of a number of processes affected by a variety of structural parameters (e.g. encounter rates, hardening rate per encounters, central velocity dispersion, ejection rate, etc.); our results show that the number of merging BBHs resulting from the complex interplay of all these processes is well described by a parameter with a simple dependence on the cluster’s mass and half-mass density. We point out the number of merging BBHs does not show any significant dependence on the galactocentric distance. For the compact clusters explored in our survey, this is to be expected as the population of merging BBHs escape from clusters as a result of either natal kicks following supernova explosion or ejection from close encounters in the cluster’s inner regions. The galactocentric distance and the strength of the tidal field, on the other hand, are relevant for the more gradual evaporation process which is not important for the BBH population studied here. Note, however, that there is a broad trend of metallicity with galactocentric distance [e.g. @1994AJ....108.1292D] that can affect the properties of BH populations and the formation of BBHs accordingly. In order to better illustrate the relative importance of the two channels, we have divided the merging BBHs escaping from our GC simulation models into two groups according to their origin. We have classified BBHs that escape through the natal kick after supernova explosions as merger candidates with the primordial origin (789 of all merging BBHs); all the other merging BBHs are classified as dynamical BBH mergers (2613 of all merging BBHs). Figs. \[F2\] and \[F3\] show the dependence of the number of primordial and dynamical BBH mergers on, respectively, $\gamma_{\rm pri}$ and $\gamma_{\rm dyn}$, the two terms already introduced in the definition of the parameter $\gamma$ in Eq. (\[F1\]), $$\label{E2} \gamma_{\rm dyn}\equiv A\cdot \frac{M_{0}}{10^{5}{\rm M}_{\odot}}\times \Big(\frac{\rho_{\rm h}}{10^{5}{\rm M}_{\odot}{\rm pc}^{-3}}\Big)^{\alpha},$$ $$\label{E3} \gamma_{\rm pri}\equiv B\cdot \frac{M_{0}}{10^{5}{\rm M}_{\odot}}\times f_{\rm b,0}.$$ The best fitting parameters are ($A$, $\alpha$) $=$ (12.30$\pm$0.44, 0.33$\pm$0.01) for dynamical BBH mergers and ($B$) $=$ (6.64$\pm$0.25) for primordial BBH mergers, separately. It is apparent that the best-fit parameters calculated for the two merging BBH population separately are very similar to those obtained from fitting all the BBH mergers together. ![Same as Fig. \[F1\] but for GC simulation models with the BBP distribution.[]{data-label="F4"}](figure_nmer_bbp.eps){width="1.0\columnwidth"} ![Same as Fig. \[F3\] but $\gamma_{\rm dyn, 12}$ is defined by using current GC properties.[]{data-label="F5"}](figure_ndy12.eps){width="1.0\columnwidth"} ![Same as Fig. \[F5\] but for the BBP distribution.[]{data-label="F6"}](figure_ndy12_bbp.eps){width="1.0\columnwidth"} For models with the BBP distribution, we find 6755 BBHs escaping from GCs in total and 2382 of them merge before 12 Gyr. In Fig. \[F4\], we show the relation between $N_{\rm merg}$ and $\gamma$ for the simulations with the BBP distribution. The best fitting parameters for this set of simulations are ($A$, $B$, $\alpha$) $=$ (14.55$\pm$0.31, -0.03$\pm$0.03, 0.38$\pm$0.01). For the BBP distribution, we obtained only 1 primordial BBH merger while all the other BBH mergers have a dynamical origin and this explains the lack of dependence of the number of merging BBHs on the initial binary fraction. This is due to the pairing of primordial binary components of massive stars. For the BBP distribution, we used a uniform pairing so that the masses of secondary stars are uniformly chosen in between \[$m_{\rm min}$, $m_{\rm pri}$\]. For the IBP distribution with mass feeding algorithm [@2013pss5.book..115K] used in the simulations, the pairing rule for the massive binaries is different. Some theoretical studies [e.g. @2000MNRAS.314...33B; @2002MNRAS.336..705B; @2007ApJ...661.1034K] for the formation of massive binaries in star forming regions show that massive binaries tend to evolve to the mass ratio $q\sim1$ during the proto-binary stage [note that, however, some mechanisms such as the magnetic breaking during proto-binary evolution can prevent the evolution of mass ratio of proto-binaries toward $q \sim 1$; see e.g. @2013ApJ...763....7Z]. GC simulation models with IBP distribution in this study adopted this condition so that the massive binaries (especially for $m>5$M$_{\odot}$) that can be the progenitors of BHs are more likely to have $q\sim1$ and evolve to BBHs.[^3] On the other hand, with the BBP distribution, massive stars are initially coupled with less massive stars and therefore require exchange encounters to become BBHs. Figs. \[F5\] and \[F6\] show the correlation between $N_{\rm merg}$ and the current (at $T = 12$ Gyr) GC properties for models with different initial binary distributions. We only present the correlation for dynamical BBH mergers in this figure. The best fitting parameters of Eq. (\[E2\]) for the correlation between the number of dynamical BBH mergers and the current GC properties are ($A$, $\alpha$) $=$ (303.1$\pm$6.1, 0.43$\pm$0.02) for the IBP distribution and ($A$, $\alpha$) $=$ (527.7$\pm$11.1, 0.52$\pm$0.02) for the BBP distribution, respectively. This relation can be used to estimate current merger rate for the MW GCs or Local Group GCs (see Section \[S4.3\]). The values of $A$ and $B$ in the relation with the current GC properties are larger than those obtained when the initial GC properties are used because the current masses of GCs are smaller than the initial ones. Time evolution of merger rates\[S3.2\] -------------------------------------- ![Time evolution of BBH merger rate normalized to the average merger rate over 12 Gyr. Different colors for histograms represent the time evolution of merger rates from GC simulation models with different initial number of stars. Solid lines show the best fitting results (see Section \[S3.2\] for further details). []{data-label="F7"}](figure_tmer.eps){width="1.0\columnwidth"} In the previous section, we provided a relation between the (initial or current) GC properties and the expected number of BBHs that escape from GCs and merge by emitting GWs within 12 Gyr. However, many numerical studies aimed at the estimation of the merger rate of BBHs from GCs have shown that the merger rate is time-dependent [e.g. @2017MNRAS.464L..36A; @2017PASJ...69...94F]. Since one of our main goals is to provide an empirical relation that allows to calculate the merger rates and the detection rates of BBH merger events from GCs, we have also calculated a model for the time dependence of the rate of BBH mergers from GCs. In Fig. \[F7\], we present the histograms of the merging time, $t_{\rm merg}$, of BBHs, with the numbers normalized to the merger rate averaged over the 12 Gyr of evolution. This figure clearly shows that the merger rate decreases with time very rapidly. Initially the merger rate is as high as 10 times the average merger rate while the merger rate at 12 Gyr is $\sim$5 times lower than the average merger rate. We found that the time evolution of the merger rate is well described by the following expression, $$\label{E4} \mathcal{R} \equiv \left<\mathcal{R}\right> ae^{-b(t/t_{12})^{c}},$$ where $\left<\mathcal{R}\right>$ is the average merger rate over 12 Gyr, and $a$, $b$ and $c$ are the fitting parameters. $\mathcal{R}$ is defined as the number of mergers per unit time bin. Our best-fit parameters for the time evolution of the merger rate are ($a$, $b$, $c$) $=$ (13.01$\pm$3.00, 4.14$\pm$0.19, 0.35$\pm$0.04) for the IBP distribution. In order to test the dependence of the best-fit parameters on the cluster’s initial number of stars, we repeated the fit for subsets of the simulation data with different initial number of stars ($N=2\times 10^5$, $5\times 10^5$ and $10^6$) and found that the time evolution of the normalized merger rates does not significantly depend on the initial number of stars in GCs. We also tested other subsets with different half-mass radii, galactocentric distances, binary fractions and we did not find any significant discrepancy from the best fitting parameters obtained for the entire survey of simulations. ![Same as Fig. \[F7\] but the BBH mergers used for the fitting are separated into the primordial and dynamical BBH mergers. Dotted line indicates the time evolution of the merger rate for GC simulation models with the BBP distribution. []{data-label="F8"}](figure_tmer_org.eps){width="1.0\columnwidth"} -------------- ------------------------- ---------------- ---------------- --------------- binary distribution $A$ $B$ $\alpha$ IBP $\gamma_{\rm tot}$ 12.53$\pm$0.22 6.89$\pm$0.84 0.33$\pm$0.02 IBP $\gamma_{\rm pri}$ - 6.64$\pm$0.25 - IBP $\gamma_{\rm dyn}$ 12.30$\pm$0.44 - 0.33$\pm$0.01 IBP $\gamma_{\rm dyn,12}$ 303.1$\pm$6.1 - 0.43$\pm$0.02 BBP $\gamma_{\rm tot}$ 14.55$\pm$0.31 -0.03$\pm$0.03 0.38$\pm$0.01 BBP $\gamma_{\rm dyn,12}$ 527.7$\pm$11.1 - 0.52$\pm$0.02 binary distribution $a$ $b$ $c$ IBP $\mathcal{R_{\rm tot}}$ 13.01$\pm$3.00 4.14$\pm$0.19 0.35$\pm$0.04 IBP $\mathcal{R_{\rm pri}}$ 21.80$\pm$8.82 5.33$\pm$0.29 0.29$\pm$0.05 IBP $\mathcal{R_{\rm dyn}}$ 6.15$\pm$1.23 3.27$\pm$0.17 0.51$\pm$0.06 BBP $\mathcal{R_{\rm tot}}$ 7.96$\pm$2.03 3.54$\pm$0.21 0.42$\pm$0.06 \[T1\] -------------- ------------------------- ---------------- ---------------- --------------- : Best fitting results for empirical relations. “pri” and “dyn” denote the best-fit results using the primordial and dynamical BBH mergers separately. $\gamma_{\rm dyn,12}$ present the expected number of dynamical BBH mergers using current GC properties. Note that for the BBP distribution, the BBH mergers originating from GCs are mostly dynamical BBHs. In Fig. \[F8\], we present the time evolution of the merger rates for BBHs with different formation origins and find some differences between the primordial and dynamical BBH merger rates. Since the progenitor BBHs for primordial mergers form in a very short time interval during a GC’s early evolution ($T<30$ Myr), the merger rate decreases more rapidly than that for dynamical BBH mergers. The best-fit results are ($a$, $b$, $c$) $=$ (21.80$\pm$8.82, 5.33$\pm$0.29, 0.29$\pm$0.05) for primordial mergers and ($a$, $b$, $c$) $=$ (6.15$\pm$1.23, 3.27$\pm$0.17, 0.51$\pm$0.06) for dynamical mergers. For the BBP distribution, the best-fit parameters are ($a$, $b$, $c$) $=$ (7.96$\pm$2.03, 3.54$\pm$0.21, 0.42$\pm$0.06), similar to the parameters found for the dynamical BBH mergers in models with the IBP distribution. We summarize our results for the correlation between $N_{\rm merg}$ and GC properties and the time evolution of the merger rates in Table \[T1\]. Chirp mass and mass ratio of BBH mergers\[S3.3\] ------------------------------------------------ In this section we investigate some fundamental properties of the merging BBHs. Fig. \[F9\] shows the distribution of chirp masses, $\mathcal{M}_{\rm chirp}\equiv (m_1m_2)^{3/5}(m_1+m_2)^{-1/5}$ of merging BBHs. It is interesting to note that the chirp mass distribution for BBH mergers throughout all look-back time, $t_{\rm lb}$ and that for BBHs merging in the local Universe with red-shift, $z\leq0.2$ (i.e., $t_{\rm lb}\lesssim2.4$ Gyr with standard cosmological parameters assumed) are slightly different because the formation and merging timescales depend on the mass of BBHs [@2017ApJ...836L..26C; @2018PhRvL.120o1101R]. In Fig. \[F9\] we also plot the values of the chirp masses of BBH mergers detected by LIGO so far and show that they approximately fall within the range of values corresponding to the broad peak in the distribution of $\mathcal{M}_{\rm chirp}$ for merging BBHs. As illustrated by this figure, for the set of simulations considered in this paper the high-$\mathcal{M}_{\rm chirp}$ events are likely to belong to the dynamical BBH groups while the low-$\mathcal{M}_{\rm chirp}$ could be either primordial or dynamical BBHs. It is also possible that these low-$\mathcal{M}_{\rm chirp}$ sources come from metal-richer environments [see e.g. @2017ApJ...836L..26C; @2018arXiv180601285A]. ![Distribution of chirp masses, $\mathcal{M}_{\rm chirp}$ of merging BBHs escaping from GC simulation models. Black line shows the distribution of all BBH mergers from GC simulation models with the IBP distribution. Blue and red lines are for primordial and dynamical BBH mergers, respectively. Dotted line shows $\mathcal{M}_{\rm chirp}$ distribution of BBH mergers from models with the BBP distribution. Thin and thick lines represent $\mathcal{M}_{\rm chirp}$ distribution for the BBH mergers throughout all look-back time and those merging in the local Universe ($z\leq0.2$), respectively. Note that the model lines correspond to merger rate densities and are not corrected for observational selection effects. The $\mathcal{M}_{\rm chirp}$ of 5 GW events that have been detected by LIGO so far are marked with the range of 90 per cent confidence intervals (data from <https://losc.ligo.org/events/>). []{data-label="F9"}](figure_chmass.eps){width="1.0\columnwidth"} We point out that the distribution of $\mathcal{M}_{\rm chirp}$ of dynamical BBH mergers for the IBP distribution is almost identical to that for the BBP distribution (which has only dynamical BBH mergers). This implies that the host GC dynamics is the key factor determining the BBH mergers’ properties and differences between the IBP and the BBP initial properties of primordial binaries do not play an important role in the chirp mass of the BBH mergers. It is interesting to note that recent numerical studies [@2018PhRvL.120o1101R; @2018ApJ...855..124S] for GCs with post-Newtonian calculations for BBHs suggested that mergers of BBHs can occur inside GCs and the merger product can form a binary with other BHs and merge again in/outside of clusters [see also @2018arXiv180201192M for the retention of in-cluster BBH mergers]. The mass distribution of merging BBHs, especially for high masses, will be affected by this process. ![Distribution of BBH mergers for the IBP distribution in $q$-$\mathcal{M}_{\rm chirp}$ plane. Grey dots show all mergers and red and blue dots represent, respectively, dynamical and primordial BBHs that merge in the local Universe ($z\leq0.2$). Black stars and boxes indicate the ranges of $\mathcal{M}_{\rm chirp}$ and $q$ based on 90 per cent confidence intervals of LIGO detections. Upper panel shows the distribution of mass ratio, $q$, of primordial (blue) and dynamical (red) BBHs merging in the local Universe.[]{data-label="F10"}](figure_q_mch.eps){width="1.0\columnwidth"} Fig. \[F10\] shows the distribution of the $\mathcal{M}_{\rm chirp}$ versus the mass ratio $q$ ($\equiv m_{2}/m_{1}$, where $m_1>m_2$) of BBH mergers from models with the IBP distribution. The sequence of blue points corresponds to the primordial BBH mergers obtained with the binary stellar evolution and the mass fall-back mechanism used in the simulations (see Section \[S2\]). Three overdense regions in this plane can be easily identified at $\mathcal{M}_{\rm chirp}\sim20$ and $q\sim1$, $\mathcal{M}_{\rm chirp}\sim13$ and $q\sim0.55$, $\mathcal{M}_{\rm chirp}\sim10$ and $q\sim1$, respectively. This features are related to the shape of the mass function of BHs produced in our simulations, which has a bi-modal distribution with peaks at $m_{\rm BH}\sim12$ and 24 M$_{\odot}$. In the upper panel in Fig. \[F10\], the distribution of $q$ shows that dynamical BBH mergers tend to have similar masses [see also @2016MNRAS.458.3075A; @2016PhRvD..93h4029R; @2017MNRAS.469.4665P]. It is interesting to point out the presence of a sequence of BBH mergers with high-$\mathcal{M}_{\rm chirp}$ and low-$q$. This group comprises BHs that may have increased their mass due to mergers with other stars or black holes. In the latter case, it may be possible that some of the merger remnants may already have been ejected from the host stellar system due to gravitational wave recoil kicks [@2018PhRvL.120o1101R; @2018arXiv180201192M]. An important general point to emphasize is that the $\mathcal{M}_{\rm chirp}$ distribution of merging BBHs strongly depends on the metallicity [@2016PhRvD..93h4029R; @2017ApJ...836L..26C; @2017MNRAS.464L..36A; @2018MNRAS.474.2959G]. The metallicity affects not only the number of BHs produced, the number of BBH mergers but also the mass range of BBHs. However, @2017ApJ...836L..26C pointed out that the $\mathcal{M}_{\rm chirp}$ distribution of BBHs formed dynamically and merging in the local Universe ($z\leq0.2$) does not depend on metallicity for $Z\leq0.001$. We note that the masses of BHs depend also on the single and binary stellar evolution recipes, [^4] and, in particular, on the fall-back prescription [e.g. @2012ApJ...749...91F; @2015MNRAS.451.4086S]. Additional observations and numerical simulations are therefore needed to constrain the values of the BH masses after supernova explosions. Merger rate density\[S4\] ========================= In this section we estimate the merger rate density of BBHs escaping from the GCs using the empirical relations obtained in the previous sections. Since the relation between the expected number, the mass distribution and the time evolution of the rates are different for primordial and dynamical BBH mergers, we calculate the merger rates separately for BBH mergers with different origins. Rate density for primordial BBH mergers\[S4.1\] ----------------------------------------------- To estimate the merger rate density, we follow the calculation of @2017MNRAS.464L..36A . For this calculation, we use the merging time and the chirp mass for each primordial BBH that will merge within 12 Gyr from all the simulation models. Having this data and the total and average initial mass of all simulated GCs, we can estimate the merger rate density per unit chirp mass using a GC star formation rate as a function of redshift and the contribution of the merger rate from individual GCs to the rate density according to the age distribution of GCs based on the GC star formation history. For this purpose, the GC star formation rate estimated by @2013MNRAS.432.3250K has been adopted in this calculation. We already pointed out that the number of primordial BBH mergers depends on the initial mass and binary fraction of GCs. If we simply assume that the initial binary fraction is universal for all GCs, the merger rate density for primordial BBHs only depends on the GC formation rate. The number of primordial BBH mergers over 12 Gyr based on the IBP distribution is $\sim$6.64$f_{\rm b,0}$ per $10^5$M$_{\odot}$ from Eq. (\[E3\]) and its best-fit parameters. On the other hand the number of primordial BBH mergers from the simulations with the BBP distribution is very small and we estimate the contribution of primordial BBH mergers to be negligible in this case. Our estimate of the number of mergers per unit mass is consistent with that in @2018MNRAS.474.2959G with similar metallicity. From this number it follows that the local merger rate density of primordial BBHs ranges from 0.18 to 1.8 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for an initial binary fraction ranging from 10 to 100 per cent. Rate density for dynamical BBH mergers from initial GC properties\[S4.2\] ------------------------------------------------------------------------- In order to estimate the merger rate density for dynamical BBH mergers from our empirical relations, we first need to calculate the number of BBH mergers per unit GC mass. For the calculation of the number of dynamical BBH mergers we need to make an assumption on the initial GC mass and size distribution and then combine these with our estimate of the number of dynamical BBH mergers from Eq. (\[E2\]). For the initial GC mass function (ICMF), we adopt a @1976ApJ...203..297S function $$dN_{\rm GC}\propto M^{-\beta}{\rm exp}(-M/M_{*})dM$$ where $\beta=2$ and $M_{*}$ is the exponential cut-off mass of the ICMF. We consider different combinations of the minimum mass of GCs, $M_{\rm min}=10^3, 10^4$M$_{\odot}$ and exponential cut-off mass $M_{*}=10^6, 10^{6.5}$M$_{\odot}$ [for the selection of $M_{*}$, see e.g. @2017ApJ...839...78J]. No firm prediction on the distribution of the initial sizes of GCs is currently available. Instead, we tried to find a realistic distribution of the initial size of GCs from the observations of young massive clusters (YMCs) in extra-galactic systems although it is possible that old GCs forming in the early Universe formed with a different size distribution. There are a number of observational studies [e.g. @2010ApJ...709..411H; @2012MNRAS.419.2606B; @2015MNRAS.452..525R] showing that the effective radii of YMCs tend to increase with the YMC’s age. This might be due to the combined effects of the primordial gas expulsion, initial mass loss by the stellar evolution and/or the presence of a significant number of retained BHs [e.g. @2008MNRAS.386...65M]. By correcting the age dependence of the effective radii of YMCs in M83 [@2015MNRAS.452..525R], we obtain a log-normal distribution of the initial half-mass radius with $\sigma=0.4$ and $\left<r_{\rm h}\right>=2.8$ pc which is comparable with the initial half-mass radii used in the numerical simulations by @2010ApJ...719..915C [@2013MNRAS.429.2881C] reproducing the distribution of GGCs. Many studies of YMCs found that there is a weak correlation between the mass and the effective radius of YMCs . We take the relation for the average value of the initial half-mass radius, $\left<r_{\rm h}\right>/{\rm pc}=2.8\times(M/10^4{\rm M}_{\odot})^{0.1}$ from . In order to investigate the effects of the initial size distribution on the merger rate density, we consider another distribution of the initial half-mass radius, $\left<r_{\rm h}\right>/{\rm pc}=0.33\times(M/10^4{\rm M}_{\odot})^{0.13}$ from , which is much smaller that the previous one [note that these “small” and “large” size distributions are roughly consistent with the half-mass radii for massive clusters and open clusters/associations from the simulations for the cluster formation done by @2016ApJ...817....4F]. We, however, ignore the effects of the host galaxy tidal field on the initial distribution of half-mass radii since there is no correlation between the effective radii and galactocentric distances of YMCs found in nearby galaxies [@2017ApJ...841...92R]. @2012ApJ...756..167M also showed that the galactocentric distance does not significantly affect the early (less than a few hundreds Myr) evolution of half-mass radii of star clusters. Using the initial mass and size distributions discussed above, we can estimate the expected number of dynamical BBH mergers per GC masses through Eq. (\[E2\]) and its fitting parameters as $$\frac{N_{\rm merg}}{M_{\rm GCSF}}=\frac{\iint \gamma_{\rm dyn} N(M)N(r_{\rm h})dMdr_{\rm h}}{\int N(M)MdM},$$ where $M_{\rm GCSF}$ is the total mass of all GCs, and $N(M)$ and $N(r_{\rm h})$ are, respectively, the mass and half-mass radius distribution of initial GC systems (and in which, as explained above, the mean of the half-mass radius distribution depends on the cluster mass). For the different ICMF we have considered, we find that the total number of dynamical BBHs mergers per unit mass over 12 Gyr based on the “large” size distribution is $\sim$2.45 (2.34) per $10^5$M$_{\odot}$ for \[$M_{\rm min}, M_{*}$\] = \[$10^4$M$_{\odot}, 10^{6.5}$M$_{\odot}$\], $\sim$2.11 (1.96) for \[$10^4$M$_{\odot}, 10^{6}$M$_{\odot}$\], $\sim$1.96 (1.82) for \[$10^3$M$_{\odot}, 10^{6.5}$M$_{\odot}$\], and $\sim$1.69 (1.53) for \[$10^3$M$_{\odot}, 10^{6}$M$_{\odot}$\], respectively for the simulation models with the IBP (BBP) distribution. From these estimates, the corresponding local merger rate densities for the different ICMF are 1.91 (1.73), 1.64 (1.45), 1.52 (1.34) and 1.31 (1.13) ${\rm Gpc}^{-3}{\rm yr}^{-1}$, respectively. We point out that the ratio of the local merger rate density to the number of mergers per GC masses for dynamical BBHs is larger than that for primordial BBHs due to the chirp mass distribution and time evolution of the merger rates (see Figs. \[F8\] and \[F9\]). We also emphasize that the local merger rate density for dynamically-formed BBHs does not show any significant dependence on the binary distributions and weakly dependent on the ICMF with a variation for the different ICMFs considered of a factor of $\lesssim2$. Along with the local merger rate density for primordial BBHs, our calculation of local merger rate density of $\sim$4 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ is consistent with that of $\sim$5 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ from other literature [e.g. @2016PhRvD..93h4029R; @2017MNRAS.464L..36A]. Some discrepancies may be due to the different distribution of GC models. For more compact initial size distribution of GCs suggested by , we obtain the merger rate density of 14.3 (17.5), 12.6 (15.1), 11.7 (13.9) and 10.3 (12.0) ${\rm Gpc}^{-3}{\rm yr}^{-1}$, respectively for GC models with the IBP (BBP) distribution with the different ICMF. The merger rate density is larger for the BBP distribution because the expected number of dynamical BBH mergers has stronger correlation with the initial density of GCs. The many more additional detections of BBH mergers will be needed to shed light on the initial binary distribution in GCs as well as the distribution of the initial properties of GCs. We point out that the mass distribution and the merger rate are nearly independent of the metallicity for $Z\leq0.001$ for either primordial [@2018MNRAS.474.2959G] or dynamical [@2017ApJ...836L..26C] BBH mergers. By combining the GC star formation history from @2013MNRAS.432.3250K and the red-shift metallicity relation from @2016Natur.534..512B, we expect that approximately $\lesssim$10% of GCs especially forming at lower red-shift ($z\sim2$–3) will be affected by the effects of metallicity. Although we fixed the metallicity to $Z=0.001$, there will not be significant effects of cosmological metallicity variation on the estimation of the local merger rate density. Rate density for dynamical BBH mergers from current GC properties\[S4.3\] ------------------------------------------------------------------------- In Section \[S3.1\], we discussed the correlation between the expected number of dynamical BBH mergers and the current GCs’ mass and half-mass density. Combining the correlation and the distribution of observed GCs, we also can estimate the merger rate density for BBH mergers originating from surviving GCs. To reproduce the mass distribution of GGCs, we generated the GCMF following an evolved Schechter function [@2007ApJS..171..101J], $$\frac{dN}{dM}\propto \frac{1}{(M+\Delta)^2}{\rm exp} \left( -\frac{M+\Delta}{M_{\rm c}}\right),$$ where $\Delta$ is a factor for the mass loss of GCs, and $M_{\rm c}$ is the exponential cut-off mass for the GCMF. We adopted the values of $\Delta=10^{5.4}$M$_{\odot}$ and $M_{\rm c}=10^{5.9}$M$_{\odot}$ from @2007ApJS..171..101J. ![Present-day distribution of GCs in the half-mass density and the mass of GCs. Red dots are the Milky Way GCs from the @1996AJ....112.1487H GGC catalog assuming that the mass-to-light ratio for all GCs is equal to 2 and the half-mass radius is equal to $\sim$1.7 of the projected half-light radius in the catalog. Grey dots are modeled GCs following an evolved Schechter function GCMF [@2007ApJS..171..101J] and a log-normal distribution with $\left<r_{\rm h}\right>=6.1$ pc and $\sigma=0.63$ for the distribution of the half-mass radius. Dashed lines indicate the expected number of dynamical BBH mergers produced in the individual GCs over 12 Gyr lifetime. []{data-label="F11"}](figure_gcdist.eps){width="1.0\columnwidth"} In Fig. \[F11\], we show the distribution of GGCs from @1996AJ....112.1487H catalog in the $M_{\rm GC}$-$\rho_{\rm h}$ plane. We simply assumed that the mass-to-light ratio $\Upsilon=2$ and the half-mass radius $r_{\rm h}= \sim 1.7 R_{\rm hl}$ [where $R_{\rm hl}$ is the projected half-light radius, see @2013MNRAS.429.2881C] for all GGCs. We then distributed the half-mass radius of GCs at present-day by using a log-normal distribution with parameters $\left<r_{\rm h}\right>=6.1$ pc and $\sigma=0.63$ which give a best-fit with the distribution of @1996AJ....112.1487H GGCs in the $M_{\rm GC}$-$\rho_{\rm h}$ plane. Dashed lines in this figure indicate the expected number of dynamical BBH mergers generated from the individual GCs over 12 Gyr cluster lifetime. Many of GGCs are expected to produce between $\sim$10 and $\sim$ 1000 BBH mergers within 12 Gyr. [c c c c c c]{}\ $f_{\rm b,0}$ &\ 10% &\ 100% &\ \ binary & ICMF &\ distribution & \[$M_{\rm min}, M_{*}$\] & &\ IBP & \[$10^4, 10^{6.5}$\] M$_{\odot}$ & &\ IBP & \[$10^4, 10^{6}$\] M$_{\odot}$ & &\ IBP & \[$10^3, 10^{6.5}$\] M$_{\odot}$ & &\ IBP & \[$10^3, 10^{6}$\] M$_{\odot}$ & &\ BBP & \[$10^4, 10^{6.5}$\] M$_{\odot}$ & &\ BBP & \[$10^4, 10^{6}$\] M$_{\odot}$ & &\ BBP & \[$10^3, 10^{6.5}$\] M$_{\odot}$ & &\ BBP & \[$10^3, 10^{6}$\] M$_{\odot}$ & &\ \ binary & GCMF &\ distribution & & 0.33 & 0.77 & 2.31 & F2017\ IBP & eSchechter & 0.63 & 1.46 & 4.39 & 5.16\ IBP & log-normal & 0.89 & 2.02 & 6.20 & 7.28\ BBP & eSchechter & 0.84 & 1.96 & 5.88 & 6.77\ BBP & log-normal & 1.15 & 2.68 & 8.04 & 9.26\ \[T2\] To compute the local merger rate density from the expected number of BBH mergers from current GCs properties, $\gamma_{\rm dyn,12}$ from Eq. (\[E2\]) and the distributions of GCs’ present-day properties introduced above, we need the number density of GCs, $\rho_{\rm GC}$ in the local Universe. We simply adopt 0.33, 0.77 and 2.31 Mpc$^{-3}$ [@2016PhRvD..93h4029R] for conservative, standard and optimistic assumptions of $\rho_{\rm GC}$, respectively. We obtain the local merger rate densities for dynamical BBH mergers are 0.63, 1.46 and 4.39 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for conservative, standard and optimistic cases, respectively, assuming the age of all GCs is 12 Gyr. The merger rate at the present-day has been corrected by using the Eq. (\[E4\]) for the case of dynamical BBH mergers (i.e. $\mathcal{R} \sim 0.24 \left<\mathcal{R}\right>$). We have also considered the time-dependent $\rho_{\rm GC}$ from @2017PASJ...69...94F for old (T $\ge$ 10 Gyr) GCs. The total $\rho_{\rm GC}$ is slightly smaller (2.2 Mpc$^{-3}$) than the optimistic case of @2016PhRvD..93h4029R. However, we obtain the merger rate density of 5.16 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ which is slightly larger than our optimistic case based on $\rho_{\rm GC}$ from @2016PhRvD..93h4029R because there are younger GCs with higher merger rates compared to GCs with ages of 12 Gyr (see Figs. \[F7\] and \[F8\]). Our estimates are similar in order of magnitude but systematically smaller than those from other studies such as $\sim$5 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ from @2016PhRvD..93h4029R (for the standard case) and 13 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ from @2017PASJ...69...94F [see also @2017MNRAS.464L..36A; @2017MNRAS.469.4665P] since our GCMF includes a larger number of GCs with lower masses which contribute less to the merger rate density compared to the more massive GCs. Using the same GCMF adopted in @2016PhRvD..93h4029R, we obtain the merger rate density of 0.89, 2.02, 6.20 and 7.28 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for conservative, standard, optimistic and time-dependent $\rho_{\rm GC}$. On the other hand, as discussed in previous sections, most of BBH mergers based on the BBP distribution are dynamical mergers. We estimate the merger rate density of 0.84, 1.96, 5.88 and 6.77 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ for conservative, standard, optimistic and time-dependent $\rho_{\rm GC}$, respectively, using the evolved Schechter function GCMF. The estimate of the merger rate density becomes 1.15, 2.68, 8.04 and 9.26 assuming that the GCMF follows a log-normal distribution as used in @2016PhRvD..93h4029R. We summarize our estimates of the local merger rate density in Table \[T2\]. We point out that the calculation of the local merger rate density based on the cluster current properties includes only the contribution of surviving clusters. GCs dissolving before 12 Gyr of course can contribute to the population of BBH mergers and to take their contribution into account, a calculation like that presented in the previous section must be carried out. Alternatively we can use a simple toy model for the evolution of a globular cluster system and assume the GC disruption proceeding from the low-mass GCs; using this simple model we can calculate the cumulative fraction of BBH mergers from surviving GCs as a function of the fraction of surviving GCs from our GC models introduced in Section \[S4.2\]. We show the result of this calculation in Fig. \[F12\]. This figure provides an approximate estimate of the fraction of BBH mergers from GCs that still survive at the present-day. If we assume that only $\sim$3 per cent of GCs survive up to now as suggested by @2014ApJ...785...71G for the Milky Way, the fraction of BBH mergers from surviving GCs varies from 0.3 to 0.7 [0.5 with $M_{\rm min}=10^4 {\rm M}_{\odot}$ used in @2014ApJ...785...71G] depending on the ICMF. The local merger rates based on the current GC properties become comparable with those based on the initial GC properties with “small” size distribution when the contribution of merging BBHs from dissolving GCs is taken into account. ![Cumulative fraction of dynamical BBH mergers from surviving GCs as a function of the fraction of surviving GCs. We simply assume that GCs are disrupted from lower-mass GCs. Different lines present the ICMF with different parameters.[]{data-label="F12"}](figure_gcdisolv.eps){width="1.0\columnwidth"} We conclude this section by pointing out that it is possible that YMCs can contribute to the local merger rate density. @2017MNRAS.467..524B [@2018MNRAS.473..909B] have performed direct $N$-body simulations for YMC-like systems with post-Newtonian approximation implemented and found that YMCs can contribute the detection rate to a similar extent as more massive GC counterpart. @2014MNRAS.441.3703Z have estimated the local merger rate density of 3.6 Gpc$^{-3}$yr$^{-1}$ for BBHs originating from YMCs in the local Universe. @2017PASJ...69...94F also have suggested from their direct $N$-body simulations that the local merger rate density can be up to a factor of $\sim$3 times larger when younger clusters with ages between 2 and 10 Gyr are included in the estimation of the merger rate density. Using the empirical relation for the time evolution of the merger rates in Eq. (\[E4\]) and the GC formation rate adopted by @2017PASJ...69...94F, we obtain a local merger rate density about $\sim$6 times higher when younger GCs are included. When younger clusters are included, an important aspect to consider is the well-known age-metallicity relation for GCs [e.g. @2013MNRAS.436..122L]. @2017PASJ...69...94F considered the effects of the metallicity by limiting the mass of BHs and found that there is no significant effect on the local merger rates. This is, however, in contrast with the findings of @2018MNRAS.474.2959G who suggested that the number of BBH mergers per unit mass strongly depends on the metallicity. The study of @2018MNRAS.474.2959G is focused on primordial BBHs, but in the dense environment like GCs, the internal dynamics can in part compensate the effects of the metallicity [see e.g. @2017MNRAS.464L..36A; @2017ApJ...836L..26C]. Finally in this study, we did not consider the contribution to the merger rate density by the BBHs that merge inside GCs through the dynamical interactions and binary evolution. However, these in-cluster mergers become more important for very young clusters [@2017MNRAS.464L..36A; @2017MNRAS.467..524B; @2018MNRAS.473..909B]. According to @2017MNRAS.464L..36A, the contribution of these in-cluster mergers is about 20 per cent of the total merger rate through the entire evolution however becomes less than 1 per cent if the host GCs are old (T $>$ 10 Gyr). It is also important to note that the rate of in-cluster mergers can increase substantially if the dissipative effects connected with GW radiation (i.e. three-body GW capture) are taken into account [@2018ApJ...855..124S]. summary and conclusions\[S5\] ============================= In this paper we have studied the formation of binary black holes (BBHs) in globular clusters (GCs) and explored the relation between the number and properties of merging BBHs and the structural properties of their host GCs. Our study is based on a large survey of Monte Carlo simulations following the dynamical evolution of GCs with a broad range of different initial masses, sizes and primordial binary properties. Our results have revealed a close correlation between the number of BBH mergers escaping from GCs and the properties of host GCs such as the initial mass, half-mass radius and the fraction of primordial binaries (Figs. \[F1\] and \[F4\]). We identified two groups of BBH mergers; one group is composed of primordial BBH mergers forming simply as a result of binary stellar evolution and escaping from GCs due to the natal kicks by supernova explosions. The second group is composed of dynamical BBH mergers forming as a result of binary-binary and binary-single interactions in the GC dense environments and ejected from GCs through the dynamical interactions. The number of primordial BBH mergers is correlated with the GC’s initial mass and binary fraction (see Eq. \[E3\]), while we found that the number of dynamical BBH mergers produced in 12 Gyr is correlated with a parameter $\gamma_{\rm dyn}$ (see Eq. \[E2\]) depending on the GC’s initial mass and half-mass density (Figs. \[F2\] and \[F3\]). Interestingly we have shown that the number of dynamical BBH mergers correlates also with the same $\gamma_{\rm dyn}$ parameter but defined in terms of the GC’s current properties (Figs. \[F5\] and \[F6\]). We provide analytic expressions describing the correlations between the number of BBH mergers and the host GC’s properties and apply them to estimate the BBH merger rate for a few different models of GC populations but the expression provided in our study can be used more in general for GC populations with initial conditions different from those adopted in our calculations. The specific properties of primordial and dynamical BBH mergers such as the merging time and the chirp mass distribution are very important for the estimate of the local merger rate and the detection rate. In general, we find that the merger rate decreases with time due to the continuous ejection of single and binary BHs from GCs (Fig. \[F7\]). We showed that the time evolution of the merger rate for primordial BBH mergers decreases more rapidly than that for dynamical BBH mergers; this difference is due to differences between the formation and ejection timescales of the two groups of BBH mergers (Fig. \[F8\]). The two groups of BBH mergers are characterized also by differences in the chirp masses. The dynamical BBH mergers contribute more massive BBH mergers compared to the primordial BBH mergers (Figs. \[F9\] and \[F10\]). Based on the analytic expressions obtained from study, we estimated the local merger rates of BBHs escaping from GCs. The local merger rate for primordial BBHs depends only on the cosmological GC formation rate and we obtained a rate of 0.18–1.8 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ (Section \[S4.1\]) depending on the primordial binary fraction. To estimate the local merger rate for dynamical BBHs, on the other hand, it is necessary make an assumption on the initial distribution of GC masses and size. As pointed out above, the analytic expressions obtained in this paper allow to calculate the local merger rate for any assumption concerning these initial distributions. We estimated a local merger rate for dynamical BBHs of 1.3–18 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ depending on a variety of combinations of the initial GC mass function and size distribution (Section \[S4.2\]). We also estimated a local rate for dynamical BBH mergers from the current properties of surviving GCs equal to 0.6–9.3 ${\rm Gpc}^{-3}{\rm yr}^{-1}$ (Section \[S4.3\]; see also Table \[T2\]), assuming all GCs have the same age and metallicity. The production of BBH mergers from GCs also can be influenced by the formation and the presence of intermediate mass black holes (IMBHs) in GCs. @2015MNRAS.454.3150G suggested that a seed BH for an IMBH can be formed by the runaway collisions of massive main-sequence (MS) stars [see also @2002ApJ...576..899P; @2017MNRAS.472.1677S]. This process will preferentially deplete the massive MS progenitors for stellar-mass BHs. Moreover, @2007MNRAS.374..857T have found that hard binaries can be disrupted by the interactions with the IMBH. These interactions between the IMBH and BBHs might result in the capture of one BH to the IMBH and the ejection of the companion BH [this IMBH-BH binary can deplete the stellar-mass BH population by ejection; see @2014MNRAS.444...29L], which is the possible source of intermediate mass ratio inspirals (IMRIs) for space-based GW detectors [e.g. @2002ApJ...581..438M; @2009ApJ...698L.129S]. Detailed investigations for the effects of the formation of IMBHs in GCs on the merger rate of stellar-mass BBHs will be studied in our forthcoming papers. acknowledgement {#acknowledgement .unnumbered} =============== We thank an anonymous referee whose suggestions helped to improve this manuscript. JH acknowledges support from the China Postdoctoral Science Foundation, Grant No. 2017M610694. AA was partially supported by the National Science Center (NCN), Poland, through the grant UMO-2015/17/N/ST9/02573. MG and AA were partially supported by NCN, Poland, through the grant UMO-2016/23/B/ST9/02732 and is currently supported by the Carl Tryggers Foundation through the grant CTS 17:113. TB was supported by the grant TEAM/2016-3/19 from the Foundation for Polish Science (FNP). This research was supported in part by Lilly Endowment, Inc., through its support for the Indiana University Pervasive Technology Institute, and in part by the Indiana METACyt Initiative. The Indiana METACyt Initiative at IU is also supported in part by Lilly Endowment, Inc. This work benefited from support by the International Space Science Institute (ISSI), Bern, Switzerland, through its International Team programme ref. no. 393 *The Evolution of Rich Stellar Populations & BH Binaries* (2017-18). Aarseth S. J. 2012, [MNRAS]{}, 422, 841 Abbott B. P. et al., 2016a, [Phys. Rev. Lett.]{}, 116, 061102 Abbott B. P. et al., 2016b, [Phys. Rev. Lett.]{}, 116, 241103 Abbott B. P. et al., 2016c, Phys. Rev. X, 6, 041015 Abbott B. P. et al., 2017a, [Phys. Rev. Lett.]{}, 118, 221101 Abbott B. P. et al., 2017b, [Phys. Rev. Lett.]{}, 119, 141101 Abbott B. P. et al., 2017c, [Phys. Rev. Lett.]{}, 119, 161101 Amaro-Seoane P., Chen X. 2016, [MNRAS]{}, 458, 3075 Antonini F., Perets H. B. 2012, [ApJ]{}, 757, 27 Antonini F., Murray N., Mikkola S. 2014, [ApJ]{}, 781, 45 Antonini F. et al., 2016, [ApJ]{}, 816, 65 Arca Sedda M., Askar A., Giersz M., 2018, MNRAS, 479, 4652 Arca Sedda M., Benacquista M. 2018, arXiv:1806.01285 Askar A., Szkudlarek M., Gondek-Rosi[ń]{}ska D., Giersz M., Bulik T., 2017, [MNRAS]{}, 464, L36 Askar A., Arca Sedda M., Giersz M. 2018, [MNRAS]{}, 478, 1844 Avni Y., 1976, [ApJ]{}, 210, 642 Bae Y.-B., Kim C., Lee H. M. 2014, [MNRAS]{}, 440, 2714 Bae Y.-B., Lee H. M., Kang G., Hansen J. 2017, [Phys. Rev. D]{}, 96, 084009 Banerjee S., 2017, [MNRAS]{}, 467, 524 Banerjee S., 2018, [MNRAS]{}, 473, 909 Banerjee S., Baumgardt H., Kroupa P. 2010, [MNRAS]{}, 402, 371 Bastian N. et al., 2012, [MNRAS]{}, 419, 2606 Bate M. R., 2000, [MNRAS]{}, 314, 33 Bate M. R., Bonnell I. A., Bromm V., 2002, [MNRAS]{}, 336, 705 Belczynski K., Kalogera V., Bulik T., 2002, [ApJ]{}, 572, 407 Belczynski K., Taam R. E., Kalogera V., Rasio F. A., Bulik T. 2007, [ApJ]{}, 662, 504 Belczynski K., Holz D. E., Bulik T., O’Shaughnessy R. 2016, Nature, 534, 512 Belczynski K. et al., 2017, arXiv:1706.07053 Belczynski K., et al., 2018, A&A, 615, 91 Belloni D. et al., 2017a, [MNRAS]{}, 468, 2429 Belloni D., Askar A., Giersz M., Kroupa P., Rocha-Pinto H. J., 2017b, [MNRAS]{}, 471, 2812 Belloni D., Kroupa P., Rocha-Pinto H. J., Giersz M., 2018, [MNRAS]{}, 474, 3740 Bik A., Lamers H. J. G. L. M., Bastian N., Panagia N., Romaniello M., 2003, [A&A]{}, 397, 473 Breen P. G., Heggie D. C. 2013, [MNRAS]{}, 432, 2779 Bulik T., Belczy[ń]{}ski K., Rudak B., 2004, [A&A]{}, 415, 407 Chatterjee S., Rodriguez C. L., Rasio F. A., 2017, [ApJ]{}, 834, 68 Chatterjee S., Fregeau J. M., Umbreit S., Rasio F. A. 2010, [ApJ]{}, 719, 915 Chatterjee S., Umbreit S., Fregeau J. M., Rasio F. A. 2013, [MNRAS]{}, 429, 2881 Chatterjee S., Rodriguez C. L., Kalogera V., Rasio F. A. 2017, [ApJ]{}, 836, L26 Djorgovski S., Meylan G. 1994, [AJ]{}, 108, 1292 Dominik M. et al., 2012, [ApJ]{}, 759, 52 Downing J. M. B., Benacquista M. J., Giersz M., Spurzem R. 2011, [MNRAS]{}, 416, 133 Duquennoy, A., Mayor, M. 1991, [A&A]{}, 248, 485 Fregeau J. M., Cheung P., Portegies Zwart S. F., Rasio F. A., 2004, [MNRAS]{}, 352, 1 Fryer C. L. et al., 2012, [ApJ]{}, 749, 91 Fujii M. S., Portegies Zwart S., 2016, [ApJ]{}, 817, 4 Fujii M. S., Tanikawa A., Makino J., 2017, [PASJ]{}, 69, 94 Fukushige T., Heggie D. C., 2000, [MNRAS]{}, 318, 753 Giacobbo N., Mapelli M., Spera M., 2018, [MNRAS]{}, 474, 2959 Giacobbo N., Mapelli M. 2018, arXiv:1806.00001 Giersz M., Heggie D. C., Hurley J. R., Hypki A., 2013, [MNRAS]{}, 431, 2184 Giersz M., Leigh N., Hypki A., L[ü]{}tzgendorf N., Askar A., 2015, [MNRAS]{}, 454, 3150 Gnedin O. Y., Ostriker J. P., Tremaine S., 2014, [ApJ]{}, 785, 71 Gond[á]{}n L., Kocsis B., Raffai P., Frei Z. 2018, [ApJ]{}, 860, 5 Harris W. E. 1996, [AJ]{}, 112, 1487 Hoang B.-M., Naoz S., Kocsis B., Rasio F. A., Dosopoulou F. 2018, [ApJ]{}, 856, 140 Hong J., Lee H. M., 2015, [MNRAS]{}, 448, 754 Hong J., Vesperini E., Belloni D., Giersz M., 2017, [MNRAS]{}, 464, 2511 Hunter D. A., Elmegreen B. G., Dupuy T. J., Mortonson M., 2003, [AJ]{}, 126, 1836 Hurley J. R., Pols O. R., Tout C. A., 2000, [MNRAS]{}, 315, 543 Hurley J. R., Tout C. A., Pols O. R., 2002, [MNRAS]{}, 329, 897 Hwang N., Lee M. G., 2010, [ApJ]{}, 709, 411 Hypki A., Giersz M., 2013, [MNRAS]{}, 429, 1221 Johnson L. C. et al., 2017, [ApJ]{}, 839, 78 Jord[á]{}n A. et al., 2007, [ApJS]{}, 171, 101 Katz H., Ricotti M., 2013, [MNRAS]{}, 432, 3250 King I. R., 1966, [AJ]{}, 71, 64 Kremer K., Ye C. S., Chatterjee S., Rodriguez C. L., Rasio F. A. 2018, [ApJ]{}, 855, L15 Kroupa P., 1995, [MNRAS]{}, 277, 1507 Kroupa P., 2001, [MNRAS]{}, 322, 231 Kroupa P. et al., 2013, Planets, Stars and Stellar Systems. Volume 5: Galactic Structure and Stellar Populations, 5, 115 Krumholz M. R., Thompson T. A., 2007, [ApJ]{}, 661, 1034 Larsen S. S., 2004, [A&A]{}, 416, 537 Leaman R., VandenBerg D. A., Mendel J. T., 2013, [MNRAS]{}, 436, 122 Leigh N. W. C. et al., 2014, [MNRAS]{}, 444, 29 Leigh N. W. C. et al., 2018, [MNRAS]{}, 474, 5672 Li, C., Hong, J. 2018, [MNRAS]{}, 476, 5274 Mackey A. D., Wilkinson M. I., Davies M. B., Gilmore G. F., 2008, [MNRAS]{}, 386, 65 Madrid J. P., Hurley J. R., Sippel A. C., 2012, [ApJ]{}, 756, 167 Marks M., Kroupa P., 2012, [A&A]{}, 543, A8 McLaughlin D. E., Fall S. M., 2008, [ApJ]{}, 679, 1272-1287 Miller M. C., 2002, [ApJ]{}, 581, 438 Morawski J., Giersz M., Askar A., Belczynski K., 2018, arXiv:1802.01192 Morscher M., Umbreit S., Farr W. M., Rasio F. A. 2013, [ApJ]{}, 763, L15 Morscher M., Pattabiraman B., Rodriguez C., Rasio F. A., Umbreit S. 2015, [ApJ]{}, 800, 9 O’Leary R. M., Rasio F. A., Fregeau J. M., Ivanova N., O’Shaughnessy R. 2006, [ApJ]{}, 637, 937 O’Leary R. M., Kocsis B., Loeb A. 2009, [MNRAS]{}, 395, 2127 Park D., Kim C., Lee H. M., Bae Y.-B., Belczynski K., 2017, [MNRAS]{}, 469, 4665 Peters P. C., 1964, Phys. Rev., 136, 1224 Portegies Zwart S. F., McMillan S. L. W. 2000, [ApJ]{}, 528, L17 Portegies Zwart S. F., McMillan S. L. W., 2002, [ApJ]{}, 576, 899 Rodriguez C. L., Antonini F. 2018, arXiv:1805.08212 Rodriguez C. L., Chatterjee S., Rasio F. A., 2016, [Phys. Rev. D]{}, 93, 084029 Rodriguez C. L. et al., 2015, [Phys. Rev. Lett.]{}, 115, 051101 Rodriguez C. L., Haster C.-J., Chatterjee S., Kalogera V., Rasio F. A. 2016a, [ApJ]{}, 824, L8 Rodriguez C. L. et al., 2016b, [MNRAS]{}, 463, 2109 Rodriguez C. L., Amaro-Seoane P., Chatterjee S., Rasio F. A. 2018, [Phys. Rev. Lett.]{}, 120, 151101 Ryon J. E. et al., 2015, [MNRAS]{}, 452, 525 Ryon J. E. et al., 2017, [ApJ]{}, 841, 92 Sakurai Y., Yoshida N., Fujii M. S., Hirano S., 2017, [MNRAS]{}, 472, 1677 Samsing J. 2018, [Phys. Rev. D]{}, 97, 103014 Samsing J., Askar A., Giersz M. 2018, [ApJ]{}, 855, 124 Sana H. et al., 2012, Science, 337, 444 Schechter P., 1976, [ApJ]{}, 203, 297 Sesana A., Gair J., Mandel I., Vecchio A., 2009, [ApJ]{}, 698, L129 Spera M., Mapelli M., Bressan A., 2015, [MNRAS]{}, 451, 4086 Tanikawa A. 2013, [MNRAS]{}, 435, 1358 Trenti M., Ardi E., Mineshige S., Hut P., 2007, [MNRAS]{}, 374, 857 Wang L. et al., 2016, [MNRAS]{}, 458, 1450 Weatherford N. C., Chatterjee S., Rodriguez C. L., & Rasio F. A. 2017, arXiv:1712.03979 Zepf S. E., Ashman K. M., English J., Freeman K. C., Sharples R. M., 1999, [AJ]{}, 118, 752 Zhang Q., Fall S. M., 1999, [ApJ]{}, 527, L81 Zhao B., Li Z.-Y., 2013, [ApJ]{}, 763, 7 Ziosi B. M., Mapelli M., Branchesi M., Tormen G., 2014, [MNRAS]{}, 441, 3703 [^1]: E-mail: [email protected] (JH) [^2]: This [*pre-main-sequence eigenevolution*]{} [@1995MNRAS.277.1507K] was originally postulated to explain the observed properties of Galactic field binary populations originating in embedded clusters. Most recently, @2017MNRAS.471.2812B [@2018MNRAS.474.3740B] have provided a modified prescription of the Kroupa IBP for the binary distributions in GCs, which is, however, not applied in this study. [^3]: However, @2017MNRAS.471.2812B [@2018MNRAS.474.3740B] provided a modified prescription of the Kroupa IBP distribution for GC environments suggesting that the pre-main-sequence eigenevolution and mass feeding algorithm are not applied to massive binaries and that the pairing rule for massive binaries is a uniform pairing based on @2012Sci...337..444S. If this is the case, the number of primordial BBH mergers from GCs will be negligible and the relation between the number of BBH mergers and GC properties will be similar to that for the BBP distribution in our study. [^4]: Note that the common-envelope phase (CEP) is also important for the binary stellar evolution and the formation of compact binaries [e.g. @2017MNRAS.468.2429B; @2018arXiv180600001G]. In this study we used the CEP parameters, $\alpha_{\rm CE}=3$ and $\lambda=0.5$. However, recent studies [e.g. @2017MNRAS.468.2429B] suggested $\alpha_{\rm CE}\sim 0.5$, and lower $\alpha_{\rm CE}$ and $\lambda$ value may lead to more binary mergers during the CEP and the subsequent production of single BHs. The uncertainty in the value of these parameters may affect the number of primordial BBHs and the mass distribution of merging BBHs from GCs.

ArXiv

--- abstract: 'We introduce a $q,t$-enumeration of Dyck paths which are forced to touch the main diagonal at specific points and forbidden to touch elsewhere and conjecture that it describes the action of the Macdonald theory $\nabla$ operator applied to a Hall-Littlewood polynomial. Our conjecture refines several earlier conjectures concerning the space of diagonal harmonics including the “shuffle conjecture" (Duke J. Math. $\mathbf {126}$ ($2005$), pp. $195-232$) for $\nabla e_n[X]$. We bring to light that certain generalized Hall-Littlewood polynomials indexed by compositions are the building blocks for the algebraic combinatorial theory of $q,t$-Catalan sequences and we prove a number of identities involving these functions.' address: - 'Department of Mathematics, University of Pennsylvania, Philadelphia, PA 19104-6395' - 'Department of Mathematics, Drexel University, Philadelphia, PA 19104' - 'Department of Mathematics and Statistics, York University, Toronto, Ontario M3J 1P3, CA' author: - 'J. Haglund' - 'J. Morse' - 'M. Zabrocki' title: - ' Dyck Paths with Forced and Forbidden Touch Points and $q,t$-Catalan building blocks' - A compositional Shuffle conjecture specifying touch points of the Dyck path --- Introduction ============ Our study concerns the combinatorics behind the character of the space of diagonal harmonics $\text{DH}_n$ and identities involving Macdonald polynomials that can be used to form expressions for this character. At the root of this theory is a linear operator $\nabla$, introduced in [@BGHT99], under which the modified Macdonald polynomials $\tilde H _{\mu}[X;q,t]$ are eigenfunctions. [@Hai02] proved that the Frobenius image of the character of $\text{DH}_n$ equals $\nabla e_n$. This gives an explicit expression involving rational functions in $q,t$ for the multiplicity of an irreducible indexed by a partition $\lambda$ in the character of $\text{DH}_n$. An important open problem in this area is the “shuffle conjecture" of [@HHLRU05] which asserts that the coefficient of $m_{\lambda}$ in $\nabla e_n$ simplifies to a $q,t$ statistic on lattice paths. A major breakthrough in this direction was made with the conjectured combinatorial formula of [@Hag03] for the coefficient of $m_{1^n}=s_{1^n}$ in $\nabla e_n$. In this case, the coefficient is a bi-graded version of the sign character and it is called the $q,t$-Catalan $C_n(q,t)$ since it reduces to the $n$th Catalan number when $q=t=1$. The combinatorial formula for $C_n(q,t)$ was proven in [@GaHa01],[@GaHa02] and pursuant work [@Hag04a] also settled the shuffle conjecture for partitions of hook-shape. However, the general case remains a mystery. An unrelated study of Macdonald polynomials led to the discovery [@LLM03] of a new family of symmetric functions called $k$-Schur functions $s_\lambda^{(k)}[X;t]$ which were conjectured to refine the special combinatorial properties held by Schur functions. The $k$-Schur functions have a number of conjecturally equivalent characterizations and it has now been established in [@LaMo08a; @Lam06b] that those introduced in [@LM07] refine combinatorial, geometric and representation theoretic aspects of Schur theory. This prompted Bergeron, Descouens, and Zabrocki to explore the role of $k$-Schur functions in the $q,t$-Catalan theory. To this end, they conjectured in [@BDZ10] that the coefficient of $s_{1^n}$ in $\nabla s_{1^n}^{(k)}[X;t]$ is a positive polynomial in $q,t$ and they proved their conjecture for the case that $t=1$. Our work here was initially motivated by a desire to find a combinatorial description for this coefficient in general, ideally in terms of a $q,t$-statistic on lattice paths as with the $q,t$-Catalans. We found such a description, but more remarkably this led us to discover that a natural setting for the combinatorial theory of $DH_n$ is created by applying $\nabla$ to the general set of Hall-Littlewood polynomials indexed by compositions. To be precise, it was proven in [@LaMo03a] that the $k$-Schur function $s_{1^n}^{(k)}[X;t]$ is merely a certain Hall-Littlewood polynomial. This led us to study $\nabla$ on a Hall-Littlewood polynomial indexed by any partition $\lambda$. But in fact, our work carries through to the family of polynomials $C_\alpha[X;q]$, for any composition $\alpha$, defined in terms of operators similar to Jing operators. A key component in the proof of the $q,t$-Catalan conjecture [@GaHa02] is the use of symmetric functions $E_{n,k}[X;q]$ which decompose $e_n$ into pieces that remain positive under the action of $\nabla$. We have discovered that the $C_\alpha[X;q]$ can be used as building blocks in the $q,t$-Catalan theory that decompose the $E_{n,k}[X;q]$ into finer pieces, still positive under the action of $\nabla$. Our conjectures on these building blocks thus refine earlier conjectures involving $E_{n,k}[X;q]$, the conjectures in [@BDZ10], the shuffle conjecture, and the conjectures in [@BGHT99] asserting that $\nabla$ applied to Hall-Littlewood functions have $q,t$-positive Schur coefficients. Loehr and Warrington [@LoWa08] introduced an intricate conjecture for the combinatorics of $\nabla$ applied to a Schur function $s_{\lambda}$. Our conjecture is extremely simple, describes the action of $\nabla$ on a larger set of symmetric functions than just a basis, and refines the conjecture of Loehr and Warrington when $\nabla$ acts on the Schur function $s_{(n-k,1^k)}$ [@LoWa08 Conjecture 3] as explained at the end of Section \[Nab\]. Garsia, Xin, and Zabrocki [@GXZ10] using work of Hicks [@Hic10] have now proven our generalized $q,t$-Catalan conjecture and expanded the result giving a “compositional $q,t$-Schröder” theorem. Definitions and notation ======================== Combinatorics ------------- A Dyck path is a lattice path in the first quadrant of the $xy$-plane from the point $(0,0)$ to the point $(n,n)$ with steps $+(0,1)$ and $+(1,0)$ which stays above the line $x=y$. For a Dyck path $D$, the cells in the $i^{th}$ row are those unit squares in the $xy$-plane that are below the path and fully above the line $x=y$ whose NE corner has a $y$ coordinate of $i$. The set of Dyck paths from $(0,0)$ to $(n,n)$ will be denoted $DP^n$ and the number of paths in this set is well known to be the Catalan number $$C_n = \frac{1}{n+1} \binom{2n}{n}~.$$ For a Dyck path $D$, let $a_i=a_i(D)$ equal the number of cells in the $i^{th}$ row of $D$. It is always true $a_1 = 0$ and $0 \leq a_{i+1} \leq a_i + 1$. We define the arm sequence $\text{arm}(D) = (a_1, a_2, \ldots, a_n)$ and note this completely determines $D$. We consider two statistics (non-negative integers) on Dyck paths. The $area$ statistic is the number of whole cells which are below the path and above the diagonal, or $$\area(D) = \sum_{i=1}^n a_i.$$ The $dinv$ statistic is defined as $$\dinv(D) = \sum_{1 \leq i < j \leq n} \chi( a_i - a_j \in \{ 0, 1\})$$ where $\chi( \text{true} ) =1$ and $\chi(\text{false}) = 0$. The Dyck path $D$ with arm sequence $(0^n)$ has $\area(D) = 0$ and $\dinv(D) = \binom{n}{2}$. The Dyck path $D'$ with arm sequence $(0, 1, 2, 3, \ldots, n-1)$ has $\area(D') = \binom{n}{2}$ and $\dinv(D') = 0$. -------------- -------------------------------------- -------------------------------------- -------------------------------------- -------------------------------------- {width=".75in"} {width=".75in"} {width=".75in"} {width=".75in"} arm sequence (0,0,0,0,0) (0,1,2,3,4) (0,1,2,2,1) (0,1,0,1,0) 0 10 6 2 10 0 4 7 -------------- -------------------------------------- -------------------------------------- -------------------------------------- -------------------------------------- The original proof of the combinatorial interpretation of the $q,t$-Catalan was stated in terms of a third statistic $bounce(D)$. Since we are more cleanly able to formulate our results in terms of the $\dinv(D)$ statistic we choose to state all the results in this paper in terms of the $\dinv(D)$ statistic however the reference [@Hag08 p. 50] describes an automorphism $\phi$ on $DP^n$ such that $\area(\phi(D)) = \text{bounce}(D)$ and $\dinv(\phi(D)) = \area(D)$. We make use of a partial order on Dyck paths; namely $D_1 \leq D_2$ if $\text{arm}(D_1) \leq \text{arm}(D_2)$, component-wise. In this case we say that $D_1$ is ‘below’ $D_2$ because $D_1$ will not cross $D_2$ and is hence weakly ‘between’ $D_2$ and the diagonal. A composition $\alpha$ of $n$, denoted $\alpha\models n$, is an integer sequence $\alpha = (\alpha_1, \alpha_2, \ldots, \alpha_r)$ with $\alpha_i \geq 1$ and where $|\alpha| = \alpha_1+ \alpha_2+ \cdots + \alpha_r = n$. The length of $\alpha$ is $\ell(\alpha)=r$. We shall also use ${{\overleftarrow{\alpha}}} = (\alpha_{\ell(\alpha)}, \alpha_{\ell(\alpha)-1}, \ldots, \alpha_2, \alpha_1)$. For any composition $\alpha$, we define $$n(\alpha) = \sum_{i=1}^{\ell(\alpha)} (i-1) \alpha_i\,.$$ The descent set of a composition $\alpha$ is defined to be $$\Des(\alpha) = \{\alpha_1 , \alpha_1 + \alpha_2 , \ldots, \alpha_1 + \alpha_2 + \cdots + \alpha_{\ell(\alpha)-1} \}~.$$ There is a common partial order defined on composition $\alpha, \beta \models n$ by letting $\alpha \leq \beta$ when $\alpha$ is ‘finer’ than $\beta$, i.e. $\Des(\beta) \subseteq \Des(\alpha)$. If $\alpha$ is a composition of $n$, $DP(\alpha)$ represents the Dyck path consisting of $\alpha_1$ steps in the North $(0, 1)$ direction followed by $\alpha_1$ steps in the East $(1, 0)$ direction, $\alpha_2$ $(0,1)$ steps followed by $\alpha_2$ $(1,0)$ steps, etc. A partition $\lambda = (\lambda_1,\ldots,\lambda_r)$ is a non-increasing sequence of positive integers. When $\lambda$ is a partition of $n$, denoted $\lambda\vdash n$, $|\lambda|=\sum \lambda_i = n$. The length of $\lambda$ is $\ell(\lambda)=r$. Given a partition $\lambda$, we set $$m(\lambda)= (m_1(\lambda), m_2(\lambda), m_3(\lambda), \cdots, m_{|\la|}(\lambda))\,,$$ where the numbers $m_i(\lambda)$ represent the number of parts of size $i$ in $\lambda$. The conjugate of a partition $\lambda$ is the partition $\lambda'= (\lambda_1',\lambda_2',\ldots,\lambda_m')$ where $\lambda_i'$ is the number of parts of $\lambda$ that are at least $i$. Partitions are generally considered to be compositions with parts arranged in non-increasing order. Hence, notions defined on compositions apply to partitions as well. Generally, we will use the symbols $\alpha, \beta, \gamma$ to represent compositions and $\la, \mu, \nu$ for partitions. For a given Dyck path $D$, $\touch(D)$ denotes the composition $\gamma = (\gamma_1, \gamma_2, \ldots, \gamma_{\ell(\gamma)}) \models n$ that specifies which rows the Dyck path ‘touches’ the diagonal. That is, for $\text{arm}(D) = (a_1, a_2, \ldots, a_n)$, $a_k = 0$ if and only if $k=1$ or $k-1 \in \Des(\gamma)$. The title of this paper comes from the notion of the $\touch$ composition. By requiring that $\touch(D) = \alpha$ for a fixed composition $\alpha$, we have specified that the Dyck path will touch the diagonal in rows $1, 1+\alpha_1, 1+\alpha_1+\alpha_2, \ldots, 1+\alpha_1+\alpha_2+ \cdots+ \alpha_{\ell(\alpha)-1}$ and is forbidden to touch the diagonal in the other rows. Note that under this definition, we view all paths as touching the diagonal in row $1$, but none touching in row $n+1$, and we say the path touches the diagonal $\ell (\alpha)$ times. The partial order on compositions is consistent with the partial order on Dyck paths in the sense that if $D_1$ and $D_2$ are Dyck paths such that $D_1 \leq D_2$, then $\touch(D_1) \leq \touch(D_2)$. Using these notions we introduce a new statistic $\doff_\alpha(D)$ for a given Dyck path $D$ with $\touch(D)\leq\alpha$. If $\text{arm}(D)=(a_1, a_2, \ldots, a_n)$, let $r_1$ be the number of rows such that $a_i=0$ for $1\leq i \leq \alpha_1$, $r_2$ be the number of rows such that $a_i=0$ for $\alpha_1< i \leq \alpha_1+\alpha_2$, and more generally $r_k = \#\{ i : a_i = 0\hbox{ and }\sum_{j=1}^{k-1} \alpha_j < i \leq\sum_{j=1}^{k} \alpha_j\}$. We then set $$\label{doffstat} \doff_\alpha(D) = \sum_{k=1}^{\ell(\alpha)}(\ell(\alpha)-k)r_k~.$$ If $\text{arm}(D) = (0,1,2,0,1,2,2,1,0,1,2,3,2,1)$, then $\touch(D) = (3,5,6)$. Taking $\alpha = (8,6)$ we have $\doff_{(8,6)}(D) = 2$. {width="2in"} The only Dyck path with $\touch(D) = (1^n)$ has $\text{arm}(D) = (0^n)$. There are $C_{n-1}$ Dyck paths with $\touch(D) = (n)$ and more generally there are $\prod_{i=1}^{\ell(\alpha)} C_{\alpha_i-1}$ Dyck paths such that $\touch(D) = \alpha$. Note that if $D$ and $E$ are two Dyck paths with $\text{touch}(D)=\text{touch}(E)\le \alpha$, then $\text{doff} _{\alpha} (D)=\text{doff} _{\alpha} (E)$. The results we mention so far are stated in terms of Dyck paths, but we will require the notion of parking functions to state the generalization of the shuffle conjecture. For a Dyck path $D$ in $DP^n$ with $\text{arm}(D)=(a_1,\ldots,a_n)$, let $\WP_D$ be the set of words of length $n$ in the alphabet $\{1, 2, \ldots , n\}$ such that $w_i < w_{i+1}$ if $a_i < a_{i+1}$. We use the notation $x^w$ to denote the monomial $x_{w_1} x_{w_2} \cdots x_{w_n}$. We also define an extension of the $\dinv$ statistic for words in $\WP_D$ by setting $$\begin{aligned} \dinv(w) = |\{&(i,j):1 \leq i < j \leq n, a_i = a_j\hbox{ and } w_i < w_j \}|\\ &+|\{(i,j): 1 \leq i < j \leq n, a_i = a_j + 1\hbox{ and }w_i > w_j\}|~.\end{aligned}$$ Symmetric functions ------------------- Let $X$ represent a sum of an infinite set of variables $X = x_1 + x_2 + x_3 + \cdots$ considered as elements of the ring of polynomial series in an infinite number of variables of bounded degree. For $r >0$, let $p_r$ represent a linear and algebraic morphism which acts on polynomial series by $p_r[x] = x^r$. That is for two polynomial series of bounded degree $A$ and $B$, $$\begin{aligned} p_r[A+B] &= p_r[A] + p_r[B]\\ p_r[A - B] &= p_r[A] - p_r[B]\\ p_r[AB] &= p_r[A] p_r[B]\\\end{aligned}$$ and in particular, $p_r[X] = x_1^r + x_2^r + x_3^r + \cdots$ represents the $r^{th}$ power sum in the variables $\{x_1, x_2, x_3, \ldots\}$. The ring of symmetric functions over the field $F$ is defined to be the polynomial ring $$\Lambda = F[ p_1[X], p_2[X], p_3[X], \ldots]~.$$ For our purposes, we choose the field $F$ to be the ring of rational power series in the variables $q,t,u,z$ over $\mathbb{Q}$ where each of the parameters $q,t,u,z$ all have the property that $p_r[a] = a^r$ for each $a = q,t,z,u$. Generally our symmetric functions $f$ will be considered as polynomials in the elements $p_r$ and then the notation $f[A]$ represents $f$ with each $p_r$ replaced by $p_r[A]$. The degree of $p_r$ is $r$ and the degree of a symmetric function $f$ is determined by the degree of the monomials in the power sums which appear in $f$. Following the notation of Macdonald [@Macdonald], we have the power sum basis $p_\lambda[X]$, Schur basis $s_\lambda[X]$, homogeneous basis $h_\lambda[X]$, and elementary basis $e_\lambda[X]$. In the expressions of variables it is useful to have a special symbol $\epsilon$ which will represent a value of negative one but behaves differently than a negative symbol. If $f$ is of homogeneous degree $r$, $$\begin{aligned} f[\epsilon X] &= (-1)^r f[X]\\ f[-\epsilon X]&= \omega (f[X])\\\end{aligned}$$ where $\omega$ is an involution on symmetric functions such that $\omega( p_\lambda[X] ) = (-1)^{|\la|+\ell(\la)}p_\lambda[X]$, $\omega( e_n[X] ) = h_n[X]$ and $\omega( s_\lambda[X] ) = s_{\lambda'}[X]$. We will also make use of the standard Hall scalar product which is defined by $$\left< p_\la[X], p_\mu[X]/z_\mu \right> = \left< s_\la[X], s_\mu[X] \right> = \chi(\la = \mu)$$ where $z_\mu = \prod_{i \geq 1} m_i(\mu)! i^{m_i(\mu)}$. For any symmetric function $f$, multiplication by $f$ is an operation on symmetric functions which raises the degree of the symmetric function by $deg(f)$. If we define $f^\perp$ to be the operation which is dual to multiplication in the sense that $$\left< f^\perp g, h \right> = \left< g, f h \right>,$$ then $f^\perp$ is an operator which lowers the degree of the symmetric function by $deg(f)$. It is not difficult to show that $$\label{hkperp} f[X + z] = \sum_{k\geq0} z^k (h_k^\perp f)[X],$$ $$\label{ekperp} f[X - z] = \sum_{k\geq0} (-z)^k (e_k^\perp f)[X].$$ In addition we will refer to the form of the Macdonald basis $\Ht_\lambda[X;q,t]$ that is relevant to the study of the $n!$ Theorem [@Hai01] and the $q,t$-Catalan numbers. The relation of this basis to the integral form $J_\mu[X;q,t]$ of [@Macdonald] is $$\Ht_\mu[X;q,t] = t^{n(\mu)} J_\mu\left[ \frac{X}{1-1/t};q,1/t \right].$$ It is also characterized as the unique basis such that $$\left< \Ht_\mu[X(1-1/t);q,t], \Ht_\la[X(1-q);q,t] \right> = 0$$ if $\la \neq \mu$ and $\left< \Ht_\mu[X;q,t], h_n[X] \right> = 1$. We are particularly interested in the Hall-Littlewood symmetric functions. Following the notation of Macdonald we define the functions $Q'_\la[X;q]$ to be the basis of the symmetric functions which satisfy $$\left< Q'_\lambda[X(1-q);q], Q'_\mu[X;q] \right> = 0$$ if $\la \neq \mu$ and $\left< Q'_\la[X;q], h_n[X] \right> = q^{n(\lambda)}$. Relating the definitions of the Hall-Littlewood and Macdonald symmetric functions, we note that $$Q'_\la[X;q] = \Ht_\la[X;0,1/q] q^{n(\la)}= \sum_{\la} K_{\la\mu}(q) s_\la[X]~.$$ The operator $\nabla$ was introduced in [@BGHT99] and is defined by $$\label{nabladef} \nabla \Ht_\la[X;q,t] = t^{n(\lambda)} q^{n(\la')} \Ht_\la[X;q,t]~.$$ This operator has been fundamental to the study of the $q,t$-combinatorial identities associated with $\text{DH}_n$ and Macdonald polynomials. Its definition is chosen so that $$\label{defqtcat} \left< \nabla( e_n[X] ), e_n[X] \right> = C_n(q,t)$$ where $C_n(q,t)$ is the $q,t$-Catalan polynomial. References [@GaHa01; @Hag03; @Hag08] showed that $$\label{CIqtcatalan} C_n(q,t) = \sum_{D \in DP^n} t^{\area(D)} q^{\dinv(D)}$$ with the sum over all Dyck paths of length $n$. A small example is $C_3(q,t) = q^3 + qt+ q^2t + qt^2 + t^3$, whose terms can be computed (in order) from the 5 Dyck paths of length $3$ with respective arm sequences $(0,0,0)$, $(0,0,1)$, $(0,1,0)$, $(0,1,1)$ and $(0,1,2)$: {width=".5in"}.2in {width=".5in"}.2in {width=".5in"}.2in {width=".5in"}.2in {width=".5in"} We will make use of the Newton element $$\begin{aligned} \label{newton} \Omega[X] = \sum_{\lambda} p_\lambda[X]/z_\lambda =\sum_{m \geq 0} h_{m}[X], \end{aligned}$$ where we have the identities $$\begin{aligned} \Omega[X + Y] &= \Omega[X]\Omega[Y]\\ \Omega[X - Y] &= \Omega[X]/\Omega[Y]\\ \Omega[x_1 + \epsilon x_2 - x_3 - \epsilon x_4] &= \frac{(1-x_3)(1+x_4)}{(1-x_1)(1+x_2)}\\ \Omega[XY] &= \sum_{\la} s_{\la}[X] s_\la[Y] = \sum_{\la} p_\la[X] p_\la[Y]/z_\la~.\end{aligned}$$ Jing [@Jin91] introduced a family of operators $\Hop_m$ indexed by $m \in \ZZ$ using the following formal power series in the parameter $z$, $$\begin{aligned} \label{hopdef} \Hop(z) P[X] &= \sum_{m \in \ZZ} z^m \Hop_m P[X] := P\left[ X - \frac{1-q}{z} \right] \Omega[zX]\\ &=\sum_{m \in \mathbb Z} z^m \sum_{r \geq 0} (-1)^r h_{m+r}[X] e_r[(1-q)X]^\perp P[X]~.\nonumber\end{aligned}$$ He proved that these operators create the Hall-Littlewood polynomials by adding rows. [@Jin91] For any partition $\lambda=(\lambda_1,\ldots,\lambda_\ell)$, $$Q'_\la[X;q] = \Hop_{\la_1} \Hop_{\la_2} \cdots \Hop_{\la_\ell}(1)~.$$ Two families of Hall-Littlewood symmetric functions =================================================== Our primary focus is the study of two families of symmetric functions and the combinatorics surrounding them. These functions arise from the following operators $\Bop_m$ and $\Cop_m$, closely related to Jing’s $\Hop_m$ operators from equation : $$\begin{aligned} \label{bopdef} \Bop(z) P[X] &= \sum_{m \in {\mathbb Z}} z^m \Bop_m P[X] := P\left[X + \epsilon \frac{(1-q)}{z}\right]\Omega[-\epsilon zX]\\ &= \sum_{m \in {\mathbb Z}} z^m \sum_{r \geq 0} (-1)^r e_{m+r}[X] h_r[X(1-q)]^\perp P[X] \nonumber\end{aligned}$$ and $$\begin{aligned} \label{copdef} \Cop(z) P[X]&= \sum_{m \in {\mathbb Z}} z^m \Cop_m P[X] := -q~ P\left[X + \epsilon\frac{(1-q)}{z}\right]\Omega[\epsilon(z/q) X]\\ &= \sum_{m \in {\mathbb Z}} (-1/q)^{m-1} z^m \sum_{r \geq 0} q^{-r} h_{m+r}[X] h_r[X(1-q)]^\perp P[X] \nonumber \,.\end{aligned}$$ The symmetric functions of particular interest here are those defined, for any composition $\alpha$, by setting $$B_\alpha[X;q] = \Bop_{\alpha_{\ell(\alpha)}} \Bop_{\alpha_{\ell(\alpha)-1}} \cdots \Bop_{\alpha_1}(1)$$ and $$C_\alpha[X;q] = \Cop_{\alpha_1} \Cop_{\alpha_2} \cdots \Cop_{\alpha_{\ell(\alpha)}}(1)~.$$ Note that the operators generating $B_\alpha$ and $C_\alpha$ are both indexed by the parts of $\alpha$, but are applied in reverse order with respect to one another. This is done so that the associated combinatorial and algebraic identities are more uniform. These operators are related by way of the equation: $$\label{HandBrel} \Bop(z) = \omega \Hop( z ) \omega$$ and $$\label{HandCrel} \Cop(z) = (-q) \Hop^{q \rightarrow 1/q}(-z/q),$$ or equivalently $$\label{operels} \Cop_m = (-1/q)^{m-1} \Hop_m^{q \rightarrow 1/q} = (-1/q)^{m-1} \omega \Bop_m^{q \rightarrow 1/q} \omega~.$$ Thus the functions themselves are related as: $$\label{BCrelation} Q'_\lambda[X;q] = \omega B_{{\overleftarrow{\lambda}}}[X;q] = (-q)^{ \ell(\lambda)-|\lambda|} C_\lambda[X;1/q] ~.$$ The Jing operators create Hall-Littlewood polynomials indexed by partitions which form a basis for the symmetric function ring. The $C_\alpha$ and $B_\alpha$ symmetric functions are indexed by compositions and are not linearly independent. The equations above detail how Hall-Littlewood symmetric functions are included in these families and therefore, Schur positive expansions of the $C_\alpha$ and $B_\alpha$ hold in certain cases. However, they are not Schur positive in complete generality. The smallest examples that are not uniformly Schur positive or Schur negative are $B_{(3,1)}[X;q]$ and $C_{(1,3)}[X;q]$. Below is a table of the symmetric functions $B_\alpha[X;q]$ and $C_\alpha[X;q]$ for $\alpha \models 4$. Notice that both $B_{(3,1)}[X;q]$ and $C_{(1,3)}[X;q]$ have mixed signs in their coefficients. $$\begin{bmatrix} B_{(1,1,1,1)}[X;q]\\ B_{(1,1,2)}[X;q]\\ B_{(1,2,1)}[X;q]\\ B_{(2,1,1)}[X;q]\\ B_{(1,3)}[X;q]\\ B_{(2,2)}[X;q]\\ B_{(3,1)}[X;q]\\ B_{(4)}[X;q] \end{bmatrix} = \begin{bmatrix} q^6&q^{3}+q^{4}+q^{5}&q^{2}+q^{4}&q+q^{2}+q^{3}&1\\ q^3&q+q^{2}&q&1&0\\ q^4&q^{2}+q^{3}&q^{2}&q&0\\ q^5&q^{3}+q^{4}&q^{3}&q^{2}&0\\ q&1&0&0&0\\ q^2&q&1&0&0\\ q^3&q^{2}&q-1&0&0\\ 1&0&0&0&0 \end{bmatrix} \begin{bmatrix} s_{(1,1,1,1)}[X;q]\\ s_{(2,1,1)}[X;q]\\ s_{(2,2)}[X;q]\\ s_{(3,1)}[X;q]\\ s_{(4)}[X;q] \end{bmatrix}$$ $$\begin{bmatrix} C_{(1,1,1,1)}[X;q]\\ C_{(1,1,2)}[X;q]\\ C_{(1,2,1)}[X;q]\\ C_{(2,1,1)}[X;q]\\ C_{(1,3)}[X;q]\\ C_{(2,2)}[X;q]\\ C_{(3,1)}[X;q]\\ C_{(4)}[X;q] \end{bmatrix} = \begin{bmatrix} 1&q^{-3}+q^{-2}+q^{-1}&q^{-4}+q^{-2}&q^{-5}+q^{-4}+q^{-3}&q^{-6}\\ 0&-q^{-3}&-q^{-4}&-q^{-5}-q^{-4}&-q^{-6}\\ 0&-q^{-2}&-q^{-3}&-q^{-4}-q^{-3}&-q^{-5}\\ 0&-q^{-1}&-q^{-2}&-q^{-3}-q^{-2}&-q^{-1}\\ 0&0&q^{-3}-q^{-2}&q^{-4}&q^{-5}\\ 0&0&q^{-2}&q^{-3}&q^{-4}\\ 0&0&0&q^{-2}&q^{-3}\\ 0&0&0&0&q^{-3} \end{bmatrix} \begin{bmatrix} s_{(1,1,1,1)}[X;q]\\ s_{(2,1,1)}[X;q]\\ s_{(2,2)}[X;q]\\ s_{(3,1)}[X;q]\\ s_{(4)}[X;q] \end{bmatrix}$$ To manipulate these symmetric functions, we derive commutation relations between the symmetric function operators. Our first result enables us to expand an element $C_\alpha[X;q]$, for any composition $\alpha$, in terms of the $C_\lambda[X;q]$ indexed by partitions $\lambda$. \[Copconjrel\] For $m, n \in \ZZ$, we have $$\label{Copconjreleq} q \Cop_m \Cop_n - \Cop_{m+1} \Cop_{n-1} = \Cop_n \Cop_m - q \Cop_{n-1} \Cop_{m+1}~.$$ In particular for $m \in \ZZ$, $$\Cop_{m} \Cop_{m+1} = \frac{1}{q} \Cop_{m+1} \Cop_m~.$$ We begin with the expressions for the $\Cop$-operators from equation . For ease of notation we shall use $h_m$ in place of $h_m[X]$ and $h_r^{q\perp}$ in place of the expression $h_r[X(1-q)]^\perp$. We compute that $$\begin{aligned} h_r^{q\perp} (h_m P[X]) &= h_m[X + (1-q)z] P[X + (1-q)z] \coeff_{z^r}\nonumber\\ &=\sum_{i\geq0} h_{m-i} h_i[1-q] P[X + (1-q)z] \coeff_{z^{r-i}} \label{hrcom}\\ &= \sum_{i \geq 0} h_{m-i} h_i[1-q] h_{r-i}^{q\perp} P[X]\nonumber.\end{aligned}$$ We also know that $$\label{hr1mqeval} h_r[1-q] = \begin{cases} 0 &\hbox{ if }r<0\,,\\ 1 &\hbox{ if }r=0\,,\\ 1-q &\hbox{ if }r>0\,. \end{cases}$$ These two identities imply that, $$\begin{aligned} q \Cop_m \Cop_n = &(-1/q)^{m+n-2} \sum_{i\geq0}\sum_{r \geq 0} \sum_{s \geq 0} q^{-r-s-i+1} h_{m+r+i} h_{n+s-i} h_i[1-q] h_{r}^{q\perp} h_s^{q\perp}\nonumber\\ = &(-1/q)^{m+n-2} \sum_{i\geq0}\sum_{r \geq 0} \sum_{s \geq 0} q^{-r-s-i+1} h_{m+r+i} h_{n+s-i} h_{r}^{q\perp} h_s^{q\perp}\nonumber\\ &- (-1/q)^{m+n-2} \sum_{i\geq1}\sum_{r \geq 0} \sum_{s \geq 0} q^{-r-s-i+2} h_{m+r+i} h_{n+s-i} h_{r}^{q\perp} h_s^{q\perp}\nonumber\\ = &(-1/q)^{m+n-2} \sum_{i\geq0}\sum_{r \geq 0} \sum_{s \geq 0} q^{-r-s-i+1} h_{m+r+i} h_{n+s-i} h_{r}^{q\perp} h_s^{q\perp}\nonumber\\ &- (-1/q)^{m+n-2} \sum_{i\geq0}\sum_{r \geq 0} \sum_{s \geq 0} q^{-r-s-i+1} h_{m+r+i+1} h_{n+s-i-1} h_{r}^{q\perp} h_s^{q\perp}\nonumber\\ = &(-1/q)^{m+n-2} \sum_{r \geq 0} \sum_{s \geq 0} \sum_{i\geq0} q^{-r-s-i+1} s_{(m+r+i,n+s-i)} h_{r}^{q\perp} h_s^{q\perp}. \label{firstterm}\end{aligned}$$ where to arrive at we have introduced the Schur function $s_{(a,b)} = h_a h_b - h_{a+1} h_{b-1}$. Similarly, $$\begin{aligned} \Cop_{m+1} \Cop_{n-1} &= (-1/q)^{m+n-2} \sum_{r \geq 0} \sum_{s \geq 0} \sum_{i\geq0} q^{-r-s-i} s_{(m+r+i+1,n+s-i-1)} h_{r}^{q\perp} h_s^{q\perp}\,.\end{aligned}$$ From these identities, we find that all terms in the difference $q \Cop_m \Cop_n - \Cop_{m+1} \Cop_{n-1}$ cancel except the $i=0$ term in : $$\begin{aligned} q \Cop_m \Cop_n - \Cop_{m+1} \Cop_{n-1} = &(-1/q)^{m+n-2} \sum_{r \geq 0} \sum_{s \geq 0} q^{-r-s+1} s_{(m+r,n+s)} h_{r}^{q\perp} h_s^{q\perp}\\ &+ (-1/q)^{m+n-2} \sum_{r \geq 0} \sum_{s \geq 0} \sum_{i\geq1} q^{-r-s-i+1} s_{(m+r+i,n+s-i)} h_{r}^{q\perp} h_s^{q\perp}\\ &- (-1/q)^{m+n-2} \sum_{r \geq 0} \sum_{s \geq 0} \sum_{i\geq0} q^{-r-s-i} s_{(m+r+i+1,n+s-i-1)} h_{r}^{q\perp} h_s^{q\perp}\\ = &(-1/q)^{m+n-2} \sum_{r \geq 0} \sum_{s \geq 0} q^{-r-s+1} s_{(m+r,n+s)} h_{r}^{q\perp} h_s^{q\perp}.\end{aligned}$$ We can then compute $\Cop_n \Cop_m - q \Cop_{n-1} \Cop_{m+1}$ from this by replacing $m \rightarrow n-1$ and $n \rightarrow m+1$. In particular, $$\begin{aligned} q \Cop_{n-1} \Cop_{m+1} - \Cop_{n} \Cop_{m} = &(-1/q)^{m+n-2} \sum_{r \geq 0} \sum_{s \geq 0} q^{-r-s+1} s_{(n-1+r,m+1+s)} h_{r}^{q\perp} h_s^{q\perp}.\end{aligned}$$ The identity $-s_{(b-1,a+1)} = s_{(a,b)}$ and the commutation of $h^\perp$ then implies our claim. An important consequence of this result is that if $\alpha$ is a composition of length $\ell$, then $C_{\alpha}[X;q]$ can be written as a linear combination of the $C_{\lambda}[X;q]$ where $\lambda$ are partitions that also have length $\ell$. The symmetric function $C_{(1,3)}[X;q]$ can be expressed in terms of $C_{(3,1)}[X;q]$ and $C_{(2,2)}[X;q]$ using this commutation relation since $q \Cop_1 \Cop_3 = \Cop_2 \Cop_2 + \Cop_3 \Cop_1 - q \Cop_2 \Cop_2$. Consequently, $C_{(1,3)}[X;q] = (1/q -1)C_{(2,2)}[X;q] + 1/qC_{(3,1)}[X;q]$. The relation of $\Cop$ to $\Bop$ given in enables us to derive an identity on $\Bop$ from Theorem \[Copconjrel\]. In particular, we simply apply $\omega$ to and replace $q$ by $1/q$. For $m \in \ZZ$, $$\Bop_m \Bop_n - q\Bop_{m+1} \Bop_{n-1} = q \Bop_n \Bop_m - \Bop_{n-1} \Bop_{m+1}\,.$$ In particular, letting $n = m+1$ gives $$\Bop_m \Bop_{m+1} = q \Bop_{m+1} \Bop_m ~.$$ In fact, we can also pin down commutation relations between the $\Bop$ and $\Cop$ operators if $m+n>0$ (note, the relation does not hold when $m+n\leq 0$). \[BCcommutation\] If $m + n > 0$, then $$\Bop_n \Cop_m = q \Cop_m \Bop_n~.$$ We use identities and and compute an expression for $\Bop_m \Cop_n$: $$\begin{aligned} \Bop_m \Cop_n = &(-1/q)^{n-1}\sum_{r \geq 0}\sum_{s \geq 0} \sum_{i \geq 0} (-1)^r q^{-s} e_{m+r} h_{n+s-i} h_i[1-q] h_{r-i}^{q\perp} h_s^{q\perp}\\ =&(-1/q)^{n-1} \sum_{r \geq 0}\sum_{s \geq 0} (-1)^{r} q^{-s} e_{m+r} h_{n+s} h_{r}^{q\perp} h_s^{q\perp}\\ &+(-1/q)^{n-1} \sum_{i \geq 1} \sum_{r \geq 0}\sum_{s \geq 0} (-1)^{r+i} q^{-s} e_{m+r+i} h_{n+s-i} (1-q) h_{r}^{q\perp} h_s^{q\perp}.\end{aligned}$$ Analogously, we also have the equations $h_r^{q\perp} e_m = \sum_{i \geq 0} e_{m-i} e_i[1-q] h_{r-i}^{q\perp}$ and $e_r[1-q] = (-q)^{r-1}h_r[1-q]\hbox{ if }r>0$. From this we derive a similar expression for $q \Cop_n \Bop_m$: $$\begin{aligned} q \Cop_n \Bop_m = &(-1/q)^{n-1} \sum_{s \geq 0}\sum_{r \geq 0}\sum_{i \geq 0} (-1)^r q^{-s+1} h_{n+s} e_{m+r-i} e_i[1-q] h_{s-i}^{q\perp} h_r^{q\perp}\\ = &(-1/q)^{n-1} \sum_{s \geq 0}\sum_{r \geq 0} (-1)^r q^{-s+1} e_{m+r} h_{n+s} h_{s}^{q\perp} h_r^{q\perp}\\ &+(-1/q)^{n-1} \sum_{i \geq 1} \sum_{s \geq 0}\sum_{r \geq 0} (-1)^{r+i+1} q^{-s} e_{m+r-i} h_{n+s+i} (1-q) h_{s}^{q\perp} h_r^{q\perp}.\end{aligned}$$ Their difference is $$\begin{aligned} \Bop_m\Cop_n - q \Cop_n \Bop_m = &(1-q)(-1/q)^{n-1} \sum_{s \geq 0}\sum_{r \geq 0} q^{-s} (-1)^r e_{m+r} h_{n+s} h_{r}^{q\perp} h_s^{q\perp}\\ &+(1-q)(-1/q)^{n-1} \sum_{r \geq 0} \sum_{s \geq 0}q^{-s} \left(\sum_{i \geq 1} (-1)^{r+i} e_{m+r-i} h_{n+s+i}\right) h_{r}^{q\perp} h_s^{q\perp}\\ &+(1-q)(-1/q)^{n-1} \sum_{r \geq 0} \sum_{s \geq 0}q^{-s} \left(\sum_{i \geq 1} (-1)^{r+i} e_{m+r+i} h_{n+s-i}\right) h_{r}^{q\perp} h_s^{q\perp}\,.\end{aligned}$$ In fact, the right hand side reduces to zero since $m+n>0$ implies that for each $r,s\geq 0$, $m+n+r+s>0$ and $$\begin{aligned} (-1)^r e_{m+r} h_{n+s} &+\sum_{i \geq 1} (-1)^{r+i} e_{m+r-i} h_{n+s+i} + \sum_{i\geq1} (-1)^{r+i} e_{m+r+i} h_{n+s-i} \\ &= \sum_{i=-n-s}^{m+r} (-1)^{r+i} e_{m+r-i} h_{n+s+i}=0\end{aligned}$$ by the identity $\sum_{i=0}^d (-1)^i e_{d-i} h_i = 0$ for all $d>0$. Jing’s operators generalize operators of Bernstein (see [@Macdonald]), defined by $$\begin{aligned} \label{Bhopdef} \Sop(z) P[X] &= \sum_{m \in \ZZ} z^m \Sop_m P[X] = P\left[ X - \frac{1}{z} \right] \Omega[zX]\\ &= \sum_{m \in \ZZ} z^m \sum_{r \geq 0} (-1)^r h_{m+r}[X] e_r[X]^\perp P[X]~.\nonumber\end{aligned}$$ These are creation operators for the Schur functions since $\Sop_{\la_1} \Sop_{\la_2} \cdots \Sop_{\la_\ell}(1) = s_{\lambda}[X]$, and they satisfy the commutation relation $\Sop_m \Sop_n = - \Sop_{n-1} \Sop_{m+1}$. We can write the Schur creation operators in terms of the $\Cop_a$ operators, which will help us in § \[Sym\] to write Schur functions in terms of the $C_{\alpha}$. For $m \in \ZZ$, $$\label{SopexpCop} \Sop_m = (-q)^{m-1} \sum_{i \geq 0} \Cop_{m+i} e_i^\perp .$$ We will use the identity $h_r[(1-q)X] = \sum_{j\geq0} h_j[X] h_{r-j}[-qX] = \sum_{j\geq0} (-q)^{r-j} h_j[X] e_{r-j}[X]$ and calculate directly, $$\begin{aligned} (-q)^{m-1} \sum_{i \geq 0} \Cop_{m+i} e_i^\perp &= (-q)^{m-1} \sum_{i \geq 0} (-1/q)^{m+i-1} \sum_{r \geq 0} q^{-r} h_{m+i +r}[X] h_r[X(1-q)]^\perp e_i^\perp\\ &= \sum_{r \geq 0} \sum_{i \geq 0} \sum_{j\geq0} (-1/q)^{i-r+j} q^{-r} h_{m+i +r}[X] h_j^\perp e_{r-j}^\perp e_i^\perp\\ &= \sum_{d \geq 0} \sum_{j=0}^d \sum_{r \geq 0} (-1)^{d-r} q^{-d} h_{m+d+r-j}[X] h_j^\perp e_{r-j}^\perp e_{d-j}^\perp\\ &= \sum_{d \geq 0} \sum_{j=0}^d \sum_{r \geq 0} (-1)^{d-r-j} q^{-d} h_{m+d+r}[X] h_j^\perp e_{r}^\perp e_{d-j}^\perp\\ &= \sum_{r \geq 0} (-1)^{-r} h_{m+r}[X] e_{r}^\perp = \Sop_m~\end{aligned}$$ where the last equality follows because $\sum_{j=0}^d (-1)^j h_j e_{d-j} = 0$ for all $d > 0$ so the remaining sum is only the part where $d=0$. In the reference [@GXZ10], the operator $\Cop_a$ is presented in a slightly different but equivalent expression. We note that a series $f(z) = \sum_{n \in \ZZ} f_n z^n$ has the property that $f(\epsilon z/q) \coeff_{z^a} = (-1/q)^a f(z) \coeff_{z^a}$. For this reason, $$\begin{aligned} \mathbb C _a P[X] &= -q P \left [ X + \epsilon \frac {(1-q)}{z} \right ] \Omega [ \epsilon (z/q) X ] \coeff_{z^a} \\ &= -q (-1/q)^a P \left [ X + \frac {(1-q)}{q z} \right ] \Omega [ z X ] \coeff_{z^a} \\ &= (-1/q)^{a-1} P \left [ X - \frac {1-1/q}{z} \right ] \Omega [zX] \coeff_{z^a}.\end{aligned}$$ The combinatorics of $\nabla$ applied to Hall-Littlewood polynomials {#Nab} ==================================================================== Recall that in the special case that $\alpha$ is a partition, $B_\alpha[X;q]$ and $C_\alpha[X;q]$ are closely related to the Hall-Littlewood symmetric functions. It was conjectured in [@BGHT99 Conjecture II and III] (partially attributed to A. Lascoux) that applying $\nabla$ to a Hall-Littlewood polynomial produces a Schur positive function. Our main discovery is that including all compositions $\alpha$ in the study of $\nabla(B_\alpha[X;q])$ and $\nabla(C_\alpha[X;q])$ leads to a natural refinement for the combinatorics of Dyck paths. Moreover, our combinatorial exploration led us to discover new symmetric function identities. One useful tool in the exploration of the operator $\nabla$ is the fact that $\nabla^{q=1}$ is a multiplicative operator. Since we can deduce from the operator definitions of our symmetric functions that $B_\alpha[X;1] = e_\alpha[X]$ and $C_\alpha[X;1] = h_\alpha[X]$, we have $$\nabla^{q=1}( B_\alpha[X;1] ) = \nabla^{q=1}( e_{\alpha_\ell}[X] )\nabla^{q=1}( e_{\alpha_{\ell-1}[X]} )\cdots \nabla^{q=1}( e_{\alpha_1}[X] )$$ and $$\nabla^{q=1}( C_\alpha[X;1] ) = \nabla^{q=1}( h_{\alpha_1}[X] )\nabla^{q=1} ( h_{\alpha_2}[X] )\cdots \nabla^{q=1}( h_{\alpha_\ell}[X] )~.$$ From this we can deduce the coefficient of $e_n[X]$. In particular, the coefficient of $e_n[X]$ in $\nabla( e_n[X] )$ and in $\nabla( h_n[X] )$ is the $q,t$-Catalan numbers $C_n(q,t)$ and $C_{n-1}(q,t)$, respectively. Thus, the coefficient of $e_n[X]$ in $\nabla^{q=1}( B_\alpha[X;1] )$ and in $\nabla^{q=1}( C_\alpha[X;1] )$ is $\prod_i C_{\alpha_i}(1,t)$ and $\prod_i C_{\alpha_i-1}(1,t)$ respectively. The combinatorial interpretation for $C_n(1,t)$ then gives combinatorial meaning to these coefficients. Namely, $\left< e_n[X], \nabla^{q=1}( B_\alpha[X;1] )\right>$ is the $t$-enumeration of Dyck paths (with weight $t$ raised to the ) which lie below the staircase consisting of $\alpha_1$ steps up and over, $\alpha_2$ steps up and over, etc. and $\left< e_n[X], \nabla^{q=1}( C_\alpha[X;1] )\right>$ is a $t$-enumeration of Dyck paths which touch the diagonal only in rows $1$, $1+\alpha_1$, $1+\alpha _1+\alpha_2$ steps, etc. For $\alpha$ a composition of $n$, $$\label{qeq1nablaBalpha} \left< \nabla^{q=1}( B_\alpha[X;1] ), e_n[X] \right> = \sum_{D \leq DP(\alpha)} t^{\area(D)}$$ and $$\label{qeq1nablaCalpha} \left< \nabla^{q=1}( C_\alpha[X;1] ), e_n[X] \right> = \sum_{\touch(D) = \alpha} t^{\area(D)}.$$ Remarkably, we have empirical evidence to suggest that in general, there is a combinatorial interpretation for the coefficient of $e_n[X]$ in $\nabla(B_\alpha[X;q])$ and in $\nabla(C_\alpha[X;q])$ that naturally generalizes the beautiful combinatorics of the $q,t$-Catalan. \[belowCI\] For $\alpha \models n$, $$\label{nBa} \left< \nabla( B_\alpha[X;q] ), e_n[X] \right> = \sum_{D \leq DP(\alpha)} t^{\area(D)} q^{\dinv(D)+\doff_\alpha(D)}~.$$ \[touchCI\] For $\alpha \models n$, $$\label{nCa} \left< \nabla(C_\alpha[X;q]), e_n[X] \right> = \sum_{\touch(D) = \alpha} t^{\area(D)} q^{\dinv(D)}.$$ Our work was inspired by the work of [@BDZ10] where they considered coefficients $\nabla( B_{{\overleftarrow{\lambda}}}[X;q])$ for $\lambda$ a hook partition (since for that case $\doff_{{\overleftarrow{\lambda}}}(D) = 0$). The innovation in these identities is to consider symmetric functions indexed by compositions which allowed us to conjecture the action of $\nabla$ on a spanning set of the symmetric functions. More generally, we have conjectures for the expansion of $\nabla( B_{\alpha}[X;q] )$ and $\nabla( C_{\alpha}[X;q] )$ into monomials. \[combinterpBa\] $$\label{nBa2} \nabla( B_\alpha[X;q] ) = \sum_{D \leq DP(\alpha)} \sum_{w \in \WP_{D}} t^{\area(w)} q^{\dinv(w)+\doff_\alpha(D)} x^w.$$ \[combinterpCa\] $$\label{nCa2} \nabla(C_\alpha[X;q]) = \sum_{\touch(D) = \alpha} \sum_{w \in \WP_{D}} t^{\area(D)} q^{\dinv(w)} x^w.$$ By the arguments in [@HHLRU05] (see also [@Hag08 p.99]) Conjecture \[combinterpBa\] and \[combinterpCa\] imply Conjecture \[belowCI\] and \[touchCI\]. The case $\alpha = (n)$ of reduces to the shuffle conjecture since $B_{(n)}[X;q] = e_n[X]$. Also, because of the expansion of $s_{(n-k,1^k)}$ in Proposition \[sn1nmkexpansion\] in the next section, implies the special case of the Loehr-Warrington conjecture [@LoWa08 Conjecture 3] involving the action of $\nabla$ on the Schur function $s_{(n-k,1^k)}$. We will prove in the next section that Conjecture \[combinterpBa\] and \[combinterpCa\] are equivalent to each other (and by consequence Conjecture \[belowCI\] and \[touchCI\] are equivalent as well). In work building on our results here, [@GXZ10] with contributions from [@Hic10] proved Conjecture \[touchCI\]. Symmetric Function Identities {#Sym} ============================= The exploration of $q,t$-Catalans led [@GaHa01] to the special symmetric function elements $E_{n,k}[X;q]$, defined by the algebraic identity $$e_n\left[X\frac{1-z}{1-q} \right] = \sum_{k=1}^n \frac{(z;q)_k}{(q;q)_k} E_{n,k}[X;q]$$ where $(z;q)_k = (1-z)(1-qz)\cdots(1-q^{k-1}z)$. These elements play a fundamental role in the proof that the $q,t$-Catalan polynomial is the $q,t$-enumeration of Dyck paths as given in . Namely, the proof follows by showing that $$\label{recur} \langle\nabla E_{n,k}[X;q], e_n[X]\rangle= q^{k\choose 2} t^{n-k} \sum_{r=0}^{n-k} \begin{bmatrix} r+k-1 \\ r \end{bmatrix}_q \left<\nabla(E_{n-k,r}[X;q]),e_{n-r}[X] \right>\,,$$ where $\begin{bmatrix}n\\k\end{bmatrix}_q = \frac{(q;q)_n}{(q;q)_k (q;q)_{n-k}}$, and the combinatorial interpretation for $\langle\nabla E_{n,k}[X;q], e_n[X]\rangle$ in terms of Dyck paths satisfies the same recurrence. In particular, the coefficient of $e_n[X]$ in $\nabla( E_{n,k}[X;q] )$ $q,t$-enumerates the Dyck paths which touch the diagonal $k$ times. From this, Conjecture \[touchCI\] leads us to expect that $$\left< e_n[X], \nabla( E_{n,k}[X;q] ) \right> = \sum_{\alpha \models n, \ell(\alpha) = k} \left< e_n[X], \nabla( C_{\alpha}[X;q] ) \right>\,.$$ In fact, we have discovered much more generally that $$E_{n,k}[X;q]=\sum_{\alpha \models n\atop\ell(\alpha)=k} C_\alpha[X;q]\,.$$ This section is devoted to proving this surprising result, which suggests that the $C_\alpha[X;q]$ are the building blocks in $q,t$-Catalan theory. A key step in the proof of our Conjectures \[belowCI\] and \[touchCI\] relies on extending the recurrence to involve Dyck paths which touch the diagonal at certain points. This is carried out in [@GXZ10]. Our point of departure is to give a simple expression for $e_n[X]$ in terms of the Hall-Littlewood symmetric functions $C_\alpha[X;q]$. \[enexpansioninCa\] $$\label{enexpansion} e_n[X] = \sum_{\alpha \models n} C_\alpha[X;q].$$ Assume by induction on $n$ that equation holds (the base cases of $n=0$ and $1$ are easily verified). Since the operator $\Sop_m$ is a creation operator for the Schur functions, by we have $$\begin{aligned} e_n[X] &= s_{(1^n)}[X] = \Sop_1( s_{(1^{n-1})}[X]) = \sum_{i=0}^{n-1} \Cop_{1+i} s_{(1^{n-i-1})}[X]\end{aligned}$$ Which, by induction, gives $$\begin{aligned} e_n[X]= \sum_{i=0}^{n-1} \sum_{\alpha\models n-i-1} \Cop_{i+1} C_{\alpha}[X;q] = \sum_{\alpha\models n} C_\alpha[X;q]~.\end{aligned}$$ Proposition \[enexpansioninCa\] can be stated in a more general form, suggesting that any Schur function may expand nicely in terms of our Hall-Littlewood spanning set. \[sn1nmkexpansion\] For $0 \leq k < n$, $$s_{(n-k,1^k)}[X] = (-q)^{n-k-1}\sum_{\substack{\alpha \models n\\\alpha_1 \geq n-k}} C_\alpha[X;q]~.$$ Again using that $\Sop$ is a Schur function creation operator, by we have $$\begin{aligned} s_{(n-k,1^k)}[X] &= \Sop_{n-k}( s_{1^k}[X] ) = (-q)^{n-k-1} \sum_{i=0}^{k} \Cop_{n-k+i} \left( s_{(1^{k-i})}[X] \right) \,.\end{aligned}$$ The previous proposition then implies $$\begin{aligned} s_{(n-k,1^k)}[X] &= (-q)^{n-k-1} \sum_{i=0}^{k} \Cop_{n-k+i} \left(\sum_{\alpha \models k-i} C_{\alpha}[X;q]\right)\\ &= (-q)^{n-k-1} \sum_{i=0}^{k} \sum_{\alpha \models k-i} C_{(n-k+i,\alpha)}[X;q] = (-q)^{n-k-1} \sum_{\substack{\alpha \models n\\ \alpha_1\geq n-k}} C_{\alpha}[X;q]~.\end{aligned}$$ We are now in the position to prove that $E_{n,k}[X;q]$ can be decomposed canonically in terms of the $C_\alpha[X;q]$. \[Enkexpansion\] For $0\leq k<n$, $$\label{Enkformula} E_{n,k}[X;q] = \sum_{\substack{\mu \vdash n\\\ell(\mu) = k}} q^{-n(\mu)-k + M(\mu)} {\begin{bmatrix}k\\ m(\mu)\end{bmatrix}}_q C_\mu[X;q], $$ where $M(\mu)=\sum_{i=1}^n {m_i(\mu)+1 \choose 2}$ and ${\begin{bmatrix}k\\ m(\mu)\end{bmatrix}}_q= (q;q)_k/\prod_{i=1}^n (q;q)_{m_i(\mu)}$. Recall the expansion of the elementary symmetric functions in the Macdonald basis is given by (see [@GaHa01]) $$e_n\left[X\frac{1-z}{1-q}\right] = \sum_{\mu \vdash n} \frac{\Ht_\mu[X;q,t] \Ht_\mu[(1-z)(1-t);q,t]}{{\tilde h}_\mu(q,t) {\tilde h}'_\mu(q,t)},$$ where ${\tilde h}_\mu(q,t) = \prod_{c \in \mu} (q^{a(c)} - t^{l(c) +1})$ and ${\tilde h}'_\mu(q,t) = \prod_{c \in \mu}(t^{l(c)} - q^{a(c) +1})$. When $t=0$, these expressions reduce to $${\tilde h}_\mu(q,0) = \prod_{c \in \mu} (q^{a(c)} - 0^{l(c) +1}) = q^{n(\mu')}$$ and $$\begin{aligned} {\tilde h}'_\mu(q,0) &= \prod_{\substack{c \in \mu\\l(c)=0}}(1 - q^{a(c) +1}) \prod_{\substack{c \in \mu\\l(c)\neq 0}}(- q^{a(c) +1}) \\ &=(-1)^{n-\mu_1} q^{n+n(\mu')-M(\mu ^{\prime})} \prod_{i=1}^n (q;q)_{m_i(\mu')}.\end{aligned}$$ Therefore since $\Ht_\mu[X;q,t] = \Ht_{\mu'}[X;t,q]$, when we set $t=0$ everywhere we have $$\label{ent0} e_n\left[X\frac{1-z}{1-q}\right] = \sum_{\mu \vdash n} (-1)^{n-\mu_1} q^{-n-2n(\mu')+M(\mu ^{\prime})} \frac{\Ht_{\mu'}[X;0,q] \Ht_{\mu}[(1-z);q,0]} {\prod_{i=1}^n (q;q)_{m_i(\mu')}}\,.$$ Now the evaluation $$\Ht_\mu[(1-z);q,t] = \prod_{c \in \mu} (1-z t^{l'(c)} q^{a'(c)})$$ also yields $$\Ht_\mu[(1-z);q,0] = (z;q)_{\mu_1}~.$$ Thus, replacing $\mu$ by $\mu ^{\prime}$ in , and thereby exchanging $\mu_1$ and $\ell(\mu)$, gives $$e_n\left[X\frac{1-z}{1-q}\right] = \sum_{\mu \vdash n} (-1)^{n-\ell(\mu)} q^{-n-2n(\mu)+M(\mu)} \frac{\Ht_{\mu}[X;0,q] (z;q)_{\ell(\mu)}} {\prod_{i=1}^n (q;q)_{m_i(\mu)}}\,.$$ Since $\Ht_\mu[X;0,q] = (-1)^{n-\ell(\mu)} q^{n(\mu) + n - \ell(\mu)} C_\mu[X;q]$, we have $$e_n\left[X\frac{1-z}{1-q}\right] = \sum_{k=1}^n \frac{(z;q)_k}{(q;q)_k} \sum_{\substack{\mu \vdash n\\\ell(\mu)=k}} q^{-k-n(\mu)+M(\mu)} \frac{C_\mu[X;q] (q;q)_{k}} {\prod_{i=1}^n (q;q)_{m_i(\mu)}}\,,$$ which implies our claim. The $q$-binomial coefficients that appear in equation suggest that there is a relation between the terms of the $C_\la[X;q]$ basis and subsets of a $k$ element set. It turns out that Proposition \[Enkexpansion\] can be more cleanly written over compositions using a different expansion. \[cleanEnkformula\] For $0\leq k<n$, $$E_{n,k}[X;q] = \sum_{\substack{\alpha\models n\\ \ell(\alpha)=k}} C_\alpha[X;q]\,.$$ Using the straightening relations of the $\Cop_m$ operators, if $\alpha$ is a composition of $n$ and $\la$ is a partition of $n$ such that $\ell(\la) \neq \ell(\alpha)$, then $$C_\alpha[X;q] \coeff_{C_\la[X;q]} = 0~.$$ Now for $\ell(\la) = k$, by Proposition \[Enkexpansion\] and the fact that $e_n[X] = \sum_{k=1}^n E_{n,k}[X;q]$, $$\begin{aligned} E_{n,k}[X;q] \coeff_{C_\la[X;q]} &= e_n[X] \coeff_{C_\la[X;q]}\\ &= \sum_{\alpha \models n} C_\alpha[X;q] \coeff_{C_\la[X;q]}\\ &= \sum_{\substack{ \alpha \models n\\ \ell(\alpha)= k}} C_\alpha[X;q] \coeff_{C_\la[X;q]}.\end{aligned}$$ Furthermore if $\ell(\la) \neq k$ then $$E_{n,k}[X;q] \coeff_{C_\la[X;q]} = 0 = \sum_{\substack{ \alpha \models n\\ \ell(\alpha)= k}} C_\alpha[X;q] \coeff_{C_\la[X;q]}.$$ Since the functions $C_\la[X;q]$ are a basis this implies that $E_{n,k}[X;q] = \sum_{\substack{ \alpha \models n\\ \ell(\alpha)= k}} C_\alpha[X;q]$. We have seen in that $\Cop$ is naturally related to $\Bop$. Here we pin down the relationship between the symmetric functions $B_\alpha[X;q]$ and $C_\alpha[X;q]$. A by-product of this identity is that Conjecture \[combinterpCa\] implies Conjecture \[combinterpBa\]. \[BatoCaexpansion\] For $n \geq 0$ and any composition $\alpha \models n$, $$\label{BatoCa} B_{\alpha}[X;q] = \sum_{\beta \leq \alpha} q^{\doff_\alpha(DP(\beta))} C_\beta[X;q]\,.$$ We show this result by induction on the number of parts of $\alpha$. The base case follows since $\Bop_m(1) = e_m[X]$ which is equal to $\sum_{\gamma\models m} C_\gamma[X;q]$ by Proposition \[enexpansioninCa\]. Assume by induction that holds for a composition $\alpha$ of length $\ell$ and consider a composition $(m, \alpha)$. We then have $$\begin{aligned} \label{Balpham} B_{(\alpha,m)}[X;q] &= \Bop_m( B_\alpha[X;q]) =\sum_{\beta \leq \alpha} q^{\doff_\alpha(DP(\beta))} \Bop_m( C_\beta[X;q] )\,. $$ Now consider $\Bop_m( C_\beta[X;q] ) = \Bop_m \circ \Cop_{\beta_1}\circ \Cop_{\beta_2} \circ \cdots \circ\Cop_{\beta_{\ell(\beta)}}(1)$. The commutation relation between the $\Cop_n$ and $\Bop_m$ from Theorem \[BCcommutation\] implies $$\begin{aligned} \Bop_m( C_\beta[X;q] ) &= q^{\ell(\beta)} \Cop_{\beta_1}\circ \Cop_{\beta_2} \circ \cdots \circ\Cop_{\beta_{\ell(\beta)}} \circ \Bop_m(1)\,. $$ By Proposition \[enexpansioninCa\], we then have $$\begin{aligned} \Bop_m( C_\beta[X;q] ) &= q^{\ell(\beta)} \Cop_{\beta_1}\circ \Cop_{\beta_2} \circ \cdots \circ\Cop_{\beta_{\ell(\beta)}} ( \sum_{\gamma \models m} C_\gamma[X;q] )\\ &= q^{\ell(\beta)} \sum_{\gamma \models m} C_{(\beta,\gamma)}[X;q] )\,. \label{BmC}\end{aligned}$$ Putting into , we thus find that $$\begin{aligned} B_{(\alpha,m)}[X;q] &=\sum_{\beta \leq \alpha} q^{\doff_\alpha(DP(\beta))+\ell(\beta)} \sum_{\gamma \models m} C_{(\beta,\gamma)}[X;q]\,.\end{aligned}$$ For each term in the sum, the composition $(\beta,\gamma)$ is finer than the composition $(\alpha, m)$. Moreover, if we let $r_i$ be the number of times that $DP(\beta)$ touches the diagonal below the $i^{th}$ bump of the Dyck path $DP(\alpha)$, then $\doff_{(\alpha,m)}( DP(\beta,\gamma) ) = \sum_{i=1}^{\ell(\alpha)} r_i (\ell(\alpha)+1-i) = \sum_{i=1}^{\ell(\alpha)} r_i (\ell(\alpha)-i) + \sum_{i=1}^{\ell(\alpha)} r_i = \doff_{\alpha}(DP) + \ell(\beta)$. To prove that that Conjectures \[combinterpCa\] and \[combinterpBa\] are in fact equivalent, we need to express $C_\alpha[X;q]$ in terms of $B_\beta[X;q]$. \[www\] Let $\gamma, \alpha$ be compositions with $\gamma \le \alpha$. Then $$\begin{aligned} \sum_{{\beta \atop \gamma \leq \beta \leq \alpha }} (-1)^{ \ell(\alpha) - \ell(\beta) } q^{ \ell(\alpha) - \ell(\beta)+\doff_{\beta}( DP(\gamma) ) - \doff_{\overleftarrow{\alpha}}(DP(\overleftarrow{\beta}))} = \begin{cases} 0 \text{ if $\gamma < \alpha$}\\ 1 \text{ if $\gamma = \alpha$} \end{cases}.\end{aligned}$$ First assume that $\gamma < \alpha$ and consider the difference of the descent sets $\Des(\gamma)\backslash \Des(\alpha) = \{ i_1, i_2, \ldots, i_d\}$. There are $2^d$ compositions $\beta$ such that $\gamma \leq \beta \leq \alpha$ and we will pair them up with a sign reversing involution. If $i_1 \in \Des(\beta)$, let $\tilde{\beta}$ be the composition with $\Des(\tilde{\beta}) = \Des(\beta)\backslash\{i_1\}$ (the terms with $i_1 \in \Des(\beta)$ will match with the terms $i_1 \notin \Des(\tilde{\beta})$). There is some $r > 1$ such that $\alpha_1 = \beta_1$, $\alpha_2 = \beta_2, \ldots, \alpha_r > \beta_r$ because $\Des(\beta)$ contains the descent $i_1$ that is not in $\Des(\alpha)$. Calculating directly we have that, $$\begin{aligned} \ell(\alpha) - \ell(\beta) &= \ell(\alpha) - \ell(\tilde{\beta}) - 1 \\ \doff_\beta(DP(\gamma)) &= \doff_{\tilde{\beta}}(DP(\gamma)) + r \\ -\doff_{\overleftarrow{\alpha}}(DP(\overleftarrow{\beta})) &= - \doff_{\overleftarrow{\alpha}}(DP(\overleftarrow{\tilde{\beta}}))-r+1.\end{aligned}$$ Therefore $$\ell(\alpha) - \ell(\beta)+\doff_\beta(DP(\gamma))- \doff_{\overleftarrow{\alpha}}(DP(\overleftarrow{\beta}))= \ell(\alpha) - \ell(\tilde{\beta})+\doff_{\tilde{\beta}}(DP(\gamma)) - \doff_{\overleftarrow{\alpha}}(DP(\overleftarrow{\tilde{\beta}}))$$ and the signs of these terms are different in the sum. This provides a sign reversing involution and hence those terms with $\touch(D) < \alpha$ sum to $0$ matching those terms with $\Des(\beta)$ which include the smallest descent. Now for those terms with $\touch(D) = \alpha$ we have $\gamma= \beta = \alpha$ and $$\ell(\alpha) - \ell(\beta) + \doff_\beta(DP(\gamma)) - \doff_{\overleftarrow{\alpha}}(DP( \overleftarrow{\beta})) = 0$$ since $\doff_\alpha(DP(\alpha)) = {\begin{pmatrix}\ell(\alpha)\\ 2\end{pmatrix}}$. \[CatoBaexpansion\] For $n \geq 0$ and for any composition $\alpha \models n$, $$\label{CatoBa} C_{\alpha}[X;q] = \sum_{\beta \leq \alpha} (-q)^{\ell(\alpha)-\ell(\beta)} q^{-\doff_{\overleftarrow\alpha}(DP(\overleftarrow\beta))} B_\beta[X;q]~.$$ By Theorem \[BatoCaexpansion\] we have $$\begin{aligned} \sum_{\beta \leq \alpha} &(-q)^{ \ell(\alpha)-\ell(\beta) } q^{ -\doff_{\overleftarrow\alpha}(DP(\overleftarrow\beta)) } B_\beta[X;q] \\ &=\sum_{\beta \leq \alpha} (-q)^{ \ell(\alpha)-\ell(\beta) } q^{ -\doff_{\overleftarrow\alpha}(DP(\overleftarrow\beta)) } \sum _{\gamma \le \beta} q^{ \text{doff}_{\beta} (DP(\gamma)) } C_{\gamma}[X;q] \\ &= \sum_{\gamma} C_{\gamma}[X;q] \sum_{\beta \atop \gamma \le \beta \le \alpha} (-q)^{\ell(\alpha)-\ell(\beta)} q^{-\doff_{\overleftarrow\alpha}(DP(\overleftarrow\beta))} q^{\text{doff}_{\beta} (DP(\gamma))}\,,\end{aligned}$$ and our claim now follows by Lemma \[www\]. \[combequiv\] Conjecture \[combinterpCa\] is true if and only if Conjecture \[combinterpBa\] is true. Theorem \[BatoCaexpansion\] gives $B_\alpha[X;q]$ in terms of $C_\beta[X;q]$, to which we apply $\nabla$: $$\begin{aligned} \nabla(B_{\alpha}[X;q]) &= \sum_{\beta \leq \alpha} q^{\doff_\alpha(DP(\beta))} \nabla(C_\beta[X;q]) \,.\end{aligned}$$ Given that Conjecture \[combinterpCa\] holds, we then have that $$\begin{aligned} \nabla(B_{\alpha}[X;q]) &= \sum_{\beta \leq \alpha} \, \sum_{ D \atop \touch(D) = \beta} \, \sum_{w \in \WP_{D}} t^{\area(D)} q^{\dinv(w)+\doff_\alpha(DP(\beta))} x^w\\ &= \sum_{D \leq DP(\beta)} \sum_{w \in \WP_{D}} t^{\area(D)} q^{\dinv(w)+\doff_\alpha(DP(\beta))} x^w~.\end{aligned}$$ On the other hand, assuming Conjecture \[combinterpBa\] holds, Theorem \[CatoBaexpansion\] gives $$\begin{aligned} \nabla(C_{\alpha}[X;q]) &= \sum_{\beta \leq \alpha} (-q)^{\ell(\alpha)-\ell(\beta)} q^{-\doff_{\overleftarrow\alpha}(DP(\overleftarrow\beta))} \nabla(B_\beta[X;q])\nonumber\\ &= \sum_{\beta \leq \alpha} \, \sum_{D \leq DP(\beta)} \, \sum_{w \in \WP_D} t^{\area(D)} (-q)^{\ell(\alpha)-\ell(\beta)} q^{\dinv(w)+\doff_\beta(D)- \doff_{\overleftarrow{\alpha}}(DP(\overleftarrow{\beta}))} x^w~.\label{tocancel}\end{aligned}$$ For a Dyck path $D$ in the sum, let $\gamma = \touch(D)$. Note that $D \leq DP(\beta)$ implies that $\gamma \leq \beta \leq \alpha$, and thus we may rearrange sums as follows: $$\begin{aligned} \sum_{\beta \leq \alpha} \, &\sum_{D \leq DP(\beta)} (-q)^{\ell(\alpha) - \ell(\beta)} q^{\doff_\beta(D) - \doff_{\overleftarrow{\alpha}}(DP(\overleftarrow{\beta}))}\\ &= \sum_{\gamma \leq \alpha} \sum_{D \atop \touch(D) = \gamma} \sum_{\beta \atop \gamma \leq \beta \leq \alpha} (-1)^{\ell(\alpha) - \ell(\beta)} q^{\ell(\alpha) - \ell(\beta)+\doff_\beta(D) - \doff_{\overleftarrow{\alpha}}(DP(\overleftarrow{\beta}))}\,.\end{aligned}$$ Lemma \[www\] allows us to conclude that reduces to Conjecture \[combinterpCa\]. Conjecture \[belowCI\] is true if and only if Conjecture \[touchCI\] is true. [99]{} N. Bergeron, F. Descouens, and M. Zabrocki, *A filtration of [$(q,t)$]{}-[C]{}atalan numbers*, Adv. in Appl. Math. **44** (2010), no. 1, 16–36. F. Bergeron, A. M. Garsia, M. Haiman, and G. Tesler, *Identities and positivity conjectures for some remarkable operators in the theory of symmetric functions*, Methods Appl. Anal. **6** (1999), 363–420. A. M. Garsia and J. Haglund, *A positivity result in the theory of [M]{}acdonald polynomials*, Proc. Nat. Acad. Sci. U.S.A. **98** (2001), 4313–4316. , *A proof of the $q,t$-[C]{}atalan positivity conjecture*, Discrete Math. **256** (2002), 677–717. A. Garsia, G. Xin, and M. Zabrocki, *[H]{}all-[L]{}ittlewood operators in the theory of parking functions and diagonal harmonics*. International Mathematics Research Notices 2011; doi:10.1093/imrn/rnr060 . J. Haglund, *Conjectured statistics for the [$q,t$]{}-[C]{}atalan numbers*, Adv. Math. **175** (2003), no. 2, 319–334. , *A proof of the [$q,t$]{}-[S]{}chr[ö]{}der conjecture*, Internat. Math. Res. Notices **11** (2004), 525–560. James Haglund, *The [$q$]{},[$t$]{}-[C]{}atalan numbers and the space of diagonal harmonics*, University Lecture Series, vol. 41, American Mathematical Society, Providence, RI, 2008, With an appendix on the combinatorics of Macdonald polynomials. M. Haiman, *Hilbert schemes, polygraphs, and the [M]{}acdonald positivity conjecture*, J. Amer. Math. Soc. **14** (2001), 941–1006. , *Vanishing theorems and character formulas for the [H]{}ilbert scheme of points in the plane*, Invent. Math. **149** (2002), 371–407. J. Haglund, M. Haiman, N. Loehr, J. B. Remmel, and A. Ulyanov, *A combinatorial formula for the character of the diagonal coinvariants*, Duke J. Math. **126** (2005), 195–232. A. Hicks, *Two parking function bijections refining the $q,t$-[C]{}atalan and [S]{}chröder recursions*, preprint, 2010. Nai Huan Jing, *Vertex operators and [H]{}all-[L]{}ittlewood symmetric functions*, Adv. Math. **87** (1991), no. 2, 226–248. T. Lam, *Schubert polynomials for the affine grassmannian*, J. Amer. Math Soc **21** (2008), no. 1, 259–281. L. Lapointe, A. Lascoux, and J. Morse, *Tableau atoms and a new [M]{}acdonald positivity conjecture*, Duke Math. J. **116** (2003), no. 1, 103–146. L. Lapointe and J. Morse, *Schur function analogs for a filtration of the symmetric function space*, J. Combin. Theory Ser. A **101** (2003), no. 2, 191–224. L. Lapointe, , and J. Morse, *A $k$-tableaux characterization of $k$-schur functions*, Adv Math **213** (2007), no. 1, 183–204. L. Lapointe and J. Morse, *Quantum cohomology and the [$k$]{}-[S]{}chur basis*, Trans. Amer. Math. Soc. **360** (2008), no. 4, 2021–2040. N. Loehr and G. Warrington, *Nested quantum [D]{}yck paths and [$\nabla(s_\lambda)$]{}*, Int. Math. Res. Not. IMRN **5** (2008), Art. ID rnm 157, 29. I. G. Macdonald, *Symmetric functions and [H]{}all polynomials*, Oxford Mathematical Monographs, second ed., Oxford Science Publications, The Clarendon Press Oxford University Press, New York, 1995.

ArXiv

--- abstract: 'We study effects from new colored states and the Higgs portal on gluon fusion production. We isolate possible loop contributions from new colored scalars, fermions, and vectors, incorporating effects from Higgs portal-induced scalar mixing, thus leading to dramatic effects on gluon fusion and branching fractions. Higgs identification must generally allow for these effects, and using our results, possible tensions from fits to the Standard Model expectation can be relieved by inclusion of New Physics effects.' author: - Kunal Kumar - 'Roberto Vega-Morales' - Felix Yu bibliography: - 'ggH.bib' title: Effects from New Colored States and the Higgs Portal on Gluon Fusion and Higgs Decays --- Introduction {#sec:Introduction} ============ The ATLAS [@:2012gk] and CMS [@:2012gu] experiments at the Large Hadron Collider (LHC) have recently presented results that indicate the observation (also supported by evidence coming from the 1.96 TeV run of the Tevatron [@TEVNPH:2012ab]) of a new resonance. When interpreted as a Standard Model (SM) Higgs, the combined channels are consistent with the SM expectation at $1 \sigma$, yet individual channels show deviations from the SM expectation in the 1–2$\sigma$ range. The discovery of the Higgs boson would certainly be one of the most exciting developments in particle physics to date and it is tempting assume this new resonance is indeed the Higgs, but establishing the true nature of this excess as the SM Higgs must still proceed with due diligence. The two main theoretical inputs in performing such a Higgs identification are the Standard Model branching fractions for each of the decay channels used in the combination and the overall Higgs production cross section. We highlight that Higgs production from gluon fusion, the dominant production mode at hadron colliders, occurs via loops of SM quarks and is hence uniquely sensitive to New Physics (NP) effects arising from new colored states [@Spira:1995rr; @Binoth:1996au; @DiazCruz:2000yi; @Manohar:2006gz; @Muhlleitner:2006wx; @Djouadi:2007fm; @Bonciani:2007ex; @Arnesen:2008fb; @Low:2009di; @Bouchart:2009vq; @Boughezal:2010kx; @Azatov:2010pf; @Boughezal:2011mh; @Ruan:2011qg; @Burgess:2009wm; @Bai:2011aa; @Dobrescu:2011aa; @Belanger:2012zg; @Ilisie:2012cc; @Carena:2012fk; @Azatov:2012rj; @Azatov:2011qy; @Berger:2012ec] or more general Higgs portal [@Schabinger:2005ei; @Barbieri:2005ri; @Patt:2006fw; @Bowen:2007ia; @Englert:2011yb; @Cheung:2011aa; @Arnold:2009ay; @Djouadi:2011aa; @Batell:2011pz; @LopezHonorez:2012kv; @Cohen:2012zz; @Englert:2012ha; @Djouadi:2012zc] and scalar mixing effects [@Ignatiev:2000yy; @Gupta:2011gd; @Frank:2012nb; @Heckman:2012nt; @deSandes:2011zs]. The $\gamma \gamma$ branching fraction, which arises at loop-level in the Standard Model [@Shifman:1979eb], is similarly sensitive to NP effects [@Kribs:2007nz; @Carena:2011aa; @Christensen:2012ei; @Casagrande:2010si; @Goertz:2011hj]. For example, in the well-studied four generation Standard Model (SM4), gluon fusion rates are enhanced while the branching ratio to diphotons is suppressed [@Arik:2002ci; @Kribs:2007nz; @Schmidt:2009kk; @Li:2010fu; @Anastasiou:2011qw; @Passarino:2011kv; @Denner:2011vt; @Guo:2011ab; @He:2011ti; @Djouadi:2012ae; @Kuflik:2012ai; @Belotsky:2002ym]. Moreover, the landmark discovery of a Standard Model Higgs crucially relies on affirming the hypothesis of the Higgs mechanism for spontaneous breaking of $SU(2)_L \times U(1)_Y$ gauge symmetry, which predicts the existence of the Higgs boson. This fundamental test can only come after directly measuring the Higgs couplings to the $SU(2)$ gauge bosons. The fact that the Higgs is responsible for chiral symmetry breaking in the Standard Model and hence gives fermion masses is only a byproduct of the Higgs mechanism, and thus, in particular, the gluon fusion Higgs production mode does not directly probe the Higgs mechanism. This implies that NP could still be hiding in the gluon fusion process without effecting EWSB. This crucial point, which has also been emphasized in several recent papers [@Bai:2011aa; @Cheung:2011aa; @Batell:2011pz; @Dobrescu:2011aa; @Englert:2012ha], means that the LHC Higgs searches can be skewed by the presence of new colored particles which positively or negatively contribute to gluon fusion. Moreover, the excess in the data should be interpreted not only in the context of a SM Higgs, but also in the more exciting scenario of a possible new scalar state which arises from a Higgs portal-induced mixing between the SM Higgs and a new scalar. We demonstrate that extended color sectors involving new colored particles will generally give rise to both effects. In particular, if the new colored particles do not get their mass from the vacuum expectation value (vev) of the Higgs boson, then a generic Higgs portal term can give rise to Higgs mixing. We can see that direct Higgs coupling and Higgs portal-induced scalar mixing are two important categories of NP contributions that can have marked effects on Higgs collider signals, and thus we consider them simultaneously. Motivated by the possibility of probing new colored states via gluon fusion, we adopt a building block approach for an arbitrary NP model. Namely, we isolate and calculate the gluon fusion amplitude for new colored scalars, fermions, and vectors. In the case of the colored vector, we present the calculation in the context of the renormalizable coloron model (ReCoM) [@Simmons:1996fz; @Hill:2002ap; @Bai:2010dj] such that the concomitant effects from maintaining UV consistency can be readily included. We also allow for Higgs mixing, where the SM Higgs is mixed with a new scalar. In addition, for a mild and well-motivated set of assumptions, we give generic expressions for branching ratios of the scalar mass eigenstates into the most sensitive SM Higgs decay modes. Since our work has some overlap with many studies in the literature, we survey several representative papers and elaborate on the differences. A few recent papers have focused on our first category of NP effects for $gg \rightarrow h$ in which the NP states couple directly to the SM Higgs. In particular, the authors of [@Cheung:2011aa] focused solely on the situation where new particle masses arise from the Higgs vev, which simultaneously sharpens their discussion of resulting gluon fusion and diphoton decay phenomenology and limits the breadth of their conclusions. Separately, the authors of [@Dobrescu:2011aa] focused on Higgs portal phenomenology with new colored scalars, while the work in [@Arnold:2009ay; @Batell:2011pz] also included new scalars transforming under the full SM gauge symmetry. A similar study, emphasizing the constraints from electroweak precision fits, was performed in [@Burgess:2009wm; @Bai:2011aa]. We go beyond these direct coupling studies by also including the effects of a colored vector. There have also been a number of recent studies [@Plehn:2012iz; @Low:2012rj] where fits are performed in order to determine how consistent the data is with a SM Higgs hypothesis. These studies find that generally the excess is largely consistent with a SM Higgs with a tantalizing, but small enhancement in the $\gamma \gamma$ channel. However, these fits still have large uncertainties, and the next data set can change the picture drastically. Also, the Higgs couplings to bosons rely heavily on the vector boson fusion channel, which has large fluctuations between 7 and 8 TeV. These uncertainties leave room for modifications to the coupling of the SM Higgs to gluons which can be either enhanced or suppressed given the sign of the Higgs portal term. We will examine this in detail below. Regarding our second category of NP effects, when the SM Higgs mixes with a new scalar via a Higgs portal term, a majority of the literature has focused on the case where the New Physics sector is completely invisible to the SM , providing a possible connection to the dark matter. In this situation, as we will see in [Sec. \[sec:HiggsPortal\]]{}, only a simple mixing angle is needed to parametrize the effects on Higgs phenomenology, if no new decays are kinematically allowed. Our work considers the more complicated scenario where the new scalar couples to new colored particles, similar to [@Batell:2011pz], as mentioned above. In addition to these renormalizable scenarios of NP effects on loop-induced SM Higgs phenomenology, a few papers have followed an effective field theory approach by constructing and constraining the size of dimension-six operators. In [@Manohar:2006gz], the authors focused on the coefficients and constraints of operators for $h \rightarrow \gamma \gamma$, $\gamma Z$, and $gg$, while [@Low:2009di] extended the discussion to include $h \rightarrow f \bar{f}$ as well. Importantly, both of these studies assume any New Physics contributions are heavy enough to be integrated out, thus there are no new particles in the low energy spectrum. In contrast to the previous literature, therefore, we discuss the general case using renormalizable interactions when both categories of NP effects are present. We isolate contributions with new colored scalars, new colored fermions, including Standard Model quark mixing, and new colored vectors, and we allow such effects to be modified by Higgs mixing. The paper is organized as follows. In [Sec. \[sec:HiggsPortal\]]{}, we discuss general aspects of the Higgs portal relevant for our analysis of Standard Model Higgs production from gluon fusion. In [Sec. \[sec:SM\]]{}, we briefly review the leading order $gg\rightarrow h$ calculation for the Standard Model as well as the trivial extension of adding a fourth generation. In [Sec. \[sec:Scalar\]]{}, we discuss gluon fusion in the presence of a new colored scalar. In [Sec. \[sec:Fermion\]]{}, we present the analogous calculation for a general new colored fermion. Lastly, in [Sec. \[sec:Vector\]]{}, we discuss the interesting case of a new colored vector and its effects on gluon fusion in the context of a UV-complete, renormalizable model. Details of this calculation are presented in [Appendix \[sec:Vector\_explicit\]]{}. We summarize and conclude in [Sec. \[sec:Conclusion\]]{}. The Higgs Portal and Higgs Mixing {#sec:HiggsPortal} ================================= In this section, we review the Higgs portal as a general framework for studying the connection between arbitrary New Physics models and Higgs physics, with a special emphasis on the resulting effect on gluon fusion. In the SM, the Higgs field is responsible for breaking $SU(2)_L \times U(1)_Y$ gauge symmetry, resulting in masses for the $W^\pm$ and $Z$ bosons as well as the chiral SM fermions. By virtue of being the only scalar field present in the SM, the Higgs also generates $H^\dagger H$, which is the lowest mass dimension operator possible in the SM that is both gauge and Lorentz invariant. Hence, arbitrary NP operators can then be tacked on to $H^\dagger H$ to give $$\label{eqn:LOhp} \mathcal{L}_{hp} \supset \lambda_{hp} \mathcal{O}_{NP} H^\dagger H \ .$$ Although $\mathcal{O}_{NP}$ can be an arbitrarily high dimension operator, with an appropriate power suppression from a high scale $\Lambda_{NP}$, a generic Higgs portal term is only typically unsuppressed when $\mathcal{O}_{NP}$ itself is dimension two and gauge and Lorentz invariant: hence, we take $\mathcal{O}_{NP} \sim \Phi^\dagger \Phi$. One exception is the case when a new scalar field is a pure SM and NP gauge singlet, but since we are focused on NP effects on gluon fusion, we will not discuss the gauge singlet case further. One class of NP effects on gluon fusion arises from new colored states that directly enter the $gg \rightarrow h$ loop diagram. The direct coupling of colored states to the Higgs via [Eq. (\[eqn:LOhp\])]{} implies the mass of the new state is shifted after electroweak symmetry breaking (EWSB), and as this direct coupling is turned off, the NP effect vanishes. This class of effects is typified by models with new colored scalars, but a new fermion with Yukawa-like couplings to the SM Higgs boson also follows this scheme, albeit not via the Higgs portal. Although the case where the mass of the NP state arises primarily from the Higgs vev was discussed in [@Cheung:2011aa], in our more general framework the NP mass scale and the new couplings to the SM Higgs are independent. Since new particle masses do not have to arise from the SM Higgs vev, a second broad class of NP effects on gluon fusion emerges. Namely, if a new scalar field obtains a vev to spontaneously break a new gauge symmetry and if a Higgs portal term is present, this new scalar field will mix with the SM Higgs. In this class, NP effects coming from new colored states can infiltrate gluon fusion through the mixing induced from the Higgs portal even if these states do not directly couple to the SM Higgs. These effects will also exhibit the familiar non-decoupling features in SM $gg \rightarrow h$ loop calculations by chiral fermions or $h \rightarrow \gamma \gamma$ loop calculations by $W$ bosons if the analogous NP states are present [@Shifman:1979eb; @Shifman:2011ri; @Marciano:2011gm]: however, this non-decoupling feature only applies to the new scalar field component of the scalar mass eigenstates. As mentioned in the introduction, we allow for both direct and Higgs mixing mediated categories of NP effects to be present simultaneously. These effects arise in many extended color sector models, and we consider isolated new colored scalars, fermions, and vectors in turn. For colored scalars, we couple them to the Higgs via the Higgs portal in [Eq. (\[eqn:LOhp\])]{}, and hence they will exhibit an example of the direct category of NP effects with $\lambda_{hp}$ as the direct coupling. For colored fermions, we consider two subcategories distinguished by the possibility of SM fermion mixing. If new fermions are introduced that mix with SM fermions, the usual SM calculation is modified to accommodate fermion mass eigenstates that do not typically couple with the SM Higgs with the usual Yukawa strength. Without such fermion mixing, the SM calculation is unchanged and the new contribution arises from direct Yukawa couplings to the Higgs, the new scalar, or both. The decoupling behavior of new colored fermions are parametrized by fermion mixing angles and the possible scalar mixing angle. Perhaps the most interesting case is that of a massive colored vector boson. Here, in order to have a theory which is tree level unitary [@Cornwall:1974km], it is natural to consider an extended color symmetry which is then spontaneously broken to $SU(3)_c$ gauge symmetry. Then the massive vectors corresponding to the broken generators form representations of the unbroken color symmetry. We are thus left with a renormalizable, unitary, spontaneously broken gauge theory [@Fujikawa:1972fe]. We remark that another class of New Physics effects via the Higgs portal operator is possible. Broadly speaking, at the renormalizable level, where $\mathcal{O}_{NP} \sim \Phi^\dagger \Phi$ in [Eq. (\[eqn:LOhp\])]{}, one class of Higgs portal effects is characterized by new colored scalars which do not obtain vevs. The second class is driven by new uncolored scalars that do obtain vevs from their scalar potential. Another possibility is colored scalars that do obtain vevs, but such color-breaking vacua are not viable phenomenologically. The last possibility consists of new uncolored scalars that do not obtain vevs from their scalar potential. Such a scalar does not enter the $gg \rightarrow h$ loop, but if $\lambda_{hp}$ is large and positive, the resulting Higgs portal-induced shift in mass squared, $-\lambda_{hp} v_h^2 / 2$ ($v_h$ is the Higgs vev) could drive the new scalar to acquire a vev. Hence, this last category of portal symmetry breaking models is unique because the Higgs portal coupling is a necessary ingredient for driving the new scalar to obtain a nonzero vev. Obviously, the roles of the new scalar and the Higgs scalar can be reversed, whereby the Higgs portal term allows a new scalar vev to drive the Higgs field to obtain a negative mass squared and hence trigger EWSB. We reserve a study of “Portal Symmetry Breaking” phenomenology for future work. Also, in the discussion above, we have delineated cases according to specific constraints on the Lagrangian parameters. A precise determination of these bounds would require an analysis of renormalization group evolution, which is beyond the scope of this work. New Physics Scalar – Standard Model Higgs Mixing {#subsec:mixing} ------------------------------------------------ We briefly discuss the second class of NP effects from the Higgs portal described above, [*i.e.*]{} a new scalar and the SM Higgs both obtain vevs in [Eq. (\[eqn:LOhp\])]{} and mix. For simplicity, we only consider one new scalar, but our discussion is readily generalized to multiple scalars. We also assume $\Phi$ transforms as a singlet under $SU(2)_L\times U(1)_Y$, but that it is charged under a new local or global symmetry in order to prevent “tadpole” terms. We let $\mathcal{O}_{NP} \sim \Phi^\dagger \Phi$ for a new scalar field $\Phi$, giving $$\label{eqn:Lhp} \mathcal{L} \supset \lambda_{hp} H^\dagger H \Phi^\dagger \Phi \sim \lambda_{hp} v_h v_\phi h \phi \ ,$$ where we have suppressed representation indices and expanded the fields $H \sim \dfrac{1}{\sqrt{2}} (h + v_h)$ and $\Phi \sim \dfrac{1}{\sqrt{2}} (\phi + v_{\phi})$. We assume the scalar potentials $V(\Phi)$ and $V(H)$ are also present and [Eq. (\[eqn:Lhp\])]{} is the only Lagrangian term involving both $\Phi$ and $H$ fields. The usual stability, triviality, and renormalizability constraints on the full scalar potential $V(H) + V(\Phi) - \lambda_{hp} |H|^2 |\Phi|^2$ are assumed to be satisfied and will be imposed when we consider explicit models in [Secs. \[sec:Fermion\] and \[sec:Vector\]]{}. Here, since $\Phi$ obtains a vev, [Eq. (\[eqn:Lhp\])]{} leads to mixing via the mass matrix $$\label{eqn:scalar_mass_matrix} m_{\text{scalar}}^2 = \left( \begin{array}{cc} m_h^2 & -\lambda_{hp} v_h v_{\phi} \\ -\lambda_{hp} v_h v_{\phi} & m_{\phi}^2 \\ \end{array} \right) \ ,$$ where $v_h$ and $v_{\phi}$ are calculated from minimizing the full potential $V(H) + V(\Phi) - \lambda_{hp} |\Phi|^2 |H|^2$ and hence determine $m_h$ and $m_{\phi}$. The functional dependence of $m_h$ and $m_{\phi}$ on their respective potential parameters can be fixed by solving the potentials $V(H)$ and $V(\Phi)$ separately, and in the limit that $\lambda_{hp} \rightarrow 0$, the exact vevs $v_h$ and $v_{\phi}$ recover their original, unperturbed values. This observation has important ramifications when calculating the exact Goldstone–Goldstone–scalar couplings needed for vector loop amplitudes in Feynman gauge, which are discussed in [Subsec. \[subsec:GoldstoneSector\]]{}. We can readily diagonalize the symmetric mass matrix [Eq. (\[eqn:scalar\_mass\_matrix\])]{} to obtain the mass eigenstates $$\label{eqn:s1s2_definition} \begin{array}{ccc} s_1 &=& h \cos \theta - \phi \sin \theta \ , \\ s_2 &=& h \sin \theta + \phi \cos \theta \ , \\ \end{array}$$ with a Jacobi rotation mixing angle $\theta$ defined by $$\label{eqn:mixingangle} \tan 2\theta = \dfrac{ -2 \lambda_{hp} v_h v_{\phi}}{m_{\phi}^2 - m_h^2} \ .$$ We will also need the inverse operations, $$\label{eqn:hphi_definition} \begin{array}{ccc} h &=& s_1 \cos \theta + s_2 \sin \theta \ , \\ \phi &=& -s_1 \sin \theta + s_2 \cos \theta \ . \\ \end{array}$$ The eigenvalues of [Eq. (\[eqn:scalar\_mass\_matrix\])]{} are $$\label{eqn:Ms1} m^2_{s_1} = \dfrac{1}{2} \left( m_h^2 + m_\phi^2 \right) - \dfrac{1}{2} \sqrt{ \left( -m_h^2 + m_\phi^2 \right)^2 + 4 \lambda_{hp}^2 v_h^2 v_\phi^2 } \ ,$$ and $$\label{eqn:Ms2} m^2_{s_2} = \dfrac{1}{2} \left( m_h^2 + m_\phi^2 \right) + \dfrac{1}{2} \sqrt{ \left( -m_h^2 + m_\phi^2 \right)^2 + 4 \lambda_{hp}^2 v_h^2 v_\phi^2 } \ ,$$ where we have taken $m_{s_1} < m_{s_2}$ without loss of generality. As mentioned before and demonstrated in [@Schabinger:2005ei; @Bowen:2007ia; @Batell:2011pz], the mixing of the scalar states from the Higgs portal can significantly affect scalar production via gluon fusion. Moreover, the mixing is driven purely by the strength of $\lambda_{hp}$, which must be real but whose sign is not fixed. New Physics Effects on Production of $s_{1,2}$ {#subsec:general_amplitudes} ---------------------------------------------- We can now readily disentangle the two categories of New Physics effects on gluon fusion. Now, because of $h$–$\phi$ mixing via the Higgs portal in [Eq. (\[eqn:Lhp\])]{}, we must calculate cross sections for $gg \rightarrow s_1$ and $gg \rightarrow s_2$ production instead of the gauge eigenstates $h$ and $\phi$. Since both $h$ and $\phi$ can couple to new colored particles, contributions to $gg \rightarrow s_{1,2}$ can manifest themselves through both the $h$ and $\phi$ components of $s_{1,2}$, leading to suppression or enhancement of the production rate relative to the SM. This also implies that partial decay widths are affected, whereas in hidden sector models, such widths are unaltered apart from a universal $\cos^2 \theta$ suppression coming from Higgs mixing. From the discussion above, we can decompose the production amplitude of $s_1$ via gluon fusion in terms of the gauge eigenstate $h$ and $\phi$ production amplitudes as, $$\label{eqn:general_s1_s2_amps} \begin{array}{ccc} \mathcal{M} (gg \rightarrow s_1) &=& \left. c_\theta \left[ \mathcal{M} (gg \rightarrow h) \right] \right|_{m_h = m_{s_1}} \left. -s_\theta \left[ \mathcal{M} (gg \rightarrow \phi) \right] \right|_{m_\phi = m_{s_1}} \\ \mathcal{M} (gg \rightarrow s_2) &=& \left. s_\theta \left[ \mathcal{M} (gg \rightarrow h) \right] \right|_{m_h = m_{s_2}} \left. +c_\theta \left[ \mathcal{M} (gg \rightarrow \phi) \right] \right|_{m_\phi = m_{s_2}} \ , \\ \end{array}$$ where $c_\theta \equiv \cos \theta$, $s_\theta \equiv \sin \theta$ are defined by [Eq. (\[eqn:mixingangle\])]{}. In the discussion below, we presume the matrix elements are evaluated at the appropriate scalar mass and will drop the notation above. Hence, given the linear combination dictated by [Eq. (\[eqn:general\_s1\_s2\_amps\])]{}, we are now free to isolate the contributions to $gg \rightarrow h$ and $gg \rightarrow \phi$. We are particularly interested in identifying, at the amplitude level, the mechanisms responsible for modifying gluon fusion and whether and how they can decouple. A completely general expression for all possible NP effects along these lines is cumbersome, so instead we write $$\label{eqn:general_s1_sfv_amp} \begin{array}{ccl} \vspace{4pt} \mathcal{M} (gg \rightarrow s_1) &=& c_\theta \left[ \mathcal{M} (gg \xrightarrow[scalars]{} h) + \mathcal{M} (gg \xrightarrow[fermions]{} h) + \mathcal{M} (gg \xrightarrow[vectors]{} h) \right] \\ &-& s_\theta \left[ \mathcal{M} (gg \xrightarrow[scalars]{} \phi) + \mathcal{M} (gg \xrightarrow[fermions]{} \phi) + \mathcal{M} (gg \xrightarrow[vectors]{} \phi) \right] \ , \\ \end{array}$$ and treat each category of loop particles separately.[^1] Each of these categories can be further subdivided into particles that couple solely to $h$, solely to $\phi$, or simultaneously to both. In the scalar case, for example, we can write $$\label{eqn:general_s1_scalar_amp} \begin{array}{ccl} \vspace{4pt} \mathcal{M} (gg \xrightarrow[scalars]{} s_1) &=& c_\theta \left[ \sum\limits_i \mathcal{M} (gg \xrightarrow[\eta_i]{} h) + \sum\limits_j \mathcal{M} (gg \xrightarrow[\eta_j]{} h) \right] \\ &-& s_\theta \left[ \sum\limits_j \mathcal{M} (gg \xrightarrow[\eta_j]{} \phi) + \sum\limits_k \mathcal{M} (gg \xrightarrow[\eta_k]{} \phi) \right] \ , \\ \end{array}$$ where the scalars $\eta_i$, $\eta_j$, $\eta_k$ couple only to $h$, both to $h$ and $\phi$, and only to $\phi$, respectively. We can now make definitive statements about the decoupling behavior of the scalars $\eta_i$, $\eta_j$ and $\eta_k$. If the masses of $\eta_i$ ($\eta_k$) arise solely from the vev $v_h$ ($v_\phi$), then these scalars will exhibit non-decoupling from $h$ ($\phi$) as their masses are taken very large: if instead their masses include sources besides $v_h$ or $v_{\phi}$, then decoupling will occur as the mass scale of these new sources is taken large. The behavior of the $\eta_j$ states are a straightforward combination of the previous arguments. For fermions, we write $$\label{eqn:general_s1_fermion_amp} \begin{array}{ccl} \vspace{4pt} \mathcal{M} (gg \xrightarrow[fermions]{} s_1) &=& c_\theta \left[ \sum\limits_i \mathcal{M} (gg \xrightarrow[\psi_i]{} h) + \sum\limits_j \mathcal{M} (gg \xrightarrow[\psi_j]{} h) \right] \\ &-& s_\theta \left[ \sum\limits_j \mathcal{M} (gg \xrightarrow[\psi_j]{} \phi) + \sum\limits_k \mathcal{M} (gg \xrightarrow[\psi_k]{} \phi) \right] \ . \\ \end{array}$$ To be more illustrative, we can take some familiar examples to demonstrate the flexibility of [Eq. (\[eqn:general\_s1\_fermion\_amp\])]{}. In the case with Higgs mixing but without new fermions $\psi_j$ or $\psi_k$, then $\psi_i$ consists of the SM quarks and we get a universal $c_\theta$ suppression of the matrix element. If instead we only add a new vector-like top partner to the SM, then $c_\theta = 1$, $s_\theta = 0$, and $\psi_i$ includes the first five SM quarks and the two fermion mass eigenstates resulting from top mixing while the $\psi_j$ and $\psi_k$ sums are absent. Finally, if Higgs mixing is present and new colored fermions are added that couple both to $h$ and $\phi$ but do not mix with the SM fermions, then $\psi_i$ will run over the SM quarks and $\psi_j$ will run over the NP colored fermions. Lastly, we can introduce massive colored vectors. We will only consider the case where these vectors couple to $\phi$, giving the relatively simple expression $$\label{eqn:general_s1_vector_amp} \begin{array}{ccl} \mathcal{M} (gg \xrightarrow[vectors]{} s_1) &=& -s_\theta \left[ \sum\limits_k \mathcal{M} (gg \xrightarrow[V_k]{} \phi) \right] \ , \end{array}$$ emphasizing that this contribution to the gluon fusion rate for $s_1$ production relies on the Higgs portal, since the SM Higgs is assumed to play no role in breaking the extended color gauge symmetry. After the above discussion, we present a parametric understanding of how production and decays of $s_{1,2}$ are affected by direct coupling and $h$–$\phi$ mixing. As we have seen, performing a completely general analysis would be overly cumbersome, and so we will make a few mild assumptions to make the analysis more intuitive and tractable. Throughout the discussion, we assume a narrow width approximation, allowing us to factorize production and decay processes. We define the overall leading order enhancement or suppression factor of $s_1$ production relative to SM Higgs production via gluon fusion as $$\label{eqn:epsilon_definition} \epsilon_{gg} \equiv \dfrac{\sigma (gg \rightarrow s_1)}{\sigma (gg \xrightarrow[SM]{} h)} = \dfrac{ \left| \mathcal{M} (gg \rightarrow s_1) \right|^2 }{ \left| \mathcal{M} (gg \xrightarrow[SM]{} h) \right|^2 } = \dfrac{ \left| c_\theta \mathcal{M} (gg \rightarrow h) - s_\theta \mathcal{M} (gg \rightarrow \phi) \right|^2}{ \left| \mathcal{M} (gg \xrightarrow[SM]{} h) \right|^2 } = c^2_\theta\left| \mathcal{Z}_{ggh} - t_\theta \mathcal{Z}_{gg \phi} \right|^2 \ ,$$ using [Eq. (\[eqn:general\_s1\_s2\_amps\])]{} and with $t_\theta = \tan \theta$. The complex amplitude ratios are given by $$\label{eqn:Zgg_definition} \mathcal{Z}_{ggh} \equiv \dfrac{\mathcal{M} (gg \rightarrow h)}{\mathcal{M} (gg \xrightarrow[SM]{} h)} \qquad \mathcal{Z}_{gg\phi} \equiv \dfrac{\mathcal{M} (gg \rightarrow \phi)}{\mathcal{M} (gg \xrightarrow[SM]{} h)} \ ,$$ and will simplify significantly for any given NP model under consideration, as we will demonstrate in [Secs. \[sec:Scalar\], \[sec:Fermion\] and  \[sec:Vector\]]{}. We see that both $\epsilon_{gg} > 1$ (signaling enhancement) and $\epsilon_{gg} < 1$ (signaling suppression) are possible with New Physics and changing the sign of $\lambda_{hp}$. In the limit that $\theta = 0$, the only effect on gluon fusion arises from the inclusion of new colored states that directly couple to the SM Higgs, which was a main focus of [@Cheung:2011aa; @Bai:2011aa; @Dobrescu:2011aa]. In the case where Higgs mixing is the only new effect, then $\mathcal{Z}_{ggh} = 1$ and $\mathcal{Z}_{gg\phi} = 0$, and we have the simple expression $\epsilon_{gg} = c^2_\theta$, as noted in [@Schabinger:2005ei]. We remark that complete suppression of gluon fusion does not correspond to vanishing LHC production for the $s_1$ state. This is because the subdominant modes of vector boson fusion, vector boson association, and $t \overline{t} h$ production comprise 12.5% of the total cross section for a SM Higgs mass at 125 GeV [@Dittmaier:2011ti]. Moreover, even if the leading order cancellation in [Eq. (\[eqn:epsilon\_definition\])]{} is exact, we expect NLO corrections, which can be as large as 20% in the case of colored stops [@Muhlleitner:2006wx], to make the cancellation imperfect. New Physics Effects on Decays of $s_{1,2}$ {#subsec:s12_decays} ------------------------------------------ We now extend our discussion to include NP effects on decay widths for our scalar state $s_1$, which we take to be dominantly SM Higgs-like. We will not detail all of the (practically infinite!) possible final states for $s_1$, but will instead focus on the $WW$, $ZZ$, $\gamma \gamma$, $b \overline{b}$ and $\tau^+ \tau^-$ decay channels. For the $WW$ final state, we write $$\label{eqn:s1_MEtoWW} \mathcal{M} (s_1 \rightarrow WW) = c_\theta \mathcal{M} (h \rightarrow WW) -s_\theta \mathcal{M} (\phi \rightarrow WW) \approx c_\theta \mathcal{M} (h \rightarrow WW) \ ,$$ and thus $$\label{eqn:s1_BRtoWW} \dfrac{\mathcal{B}(s_1 \rightarrow WW)}{\mathcal{B} (h \xrightarrow[SM]{} WW)} \approx c_\theta^2 \dfrac{\Gamma_h}{\Gamma_{s_1}} \ ,$$ where we have assumed the tree-level coupling of $h WW$ dominates the (typically loop-induced) coupling of $\phi WW$, and $\Gamma_h$ and $\Gamma_{s_1}$ are the total width of the purely SM Higgs and the mass eigenstate $s_1$, respectively. Under the same assumption that $h ZZ$ dominates the $\phi ZZ$ coupling, the same result in [Eq. (\[eqn:s1\_BRtoWW\])]{} also applies to the $ZZ$ final state, and so branching ratios of $s_1$ to $WW$ or $ZZ$ diboson states are typically suppressed in Higgs mixing models. For the diphoton final state, we can adapt our gluon fusion discussion, replacing colored particles with electromagnetically charged particles. Following the guide of [Eq. (\[eqn:general\_s1\_s2\_amps\])]{}, this gives $$\label{eqn:s1_MEtogammagamma} \begin{array}{ccc} \mathcal{M} (s_1 \rightarrow \gamma \gamma) = \left. c_\theta \left[ \mathcal{M} (h \rightarrow \gamma \gamma) \right] \right|_{m_h = m_{s_1}} - \left. s_\theta \left[ \mathcal{M}(\phi \rightarrow \gamma \gamma) \right] \right|_{m_\phi = m_{s_1}} \ . \end{array}$$ Unlike the $WW$ or $ZZ$ decay modes, the $h \rightarrow \gamma \gamma$ decay is induced at loop level in the SM and new contributions can easily cancel against or add to the SM contributions. Using [Eq. (\[eqn:s1\_MEtogammagamma\])]{}, we can write the relative branching ratio as, $$\label{eqn:s1_BRtogammagamma} \dfrac{ \mathcal{B}(s_1 \rightarrow \gamma \gamma)}{ \mathcal{B}(h \xrightarrow[SM]{} \gamma \gamma)} = \epsilon_{\gamma\gamma} \dfrac{\Gamma_h}{\Gamma_{s_1}} \ ,$$ where $\epsilon_{\gamma\gamma}$ is analogous to $\epsilon_{gg}$ in [Eq. (\[eqn:epsilon\_definition\])]{} and $\mathcal{Z}_{h\gamma\gamma}$ and $\mathcal{Z}_{\phi\gamma\gamma}$ are defined similarly. The relative rate for $gg \rightarrow s_1 \rightarrow \gamma \gamma$ versus $gg \rightarrow h \rightarrow \gamma \gamma$ is now given by $$\label{eqn:totalrate} \mathcal{R} = \epsilon_{gg}\epsilon_{\gamma\gamma} \dfrac{\Gamma_h}{\Gamma_{s_1}} \ .$$ In many models, though, the various inputs for [Eq. (\[eqn:totalrate\])]{} reduce to simple expressions. For example, in Higgs mixing scenarios where $\phi$ only couples to hidden sector particles, we obtain $\mathcal{Z}_{h\gamma \gamma} = 1$, $\mathcal{Z}_{\phi \gamma \gamma} = 0$, and so $$\label{eqn:s1_BRtogammagamma2} \dfrac{ \mathcal{B}(s_1 \rightarrow \gamma \gamma)}{ \mathcal{B}(h \xrightarrow[SM]{} \gamma \gamma)} = c_\theta^2 \dfrac{\Gamma_h}{\Gamma_{s_1}} \ ,$$ which agrees with the universal $c_\theta^2$ suppression noted in [@Schabinger:2005ei]. Another simple limiting case arises if we take $\theta = 0$ and introduce new charged particles in the $\gamma \gamma$ loop coupling to the Higgs. In this case, $h \equiv s_1$ and we can write $$\label{eqn:s1_BRtogammagamma3} \dfrac{ \mathcal{B}(s_1 \rightarrow \gamma \gamma)}{ \mathcal{B}(h \xrightarrow[SM]{} \gamma \gamma)} = \dfrac{\Gamma_h}{\Gamma_{s_1}} \left| \mathcal{Z}_{h\gamma\gamma} \right|^2 \ ,$$ so that only the direct NP effects contribute. Finally, we can calculate the $s_1$ branching ratio to $b \bar{b}$ or $\tau^+ \tau^-$. If Higgs mixing is present, if $\phi$ does not appreciably couple to the SM fermions, and if the SM fermions are not mixed with NP fermions, then the same results from [Eq. (\[eqn:s1\_BRtoWW\])]{} apply, substituting $f \bar{f}$ for $WW$. A completely general expression, however, because of the possible presence of all of these effects, is unwieldy. As an explicit case, for the $b \bar{b}$ final state, if we allow for $h$–$\phi$ mixing and introduce a coupling between $\phi$ and $b \bar{b}$, we obtain $$\label{eqn:s1_MEtoffbar} \dfrac{ \mathcal{B}(s_1 \rightarrow b\bar{b})}{ \mathcal{B}(h \xrightarrow[SM]{} b\bar{b})} = c_\theta^2 \dfrac{\Gamma_h}{\Gamma_{s_1}} \left| 1 - t_\theta \mathcal{Z}_{\phi b\bar{b}} \right|^2 \ ,$$ where $\mathcal{Z}_{\phi b\bar{b}} = \mathcal{M} (\phi \rightarrow b \bar{b}) / \mathcal{M} (h \xrightarrow[SM]{} b \bar{b})$. We can see that interference effects from $\mathcal{Z}_{\phi b\bar{b}}$, although diluted by $t_\theta$, can lead to an overall increase in the branching fraction of $s_1 \rightarrow b \bar{b}$. In summary, we have disentangled the effects from Higgs portal-induced mixing of $h$ and $\phi$ from NP effects caused by direct coupling to $h$, $\phi$, or both. For gluon fusion, we have explicitly identified the decoupling behavior of New Physics states in [Eqs. (\[eqn:general\_s1\_scalar\_amp\]), (\[eqn:general\_s1\_fermion\_amp\]) and (\[eqn:general\_s1\_vector\_amp\])]{}. If we assume NP couplings to be small or negligible, then the resulting $s_1$ branching ratio has a universal $c_\theta^2$ suppression and a universal total width ratio suppression. On the other hand, interference effects resulting from couplings to $h$ and/or $\phi$ lead to a myriad of effects and possibilities for both suppression and enhancement of relative rates can be achieved. We note that all of these expressions can readily be adapted for $s_2$ decay with an appropriate $c_\theta \rightarrow s_\theta$, $-s_\theta \rightarrow c_\theta$ exchange and $m_{s_1} \rightarrow m_{s_2}$. In addition, if $m_{s_2} > 2 m_{s_1}$, there is the additional decay mode $s_2 \rightarrow s_1 s_1$, as emphasized in [@Bowen:2007ia]. Also, if any of the new states are lighter than $m_{s_1} /2$ or $m_{s_2} / 2$, then additional non-standard decay modes open up. This effect is manifest in the above expressions through the ratio of total widths $\Gamma_h / \Gamma_{s_1}$. The $gg\rightarrow h$ Process in SM {#sec:SM} =================================== Here we briefly review the leading order Standard Model calculation for Higgs production via gluon fusion. As shown in [Fig. \[fig:SM\_ggh\]]{}, gluon fusion arises in the SM via quark loops, with the dominant contribution coming from the top quark with its large Yukawa coupling. We again highlight the fact that since neither the $W$ or $Z$ boson couplings are probed in this production mode, large effects can be present in this loop process that strongly change Higgs production but do not affect EWSB. ![\[fig:SM\_ggh\][The Standard Model contribution to $gg \rightarrow h$. ]{}](FigSMfermions.pdf) The total Standard Model amplitude is $$\label{eqn:fermionamp} i \mathcal{M}_{SM}^{ad} = \sum\limits_f i \mathcal{M}_f^{ad} = \sum\limits_f i \left( \dfrac{\alpha_s}{\pi} \right) \dfrac{C(r_f)}{2v_h} \delta^{ad} \epsilon_{1\mu} \epsilon_{2\nu} \left( p_1^\nu p_2^\mu - \dfrac{m_h^2}{2} g^{\mu \nu} \right) F_F(\tau_f) \ ,$$ where $f$ runs over the SM quarks, $C(r_f)$ is the Casimir invariant ($C(r_f) = 1/2$ for SM quarks), $a$ and $d$ are color indices, $p_1 \cdot p_2 = \dfrac{m_h^2}{2}$ for an on-shell Higgs, $\tau_f \equiv m_h^2 / (4 m_f^2)$ and $F_F(\tau)$ is the well-known loop function $$\label{eqn:F_F} F_F(\tau) = \dfrac{2}{\tau^2} \left( \tau + (\tau - 1) Z(\tau) \right) \ ,$$ using $$\label{eqn:Z_SM} Z(\tau) = \left\{ \begin{array}{c} \arcsin^2 \sqrt{\tau} \qquad \tau \leq 1 \\ \dfrac{-1}{4} \log \left[ \dfrac{ 1 + \sqrt{1 - \tau^{-1}} }{ 1 - \sqrt{1 - \tau^{-1}} } - i \pi \right]^2 \qquad \tau > 1 \ . \\ \end{array} \right.$$ Because the SM quarks obtain their mass purely from EWSB, they do not decouple even as we take the limit $\tau \rightarrow 0$ (equivalent to $m_f \rightarrow \infty$). In the case of the SM4, this sum would include $t'$ and $b'$ contributions. In the limit that the SM Higgs is well below the threshold for $t$, $t'$, and $b'$ to propagate on-shell in [Fig. \[fig:SM\_ggh\]]{}, we obtain the usual factor of 3 enhancement in the SM4 matrix element for $gg \rightarrow h$, which gives, at leading order, a factor of 9 enhancement for gluon fusion production in SM4 relative to SM3 [@Denner:2011vt]. Recent literature that has attempted to resolve the quandary of a sequential fourth generation of fermions with the lack of enhancement in gluon fusion include Refs. [@Schmidt:2009kk; @Guo:2011ab; @He:2011ti; @Djouadi:2012ae; @Kuflik:2012ai]. We can anticipate, in the presence of new additions to gluon fusion coming from New Physics, that the main structure of [Eq. (\[eqn:fermionamp\])]{} will not change apart from possible differences in the scalar vertex, $C(r)$, and the loop function $F(\tau)$. In particular, the $p_1^\nu p_2^\mu - p_1 \cdot p_2 g^{\mu \nu}$ structure of the matrix element is assured by $SU(3)_c$ gauge invariance (or equivalently, the Ward identity). This is analogous to the situation in the $h \rightarrow \gamma \gamma$ calculation, where electromagnetic gauge invariance requires the same momentum structure [@Marciano:2011gm]. New Colored Scalar {#sec:Scalar} ================== In this section, we isolate and calculate the effect of a colored complex scalar $S$ propagating in the $gg \rightarrow h$ loop. We use the Higgs portal in [Eq. (\[eqn:Lhp\])]{} to couple $S$ to the SM Higgs, and we write a (positive) tree-level mass squared for $S$ such that $SU(3)_c$ remains unbroken and Higgs mixing is absent. Depending on the sign and strength of $\lambda_{hp}$, we can achieve significant suppression or enhancement of gluon fusion as a result of the interference between the SM fermions and the colored scalar. The Lagrangian involving $S$ is $$\label{eqn:ScalarL} \mathcal{L}_S = \left|D_\mu S \right|^2 - m_0^2 S^\dagger S - \kappa |S^\dagger S|^2 + \lambda_{hp} S^\dagger S H^\dagger H \ ,$$ where color indices have been suppressed and we assume $m_0^2 > 0$ and $\kappa > 0$ to ensure stability. As discussed in [Sec. \[sec:HiggsPortal\]]{}, $\lambda_{hp}$ must be real: for positive (negative) $\lambda_{hp}$, we will get destructive (constructive) interference with the SM loop calculation, in agreement with [@Dobrescu:2011aa; @Batell:2011pz] (note we have a different sign convention for $\lambda_{hp}$). After EWSB, the physical scalar mass is $$\label{eqn:mS_definition} m_S^2 \equiv m_0^2 - \frac{1}{2} \lambda_{hp} v_h^2 \ ,$$ which imposes the constraint that $m_0^2 > \frac{1}{2} \lambda_{hp} v_h^2$ to avoid portal symmetry breaking of $SU(3)_c$. The two diagrams to calculate are shown in [Fig. \[fig:scalarloop\]]{}. Since $S$ is complex, the matrix element for [Fig. \[fig:scalarloop\]]{}A needs to be multiplied by 2 to account for the charge conjugate diagram: if $S$ were real, no factor of 2 is used and instead the matrix element for [Fig. \[fig:scalarloop\]]{}B must include a symmetry factor of (1/2). ![\[fig:scalarloop\][Feynman diagrams for scalar loop contributions to $gg \rightarrow h$. For a complex scalar one must also include the charge conjugate equivalent of diagram (A).]{}](scalarFigA.pdf "fig:"){width="45.00000%"} ![\[fig:scalarloop\][Feynman diagrams for scalar loop contributions to $gg \rightarrow h$. For a complex scalar one must also include the charge conjugate equivalent of diagram (A).]{}](scalarFigB.pdf "fig:"){width="45.00000%"} The total amplitude corresponding to the diagrams in [Fig. \[fig:scalarloop\]]{} for a complex scalar field propagating in the loop is $$\label{eqn:scalar_ME} i \mathcal{M}_S^{ad} = i \left( \dfrac{\alpha_s}{\pi} \right) \left( \dfrac{ C(r_S) \lambda_{hp} v_h }{4 m_S^2} \right) \delta^{ad} \epsilon_{1\mu} \epsilon_{2\nu} (p_1^\nu p_2^\mu - \dfrac{m_h^2}{2} g^{\mu \nu}) F_S(\tau_S) \ ,$$ where $C(r_S)$ is the $SU(3)_c$ Casimir invariant for $S$, $\tau_S = m_h^2 / (4 m_S^2)$ and the loop function $F_S$ is defined to be $$\label{eqn:F_S} F_S(\tau) = \tau^{-1} - \tau^{-2} Z(\tau) \ ,$$ with $Z(\tau)$ defined in [Eq. (\[eqn:Z\_SM\])]{}. Note that for fixed $\lambda_{hp}$ the amplitude decouples as $m_S \rightarrow \infty$ because of the tree-level mass squared $m_0^2$. Now, the summed amplitude for $\mathcal{M} (gg \xrightarrow[SM+S]{} h)$ is $$\label{eqn:SM3+S_ME} \begin{array}{ccl} i \mathcal{M}_{SM+S}^{ad} &=& i \left( \sum\limits_{f} \mathcal{M}_f^{ad} \right) + i \mathcal{M}_S^{ad} \\ &=& i \left( \dfrac{\alpha_s}{\pi} \right) \delta^{ad} \epsilon_{1\mu} \epsilon_{2\nu} (p_1^\nu p_2^\mu - \dfrac{m_h^2}{2} g^{\mu \nu}) \left( \sum\limits_{f} \left( \dfrac{C(r_f)}{2v_h} F_F(\tau_f) \right) + \dfrac{C(r_S) \lambda_{hp} v_h}{4 m_S^2} F_S(\tau_S) \right) \ . \\ \end{array}$$ If $m_S$, $m_t > m_h/2$, then $F_S$ is strictly real and negative and $F_F$ is strictly real and positive, which implies that for $\lambda_{hp} > 0$ ($\lambda_{hp} < 0$) the interference between the colored scalar amplitude and the SM is destructive (constructive). Since the phase space integration needed to calculate the $s_1$ cross section is identical to the SM Higgs case, we can write the ratio $\epsilon_{gg}$ from [Eq. (\[eqn:epsilon\_definition\])]{} as $$\label{eqn:SM+S_epsilon} \left. \epsilon_{gg} \right|_{SM+S} = \dfrac{ \left| \sum\limits_{f} \left( \dfrac{C(r_f)}{2v_h} F_F(\tau_f) \right) + \dfrac{C(r_s) \lambda_{hp} v_h}{4 m_S^2} F_S(\tau_S) \right|^2 }{ \left| \sum\limits_{f} \left( \dfrac{C(r_f)}{2v_h} F_F(\tau_f) \right) \right|^2 } \ .$$ ![\[fig:SM3LMsContour\_fixlhp\][Relative rate $\epsilon_{gg}$ in [Eq. (\[eqn:SM+S\_epsilon\])]{}, showing the effect of the inclusion of a real color octet scalar (top row) or complex color triplet scalar (bottom row), for $m_h = 125$ GeV (left panels) or $m_h = 500$ GeV (right panels). At the center of each plot, from top to bottom, the solid lines correspond to $\lambda_{hp} =-0.1$, $-0.025$, $0.025$, $0.1$. The left (right) gray bands in the octet scalar plots come from the ATLAS (CMS) search for pair produced dijet resonances. For the triplet case, the CMS bound still applies but the ATLAS bound is unconstraining after rescaling cross sections. ]{}](hprodColOctmh125GeV.pdf "fig:"){width="48.00000%"} ![\[fig:SM3LMsContour\_fixlhp\][Relative rate $\epsilon_{gg}$ in [Eq. (\[eqn:SM+S\_epsilon\])]{}, showing the effect of the inclusion of a real color octet scalar (top row) or complex color triplet scalar (bottom row), for $m_h = 125$ GeV (left panels) or $m_h = 500$ GeV (right panels). At the center of each plot, from top to bottom, the solid lines correspond to $\lambda_{hp} =-0.1$, $-0.025$, $0.025$, $0.1$. The left (right) gray bands in the octet scalar plots come from the ATLAS (CMS) search for pair produced dijet resonances. For the triplet case, the CMS bound still applies but the ATLAS bound is unconstraining after rescaling cross sections. ]{}](hprodColOctmh500GeV.pdf "fig:"){width="48.00000%"} ![\[fig:SM3LMsContour\_fixlhp\][Relative rate $\epsilon_{gg}$ in [Eq. (\[eqn:SM+S\_epsilon\])]{}, showing the effect of the inclusion of a real color octet scalar (top row) or complex color triplet scalar (bottom row), for $m_h = 125$ GeV (left panels) or $m_h = 500$ GeV (right panels). At the center of each plot, from top to bottom, the solid lines correspond to $\lambda_{hp} =-0.1$, $-0.025$, $0.025$, $0.1$. The left (right) gray bands in the octet scalar plots come from the ATLAS (CMS) search for pair produced dijet resonances. For the triplet case, the CMS bound still applies but the ATLAS bound is unconstraining after rescaling cross sections. ]{}](hprodColTriplet125GeV.pdf "fig:"){width="48.00000%"} ![\[fig:SM3LMsContour\_fixlhp\][Relative rate $\epsilon_{gg}$ in [Eq. (\[eqn:SM+S\_epsilon\])]{}, showing the effect of the inclusion of a real color octet scalar (top row) or complex color triplet scalar (bottom row), for $m_h = 125$ GeV (left panels) or $m_h = 500$ GeV (right panels). At the center of each plot, from top to bottom, the solid lines correspond to $\lambda_{hp} =-0.1$, $-0.025$, $0.025$, $0.1$. The left (right) gray bands in the octet scalar plots come from the ATLAS (CMS) search for pair produced dijet resonances. For the triplet case, the CMS bound still applies but the ATLAS bound is unconstraining after rescaling cross sections. ]{}](hprodColTriplet500GeV.pdf "fig:"){width="48.00000%"} We consider the addition of a real color octet scalar ($C(r_s)=3$, symmetry factor of $1/2$) and a complex color triplet scalar ($C(r_s) = 1/2$) and plot $\epsilon_{gg}$ in [Fig. \[fig:SM3LMsContour\_fixlhp\]]{} as a function of $m_S$ for some representative choices of parameters $m_h$ and $\lambda_{hp}$. For the SM calculation, we sum over bottom and top quark contributions, using $m_b = 4.20$ GeV and $m_t = 172.5$ GeV. We adopt the results of [@Altmannshofer:2012ur] to draw vertical exclusion bands on light color octet scalars from ATLAS [@Aad:2011yh] and CMS [@CMS-PAS-EXO-11-016] in dijet pair resonance searches. The gap in sensitivity from 200 GeV to 320 GeV between the 34 pb$^{-1}$ ATLAS search and the 2.2 fb$^{-1}$ CMS search is a result of the increased multijet trigger threshold to handle more difficult run conditions. In particular, for the CMS study, turn-on effects of the QCD multijet trigger made the background modeling unreliable below 320 GeV. We see that both enhancement and suppression are possible, typically delineated by the choice of the sign of $\lambda_{hp}$. This is manifest in the region where $m_S > m_h/2$ where $\lambda_{hp} > 0$ corresponds to a suppression and $\lambda_{hp} < 0$ corresponds to an enhancement. In the region where $m_S < m_h/2$, we see enhancement and suppression for both signs of $\lambda_{hp}$ since the scalars can go on-shell in the loop, leading to an additional imaginary contribution to the scalar amplitude. The resulting interference is complicated by our inclusion of the bottom quark and its imaginary contribution, so the overall magnitude has competing cancellations among real and imaginary amplitude pieces. We note that [Fig. \[fig:SM3LMsContour\_fixlhp\]]{} shows the expected decoupling of $S$ as $m_S$ grows. We also remark that for negative values of $\lambda_{hp}$, our results are consistent with [@Dobrescu:2011aa], where the finite difference in our results is a result of our inclusion of the bottom quark. Lastly, with regards to the ATLAS dijet pair search, we note that the complex triplet scalar is 1/9 the production cross section of the real octet scalar, if their masses are equal, rendering the search insensitive to complex triplet scalars. New Colored Fermion {#sec:Fermion} =================== Adding new colored fermions to the Standard Model can greatly affect gluon fusion SM Higgs production in a number of unique ways. On one hand, new sequential generations of chiral fermions will add constructively with the SM fermion loops and, at leading order, scale the top quark loop by a multiplicative factor, as discussed in [Sec. \[sec:SM\]]{}. On the other hand, a new vector-like colored fermion that does not mix with SM fermions has no effect on gluon fusion. In general, a new colored fermion mass eigenstate comprised of chiral and vector-like components will enhance the SM Higgs gluon fusion rate according to the chiral projection of the mass eigenstate. Because we also allow for Higgs portal-induced scalar mixing, though, the general situation can lead to either an overall suppression or enhancement of the gluon fusion rate. A model demonstrating the myriad of competing effects is straightforward to construct but only illuminating in its limiting cases. Hence, we will initially consider only mixing between a NP fermion and a SM fermion, neglecting the Higgs portal and Higgs mixing. We add new vector-like top partners [@Choudhury:2001hs; @Kumar:2010vx], given by $$\label{eqn:psiRrep} \chi_{L,R} \sim \left( 3, 1 \right)_{2/3} \ .$$ This leads to the following mass terms, $$\label{eqn:masslag} \mathcal{L} \supset -y_t \tilde{H} \bar{Q}_L t_R - y_L \tilde{H} \bar{Q}_L \chi_R - M \bar{\chi}_L \chi_R + \text{ h.c.} \ ,$$ where $M$ is a free parameter and $y_L$ induces mixing between the SM top quark and $\chi$. In the $(t, \chi)$ gauge basis, we have mass $\hat{M}$ and interaction $\hat{N}_h$ matrices given by $$\label{eqn:MN_matrices} \hat{M} = \left( \begin{array}{cc} M_t & \xi_L \\ 0 & M \\ \end{array} \right) \ , \qquad \hat{N}_h = \left( \begin{array}{cc} M_t & \xi_L \\ 0 & 0 \\ \end{array} \right) \ ,$$ with $\xi_L = \dfrac{y_L v_h}{\sqrt{2}}$ and $M_t = \dfrac{y_t v_h}{\sqrt{2}}$. Note the $0$ entry in $\hat{M}$ can always be ensured since it corresponds to the $M^\prime \bar{\chi}_L t_R$ operator, which can be trivially rotated away since $\chi_R$ and $t_R$ have the same quantum numbers. The mass basis rotation matrices are defined in the usual way from $\hat{R} (\hat{M}^\dagger\hat{M}) \hat{R}^\dagger = \left| \hat{M}_D \right|^2$ and $\hat{L} (\hat{M} \hat{M}^\dagger) \hat{L}^\dagger = \left| \hat{M}_D \right|^2$. After rotating $\hat{M}$ and $\hat{N}_h$ on the left (right) by a left-handed (right-handed) rotation matrix, we obtain $$\label{eqn:mass_basis_lagrangian} \begin{array}{ccc} \mathcal{L} \supset - \bar{\textbf{t}} \left( \hat{M}_D + \dfrac{h}{v_h} \hat{V}_h \right) P_R~\textbf{t} + \text{ h.c.} \ , \end{array}$$ where $\textbf{t} \equiv (t_1, t_2)$ and $\hat{M}_D = \hat{L} \hat{M} \hat{R}^\dagger = \text{ diag}(m_{t_1}, m_{t_2})$, $\hat{V}_h = \hat{L} \hat{N}_h \hat{R}^\dagger$. The coupling matrix $\hat{V}_h$ dictates the couplings of the top sector to the SM Higgs and, in principle, can have off diagonal entries; however, only the diagonal entries contribute to $gg \rightarrow h$, because the $\hat{L}$ and $\hat{R}$ rotations leave the gauge interactions diagonal in the mass basis. ![\[fig:mixed\_fermion\_diagrams\][Exotic fermion contribution in the mass eigenbasis.]{}](Exoticfloop_t.pdf) In this top partner scenario, each mass eigenstate gives its own contribution to the diagrams in [Fig. \[fig:mixed\_fermion\_diagrams\]]{}. Since $SU(3)_c$ gauge invariance guarantees these two contributions differ only in their coupling to the Higgs via the element of $\hat{V}_h$, we can take the SM result for $gg \rightarrow h$ through a top quark and insert the appropriate element of $\hat{V}_h$ in place of the usual Yukawa coupling. This approach also encompasses more complicated mixing scenarios whereby the matrix element will exhibit different combinations of mixing angles and couplings for the various $\hat{V}_h$ entries as a replacement for the $h f \bar{f}$ vertex in the $gg \rightarrow h$ amplitude. Since we are focused on exotic fermion effects on $gg \rightarrow h$, we take the $(\hat{V}_h)_{ij}$ entry to be a free parameter, which can be readily calculated in any complete model. The amplitudes involving exotic fermions in the loop are $$\label{eqn:exotic_fermionamp} i \mathcal{M}_{F}^{ad} = i \sum\limits_i \left( \dfrac{\alpha_s}{\pi} \right) \left( \dfrac{(\hat{V}_h)_{ii}}{m_{F_i}} \right) \left( \dfrac{C(r_{F_i})}{2v_h} \right) \delta^{ad} \epsilon_{1\mu} \epsilon_{2\nu} \left( p_1^\nu p_2^\mu - \dfrac{m_h^2}{2} g^{\mu \nu} \right) F_F(\tau_{F_i}) \ ,$$ where the repeated indices on $(\hat{V}_h)_{ii}$ are not summed, $F_F(\tau)$ is given by [Eq. (\[eqn:F\_F\])]{}, $\tau_{F_i} \equiv m_h^2 / (4m_{F_i}^2)$, and $F_i \in \{ t_1, t_2 \}$. We see that the amplitude decouples as $m_{F_i} \rightarrow \infty$, unless $m_{F_i}$ and $(\hat{V}_h)_{ii}$ are generated by a common scale such as the Higgs vev. These direct new physics contributions will alter $gg \rightarrow h$ even in the absence of Higgs mixing. Generally these contributions will add constructively if $(\hat{V}_h)_{ii} > 0$. Now, we augment the previous discussion to include Higgs mixing between $h$ and a new scalar $\phi$. We replace the vector-like mass $M$ in [Eq. (\[eqn:masslag\])]{} by a Yukawa term which generates the desired mass term and a new interaction term involving $\phi$ after $\phi$ obtains a vev, giving $$\label{eqn:vectormass_generation} y_\phi \phi \bar{\chi}_L \chi_R \Rightarrow M (1 + \dfrac{\phi}{v_\phi}) \bar{\chi}_L \chi_R \ ,$$ where $M = \dfrac{y_\phi v_\phi}{\sqrt{2}}$. We now have a second interaction matrix $\hat{V}_\phi$, which is added to [Eq. (\[eqn:mass\_basis\_lagrangian\])]{} and defined analogously to $\hat{V}_h$, where $$\label{eqn:Mmatrix} \hat{N}_\phi = \left( \begin{array}{cc} 0 & 0 \\ 0 & M \\ \end{array} \right) \ .$$ As discussed in [Subsec. \[subsec:general\_amplitudes\]]{}, the $t_1$, $t_2$ loops will give an enhancement factor for gluon fusion production given by $$\label{eqn:SM+F_epsilon} \begin{array}{l} \left. \epsilon_{gg} \right|_{SM + \chi_{L,R}} = \\ \dfrac{ c^2_\theta \left| \sum\limits_{f, \text{ no }t} \left( \dfrac{C(r_f)}{v_h} F_F(\tau_f) \right) + \sum\limits_i \left( \dfrac{C(r_{F_i})}{v_h} \left( \dfrac{ (\hat{V}_h)_{ii}}{m_{F_i}} \right) F_F(\tau_{F_i}) - t_\theta \dfrac{C(r_{F_i})}{v_\phi} \left( \dfrac{ (\hat{V}_\phi)_{ii}}{m_{F_i}} \right) F_F(\tau_{F_i}) \right) \right|^2 }{ \left| \sum\limits_{f} \left( \dfrac{C(r_f)}{v_h} F_F(\tau_f) \right) \right|^2 } \ . \\ \end{array}$$ Note that as $m_{t_1,t_2}\rightarrow \infty$, $t_1$ will decouple from the $\phi$ component of $s_1$ but not the $h$ component, and vice versa for $t_2$. We recognize that these new top partners will induce shifts in the electroweak oblique parameters $S$ and $T$, but a full analysis of the allowed top partner parameter space is beyond the scope of this study. Therefore, we adapt the results from Ref. [@Bai:2011aa], which studied top partner effects on Higgs production and included the constraints from the $S$ and $T$ oblique corrections. We set $m_\phi^2 > m_h^2$, and for the vevs we are considering, $-\lambda_{hp} v_h v_\phi$ is a small perturbation to the diagonal mass terms in [Eq. (\[eqn:scalar\_mass\_matrix\])]{}, allowing us to approximate the $s_1$ contribution to $S$ and $T$ by the Higgs contribution considered in [@Bai:2011aa] for equal masses. We can thus illustrate our main point, suppression of gluon fusion, in this phenomenologically viable top partner scenario. In [Fig. \[fig:mixed\_fermion\_relative\_rate\]]{} we plot contours of $\epsilon_{gg}$ as a function of the left-handed fermion mixing angle and the heavy fermion mass eigenstate, $m_{t_2}$ for representative values of $\lambda_{hp}$. The shaded bands correspond to regions consistent with the oblique parameters at the 68 % and 95 % C.L., taken from [@Bai:2011aa]. Separately, we can also consider new colored fermions which do not mix with the SM quarks. As a final example, we consider new electroweak singlet fermions $\psi$ in the adjoint and fundamental representations of $SU(3)_c$ which obtain mass from the new Yukawa term of [Eq. (\[eqn:vectormass\_generation\])]{} (with $\chi \rightarrow \psi$). In particular, these fermions do not couple to $h$, and hence their effects will be suppressed by the scalar mixing angle in [Eq. (\[eqn:mixingangle\])]{}. The relative rate is now $$\label{eqn:SM+Fadj_epsilon} \left. \epsilon_{gg} \right|_{SM + \psi} = \dfrac{ c^2_\theta \left| \sum\limits_{f} \left( \dfrac{C(r_f)}{v_h} F_F(\tau_f) \right) - t_\theta \left( \dfrac{C(r_\psi)}{v_\phi} F_F(\tau_\psi) \right) \right|^2 }{ \left| \sum\limits_{f} \left( \dfrac{C(r_f)}{v_h} F_F(\tau_f) \right) \right|^2 } \ .$$ We can see that $\psi$ does not decouple from the $gg \rightarrow s_1$ amplitude as its mass is taken very large because $F_F (\tau_\psi)$ asymptotes to a finite value. We show $\epsilon_{gg}$ in [Fig. \[fig:unmixed\_fermion\_relative\_rate\]]{} for two choices of color representations. We see from [Fig. \[fig:unmixed\_fermion\_relative\_rate\]]{} that the octet fermion (which includes a 1/2 to account for lack of conjugate diagram for a real fermion) produces larger suppression or enhancement than the triplet fermion for identical $\lambda_{hp}$ values, as expected from the difference in their respective $C(r_{\psi})$. Collider constraints on these new fermions will require model dependent assumptions about their decay channels, and since our focus is on the model independent effects on gluon fusion, we do not consider such constraints here. ![\[fig:mixed\_fermion\_relative\_rate\][Contours of the relative rate of $s_1$ production as a function of $m_{t_2}$ and the left-handed mixing angle in the top partner scenario for $m_{s_1} = 800$ GeV, $v_\phi = 500$ GeV, and $\lambda_{hp} = -1$ (left) and $0.75$ (right). We adapt the analysis and results of Ref. [@Bai:2011aa] to show shaded contours which are consistent with the oblique parameters $S$ and $T$ at the 68% and 95% C.L.. ]{}](toppartnermin1.pdf "fig:"){width="48.00000%"} ![\[fig:mixed\_fermion\_relative\_rate\][Contours of the relative rate of $s_1$ production as a function of $m_{t_2}$ and the left-handed mixing angle in the top partner scenario for $m_{s_1} = 800$ GeV, $v_\phi = 500$ GeV, and $\lambda_{hp} = -1$ (left) and $0.75$ (right). We adapt the analysis and results of Ref. [@Bai:2011aa] to show shaded contours which are consistent with the oblique parameters $S$ and $T$ at the 68% and 95% C.L.. ]{}](toppartnerpt75.pdf "fig:"){width="48.00000%"} ![\[fig:unmixed\_fermion\_relative\_rate\][Relative rate of $s_1$ production with Higgs mixing and a new color octet fermion (left) and a new color triplet fermion (right). We choose $v_{\phi}=500$ GeV, $m_{s_1}=500$ GeV and $m_{s_2} = 700$ GeV. From top to bottom, the solid lines in each plot correspond to $\lambda_{hp}=0.015$, $0.005$, $-0.005$, $-0.015$. ]{}](SMEFOctetms1500GeV.pdf "fig:"){width="48.00000%"} ![\[fig:unmixed\_fermion\_relative\_rate\][Relative rate of $s_1$ production with Higgs mixing and a new color octet fermion (left) and a new color triplet fermion (right). We choose $v_{\phi}=500$ GeV, $m_{s_1}=500$ GeV and $m_{s_2} = 700$ GeV. From top to bottom, the solid lines in each plot correspond to $\lambda_{hp}=0.015$, $0.005$, $-0.005$, $-0.015$. ]{}](SMEFTripletms1500GeV.pdf "fig:"){width="48.00000%"} New Colored Vector {#sec:Vector} ================== The last type of New Physics contribution to gluon fusion we will consider is the addtion of a new massive colored vector. In a renormalizable theory, the massive vector must arise from a spontaneously broken gauge theory, which necessarily entails the addition of a new scalar that acquires a vev and can mix with the SM Higgs via [Eq. (\[eqn:Lhp\])]{}. Correspondingly, we will consider an extended gauge symmetry $SU(3)_1 \times SU(3)_2$, known as the renormalizable coloron model (ReCoM) [@Hill:2002ap; @Bai:2010dj]. In this model, the complex scalar field $\Phi$ transforms as $(3, \overline{3})$ and obtains a diagonal vev (when written as a matrix-valued field), which breaks $SU(3)_1 \times SU(3)_2$ to the diagonal subgroup, which is identified with the SM $SU(3)_c$ symmetry. The $\Phi$ field has 18 degrees of freedom: 8 are “eaten” by the broken gauge generators to make the massive color vector $G'$ known as the coloron, 8 become a real scalar $SU(3)_c$ octet $G_H$, and the remaining 2 are the real scalar $\phi_R$ and pseudoscalar $\phi_I$ color singlet fields. Hence, in this construction and a consequence of the Higgs portal, the addition of a massive color vector $G^\prime$ concomitantly includes a new scalar octet and two new scalar singlets, of which $G_H$ necessarily affects gluon fusion and $\phi_R$ mixes with the SM Higgs. The Renormalizable Coloron Model {#subsec:ReCoM} -------------------------------- We analyze the total scalar potential including the SM, the ReCoM, and the Higgs portal. Our analysis mirrors that found in [@Bai:2010dj], except our addition of the Higgs portal operator creates a link between the two scalar potentials $V(H)$ and $V(\Phi)$ and hence the two vevs must be solved for simultaneously. The full scalar potential is $$\label{eqn:fullscalarpotential} \begin{array}{ccl} V_{tot} &=& V(\Phi) + V(H) + V_{hp} \\ &=& -m_\Phi^2 \text{ Tr}(\Phi^\dagger \Phi) - \mu_\Phi(\det \Phi + \text{ h.c.}) + \dfrac{\lambda_\Phi}{2} \left[ \text{Tr}( \Phi \Phi^\dagger ) \right]^2 + \dfrac{\kappa_\Phi}{2} \text{ Tr}( \Phi \Phi^\dagger \Phi \Phi^\dagger) \\ &-& m_H^2 |H|^2 + \lambda_H |H|^4 \\ &-& \lambda_{hp} |H|^2 \text{ Tr}(\Phi^\dagger \Phi) \ , \\ \end{array}$$ where we assume $\mu_\Phi > 0$ without loss of generality. We require $m_\Phi^2 > 0$ and $m_H^2 > 0$ such that $\Phi$ and $H$ will acquire vevs. We also require $3 \lambda_\Phi + \kappa_\Phi > 0$ and $\lambda_H > 0$ so each individual potential is bounded from below: we neglect renormalization group effects when discussing bounds on potential parameters. It is straightforward to find the vevs for $\Phi$ and $H$ by decoupling the two equation system. We find, in analogy with [@Bai:2010dj], $$\label{eqn:vPhi} \langle \Phi \rangle = \dfrac{v_\phi}{\sqrt{6}} \mathbb{I}_3 = \dfrac{ \mu_\Phi + \sqrt{\mu_\Phi^2 + \left( 2 (3\lambda_\Phi+\kappa_\Phi) - \dfrac{3 \lambda_{hp}^2}{\lambda_H} \right) \left( 2 m_\Phi^2 + \dfrac{\lambda_{hp} m_H^2}{\lambda_H} \right) }}{ \left( 2 (3\lambda_\Phi + \kappa_\Phi) - \dfrac{ 3\lambda_{hp}^2}{\lambda_H} \right)} \mathbb{I}_3 \ .$$ If $\lambda_{hp}$ is too large, then it can generate a new ground state at large field values of $h$ and $\phi$. The resulting upper bound on $\lambda_{hp}$ is $\lambda_{hp}^2 < \dfrac{2}{3} \lambda_H (3 \lambda_{\Phi} + \kappa_{\Phi})$, which we satisfy by requiring $v_{\phi} > 0$. Given [Eq. (\[eqn:vPhi\])]{}, the Higgs vev is most easily written as $$\label{eqn:vH} \langle H \rangle = \dfrac{1}{\sqrt{2}} \left( \begin{array}{c} 0 \\ v_h \\ \end{array} \right) = \dfrac{1}{\sqrt{2}} \left( \begin{array}{c} 0 \\ \sqrt{ \dfrac{m_H^2}{\lambda_H} + \dfrac{\lambda_{hp} v_\phi^2}{2 \lambda_H}} \\ \end{array} \right) \ ,$$ and $v_h$ is fixed to be 246 GeV. Expanding $\Phi$ around its vev, we get $$\label{eqn:Phi} \Phi = \dfrac{1}{\sqrt{6}} \left( v_\phi + \phi_R + i \phi_I \right) \mathbb{I}_3 + \left( G_H^a + iG_G^a \right) T^a \ ,$$ where $\phi_R$ and $\phi_I$ are $SU(3)_c$ singlets and $G_H$ and $G_G$ are $SU(3)_c$ octets [@Bai:2010dj]. The $G_G$ comprise the Goldstone bosons which give mass to the coloron, $G^\prime$. The Higgs is decomposed in the usual way, $$\label{eqn:Higgs} H = \dfrac{1}{\sqrt{2}} \left( \begin{array}{c} G^{\pm} \\ v_h + h + iG_o \\ \end{array} \right)$$ where $G_o$ and $G^{\pm}$ are the Goldstone bosons eaten by the electroweak gauge bosons. After the spontaneous symmetry breaking $SU(3)_1\times SU(3)_2\rightarrow SU(3)_c$ and EWSB, mixing is induced between the singlets $\phi_R$ and $h$. This leads to the mass squared matrix in the $(h, \phi_R)$ interaction basis given in [Eq. (\[eqn:scalar\_mass\_matrix\])]{} but with $m_{\phi}^2 \rightarrow m_{\phi_R}^2$ and $$\label{eqn:mass_matrix_terms} m_h^2 = 2\lambda_H v_h^2 \ , \quad m_{\phi_R}^2 = \dfrac{v_\phi^2}{3}(3\lambda_\Phi + \kappa_\Phi) - \dfrac{\mu_\Phi v_\phi}{\sqrt{6}} \ ,$$ where $v_h$ and $v_\phi$ depend on $\lambda_{hp}$. The assumption of $v_{\phi} > 0$ and the conditions $0 \leq m_h^2 \leq m^2_{\phi_R}$ imply $$\label{eqn:condition5} \mu_\Phi < \sqrt{\dfrac{2}{3}}(3\lambda_\Phi + \kappa_\Phi) v_\phi \ .$$ By our assumptions, the right hand side of [Eq. (\[eqn:condition5\])]{} is positive definite and thus bounds $\mu_\Phi$ from above. Our analysis follows exactly the same steps as [Subsec. \[subsec:mixing\]]{}, giving the following results: $$\label{eqn:hphiR_mixing} \begin{array}{ccc} \tan 2\theta &=& \dfrac{-2 \lambda_{hp} v_h v_{\phi} }{m_{\phi_R}^2 - m_h^2} \ , \\ s_1 &=& h \cos \theta - \phi_R \sin \theta \ , \\ s_2 &=& h \sin \theta + \phi_R \sin \theta \ . \\ \end{array}$$ For the physical masses of $s_1$ and $s_2$, we obtain $$\label{eqn:ms1_ms2} \begin{array}{ccl} \vspace{4pt} m^2_{s_1} &=& \dfrac{1}{2}(m_h^2 + m_{\phi_R}^2) - \dfrac{1}{2}\sqrt{(-m_h^2 + m_{\phi_R}^2)^2 + 4 \lambda_{hp}^2 v_h^2 v_{\phi}^2} \ , \\ m^2_{s_2} &=& \dfrac{1}{2}(m_h^2 + m_{\phi_R}^2) + \dfrac{1}{2}\sqrt{(-m_h^2 + m_{\phi_R}^2)^2 + 4 \lambda_{hp}^2 v_h^2 v_{\phi}^2} \ . \\ \end{array}$$ For the physical masses of the remaining scalars in the spectrum we find, $$\label{eqn:scalarmasses} m_{\phi_I}^2 = \sqrt{\dfrac{3}{2}}\mu_\Phi v_\phi \ , \quad m_{G_H}^2 = \dfrac{1}{3}(2m_{\phi_I}^2 + \kappa_\Phi v^2_\phi) \ ,$$ which agrees with [@Bai:2010dj] in the limit $\lambda_{hp} \rightarrow 0$. The constraint $m_{\phi_I}^2 > 0$ is satisfied since we assumed $\mu_\Phi > 0$ and $v_\phi > 0$. Requiring $m_{G_H}^2 > 0$ implies $m_{\phi_I}^2 > -\kappa_\Phi v_\phi^2 /2$, which augments the previous condition [Eq. (\[eqn:condition5\])]{} to give $$\label{eqn:mu_bounds} -\frac{\kappa_\Phi v_\Phi}{\sqrt{6}} < \mu_\Phi < \sqrt{\frac{2}{3}} \left( 3 \lambda_\Phi + \kappa_\Phi \right) v_\phi$$ In order for a valid range of $\mu_\Phi$ to exist, we thus require $$\label{eqn:lambda_kappa_inequality} 2 \lambda_\Phi + \kappa_\Phi > 0 \ .$$ Our subsequent analysis ensures these constraints are satisfied. Phenomenology {#subsec:SM3pVectorPheno} ------------- The diagrams for the colored vector loop in unitary gauge are shown in [Fig. \[fig:Unitaryvectorloop\]]{}. Although the ReCoM model includes the $G_H$ scalar octet contribution, we have isolated colored scalar amplitudes in [Sec. \[sec:Scalar\]]{}, and so we focus here on the colored vector contribution. ![\[fig:Unitaryvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow \phi_R$ in the unitary gauge.]{}](UnitaryFigA.pdf "fig:") ![\[fig:Unitaryvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow \phi_R$ in the unitary gauge.]{}](UnitaryFigB.pdf "fig:") The full amplitude for a real vector field propagating in the loop is $$\label{eqn:UGampsum} \begin{array}{ccl} i \mathcal{M}^{ad}_V &=& i \left. \mathcal{M}^{ad}_V (gg \rightarrow \phi_R) \right|_{m_{\phi_R} = m_{s_1}} \\ &=& i \left( \dfrac{\alpha_s}{\pi} \right) \left( \dfrac{C(r_{G^\prime})}{4 v_\phi} \right) \delta^{ad} \epsilon_{1\mu} \epsilon_{2\gamma} \left( p_1^\gamma p_2^\mu - \dfrac{m_{s_1}^2}{2} g^{\mu\gamma} \right) F_V(\tau_{G^\prime}) \\ \end{array}$$ where $$\label{eqn:FV_definition} F_V(\tau) \equiv - \left( \tau^{-1} (3 + 2 \tau) + 3 \tau^{-2} (-1 + 2 \tau)Z(\tau) \right) \ .$$ A full derivation of this amplitude in both unitary and Feynman gauge is presented in [Appendix \[sec:Vector\_explicit\]]{}. The resulting enhancement factor is $$\label{eqn:SM+V_epsilon} \begin{array}{l} \left. \epsilon_{gg} \right|_{SM + ReCoM} = \\ \dfrac{ c^2_\theta \left| \sum\limits_{f} \left( \dfrac{C(r_{f})}{2v_h} F_F(\tau_{f}) \right) + \dfrac{1}{2} \dfrac{C(r_{G_H})}{4v_\phi} \left( \dfrac{ \lambda_{hp} v_h v_\phi - t_\theta x_{G_H} }{m_{G_{H}}^2} \right) F_S (\tau_{G_H}) - t_\theta \left(\dfrac{C(r_{G^\prime})}{4v_\phi} F_V(\tau_{G^\prime}) \right) \right|^2 }{ \left| \sum\limits_{f} \left( \dfrac{C(r_f)}{2v_h} F_F(\tau_f) \right) \right|^2 } \ , \\ \end{array}$$ where $x_{G_H} / v_\phi = \left( -m_{\phi_R}^2 + \frac{2}{3} m_{\phi_I}^2 - 2 m_{G_H}^2 \right) / v_\phi$ evaluated at $m_{\phi_R} = m_{s_1}$ in [Eq. (\[eqn:SM+V\_epsilon\])]{} is the $G_H$ coupling to $\phi_R$ and we have included both $G_H$ (with an explicit symmetry factor of 1/2) and $G^\prime$ contributions. We note that the real colored vector loop function $F_{G^\prime}$ is numerically about a factor of 5 larger and of the opposite sign than the usual SM loop function $F_F$. The scalar loop function $F_S(\tau)$ is also of the opposite sign and roughly a third of $F_F$: the loop functions are plotted in [Fig. \[fig:loop\_functions\]]{}. We comment that, as a result of the large loop function for the colored vector, moderate values of $\lambda_{hp}$ can have large effects on gluon fusion production. This provides a straightforward construction, for example, to counteract the enhancement from a fourth generation of chiral fermions in the Standard Model. If such a large cancellation of $gg \rightarrow h$ amplitudes was present, however, we expect di-Higgs production via $gg \rightarrow hh$ to be correspondingly enhanced if the $gg \rightarrow hh$ triangle and bubble amplitudes are negligible: this is because the individual signs of direct Higgs couplings that lead to suppression become squared in the $gg \rightarrow hh$ box amplitude. In this case, the di-Higgs gluon fusion production channel may be a promising discovery mode, and a more careful study is certainly warranted. ![\[fig:loop\_functions\][Loop functions $F_L = F_F$ (top), $F_L = F_S$ (middle), and $F_L = F_V$ (bottom) for fermion, scalar, and vector particles of mass $M$, respectively. The loop functions develop an imaginary part for $\tau > 1$, which corresponds to the particles in the loop going on-shell. ]{}](Loopfunctions.pdf){width="70.00000%"} We first present results in [Fig. \[fig:vector\_LMsContour\_fixmh\]]{} for $\epsilon_{gg}$ with the sole addition of the colored vector for $v_\phi = 200$ GeV and $3$ TeV. The $v_\phi = 200$ GeV choice is interesting because bounds on colorons in the low mass region coming from pairs of dijet resonances from ATLAS [@Aad:2011yh] and CMS [@CMS-PAS-EXO-11-016] are not currently sensitive in the 200 to 320 GeV range, as discussed in [Sec. \[sec:Scalar\]]{}. There are numerous recent studies which have done fits of the Higgs couplings to the data, including the Higgs-gluon effective coupling [@Carmi:2012yp; @Carmi:2012zd; @Giardino:2012ww; @Giudice:2007fh]. For $v_{\phi}=200$ GeV, a coloron mass of $250$ GeV and mixing angle of $s_\theta \sim -0.04$ one can reproduce the best fit point ($c_g \sim 0.5$) for the scalar-gluon effective coupling found in [@Carmi:2012zd] and as can be seen from [Fig. \[fig:vector\_LMsContour\_fixmh\]]{} one can also easily reproduce values which give rates close to the SM value should the effective coupling migrate towards the SM prediction with more data. We also consider $v_\phi = 3$ TeV, which is the scale probed in dijet resonance searches using 4 fb$^{-1}$ of 8 TeV LHC data at CMS [@CMS-PAS-EXO-12-016]. The observed limit from this search on the coloron mass, $m_{G^\prime}$, is 3.28 TeV. The bounds are indicated by gray vertical bands as described in the caption. The latest dijet resonance search done by the ATLAS collaboration [@ATLAS-CONF-2012-088] with 5.8 fb$^{-1}$ of 8 TeV LHC data does not report an observed limit on the coloron mass but we expect the limit to be just a little higher because of the increased luminosity. An additional constraint on the coloron mass arises from the constraints imposed by ReCoM, whereby perturbativity restrictions on the original $SU(3)_1 \times SU(3)_2$ gauge couplings give an upper limit and requirements on generating the correct $SU(3)_c$ coupling give a lower limit. In deriving these bounds, which are given by dashed vertical lines in [Fig. \[fig:vector\_LMsContour\_fixmh\]]{}, we have neglected renormalization group running of $\alpha_s$. In general, the ReCoM model includes contributions from the color vector $G^\prime$ and the scalar octet $G_H$. We can see from [Eq. (\[eqn:SM+V\_epsilon\])]{} that the contribution from $G_H$ coming through the $h$ component of $s_1$ always adds constructively with the $G^\prime$ contribution. Whether the contribution from $G_H$ entering through $\phi_R$ also adds constructively with the $G^\prime$ contribution depends on the sign of $x_{G_H}$, which in turn depends on the mass hierarchy between the various scalars. We present the effect arising from only the color vector in the top row and from both new colored states in the bottom row of [Fig. \[fig:vector\_LMsContour\_fixmh\]]{}. For these plots, we have set $m_{s_1} = 125$ GeV, $m_{s_2} = 225$ GeV, $m_{\phi_I} = 160$ GeV, and $m_{G_H} = 140$ GeV.[^2] We remark that the flat behavior of $\epsilon_{gg}$ in each plot arises because for $m_{G^\prime} > m_{s_1}$, the loop function dependence of $m_{G^\prime}$ asymptotes quickly. This reflects the fact that as $m_{G^\prime}$ is taken large, its effects (which enter only through the $\phi_R$ component of $s_1$) do not decouple from the $s_1$ production amplitude, which is reminiscent of the non-decoupling of $W$ bosons from the SM Higgs in $h \rightarrow \gamma \gamma$. In addition, for the scalar octet $G_H$, which couples to both $h$ and $\phi_R$ components, we find that as $m_{G_H}$ is taken large, its effects do not decouple from the $\phi_R$ component but do decouple from $h$, see [Sec. \[sec:Scalar\]]{}. Finally, we note the small reduction of $\epsilon_{gg}$ in ReCoM is a result of the $G_H$ contribution slightly cancelling the $G^\prime$ contribution, given our chosen parameter point for which $x_{G_H} < 0$, and we see that the overall effect is dominated by the coloron contribution, as expected from the magnitudes of the loop functions shown in [Fig. \[fig:loop\_functions\]]{}. ![\[fig:vector\_LMsContour\_fixmh\][Relative rate of $s_1$ production showing the effect of adding only the coloron (top row) and the effect of adding the coloron and $G_H$, with $m_{G_H} = 140$ GeV (bottom row). From top to bottom, the solid lines correspond to $\lambda_{hp} = -7.5 \times 10^{-4}$, $-5 \times 10^{-4}$, $5 \times 10^{-4}$, $7.5 \times 10^{-4}$. In the left panels, the gray bands correspond to the pair-produced dijet bounds from ATLAS (left) and CMS (right). In the right panels, the left gray band corresponds to the CMS exclusion on coloron production in dijet resonance searches. The left vertical dotted line corresponds to the minimum $m_{G'}$ mass allowed given the specified choice of $v_\phi$, and the right vertical dotted line corresponds to the perturbativity constraint. ]{}](s1prodSMpColoron200GeV.pdf "fig:"){width="48.00000%"} ![\[fig:vector\_LMsContour\_fixmh\][Relative rate of $s_1$ production showing the effect of adding only the coloron (top row) and the effect of adding the coloron and $G_H$, with $m_{G_H} = 140$ GeV (bottom row). From top to bottom, the solid lines correspond to $\lambda_{hp} = -7.5 \times 10^{-4}$, $-5 \times 10^{-4}$, $5 \times 10^{-4}$, $7.5 \times 10^{-4}$. In the left panels, the gray bands correspond to the pair-produced dijet bounds from ATLAS (left) and CMS (right). In the right panels, the left gray band corresponds to the CMS exclusion on coloron production in dijet resonance searches. The left vertical dotted line corresponds to the minimum $m_{G'}$ mass allowed given the specified choice of $v_\phi$, and the right vertical dotted line corresponds to the perturbativity constraint. ]{}](s1prodSMpColoron3TeV.pdf "fig:"){width="48.00000%"} ![\[fig:vector\_LMsContour\_fixmh\][Relative rate of $s_1$ production showing the effect of adding only the coloron (top row) and the effect of adding the coloron and $G_H$, with $m_{G_H} = 140$ GeV (bottom row). From top to bottom, the solid lines correspond to $\lambda_{hp} = -7.5 \times 10^{-4}$, $-5 \times 10^{-4}$, $5 \times 10^{-4}$, $7.5 \times 10^{-4}$. In the left panels, the gray bands correspond to the pair-produced dijet bounds from ATLAS (left) and CMS (right). In the right panels, the left gray band corresponds to the CMS exclusion on coloron production in dijet resonance searches. The left vertical dotted line corresponds to the minimum $m_{G'}$ mass allowed given the specified choice of $v_\phi$, and the right vertical dotted line corresponds to the perturbativity constraint. ]{}](s1prodSMpfullColoron200GeV.pdf "fig:"){width="48.00000%"} ![\[fig:vector\_LMsContour\_fixmh\][Relative rate of $s_1$ production showing the effect of adding only the coloron (top row) and the effect of adding the coloron and $G_H$, with $m_{G_H} = 140$ GeV (bottom row). From top to bottom, the solid lines correspond to $\lambda_{hp} = -7.5 \times 10^{-4}$, $-5 \times 10^{-4}$, $5 \times 10^{-4}$, $7.5 \times 10^{-4}$. In the left panels, the gray bands correspond to the pair-produced dijet bounds from ATLAS (left) and CMS (right). In the right panels, the left gray band corresponds to the CMS exclusion on coloron production in dijet resonance searches. The left vertical dotted line corresponds to the minimum $m_{G'}$ mass allowed given the specified choice of $v_\phi$, and the right vertical dotted line corresponds to the perturbativity constraint. ]{}](s1prodSMpfullColoron3TeV.pdf "fig:"){width="48.00000%"} Discussion and Conclusion {#sec:Conclusion} ========================= We have seen that new colored particles can have significant effects on gluon fusion production. Also, the inclusion of Higgs portal-induced scalar mixing readily leads to new possibilities for suppressing or enhancing the gluon fusion rate. We have isolated contributions arising from new colored scalars, new colored fermions, including quark mixing, and new colored vectors. With the amplitudes in [Eq. (\[eqn:scalar\_ME\])]{}, [Eq. (\[eqn:exotic\_fermionamp\])]{}, and [Eq. (\[eqn:UGampsum\])]{}, we can immediately calculate the interference effects present in general new physics models. We have demonstrated these effects can easily run from $\mathcal{O}(1\%)$ to $\mathcal{O}(10\%)$, and both suppression and enhancement of gluon fusion can occur. Moreover, such large deviations are possible by colored states at mass scales that can be directly probed at the LHC. In particular, when the effects on gluon fusion scale with the Higgs portal coupling $\lambda_{hp}$, the dearth of restrictions on $\lambda_{hp}$ from the underlying theory can lead to very large modifications. If many competing effects are present in the $gg \rightarrow h$ amplitude, we expect that di-Higgs production will be correspondingly altered and a promising discovery mode, but we leave a more careful study for future work. Since gluon fusion production does not directly probe the Higgs mechanism or the phenomenon of electroweak symmetry breaking, Higgs identification studies should generally allow for mixing with a separate scalar state as well as competing effects from multiple new colored states running in the loop. Our general framework and analysis can be easily mapped onto the various recent studies which attempt to fit the data with non-SM Higgs couplings to two gluons. In particular we have shown that the various new physics effects can conspire to give rates very close to the SM expectation as well as easily accounting for any slight deviations as suggested by recent fits of the scalar effective couplings [@Carmi:2012yp; @Carmi:2012zd; @Giardino:2012ww; @Giudice:2007fh]. Acknowledgements {#sec:acknowledgements .unnumbered} ================ The authors are grateful to Bill Bardeen, Marcela Carena, André de Gouvêa, Bogdan Dobrescu, Jennifer Kile, Ian Low, Adam Martin, Reinard Primulando, Pedro Schwaller, Daniel Stolarski, and Tim Tait for useful discussions. Fermilab is operated by Fermi Research Alliance, LLC under Contract No. De-AC02-07CH11359 with the United States Department of Energy. Vector Loop Calculation {#sec:Vector_explicit} ======================= In this appendix, we present the explicit calculation of the vector loop contribution to gluon fusion. As the loop calculation involves a particular choice of $R_\xi$ gauge, we perform the calculation in both unitary ($R_{\xi} \rightarrow \infty$) and Feynman ($R_{\xi} = 1$) gauges. This calculation generalizes the well-known Standard Model calculation for $h \rightarrow \gamma \gamma$ [@Ellis:1975ap; @Shifman:1979eb; @Marciano:2011gm] to situations where a new “Higgs” field acquires a vev and leaves a non-Abelian gauge symmetry unbroken. In the SM, the Higgs field is responsible for spontaneously breaking $SU(2)_L \times U(1)_Y$, leaving the photon as the gauge field of the remaining Abelian $U(1)_{em}$ gauge symmetry. In contrast, in the renormalizable coloron model (ReCoM), $\Phi$ is responsible for spontaneously breaking $SU(3)_1 \times SU(3)_2$, leaving the gluon as the gauge field of the remaining non-Abelian $SU(3)_c$ gauge symmetry. Then, $gg \rightarrow \phi_R$ is the non-Abelian mirror version of $h \rightarrow \gamma \gamma$. We intuit that $\mathcal{M}(gg \rightarrow \phi_R)$ is enhanced by a color factor over the mirror process $\mathcal{M}(h \rightarrow \gamma \gamma)$, which is borne out from our calculation. Vector Loop Amplitude: Unitary Gauge {#subsec:Vector_unitary_explicit} ------------------------------------ We present the unitary gauge calculation of a colored vector contribution to gluon fusion. As mentioned above, we assume an extended color gauge symmetry that is partially broken by the vev of a new scalar field $\Phi$. After Higgs portal-induced mixing of $h$ and $\phi_R$, the matrix elements $\mathcal{M}(gg \rightarrow s_1)$ and $\mathcal{M}(gg \rightarrow s_2)$ are simply related to $\mathcal{M}(gg \rightarrow h)$ and $\mathcal{M}(gg \rightarrow \phi_R)$ by $$\label{eqn:ME_projections} \begin{array}{ccc} \mathcal{M} (gg \rightarrow s_1) &=& \left. c_\theta \left[ \mathcal{M} (gg \rightarrow h) \right] \right|_{m_h = m_{s_1}} \left. -s_\theta \left[ \mathcal{M} (gg \rightarrow \phi_R) \right] \right|_{m_{\phi_R} = m_{s_1}} \\ \mathcal{M} (gg \rightarrow s_2) &=& \left. s_\theta \left[ \mathcal{M} (gg \rightarrow h) \right] \right|_{m_h = m_{s_2}} \left. +c_\theta \left[ \mathcal{M} (gg \rightarrow \phi_R) \right] \right|_{m_{\phi_R} = m_{s_2}} \ . \\ \end{array}$$ There are two diagrams which contribute to $\mathcal{M} (gg \rightarrow \phi_R)$ in the unitary gauge, shown in [Fig. \[fig:Unitaryvectorloop\]]{} of the main text. The triangle diagram in [Fig. \[fig:Unitaryvectorloop\]]{}A for the coloron insertion gives $$\begin{aligned} \label{eqn:MaAmplitude} i{\mathcal M}^{ad}_A &=& -g_s^2 \left( \dfrac{2m_{G^\prime}^2}{v_\phi}\right) f^{abc}f^{dcb} \epsilon_{1\mu} \epsilon_{2\gamma} \nonumber \\ & & \int \dfrac{d^dk}{(2\pi)^d} \dfrac{V^{\mu\nu\rho} \left( g_{\alpha \nu} - \dfrac{k_\alpha k_\nu}{m_{G^\prime}^2}\right) V^{\gamma\beta\alpha} \left(g_{\beta \delta} - \dfrac{{p_b}_\beta {p_b}_\delta}{m_{G^\prime}^2}\right) g^{\sigma\delta} \left(g_{\sigma \rho} - \dfrac{{p_a}_\sigma {p_a}_\rho}{m_{G^\prime}^2}\right) }{(k^2 - m_{G^\prime}^2)(p_a^2 - m_{G^\prime}^2)(p_b^2 - m_{G^\prime}^2)} \ ,\end{aligned}$$ where $p_a = k + p_1$, $p_b = k - p_2$, and the three-point vector vertex is $$\begin{array}{ccc} \label{eqn:three_vertex} V^{\mu\nu\rho} = (k + p_a)^\mu g^{\nu \rho} + (-p_a - p_1)^\nu g^{\rho \mu} +(p_1 -k)^\rho g^{\mu \nu} \\ V^{\gamma\beta\alpha} = (p_b + k)^\gamma g^{\alpha \beta} + (p_2 - p_b)^\alpha g^{\gamma \beta} +(-k - p_2)^\beta g^{\alpha \gamma} \ . \\ \end{array}$$ The amplitude for the bubble loop in [Fig. \[fig:Unitaryvectorloop\]]{}B is $$\label{eqn:MbAmplitude} i{\mathcal M}^{ad}_B = - \left( \dfrac{1}{2} \right) g_s^2 \left(\dfrac{2m_{G^\prime}^2}{v_\phi}\right) \epsilon_{1\mu} \epsilon_{2\gamma} \int \dfrac{d^dk}{(2\pi)^d} \dfrac{V_{acdb}^{\mu\rho\gamma\beta} g^{\sigma \delta} \left(g_{\delta \beta} - \dfrac{p_{a \delta} p_{a \beta}}{m_{G^\prime}^2}\right) \left(g_{\rho \sigma} - \dfrac{k_\rho k_\sigma}{m_{G^\prime}^2}\right)}{(p_a^2 - m_{G^\prime}^2)(k^2 - m_{G^\prime}^2)} \ ,$$ where $p_a = p_1 + p_2 - k$, a symmetry factor of $\dfrac{1}{2}$ has been included, and the four-point vector vertex is $$\label{eqn:four_vertex} \begin{array}{ccl} \delta^{bc} V_{acdb}^{\mu\rho\gamma\beta} &=& -ig_s^2 \delta^{bc} \left(f^{ace}f^{dbe} \left(g^{\mu \gamma}g^{\rho \beta}-g^{\mu \beta}g^{\rho \gamma}\right) + f^{ade}f^{cbe} \left(g^{\mu \rho}g^{\gamma \beta}-g^{\mu \beta}g^{\gamma \rho}\right) \right. \\ &+& f^{abe}f^{cde} \left. \left(g^{\mu \rho}g^{\gamma \beta} - g^{\mu \gamma}g^{\rho \beta}\right) \right) \\ &=& -ig_s^2 f^{abe}f^{dbe} \left( 2 g^{\mu \gamma} g^{\rho \beta} - g^{\mu \beta} g^{\rho \gamma} - g^{\mu \rho} g^{\gamma \beta} \right) \ . \end{array}$$ After expanding both [Eqs. (\[eqn:MaAmplitude\]) and (\[eqn:MbAmplitude\])]{} using Feynman parameters, performing the loop momentum integration using Dimensional Regularization [@'tHooft:1972fi], and some simpification, we arrive at the summed amplitude $$\label{eqn:Unitary_amplitude} i{\mathcal M}^{ad}_V = i \left( \dfrac{\alpha_s}{\pi} \right) \left( \dfrac{C(r_{G^\prime})}{4 v_\phi} \right) \delta^{ad} \epsilon_{1\mu} \epsilon_{2\gamma} \left( p_1^\gamma p_2^\mu - \dfrac{m_{s_1}^2}{2} g^{\mu\gamma} \right) F_V(\tau_{G^\prime}) \ ,$$ where $C(r_{G^\prime}) = 3$ for the coloron, $\tau_{G^\prime} = m_{s_1}^2 / (4 m_{G^\prime}^2)$, and the loop function $F_V$ is given in the main text in [Eq. (\[eqn:FV\_definition\])]{}. We remark that this result also agrees with the analogous SM calculation for $h \xrightarrow[W]{} \gamma \gamma$ with the appropriate substitutions $\alpha_s \rightarrow \alpha$, $C(r) \rightarrow 1$, $v_\phi \rightarrow v_h$, and a factor of 2 included for the charge conjugate process [@Ellis:1975ap; @Shifman:1979eb; @Marciano:2011gm]. Calculation: Feynman Gauge {#subsec:Vector_Feynman_explicit} -------------------------- ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigA.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigB.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigC.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigD.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigE.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigF.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigG.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigH.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigI.pdf "fig:") ![\[fig:Feynmanvectorloop\][Feynman diagrams for vector loop contributions to $gg \rightarrow s_1$ in the Feynman gauge.]{}](FigJ.pdf "fig:") As a cross-check of our result in [Subsec. \[subsec:Vector\_unitary\_explicit\]]{}, we perform the same calculation in Feynman gauge, setting $R_\xi = 1$. Again, we adopt the ReCoM with the Higgs portal, as discussed in [Subsec. \[subsec:ReCoM\]]{}. In contrast to the calculation done in the unitary gauge, here we perform the calculation in the scalar mass basis, explicitly deriving the corresponding Goldstone couplings to $s_1$ after taking into account the $h$–$\phi_R$ mixing. This motivates an interesting discussion of the coupling of Goldstone bosons to their partner Higgs field when the partner Higgs field is mixed with spectator scalars and provides a further check that Higgs mixing and NP contributions to $s_{1,2}$ production can be factored as in [Eq. (\[eqn:ME\_projections\])]{}. Goldstone couplings {#subsec:GoldstoneSector} ------------------- Beginning with the full scalar potential in [Eq. (\[eqn:fullscalarpotential\])]{} and the exact vacuum expectation values given in [Eqs. (\[eqn:vPhi\]) and (\[eqn:vH\])]{}, we examine the Goldstone couplings to the scalars $h$ and $\phi$. After expanding the potential, we have $$\label{eqn:Goldstone_Couplings} V(\Phi) \supset \dfrac{1}{2}G_G^a G_G^d \delta^{ad} \left( \phi_R \left( -\dfrac{\mu}{\sqrt{6}} + \dfrac{v_\phi}{3}(3\lambda_\Phi + \kappa_\Phi) \right) + h (-\lambda_H v_h) \right) + \dfrac{G_0^2}{2} \left( -\lambda_{hp} v_\phi \phi_R + 2\lambda_H v_h h \right) \ .$$ This can be written in matrix form as, $$\label{eqn:Goldstone_Couplings_MatrixForm} \begin{array}{ccl} V(\Phi) &\supset& \left( \begin{array}{cc} \dfrac{h}{\sqrt{2}} & \dfrac{\phi_R}{\sqrt{2}} \end{array} \right) \left( \begin{array}{cc} 2\lambda_H v_h^2 & -\lambda_{hp} v_h v_\phi \\ -\lambda_{hp} v_h v_\phi & \dfrac{-\mu}{\sqrt{6}} v_\phi + \dfrac{v^2_\phi}{3} (3\lambda_\Phi+\kappa_\Phi) \end{array} \right) \left( \begin{array}{c} \dfrac{G_0^2}{v_h\sqrt{2}} \\ \dfrac{G_G^a G_G^d \delta^{ad}}{v_\phi \sqrt{2}} \\ \end{array} \right) \\ &=& \left( \begin{array}{cc} \dfrac{h}{\sqrt{2}} & \dfrac{\phi_R}{\sqrt{2}} \end{array} \right) \hat{M} \left( \begin{array}{c} \dfrac{G_0^2}{v_h\sqrt{2}} \\ \dfrac{G_G^a G_G^d \delta^{ad}}{v_\phi \sqrt{2}} \\ \end{array} \right) \ . \\ \end{array}$$ Note if we set $\lambda_{hp} = 0$, then $v_h$ and $v_\phi$ become the unperturbed vevs and the Goldstone couplings become $m_h^2 / v_h$ and $m^2_{\phi_R}/v_\phi$ for the original $m^2_{\phi_R}$, $m_h^2$, $v_\phi$, and $v_h$. From [Eq. (\[eqn:Goldstone\_Couplings\_MatrixForm\])]{} and [Eq. (\[eqn:mass\_matrix\_terms\])]{} we see that the Goldstone–Goldstone–scalar interaction matrix $\hat{M}$ is the same as the scalar mass matrix. Thus when we diagonalize the mass matrix, we will simultaneously diagonalize the Goldstone interaction matrix in [Eq. (\[eqn:Goldstone\_Couplings\_MatrixForm\])]{}. Explicitly, we write the scalar mass and Goldstone interaction terms as $$\label{eqn:Mass_Plus_Goldstone_Lagrangian} \mathcal{L} \supset - \left( \begin{array}{cc} \dfrac{h}{\sqrt{2}} & \dfrac{\phi_R}{\sqrt{2}} \end{array} \right) \hat{M} \left( \begin{array}{c} \dfrac{h}{\sqrt{2}} \\ \dfrac{\phi_R}{\sqrt{2}} \end{array} \right) - \left( \begin{array}{cc} \dfrac{h}{\sqrt{2}} & \dfrac{\phi_R}{\sqrt{2}} \end{array} \right) \hat{M} \left( \begin{array}{c} \dfrac{G_0^2}{v_h\sqrt{2}} \\ \dfrac{G_G^a G_G^d \delta^{ad}}{v_\phi \sqrt{2}} \\ \end{array} \right) \ .$$ After diagonalization we obtain, for the interaction term, $$\mathcal{L} \supset - \left( \begin{array}{cc} \dfrac{s_1}{\sqrt{2}} & \dfrac{s_2}{\sqrt{2}} \\ \end{array} \right) \hat{M}_D \hat{U}^{-1} \left( \begin{array}{c} \dfrac{G_0^2}{v_h\sqrt{2}} \\ \dfrac{G_G^a G_G^d \delta^{ad}}{v_\phi \sqrt{2}} \\ \end{array} \right)$$ where $\hat{M}_D = \hat{U}^{-1} \hat{M} \hat{U}$ is the diagonalized mass matrix with eigenvalues given in [Eq. (\[eqn:ms1\_ms2\])]{}, and $\hat{U}$ is the unitary rotation matrix with its mixing angle defined in [Eq. (\[eqn:hphiR\_mixing\])]{}. We see that the Goldstone-Goldsone-$s_1$ interaction is $$\label{GGphi} \mathcal{L} \supset -s_\theta \dfrac{m_{s_1}^2}{v_\phi} s_1 G_G^a G_G^d \delta^{ad} \ ,$$ and so the Goldstone–Goldstone coupling to the scalar mass eigenstate is proportional to the scalar mass squared and a mixing angle. This justifies our ability to factor out the Higgs mixing angle when considering $s_{1,2}$ production. Continuation of the Feynman Gauge Calculation {#subsec:Feynman_gauge_continued} --------------------------------------------- Returning to the calculation, there are ten diagrams in the Feynman gauge that must be calculated, as shown in [Fig. \[fig:Feynmanvectorloop\]]{}. Using $p_a = k + p_1$ and $p_b = k -p_2$, $p_c = p_1 + p_2 - k$,  [Eq. (\[eqn:three\_vertex\])]{}, and [Eq. (\[eqn:four\_vertex\])]{}, we obtain the amplitudes $$\begin{aligned} \label{eqn:Feynman_ME} \vspace{4pt} i{\mathcal M}^{ad}_A &=& -2\left( -s_\theta \dfrac{m_{G^\prime}^2}{v_\phi} \right) g_s^2 f^{abc} f^{dcb} \epsilon_{1\mu} \epsilon_{2\gamma} \int \dfrac{d^dk}{(2\pi)^d} \dfrac{g_{\alpha \nu} g_{\beta \delta} g_{\sigma \rho} g^{\delta \sigma} V^{\mu\nu\rho} V^{\gamma\beta\alpha}}{ (p_a^2 - m_{G^\prime}^2) (p_b^2 - m_{G^\prime}^2)(k^2 - m_{G^\prime}^2)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_B &=& \left( \dfrac{1}{2} \right) \left( -s_\theta \dfrac{m_{G^\prime}^2}{v_\phi} \right) \epsilon_{1\mu} \epsilon_{2\gamma} \int \dfrac{d^dk}{(2\pi)^d} \left (-i g_{\rho \beta} \delta^{bc} \right) V_{acdb}^{\mu\rho\gamma\beta} \dfrac{1}{ (p_c^2 - m_{G^\prime}^2)(k^2 - m_{G^\prime}^2)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_C &=& -g_s^2 f^{abc} f^{dcb} \left( -s_\theta \dfrac{m_{s_1}^2}{v_\phi}\right) \epsilon_{1\mu} \epsilon_{2\gamma} \int \dfrac{d^dk}{(2\pi)^d} \dfrac{ \left(p_a^\mu + k^\mu\right) \left(p_b^\gamma + k^\gamma\right)}{ \left(p_a^2-m_{G^\prime}^2\right) \left(k^2-m_{G^\prime}^2\right) \left(p_b^2-m_{G^\prime}^2\right)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_D &=& - g_s^2 f^{bde}f^{bae} \left( -s_\theta \dfrac{m_{s_1}^2}{v_\phi}\right) \epsilon_{1\mu} \epsilon_{2\gamma} g^{\mu\gamma} \int \dfrac{d^dk}{(2\pi)^d} \dfrac{1}{(k^2-m_{G^\prime}^2)(p_c^2-m_{G^\prime}^2)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_E &=& g_s^2 m_{G^\prime}^2 \left( -s_\theta \dfrac{m_{s_1}^2}{v_\phi}\right)f^{abc}f^{dcb} g^{\mu \gamma} \epsilon_\mu \epsilon_\gamma \int \dfrac{d^dk}{(2\pi)^d} \dfrac{1}{(k^2-m_{G^\prime}^2)(p_a^2-m_{G^\prime}^2)(p_b^2-m_{G^\prime}^2)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_F &=& 2 g_s^2 m_{G^\prime}^2 \left( -s_\theta \dfrac{m_{G^\prime}^2}{v_\phi} \right) f^{abc}f^{dcb} g^{\mu \gamma} \epsilon_\mu \epsilon_\gamma \int \dfrac{d^dk}{(2\pi)^4} \dfrac{1}{(k^2-m_{G^\prime}^2)(p_a^2-m_{G^\prime}^2)(p_b^2-m_{G^\prime}^2)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_G &=& 2 g_s^2 \left( -s_\theta \dfrac{m_{G^\prime}^2}{v_\phi}\right) f^{abc} f^{dcb} \epsilon_{1\mu} \epsilon_{2\gamma} \int \dfrac{d^dk}{(2\pi)^d} \dfrac{k^\mu k^\gamma}{ (p_a^2 - m_{G^\prime}^2)(p_b^2-m_{G^\prime}^2)(k^2 - m_{G^\prime}^2)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_H &=& 2 g_s^2 f^{abc} f^{dcb} g^{\mu \gamma} \left( -s_\theta \dfrac{m_{G^\prime}^2 }{v_\phi} \right) \epsilon_{1\mu} \epsilon_{2\gamma} \int \dfrac{d^dk}{(2\pi)^d} \dfrac{1}{ (k^2-m_{G^\prime}^2)(p_b^2-m_{G^\prime}^2)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_I &=& -2 g_s^2 f^{abc} f^{dcb} \left( -s_\theta \dfrac{m_{G^\prime}^2}{v_\phi} \right) \epsilon_{1\mu} \epsilon_{2\gamma} \int \dfrac{d^dk}{(2\pi)^d} \dfrac{g_{\alpha \nu} g_{\sigma \rho} g^{\alpha \gamma} V^{\mu\nu\rho} (k-p_1-2p_2)^\sigma}{ (p_a^2-m_{G^\prime}^2)(p_b^2-m_{G^\prime}^2)(k^2-m_{G^\prime}^2)} \ , \\ \vspace{4pt} i{\mathcal M}^{ad}_J &=& -2 g_s^2 f^{abc} f^{dcb} \left( -s_\theta \dfrac{m_{G^\prime}^2}{v_\phi} \right) \epsilon_{1\mu} \epsilon_{2\gamma} \int \dfrac{d^dk}{(2\pi)^d} \dfrac{g_{\beta \delta} g^{\beta \gamma} (k-p_2)^\delta (2k+p_1)^\mu}{ (p_a^2-m_{G^\prime}^2) (p_b^2-m_{G^\prime}^2)(k^2-m_{G^\prime}^2)} \ .\end{aligned}$$ After converting to Feynman parameters, calculating the loop integrals, summing all of the amplitudes, and some simplification, the divergences cancel and we obtain for the production of $s_1$, $$\label{eqn:Feynman_amplitude} i{\mathcal M}^{ad}_V = -i s_\theta \left( \dfrac{\alpha_s}{\pi} \right) \left(\dfrac{C(r_{G^\prime})}{4 v_\phi} \right) \delta^{ad} \epsilon_{1\mu} \epsilon_{2\gamma} \left( p_1^\gamma p_2^\mu - \dfrac{m_{s_1}^2}{2} g^{\mu\gamma} \right) F_V(\tau_{G^\prime}) \ ,$$ in agreement with [Eq. (\[eqn:Unitary\_amplitude\])]{} and [Eq. (\[eqn:ME\_projections\])]{}. [^1]: For the vector loop calculation, we implicitly assume a unitary gauge calculation where only vectors appear in the loop. If working in Feynman gauge, the associated Goldstone and ghosts would also be part of the vector category. [^2]: The ATLAS search for dijet pairs has an upward fluctuation above their expected limit in the 140 GeV range, which leaves the window of a scalar octet open for this mass point.

ArXiv

--- abstract: 'We consider a one-dimensional infinite chain of coupled charged harmonic oscillators in a magnetic field with a small stochastic perturbation of order $\epsilon$. We prove that for a space-time scale of order $\epsilon^{-1}$ the density of energy distribution (Wigner distribution) evolves according to a linear phonon Boltzmann equation. We also prove that an appropriately scaled limit of solutions of the linear phonon Boltzmann equation is a solution of the fractional diffusion equation with exponent $5/6$.' address: - 'Department of Physics, Faculty of Science and Technology, Keio University, 3-14-1, Hiyoshi, Kohoku-ku, Yokohama, Kanagawa, 223-8522, Japan' - 'Graduate School of Mathematical Sciences, University of Tokyo, 3-8-1, Komaba, Meguro-ku, Tokyo, 153–8914, Japan' - 'Graduate School of Mathematical Sciences, University of Tokyo, 3-8-1, Komaba, Meguro-ku, Tokyo, 153–8914, Japan' author: - Keiji Saito - Makiko Sasada - Hayate Suda title: '$5/6$-Superdiffusion of energy for coupled charged harmonic oscillators in a magnetic field' --- Introduction ============ There has been much progress during the past decades in the understanding of superdiffusion in one dimensional systems with several conservation laws. Chains of coupled oscillators are typical models showing superdiffusive transport of energy. They are the one-dimensional Hamiltonian systems $$\begin{aligned} \begin{cases} \frac{d}{dt} q_x(t) & = \partial_{v_x}\mathcal{H}(v_x(t),q_x(t)) \\ \frac{d}{dt} v_x(t) & = -\partial_{q_x}\mathcal{H}(v_x(t),q_x(t)) , \end{cases}\end{aligned}$$ with Hamiltonian $$\begin{aligned} &\mathcal{H} = \sum_{x \in \Z} \left( \frac{|v_x|^{2}}{2} + V(q_x- q_{x+1}) \right). $$ Here $v_x(t)$ is the velocity of the oscillator $x$ at time $t$ and $q_x(t)$ is the displacement from its equilibrium position of the oscillator $x$ at time $t$. In the case where the potential $V$ is quadratic, the dynamics is linear and the chain is said to be harmonic and otherwise anharmonic. The Fermi-Pasta-Ulam chain (FPU chain) has possibly cubic and/or quartic terms in the potential. Super diffusion of energy and the divergence of the corresponding thermal conductivity have been observed numerically in the dynamics of FPU chains ([@D],[@Ls],[@LLP]). Strong efforts are made to identify the exponent of the divergence and the nature of superdiffusion in FPU chains numerically and theoretically in recent years. In an innovative article [@S], Spohn discussed an asymptotic behavior of time-dependent correlation functions of heat mode applying the method of fluctuating hydrodynamics. His argument suggests that for general anharmonic chains the macroscopic diffusion of energy is governed by the fractional diffusion equation $$\begin{aligned} \label{fractional} \partial_{t} \mathbf{e}(y,t) = - (- \Delta_{y})^{\frac{s}{2}}\mathbf{e}(y,t).\end{aligned}$$ Moreover, Spohn’s theory suggests that there are only two universality classes, $s = \frac{3}{2}$ or $\frac{5}{3}$. However, a rigorous mathematical analysis of the energy transport in the anharmonic chains is too hard to justify Spohn’s theory. Recently as an analytically tractable model, the *harmonic* chains of oscillators with a stochastic exchange of momentum between neighboring sites, which we call the momentum exchange model, was introduced [@BBO]. In [@BBO] the authors prove the divergence of the thermal conductivity for this model and obtain an explicit exponent of the divergence of Green-Kubo formula. To understand the nature of superdiffusion for this model, a weak noise limit is studied in [@BOS]. They show that in the weak noise limit the time evolution of the local density of the energy is governed by the Boltzmann equation $$\begin{aligned} \label{BBOBoltzmann} \partial_{t} u(y,k,t) + \frac{1}{2 \pi} \omega'(k) \partial_{y} u(y,k,t) = (\mathcal{L}u)(y,k,t) , \\ (\mathcal{L}u)(y,k,t) = \int_{\T} dk' ~ R(k,k') (u(y,k',t) - u(y,k,t)). \notag \end{aligned}$$ Here, the local density of energy $u(y,k,t)$ depends on the position $y \in \R$ along the chain, the wave number $k \in \T = [-\frac{1}{2},\frac{1}{2})$ and time $t \ge 0$. ${\omega}(k)$ is the dispersion relation. Later in [@KJO], it is shown that a properly scaled solution of the Boltzmann equation converges to the solution of the fractional diffusion equation with $s=\frac{3}{2}$. The main idea of the proof of this convergence is the following: Since the scattering kernel $R(k,k')$ is positive, can be interpreted as the forward equation for the probability density of a Markov process $(z(t),k(t))$ on $\R \times \T$. Applying a limit theorem for additive functionals of Markov processes, the scaled process $N^{ - \frac{2}{3}} z(Nt)$ converges to a Lévy process generated by $- (-\Delta)^{\frac{3}{4}}$ (up to a constant). By this two-step scaling limit, the $3/4$-fractional diffusion equation is derived from the momentum exchange model rigorously. Recently the $3/4$-fractional diffusion equation is derived by a direct limit (namely one-step scaling limit) in [@JKO]. For a variant of the momentum exchange model, a skew $3/4$-fractional diffusion equation is derived by a direct space-time scaling limit in [@BGJ]. Most recently in [@SS; @TSS] two of the authors introduced another variant of the momentum exchange model which also shows the superdiffusive behavior of the energy but the exponent of the divergence of Green-Kubo formula is different from the original one. The model is a chain of coupled charged harmonic oscillators in a magnetic field with a stochastic exchange of velocity between neighboring sites. The goal of the present paper is to understand the nature of the superdiffusion for this coupled charged harmonic chain of oscillators in a magnetic field with noise. We apply the two-step scaling limits. Following the idea of [@BOS], we first show as Theorem \[thm:main\] that in the weak noise limit the local density of energy is governed by the phonon linear Boltzmann equation $$\begin{aligned} \label{ourBoltzmann} \partial_{t} u(y,k,i,t) & + \frac{1}{2\pi} \omega'(k) \partial_{y} u(y,k,i,t) = \mathcal{L} u(y,k,i,t) , \\ \mathcal{L}u(y,k,i,t) & = \sum_{j= 1,2} \int_{\mathbf{T}} dk' ~ R(k,i,k',j) (u(y,k',j,t)-u(y,k,i,t) ). \notag \end{aligned}$$ Here, the local density of energy $u(y,k,i,t)$ depends on position $y \in \R$ along the chain, the wave number $k \in \T$, the type of phonon $i =1,2$ and time $t \ge 0$. Then, we consider a properly scaled solution of the Boltzmann equation and show that it converges to the solution of the fractional diffusion equation with $s=\frac{5}{3}$ as Theorem \[thm:main2\]. This provides a first rigorous example of the $5/6$-superdiffusion of energy in a chain of oscillators. A key ingredient of the proof of Theorem \[thm:main2\] is the scaling limit of an additive functional of a Markov process as the prior work. Actually, since the scattering kernel $R(k,i,k',j)$ is positive, $(\ref{ourBoltzmann})$ can be interpreted as the time evolution of the density for a Markov process $(Z(t),K(t),I(t))$ on $\R \times \T \times \{ 1,2 \}$. Applying a general limit theorem in [@KJO], we show that the scaled process $N^{ - \frac{3}{5}} Z(Nt)$ converges to a Lévy process generated by $- (-\Delta)^{\frac{5}{6}}$ (up to a constant) as Theorem \[thm:main3\]. The difference of the exponents between $\frac{3}{4}$ (obtained in [@KJO; @JKO] for the original momentum exchange model) and $\frac{5}{6}$ is explained by the asymptotic behavior of the derivative of the dispersion relation $\omega'(k)$ and the mean value of the scattering kernel $R(k) = \int_{\T} R(k,k')dk'$ as $k \to 0$. (We abbreviate the term $i,j$.) Roughly speaking, if $$\begin{aligned} \omega'(k) \sim k^{a}, ~ R(k) \sim k^{b} ~ \textit{as} ~ k \to 0\end{aligned}$$ for some $a,b \in \N_{\ge 0}$, by applying the argument in [@KJO] formally, one will obtain a Lévy process generated by $- (- \Delta)^{\frac{b+1}{2(b-a)}}$ as a proper scaling limit if $0 < \frac{b+1}{2(b-a)} < 1$ and by $\Delta$ if $\frac{b+1}{2(b-a)} \ge 1$. For the original momentum exchange model presented in [@BOS] and [@JKO] $$\begin{aligned} \omega'(k) \sim 1, ~ R(k) \sim k^{2} ~ \textit{as} ~ k \to 0,\end{aligned}$$ while in our model $$\begin{aligned} \omega'(k) \sim k, ~ R(k) \sim k^{4} ~ \textit{as}~ k \to 0.\end{aligned}$$ In particular, our model has the vanishing sound speed since $\lim_{k \to 0}\omega'(k)=0$. To be more precise, in our model $R(k,i) = \sum_{j=1}^2\int_{\T} R(k,i,k',j)dk'$ satisfies $R(k,1) \sim k^{2}$ and $R(k,2) \sim k^{4}$ (or $R(k,2) \sim k^{2}$ and $R(k,1) \sim k^{4}$ depending on the sign of the magnetic field) and the latter dominates the macroscopic evolution. Note that for a class of non-acoustic chains introduced in [@KO], $$\begin{aligned} \omega'(k) \sim k, ~ R(k) \sim k^{2} ~ \textit{as} ~ k \to 0\end{aligned}$$ and so its macroscopic evolution is diffusive. A technically crucial idea of our proof of Theorem \[thm:main\] is that we consider the microscopic local density of energy, called the Wigner distribution in physics, associated to the eigenvectors of the deterministic dynamics including the effect of the magnetic field. If we employ the classical wave functions which are the eigenvectors of the harmonic Hamiltonian dynamics (without a magnetic field) and study its associated Wigner distribution, then we obtain a *system* of Boltzmann equations as the weak noise limit. However, so far we do not know how to rescale the solutions of the system and derive the fractional diffusion equation from it. By employing the modified wave functions, instead of the classical wave functions, we obtain a *single* limiting Boltzmann equation which is much easier to analyze. This strategy can be applied to derive the limiting equation from other Hamiltonian systems with some energy-conservative external field. Our paper is organized as follows: In Section 2 we prepare some notations. In Section 3 we introduce our model, wave functions and its associated Wigner distribution. Note that since we consider the infinite system, we need to define our model in terms of wave functions to make the argument rigorous. In Section 4 we state our main results, Theorem \[thm:main\] and \[thm:main2\]. We study a Markov process associated to our Boltzmann equation and its scaling limit in Section 5. Proofs of Theorem \[thm:main\] and \[thm:main2\] are given in Sections 6 and 7 respectively. Notations ========= Let $\T \cong [-\frac{1}{2},\frac{1}{2})$ be the one-dimensional torus. For $f \in \ell^{2}(\Z)$, we introduce the discrete Laplacian $\Delta f : \Z \to \R$ defined by $$\begin{aligned} \Delta f (x) &= f(x+1) + f(x-1) - 2 f(x)\end{aligned}$$ and its Fourier transform $\widehat{f} \in L^{2}(\T) $ defined by $$\begin{aligned} \widehat{f}(k) &= \sum_{x \in \Z} e^{-2 \pi \sqrt{-1} k x} f(x).\end{aligned}$$ For functions $f,g \in \ell^{2}(\Z)$, the discrete convolution $f * g :\Z \to \R$ is defined by $$\begin{aligned} f *g(x) &= \sum_{z \in \Z} f(x-z)g(z).\end{aligned}$$ For $J : \R \times \T \to \mathbb{C}$ such that $J(y,k)$ is rapidly decreasing in $y \in \R$, we define $\widehat{J} : \R \times \T \to \mathbb{C}$ as $$\begin{aligned} \widehat{J}(p,k) = \int_{\R} dy ~ e^{-2 \pi \sqrt{-1} p y} J(y,k).\end{aligned}$$ Let $\mathbf{S}$ be the space of rapidly decreasing functions on $ \R \times \T$ defined by $$\begin{aligned} \mathbf{S} =\{ J \in C^{\infty}( \R \times \T ,\mathbb{C}) \ ; \ |J|_{m,n} < \infty \quad \forall m, n \in \Z_{\ge 0} \}\end{aligned}$$ where $$\begin{aligned} |J|_{m,n} = \sup_{r,s \le m} \sup_{y \in \R , k \in \T} (1 + y^{2})^{n}|\partial_{y}^{r} \partial_{k}^{s} J(y, k)|.\end{aligned}$$ We introduce a norm $||\cdot||$ on $\mathbf{S}^{2} = \mathbf{S} \times \mathbf{S}$ defined by $$\begin{aligned} ||\boldsymbol{J}|| = \sum_{i = 1 , 2} \int_{\R} dp \sup_{k} |\widehat{J_{i}}(p,k)| \end{aligned}$$ for $\boldsymbol{J}=(J_{1},J_{2}) \in \mathbf{S}^{2}$ and define a topology on $\mathbf{S}^{2}$ induce by the norm $||\cdot||$. By $(\mathbf{S}^2)'$ we denote the dual space of $\mathbf{S}^2$ equipped with the weak-$*$ topology. For two functions $f(k)$ and $g(k)$ defined on $\T$ or $\T \setminus \{0\}$, we denote by $f(k) \sim g(k)$ as $k \to 0$ if there exists a constant $C>0$ such that for all $k$ whose absolute value is small enough, $\frac{1}{C}|g(k)| \le |f(k)| \le C|g(k)|$. The Dynamics ============ We consider the one-dimensional infinite chain of coupled charged harmonic oscillators in two-dimensional space with weak continuous noise. Since the dynamics involves infinite number of particles, we give a formal description of the deterministic dynamics in Section 3.1, a formal construction of the associated wave functions in Section 3.2 and a formal description of the stochastic perturbation in Section 3.3. They are rigorous when we consider a finite chain. Then we present a proper definition of the dynamics in Section 3.4. In Section 3.5 we introduce the Wigner distribution associated to our wave functions. Deterministic dynamics. ----------------------- We consider a one-dimensional chain of oscillators in a magnetic field. Our deterministic dynamics $ ( \mathbf{v}_x(t), \mathbf{q}_x(t) ) \in \R^{2} \times \R^{2}$ is formally given as follows: $$\begin{aligned} \label{eq:dynamics} \begin{cases} \frac{d}{dt} q_x^i & =v_x^i \\ \frac{d}{dt} v_x^i & =[\Delta q^i]_x +\delta_{i,1}Bv^2_x-\delta_{i,2}B v^1_x \end{cases}\end{aligned}$$ for $x \in \Z, i =1,2$ where $B \in \R \setminus \{0\}$ is the strength of the magnetic field. The total energy $E$ of the system is formally given by $$\begin{aligned} E = \sum_{i=1,2} \sum_{x \in \Z } \left( \frac{|v_x^i|^{2}}{2} + \frac{|q_x^i - q_{x+1}^i|^{2}}{2} \right).\end{aligned}$$ We introduce operators $A$ and $G$ as follows: $$\begin{aligned} A & = \sum_{i=1,2} \sum_{x \in \Z}( v_{x}^i \partial_{q_{x}^i} + [\Delta q^{i}]_x \partial_{v_{x}^i}) , \\ G & = \sum_{x \in \Z} \big( v_{x}^2 \partial_{v_{x}^1} - v_{x}^1 \partial_{v_{x}^2} \big). \end{aligned}$$ Then our deterministic dynamics formally satisfies $\frac{d}{dt}f( \mathbf{v}, \mathbf{q})=(A+BG)f( \mathbf{v}, \mathbf{q})$ for any smooth cylinder function $f$, that is, $f$ depends on the configuration $( \mathbf{v}, \mathbf{q})$ only through a finite set of coordinates. Let ${\alpha}: \Z \to \R$ be a function that ${\alpha}(0) = 2 $, ${\alpha}(1)= {\alpha}(-1) = -1$ and ${\alpha}(x) = 0, |x| \ge 2$. Using this function, the total energy $E$ and the operator $A$ are also written as follows: $$\begin{aligned} E &= \sum_{i=1,2} \left(\sum_{x \in \Z } \frac{|v_{x}^i|^{2}}{2} + \sum_{ x , x' \in \Z } \frac{{\alpha}(x-x')}{2} q_{x}^iq_{x'}^i \right), \\ A & = \sum_{i=1,2} \left(\sum_{x \in \Z} v_{x}^i \partial_{q_{x}^i} - \sum_{x , x' \in \Z} {\alpha}(x-x')q_{x'}^i \partial_{v_{x}^i}\right).\end{aligned}$$ \[generala\] Suppose that ${\alpha}_* : \Z \to \R$ is a function satisfying the following conditions $(a.1) - (a.4)$. $(a.1) ~ {\alpha}_* (x) \neq 0 $ for some $x \in \Z. $ $(a.2) ~ {\alpha}_* (x) = {\alpha}_* (-x) $ for all $ x \in \Z.$ $(a.3) ~ $ There exist some positive constants $C_{1} , C_{2}$ such that $|{\alpha}_* (x)| \le C_{1}e^{-C_{2}|x|} $ for all $x \in \Z$. $(a.4) ~ \widehat{{\alpha}_* }(k) >0 $ for all $k \neq 0$ , $\widehat{{\alpha}_* }(0) = 0 , \widehat{{\alpha}_* }''(0) > 0$. We can consider the dynamics associated to ${\alpha}_*$, or precisely that given by $A_* + BG$ where $$\begin{aligned} A_* = \sum_{i=1,2} \left( \sum_{x \in \Z} v_{x}^i \partial_{q_{x}^i} - \sum_{x , x' \in \Z} {\alpha}_* (x-x')q_{x'}^i\partial_{v_{x}^i}\right).\end{aligned}$$ Then, Theorem 1, 2, and 3 are generalized to this dynamics (with stochastic perturbation) by replacing ${\alpha}$ with ${\alpha}_* $. The generalization from ${\alpha}$ to ${\alpha}_* $ is straightforward, so we omit the proof. Wave functions -------------- To define our dynamics rigorously and then introduce the Wigner distribution, we consider the Fourier transform of the configuration $( \mathbf{v}, \mathbf{q})$. From the formal description of the dynamics , the time evolution of the deterministic process $( \widehat{\mathbf{v}}(k,t), \widehat{\mathbf{q}}(k,t) )$ should be given by $$\begin{aligned} \label{eq:dynamicsinF} \partial_{t} ~ \begin{pmatrix} \widehat{q^1}(k,t) \\ \widehat{q^2}(k,t) \\ \widehat{v^1}(k,t) \\ \widehat{v^2}(k,t) \end{pmatrix} = M(k) ~ \begin{pmatrix} \widehat{q^1}(k,t) \\ \widehat{q^2}(k,t) \\ \widehat{v^1}(k,t) \\ \widehat{v^2}(k,t) \end{pmatrix} , \\ M(k) = \begin{pmatrix} 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ -\widehat{{\alpha}}(k) & 0 & 0 & B \\ 0 & -\widehat{{\alpha}}(k) & - B & 0 \end{pmatrix} \notag,\end{aligned}$$ for each $k \in \T$ where $\widehat{{\alpha}}(k) = 2 - 2\cos{2 \pi k}$. Note that the dynamics is well-defined for any initial condition $( \widehat{\mathbf{v}}(k,0), \widehat{\mathbf{q}}(k,0) )$ for each $k \in \T$. We denote the eigenvalues of the matrix $M(k)$ by $\{ \pm \sqrt{-1} {\omega}_i(k) , i = 1,2 \}$, which are explicitly given as $$\begin{aligned} {\omega}_1(k) &= \sqrt{\widehat{{\alpha}}(k) + \frac{B^{2}}{4}} + \frac{B}{2} , \\ {\omega}_2(k) &= \sqrt{\widehat{{\alpha}}(k) + \frac{B^{2}}{4}} - \frac{B}{2} .\end{aligned}$$ Note that ${\omega}_i(k), \omega'_i(k), ~ i=1,2$ are bounded in $k \in \T$ and $\omega'_1 = \omega'_2$. Denote by $\omega'(k)$ the common value of $\omega'_i(k)$. We introduce the corresponding wave functions $\{ \widehat{\psi_{i}}(k,t) ; i = 1,2 \}$ given by $$\begin{aligned} \label{defofpsi1} \widehat{\psi_{1}}(k,t) &= \theta_{1}(k)(\widehat{v^1}(k,t) - \sqrt{-1}{\omega}_2(k)\widehat{q^1}(k,t) + \sqrt{-1}\widehat{v^2}(k,t) + {\omega}_2(k)\widehat{q^2}(k,t)) ,\notag \\ \widehat{\psi_{2}}(k,t) &= \theta_{2}(k)(\widehat{v^1}(k,t) - \sqrt{-1}{\omega}_1(k)\widehat{q^1}(k,t) - \sqrt{-1}\widehat{v^2}(k,t) - {\omega}_1(k)\widehat{q^2}(k,t))\end{aligned}$$ with $$\begin{aligned} \theta_{i}(k) &= \sqrt{\frac{{\omega}_i(k)}{{\omega}_1(k)+{\omega}_2(k)}} , ~ i=1,2. \end{aligned}$$ $\widehat{\psi_{i}}(k)$ is the eigenfunction associated to the eigenvalue $- \sqrt{-1} {\omega}_i(k)$ : $$\begin{aligned} \partial_t \widehat{\psi_{i}}(k) = - \sqrt{-1} {\omega}_i(k) \widehat{\psi_{i}}(k) , ~ i=1,2.\end{aligned}$$ We normalize $\widehat{\psi}$ by multiplying $\theta_{i}$ so that the total energy $E$ is given by the integral of the $L^2$ norm of the wave functions as $$\begin{aligned} E & = \frac{1}{2} \int_{\T} dk ~ \left(|\widehat{v^{1}}(k)|^{2}+|\widehat{v^{2}}(k)|^{2}+\widehat{{\alpha}}(k)(|\widehat{q^{1}}(k)|^{2}+|\widehat{q^{2}}(k)|^{2}) \right) \\ &= \frac{1}{2}\int_{\T} dk ~ \left(|\widehat{\psi_{1}}(k)|^{2} + |\widehat{\psi_{2}}(k)|^{2}\right).\end{aligned}$$ By a direct computation we have $$\begin{aligned} \label{repbypsi} \widehat{v_{1}}(k) &= \frac{\theta_{1}(k)}{2}(\widehat{\psi_{1}}(k) + \widehat{\psi_{1}}(-k)^{*}) + \frac{\theta_{2}(k)}{2}(\widehat{\psi_{2}}(k) + \widehat{\psi_{2}}(-k)^{*}) , \notag \\ \widehat{v_{2}}(k) &= - \frac{\sqrt{-1} \theta_{1}(k)}{2}(\widehat{\psi_{1}}(k) - \widehat{\psi_{1}}(-k)^{*}) + \frac{\sqrt{-1} \theta_{2}(k)}{2}(\widehat{\psi_{2}}(k) - \widehat{\psi_{2}}(-k)^{*}) , \notag \\ \widehat{q_{1}}(k) &= \frac{\sqrt{-1} \theta_{1}(k)}{2{\omega}_1(k)}(\widehat{\psi_{1}}(k) - \widehat{\psi_{1}}(-k)^{*}) + \frac{\sqrt{-1} \theta_{2}(k)}{2{\omega}_2(k)}(\widehat{\psi_{2}}(k) - \widehat{\psi_{2}}(-k)^{*}) , \notag \\ \widehat{q_{2}}(k) &= \frac{\theta_{1}(k)}{2{\omega}_1(k)}(\widehat{\psi_{1}}(k) + \widehat{\psi_{1}}(-k)^{*}) - \frac{\theta_{2}(k)}{2{\omega}_2(k)}(\widehat{\psi_{2}}(k) + \widehat{\psi_{2}}(-k)^{*}). \end{aligned}$$ Stochastic perturbation. ------------------------ We consider a local stochastic perturbation of the dynamics which conserves the total energy. We introduce an operator $S$ as follows: $$\begin{aligned} S &= \frac{1}{2} \sum_{x \in \Z} (Y_{x,x+1})^{2} = \frac{1}{4} \sum_{x \in \Z} \sum_{z \in \Z;|x-z| = 1} (Y_{x,z})^{2}, \\ Y_{x , z} &= (v^2_z - v_x^2) (\partial_{v^1_z}-\partial_{v_{x}^1}) -(v^{1}_z - v_{x}^1)(\partial_{v^2_z}-\partial_{v_{x}^2}). \end{aligned}$$ We consider a Markov process $( \mathbf{v}_x(t) , \mathbf{q}_x(t) )$ generated by $L := A + BG +\epsilon \gamma S$. $\gamma > 0$ is the strength of the stochastic noise and $0< \epsilon <1$ is a scale parameter. The dynamics can be also given by the stochastic differential equation $$\begin{aligned} \label{formalz} \begin{cases} d q_x^i & =v_x^i dt \\ d v_x^i & = ( - [{\alpha}* q^{i}]_x +\delta_{i,1}Bv^{2}_x-\delta_{i,2}B v^{1}_x + \epsilon \gamma [\Delta v^{i}]_x ) dt \\ & ~ + \sqrt{\epsilon \gamma} \sum_{z; |z-x| = 1} (Y_{x , z} v_x^i) dw_{x,z} , \end{cases}\end{aligned}$$ for $x \in \Z$, $i=1,2$ where $\{ w_{x,z}(t) = w_{z,x}(t) ; x,z \in \Z, |z-x| = 1 \}$ are independent standard Wiener processes on $\R$. Note that $L$ formally conserves the total energy and the total pseudomomentum $\sum_{x} v_{x}^1 - Bq^2_x, \sum_{x} v_{x}^2 + Bq^1_x$. For more details about the conserved quantities, see [@SS]. This specific choice of noise is not important. Our proof is also applicable for the velocity exchange noise used in [@SS] and yields the same scaling limits. For the construction of this jump-type process, we can follow the argument in Chapter 5 of [@FFL]. Rigorous definition of the dynamics ----------------------------------- In this subsection, we define the dynamics rigorously. First, we calculate the time evolution of the wave functions $\widehat{\psi_{i}}(k,t)$ obtained from the formal description : $$\begin{aligned} \label{formalk} d \widehat{q^i}(k,t) & = \widehat{v^i}(k,t) dt ~ , i=1,2 \notag ,\\ d \widehat{v^1}(k,t) &= (- \widehat{{\alpha}}(k) \widehat{q^1}(k,t) + B\widehat{v^2}(k,t) + \epsilon \gamma \widehat{{\beta}}(k) \widehat{v^1}(k,t) ) dt \notag \\ & ~ - \sqrt{\epsilon \gamma} \int_{\T} r(k,k') \widehat{v^{2}} (k-k',t) W(dk',dt) \notag ,\\ d \widehat{v^2}(k,t) &= (- \widehat{{\alpha}}(k) \widehat{q^2}(k,t) - B\widehat{v^1}(k,t) + \epsilon \gamma \widehat{{\beta}}(k) \widehat{v^2}(k,t) ) dt \notag \\ & ~ + \sqrt{\epsilon \gamma} \int_{\T} r(k,k') \widehat{v^{1}} (k-k',t) W(dk',dt) ,\end{aligned}$$ where $$\begin{aligned} \widehat{{\beta}}(k) &= 2 \cos{2 \pi k} - 2 ,\\ r(k,k') &= (e^{-2 \pi \sqrt{-1} k'} - e^{-2 \pi \sqrt{-1} k})(e^{2 \pi \sqrt{-1} k} - 1) ,\\ W(k,t) &= \sum_{x \in \Z} w_{x,x+1}(t) e^{-2 \pi \sqrt{-1} k x}.\end{aligned}$$ The term with $\widehat{{\beta}}(k)$ comes from the stochastic perturbation. In our case $\widehat{{\alpha}}(k) = - \widehat{{\beta}}(k)$, but in general (cf. Remark \[generala\]) there is no such relation between $\widehat{{\alpha}}$ and $\widehat{{\beta}}$, and so we keep $\widehat{{\alpha}}$ and $\widehat{{\beta}}$ for the generalization. $W$ is called a cylindrical Wiener process on $\mathbb{L}^{2}(\T)$. A precise derivation of from is given in Appendix \[derivation\]. Combining $(\ref{defofpsi1})$ and $(\ref{formalk})$ we have $$\begin{aligned} \label{defofpsi} & d \widehat{\psi_{1}}(k,t) = (- \sqrt{-1}{\omega}_1(k) \widehat{\psi_{1}}(k,t) + \epsilon \gamma \theta_{1}(k) {\beta}(k) ( \theta_{1}(k) \widehat{\psi_{1}}(k,t) + \theta_{2}(k) \widehat{\psi_{2}}(-k,t)^{*}) )dt \notag \\ & ~ + \sqrt{-1} \theta_{1}(k) \sqrt{\epsilon \gamma} \int_{\T} r(k,k') (\theta_{1}(k-k') \widehat{\psi_{1}}(k-k',t) + \theta_{2}(k-k') \widehat{\psi_{2}}(k'-k,t)^{*} ) W(dk',dt) , \notag \\ & d \widehat{\psi_{2}}(k,t) = (- \sqrt{-1}{\omega}_2(k) \widehat{\psi_{2}}(k,t) + \epsilon \gamma \theta_{2}(k) {\beta}(k) ( \theta_{1}(k) \widehat{\psi_{1}}(-k,t)^{*} + \theta_{2}(k) \widehat{\psi_{2}}(k,t)) ) dt \notag \\ & ~ - \sqrt{-1} \theta_{2}(k) \sqrt{\epsilon \gamma} \int_{\T} r(k,k') (\theta_{1}(k-k') \widehat{\psi_{1}}(k'-k,t)^{*} + \theta_{2}(k-k') \widehat{\psi_{2}}(k-k',t) ) W(dk',dt) . \end{aligned}$$ Now we define a stochastic process $\{ \widehat{\boldsymbol{\psi}}(\cdot,t) \in (\mathbb{L}^{2}(\T))^{2} ; t \ge 0 \} $ as the unique solution of . We can show the existence of the solution by using a classical technique, called a fixed-point theorem. For the sketch of the proof, see Appendix \[existence\]. Once we define the dynamics $ \widehat{\boldsymbol{\psi}}(\cdot,t) \in (\mathbb{L}^{2}(\T))^{2} $, then we can also define $\widehat{\mathbf{v}}(k,t)$ by and then define a stochastic process $\{ \mathbf{v}_x(t), \boldsymbol{\psi}(x,t) ; x \in \Z , t \ge 0 \}$ by $$\begin{aligned} &v_x^i(t) = \int_{\T} dk ~ e^{2\pi \sqrt{-1} k x} \widehat{v_{i}}(k,t) , \\ &\psi_i(x,t) = \int_{\T} dk ~ e^{2\pi \sqrt{-1} k x} \widehat{\psi_{i}}(k,t)\end{aligned}$$ for $x \in \Z, i= 1,2$. On the other hand, $\widehat{\mathbf{q}}(\cdot,t)$ is not necessarily well-defined as an element of $(\mathbb{L}^{2}(\T))^2$ because ${\omega}_2(k) \sim k^{2}$ as $k \to 0$ if $B>0$ and ${\omega}_1(k) \sim k^{2}$ as $k \to 0$ if $B<0$. Hence, $\mathbf{q}_x(t)$ are also not necessarily well-defined. Hereafter we do not use the variables $\mathbf{q}_x$. Wigner distribution. -------------------- Let $Q_{\epsilon}$ be a probability measure on $(\mathbb{L}^{2}(\T))^{2} $ which satisfies the following condition: $$\begin{aligned} \label{cond.initial} K_{0} = \sup_{0 < \epsilon < 1} \sum_{i=1,2} \epsilon \int_{\T} dk ~ E_{Q_{\epsilon}}[ |\widehat{\psi_{1}}(k)|^{2} + |\widehat{\psi_{2}}(k)|^{2}] ~ < ~ \infty. \end{aligned}$$ Denote by $\mathbb{E}_{\epsilon}$ the expectation with respect to the distribution of $\{\widehat{\psi_{i}}(\cdot,t)\}_{t \ge 0}$ which starts from $Q_{\epsilon}$. In Appendix \[conservation\], we show that $$\sum_{i=1,2} \mathbb{E}_{\epsilon}[ ||\widehat{\psi_{i}}(\cdot,t)||_{\mathbb{L}_{2}}^{2} ] = \sum_{i=1,2} \mathbb{E}_{\epsilon} [ ||\widehat{\psi_{i}}(\cdot,0)||_{\mathbb{L}_{2}}^{2} ]$$ for any $t \ge 0$. In particular, under the condition $(\ref{cond.initial})$ $$\begin{aligned} \label{ebound} \sup_{0 < \epsilon < 1} \sum_{i=1,2} \epsilon \int_{\T} dk ~ \mathbb{E}_{\epsilon}[ |\widehat{\psi_{1}}(k,t)|^{2} + |\widehat{\psi_{2}}(k,t)|^{2} ] ~ = K_{0} ~ < ~ \infty\end{aligned}$$ for any time $t \ge 0$. For the wave function $\boldsymbol{\psi}$, we introduce the averaged Wigner function as in Section 3 of [@BOS]. We denote the Wigner distribution on the time scale $\epsilon^{-1}t$ by $\Omega^{\epsilon}(t)$ with $\epsilon$ the small semiclassical parameter. Namely, we define $\Omega^{\epsilon}(t) \in (\mathbf{S}^2)'$ by $$\begin{aligned} &<\Omega^{\epsilon}(t),\boldsymbol{J}> ~ = \sum_{i=1,2} <\Omega_{i}^{\epsilon}(t),J_{i}> $$ for $\boldsymbol{J}=(J_{1},J_{2}) \in \mathbf{S}^{2}$ with $$\begin{aligned} \label{defofWigner} &<\Omega_{i}^{\epsilon}(t),J> \notag \\ &= \frac{\epsilon}{2} \sum_{x,x' \in \Z} \mathbb{E}_{\epsilon}[\psi_{i}(x',\frac{t}{\epsilon})^{*} \psi_{i}(x,\frac{t}{\epsilon}) ] \int_{\T} dk ~ e^{2\pi \sqrt{-1} (x'-x) k} J(\frac{\epsilon}{2} (x+x'),k)^{*} \notag \\ &= \frac{\epsilon}{2} \int_{\R} dp \int_{\T} dk ~ \mathbb{E}_{\epsilon} [\widehat{\psi_{i}}(k-\frac{\epsilon p}{2},\frac{t}{\epsilon})^{*} ~ \widehat{\psi_{i}}(k+\frac{\epsilon p}{2},\frac{t}{\epsilon}) ] \widehat{J}(p,k)^{*} $$ for $J \in \mathbf{S}$. By the Cauchy-Schwarz inequality and $(\ref{ebound})$, $$\begin{aligned} \label{lbb} \sup_{0 < \epsilon < 1} \sup_{t \ge 0} |<\Omega^{\epsilon}(t),\boldsymbol{J}>| ~ \le ~ \frac{1}{2} K_{0}||\boldsymbol{J}||\end{aligned}$$ under the condition $(\ref{cond.initial})$. As discussed in [@BOS], $\Omega^{\epsilon}(\cdot)$ is well-defined on a wider class of test functions than $\mathbf{S}^2$. In particular we can take $\boldsymbol{J}(y,k) =(J(k),J(k))$ with a bounded function $J(k)$ on $\T$, and then we have $$\begin{aligned} <\Omega^{\epsilon}(t),\boldsymbol{J}> = \frac{\epsilon}{2} \int_{\T} dk ~ \sum_{i=1,2}\mathbb{E}_{\epsilon} [|\widehat{\psi_{i}}(k,\frac{t}{\epsilon})|^{2} ] J(k).\end{aligned}$$ From this representation one can see that $\Omega^{\epsilon}(\cdot)$ is the distribution of the spectral density of the energy. Also if we take $\boldsymbol{J}(y,k) =(J(y),J(y))$ with a rapidly decreasing function $J(y)$ on $\R$ as a test function, then we have $$\begin{aligned} <\Omega^{\epsilon}(t),\boldsymbol{J}> = \frac{\epsilon}{2} \sum_{x \in \Z} \sum_{i=1,2}\mathbb{E}_{\epsilon} [ |\psi_{i}(x,\frac{t}{\epsilon})|^{2} ] J(\epsilon x).\end{aligned}$$ This is the integral of $J$ with respect to the averaged empirical measure of $\frac{1}{2} \sum_{i=1,2}|\psi_{i}(x,\frac{t}{\epsilon})|^2$. Namely, $ \Omega^{\epsilon}(t)$ is a rescaled microscopic local spectral density. Main results ============ As mentioned in the Introduction, the main purpose of the present paper is to understand the nature of the superdiffusion for the coupled charged harmonic chain of oscillators in a magnetic field with noise defined in the last section, and we apply the two-step scaling limits. In Subsection 4.1, following the idea of [@BOS], we claim that in the weak noise limit the local density of energy is governed by a phonon linear Boltzmann equation. In Subsection 4.2, we consider a properly scaled solution of the Boltzmann equation and state that it converges to the solution of the fractional diffusion equation with $s=\frac{5}{3}$, which is our main result. Boltzmann equation ------------------ In this subsection we state the limiting behavior of the Wigner distribution. \[thm:main\] Suppose the condition holds. If $\Omega^{\epsilon}(0)$ converges to $\Omega_0$ in $(\mathbf{S}^2)'$ as $\epsilon \to 0$, then for all $t \ge 0$, $\Omega^{\epsilon}(t)$ converges to a vector-valued finite positive measure $\boldsymbol{\mu}(t)=(\mu_{1}(t),\mu_{2}(t))$ in $(\mathbf{S}^2)'$ as $\epsilon \to 0$, which is the unique solution of the following Boltzmann equation $$\begin{aligned} \label{mboltzmann} \begin{cases} \partial_{t} \int d \boldsymbol{\mu}(t) \cdot \boldsymbol{J} = \frac{1}{2\pi} \int d\boldsymbol{\mu}(t) \cdot \omega' \partial_{y} \boldsymbol{J} + \gamma \int d\boldsymbol{\mu}(t)\cdot C\boldsymbol{J} \\ \int d \boldsymbol{\mu}(0) \cdot \boldsymbol{J} = < \Omega_0 , \boldsymbol{J}>, \end{cases}\end{aligned}$$ where $$\begin{aligned} \int d \boldsymbol{\mu} \cdot \boldsymbol{J} &= \sum_{i=1,2} \int_{\R \times \T} \mu_{i} (dy,dk) ~ J_{i}(y,k)^* \quad \text{for} \quad \boldsymbol{\mu}=(\mu_1,\mu_2),\\ (C\boldsymbol{J})_{i}(x,k) &= \sum_{j = 1,2} \int_{\T} dk' \theta_{i}(k)^{2} R(k,k') \theta_{j}(k')^{2} (J_{j}(x,k')-J_{i}(x,k)) \end{aligned}$$ for $\boldsymbol{J}=(J_1,J_2)\in \mathbf{S}^{2}$ with $R(k,k') = 16\sin^{2}{\pi k}\sin^{2}{\pi k'}$. In the case $B = 0$, if we assume an additional assumption $$\begin{aligned} \lim_{\rho \to 0} \limsup_{\epsilon \to 0} \frac{\epsilon}{2} \sum_{i=1}^2\int_{|k| < \rho} dk ~ E_{Q_{\epsilon}}[|\widehat{\psi_{i}}(k)|^{2}] = 0 \end{aligned}$$ on the initial measure $Q_{\epsilon}$, the same statement of Theorem \[thm:main\] holds. For this case, the proof is essentially given in [@BOS]. Suppose that the solution of (\[mboltzmann\]) has the density $u(y,k,i,t)$ for all $t \in [0,T]$ , that is, $$\begin{aligned} \mu_{i}(t)(dy,dk) &= u(y,k,i,t) dy dk ~ , ~ i = 1,2 , \\ \mu_{i}(0)(dy,dk) &= u_{0}(y,k,i) dy dk ~ , ~ i = 1,2 .\end{aligned}$$ Then $u(y,k,i,t)$ is a weak solution of the linear Boltzmann equation $$\begin{aligned} \label{boltzmann} \begin{cases} \partial_{t} u(y,k,i,t) + \frac{1}{2\pi} \omega'(k) \partial_{y} u(y,k,i,t) = \gamma \mathcal{L} u(y,k,i,t) \\ u(y,k,i,0) = u_{0}(y,k,i) , \\ \end{cases}\end{aligned}$$ where $$\begin{aligned} \mathcal{L}u(y,k,i,t) = \sum_{j= 1,2} \int_{\mathbf{T}} dk' \theta_{i}(k)^2 R(k,k') \theta_{j}(k')^2 (u(y,k',j,t)-u(y,k,i,t) ) .\end{aligned}$$ We prove Theorem \[thm:main\] in Section \[sec:proof1\]. The strategy of our proof is as follows: First we derive a microscopic evolution equation of $\Omega^{\epsilon}$, which is not closed in terms of $\Omega^{\epsilon}$. Then, with this expression of the time evolution, we show that for any fixed $T>0$, $\{ \Omega^{\epsilon}(t) , 0 \le t \le T \}_{0 < \epsilon < 1}$ is sequentially compact in $C([0,T];(\mathbf{S}^2)')$ in a certain weak-$*$ sense. See its precise meaning in Section \[sec:proof1\]. We verify that any limit of a convergent subsequence is extended to a vector-valued finite positive measure in Appendix \[positivity\]. The uniqueness of the bounded solution of $(\ref{mboltzmann})$ in the class of vector-valued finite positive measures is shown in Appendix \[uniqueness\]. Finally we show that any limit of a convergent subsequence satisfies $(\ref{mboltzmann})$, which is a closed equation in terms of $\boldsymbol{\mu}$. Summarizing the above we can show that $( \Omega^{\epsilon}(\cdot) )_{\epsilon}$ is convergent and the limit satisfies $(\ref{mboltzmann})$. Derivation of the $\frac{5}{6}$ fractional diffusion equation. -------------------------------------------------------------- In this subsection we study a macroscopic behavior of a solution of properly scaled Boltzmann equation . Consider a spatially scaled linear Boltzmann equation with a scaling parameter $N$ as $$\begin{aligned} \label{boltzmannN} \begin{cases} \partial_{t} u(y,k,i,t) + \frac{1}{N^{3/5}}\frac{1}{2\pi} \omega'(k) \partial_{y} u(y,k,i,t) = \gamma \mathcal{L} u(y,k,i,t) \\ u(y,k,i,0) = u_{0}(y,k,i) , \\ \end{cases}\end{aligned}$$ and denote its solution by $u_N$. For any given $u_{0}(y,k,i) \in C_{0}^{\infty}(\R \times \T) , ~ i=1,2$, a solution of is constructed explicitly using a Markov process associated to the Boltzmann equation in the next section. The uniqueness of solutions in a certain class follows from that of (\[mboltzmann\]). The argument also applies to and so the existence and uniqueness of $u_N$ follows. \[thm:main2\] Suppose $u_{0}(y,k,i) \in C_{0}^{\infty}(\R \times \T) , ~ i=1,2$. Define the initial local density of energy at $y \in \R$ as $\bar{u}_0(y) =\sum_{i=1,2} \int_{\T \times \{ 1,2 \} } dk ~ u_{0}(y,k,i)$. Then, for all $y \in \R, ~ t \ge 0$, $$\begin{aligned} \lim_{N \to \infty} \sum_{i=1,2} \int_{\T} dk ~ |u_{N}(y,k,i,Nt) - \frac{1}{2}\bar{u}(y,t)|^2 = 0 ,\end{aligned}$$ where $\bar{u}$ is a solution of $$\begin{aligned} \label{boltzmann2} \begin{cases} \partial_{t} \bar{u}(y,t) = - D (-\Delta_{y})^{\frac{5}{6}} \bar{u}(y,t) \\ \bar{u}(y,0) = \bar{u}_0(y) \end{cases}\end{aligned}$$ and $D = D(B,\gamma,{\alpha}) $ is a positive constant such that $$\begin{aligned} D = C |B|^{-\frac{1}{3}} \gamma^{-\frac{2}{3}} \widehat{{\alpha}}''(0) \end{aligned}$$ with a universal constant $C$. In particular, $$\begin{aligned} \lim_{N \to \infty} ~ | \sum_{i=1,2} \int_{\T} dk \ u_{N}(y,k,i,Nt) - \bar{u}(y,t)|^2 = 0.\end{aligned}$$ In the case $B = 0$, if we denote by $u_N(y,k,i,t)$ the solution of a scaled linear Boltzmann equation $$\begin{aligned} \begin{cases} \partial_{t} u(y,k,i,t) + \frac{1}{N^{2/3}} \frac{1}{2\pi} \omega'(k) \partial_{y} u(y,k,i,t) = \gamma \mathcal{L} u(y,k,i,t) \\ u(y,k,i,0) = u_{0}(y,k,i) , \end{cases}\end{aligned}$$ then for all $y \in \R,~ t \ge 0$, $$\begin{aligned} \lim_{N \to \infty} \sum_{i=1,2} \int_{\T} dk ~ |u_{N}(y,k,i,Nt) - \frac{1}{2}\bar{u}(y,t)|^2 = 0 \end{aligned}$$ where $\bar{u}$ is the solution of $$\begin{aligned} \begin{cases} \partial_{t} \bar{u}(y,t) = - D' (-\Delta_{y})^{\frac{3}{4}} \bar{u}(y,t) \\ \bar{u}(y,0) = \bar{u}_0(y) . \end{cases}\end{aligned}$$ and $D' = D'(\gamma,{\alpha}) $ is a positive constant such that $$\begin{aligned} D' = C'\gamma^{-\frac{1}{2}} (\widehat{{\alpha}}''(0))^{\frac{3}{4}}\end{aligned}$$ with a universal constant $C'$. The result is essentially proved in [@KJO]. For the proof, we follow the strategy of [@KJO]. Namely, we consider a long-time asymptotic behavior of a Markov process associated to the Boltzmann equation and then use a functional limit theorem for a continuous time random walk. To apply a general theorem in [@KJO], we need to check several conditions. This is the main subject of the next section, where we conclude all the required conditions are satisfied and then Theorem \[thm:main3\] on the asymptotic behavior of a Markov process is obtained. We apply it to the study of the limit of $u_N$ and prove Theorem \[thm:main2\] in Section \[sec:proof2\]. Markov process associated to the Boltzmann equation {#markov} =================================================== In this section we construct a solution of $(\ref{boltzmann})$ probabilistically. We will see that there exists a Markov process associated to $(\ref{boltzmann})$ and study its long-time asymptotic behavior. Let $\{ ( K_{n} , I_{n} ) ; n \in \Z_{\ge 0} \}$ be a Markov chain on $\T \times \{ 1 , 2 \}$ whose transition probability is given by $$\begin{aligned} P(k,i,dk',j) = t(k,i) \gamma \theta_{i}(k)^2 R(k,k') \theta_{j}(k')^2 dk',\end{aligned}$$ where $$\begin{aligned} t(k,i) &= [ \gamma \theta_{i}(k)^2 R(k) ]^{-1} , \quad R(k) = \int_{\T} dk' R(k,k').\end{aligned}$$ Since $R(k,k')$ is a product of functions of $k$ and $k'$, we have $$\begin{aligned} P(k,i,dk',j) = \pi(dk',j)\end{aligned}$$ where $\pi(dk,di)$ is a reversible measure for this Markov chain given as $$\begin{aligned} \pi(dk,di) = \sum_{j=1,2} \frac{t(k,j)^{-1}}{\gamma \overline{R}} dk \delta_{ \{ j \} }(di) , \quad \overline{R} = \int_{\T} dk ~ R(k).\end{aligned}$$ In particular, $\{ ( K_{n} , I_{n} ) ; n \ge 1 \}$ is an i.i.d. sequence of random variables on $\T \times \{ 1 , 2 \}$ with distribution $\pi$. Now we construct a continuous time random walk generated by $\mathcal{L}$. Let $\{ \tau_{n} , n \ge 1 \}$ be an i.i.d. sequence of random variables such that $\tau_{1}$ is exponentially distributed with intensity 1 and $\{ ( K_{n} , I_{n} ) ; n \in \Z_{\ge 0} \}$ and $\{ \tau_{n} , n \ge 1 \}$ are independent. Set $t_{n} := \sum_{m = 1}^{n} t(K_{m-1},I_{m-1}) \tau_{m} , n \ge 1 , ~ t_{0} = 0 $ and define a stochastic process $( K(t) , I(t) ) $ as $K(t) = K_{n} , I(t) = I_{n} $ if $t \in [t_{n},t_{n+1})$. Then, by the construction $\{ ( K(t) , I(t) ) \}_{t \ge 0}$ is a continuous time random walk generated by $\mathcal{L}$. With this process we can construct an explicit solution of the equation (\[boltzmann\]) by $$\begin{aligned} u(y,k,i,t) &= \mathbb{E}_{(k,i)}[u_{0}(Z(t),K(t),I(t))] , \intertext{where} Z(t) &= y + \frac{1}{2\pi} \int_{0}^{t} ds ~ \omega'(K(s)) ,\end{aligned}$$ and $K(0) = k , I(0) = i$. For this process, we have the following result. \[thm:main3\] Suppose $(K(0), I(0))=(k,i)$ for some $k \neq 0$ and $i=1$ or $2$. Then as $N \to \infty$, the finite-dimensional distribution of scaled processes converge weakly to a Lévy process generated by $- D (-\Delta_{y})^{\frac{5}{6}}$, where $D = D(B,\gamma,{\alpha}) $ is a positive constant such that $$\begin{aligned} D = C |B|^{-\frac{1}{3}} \gamma^{-\frac{2}{3}} \widehat{{\alpha}}''(0),\end{aligned}$$ and C is a positive constant which does not depend on B, $\gamma$, ${\alpha}$. In the case of $B = 0$,the finite-dimensional distribution of scaled processes converge weakly to a Lévy process generated by $ - D' (-\Delta_{y})^{\frac{3}{4}}$, where $D' = D'(\gamma,{\alpha}) $ is a positive constant such that $$\begin{aligned} D' = C'\gamma^{-\frac{1}{2}} (\widehat{{\alpha}}''(0))^{\frac{3}{4}} ,\end{aligned}$$ and $C'$ is a positive constant which does not depend on $\gamma$, ${\alpha}$. It is essentially shown in [@KJO]. Proof of Theorem \[thm:main3\]. {#sec:proof3} ------------------------------- We apply [@KJO Theorem 2.8 (i)] to our process with ${\alpha}=\frac{5}{3}$. For this, it is enough to show that Conditions 2.1, 2.2, 2.3 and (2.12) of [@KJO] are satisfied. First we verify that Condition 2.1 is satisfied. Define $$\begin{aligned} \Psi(k,i) = \omega'(k) t(k,i) .\end{aligned}$$ The tail of $\Psi$ under $\pi$ is $$\begin{aligned} \pi(\{ (k,i) ; \Psi(k,i) \ge \lambda \}) & = \sum_{i = 1,2} \int_{ \{ k ; \Psi(k,i) \ge \lambda \} } dk \ \frac{\theta_{i}(k)^2R(k)}{\overline{R}} \\ &= C |B|^{- \frac{1}{3}} \gamma^{- \frac{5}{3}} \widehat{{\alpha}}''(0) \lambda^{-\frac{5}{3}}(1 + O(\lambda^{-\frac{4} {3}}) ) ,\end{aligned}$$ as $\lambda \to \infty$ because $$\begin{aligned} &\theta_{1}(k)^2 \sim 1, \ \textit{and } \ \theta_{2}(k)^2 \sim \frac{\widehat{{\alpha}}''(0)k^{2}}{|B|^{2}} \textit{as} \ k \to 0 \ \textit{if} \ B>0 \\ &\theta_{1}(k)^2 \sim \frac{\widehat{{\alpha}}''(0)k^{2}}{|B|^{2}} \ \textit{and } \ \theta_{2}(k)^2 \sim 1 \textit{as} \ k \to 0 \ \textit{if} \ B<0 \ \\\end{aligned}$$ and $$\begin{aligned} \omega'(k) \sim \frac{\widehat{{\alpha}}(0)'' k}{|B|}, \quad R(k) \sim k^{2} \quad \textit{as} \ k \to 0.\end{aligned}$$ $C$ is a positive constant which does not depend on $B,\gamma,{\alpha}$. $\Psi$ is odd for $k$ and the density of $\pi(\cdot,i)$ with respect to the Lebesgue measure is even for k, so $$\pi(\{ (k,i) ; \Psi(k,i) \ge \lambda \}) = \pi(\{ (k,i) ; \Psi(k,i) \le -\lambda \})$$ and $ \int \Psi d\pi = 0$. Next we check Condition 2.2. It is obvious that $$\begin{aligned} \sup \{ ||Pf||_{\mathbb{L}^{2}(\pi)} ; \int d\pi ~ f = 0 , ||f||_{\mathbb{L}^{2}(\pi)} = 1 \} = 0\end{aligned}$$ because $Pf = \int d\pi ~ f$. Finally we show that Condition 2.3 and (2.12) hold. Condition 2.3 is obviously satisfied with $Q \equiv 0$ and $p \equiv 1$. Also, we have $$\begin{aligned} ||P\Psi||_{\mathbb{L}^{2}(\pi)}^{2} &= \sum_{i=1,2} \int_{\T} dk \left(\sum_{j=1,2} \int_{\T} dk' \Psi(k',j) \frac{t(k',j)^{-1}}{\gamma \overline{R}}\right)^{2} \frac{t(k,i)^{-1}}{\gamma \overline{R}} \\ &= \sum_{i=1,2} \int_{\T} dk \left(\sum_{j=1,2} \int_{\T} dk' \frac{\omega'(k')}{\gamma \overline{R}} \right)^{2} \frac{t(k,i)^{-1}}{\gamma \overline{R}} < \infty.\end{aligned}$$ Therefore, by [@KJO Theorem 2.8 (i)], the finite-dimensional distributions of the scaled process $\{ N^{-\frac{3}{5}} Z(Nt)\}_{t \ge 0 }$ under $\mathbb{P}_{\pi}$ converge to a stable process with exponent $\frac{5}{3}$ whose characteristic function at time 1, denoted by $\phi(x)$ is $$\begin{aligned} \phi(x) = \exp{(\int_{\R} d\lambda ~ e^{\sqrt{-1}\lambda x} c_{*}(\lambda) |\lambda|^{-\frac{8}{3}})},\end{aligned}$$ where $$\begin{aligned} c_{*}(\lambda) = \frac{5C |B|^{- \frac{1}{3}} \gamma^{- \frac{5}{3}} \widehat{{\alpha}}''(0) A_{\frac{5}{3}}}{\bar{t}}\end{aligned}$$ for all $\lambda \neq 0$, C is a positive constant appeared in the tail estimate of $\Psi$ and $$\begin{aligned} A_{\frac{5}{3}} = \int_{0}^{\infty} dy ~ e^{-y} y^{\frac{5}{3}}, \\ \bar{t} = \int d\pi ~ t(k,i) ~ = \frac{1}{2\gamma}. \end{aligned}$$ Finally we show that the finite-dimensional distributions of $\{ N^{-\frac{3}{5}} Z(Nt) ; t \ge 0 \}$ under $\mathbb{P}_{(k,i)}$ also converge to the same stable process for $k \in \T \setminus \{ 0 \} , ~ i=1,2$. For $t \ge 0$ define $n(t)$ as the nonnegative integer such that $$\begin{aligned} t_{n(t)} \le t < t_{n(t) + 1}.\end{aligned}$$ Then we have $$\begin{aligned} N^{-\frac{3}{5}} Z(Nt) = N^{-\frac{3}{5}} \sum_{n = 0}^{n(Nt)} \Psi(K_{n},I_{n}) \tau_{n+1}.\end{aligned}$$ If $k \neq 0$ then $N^{-\frac{3}{5}} \Psi(k,i) \tau_{1} \to 0 $ as $N \to \infty$ $\mathbb{P}_{(k,i)}$ - almost surely. Moreover, under $\mathbb{P}_{(k,i)}$, the distribution of $\{ ( K_{n} , I_{n} )\}_{n \ge 1 }$ is an i.i.d. sequence with distribution $\pi$. By Theorem 6.1 and Lemma 6.2 of [@KJO], the finite-dimensional distributions of $\{ N^{-\frac{3}{5}} \sum_{n = 1}^{n(Nt)} \Psi(K_{n},I_{n}) \tau_{n+1} ; t \ge 0 \}$ under $\mathbb{P}_{(k,i)}$ converge to the stable process, so the finite-dimensional distributions of $\{ N^{-\frac{3}{5}} Z(Nt) ; t \ge 0 \}$ under $\mathbb{P}_{(k,i)}$ also converge to the same stable process if $k \neq 0$. Proof of the Theorem \[thm:main\]. {#sec:proof1} ================================== To simplify the notation, we define functions $\widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k)$, $\widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k)$ on $\R \times \T$ by $$\begin{aligned} \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k) &= \frac{\epsilon}{2} \mathbb{E}_{\epsilon} [ \widehat{\psi_{i}}(k-\frac{\epsilon p}{2},\frac{t}{\epsilon})^{*} ~ \widehat{\psi_{i}}(k+\frac{\epsilon p}{2},\frac{t}{\epsilon}) ] , \notag \\ \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k) &= \frac{\epsilon}{2} \mathbb{E}_{\epsilon} [ \widehat{\psi_{i}}(-k+\frac{\epsilon p}{2},\frac{t}{\epsilon}) ~ \widehat{\psi_{i^{*}}}(k+\frac{\epsilon p}{2},\frac{t}{\epsilon}) ] \end{aligned}$$ for $i=1,2$ where $i^{*} = 3-i$. We use the notation $i^*$ throughout the rest of the paper. We also define $\widehat{\Omega_{i}^{\epsilon}}_{-}(t)(p,k)$, $\widehat{\Gamma_{i}^{\epsilon}}_{-}(t)(p,k)$ as $$\begin{aligned} \widehat{\Omega_{i}^{\epsilon}}_{-}(t)(p,k) &= \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,-k), \notag \\ \widehat{\Gamma_{i}^{\epsilon}}_{-}(t)(p,k) &= \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)^{*}(-p,k).\end{aligned}$$ Note that for all $p \in \R$ these functions satisfy $$\begin{aligned} ||\widehat{\Omega_{i}^{\epsilon}}_{\iota}(t)(p,\cdot)||_{\mathbb{L}^{1}(\T)} \le \frac{1}{2} K_{0}, \quad ||\widehat{\Gamma_{i}^{\epsilon}}_{\iota}(t)(p,\cdot)||_{\mathbb{L}^{1}(\T)} \le \frac{1}{2} K_{0} \end{aligned}$$ for $i=1,2 $, $\iota = +,-$ under the condition $(\ref{cond.initial})$. With this notation, Wigner distribution is rewritten as $$\begin{aligned} \label{rewriteWigner} <\Omega^{\epsilon}(t),\boldsymbol{J}> &= \sum_{i=1,2} \int_{\R} dp \int_{\T} dk ~ \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k) \widehat{J_{i}}(p,k)^{*}.\end{aligned}$$ From now on we will show that the time evolution of $\Omega^{\epsilon}(\cdot)$ satisfies the following equation $$\begin{aligned} \label{timeevo} &\partial_{t} <\Omega^{\epsilon}(t),\boldsymbol{J}> \notag \\ &= \frac{1}{2\pi} <\Omega^{\epsilon}(t),\omega'(k) \partial_{y} \boldsymbol{J}> + \gamma<\Omega^{\epsilon}(t),C\boldsymbol{J}> \notag \\ & ~ + \gamma ( <\Gamma^{\epsilon}(t),C'\boldsymbol{J}> + <(\Gamma^{\epsilon})^{*}(t),C'\boldsymbol{J}>) + O_{\boldsymbol{J}}(\epsilon)\end{aligned}$$ for $\boldsymbol{J} \in \mathbf{S}^{2}$ where $$\begin{aligned} \label{rewriteWigner2} <\Gamma^{\epsilon}(t),\boldsymbol{J}> &= \sum_{i=1,2} \int_{\R} dp \int_{\T} dk ~ \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k) \widehat{J_{i}}(p,k)^{*} ,\notag \\ <(\Gamma^{\epsilon})^{*}(t),\boldsymbol{J}> &= \sum_{i=1,2} \int_{\R} dp \int_{\T} dk ~ \widehat{\Gamma_{i}^{\epsilon}}_{-}(t)(p,k) \widehat{J_{i}}(p,k)^{*}\end{aligned}$$ and $$\begin{aligned} (C'\boldsymbol{J})_{i}(p,k) = \int_{\T} dk' ~ \theta_{1}(k) \theta_{2}(k) R(k,k') \theta_{i^{*}}^{2}(k') J_{i^*}(p,k') + \widehat{{\beta}}(k) \theta_{1}(k) \theta_{2}(k) J_{i}(p,k).\end{aligned}$$ Here, $O_{\boldsymbol{J}}(\epsilon)$ is a term which satisfies $$\begin{aligned} \frac{O_{\boldsymbol{J}}(\epsilon)}{\epsilon} \le C_{\boldsymbol{J}} \end{aligned}$$ for all $0 < \epsilon < 1$ with a positive constant $C_{\boldsymbol{J}}$ which depends on $\boldsymbol{J}$. By $(\ref{defofpsi})$ the time evolution of $\widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k)$ is $$\begin{aligned} \label{evoofWigner} &\partial_{t} \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k) \notag \\ &= - \frac{\sqrt{-1}}{\epsilon}({\omega}_i(k+\frac{\epsilon p}{2}) - {\omega}_i(k-\frac{\epsilon p}{2})) \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k) \notag \\ & ~ + \gamma (\widehat{{\beta}}(k+\frac{\epsilon p}{2}) \theta_{i}^{2}(k+\frac{\epsilon p}{2}) + \widehat{{\beta}}(k-\frac{\epsilon p}{2}) \theta_{i}^{2}(k-\frac{\epsilon p}{2})) \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k) \notag \\ & ~ + \gamma \widehat{{\beta}}(k+\frac{\epsilon p}{2}) \theta_{i}(k+\frac{\epsilon p}{2}) \theta_{i^{*}}(k+\frac{\epsilon p}{2}) \widehat{\Gamma_{i^{*}}^{\epsilon}}_{-}(t)(p,k) \notag \\ & ~ + \gamma \widehat{{\beta}}(k-\frac{\epsilon p}{2}) \theta_{i}(k-\frac{\epsilon p}{2}) \theta_{i^{*}}(k-\frac{\epsilon p}{2})\widehat{\Gamma_{i^{*}}^{\epsilon}}_{+}(t)(p,k) \notag \\ & + \gamma \theta_{i}(k-\frac{\epsilon p}{2}) \theta_{i}(k+\frac{\epsilon p}{2}) \int_{\T} dk' R_{\epsilon p}(k,k') \notag \\ & ~ ~ \times [\theta_{i}(k'-\frac{\epsilon p}{2}) \theta_{i}(k'+\frac{\epsilon p}{2}) \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k') + \theta_{i^{*}}(k'-\frac{\epsilon p}{2}) \theta_{i^{*}}(k'+\frac{\epsilon p}{2}) \widehat{\Omega_{i^{*}}^{\epsilon}}_{+}(t)(p,k')) \notag \\ & ~ ~ ~ + \theta_{i}(k'-\frac{\epsilon p}{2}) \theta_{i^{*}}(k'+\frac{\epsilon p}{2}) \widehat{\Gamma_{i^{*}}^{\epsilon}}_{-}(t)(p,k') + \theta_{i}(k'+\frac{\epsilon p}{2}) \theta_{i^{*}}(k'-\frac{\epsilon p}{2}) \widehat{\Gamma_{i^{*}}^{\epsilon}}_{+}(t)(p,k')], \end{aligned}$$ where $$\begin{aligned} R_{\epsilon p}(k,k') = 16 \sin{(k+\frac{\epsilon p}{2})}\sin{(k-\frac{\epsilon p}{2})}\sin{(k'+\frac{\epsilon p}{2})}\sin{(k'-\frac{\epsilon p}{2})}.\end{aligned}$$ For the derivation of $(\ref{evoofWigner})$, see Appendix \[evolution\]. Since $\widehat{{\beta}}, \theta_{i}$ and ${\omega}_i$ are smooth on $\T$, the term $(\ref{evoofWigner})$ is rewritten as $$\begin{aligned} \label{evoofWigner2} &\partial_{t} \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k) \notag \\ &= - \sqrt{-1} p \omega'_i(k) \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k) + 2 \gamma \widehat{{\beta}}(k) \theta_{i}^{2}(k) \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k) \notag \\ & ~ + \gamma \widehat{{\beta}}(k) \theta_{i}(k) \theta_{i^{*}}(k) (\widehat{\Gamma_{i^{*}}^{\epsilon}}_{-}(t)(p,k) + \widehat{\Gamma_{i^{*}}^{\epsilon}}_{+}(t)(p,k)) \notag \\ & + \gamma \theta_{i}^{2}(k) \int_{\T} dk' R(k,k') [\theta_{i}^{2}(k') \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k') + \theta_{i^{*}}^{2}(k') \widehat{\Omega_{i^{*}}^{\epsilon}}_{+}(t)(p,k') \notag \\ & ~ ~ + \theta_{i}(k') \theta_{i^{*}}(k') \widehat{\Gamma_{i^{*}}^{\epsilon}}_{-}(t)(p,k') + \theta_{i}(k') \theta_{i^{*}}(k') \widehat{\Gamma_{i^{*}}^{\epsilon}}_{+}(t)(p,k')] + \mathcal{R}_{i}(p,k), \end{aligned}$$ where $\mathcal{R}_{i}, i=1,2$ are the remainder terms and these satisfy $$\begin{aligned} ||\mathcal{R}_{i}(p,\cdot)||_{\mathbb{L}^{1}(\T)} \le C(T,B,\gamma,{\alpha},K_{0}) |p| \epsilon \end{aligned}$$ for all $p \in \R$. Then for any $\boldsymbol{J} \in \mathbf{S}^{2}$, $$\begin{aligned} \label{remainder} \int_{\R} dp \int_{\T} dk ~ \mathcal{R}_{i}(p,k) \widehat{J_{i}}(p,k)^{*} = O_{\boldsymbol{J}}(\epsilon).\end{aligned}$$ Combining , , and with the relation $\int_{\T} dk' R(k,k') = -2 \widehat{{\beta}}(k)$, we conclude that holds. From and , for any fixed $T>0$ and $\boldsymbol{J} \in \mathbf{S}^{2}$, $\{ < \Omega^{\epsilon}(\cdot),\boldsymbol{J}> \}_{ 0<\epsilon <1} \subset C([0,T],\C) $ is uniformly bounded and equicontinuous. Hence, for each $\boldsymbol{J} \in \mathbf{S}^{2}$, there exists a subsequence $\epsilon_N \downarrow 0$ such that $ < \Omega^{\epsilon_N}(\cdot),\boldsymbol{J}>$ converges to a function in $C([0,T], \C)$ uniformly as $N \to \infty$. Since $ \mathbf{S}^{2}$ is separable, there is a dense countable subset $\{ \boldsymbol{J}^{(m)} ; m \in \N \}$ of $\mathbf{S}^2$ and by the diagonal argument we can find a sequence $\epsilon_N \downarrow 0$ such that $ < \Omega^{\epsilon_N}(\cdot),\boldsymbol{J}^{(m)}> $ converges for all $m \in \N$. Now, we show that for all $J \in \mathbf{S}^2$, $ < \Omega^{\epsilon_N}(\cdot),\boldsymbol{J}> $ converges uniformly to a continuous function as $N \to \infty$. Fix $\boldsymbol{J} \in \mathbf{S}^2$ and $\delta > 0$. Since $\{ \boldsymbol{J}^{(m)} ; m \in \N \}$ is dense we can take some $\boldsymbol{J}^{(l)}$ so that $||\boldsymbol{J} - \boldsymbol{J}^{(l)}|| < \delta$. Then for any $n,m \in \N$ we have $$\begin{aligned} &\sup_{t \in [0,T]} | <\Omega^{\epsilon_n}(t),\boldsymbol{J}> - <\Omega^{\epsilon_m}(t),\boldsymbol{J}> | \\ &\le \sup_{t \in [0,T]} | <\Omega^{\epsilon_n}(t),\boldsymbol{J}> - <\Omega^{\epsilon_n}(t),\boldsymbol{J}^{(l)}> | \\ & ~ + \sup_{t \in [0,T]} | <\Omega^{\epsilon_n}(t),\boldsymbol{J}^{(l)}> - <\Omega^{\epsilon_m}(t),\boldsymbol{J}^{(l)}> | \\ & ~ + \sup_{t \in [0,T]} | <\Omega^{\epsilon_m)}(t),\boldsymbol{J}^{(l)}> - <\Omega^{\epsilon_m}(t),\boldsymbol{J}> | \\ &\le K_{0} \delta + \sup_{t \in [0,T]} | <\Omega^{\epsilon_n}(t),\boldsymbol{J}^{(l)}> - <\Omega^{\epsilon_m}(t),\boldsymbol{J}^{(l)}> |.\end{aligned}$$ by . Hence, for sufficiently large $n,m$, $$\sup_{t \in [0,T]} | <\Omega^{\epsilon_n}(t),\boldsymbol{J}> - <\Omega^{\epsilon_m}(t),\boldsymbol{J}> |\le (1+K_0) \delta$$ and so $ < \Omega^{\epsilon_N}(\cdot),\boldsymbol{J}> $ converges uniformly. In Appendix \[positivity\], we prove that for any $t \ge 0$, any limit of a weak-\* convergent subsequence of $\{ \Omega^{\epsilon}(t) \}_{\epsilon}$ can be extended to a vector-valued finite positive measures on $\R \times \T$. The uniqueness of solutions of the equation (\[mboltzmann\]) is shown in Appendix \[uniqueness\]. Hence, noting that $\omega'(k) \partial_y\boldsymbol{J}, C\boldsymbol{J}, C'\boldsymbol{J} \in \mathbf{S}^{2}$ for any $\boldsymbol{J} \in \mathbf{S}^{2}$, by and the following lemma we conclude the proof of Theorem \[thm:main\]. For any $T>0$ and $\boldsymbol{J} \in \mathbf{S}^{2}$, $$\begin{aligned} \lim_{\epsilon \to 0} |\int_{0}^{T} dt ~ <\Gamma^{\epsilon}(t),\boldsymbol{J}> | = 0 , \\ \lim_{\epsilon \to 0} |\int_{0}^{T} dt ~ <(\Gamma^{\epsilon})^{*}(t),\boldsymbol{J}> | = 0.\end{aligned}$$ By $(\ref{defofpsi})$ the time evolution of $\widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k), i =1,2$ is $$\begin{aligned} &\partial_{t} \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k) \\ &= - \frac{\sqrt{-1}}{\epsilon} ({\omega}_i(k-\frac{\epsilon p}{2}) + \omega_{i^{*}}(k+\frac{\epsilon p}{2})) \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k) \\ & ~ + \gamma (\widehat{{\beta}}(k-\frac{\epsilon p}{2}) \theta_{i}^{2}(k-\frac{\epsilon p}{2}) + \widehat{{\beta}}(k+\frac{\epsilon p}{2}) \theta_{i^{*}}^{2}(k+\frac{\epsilon p}{2}))\widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k) \\ & ~ + \gamma \widehat{{\beta}}(k+\frac{\epsilon p}{2}) \theta_{i}(k+\frac{\epsilon p}{2}) \theta_{i^{*}}(k+\frac{\epsilon p}{2})\widehat{\Omega_{i}^{\epsilon}}_{-}(t)(p,k) \\ & ~ + \gamma \widehat{{\beta}}(k-\frac{\epsilon p}{2}) \theta_{i}(k-\frac{\epsilon p}{2}) \theta_{i^{*}}(k-\frac{\epsilon p}{2})\widehat{\Omega_{i^{*}}^{\epsilon}}_{+}(t)(p,k) \\ & ~ + \gamma \theta_{i}(k-\frac{\epsilon p}{2}) \theta_{i^{*}}(k+\frac{\epsilon p}{2}) \int_{\T} dk' R_{\epsilon p}(k,k') \notag \\ & ~ ~ \times [\theta_{i}(k'-\frac{\epsilon p}{2}) \theta_{i}(k'+\frac{\epsilon p}{2}) \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k') + \theta_{i^{*}}(k'-\frac{\epsilon p}{2}) \theta_{i^{*}}(k'+\frac{\epsilon p}{2}) \widehat{\Omega_{i^{*}}^{\epsilon}}_{+}(t)(p,k')) \notag \\ & ~ ~ ~ + \theta_{i}(k'-\frac{\epsilon p}{2}) \theta_{i^{*}}(k'+\frac{\epsilon p}{2}) \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k') + \theta_{i}(k'+\frac{\epsilon p}{2}) \theta_{i^{*}}(k'-\frac{\epsilon p}{2}) \widehat{\Gamma_{i}^{\epsilon}}_{-}(t)(p,k')] .\end{aligned}$$ Since $\widehat{{\beta}}$, $\theta_{i}$ and ${\omega}_i$ are smooth on $\T$, the above term is rewritten as $$\begin{aligned} \label{evoofWigner3} &\partial_{t} \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k) \notag \\ &= - \frac{\sqrt{-1}}{\epsilon} ({\omega}_i(k) + \omega_{i^{*}}(k)) \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k) + \gamma (\widehat{{\beta}}(k) \theta_{i}^{2}(k) + \widehat{{\beta}}(k) \theta_{i^{*}}^{2}(k))\widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k) \notag \\ & ~ + \gamma \widehat{{\beta}}(k) \theta_{i}(k) \theta_{i^{*}}(k)\widehat{\Omega_{i}^{\epsilon}}_{-}(t)(p,k) + \gamma \widehat{{\beta}}(k) \theta_{i}(k) \theta_{i^{*}}(k)\widehat{\Omega_{i^{*}}^{\epsilon}}_{+}(t)(p,k) \notag \\ & ~ + \gamma \theta_{i}(k) \theta_{i^{*}}(k) \int_{\T} dk' R(k,k') [\theta_{i}^{2}(k') \widehat{\Omega_{i}^{\epsilon}}_{+}(t)(p,k') + \theta_{i^{*}}^{2}(k') \widehat{\Omega_{i^{*}}^{\epsilon}}_{+}(t)(p,k')) \notag \\ & ~ ~ ~ + \theta_{i}(k') \theta_{i^{*}}(k') \widehat{\Gamma_{i}^{\epsilon}}_{+}(t)(p,k') + \theta_{i}(k') \theta_{i^{*}}(k') \widehat{\Gamma_{i}^{\epsilon}}_{-}(t)(p,k')] + \mathcal{R}_{i + 2}(p,k) \end{aligned}$$ for $i =1,2$ where $\mathcal{R}_{i}, i=3,4$ are the remainder terms which satisfy $$\begin{aligned} \label{remainder2} ||\mathcal{R}_{i}(p,\cdot)||_{\mathbb{L}^{1}(\T)} \le C(T,B,\gamma,{\alpha},K_{0}) |p| (1+\epsilon)\end{aligned}$$ for all $p \in \R$. Hence, for any $\boldsymbol{J} \in \mathbf{S}^{2}$ and $i=1,2$, $$\begin{aligned} \int_{\R} dp \int_{\T} dk ~ \mathcal{R}_{i+2}(p,k) \widehat{J_{i}}(p,k)^{*} = O_{\boldsymbol{J} }(1).\end{aligned}$$ Combining $(\ref{rewriteWigner})$, $(\ref{evoofWigner3})$ and $(\ref{remainder2})$ we have $$\begin{aligned} &\partial_{t} <\Gamma^{\epsilon}(t),\boldsymbol{J}> \\ &= - \frac{\sqrt{-1}}{\epsilon} <\Gamma^{\epsilon},({\omega}_1+{\omega}_2)\boldsymbol{J} > + <\Omega^{\epsilon},R'\boldsymbol{J} > + <\Omega^{\epsilon},R'\boldsymbol{J}^{t}> + <\Gamma^{\epsilon},R''\boldsymbol{J}> \\ & ~ + <(\Gamma^{\epsilon})^{*},R''\boldsymbol{J}> + <\Omega^{\epsilon},{\beta}'\boldsymbol{J}^{t}> + <(\Omega^{\epsilon})^{*},{\beta}'\boldsymbol{J}> + <\Gamma^{\epsilon},{\beta}\boldsymbol{J}> + ~ O_{\boldsymbol{J}}(1) , \end{aligned}$$ where $\boldsymbol{J}^{t} = (J_{2},J_{1})$ and $$\begin{aligned} & <(\Omega^{\epsilon})^{*}(t),\boldsymbol{J}> = \sum_{i=1,2} \int_{\R} dp \int_{\T} dk ~ \widehat{\Omega_{i}^{\epsilon}}_{-}(t)(p,k) \widehat{J_{i}}(p,k)^{*}, \\ &{\beta}'(k) = \theta_{1}(k) \theta_{2}(k) {\beta}(k), \\ &(R'\boldsymbol{J})_{i}(p,k) = \int_{\T} dk' \theta_{i}(k)^2 R(k,k') \theta_{1}(k')\theta_{2}(k') J_{i}(p,k'), \\ &(R''\boldsymbol{J})_{i}(p,k) = \int_{\T} dk' \theta_{1}(k) \theta_{2}(k) R(k,k') \theta_{1}(k')\theta_{2}(k') J_{i}(p,k').\end{aligned}$$ Therefore, we have $$\begin{aligned} \lim_{\epsilon \to 0} |\int_{0}^{T} dt ~ <\Gamma^{\epsilon}(t),({\omega}_1+{\omega}_2)\boldsymbol{J} > | = 0 \end{aligned}$$ for all $\boldsymbol{J} \in \mathbf{S}^{2}$. Since ${\omega}_1(k) + {\omega}_2(k) $ is uniformly bounded by positive constants from above and below, $({\omega}_1 + {\omega}_2)^{-1}\boldsymbol{J} \in \mathbf{S}^2$ for all $\boldsymbol{J} \in \mathbf{S}^2$. Hence we conclude that $$\begin{aligned} \lim_{\epsilon \to 0} |\int_{0}^{T} dt ~ <\Gamma^{\epsilon}(t),\boldsymbol{J}> | = 0 \end{aligned}$$ for all $\boldsymbol{J} \in \mathbf{S}^{2}$. For $(\Gamma^{\epsilon})^{*}$ we can apply the same proof. Proof of Theorem \[thm:main2\]. {#sec:proof2} =============================== We use the Markov chain $(K(t),I(t))$ introduced in Section \[markov\]. First note that for any $u_0 \in C^{\infty}_0(\R \times \T \times \{1,2\})$, $$u_N(y,k,i,t) = \mathbb{E}_{(k,i)}[u_{0}(Z_{N}(t),K(t),I(t))].$$ where $$\begin{aligned} Z_{N}(t) = y + \frac{1}{2\pi N^{\frac{3}{5}}} \int_{0}^{t} ds ~ \omega'(K(s)).\end{aligned}$$ Then, by using the Fourier transform we can write $$\begin{aligned} u_N(y,k,i,Nt) &= \mathbb{E}_{(k,i)}[u_{0}(Z_{N}(Nt),K(Nt),I(Nt))] \\ &= \sum_{x \in \Z} \int_{\R} dp \sum_{j = 1,2} \widetilde{u_{0}} (p,x,j) \mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt)} e^{\sqrt{-1} x K(Nt)} 1_{\{ I(Nt) = j \} } ] ,\end{aligned}$$ where $\widetilde{u}(p,x,i)$ is the Fourier transform of $u(y,k,i)$. Denote by $di$ the counting measure on $\{1,2 \}$. Let $P^{t} , t \ge 0$ be the semigroup generated by $\mathcal{L}$. Since $\frac{1}{2}dkdi$ is a reversible probability measure of the process $\{ ( K(t) , I(t) )\}_{t \ge 0}$ and $0$ is a simple eigenvalue for the generator $\mathcal{L}$, we have $$\begin{aligned} \lim_{t \to \infty} ||P^{t}f||_{\mathbb{L}^{2}(\T \times \{ 1,2 \} )} = 0 \end{aligned}$$ for any $f \in \mathbb{L}^{2}(\T \times \{ 1,2 \} )$ satisfying $\int_{\T \times \{1,2 \}} dkdi f(k,i) = 0$ by the ergodicity and the reversibility (cf. Theorem 1.6.1, 1.6.3 and Exercise 4.7.2 of [@FOT]). Let $\{ m_{N}\}_{N \in \N } $ be an increasing sequence of positive numbers such that $$\begin{aligned} \lim_{N \to \infty} m_{N} = \infty , \\ \lim_{N \to \infty} m_{N} N^{-\frac{3}{5}} = 0.\end{aligned}$$ Then for any $t \ge 0 , p \in \R , x \in \Z$ and $j = 1,2$ $$\begin{aligned} & \left|\mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt)} e^{\sqrt{-1} x K(Nt)} 1_{\{ I(Nt) = j \} }] - \mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt - m_{N}t)} e^{\sqrt{-1} x K(Nt)} 1_{\{ I(Nt) = j \} }] \right|\\ &\le \mathbb{E}_{(k,i)}[|1 - e^{\sqrt{-1} p (Z_{N}(Nt) - Z_{N}(Nt - m_{N}t))}|] \\ &\le \mathbb{E}_{(k,i)}[|p (Z_{N}(Nt) - Z_{N}(Nt - m_{N}t))|] \end{aligned}$$ since $|1-e^{\sqrt{-1}a}| \le |a|$ for any $a \in \R$. The last expression converges to $0$ as $N \to \infty$ since $$\begin{aligned} \mathbb{E}_{(k,i)}[|p (Z_{N}(Nt) - Z_{N}(Nt - m_{N}t))|] &= \mathbb{E}_{(k,i)}[|p \frac{1}{2\pi N^{\frac{3}{5}}} \int_{Nt - m_{N}t}^{Nt} ds ~ \omega'(K(s))|] \\ &\le \|\omega'\|_{\infty} t |p| m_{N} N^{-\frac{3}{5}} \to 0 \end{aligned}$$ where $ \|\omega'\|_{\infty}=\sup_{k} | \omega'(k)|$. By the Markov property $$\begin{aligned} &\mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt - m_{N}t)} e^{\sqrt{-1} x K(Nt)} 1_{\{ I(Nt) = j \} }] \\ &= \mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt - m_{N}t)} \mathbb{E}_{(K(Nt - m_{N}t),I(Nt - m_{N}t))} [e^{\sqrt{-1} x K(m_{N}t)} 1_{\{ I(m_{N}t) = j \} }] ] .\end{aligned}$$ By the Schwarz’s inequality, $$\begin{aligned} \label{0inl^{2}} & \big|\mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt - m_{N}t)} \mathbb{E}_{(K(Nt - m_{N}t),I(Nt - m_{N}t))} [e^{\sqrt{-1} x K(m_{N}t)} 1_{\{ I(m_{N}t) = j \} }] ] \notag \\ & ~ - \mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt - m_{N}t)} \mathbb{E}_{(K(Nt - m_{N}t),I(Nt - m_{N}t))}[ \frac{1}{2} \int_{\T} dk' e^{\sqrt{-1} x k'}]] \big| \notag \\ &\le \mathbb{E}_{(k,i)}[|\mathbb{E}_{(K(Nt - m_{N}t),I(Nt - m_{N}t))}[e^{\sqrt{-1} x K(m_{N}t)} 1_{\{ I(m_{N}t) = j \} } - \frac{1}{2} \int_{\T} dk' e^{\sqrt{-1} x k'}]|^{2}]^{\frac{1}{2}} .\end{aligned}$$ Let $g(k,i) = e^{\sqrt{-1} x k} 1_{\{ j \} }(i) - \frac{1}{2} \int_{\T} dk' e^{\sqrt{-1} x k'}$. Since $\frac{1}{2}dkdi$ is the reversible probability measure we have $$\begin{aligned} &\int_{\T \times \{1,2 \}} dkdi ~ \mathbb{E}_{(k,i)}[|\mathbb{E}_{(K(Nt - m_{N}t),I(Nt - m_{N}t))}[e^{\sqrt{-1} x K(m_{N}t)}1_{\{ I(m_{N}t) = j \} } - \frac{1}{2} \int_{\T} dk' e^{\sqrt{-1} x k'}]|^{2}] \\ &= ||P^{m_{N}t}g||_{\mathbb{L}^{2}(\T \times \{ 1,2 \} )}^{2} .\end{aligned}$$ Hence we conclude that converges to 0 in $\mathbb{L}^{2}(\T \times \{ 1,2 \} )$ as $N \to \infty$ since $\int_{\T \times \{1,2 \}} dkdi ~ g(k,i) = 0$. Summarizing the above and applying the dominated convergence theorem, we have $$\begin{aligned} &\lim_{N \to \infty} \int_{\T \times \{1,2 \}} dk di \sum_{x \in \Z} \int_{\R} dp \sum_{j = 1,2} | \widetilde{u_{0}} (p,x,j) | \\ &\times ~ |\mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt)} e^{\sqrt{-1} x K(Nt)} 1_{\{ I(Nt) = j \} } ] - \mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt - m_{N}t)} \frac{1}{2} \int_{\T} dk' e^{\sqrt{-1} x k'}] |^{2} \\ &= 0.\end{aligned}$$ Note that $$\begin{aligned} & \sum_{x \in \Z} \int_{\R} dp \sum_{j = 1,2} \widetilde{u_{0}} (p,x,j) \mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt - m_{N}t)} \frac{1}{2} \int_{\T} dk' e^{\sqrt{-1} x k'}] \\ &= \mathbb{E}_{(k,i)} [ \frac{1}{2} \int_{\T \times \{1,2 \}} dk' dj ~ u_{0}(Z_{N}(Nt - m_{N}t),k',j) ], \\ & = \frac{1}{2} \mathbb{E}_{(k,i)} [\bar{u}_0(Z_{N}(Nt - m_{N}t))].\end{aligned}$$ By Theorem \[thm:main3\], $Z_{N}(Nt - m_{N}t)$ converges to a Lévy process starting from $y$ and generated by $D (-\Delta_{y})^{\frac{5}{6}}$ and so the last term converges to $\bar{u}(y,t)$ given in for $k \neq 0, ~ i=1,2$. Therefore, $$\frac{1}{2}\mathbb{E}_{(k,i)} [\bar{u}_0(Z_{N}(Nt - m_{N}t))] \to \frac{1}{2}\bar{u}(y,t) \quad a.e.$$ and by the dominated convergence theorem, $$\begin{aligned} & \limsup_{N \to \infty} \int_{\T \times \{1,2 \}} dkdi | u_N(y,k,i,Nt) - \frac{1}{2}\bar{u}(y,t) |^{2} \\ & \le \limsup_{N \to \infty}\int_{\T \times \{1,2 \}} dkdi | u_N(y,k,i,Nt) - \frac{1}{2} \mathbb{E}_{(k,i)} [\bar{u}_0(Z_{N}(Nt - m_{N}t))] |^{2}.\end{aligned}$$ Applying the Fourier transform, the last term is bounded from above by $$\begin{aligned} & \limsup_{N\to \infty} \left(\sum_{x \in \Z} \int_{\R} dp \sum_{j = 1,2} | \widetilde{u_{0}} (p,x,j) |\right) \int_{\T \times \{1,2 \}} dkdi \sum_{x \in \Z} \int_{\R} dp \sum_{j = 1,2} | \widetilde{u_{0}} (p,x,j) | \\ & \quad \quad \times |\mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt)} e^{\sqrt{-1} x K(Nt)} 1_{\{ I(Nt) = j \} } ] - \mathbb{E}_{(k,i)}[e^{\sqrt{-1} p Z_{N}(Nt - m_{N}t)} \frac{1}{2} \int_{\T} dk' e^{\sqrt{-1} x k'}] |^{2}\end{aligned}$$ and so we complete the proof. Acknowledgement {#acknowledgement .unnumbered} =============== KS was supported by JSPS Grants-in-Aid for Scientific Research No. JP17K05587 and No. JP16H02211. MS was supported by JSPS Grants-in-Aid for Scientific Research No. JP16KT0021. Derivation of {#derivation} ============== We only consider the time evolution of $\widehat{v}_{1}(k,t)$. By the same calculation one can get the time evolution of $\widehat{v}_{2}(k,t)$. From $(\ref{formalz})$ we have $$\begin{aligned} d \widehat{v}_{1}(k,t) &= \sum_{x \in \Z} e^{-2 \pi \sqrt{-1} k x} d v_{1}(x,t) \\ &=( - \widehat{{\alpha}}(k) \widehat{q^1}(k,t) + B\widehat{v^2}(k,t) + \epsilon \gamma \widehat{{\beta}}(k) \widehat{v^1}(k,t) ) dt\\ & ~ + \sqrt{\epsilon \gamma} \sum_{x \in \Z} \sum_{z;|z-x|=1} e^{-2 \pi \sqrt{-1} k x} (Y_{x,z} v_{1}(x,t)) dw_{x,z}. \end{aligned}$$ Now we compute the last term. By summation by parts we have $$\begin{aligned} &-\sum_{x \in \Z} \sum_{z;|z-x|=1} e^{-2 \pi \sqrt{-1} k x} (Y_{x,z} v_{1}(x,t)) dw_{x,z} \\ &= \sum_{x \in \Z} \sum_{z \in \Z} h(z) v_{2}(x+z) e^{-2 \pi \sqrt{-1} k (x+z)} dw_{x,x+1} ,\end{aligned}$$ where $h: \Z \to \Z$ is defined as $$\begin{aligned} h(z)= \begin{cases} e^{2 \pi \sqrt{-1} k} -1 , \ z=1 \\ e^{-2 \pi \sqrt{-1} k} -1 , \ z= 0\\ 0 , \ z \neq 0,1. \end{cases}\end{aligned}$$ By the change of variables, the last term is rewritten as $$\begin{aligned} &\sum_{x \in \Z} \sum_{z \in \Z} h(z) v_{2}(x+z) e^{-2 \pi \sqrt{-1} k (x+z)} dw_{x,x+1} \\ &= \sum_{x \in \Z} \sum_{x' \in \Z} h(x'-x) v_{2}(x') e^{-2 \pi \sqrt{-1} k x'} dw_{x,x+1} \\ &= \sum_{x,x' \in \Z} (\int_{\T} dk' e^{2 \pi \sqrt{-1} k' (x'-x)} \sum_{y \in \Z} e^{-2 \pi \sqrt{-1} k' y} h(y)) v_{2}(x') e^{-2 \pi \sqrt{-1} k x'} dw_{x,x+1} ,\end{aligned}$$ and $$\begin{aligned} \sum_{y \in \Z} e^{-2 \pi \sqrt{-1} k' y} h(y) &= e^{-2 \pi \sqrt{-1} k} -1 + e^{- 2 \pi \sqrt{-1} k'}(e^{2 \pi \sqrt{-1} k} -1) \\ &= (e^{-2 \pi \sqrt{-1} k'} - e^{-2 \pi \sqrt{-1} k})(e^{2 \pi \sqrt{-1} k} - 1) \\ &= r(k,k').\end{aligned}$$ Therefore we have $(\ref{formalk})$. Existence and uniqueness of the solution of {#existence} ============================================ We follow the strategy of [@DZ] to show the existence by classical fixed point theorem. First we prepare some notations. We introduce a norm on $(\mathbb{L}^{2}(\T))^{2}$ defined as $$\begin{aligned} ||\mathbf{f}||_{(\mathbb{L}^{2}(\T))^{2}}^{2} = ||f_{1}||_{\mathbb{L}^{2}(\T)}^{2} + ||f_{2}||_{\mathbb{L}^{2}(\T)}^{2}\end{aligned}$$ for $\mathbf{f} = (f_{1},f_{2}) \in (\mathbb{L}^{2}(\T))^{2}$. Let $(E,\mathcal{F},\mathbb{P})$ be a probability space and $W$ be a cylindrical Wiener process on $(E,\mathcal{F},\mathbb{P})$. Fix $T>0$. Denote by $(\mathcal{H},||\cdot||_{\mathcal{H}})$ the Banach space of $(\mathbb{L}^{2}(\T))^{2}$-valued measurable processes $\mathbf{f}(k,t) , ~ k \in \T, t \in [0,T]$ such that $$\begin{aligned} ||\mathbf{f}||_{\mathcal{H}} = (\sup_{t \in [0,T]} \mathbb{E}[||\mathbf{f}(\cdot,t)||_{(\mathbb{L}^{2}(\T))^{2}}^{2}])^{\frac{1}{2}} < \infty ,\end{aligned}$$ where two processes are identified if they are $\mathbb{P} \times dt$ almost surely equal. Next we rewrite as $$\begin{aligned} d \begin{pmatrix} \widehat{\psi_{1}}(k,t) \\ \widehat{\psi_{2}}(k,t) \end{pmatrix} &= A_{1}(\widehat{\boldsymbol{\psi}}(\cdot,t))(k) dt + A_{2}(k')(\widehat{\boldsymbol{\psi}}(\cdot,t))(k) W(dk',dt) ,\end{aligned}$$ where $A_{1}$ and $A_{2}(k') , k' \in \T$ are bounded linear operators on $(\mathbb{L}^{2}(\T))^{2}$ defined as $$\begin{aligned} A_{1}(\mathbf{f})(k)&=\begin{pmatrix} \{ - \sqrt{-1} {\omega}_1(k) + \epsilon \gamma {\beta}(k) \theta_{1}(k)^2 \} f_{1}(k) + \epsilon \gamma {\beta}(k) \theta_{1}(k) \theta_{2}(k) f_{2}^{*}(k) \\ \epsilon \gamma {\beta}(k) \theta_{1}(k) \theta_{2}(k) f_{1}^{*}(k) + \{ - \sqrt{-1} {\omega}_2(k) + \epsilon \gamma {\beta}(k) \theta_{2}(k)^2 \} f_{2}(k) \end{pmatrix}, \\ A_{2}(k')(\mathbf{f})(k) &= \begin{pmatrix} r(k,k') (\theta_{1}(k-k') f_{1}(k-k') + \theta_{2}(k-k') f_{2}^*(k'-k) ) \\ r(k,k') (\theta_{1}(k-k') f_{1}^{*}(k'-k) + \theta_{2}(k-k') f_{2}(k-k') ) \end{pmatrix}\end{aligned}$$ for $\mathbf{f} = (f_{1},f_{2}) \in (\mathbb{L}^{2}(\T))^{2}$. Fix $\widehat{\boldsymbol{\psi}}_{0} \in (\mathbb{L}^{2}(\T))^{2}$. We define a functional $I:\mathcal{H} \to \mathcal{H}$ as $$\begin{aligned} I(\mathbf{f})_t= \widehat{\boldsymbol{\psi}}_{0} + \int_{[0,t]} A_{1}(\mathbf{f}(\cdot,s)) ds + \int_{[0,t]} A_{2}(k')(\mathbf{f}(\cdot,s)) W(dk',ds). \end{aligned}$$ For sufficiently small $T > 0$, $I$ is contractive and so there exists the unique fixed point $\widehat{\psi} \in \mathcal{H}$ such that $I(\widehat{\boldsymbol{\psi}}) = \widehat{\boldsymbol{\psi}}$. In this way we can construct a solution on the time interval $[0,T]$, and then we can construct a solution on the time interval $[T,2T]$ by the same argument and so on. Finally we check the uniqueness of the solution in the sense of the distribution. Suppose that $\mathbf{f}^{(1)}$ and $\mathbf{f}^{(2)} \in \mathcal{H}$ are two solutions of with a same initial condition. By the Cauchy-Schwarz inequality and the boundedness of $A_{1}$ and $A_{2}$, we have $$\begin{aligned} &\mathbb{E}[||\mathbf{f}^{(1)}(\cdot,t) - \mathbf{f}^{(2)}(\cdot,t)||_{(\mathbb{L}^{2}(\T))^{2}}^{2}] \\ &\le C(T) \int_{[0,t]} ds ~ \mathbb{E}[||\mathbf{f}^{(1)}(\cdot,s) - \mathbf{f}^{(2)}(\cdot,s)||_{(\mathbb{L}^{2}(\T))^{2}}^{2}]\end{aligned}$$ for all $t \ge 0$. By the Gronwall’s inequality we have $\mathbb{E}[||\mathbf{f}^{(1)}(\cdot,t) - \mathbf{f}^{(2)}(\cdot,t)||_{(\mathbb{L}^{2}(\T))^{2}}^{2}] = 0$. Conservation of the total energy {#conservation} ================================ By , the time evolution of $\int_{\T} dk ~ \mathbb{E}_{\epsilon} [ |\widehat{\psi_{i}}(k,t)|^{2} ] $ is given by $$\begin{aligned} &\frac{d}{dt} \int_{\T} dk ~ \mathbb{E}_{\epsilon} [ |\widehat{\psi_{i}}(k,t)|^{2} ] \\ &= \int_{\T} dk ~ 2 \gamma \widehat{{\beta}}(k) \theta_{i}^{2}(k) \mathbb{E}_{\epsilon} [ |\widehat{\psi_{i}}(k,t)|^{2} ] + 2 \gamma \widehat{{\beta}}(k) \theta_{i}(k) \theta_{i^{*}}(k) \Re(\mathbb{E}_{\epsilon} [\widehat{\psi_{i}}(k,t)\widehat{\psi_{i^{*}}}(k,t) ] ) \\ & ~ + \gamma \theta_{i}^{2}(k) \int_{\T} dk' R(k,k') \{ \theta_{i}^{2}(k') \mathbb{E}_{\epsilon} [ |\widehat{\psi_{i}}(k',t)|^{2} ] + \theta_{i^{*}}^{2}(k') \mathbb{E}_{\epsilon} [ |\widehat{\psi_{i^{*}}}(k',t)|^{2} ] \notag \\ & ~ ~ + 2 \theta_{i}(k') \theta_{i^{*}}(k') \Re(\mathbb{E}_{\epsilon} [ \widehat{\psi_{i}}(k,t)\widehat{\psi_{i^{*}}}(k,t)) ] ) \}\end{aligned}$$ where $\Re a$ is the real part of $a \in \C$. Since $ \sum_{i=1,2} \theta_{i}(k)^2 = 1$ and $\int_{\T} dk' ~ R(k,k') = -2 \widehat{{\beta}}(k)$, we have $$\begin{aligned} &\frac{d}{dt} \sum_{i=1,2} \int_{\T} dk ~ \mathbb{E}_{\epsilon} [ |\widehat{\psi_{i}}(k,t)|^{2} ] \\ &= \int_{\T} dk ~ 2 \gamma \widehat{{\beta}}(k) \sum_{i=1,2} \theta_{i}^{2}(k) \mathbb{E}_{\epsilon} [ |\widehat{\psi_{i}}(k,t)|^{2} ] + 4 \gamma \widehat{{\beta}}(k) \theta_{1}(k) \theta_{2}(k) \Re(\mathbb{E}_{\epsilon} [ \widehat{\psi_{1}}(k,t)\widehat{\psi_{2}}(k,t) ] ) \\ & ~ + \gamma \int_{\T} dk' R(k,k') \{ \theta_{1}^{2}(k') \mathbb{E}_{\epsilon} [ |\widehat{\psi_{1}}(k',t)|^{2} ] + \theta_{2}^{2}(k') \mathbb{E}_{\epsilon} [ |\widehat{\psi_{2}}(k',t)|^{2} ] \notag \\ & ~ ~ + 2 \theta_{1}(k') \theta_{2}(k') \Re(\mathbb{E}_{\epsilon} [ \widehat{\psi_{1}}(k,t)\widehat{\psi_{2}}(k,t) ] ) \} \\ & = 0.\end{aligned}$$ Uniqueness of solutions of the linear kinetic equation {#positivity} ====================================================== \[Riesz\] Let $\{ \Omega^{\epsilon_N}(t) \}_{N \in \N}$ be a convergent subsequence with its limit $\Omega(t)$. Then there exists a vector-valued finite positive measure $\boldsymbol{\mu}(t)=(\mu_1(t),\mu_2(t))$ such that $$\begin{aligned} \int_{\R \times \T} \mu_i(t)(dy,dk) J_i(y,k)^* = <\Omega(t),J_i>, \quad i=1,2\end{aligned}$$ for all $\boldsymbol{J} \in \mathbf{S}^2$. Let $\Omega_{1}(t) \in \mathbf{S}'$ as $<\Omega_{1}(t), J>:= <\Omega(t), \boldsymbol{J}>$ for $\boldsymbol{J}=(J,0)$, $J \in \mathbf{S}$. First we show that $\Omega_1(\cdot)$ is multiplicatively positive, that is, $$\begin{aligned} <\Omega_{1}(t),|J|^{2}> ~ \ge 0 \end{aligned}$$ for all $t \ge 0$ and $J \in \mathbf{S}$. Fix $t \ge 0$ and $J \in \mathbf{S}$. Since $J$ is smooth, $$\begin{aligned} J(\frac{\epsilon}{2}(x+x'),k) - J(\epsilon x,k) &= \frac{\epsilon}{2} \int_{0}^{1} dr ~ (x' - x) \partial_{y} J(\epsilon x + r \frac{\epsilon}{2}(x' - x),k) \end{aligned}$$ for all $x,x' \in \Z$ and so $$\begin{aligned} &\left| \int_{\T} dk e^{2\pi \sqrt{-1} (x'-x) k} \left(J(\frac{\epsilon}{2} (x+x'),k)J(\frac{\epsilon}{2} (x+x'),k)^* - J(\epsilon x,k) J(\frac{\epsilon}{2} (x+x'),k)^{*} \right) \right| \\ &= \left| \frac{\epsilon}{2}(x'-x) \int_{\T} dk e^{2\pi \sqrt{-1} (x'-x) k} J(\frac{\epsilon}{2} (x+x'),k)^{*} \int_{0}^{1} dr \partial_{y} J(\epsilon x + r \frac{\epsilon}{2} (x' - x) , k) \right|.\end{aligned}$$ By repeating the integration by parts we have $$\begin{aligned} & \left| \int_{\T} dk e^{2\pi \sqrt{-1} (x'-x) k} J(\frac{\epsilon}{2} (x+x'),k)^{*} \int_{0}^{1} dr \partial_{y} J(\epsilon x + r \frac{\epsilon}{2} (x' - x) , k) \right| \\ & = \left | \int_{\T} dk \left(\frac{1}{2 \pi \sqrt{-1} (x' - x)}\right)^{3} e^{2\pi \sqrt{-1} (x'-x) k} \partial_{k}^{3}[ J(\frac{\epsilon}{2} (x+x'),k)^{*} \int_{0}^{1} dr \partial_{y} J(\epsilon x + r \frac{\epsilon}{2} (x' - x) , k) ] \right| \\ &\le \frac{1}{8\pi^3|x - x'|^{3}} \int_{\T} dk \ | \partial_{k}^{3}[ J(\frac{\epsilon}{2} (x+x'),k)^{*} \int_{0}^{1} dr \partial_{y} J(\epsilon x + r \frac{\epsilon}{2} (x' - x) , k) ] |. \end{aligned}$$ Hence, we have $$\begin{aligned} & \left| \int_{\T} dk e^{2\pi \sqrt{-1} (x'-x) k} \left(J(\frac{\epsilon}{2} (x+x'),k)J(\frac{\epsilon}{2} (x+x'),k)^* - J(\epsilon x,k) J(\frac{\epsilon}{2} (x+x'),k)^{*} \right) \right| \\ & \le \frac{\epsilon}{16\pi^3|x - x'|^{2}} \int_{\T} dk \ | \partial_{k}^{3}[ J(\frac{\epsilon}{2} (x+x'),k)^{*} \int_{0}^{1} dr \partial_{y} J(\epsilon x + r \frac{\epsilon}{2} (x' - x) , k) ] | \\ & \le \frac{1}{|x - x'|^{2}}O_{J}(\epsilon)\end{aligned}$$ for all $x \neq x' \in \Z$. In the same way, we can show that $$\begin{aligned} \left| \int_{\T} dk e^{2\pi \sqrt{-1} (x'-x) k} \left(J(\epsilon x,k) J(\frac{\epsilon}{2} (x+x'),k)^{*} - J(\epsilon x,k) J(\epsilon x',k)^{*} \right) \right| \le \frac{1}{|x - x'|^{2}}O_{J}(\epsilon).\end{aligned}$$ On the other hand we have $$\begin{aligned} &\frac{\epsilon}{2} \sum_{x,x' \in \Z} <\psi_{1}(x',\frac{t}{\epsilon})^{*} \psi_{1}(x,\frac{t}{\epsilon})>_{\epsilon} \int_{\T} dk ~ e^{2\pi \sqrt{-1} (x'-x) k} J(\epsilon x,k) J(\epsilon x',k)^{*} \\ &= \frac{\epsilon}{2} \int_{\T} dk ~ <|\sum_{x \in \Z} e^{- 2\pi \sqrt{-1} x k} \psi(x,\frac{t}{\epsilon}) J(\epsilon x,k)|^{2} >_{\epsilon} ~ \ge 0.\end{aligned}$$ Since $$\begin{aligned} & \left | \int_{\T} dk e^{2\pi \sqrt{-1} (x'-x) k} |J(\frac{\epsilon}{2} (x+x'),k)|^{2} - J(\epsilon x,k) J(\epsilon x',k)^{*} \right| \\ &\le \left| \int_{\T} dk e^{2\pi \sqrt{-1} (x'-x) k} \left(J(\frac{\epsilon}{2} (x+x'),k)J(\frac{\epsilon}{2} (x+x'),k)^* - J(\epsilon x,k) J(\frac{\epsilon}{2} (x+x'),k)^{*} \right) \right| \\ & ~ + \left| \int_{\T} dk e^{2\pi \sqrt{-1} (x'-x) k} \left(J(\epsilon x,k) J(\frac{\epsilon}{2} (x+x'),k)^{*} - J(\epsilon x,k) J(\epsilon x',k)^{*} \right) \right|,\end{aligned}$$ combining the above calculations we have $$\begin{aligned} <\Omega_{1}^{\epsilon}(t),|J|^{2}> = \frac{\epsilon}{2} \int_{\T} dk ~ <|\sum_{x \in \Z} e^{- 2\pi \sqrt{-1} x k} \psi_{1}(x,\frac{t}{\epsilon}) J(\epsilon x,k)|^{2} >_{\epsilon} + O_{J}(\epsilon).\end{aligned}$$ Therefore $\Omega_{1}(\cdot)$ is multiplicatively positive. Next we show that $\Omega_{1}(\cdot)$ is positive, that is, $$\begin{aligned} <\Omega_{1}(t),J> ~ \ge 0\end{aligned}$$ for all $t \ge 0$ and $J \in \mathbf{S} , J \ge 0$. Since $\{ J \in \mathbf{S} ; J \in C_{0}^{\infty}(\R \times \T) , J \ge 0 \}$ is a dense subset of $\{ J \in \mathbf{S} ; J \ge 0 \} $, it is sufficient to show the positivity on $C_{0}^{\infty}(\R \times \T)$. Fix $t \ge 0$ and a positive function $J \in C_{0}^{\infty}(\R \times \T)$. There exists a positive constant $M > 0$ such that the support of $J$ is a subset of $[-M,M] \times \T$. Let $a(y) \in C_{0}^{\infty}(\R)$ be a function such that $a(y) = 1 , ~ y \in [-M,M]$. Define $J^{(m)}(y,k) \in C_{0}^{\infty}(\R \times \T) , ~ m \in \N$ as $$\begin{aligned} J^{(m)}(y,k) = a(y) \sqrt{J(y,k) + \frac{1}{m} }.\end{aligned}$$ Then the sequence $\{ |J^{(m)}|^{2} , ~ m \in \N \}$ converges to $J(y,k)$ in the topology of $C_{0}^{\infty}(\R \times \T)$. Since the embedding of the space $C_{0}^{\infty}(\R \times \T)$ into the space $\mathbf{S}$ is continuous, $\{ |J^{(m)}|^{2} , ~ m \in \N \}$ also converges to $J(y,k)$ in the topology of $\mathbf{S}$. By the continuity of $\Omega_{1}(t)$, we have $$\begin{aligned} <\Omega_{1}(t),J> = \lim_{m \to \infty} <\Omega_{1}(t),|J^{(m)}|^{2}> \ \ge 0.\end{aligned}$$ Therefore $\Omega_{1}(\cdot)$ is positive. In the same way we can show that $\Omega_{2}(\cdot)$ is also positive. By the usual method, for example see Lemma 1 in Chapter 2 of [@GV], we can extend the domain of $\Omega_{i}(\cdot) , ~ i=1,2$ to the space $C_{0}(\R \times \T)$. By the Riesz representation theorem there exists a finite positive measure $\mu_{i}(\cdot) , ~ i=1,2$ such that $$\begin{aligned} <\Omega_{i}(t),J> ~ = \int_{\R \times \T} \mu_{i}(t)(dy,dk) ~ J(y,k) , ~ i=1,2\end{aligned}$$ for all $t \ge 0$ and $J \in C_{0}(\R \times \T)$. By the linearity and the definition of $\Omega_{i}^{\epsilon} ( \cdot)$, $$\begin{aligned} <\Omega_{i}(t),J> ~ = \int_{\R \times \T} \mu_{i}(t)(dy,dk) ~ J(y,k)^* , ~ i=1,2\end{aligned}$$ for all $J \in \mathbf{S}$. Uniqueness of the solution of the Boltzmann equation {#uniqueness} ==================================================== Suppose that a vector-valued finite positive measure $\boldsymbol{\mu}(t)$ is a solution of the Boltzmann equation . Then $\tilde{\boldsymbol{\mu}}(t)(dy,dk) := \boldsymbol{\mu}(t)(dy + \frac{1}{2\pi} \omega'(k)t,dk)$ is a solution of the following space-homogeneous Boltzmann equation $$\begin{aligned} \partial_{t} \int d\tilde{\boldsymbol{\mu}}(t) \cdot \boldsymbol{J} = \int d\tilde{\boldsymbol{\mu}}(t) \cdot (C\boldsymbol{J})\end{aligned}$$ where $$\begin{aligned} \int d\tilde{\boldsymbol{\mu}}(t) \cdot \boldsymbol{J} &= \int \boldsymbol{\mu}(t)(dy + \frac{1}{2\pi} \omega'(k)t,dk) \cdot \boldsymbol{J} \\ &:= \int \boldsymbol{\mu}(t)(dy,dk) \cdot \boldsymbol{J}(y - \frac{1}{2\pi} \omega'(k)t,k). \end{aligned}$$ Conversely, if $\tilde{\boldsymbol{\mu}}(t)$ is a solution of the space-homogeneous Boltzmann equation, then $\boldsymbol{\mu}(t)(dy,dk) := \tilde{\boldsymbol{\mu}}(t)(dy - \frac{1}{2\pi} \omega'(k)t,dk)$ is a solution of the Boltzmann equation . Therefore, it is sufficient to show the uniqueness of the solution for the space-homogeneous Boltzmann equation. Suppose that $J^{1}(y,k) = f^{\lambda,y^{*},r}(y)G(k) , J^{2}(y,k) = 0$, where $$\begin{aligned} f^{\lambda,y^{*},r}(y) &=\exp \left( - \frac{\lambda}{r^{2} - |y -y^{*}|^{2}} \right) 1_{B(y^{*},r)}(y), \\ B(y^{*},r) &= \{ y \in \R \ ; \ |y - y^{*}| < r \}, \end{aligned}$$ $y^{*} \in \R$ , $r > 0$ and $G(\cdot) \in C^{\infty}(\T)$. Note that $f^{\lambda,y^{*},r} \in C^{\infty}_{0}(\R)$, $\|f^{\lambda,y^{*},r} \|_{\infty} \le 1$ and $$\begin{aligned} \lim_{\lambda \to 0} f^{\lambda,y^{*},r}(y) = 1_{B(y^{*},r)}(y).\end{aligned}$$ Let $\boldsymbol{\mu}(t), \boldsymbol{\nu}(t)$ be solutions of the space-homogeneous Boltzmann equation with a same initial condition. Then $$\begin{aligned} &|\int d\boldsymbol{\mu}(t) \cdot \boldsymbol{J} - \int d\boldsymbol{\nu}(t) \cdot \boldsymbol{J} | \\ & = |\int d\boldsymbol{\mu}(t) \cdot f^{\lambda,y^{*},r}(y) \boldsymbol{G} - \int d\boldsymbol{\nu}(t) \cdot f^{\lambda,y^{*},r}(y) \boldsymbol{G} |\\ &\le \int_{0}^{t} ds \left| \int d(\boldsymbol{\mu}(s) - \boldsymbol{\nu}(s)) \cdot (f^{\lambda,y^{*},r}(y) C\mathbf{G}(k)) \right|\end{aligned}$$ where $\mathbf{G}=(G(k),0)$. By taking the limit $\lambda \to 0$, we have $$\begin{aligned} &|\int_{\T} G(k) d(\mu_{1}(t)(B(y^{*},r),dk) - \nu_{1}(t)(B(y^{*},r),dk))| \\ &\le \int_{0}^{t} ds |\int_{\T} d(\mu(s)(B(y^{*},r),dk) - \nu(s)(B(y^{*},r),dk)) \cdot (C\mathbf{G})| \\ &\le \int_{0}^{t} ds \sum_{i=1,2} |\int_{\T} d(\mu_{i}(s)(B(y^{*},r),dk) - \nu_{i}(s)(B(y^{*},r),dk)) (C\mathbf{G})_{i}| \\ &\le 32 \sup_{k}|G(k)| \int_{0}^{t} ds \sum_{i=1,2} ||\mu_{i}(s)(B(y^{*},r),dk) - \nu_{i}(s)(B(y^{*},r),dk)|| \\\end{aligned}$$ where $||\cdot||$ is the total variation for a bounded signed measure on $\T$. Hence, $$\begin{aligned} & ||\mu_{1}(t)(B(y^{*},r),dk) - \nu_{1}(t)(B(y^{*},r),dk)|| \\ & \quad \quad \le 32 \int_{0}^{t} ds \sum_{i=1,2} ||\mu_{i}(s)(B(y^{*},r),dk) - \nu_{i}(s)(B(y^{*},r),dk)||.\end{aligned}$$ By the same proof, we have $$\begin{aligned} &||\mu_{2}(t)(B(y^{*},r),dk) - \nu_{2}(t)(B(y^{*},r),dk)|| \\ & \quad \le 32 \int_{0}^{t} ds \sum_{i=1,2} ||\mu_{i}(s)(B(y^{*},r),dk) - \nu_{i}(s)(B(y^{*},r),dk)||. \\ &\therefore \sum_{i=1,2} ||\mu_{i}(t)(B(y^{*},r),dk) - \nu_{i}(t)(B(y^{*},r),dk)|| \\ & \quad \quad \le 64 \int_{0}^{t} ds \sum_{i=1,2} ||\mu_{i}(s)(B(y^{*},r),dk) - \nu_{i}(s)(B(y^{*},r),dk)|| . \end{aligned}$$ Therefore $\mu_{i}(t)(B(y^{*},r),dk) = \nu_{i}(t)(B(y^{*},r),dk)$ on $\T$ for any ball $B(y^{*},r) \subset \R$, which concludes $\boldsymbol{\mu}(t)=\boldsymbol{\nu}(t)$ for any $t \ge 0$. Derivation of $(\ref{evoofWigner})$ {#evolution} =================================== We only consider the time evolution of $\widehat{\Omega_{1}^{\epsilon}}_{+}(t)(p,k)$. By the same calculation we can obtain the time evolution of $\widehat{\Omega_{2}^{\epsilon}}_{+}(t)(p,k)$. From $(\ref{defofpsi})$ we have $$\begin{aligned} &\partial_{t} \widehat{\Omega_{1}^{\epsilon}}_{+}(t)(p,k) \notag \\ &= - \frac{\sqrt{-1}}{\epsilon}\left({\omega}_1(k+\frac{\epsilon p}{2}) - {\omega}_1(k-\frac{\epsilon p}{2})\right) \widehat{\Omega_{1}^{\epsilon}}_{+}(t)(p,k) \notag \\ & ~ + \gamma \left({\beta}(k+\frac{\epsilon p}{2}) \theta_{1}(k+\frac{\epsilon p}{2})^2 + {\beta}(k-\frac{\epsilon p}{2}) \theta_{1}(k-\frac{\epsilon p}{2})^2 \right) \widehat{\Omega_{1}^{\epsilon}}_{+}(t)(p,k) \notag \\ & ~ + \gamma {\beta}(k+\frac{\epsilon p}{2}) \theta_{1}(k+\frac{\epsilon p}{2}) \theta_{2}(k+\frac{\epsilon p}{2}) \widehat{\Gamma_{2}^{\epsilon}}_{-}(t)(p,k) \notag \\ & ~ + \gamma {\beta}(k-\frac{\epsilon p}{2}) \theta_{1}(k-\frac{\epsilon p}{2}) \theta_{2}(k-\frac{\epsilon p}{2})\widehat{\Gamma_{2}^{\epsilon}}_{+}(t)(p,k) \notag \\ & + \gamma \theta_{1}(k-\frac{\epsilon p}{2}) \theta_{1}(k+\frac{\epsilon p}{2}) \int_{\T} dk' r(k-\frac{\epsilon p}{2},k')^{*} r(k+\frac{\epsilon p}{2},k') \notag \\ & ~ ~ \times [\theta_{1}(k-k'-\frac{\epsilon p}{2}) \theta_{1}(k-k'+\frac{\epsilon p}{2}) \widehat{\Omega_{1}^{\epsilon}}_{+}(t)(p,k-k') \\ & ~ ~ ~ + \theta_{2}(k-k'-\frac{\epsilon p}{2}) \theta_{2}(k-k'+\frac{\epsilon p}{2}) \widehat{\Omega_{2}^{\epsilon}}_{+}(t)(p,k-k')) \notag \\ & ~ ~ ~ + \theta_{1}(k-k'-\frac{\epsilon p}{2}) \theta_{2}(k-k'+\frac{\epsilon p}{2}) \widehat{\Gamma_{2}^{\epsilon}}_{-}(t)(p,k-k') \\ & ~ ~ ~ + \theta_{1}(k-k'+\frac{\epsilon p}{2}) \theta_{2}(k-k'-\frac{\epsilon p}{2}) \widehat{\Gamma_{2}^{\epsilon}}_{+}(t)(p,k-k')]. \notag \end{aligned}$$ By the change of variables $k-k' \to k'$, the last integral is rewritten as $$\begin{aligned} &\int_{\T} dk' r(k-\frac{\epsilon p}{2},k-k')^{*} r(k+\frac{\epsilon p}{2},k-k') \notag \\ & ~ ~ \times [\theta_{1}(k'-\frac{\epsilon p}{2}) \theta_{1}(k'+\frac{\epsilon p}{2}) \widehat{\Omega_{1}^{\epsilon}}_{+}(t)(p,k') + \theta_{2}(k'-\frac{\epsilon p}{2}) \theta_{2}(k'+\frac{\epsilon p}{2}) \widehat{\Omega_{2}^{\epsilon}}_{+}(t)(p,k')) \notag \\ & ~ ~ ~ + \theta_{1}(k'-\frac{\epsilon p}{2}) \theta_{2}(k'+\frac{\epsilon p}{2}) \widehat{\Gamma_{2}^{\epsilon}}_{-}(t)(p,k') + \theta_{1}(k'+\frac{\epsilon p}{2}) \theta_{2}(k'-\frac{\epsilon p}{2}) \widehat{\Gamma_{2}^{\epsilon}}_{+}(t)(p,k')]. \notag\end{aligned}$$ Hence, it is sufficient to show that $r(k-\frac{\epsilon p}{2},k-k')^{*} r(k+\frac{\epsilon p}{2},k-k') = R_{\epsilon p}(k,k')$. By the following direct calculations $$\begin{aligned} &r(k-\frac{\epsilon p}{2},k-k')^{*} r(k+\frac{\epsilon p}{2},k-k') \\ &= (e^{2 \pi \sqrt{-1} (k-k')} - e^{2 \pi \sqrt{-1} (k-\frac{\epsilon p}{2})})(e^{- 2 \pi \sqrt{-1} (k-\frac{\epsilon p}{2})} - 1) \\ & ~ \times (e^{- 2 \pi \sqrt{-1} (k-k')} - e^{- 2 \pi \sqrt{-1} (k+\frac{\epsilon p}{2})})(e^{2 \pi \sqrt{-1} (k+\frac{\epsilon p}{2})} - 1) \\ &= (1 - e^{- \pi \sqrt{-1} \epsilon p}(e^{2 \pi \sqrt{-1} k'} + e^{-2 \pi \sqrt{-1} k'}) + e^{- 2 \pi \sqrt{-1} \epsilon p}) \\ & ~ \times (1 - e^{ \pi \sqrt{-1} \epsilon p}(e^{2 \pi \sqrt{-1} k} + e^{-2 \pi \sqrt{-1} k}) + e^{ 2 \pi \sqrt{-1} \epsilon p}) \\ &= (e^{ \pi \sqrt{-1} \epsilon p} - (e^{2 \pi \sqrt{-1} k'} + e^{-2 \pi \sqrt{-1} k'}) + e^{- \pi \sqrt{-1} \epsilon p}) \\ & ~ \times (e^{- \pi \sqrt{-1} \epsilon p} - (e^{2 \pi \sqrt{-1} k} + e^{-2 \pi \sqrt{-1} k}) + e^{\pi \sqrt{-1} \epsilon p}) \end{aligned}$$ and $$\begin{aligned} &e^{- \pi \sqrt{-1} \epsilon p} - (e^{2 \pi \sqrt{-1} k} + e^{-2 \pi \sqrt{-1} k}) + e^{\pi \sqrt{-1} \epsilon p} \\ &= 2\cos{\pi \epsilon p} - 2\cos{2 \pi k} \\ &= 4 \sin{(k+\frac{\epsilon p}{2})}\sin{(k-\frac{\epsilon p}{2})} ,\end{aligned}$$ we can verify the equation $r(k-\frac{\epsilon p}{2},k-k')^{*} r(k+\frac{\epsilon p}{2},k-k') = R_{\epsilon p}(k,k')$. [15]{} : [*Thermal Conductivity for a Momentum Conservative Model*]{}. Commun.  Math.  Phys. **287**, 67-98 (2009) : [*Energy transport in stochastically perturbed lattice dynamics*]{}. Arch.  Ration.  Mech. **195**, 171–203 (2009) , [*$3/4$-fractional superdiffusion in a system of harmonic oscillators perturbed by a conservative noise*]{}, Arch. Rational Mech. Anal. **220** (2016), 505–542. : [*Stochastic equations in infinite dimensions*]{}. Cambridge University Press, Cambridge (1992) : [*Heat transport in low-dimensional systems*]{}. Adv.  Phys. **57**(5), 457–537 (2008) : [*Stationary states of random Hamiltonian systems*]{}, Probab. Theory relat. Fields **99**, 211–236 (1994) : [*Dirichlet Forms and Symmetric Markov Processes*]{}, Berlin, Boston: De Gruyter, 2nd ed. (2010) : [*Generalized Functions volume 4*]{}. Academic Press, New York (1964) : [*A limit theorem for an additive functionals of Markov chains*]{}. Ann.  Appl.  Probab. **19**, 2270–2230 (2009) : [*Superdiffusion of Energy in a Chain of Harmonic Oscillators with Noise*]{}. Commun.  Math.  Phys. **339**, 407–453 (2015) , [*Diffusive Propagation of Energy in a Non-acoustic Chain*]{}, Arch. Rational Mech. Anal., **223** (2017), 95–139 . edited by S. Lepri (Springer, New York), (2016) : [*Thermal conduction in classical low-dimensional lattices*]{}. Phys. Rep. **377**(1) 1–80 (2003) : [*Thermal conductivity for a stochastic dynamics in a magnetic field*]{}. Commun.  Math.  Phys. **361**, 951–995 (2018) : [*Nonlinear fluctuating hydrodynamics for anharmonic chains*]{}. J. Stat. Phys. **154**(5), 1191-1227 (2014) : [*Heat Transport via Low-Dimensional Systems with Broken Time-Reversal Symmetry*]{}. Phys. Rev. Lett. **119** (2017)

ArXiv

--- abstract: 'We have carried out scanning tunneling spectroscopy measurements on exfoliated monolayer graphene on SiO$_2$ to probe the correlation between its electronic and structural properties. Maps of the local density of states are characterized by electron and hole puddles that arise due to long range intravalley scattering from intrinsic ripples in graphene and random charged impurities. At low energy, we observe short range intervalley scattering which we attribute to lattice defects. Our results demonstrate that the electronic properties of graphene are influenced by intrinsic ripples, defects and the underlying SiO$_2$ substrate.' author: - 'A. Deshpande' - 'W. Bao' - 'F. Miao' - 'C.N. Lau' - 'B.J. LeRoy' title: 'Spatially resolved spectroscopy of monolayer graphene on SiO$_2$' --- Graphene is the two dimensional form of carbon characterized by a honeycomb lattice with two inequivalent lattice sites. It is unusual in many aspects, one of them being the linear energy dispersion relation that causes the electrons to obey the relativistic Dirac equation instead of the Schrödinger equation [@novoselov2007]. The theoretical prediction of the instability of 2D crystals along with unsuccessful efforts at their synthesis on insulating substrates kept graphene far from the realm of experiments. A breakthrough in the isolation of graphene from 3D graphite opened up a new frontier for experimental investigation [@novoselov2004; @novoselovPNAS]. Electrical transport measurements have brought to the fore some of the exotic properties of graphene like a novel quantum Hall effect, high carrier mobility and minimum conductivity [@novoselov2005; @zhang2005; @miao2007]. Transmission electron microscopy characterization (TEM) of graphene revealed intrinsic ripples [@meyer2007] which were confirmed by Monte Carlo simulations [@fasolino2007]. These structural findings hinted that electron scattering and carrier localization could be highly influenced not only by external effects such as impurities but also by intrinsic ripples. Furthermore, electrical transport measurements showed the suppression of electron localization [@morozov2006] and that the mobility was insensitive to doping [@schedin2007]. Both these results, unexpected for a two dimensional system like graphene, were attributed to corrugations in graphene. Hence, to investigate phenomena involving electronic and structural aspects of graphene we have performed spatially resolved spectroscopy measurements. A single atom thick graphene sheet can be studied by suspending it on a micron-sized metal grid or by supporting it on a substrate like SiO$_2$ [@meyer2007; @ishigami2007; @stolyarova2007]. The sheet is not perfectly flat but contoured with intrinsic ripples with deformations up to 1 nm normal to the plane clearly evident from TEM of suspended graphene and Monte Carlo simulations [@meyer2007; @fasolino2007]. These ripples are due to the stability requirement of the two dimensional lattice and the ability of carbon to bond with asymmetric bond lengths [@meyer2007; @fasolino2007]. Supported sheets of graphene have an additional constraint on their morphology due to the SiO$_2$ substrate. They have been found to partially conform to the substrate while still having intrinsic ripples and therefore areas not in contact with SiO$_2$ [@ishigami2007; @stolyarova2007; @geringer2009]. These previous measurements showed the topographic structure of monolayer graphene. Spectroscopy measurements of graphene on SiO$_2$ revealed the contribution of phonons to the tunneling process [@zhang2008]. However, none of these measurements probed the effects of graphene morphology on its electronic properties. Theoretical calculations show that variations in the local curvature of graphene, ripples, lead to a $\pi-\sigma$ orbital mixing which changes the local electrochemical potential breaking particle hole symmetry and causing charge inhomogeneities [@kim2008]. The shift in the local chemical potential $\Phi(r)$ is proportional to the square of the curvature of graphene $$\Phi(r)= -\alpha\frac{3a^{2}}{4}(\nabla^2(h(r))^2$$ where $h(r)$ is the local height referenced to a flat configuration, $a$ is the nearest neighbor distance and $\alpha \approx$ 9.2 eV is a constant [@kim2008]. This implies that curvature of the graphene sheet leads to density variations due to the shifting chemical potential. In addition to the intrinsic rippling in graphene, its two dimensional nature makes its entire surface area susceptible to adsorption and defects. Given the intricacies involved in graphene device preparation the presence of random impurities, adsorbates or defects from various sources is difficult to prevent. Depending on the type of imperfection, they can either act as long or short range scattering sites. Coulomb scattering by random charged impurities at the graphene-SiO$_2$ boundary is one example of long range scattering. This long range, intravalley, scattering tends to create inhomogeneities in the carrier density forming electron and hole puddles [@hwang2007; @martin2008]. When the scattering potential has components shorter than the lattice constant, typically due to lattice defects, intervalley scattering takes place which mixes the two sublattices of graphene [@ando]. Here we use a low temperature scanning tunneling microscope (STM) to probe the morphology, local electronic properties and scattering phenomena in monolayer graphene. Graphene was prepared by mechanical exfoliation of graphite on SiO$_2$ [@novoselov2004; @novoselov2005]. Monolayer areas were identified using an optical microscope and then Ti/Au electrodes were deposited using standard electron beam lithography. The lithography process leaves some PMMA resist on the surface of graphene. To eliminate the resist the device was annealed in argon and hydrogen at 400 °C for 1 hour [@meyer2007] followed by annealing in air at 300 °C for 30 minutes. The device was then immediately transferred to the STM (Omicron low temperature STM operating at T = 4.5 K in ultrahigh vacuum (p $\leq 10^{-11}$ mbar)). Electrochemically etched tungsten tips were used for imaging and spectroscopy. All of the tips used were first checked on an Au surface to ensure that their density of states was constant. A schematic of the measurement set-up showing the graphene flake on SiO$_2$ with a gold electrode for electrical contact is shown in Fig. \[schematic\](a). A typical STM image of the monolayer graphene showing the complete hexagonal lattice along with surface corrugations due to the underlying SiO$_2$ substrate is shown in Fig. \[schematic\](b). The observation of both sublattices, giving the hexagonal structure, is characteristic of monolayer graphene as opposed to bulk graphite where only a single sublattice is usually observed giving a triangular pattern. The hexagonal lattice in our images extends over areas as large as 40 nm demonstrating the cleanliness of the surface. The uneven surface underneath the hexagons is a distinctive feature of all our images. This additional corrugation has a height variation of $\sim$ 5Å over an area of 30 $\times$ 30 nm$^2$. The lateral extent of these corrugations is in the range of a few nanometers mimicking the SiO$_2$ corrugation [@ishigami2007; @stolyarova2007]. We have recorded the local density of states of graphene using dI/dV point spectroscopy measurements. In this case, the tip is fixed at a specific location on the sample, the feedback loop is turned off and the sample voltage is ramped within a specific energy window. An ac modulation voltage of 4 mV rms, 574 Hz is applied to the sample and the resulting dI/dV spectrum is recorded using lockin detection. One such spectrum is shown in Fig. \[pointspec\]. The dI/dV spectrum is approximately linear in energy and featureless. Neither a gap nor a dip is seen around the Fermi level. To understand the effect of the corrugations seen in Fig. \[schematic\](b) on the density of states of graphene, we have performed spatially resolved scanning tunneling spectroscopy (STS) measurements. By measuring the differential conductance, dI/dV as a function of energy and position, we obtain maps of the local density of states (LDOS). Figure \[puddles\] shows atomically resolved topography and spectroscopy over a 40 $\times$ 40 nm$^2$ area. The topography of the area, Fig. \[puddles\](a), demonstrates the atomic resolution as well as the corrugations. Fig. \[puddles\](b)-(e) are maps of the LDOS over the same area for four different sample voltages ranging from -0.2 V to 0.4 V. The maps show areas of high and low differential conductance which change as a function of energy. These changes demonstrate that the Dirac point varies as a function of position as discussed below. At low energy, the shifting of the Dirac point leads to electron and hole puddles. When performing the spectroscopy measurements, the feedback circuit is turned off and the tip height is held constant. The height is determined by the setpoint current, $I_s$, (100 pA in this case) and the setpoint voltage, $V_s$, (0.4 V in this case). If we assume that the Dirac point is at the Fermi energy, then the dI/dV curve can be written as $\frac{dI}{dV}={\alpha}{|V|}$ where $V$ is the sample voltage (green curve in Fig. \[cross\]). The value of $\alpha$ is determined by the setpoint parameters, $I_s$ and $V_s$. This is because the feedback circuit adjusts the tip height such that the integrated current from 0 to $V_s$ is $I_s$. The value of $\alpha$ is determined by setting the area under the dI/dV curve equal to $I_s$, $I_s=\frac{\alpha V_s^2}{2}$. This gives a value of $\alpha=\frac{2I_s}{V_s^2}$ and therefore $\frac{dI}{dV}=\frac{2I_s}{V_s^2}|V|$ (green curve in Fig. \[cross\](a)). However, in a region where the Dirac point shifts by an amount $\Delta<0$ the sample essentially becomes more conductive. Therefore, the tip must move farther from the surface to maintain the setpoint current causing the dI/dV curve to have a lower slope. Now, $\frac{dI}{dV}=\beta|V-\Delta|$ as shown by the blue curve in Fig. \[cross\](a). The value of $\beta$ is once again determined by the requirement that the area under the $\frac{dI}{dV}$ curve from 0 to $V_s$ is $I_s$. This gives $\beta=\frac{2I_s}{V_s^2-2\Delta V_s}$. The point where two dI/dV curves with slopes $\alpha$ and $\beta$ cross can be obtained by setting the dI/dV curves to be equal and solving for V. The result is that the two curves intersect at $V_s/2$. When the Dirac point shifts by an amount $\Delta>0$, $\frac{dI}{dV}=\gamma|V-\Delta|$ and a similar effect occurs except the tip moves closer to the surface and the slope of the dI/dV curve increases (red curve in Fig. \[cross\](a)). The value of $\gamma$ is given by $$\gamma=\frac{2I_s}{V_s^2-2\Delta V_s+ 2\Delta^2}$$ Again the dI/dV curves cross at $V_s /2$ if $\Delta << V_s$ and at negative sample voltages, the curves have different slopes based on the energy of the Dirac point but they do not cross. This analysis makes two predictions for the LDOS maps. (1) The variation in the maps should be smallest at $V_s/2$ because regardless of the energy of the Dirac point, all the dI/dV curves have the same value at $V_s/2$. (2) The value of dI/dV at $V_s$ is related to the energy of the Dirac point by the relationship $\Delta=V_s(\frac{1-\eta/2}{1-\eta})$, where $\eta$ is the normalized differential conductance given by $\eta=\frac{dI/dV}{I/V}$. We have studied the variation in the LDOS maps for a set of three different locations and tunneling parameters. The results are shown in Fig. \[cross\](b). The blue curve is for the images shown in Fig. \[puddles\] where $V_s$ was 0.4 V. From the curve it is clear that the minimum variation occurs at 0.2 V as can also be seen in Fig. \[puddles\](d). The green curve is taken in a different region of the sample with $V_s= 0.25$ V. The minimum has shifted to lower energy compared to the blue curve and it is now at 0.125 V. Lastly, the yellow curve has $V_s = 0.15$ V and the minimum is at the lowest energy of the three curves. This shows that the location of the minimum in the LDOS maps is dependent on the parameters used for the measurement but the value of dI/dV at $V_s$ still gives information on the Dirac point. Fig. \[puddles\](f) shows the dI/dV curves taken from the LDOS maps for five different areas of the image as shown in Fig. \[puddles\](a). These curves exhibit a range of dI/dV values at $V_s$ which is evidence for the formation of electron and hole puddles. In Fig. \[puddles\](e), the red regions (large dI/dV) are locations where the Dirac point has shifted towards positive sample voltage while in the blue regions (smaller dI/dV) the Dirac point has shifted towards negative sample voltage. From the changes in the dI/dV curves we estimate that the shift in the Dirac point, $\Delta E_d$, is about 77 mV in our images. As discussed earlier in this manuscript intrinsic ripples in graphene give rise to a spatially varying electrochemical potential (equation (1)). This equation predicts that highly curved regions of the sample will be electron doped while flat regions will be hole doped. Using a STM topographic image, Fig. \[curve\](a), we have calculated the shift in the electrochemical potential caused by the graphene curvature using equation (1). The results in Fig. \[curve\](b) show areas of high curvature, large negative change in chemical potential, in blue while relatively flat regions are brown. A comparison of the shift in Dirac point from the dI/dV map at 0.15 V, Fig. \[curve\](c) and the electrochemical potential landscape Fig. \[curve\](b) shows there are limited regions which are in agreement between them. For example the curved region in the bottom right side is electron doped (blue region in Fig. \[curve\](c) and Fig. \[curve\](b)). However, this is not always the case; there are curved regions such as the blue patch in the center of Fig. \[curve\](b) that is hole doped as seen in the center of Fig. \[curve\](c). Thus from our measurements we conclude that the curvature in the graphene flake contributes to a variation in the electrochemical potential but it is not the main factor responsible for the features in the dI/dV map. Instead, the potential variation is due to a combination of the ripples and long range scatterers such as random charged impurities present on the graphene sheet [@martin2008]. These electron and hole puddles, a signature of disorder in graphene, are also responsible for the finite minimum conductivity at the Dirac point [@hwang2007]. A second source of potential variation in graphene is long range scattering from random charged impurities present at the graphene substrate interface. From our measured shift in the Dirac point, $\Delta E_d$, we calculate the density variation using $\Delta n =\frac{(\Delta E_d)^2}{\pi \hbar^2 v_F^2}$ to be $4.3\times 10^{11}$cm$^{-2}$. An independent measurement of the impurity density can be done using the voltage dependence of the conductivity, $\sigma=20 e \epsilon V_g / (h n_i t)$ where $\epsilon$ is the dielectric constant of SiO$_2$, $t$ is the oxide thickness and $V_g$ is the gate voltage [@adam2007]. We measured another flake of graphene processed in the same manner as the one used for the STS measurements and found $n_i = 4\times 10^{11}$cm$^{-2}$. This impurity density can be converted to local electron density fluctuations using $\delta n^2 = n_i / (8\pi d^2)$ where $d$ is the distance of the impurities from the graphene[@hwang2007]. This agreement between our spectroscopy measurements and transport measurements as well as transport measurements by other groups [@hwang2007] indicates that the role of charged impurities is critical for understanding current graphene devices. While long range scatterers give rise to intravalley scattering, within one sublattice, creating the electron and hole puddles seen in the LDOS images, short range scatterers such as lattice defects can also be present in graphene. These short range scatterers induce intervalley scattering from one Dirac cone to the other. The mobility of graphene was found to be insensitive to doping with gas molecules, long range scatterers [@schedin2007]. This implies that there is another source which may already limit the mobility of the graphene such as short range scatterers which tend to give a constant resistivity and hence low mobility [@katsnelson2008; @chen2008]. We observe lattice defects in the monolayer graphene images and analyze the resulting Fourier transforms of the LDOS maps and topographs for an insight into intervalley scattering. In the case of epitaxial graphene and graphite, lattice defects have been shown to give rise to scattering and interference [@rutter2007; @ruffieux2005]. Figure \[transforms\] shows the results for graphene on SiO$_2$. Atomically resolved topography, Fig. \[transforms\](a), shows the hexagonal lattice of graphene and the corresponding Fourier transform, Fig. \[transforms\](c), contains the reciprocal lattice points (red circles) and components due to the C-C bonds (brown circles) which are longer and rotated by 30 degrees. Figure \[transforms\](b) is an image of the LDOS at the Fermi level showing the defect induced interference pattern superimposed on the electron and hole puddles due to the long range scattering. The Fourier transform of this image, Fig. \[transforms\](d), shows two hexagonal patterns. The outer hexagon, red circles, is due to the reciprocal lattice and is located at the same points as in Fig. \[transforms\](c). There is also an additional inner hexagon, blue circles, that is rotated by 30 degrees due to scattering. This is the $\sqrt{3} \times \sqrt{3}$ R30$^\circ$ interference pattern. The six peaks at the K$_\pm$ points in Fig. \[transforms\](d), are a result of intervalley, short range, scattering events. To characterize the strength of intervalley scattering as a function of energy, we have fit the six peaks at the K$_\pm$ points of the $\sqrt{3} \times \sqrt{3}$ R30$^\circ$ interference pattern with a 2D Lorentzian to determine the area under the peak. We do not observe any structures around these K$_\pm$ points unlike those observed in case of graphene on SiC [@brihuega2008]. Hence we have chosen a Lorentzian for the fit. This area is then divided by the average value of the Fourier transform to get the relative amount of intervalley scattering at a given energy. As an example, Figure \[scattering\](a) shows one such Fourier transform with six peaks that are fit with a 2D Lorentzian. Figure \[scattering\](b) shows the linescan across one of the 6 peaks in the Fourier transform and the Lorentzian fit corresponding to that peak. Figure \[scattering\](c) plots the relative strength of these peaks as a function of sample voltage for three different regions of the sample. The broad peak in scattering at the Fermi energy is independent of the imaging parameters. It is consistently observed in areas of the flake with scatterers or defects. As we move away from the Fermi energy the strength of scattering decreases sharply. This enhanced intervalley scattering at low energy is evidence of weak localization of carriers [@rutter2007; @morgenstern2002]. In conclusion, we have presented an extensive topographic and spectroscopic investigation of monolayer graphene on SiO$_2$ using a STM at 4.5 K. We were able to atomically resolve large areas of exfoliated monolayer graphene, record energy resolved local density of states maps and interpret the maps to be a signature of the shifting Dirac point. Also we identified and analyzed intervalley scattering mechanisms on exfoliated monolayer graphene. Thus, the electronic properties of monolayer graphene are closely related to intrinsic ripples, SiO$_2$ substrate morphology and random impurities. Investigation of monolayer graphene flakes can be extended to different geometries and substrates to advance our understanding of the subtleties of the electronic properties of graphene. CNL, WB and FM acknowledge the support of NSF CAREER DMR/0748910, NSF CBET/0756359 and ONR/DMEA Award H94003-07-2-0703. [10]{} K.S. Novoselov and A.K. Geim, Nat. Mater. [**6**]{}, 183 (2007). K.S. Novoselov *et al.*, Proc. Natl. Acad. Sci. USA [**102**]{}, 10451 (2005). K.S. Novoselov *et al.*, Science [**306**]{}, 666 (2004). K.S. Novoselov *et al.*, Nature (London) [**438**]{}, 197 (2005). Y.B. Zhang *et al.*, Nature (London) [**438**]{}, 201 (2006). F. Miao *et al.*, Science [**317**]{}, 1530 (2007). J.C. Meyer *et al.*, Nature (London) [**446**]{}, 60 (2007). A. Fasolino, J.H. Los, and M.I. Katsnelson, Nature Mater. [**6**]{}, 858 (2007). S.V. Morozov *et. al*,.Phys. Rev. Lett. [**97**]{}, 016801 (2006). F. Schedin *et. al*, Nature Mater. [**6**]{}, 652 (2007). M.J. Ishigami *et al.* Nano Lett. [**7**]{}, 1643 (2007). E. Stolyarova *et al.* Proc. Natl. Acad. Sci. USA [**104**]{}, 9209 (2007). V. Geringer *et. al*, Phys. Rev. Lett. [**102**]{}, 076102 (2009). Y. Zhang *et. al*, Nature Phys. [**4**]{}, 627 (2008). E.A. Kim and A.H. Castro Neto, EuroPhys. Lett. [**84**]{}, 57007 (2008). E.H. Hwang, S. Adam, and S. Das Sarma, Phys. Rev. Lett. [**98**]{}, 186806 (2007). J. Martin *et. al*, Nature Phys. [**4**]{}, 144 (2008). T. Ando and T. Nakanishi, J. Phys. Soc. Jpn [**67**]{}, 1704 (1998); T. Ando, Phil. Trans. R. Soc. A [**366**]{}, 221 (2008). S. Adam *et al.*, Proc. Natl. Acad. Sci. U.S.A. [**104**]{}, 18392 (2007). M.I. Katsnelson and A.K. Geim, Phil. Trans. R. Soc. A [**366**]{}, 195 (2008). J.H. Chen *et al.*, Nature Phys. [**4**]{}, 377 (2008). G.M. Rutter *et al.*, Science [**317**]{}, 219 (2007). P. Ruffieux *et al.*, Phys. Rev. B [**71**]{}, 153403 (2005). I. Brihuega *et al.*, Phys. Rev. Lett. [**101**]{}, 206802 (2008). M. Morgenstern *et al.*, Phys. Rev. Lett. [**89**]{}, 136806 (2002).

ArXiv

LPENSL - 2014\ IMB - 2014 [Open spin chains with generic integrable boundaries:\ Baxter equation and Bethe ansatz completeness from SOV]{} [**N. Kitanine**,   **J.-M. Maillet**,   **G. Niccoli** ]{} **Abstract** Introduction ============ The functional characterization of the complete transfer matrix spectrum associated to the most general spin-1/2 representations of the 6-vertex reflection algebra on general inhomogeneous chains is a longstanding open problem. It has attracted much attention in the framework of quantum integrability producing so far only partial results. The interest in the solution of this problem is at least twofold. On the one hand, the quantum integrable system associated to the limit of the homogeneous chain, i.e. the open spin-1/2 XXZ quantum chain with arbitrary boundary magnetic fields, is an interesting physical quantum model. It appears, in particular, in the context of out-of-equilibrium physics ranging from the relaxation behavior of some classical stochastic processes, as the asymmetric simple exclusion processes [@EssD05; @EssD06], to the transport properties of the quantum spin systems [@SirPA09; @Pro11]. Their solution can lead to non-perturbative physical results and a complete and manageable functional characterization of their spectrum represents the first fundamental steps in this direction. On the other hand, it is important to remark that the analysis of the spectral problem of these integrable quantum models turned out to be quite involved by standard Bethe ansatz [@Bet31; @FadST79] techniques. Therefore, these quantum models are natural laboratories where to define alternative non-perturbative approach to their exact solution. Indeed, the algebraic Bethe ansatz, introduced for open systems by Sklyanin [@Skl88] based on the Cherednik’s reflection equation [@Che84], in the case of open XXZ quantum spin chains can be applied directly only in the case of parallel z-oriented boundary magnetic fields. Under these special boundary conditions the spectrum is naturally described by a finite system of Bethe ansatz equations. Moreover the dynamics of such systems can be studied by exact computation of correlation functions [@KitKMNST07; @KitKMNST08], derived from a generalisation of the method introduced in [@KitMT99; @KitMT00; @MaiT00] for periodic spin chains. Introducing a Baxter $T$-$Q$ equation, Nepomechie [@Nep02; @Nep04] first succeeded to describe the spectrum of the XXZ spin chain with non-diagonal boundary terms in the case of an anisotropy parameter associated to the roots of unity; furthermore, the result was obtained there only if the boundary terms satisfied a very particular constraint relating the magnetic fields on the two boundaries. This last constraint was also used in [@CaoLSW03] to introduce a generalized algebraic Bethe ansatz approach to this problem inspired by papers of Baxter [@Bax72; @Bax72a] and of Faddeev and Takhtadjan [@FadT79] on the XYZ spin chain. This method has led to the first construction of the eigenstates of the XXZ spin chain with non z-oriented boundary magnetic fields and this construction has been obtained for a general anisotropy parameter, i.e., not restricted to the roots of unity cases[^1]. In [@YanZ07] a different version of this technique based on the vertex-IRF transformation was proposed but in fact it required one additional constraint on the boundary parameters to work. It is worth mentioning that even if these constrained boundary conditions are satisfied and generalized Bethe ansatz method gives a possibility to go beyond the spectrum, as it was done for the diagonal boundary conditions, no representation for the scalar product of Bethe vectors[^2] and hence for the correlation functions were obtained. This spectral problem in the most general setting has then been also addressed by other approaches. It is worth mentioning a new functional method leading to nested Bethe ansatz equations presented in [@Gal08] for the eigenvalue characterization and analogous to those previously introduced in [@MurN05] by a generalized $T$-$Q$ formalism. The eigenstate construction has been considered in these general settings in [BasK05a,Bas06]{} by developing the so-called $q$-Onsager algebra formalism. In this last case the characterization of the spectrum is given by classifying the roots of some characteristic polynomials. More recently, in [@CaoYSW13-4] an ansatz $T$-$Q$ functional equations for the spin chains with non-diagonal boundaries has been proposed[^3]. It is extremely important to remark that in general all methods based on Bethe ansatz (or generalized Bethe ansatz) are lacking proofs of the completeness of the spectrum and in most cases the only evidences of completeness are based on numerical checks for short length chains. This is the case for the XXZ chain with non-diagonal boundary matrices with the boundary constraint for which the completeness of the spectrum description by the associated system of Bethe ansatz equations has been studied numerically [@Nep-R-2003; @Nep-R-2003add]. In the case of the XXZ chain with completely general non-diagonal boundary matrices some numerical analysis is also presented in [@CaoYSW13-4]. Further numerical analysis have been developed in a much simpler case of the isotropic XXX spin chain where the most general boundary conditions can be always reduced by using the $SU(2)$ symmetry to one diagonal and one non-diagonal boundary matrices. For the XXX chains the ansatz introduced in [@CaoYSW13-2] was also applied and the completeness of the Bethe ansatz spectrum was checked numerically [@CaoJYW2013]. It is also important to mention a simplified ansatz proposed by Nepomechie based on a standard second order difference functional $T$-$Q$ equation with an additional inhomogeneous term. The completeness of the Bethe ansatz spectrum has been verified numerically for small XXX chains in [@Nep-2013] while in [@Nep-W-2013] the problem of the description of some thermodynamical properties has been addressed. These interesting developments attracted our attention in connection to the quantum separation of variables (SOV) method pioneered by Sklyanin [@Skl85; @Skl92]. The first analysis of the spin chain in the classical limit from this point of view was performed in [@Skl89a; @Skl89b]. This alternative approach allows to obtain (mainly by construction) the complete set of eigenvalues and eigenvectors of quantum integrable systems. In particular, it was recently developed [NicT10,N10-1,N10-2,GroMN12,GroN12,Nic12b,Nic13a,N13-1,Nic13b,Nic13c,Fald-KN13,FaldN13,GroMN13]{} for a large variety of quantum models not solvable by algebraic Bethe ansatz. Moreover it has been shown first in [@GroMN12] that once the SOV spectrum characterization is achieved manageable and rather universal determinant formulae can be derived for matrix elements of local operators between transfer matrix eigenstates. In particular, this SOV method was first developed in [@Nic12b] for the spin-1/2 representations of the 6-vertex reflection algebra with quite general non-diagonal boundaries and then generalized to the most general boundaries in [@Fald-KN13]. There, it gives the complete spectrum (eigenvalues and eigenstates) and already allows to compute matrix elements of some local operators within this most general boundary framework. However, it is important to remark that this SOV characterization of the spectrum is somehow unusual in comparison to more traditional characterizations like those obtained from Bethe ansatz techniques. More precisely, the spectrum is described not in terms of the set of solutions to a standard system of Bethe ansatz equations but is given in terms of sets of solutions to a characteristic system of $\mathsf{N}$ quadratic equations in $\mathsf{N}$ unknowns, $\mathsf{N}$ being the number of sites of the chain. While the clear advantage of this SOV characterization is that it permits to characterize completely the spectrum without introducing any ansatz one has to stress that the classification of the sets of solutions of the SOV system of quadratic equations represents a new problem in quantum integrability which requires a deeper and systematic analysis. The aim of the present article is to show that the SOV analysis of the transfer matrix spectrum associated to the most general spin-1/2 representations of the 6-vertex reflection algebra on general inhomogeneous chains is strictly equivalent to a system of *generalized* Bethe ansatz equations. This ensures that this system of Bethe equations characterizes automatically the entire spectrum of the transfer matrix. More in detail, we prove that the SOV characterization is equivalent to a second order finite difference functional equation of Baxter type: $$\tau (\lambda )Q(\lambda )=\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A} (-\lambda )Q(\lambda +\eta )+F(\lambda ),$$ which contains an inhomogeneous term $F(\lambda )$ independent on the $\tau $ and $Q$-functions and entirely fixed by the boundary parameters. It vanishes only for some special but yet interesting non-diagonal boundary conditions (corresponding to the boundary constraints mentioned above). One central requirement in our construction of this functional characterization is the polynomial character of the $Q$-function. Indeed, it is this requirement that allows then to show that a finite system of equations of generalized Bethe ansatz type can be used to describe the complete transfer matrix spectrum. Note that similar results on the reformulation of the SOV spectrum characterization in terms of functional $T$-$Q$ equations with $Q$-function solutions in a well defined model dependent set of polynomials were previously derived [@N10-1; @N10-2; @GroN12] for the cases of transfer matrices associated to cyclic representations of the Yang-Baxter algebra. The analysis presented here is also interesting as it introduces the main tools to generalize this type of reformulation to other classes of integrable quantum models. The article is organized as follows. In Section 2 we set the main notations and we recall the main results of previous papers on SOV necessary for our purposes. Section 3 contains the main results of the paper with the reformulation of the SOV characterization of the transfer matrix spectrum in terms of the inhomogeneous Baxter functional equation and the associated finite system of generalized Bethe ansatz equations. In Section 4 we define the boundary conditions for which the inhomogeneity in the Baxter equation identically vanishes, in this way deriving the completeness of standard Bethe ansatz equations. There, we moreover derive the SOV spectrum functional reformulation for the remaining boundary conditions compatibles with homogeneous Baxter equations. Section 5 contains the description of a set of discrete transformations which leave unchanged the SOV characterization of the spectrum in this way proving the isospectrality of the transformed transfer matrices. These symmetries are used to find equivalent functional equation characterizations of the spectrum which allow to generalize the results described in Section 3 and 4. In Section 6 we present the SOV characterization of the spectrum for the rational 6-vertex representation of the reflection algebra and the reformulation of the spectrum by inhomogeneous Baxter equation. Finally, in Section 7, we present a comparison with the known numerical results in the literature for both the XXZ and XXX chains; the evidenced compatibility suggests that even in the homogenous chains our spectrum description is still complete. Separation of variable for spin-1/2 representations of the reflection algebra ============================================================================= Spin-1/2 representations of the reflection algebra and open XXZ quantum chain ----------------------------------------------------------------------------- The representation theory of the reflection algebra can be studied in terms of the solutions $\mathcal{U}(\lambda )$ (monodromy matrices) of the following reflection equation:$$R_{12}(\lambda -\mu )\,\mathcal{U}_{1}(\lambda )\,R_{21}(\lambda +\mu -\eta )\,\mathcal{U}_{2}(\mu )=\mathcal{U}_{2}(\mu )\,R_{12}(\lambda +\mu -\eta )\,\mathcal{U}_{1}(\lambda )\,R_{21}(\lambda -\mu ). \label{bYB}$$Here we consider the reflection equation associated to the 6-vertex trigonometric $R$ matrix $$R_{12}(\lambda )=\left( \begin{array}{cccc} \sinh (\lambda +\eta ) & 0 & 0 & 0 \\ 0 & \sinh \lambda & \sinh \eta & 0 \\ 0 & \sinh \eta & \sinh \lambda & 0 \\ 0 & 0 & 0 & \sinh (\lambda +\eta )\end{array}\right) \in \text{End}(\mathcal{H}_{1}\otimes \mathcal{H}_{2}),$$where $\mathcal{H}_{a}\simeq \mathbb{C}^{2}$ is a 2-dimensional linear space. The 6-vertex trigonometric $R$-matrix is a solution of the Yang-Baxter equation:$$R_{12}(\lambda -\mu )R_{13}(\lambda )R_{23}(\mu )=R_{23}(\mu )R_{13}(\lambda )R_{12}(\lambda -\mu ).$$The most general scalar solution ($2\times 2$ matrix) of the reflection equation reads$$K(\lambda ;\zeta ,\kappa ,\tau )=\frac{1}{\sinh \zeta }\left( \begin{array}{cc} \sinh (\lambda -\eta /2+\zeta ) & \kappa e^{\tau }\sinh (2\lambda -\eta ) \\ \kappa e^{-\tau }\sinh (2\lambda -\eta ) & \sinh (\zeta -\lambda +\eta /2)\end{array}\right) \in \text{End}(\mathcal{H}_{0}\simeq \mathbb{C}^{2}), \label{ADMFKK}$$where $\zeta ,$ $\kappa $ and $\tau $ are arbitrary complex parameters. Using it and following [@Skl88] we can construct two classes of solutions to the reflection equation (\[bYB\]) in the 2$^{\mathsf{N}}$-dimensional representation space:$$\mathcal{H}=\otimes _{n=1}^{\mathsf{N}}\mathcal{H}_{n}.$$Indeed, starting from$$K_{-}(\lambda )=K(\lambda ;\zeta _{-},\kappa _{-},\tau _{-}),\text{ \ \ \ \ }K_{+}(\lambda )=K(\lambda +\eta ;\zeta _{+},\kappa _{+},\tau _{+}),$$where $\zeta _{\pm },\kappa _{\pm },\tau _{\pm }$ are the boundary parameters, the following boundary monodromy matrices can be introduced $$\begin{aligned} \mathcal{U}_{-}(\lambda ) &=&M_{0}(\lambda )K_{-}(\lambda )\widehat{M}_{0}(\lambda )=\left( \begin{array}{cc} \mathcal{A}_{-}(\lambda ) & \mathcal{B}_{-}(\lambda ) \\ \mathcal{C}_{-}(\lambda ) & \mathcal{D}_{-}(\lambda )\end{array}\right) \in \text{End}(\mathcal{H}_{0}\otimes \mathcal{H}), \\ \mathcal{U}_{+}^{t_{0}}(\lambda ) &=&M_{0}^{t_{0}}(\lambda )K_{+}^{t_{0}}(\lambda )\widehat{M}_{0}^{t_{0}}(\lambda )=\left( \begin{array}{cc} \mathcal{A}_{+}(\lambda ) & \mathcal{C}_{+}(\lambda ) \\ \mathcal{B}_{+}(\lambda ) & \mathcal{D}_{+}(\lambda )\end{array}\right) \in \text{End}(\mathcal{H}_{0}\otimes \mathcal{H}).\end{aligned}$$ These matrices $\mathcal{U}_{-}(\lambda )$ and $\mathcal{V}_{+}(\lambda )=\mathcal{U}_{+}^{t_{0}}(-\lambda )$ define two classes of solutions of the reflection equation (\[bYB\]). Here, we have used the notations:$$M_{0}(\lambda )=R_{0\mathsf{N}}(\lambda -\xi _{\mathsf{N}}-\eta /2)\dots R_{01}(\lambda -\xi _{1}-\eta /2)=\left( \begin{array}{cc} A(\lambda ) & B(\lambda ) \\ C(\lambda ) & D(\lambda )\end{array}\right) \label{T}$$and $$\widehat{M}(\lambda )=(-1)^{\mathsf{N}}\,\sigma _{0}^{y}\,M^{t_{0}}(-\lambda )\,\sigma _{0}^{y}, \label{Mhat}$$where $M_{0}(\lambda )\in $ End$(\mathcal{H}_{0}\otimes \mathcal{H})$ is the bulk inhomogeneous monodromy matrix (the $\xi _{j}$ are the arbitrary inhomogeneity parameters) satisfing the Yang-Baxter relation:$$R_{12}(\lambda -\mu )M_{1}(\lambda )M_{2}(\mu )=M_{2}(\mu )M_{1}(\lambda )R_{12}(\lambda -\mu ). \label{YB}$$The main interest of these boundary monodromy matrices is the property shown by Sklyanin [@Skl88] that the following family of transfer matrices:$$\mathcal{T}(\lambda )=\text{tr}_{0}\{K_{+}(\lambda )\,M(\lambda )\,K_{-}(\lambda )\widehat{M}(\lambda )\}=\text{tr}_{0}\{K_{+}(\lambda )\mathcal{U}_{-}(\lambda )\}=\text{tr}_{0}\{K_{-}(\lambda )\mathcal{U}_{+}(\lambda )\}\in \text{\thinspace End}(\mathcal{H}), \label{transfer}$$defines a one parameter family of commuting operators in End$(\mathcal{H})$. The Hamiltonian of the open XXZ quantum spin 1/2 chain with the most general integrable boundary terms can be obtained in the homogeneous limit ($\xi _{m}=0$ for $m=1,\ldots ,\mathsf{N}$) from the following derivative of the transfer matrix (\[transfer\]):$$H=\frac{2(\sinh \eta )^{1-2\mathsf{N}}}{\text{tr}\{K_{+}(\eta /2)\}\,\text{tr}\{K_{-}(\eta /2)\}}\frac{d}{d\lambda }\mathcal{T}(\lambda )_{\,\vrule height13ptdepth1pt\>{\lambda =\eta /2}\!}+\text{constant,} \label{Ht}$$and its explicit form reads: $$\begin{aligned} H& =\sum_{i=1}^{\mathsf{N}-1}(\sigma _{i}^{x}\sigma _{i+1}^{x}+\sigma _{i}^{y}\sigma _{i+1}^{y}+\cosh \eta \sigma _{i}^{z}\sigma _{i+1}^{z}) \notag \\ & +\frac{\sinh \eta }{\sinh \zeta _{-}}\left[ \sigma _{1}^{z}\cosh \zeta _{-}+2\kappa _{-}(\sigma _{1}^{x}\cosh \tau _{-}+i\sigma _{1}^{y}\sinh \tau _{-})\right] \notag \\ & +\frac{\sinh \eta }{\sinh \zeta _{+}}[(\sigma _{\mathsf{N}}^{z}\cosh \zeta _{+}+2\kappa _{+}(\sigma _{\mathsf{N}}^{x}\cosh \tau _{+}+i\sigma _{\mathsf{N}}^{y}\sinh \tau _{+}). \label{H-XXZ-Non-D}\end{aligned}$$Here $\sigma _{i}^{a}$ are local spin $1/2$ operators (Pauli matrices), $\Delta =\cosh \eta $ is the anisotropy parameter and the six complex boundary parameters $\zeta _{\pm }$, $\kappa _{\pm }$ and $\tau _{\pm }$ define the most general integrable magnetic interactions at the boundaries. Some relevant properties ------------------------ The following quadratic linear combination of the generators $\mathcal{A}_{-}(\lambda ),$ $\mathcal{B}_{-}(\lambda ),$ $\mathcal{C}_{-}(\lambda )$ and $\mathcal{D}_{-}(\lambda )$ of the reflection algebra: $$\begin{aligned} \frac{\mathrm{det}_{q}\,\mathcal{U}_{-}(\lambda )}{\sinh (2\lambda -2\eta )}& =\mathcal{A}_{-}(\epsilon \lambda +\eta /2)\mathcal{A}_{-}(\eta /2-\epsilon \lambda )+\mathcal{B}_{-}(\epsilon \lambda +\eta /2)\mathcal{C}_{-}(\eta /2-\epsilon \lambda ) \label{q-detU_1} \\ & =\mathcal{D}_{-}(\epsilon \lambda +\eta /2)\mathcal{D}_{-}(\eta /2-\epsilon \lambda )+\mathcal{C}_{-}(\epsilon \lambda +\eta /2)\mathcal{B}_{-}(\eta /2-\epsilon \lambda ), \label{q-detU_2}\end{aligned}$$where $\epsilon =\pm 1$, is the *quantum determinant* . It was shown by Sklyanin that it is a central element of the reflection algebra$$\lbrack \mathrm{det}_{q}\,\mathcal{U}_{-}(\lambda ),\mathcal{U}_{-}(\mu )]=0.$$The quantum determinant plays a fundamental role in the characterization of the transfer matrix spectrum and it admits the following explicit expressions:$$\begin{aligned} \mathrm{det}_{q}\,\mathcal{U}_{-}(\lambda ) &=&\mathrm{det}_{q}K_{-}(\lambda )\mathrm{det}_{q}M_{0}(\lambda )\mathrm{det}_{q}M_{0}(-\lambda ) \label{q-detU_-exp} \\ &=&\sinh (2\lambda -2\eta )\mathsf{A}_{-}(\lambda +\eta /2)\mathsf{A}_{-}(-\lambda +\eta /2),\end{aligned}$$where: $$\mathrm{det}_{q}M(\lambda )=a(\lambda +\eta /2)d(\lambda -\eta /2), \label{bulk-q-det}$$is the bulk quantum determinant and$$\mathrm{det}_{q}K_{\pm }(\lambda )=\mp \sinh (2\lambda \pm 2\eta )g_{\pm }(\lambda +\eta /2)g_{\pm }(-\lambda +\eta /2).$$Here, we used the following notations:$$\mathsf{A}_{-}(\lambda )=g_{-}(\lambda )a(\lambda )d(-\lambda ),\text{ \ }d(\lambda )=a(\lambda -\eta ),\text{ \ \ }a(\lambda )=\prod_{n=1}^{\mathsf{N}}\sinh (\lambda -\xi _{n}+\eta /2), \label{eigenA}$$and$$g_{\pm }(\lambda )=\frac{\sinh (\lambda +\alpha _{\pm }-\eta /2)\cosh (\lambda \mp \beta _{\pm }-\eta /2)}{\sinh \alpha _{\pm }\cosh \beta _{\pm }}, \label{g_PM}$$where $\alpha _{\pm }$ and $\beta _{\pm }$ are defined in terms of the boundary parameters by:$$\sinh \alpha _{\pm }\cosh \beta _{\pm }=\frac{\sinh \zeta _{\pm }}{2\kappa _{\pm }},\text{ \ \ \ \ \ }\cosh \alpha _{\pm }\sinh \beta _{\pm }=\frac{\cosh \zeta _{\pm }}{2\kappa _{\pm }}. \label{alfa-beta}$$ \[normality\]The transfer matrix $\mathcal{T}(\lambda )$ is an even function of the spectral parameter $\lambda $:$$\mathcal{T}(-\lambda )=\mathcal{T}(\lambda ), \label{even-transfer}$$and it is central for the following special values of the spectral parameter: $$\begin{aligned} \lim_{\lambda \rightarrow \pm \infty }e^{\mp 2\lambda (\mathsf{N}+2)}\mathcal{T}(\lambda ) &=&2^{-(2\mathsf{N}+1)}\frac{\kappa _{+}\kappa _{-}\cosh (\tau _{+}-\tau _{-})}{\sinh \zeta _{+}\sinh \zeta _{-}}, \label{Central-asymp} \\ \mathcal{T}(\pm \eta /2) &=&(-1)^{\mathsf{N}}2\cosh \eta \mathrm{det}_{q}M(0), \label{Central-1} \\ \mathcal{T}(\pm (\eta /2-i\pi /2)) &=&-2\cosh \eta \coth \zeta _{-}\coth \zeta _{+}\mathrm{det}_{q}M(i\pi /2). \label{Central-2}\end{aligned}$$Moreover, the monodromy matrix $\mathcal{U}_{\pm }(\lambda )$ satisfy the following transformation properties under Hermitian conjugation: - Under the condition $\eta \in i\mathbb{R}$ (massless regime), it holds: $$\mathcal{U}_{\pm }(\lambda )^{\dagger }=\left[ \mathcal{U}_{\pm }(-\lambda ^{\ast })\right] ^{t_{0}}, \label{ml-Hermitian_U}$$ for $\{i\tau _{\pm },i\kappa _{\pm },i\zeta _{\pm },\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{R}^{\mathsf{N}+3}.$ - Under the condition $\eta \in \mathbb{R}$ (massive regime), it holds: $$\mathcal{U}_{\pm }(\lambda )^{\dagger }=\left[ \mathcal{U}_{\pm }(\lambda ^{\ast })\right] ^{t_{0}}, \label{m-Hermitian_U}$$for $\{\tau _{\pm },\kappa _{\pm },\zeta _{\pm },i\xi _{1},...,i\xi _{\mathsf{N}}\}\in \mathbb{R}^{\mathsf{N}+3}.$ So under the same conditions on the parameters of the representation it holds: $$\mathcal{T}(\lambda )^{\dagger }=\mathcal{T}(\lambda ^{\ast }), \label{I-Hermitian_T}$$i.e. $\mathcal{T}(\lambda )$ defines a one-parameter family of normal operators which are self-adjoint both for $\lambda $ real and purely imaginary. SOV representations for $\mathcal{T}(\protect\lambda )$-spectral problem ------------------------------------------------------------------------ Let us recall here the characterization obtained in [@Nic12b; @Fald-KN13] by SOV method of the spectrum of the transfer matrix $\mathcal{T}(\lambda )$. First we introduce the following notations:$$X_{k,m}^{(i,r)}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })\equiv \left( -1\right) ^{i}\left( 1-r\right) \eta +\tau _{-}-\tau_{+}+(-1)^{k}(\alpha _{-}+\beta _{-})-(-1)^{m}(\alpha _{+}-\beta _{+})+i\pi (k+m), \label{SOV-cond-}$$and by using these linear combinations of the boundary parameters we introduce the set $N_{SOV}\subset\mathbb{C}^6$ of boundary parameters for which the separation of variables cannot be applied directly. More precisely $$(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha_{-},\beta _{-})\in N_{SOV},$$ if $\exists (k,h,m,n)\in \left\{ 0,1\right\} $ such that $$X_{k,m}^{(0,\mathsf{N})}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })=0 \quad\text{and}\quad X_{h,n}^{(1,\mathsf{N})}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })=0.$$All the results in the following will be obtained for the generic values of the boundary parameters, not belonging to this set. The SOV method applicability can be further extended applying the discrete symmetries discussed in the Section \[sect-descretesym\]. Following [@Fald-KN13] we define the functions:$$\begin{aligned} g_{a}(\lambda ) &=&\frac{\cosh ^{2}2\lambda -\cosh ^{2}\eta }{\cosh ^{2}2\zeta _{a}^{(0)}-\cosh ^{2}\eta }\,\prod_{\substack{ b=1 \\ b\neq a}}^{\mathsf{N}}\frac{\cosh 2\lambda -\cosh 2\zeta _{b}^{(0)}}{\cosh 2\zeta _{a}^{(0)}-\cosh 2\zeta _{b}^{(0)}}\quad \text{ \ for }a\in \{1,...,\mathsf{N}\}, \\ \mathbf{A}(\lambda ) &=&(-1)^{\mathsf{N}}\frac{\sinh (2\lambda +\eta )}{\sinh 2\lambda }g_{+}(\lambda )g_{-}(\lambda )a(\lambda )d(-\lambda ),\end{aligned}$$and$$\begin{aligned} f(\lambda )=& \frac{\cosh 2\lambda +\cosh \eta }{2\cosh \eta }\prod_{b=1}^{\mathsf{N}}\frac{\cosh 2\lambda -\cosh 2\zeta _{b}^{(0)}}{\cosh \eta -\cosh 2\zeta _{b}^{(0)}}\mathbf{A}(\eta /2) \notag \\ & -(-1)^{\mathsf{N}}\frac{\cosh 2\lambda -\cosh \eta }{2\cosh \eta }\prod_{b=1}^{\mathsf{N}}\frac{\cosh 2\lambda -\cosh 2\zeta _{b}^{(0)}}{\cosh \eta +\cosh 2\zeta _{b}^{(0)}}\mathbf{A}(\eta /2+i\pi /2) \notag \\ & +2^{(1-\mathsf{N})}\frac{\kappa _{+}\kappa _{-}\cosh (\tau _{+}-\tau _{-})}{\sinh \zeta _{+}\sinh \zeta _{-}}(\cosh ^{2}2\lambda -\cosh ^{2}\eta )\prod_{b=1}^{\mathsf{N}}(\cosh 2\lambda -\cosh 2\zeta _{b}^{(0)}), \label{f-function}\end{aligned}$$where$$\zeta _{n}^{(h_{n})}=\xi _{n}+(h_{n}-\frac{1}{2})\eta \quad \forall n\in \{1,...,\mathsf{N}\},\text{ }h_{n}\in \{0,1\}\text{.}$$We can now recall the main result on the characterization of the set $\Sigma _{\mathcal{T}}$ formed by all the eigenvalue functions of the transfer matrix $\mathcal{T}(\lambda )$. \[C:T-eigenstates-\]Let $(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}\backslash N_{SOV}$ and let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C}$ $^{\mathsf{N}}$ be generic:$$\xi _{a}\neq \pm\xi _{b}+r\eta\text{\ \ mod\,}2\pi \text{ \ }\forall a\neq b\in \{1,...,\mathsf{N}\}\,\,\text{and\thinspace \thinspace }r\in \{-1,0,1\}, \label{xi-conditions}$$then $\mathcal{T}(\lambda )$ has simple spectrum and the set of its eigenvalues $\Sigma _{\mathcal{T}}$ is characterized by:$$\Sigma _{\mathcal{T}}=\left\{ \tau (\lambda ):\tau (\lambda )=f(\lambda )+\sum_{a=1}^{\mathsf{N}}g_{a}(\lambda )x_{a},\text{ \ \ }\forall \{x_{1},...,x_{\mathsf{N}}\}\in \Sigma _{T}\right\} , \label{Interpolation-Form-T}$$where $\Sigma _{T}$ is the set of solutions to the following inhomogeneous system of $\mathsf{N}$ quadratic equations:$$x_{n}\sum_{a=1}^{\mathsf{N}}g_{a}(\zeta _{n}^{(1)})x_{a}+x_{n}f(\zeta _{n}^{(1)})=q_{n},\text{ \ \ \ }q_{n}=\frac{\mathrm{det}_{q}K_{+}(\xi _{n})\mathrm{det}_{q}\,\mathcal{U}_{-}(\xi _{n})}{\sinh (\eta +2\xi _{n})\sinh (\eta -2\xi _{n})},\text{ \ \ }\forall n\in \{1,...,\mathsf{N}\}, \label{Quadratic System}$$in $\mathsf{N}$ unknowns $\{x_{1},...,x_{\mathsf{N}}\}$. Inhomogeneous Baxter equation ============================== Here we show that the SOV characterization of the spectrum admits an equivalent formulation in terms of a second order functional difference equation of Baxter type:$$\tau (\lambda )Q(\lambda )=\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta )+F(\lambda ), \label{Inhom-Baxter-Eq}$$which contains a non-zero inhomogeneous term $F(\lambda )$ non-zero for generic integrable boundary conditions and the $Q$-functions are [*trigonometric polynomials*]{}. In this paper we will call $f(\lambda)$ a trigonometric polynomial of degree $\mathsf{M}$ if $e^{\mathsf{M}\lambda}\,f(\lambda)$ is a polynomial of $e^{2\lambda}$ of degree $\mathsf{M}$. Most trigonometric polynomials we will consider in the following sections will be even functions of $\lambda$ and will satisfy an additional condition $f({\lambda}+i\pi)=f({\lambda})$. It is easy to see in this situation that such functions can be written as polynomials of $\cosh 2\lambda$. Main functions in the functional equation ----------------------------------------- Let $Q(\lambda )$ be an even trigonometric polynomial of degree $2\mathsf{N}$. It can be written in the following form:$$\begin{aligned} Q(\lambda )& =\sum_{a=1}^{\mathsf{N}}\prod_{\substack{ b=1 \\ b\neq a}}^{\mathsf{N}}\frac{\cosh 2\lambda -\cosh 2\zeta _{b}^{(0)}}{\cosh 2\zeta _{a}^{(0)}-\cosh 2\zeta _{b}^{(0)}}Q(\zeta _{a}^{(0)})+2^{\mathsf{N}}\prod_{a=1}^{\mathsf{N}}\left( \cosh 2\lambda -\cosh 2\zeta _{a}^{(0)}\right) \label{Q-form1} \\ & =2^{\mathsf{N}}\prod_{a=1}^{\mathsf{N}}\left( \cosh 2\lambda -\cosh 2\lambda _{a}\right) , \label{Q-form2}\end{aligned}$$where from now on the $Q(\zeta _{a}^{(0)})$ are arbitrary complex numbers or similarly the $\lambda _{a}$ are arbitrary complex numbers. Then, introducing the function:$$Z_{Q}(\lambda )=\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta )$$we can prove the following Lemma Let $Q(\lambda )$ be any function of the form $\left( \ref{Q-form2}\right) $ then the associated function $Z_{Q}(\lambda )$ is an even trigonometric polynomial of degree $4\mathsf{N}+4$ of the following form:$$Z_{Q}(\lambda )=\sum_{a=0}^{2(\mathsf{N}+1)}z_{a}\cosh ^{a}2\lambda ,\text{ with }z_{2(\mathsf{N}+1)}=\frac{2\kappa _{+}\kappa _{-}\cosh (\alpha _{+}+\alpha _{-}-\beta _{+}+\beta _{-}-(\mathsf{N}+1)\eta )}{\sinh \zeta _{+}\sinh \zeta _{-}}.$$ The fact that the function $Z_{Q}(\lambda )$ is even in $\lambda $ is a trivial consequence of the fact that $Q(\lambda )$ is even; in fact, it holds:$$\begin{aligned} Z_{Q}(-\lambda ) &=&\mathbf{A}(-\lambda )Q(-\lambda -\eta )+\mathbf{A}(\lambda )Q(-\lambda +\eta ) \notag \\ &=&\mathbf{A}(-\lambda )Q(\lambda +\eta )+\mathbf{A}(\lambda )Q(\lambda -\eta )=Z_{Q}(\lambda ).\end{aligned}$$The fact that $Z_{Q}(\lambda )$ is indeed a trigonometric polynomial follows from its definition once we observe that $\lambda=0$ is not a singular point and the following identity holds:$$\lim_{\lambda \rightarrow 0}Z_{Q}(\lambda )=2g_{+}(0)g_{-}(0)a(0)a(-\eta )Q(0)\cosh \eta .$$Now the functional form of $Z_{Q}(\lambda )$ is a consequence of the following identities:$$Z_{Q}(\lambda +i\pi )=Z_{Q}(\lambda ),\text{ \ }\lim_{\lambda \rightarrow \pm \infty }\frac{Z_{Q}(\lambda )}{e^{\pm 4(\mathsf{N}+1)\lambda }}=\frac{\kappa _{+}\kappa _{-}\cosh (\alpha _{+}+\alpha _{-}-\beta _{+}+\beta _{-}-(\mathsf{N}+1)\eta )}{2^{(2\mathsf{N}+1)}\sinh \zeta _{+}\sinh \zeta _{-}},$$where the second identity follows from:$$\begin{aligned} \lim_{\lambda \rightarrow \pm \infty }e^{\mp (2\mathsf{N}+4)\lambda }\mathbf{A}(\lambda )& =2^{-2(\mathsf{N}+1)}\frac{\kappa _{+}\kappa _{-}\exp \pm (\alpha _{+}+\alpha _{-}-\beta _{+}+\beta _{-}+(\mathsf{N}-1)\eta )}{\sinh \zeta _{+}\sinh \zeta _{-}}, \\ \lim_{\lambda \rightarrow \pm \infty }e^{\mp 2\mathsf{N}\lambda }Q(\lambda )& =1.\end{aligned}$$ On the need of an inhomogeneous term in the functional equation --------------------------------------------------------------- Here, we would like to point out that it is simple to define the boundary conditions for which one can prove that the homogeneous version of the Baxter equation $\left( \ref{Inhom-Baxter-Eq}\right) $ does not admit trigonometric polynomial solutions for $\tau (\lambda )\in \Sigma _{\mathcal{T}}$. \[impossible\_hom\] Assume that the boundary parameters satisfy the following conditions:$$\kappa _{+}\neq 0,\kappa _{-}\neq 0,\text{ \ }Y^{(i,r)}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })\neq 0\text{ \ }\forall i\in \left\{ 0,1\right\} ,r\in \mathbb{Z} \label{Inhomogeneous-boundary conditions}$$where we have defined:$$Y^{(i,r)}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })\equiv \tau _{-}-\tau _{+}+\left( -1\right) ^{i}\left[ \left( \mathsf{N}-1-r\right) \eta +(\alpha _{-}+\alpha _{+}+\beta _{-}-\beta _{+})\right] ,$$then for any $\tau (\lambda )\in \Sigma _{\mathcal{T}}$  the homogeneous Baxter equation:$$\tau (\lambda )Q(\lambda )=\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta ),$$does not admit any (non identically zero) $Q(\lambda )$ of Laurent polynomial form in $e^{\lambda }$. If we consider the following function:$$Q(\lambda )=\sum_{a=-s}^{r}y_{a}e^{a\lambda },\text{ \ with }r,s\in \mathbb{N}$$we can clearly always chose the coefficients $y_{a}$ such that the r.h.s. of the homogeneous Baxter equation has no poles as required. However, it is enough to consider now the asymptotics:$$\begin{aligned} \lim_{\lambda \rightarrow +\infty }\frac{\left[ \mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta )\right] }{e^{(2\mathsf{N}+4+r)\lambda }}& =\frac{y_{r}\kappa _{+}\kappa _{-}\cosh (\alpha _{+}+\alpha _{-}-\beta _{+}+\beta _{-}+(\mathsf{N}-1-r)\eta )}{2^{2(\mathsf{N}+1)}\sinh \zeta _{+}\sinh \zeta _{-}} \\ \lim_{\lambda \rightarrow +\infty }e^{-(2\mathsf{N}+4+r)\lambda }\tau (\lambda )Q(\lambda )& =\frac{y_{r}\kappa _{+}\kappa _{-}\cosh (\tau _{+}-\tau _{-})}{2^{2(\mathsf{N}+1)}\sinh \zeta _{+}\sinh \zeta _{-}}\end{aligned}$$and use the conditions $\left( \ref{Inhomogeneous-boundary conditions}\right) $ to observe that for any $r\in \mathbb{Z}$ the asymptotic of the homogeneous Baxter equation cannot be satisfied which implies the validity of the lemma. SOV spectrum in terms of the inhomogeneous Baxter equation ---------------------------------------------------------- We introduce now the following function of the boundary parameters:$$F_{0}=\frac{2\kappa _{+}\kappa _{-}\left( \cosh (\tau _{+}-\tau _{-})-\cosh (\alpha _{+}+\alpha _{-}-\beta _{+}+\beta _{-}-(\mathsf{N}+1)\eta )\right) }{\sinh \zeta _{+}\sinh \zeta _{-}},$$and then the function:$$\begin{aligned} F(\lambda ) &=&2^{\mathsf{N}}\,F_{0}\,(\cosh ^{2}2\lambda -\cosh ^{2}\eta )a(\lambda )a(-\lambda )d(-\lambda )d(\lambda ) \\ &=&F_{0}\, (\cosh ^{2}2\lambda -\cosh ^{2}\eta )\prod_{b=1}^{\mathsf{N}}\prod_{i=0}^{1}(\cosh 2\lambda -\cosh 2\zeta _{b}^{(i)}).\end{aligned}$$We introduce also the set of functions $\Sigma _{\mathcal{Q}}$ such that $Q(\lambda)\in\Sigma _{\mathcal{Q}}$ if it has a form $\left( \ref{Q-form2}\right) $ and $$\tau(\lambda)=\frac{Z_{Q}(\lambda )+F(\lambda )}{Q(\lambda )}$$ is a trigonometric polynomial. We are now ready to prove the main theorem of this article: \[T-eigenvalue-F-eq\]Let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C}$ $^{\mathsf{N}}$ be generic [(\[xi-conditions\])]{} and let the boundary parameters $(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}\backslash N_{SOV}$ satisfy the following conditions:$$\kappa _{+}\neq 0,\kappa _{-}\neq 0,\text{ \ }Y^{(i,2r)}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })\neq 0\text{ \ }\forall i\in \left\{ 0,1\right\} ,r\in \left\{ 0,...,\mathsf{N}-1\right\} , \label{Inhom-cond-BaxEq}$$then $\tau (\lambda )\in \Sigma _{\mathcal{T}}$ if and only if $\exists !Q(\lambda )\in \Sigma _{\mathcal{Q}}$ such that $$\tau (\lambda )Q(\lambda )=Z_{Q}(\lambda )+F(\lambda ).$$ First we prove that if $\tau (\lambda )\in \Sigma _{\mathcal{T}}$  then there is a trigonometric polynomial $Q(\lambda )\in \Sigma _{\mathcal{Q}}$ satisfying the inhomogeneous functional Baxter equation: $$\tau (\lambda )Q(\lambda )=\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta )+F(\lambda ).$$To prove it we will show that there is the unique set of values $Q(\zeta _{b}^{(0)})$ such that $Q(\lambda)$ of the form (\[Q-form1\]) satisfies this equation. It is straightforward to verify that if $\tau (\lambda )\in \Sigma _{\mathcal{T}}$ and $Q(\lambda )$ has the form $\left( \ref{Q-form2}\right) $ then the left and right hand sides of the above equation are both even trigonometric polynomials of $\lambda $ and both can be written (using the asymptotic behavior) in the form:$$\frac{2\kappa _{+}\kappa _{-}\cosh (\tau _{+}-\tau _{-})\prod_{b=1}^{2\mathsf{N}+2}(\cosh 2\lambda -\cosh 2y_{b}^{\left( lhs/rhs\right) })}{\sinh \zeta _{+}\sinh \zeta _{-}}.$$Then to prove that we can introduce a $Q(\lambda )$ of the form $\left( \ref{Q-form2}\right) $ which satisfies the inhomogeneous Baxter equation $\left( \ref{Inhom-Baxter-Eq}\right) $ with $\tau (\lambda )\in \Sigma _{\mathcal{T}} $, we have only to prove that $\left( \ref{Inhom-Baxter-Eq}\right) $ is satisfied in $4\mathsf{N}+4$ different values of $\lambda $. As the [*r.h.s*]{} and [*l.h.s*]{} of $\left( \ref{Inhom-Baxter-Eq}\right) $ are even functions we need to check this identity only for $2N+2$ non-zero points $\mu_j$ such that $\mu_j\neq \pm \mu_k$. It is a simple exercise verify that the equation $\left( \ref{Inhom-Baxter-Eq}\right) $ is satisfied automatically for any $Q(\lambda )$ of the form $\left( \ref{Q-form2}\right) $ in the following two points, $ \eta /2$ and $ \eta /2+i\pi /2$:$$\tau (\eta /2)Q(\eta /2)=\mathbf{A}(\eta /2)Q(\eta /2-\eta )=\mathbf{A}(\eta /2)Q(\eta /2), \label{System-A}$$and:$$\tau (\eta /2+i\pi /2)Q(\eta /2+i\pi /2)=\mathbf{A}(\eta /2+i\pi /2)Q(i\pi /2-\eta /2) \\ =\mathbf{A}(\eta /2+i\pi /2)Q(\eta /2+i\pi /2) . \label{System-B}$$Indeed, these equations reduce to:$$\tau (\eta /2)=\mathbf{A}(\eta /2),\text{ \ \ \ }\tau (\eta /2+i\pi /2)=\mathbf{A}(\eta /2+i\pi /2)$$and so they are satisfied by definition for any $\tau (\lambda )\in \Sigma _{\mathcal{T}}$. Then we check the explicit form of the equation $\left( \ref{Inhom-Baxter-Eq}\right) $ in the $2\mathsf{N}$ points $ \zeta _{b}^{(0)}$ and $\zeta _{b}^{(1)}$:$$\tau (\zeta _{b}^{(0)})Q(\zeta _{b}^{(0)})=\mathbf{A}(-\zeta _{b}^{(0)})Q(\zeta _{b}^{(0)}+\eta )=\mathbf{A}(-\zeta _{b}^{(0)})Q(\zeta _{b}^{(1)}),$$and:$$\tau (\zeta _{b}^{(1)})Q(\zeta _{b}^{(1)})=\mathbf{A}(\zeta _{b}^{(1)})Q(\zeta _{b}^{(1)}-\eta )=\mathbf{A}(\zeta _{b}^{(1)})Q(\zeta _{b}^{(0)}).$$They are equivalent to the following system of equations:$$\begin{aligned} \frac{\mathbf{A}(\zeta _{b}^{(1)})}{\tau (\zeta _{b}^{(1)})}& =\frac{\tau (\zeta _{b}^{(0)})}{\mathbf{A}(-\zeta _{b}^{(0)})}\text{ \ \ \ \ \ }\forall b\in \{1,...,\mathsf{N}\} \label{System1} \\ \frac{Q(\zeta _{b}^{(0)})\tau (\zeta _{b}^{(0)})}{\mathbf{A}(-\zeta _{b}^{(0)})}& =\sum_{a=1}^{\mathsf{N}}\prod_{\substack{ c=1 \\ c\neq a}}^{\mathsf{N}}\frac{\cosh 2\zeta _{b}^{(1)}-\cosh 2\zeta _{c}^{(0)}}{\cosh 2\zeta _{a}^{(0)}-\cosh 2\zeta _{c}^{(0)}}Q(\zeta _{a}^{(0)})+2^{\mathsf{N}}\prod_{a=1}^{\mathsf{N}}\left( \cosh 2\zeta _{b}^{(1)}-\cosh 2\zeta _{a}^{(0)}\right) \label{System2}\end{aligned}$$Now using the following quantum determinant identity $$\frac{\det_{q}K_{+}(\lambda-\eta/2)\det_{q} \mathcal{U}_{-}(\lambda -\eta /2)}{\sinh (2\lambda +\eta )\sinh (2\lambda -\eta )}=\mathbf{A}(\lambda )\mathbf{A}(-\lambda +\eta ).\label{Tot-q-det-tt}$$ it is easy to see that the system of equations $\left( \ref{System1}\right) $ is certainly satisfied as $\tau (\lambda )\in \Sigma _{\mathcal{T}}$, once we recall the SOV characterization (\[Interpolation-Form-T\]) of $\Sigma _{\mathcal{T}}$. Indeed there is a set $\{x_1,\dots,x_n\}$ satisfying the equations (\[Quadratic System\]) and $\tau(\zeta _{b}^{(0)})=x_b$. So we are left with $\left( \ref{System2}\right) $ a linear system of $\mathsf{N}$ inhomogeneous equations with $\mathsf{N}$ unknowns $Q(\zeta _{a}^{(0)})$. Here, we prove that the matrix of this linear system$$c_{a b}\equiv \prod_{\substack{ c=1 \\ c\neq a}}^{\mathsf{N}}\frac{\cosh 2\zeta _{b}^{(1)}-\cosh 2\zeta _{c}^{(0)}}{\cosh 2\zeta _{a}^{(0)}-\cosh 2\zeta _{c}^{(0)}}-\delta _{a b}\frac{\tau (\zeta _{b}^{(0)})}{\mathbf{A}(-\zeta _{b}^{(0)})}\text{ \ \ \ \ \ }\forall a,b\in \{1,...,\mathsf{N}\}$$ has nonzero determinant for the given $\tau (\lambda )\in \Sigma _{\mathcal{T}}$. Indeed, let us suppose that for some $\tau (\lambda )\in \Sigma _{\mathcal{T}}$:$$\mathrm{det}_{\mathsf{N}}\left[ c_{a b}\right] =0. \label{det-coeff}$$Then there is at least one nontrivial solution $\{Q(\zeta _{1}^{(0)}),...,Q(\zeta _{\mathsf{N}}^{(0)})\}\neq \{0,...,0\}$ to the homogeneous system of equations:$$\frac{Q(\zeta _{b}^{(0)})\tau (\zeta _{b}^{(0)})}{\mathbf{A}(-\zeta _{b}^{(0)})}=\sum_{a=1}^{\mathsf{N}}\prod_{\substack{ c=1 \\ c\neq a}}^{\mathsf{N}}\frac{\cosh 2\zeta _{b}^{(1)}-\cosh 2\zeta _{c}^{(0)}}{\cosh 2\zeta _{a}^{(0)}-\cosh 2\zeta _{c}^{(0)}}Q(\zeta _{a}^{(0)}) \label{System2-homo}$$and hence we can define:$$Q_{\mathsf{M}}(\lambda )=\sum_{a=1}^{\mathsf{N}}\prod_{\substack{ b=1 \\ b\neq a}}^{\mathsf{N}}\frac{\cosh 2\lambda -\cosh 2\zeta _{b}^{(0)}}{\cosh 2\zeta _{a}^{(0)}-\cosh 2\zeta _{b}^{(0)}}Q(\zeta _{a}^{(0)})=\lambda _{\mathsf{M}+1}^{(\mathsf{M})}\prod_{b=1}^{\mathsf{M}}\left( \cosh 2\lambda -\cosh 2\lambda _{b}^{(\mathsf{M})}\right) .$$It is an even trigonometric polynomial of degree $2\mathsf{M}$ such that $0\leq \mathsf{M}\leq \mathsf{N}-1$ fixed by the solution $\{Q(\zeta _{1}^{(0)}),...,Q(\zeta _{\mathsf{N}}^{(0)})\}$. Now using the $Q_{\mathsf{M}}(\lambda )$ and $\tau (\lambda )\in \Sigma _{\mathcal{T}}$ we can define two functions:$$W_{1}(\lambda )=Q_{\mathsf{M}}(\lambda )\tau (\lambda )\text{ \, and \, }W_{2}(\lambda )=\mathbf{A}(\lambda )Q_{\mathsf{M}}(\lambda -\eta )+\mathbf{A}(-\lambda )Q_{\mathsf{M}}(\lambda +\eta )$$which are both even trigonometric polynomials of degree $2\mathsf{M}+2\mathsf{N}+4$. Then it is straightforward to observe that the systems of equations $\left( \ref{System1}\right) $ and $\left( \ref{System2-homo}\right) $ plus the conditions $\left( \ref{System-A}\right) $ and $\left( \ref{System-B}\right) $, which are also satisfied with the function $Q_{\mathsf{M}}(\lambda )$, imply that $W_{1}(\lambda )$ and $W_{2}(\lambda )$ coincide in $4\mathsf{N}+4 $ different values of $\lambda $ ($\pm \eta /2$, $\pm (\eta /2+i\pi /2)$, $\pm \zeta _{b}^{(0)}$ and $\pm \zeta _{b}^{(1)}$). It means that $W_{1}(\lambda )\equiv W_{2}(\lambda )$, as these are two polynomials of maximal degree $4\mathsf{N}+2$. So, we have shown that from the assumption $\exists \tau (\lambda )\in \Sigma _{\mathcal{T}}$ such that $\left( \ref{det-coeff}\right) $ holds it follows that $\tau (\lambda )$ and $Q_{\mathsf{M}}(\lambda )$ have to satisfy the following homogeneous Baxter equations: $$\tau (\lambda )Q_{\mathsf{M}}(\lambda )=\mathbf{A}(\lambda )Q_{\mathsf{M}}(\lambda -\eta )+\mathbf{A}(-\lambda )Q_{\mathsf{M}}(\lambda +\eta ). \label{Baxter-eq-homo}$$Now we can apply the Lemma \[impossible\_hom\] which implies that $Q_{\mathsf{M}}(\lambda )=0$ for any $\lambda$, which contradicts the hypothesis of the existence of a nontrivial solution to the homogeneous system [(\[System2-homo\])]{}. Hence, we have proven that $\mathrm{det}_{\mathsf{N}}\left[ c_{a b}\right] \neq 0.$ Therefore there is a unique solution $\{Q(\zeta _{1}^{(0)}),...,Q(\zeta _{\mathsf{N}}^{(0)})\}$ of the inhomogeneous system $\left( \ref{System2}\right) $ which defines one and only one $Q(\lambda )$ of the form $\left( \ref{Q-form1}\right) $ satisfying the functional inhomogeneous Baxter’s equation $\left( \ref{Inhom-Baxter-Eq}\right) $. We prove now that if $Q(\lambda )\in \Sigma _{\mathcal{Q}}$ then $\tau (\lambda )=\left( Z_{Q}(\lambda )+F(\lambda )\right) /Q(\lambda )\in \Sigma _{\mathcal{T}}$. By definition of the functions $Z_{Q}(\lambda ),$ $F(\lambda )$ and $Q(\lambda )$ the function $\tau (\lambda )$ has the desired form:$$\tau (\lambda )=f(\lambda )+\sum_{a=1}^{\mathsf{N}}g_{a}(\lambda )\tau (\zeta _{a}^{(0)}).$$To prove now that $\tau (\lambda )\in \Sigma _{\mathcal{T}}$ we have to write the inhomogeneous Baxter equation $\left( \ref{Inhom-Baxter-Eq}\right) $ in the $2\mathsf{N}$ points $ \zeta _{b}^{(0)}$ and $ \zeta _{b}^{(1)}$. Indeed, we have already proved that this reproduce the systems $\left( \ref{System1}\right) $ and $\left( \ref{System2}\right) $ and it is simple to observe that the system of equations $\left( \ref{System1}\right) $ just coincides with the inhomogeneous system of $\mathsf{N}$ quadratic equations:$$x_{n}\sum_{a=1}^{\mathsf{N}}g_{a}(\zeta _{n}^{(1)})x_{a}+x_{n}f(\zeta _{n}^{(1)})=q_{n},\text{ \ \ \ }\forall n\in \{1,...,\mathsf{N}\},$$once we define $x_{a}=\tau (\zeta _{a}^{(0)})$ for any $a\in \{1,...,\mathsf{N}\}$ and we write $\tau (\zeta _{n}^{(1)})$ in terms of the $x_{a}$. Thus we show that $$\tau (\lambda )=\left( Z_{Q}(\lambda )+F(\lambda )\right) /Q(\lambda )\in \Sigma _{\mathcal{T}},$$ completing the proof of the theorem. Completeness of the Bethe ansatz equations ------------------------------------------ In the previous section we have shown that to solve the transfer matrix spectral problem associated to the most general representations of the trigonometric 6-vertex reflection algebra we have just to classify the set of functions $Q(\lambda )$ of the form $\left( \ref{Q-form2}\right) $ for which $\left( Z_{Q}(\lambda )+F(\lambda )\right) /Q(\lambda )$ is a trigonometric polynomial; i.e. the set of functions $\Sigma _{\mathcal{Q}}$ completely fixes the set $\Sigma _{\mathcal{T}}$. We can show now that the previous characterization of the transfer matrix spectrum allows to prove that $\Sigma _{InBAE}\subset \mathbb{C}^{\mathsf{N}}$ the set of all the solutions of inhomogeneous Bethe equations $$\{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in \Sigma _{InBAE}$$ if $$\mathbf{A}(\lambda _{a})Q_{\mathbf{\lambda }}(\lambda _{a}-\eta )+\mathbf{A}(-\lambda _{a})Q_{\mathbf{\lambda }}(\lambda _{a}+\eta )=-F(\lambda _{a}),\text{ \ }\forall a\in \{1,...,\mathsf{N}\} , \label{I-BAE}$$ defines the complete set of transfer matrix eigenvalues. In particular, the following corollary follows: \[Theo-InBAE\] Let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C}$ $^{\mathsf{N}}$ be generic $\left( \ref{xi-conditions}\right) $ and let the boundary parameters $(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}\backslash N_{SOV}$ satisfy $\left( \ref{Inhom-cond-BaxEq}\right) $ then $\tau (\lambda )\in \Sigma _{\mathcal{T}}$  if and only if $\exists !\{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in \Sigma _{InBAE}$ such that:$$\tau (\lambda )=\frac{Z_{Q}(\lambda )+F(\lambda )}{Q(\lambda )}\text{ \ \ with \ \ }Q(\lambda )=2^{\mathsf{N}}\prod_{a=1}^{\mathsf{N}}\left( \cosh 2\lambda -\cosh 2\lambda _{a}\right) .$$Moreover, under the condition of normality defined in Proposition [normality]{}, the set $\Sigma _{InBAE}$ of all the solutions to the inhomogeneous system of Bethe equations $\left( \ref{I-BAE}\right) $ contains $2^{\mathsf{N}}$ elements. Homogeneous Baxter equation =========================== Boundary conditions annihilating the inhomogeneity of the Baxter equation ------------------------------------------------------------------------- The description presented in the previous sections can be applied to completely general integrable boundary terms including as a particular case the boundary conditions for which the inhomogeneous term in the functional Baxter equation vanishes. As these are still quite general boundary conditions it is interesting to point out how the previous general results explicitly look like in these cases. \[homogeneousBE\_N\] Let $(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}\backslash N_{SOV}$ satisfying the condition:$$\kappa _{+}\neq 0,\kappa _{-}\neq 0,\text{ \ }\exists i\in \left\{ 0,1\right\} \text{\ }:Y^{(i,2\mathsf{N})}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })=0 \label{ond-homo-boundary}$$and let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C} $ $^{\mathsf{N}}$ be generic [(\[xi-conditions\])]{}, then $\tau (\lambda )\in \Sigma _{\mathcal{T}}$ if and only if $ \exists !Q(\lambda )\in \Sigma _{\mathcal{Q}}$ such that $$\tau (\lambda )Q(\lambda )=\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta ).$$Or equivalently, $\tau (\lambda )\in \Sigma _{\mathcal{T}}$  if and only if $\exists !\{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in \Sigma _{BAE}$ such that:$$\tau (\lambda )=\frac{\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta )}{Q(\lambda )}\text{ \ \ with \ \ }Q(\lambda )=2^{\mathsf{N}}\prod_{a=1}^{\mathsf{N}}\left( \cosh 2\lambda -\cosh 2\lambda _{a}\right) .$$where:$$\Sigma _{BAE}=\left\{ \{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in \mathbb{C}^{\mathsf{N}}:\mathbf{A}(\lambda _{a})Q_{\mathbf{\lambda }}(\lambda _{a}-\eta )+\mathbf{A}(-\lambda _{a})Q_{\mathbf{\lambda }}(\lambda _{a}+\eta )=0,\text{ \ }\forall a\in \{1,...,\mathsf{N}\}\right\} . \label{BAE}$$Moreover, under the condition of normality defined in Proposition [normality]{}, the set $\Sigma _{BAE}$ of the solutions to the homogeneous system of Bethe ansatz type equations $\left( \ref{BAE}\right) $ contains $2^{\mathsf{N}}$ elements. This theorem is just a rewriting of the results presented in the Theorem \[T-eigenvalue-F-eq\] and Corollary \[Theo-InBAE\] for the case of vanishing inhomogeneous term. Indeed if the conditions $\left( \ref{BAE}\right) $ are satisfied then automatically the conditions of the main theorem $\left( \ref{Inhom-cond-BaxEq}\right) $ are satisfied too that implies that the map from the $\tau (\lambda )\in \Sigma _{\mathcal{T}}$  to the $\{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in \Sigma _{BAE}$ is indeed an isomorphism. More general boundary conditions compatibles with homogeneous Baxter equations ------------------------------------------------------------------------------ We address here the problem of describing the boundary conditions:$$\kappa _{+}\neq 0,\kappa _{-}\neq 0,\text{ \ }\exists i\in \left\{ 0,1\right\} ,\mathsf{M}\in \left\{ 0,...,\mathsf{N}-1\right\} :Y^{(i,2\mathsf{M})}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })=0, \label{Cond-homo-M}$$for which the conditions $\left( \ref{Inhom-cond-BaxEq}\right) $ are not satisfied and then the Theorem \[T-eigenvalue-F-eq\] cannot be directly applied. In these $2\mathsf{N}$ hyperplanes in the space of the boundary parameters we have just to modify this theorem to take into account that the Baxter equation associated to the choice of coefficient $\mathbf{A}(\lambda ) $ is indeed compatible with the homogeneous Baxter equation for a special choice of the polynomial $Q(\lambda )$. First we define the following functions$$Q_{\mathsf{M}}(\lambda )=2^{\mathsf{M}}\prod_{b=1}^{\mathsf{M}}\left( \cosh 2\lambda -\cosh 2\lambda _{b}^{(\mathsf{M})}\right) . \label{Q-form-M}$$We introduce also the set of polynomials $\Sigma _{\mathcal{Q}}^{\mathsf{M}}$ such that $Q_{\mathsf{M}}(\lambda )\in \Sigma _{\mathcal{Q}}^{\mathsf{M}}$ if $Q_{\mathsf{M}}(\lambda )$ has a form $\left( \ref{Q-form-M}\right)$ and $$\tau (\lambda )=\frac{\mathbf{A}(\lambda )Q_{\mathsf{M}}(\lambda -\eta )+\mathbf{A}(-\lambda )Q_{\mathsf{M}}(\lambda +\eta )}{Q_{\mathsf{M}}(\lambda )}$$is a trigonometric polynomial. Then we can define the corresponding set $\Sigma _{\mathcal{T}}^{\mathsf{M}}$$$\Sigma _{\mathcal{T}}^{\mathsf{M}}=\left\{ \tau (\lambda ):\tau (\lambda )\equiv \frac{\mathbf{A}(\lambda )Q_{\mathsf{M}}(\lambda -\eta )+\mathbf{A}(-\lambda )Q_{\mathsf{M}}(\lambda +\eta )}{Q_{\mathsf{M}}(\lambda )}\text{ \, if }Q_{\mathsf{M}}(\lambda )\in \Sigma _{\mathcal{Q}}^{\mathsf{M}}\right\} .$$It is simple to prove the validity of the following: \[mixed-condition\] Let the boundary conditions $\left( \ref{Cond-homo-M}\right) $ be satisfied, then $\Sigma _{\mathcal{T}}^{\mathsf{M}}\subset\Sigma _{\mathcal{T}}$ and moreover for any $\tau (\lambda )\in \Sigma _{\mathcal{T}}^{\mathsf{M}}$ there exists one and only one $Q_{\mathsf{M}}(\lambda )\in \Sigma _{\mathcal{Q}}^{\mathsf{M}}$ such that:$$\label{homogen-Bax-eq-M} \tau (\lambda )Q_{\mathsf{M}}(\lambda )=\mathbf{A}(\lambda )Q_{\mathsf{M}}(\lambda -\eta )+\mathbf{A}(-\lambda )Q_{\mathsf{M}}(\lambda +\eta ),$$and for any $\tau (\lambda )\in \Sigma _{\mathcal{T}}\backslash \Sigma _{\mathcal{T}}^{\mathsf{M}}$ there exists one and only one $Q(\lambda )\in \Sigma _{\mathcal{Q}}$ such that:$$\label{inhomogen-Bax-eq-M} \tau (\lambda )Q(\lambda )=\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta )+F(\lambda ).$$ The proof follows the one given for the main Theorem \[T-eigenvalue-F-eq\] we have just to observe that thanks to the boundary conditions $\left( \ref{Cond-homo-M}\right) $ the set $\Sigma _{\mathcal{T}}^{\mathsf{M}}$ is formed by transfer matrix eigenvalues as the Baxter equation implies that for any $\tau (\lambda )\in \Sigma _{\mathcal{T}}^{\mathsf{M}}$ the systems of equations $\left( \ref{System-A}\right) ,$ $\left( \ref{System-B}\right) $ and $\left( \ref{System1}\right) $ are satisfied and moreover that the asymptotics of the $\tau (\lambda )\in \Sigma _{\mathcal{T}}^{\mathsf{M}}$ is exactly that of the transfer matrix eigenvalues. Finally, it is interesting to remark that under the boundary conditions $\left( \ref{Inhom-cond-BaxEq}\right) $ the complete characterization of the spectrum of the transfer matrix is given in terms of the even polynomials $Q(\lambda )$ all of fixed degree $2\mathsf{N}$ and form $\left( \ref{Q-form2}\right) $ which are solutions of the inhomogeneous/homogeneous Baxter equation. However, in the cases when the boundary parameters satisfy the constraints $\left( \ref{Cond-homo-M}\right) $ for a given $\mathsf{M}\in \left\{ 0,...,\mathsf{N}-1\right\} $ a part of the transfer matrix spectrum can be defined by polynomials of smaller degree; i.e. the $Q_{\mathsf{M}}(\lambda )\in \Sigma _{\mathcal{Q}}^{\mathsf{M}}$ for the fixed $\mathsf{M}\in \left\{ 0,...,\mathsf{N}-1\right\} $. Discrete symmetries and equivalent Baxter equations {#sect-descretesym} =================================================== It is important to point out that we have some large amount of freedom in the choice of the functional reformulation of the SOV characterization of the transfer matrix spectrum. We have reduced it looking for trigonometric polynomial solutions $Q(\lambda )$ of the second order difference equations with coefficients $\mathbf{A}(\lambda )$ which are rational trigonometric functions. It makes the finite difference terms $\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta )$ in the functional equation a trigonometric polynomial. Indeed, this assumption reduces the possibility to use the following gauge transformations of the coefficients allowed instead by the SOV characterization:$$\mathbf{A}_{\alpha }(\lambda )=\alpha (\lambda )\mathbf{A}(\lambda ),\text{ \ }\mathbf{D}_{\alpha }(\lambda )=\frac{\mathbf{A}(-\lambda )}{\alpha (\lambda +\eta )}.$$In the following we discuss simple transformations that do not modify the functional form of the coefficients allowing equivalent reformulations of the SOV spectrum by Baxter equations. Discrete symmetries of the transfer matrix spectrum --------------------------------------------------- It is not difficult to see that the spectrum (eigenvalues) of the transfer matrix presents the following invariance: \[Lem-invariance\]We denote explicitly the dependence from the boundary parameters in the set of boundary parameters $\Sigma _{\mathcal{T}}^{(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})}$ of the eigenvalue functions of the transfer matrix $\mathcal{T}(\lambda )$, then this set is invariant under the following $Z_{2}^{\otimes 3}$ transformations of the boundary parameters:$$\begin{aligned} &\Sigma _{\mathcal{T}}^{(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})}\equiv \Sigma _{\mathcal{T}}^{(\epsilon _{\tau }\tau _{+},\epsilon _{\alpha }\alpha _{+},\epsilon _{\beta }\beta _{+},\epsilon _{\tau }\tau _{-},\epsilon _{\alpha }\alpha _{-},\epsilon _{\beta }\beta _{-})}\ \\ & \forall (\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })\in \{-1,1\}\times \{-1,1\}\times \{-1,1\}.\nonumber\end{aligned}$$ To prove this statement it is enough to look at the SOV characterization which defines completely the transfer matrix spectrum, i.e. the set $\Sigma _{\mathcal{T}}$, and to prove that it is invariant under the above considered $Z_{2}^{\otimes 3}$ transformations of the boundary parameters. We have first to remark that the central values $\left( \ref{Central-asymp}\right) $-$\left( \ref{Central-2}\right) $ of the transfer matrix $\mathcal{T}(\lambda )$ are invariant under these discrete transformations and then the function $f(\lambda )$, defined in $\left( \ref{f-function}\right) $, is invariant too and the same is true for the form $\left( \ref{Interpolation-Form-T}\right) $ of the interpolation polynomial describing the elements of $\Sigma _{\mathcal{T}}$. Then the invariance of the SOV characterization $\left( \ref{Quadratic System}\right) $ follows from the invariance of the quantum determinant$$\begin{aligned} \mathrm{det}_{q}K_{+}(\lambda )\mathrm{det}_{q}\,\mathcal{U}_{-}(\lambda ) &=&\sinh (2\eta -2\lambda )\sinh (2\lambda +2\eta )g_{+}(\lambda +\eta /2)g_{+}(-\lambda +\eta /2)g_{-}(\lambda +\eta /2) \notag \\ &&\times g_{-}(-\lambda +\eta /2)a(\lambda +\eta /2)d(\lambda -\eta /2)a(-\lambda +\eta /2)d(-\lambda -\eta /2)\end{aligned}$$ under these discrete transformations. It is important to underline that the above $Z_{2}^{\otimes 3}$ transformations of the boundary parameters do indeed change the transfer matrix $\mathcal{T}(\lambda )$ and the Hamiltonian and so this invariance is equivalent to the statement that these different transfer matrices are all isospectral. In particular, it is simple to find the similarity matrices implementing the following $Z_{2}$ transformations of the boundary parameters:$$\begin{aligned} \mathcal{T}(\lambda |-\tau _{+},-\zeta _{+},\kappa _{+},-\tau _{-},-\zeta _{-},\kappa _{-}) &=&\Gamma _{y}\mathcal{T}(\lambda |\tau _{+},\zeta _{+},\kappa _{+},\tau _{-},\zeta _{-},\kappa _{-})\Gamma _{y},\text{ \ \ \ }\Gamma _{y}\equiv \otimes _{n=1}^{\mathsf{N}}\sigma _{n}^{y}, \\ \mathcal{T}(\lambda |\tau _{+},\zeta _{+},-\kappa _{+},\tau _{-},\zeta _{-},-\kappa _{-}) &=&\Gamma _{z}\mathcal{T}(\lambda |\tau _{+},\zeta _{+},\kappa _{+},\tau _{-},\zeta _{-},\kappa _{-})\Gamma _{z},\text{ \ \ \ }\Gamma _{z}\equiv \otimes _{n=1}^{\mathsf{N}}\sigma _{n}^{z}.\end{aligned}$$ Equivalent Baxter equations and the SOV spectrum ------------------------------------------------ The invariance of the spectrum $\Sigma _{\mathcal{T}}$ under these $Z_{2}^{\otimes 3}$ transformations of the boundary parameters can be used to define equivalent Baxter equation reformulation of $\Sigma _{\mathcal{T}}$. More precisely, let us introduce the following functions $\mathbf{A}_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda )$ and $F_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda )$ obtained respectively by implementing the $Z_{2}^{\otimes 3}$ transformations:$$(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\rightarrow (\epsilon _{\tau }\tau _{+},\epsilon _{\alpha }\alpha _{+},\epsilon _{\beta }\beta _{+},\epsilon _{\tau }\tau _{-},\epsilon _{\alpha }\alpha _{-},\epsilon _{\beta }\beta _{-}),$$then the following characterizations hold for any fixed $(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })\in \{-1,1\}\times \{-1,1\}\times \{-1,1\}$: \[T-eigenvalue-F-eq-gen\]Let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C}$ $^{\mathsf{N}}$ be generic [(\[xi-conditions\])]{} and let the boundary parameters $(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}\backslash N_{SOV}$ satisfy the following conditions:$$\kappa _{+}\neq 0,\kappa _{-}\neq 0,\text{ \ }Y^{(i,2r)}(\epsilon _{\tau }\tau _{\pm },\epsilon _{\alpha }\alpha _{\pm },\epsilon _{\beta }\beta _{\pm })\neq 0\text{ \ }\forall i\in \left\{ 0,1\right\} ,r\in \left\{ 0,...,\mathsf{N}-1\right\} , \label{Inhom-cond-BaxEq-gen}$$then $\tau (\lambda )\in \Sigma _{\mathcal{T}}$ if and only if $\exists !Q(\lambda )\in \Sigma _{\mathcal{Q}}$ such that $$\tau (\lambda )Q(\lambda )=Z_{Q,(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda )+F_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda ),$$where:$$Z_{Q,(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda )=\mathbf{A}_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda )(\lambda )Q(\lambda -\eta )+\mathbf{A}_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(-\lambda )Q(\lambda +\eta ).$$ The proof follows step by step the one given for the main Theorem \[T-eigenvalue-F-eq\]. General validity of the inhomogeneous Baxter equations ------------------------------------------------------ The previous reformulations of the spectrum in terms of different inhomogeneous Baxter equations and the observation that the conditions under which the Theorem does not apply are related to the choice of the $(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })\in \{-1,1\}\times \{-1,1\}\times \{-1,1\}$ allow us to prove that unless the boundary parameters are lying on a finite lattice of step $\eta $ we can always use an inhomogeneous Baxter equations to completely characterize the spectrum of the transfer matrix. More precisely, let us introduce the following hyperplanes in the space of the boundary parameters:$$M\equiv \left\{ \begin{array}{l} (\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}:\exists (r_{+,+},r_{-,+},r_{-,-})\in \{0,...,\mathsf{N}-1\} \\ \text{ such that: \ \ \ }\left\{ \begin{array}{l} r_{+,+}+r_{-,-}-r_{-,+}\in \{0,...,\mathsf{N}-1\} \\ \alpha _{+}+\alpha _{-}=(r_{-,+}-r_{+,+})\eta \\ \beta _{-}-\beta _{+}=(r_{-,-}-r_{-,+})\eta \\ \tau _{-}-\tau _{+}=(\mathsf{N}-1+r_{-,-}-3r_{+,+})\eta\end{array}\right.\end{array}\right\} \label{Def-M}$$then the following theorem holds: Let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C}$ $^{\mathsf{N}}$ satisfy the conditions [(\[xi-conditions\])]{} and let $(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}\backslash \left( M\cup N_{SOV}\right) $ then we can always find a $(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })\in \{-1,1\}\times \{-1,1\}\times \{-1,1\}$ such that $\tau (\lambda )\in\Sigma _{\mathcal{T}}$ if and only if $\exists !Q(\lambda )\in \Sigma _{\mathcal{Q}}$ such that $$\tau (\lambda )Q(\lambda )=Z_{Q,(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda )+F_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda ).$$ The Theorem \[T-eigenvalue-F-eq-gen\] does not apply if $\exists i\in \left\{ 0,1\right\} $ and $\exists r\in \left\{ 0,...,\mathsf{N}-1\right\} $ such that the following system of conditions on the boundary parameters are satisfied:$$Y^{(i,2r)}(\epsilon _{\tau }\tau _{\pm },\epsilon _{\alpha }\alpha _{\pm },\epsilon _{\beta }\beta _{\pm })=0\text{ \ }\forall (\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })\in \{-1,1\}^{\otimes 3} \label{Cond-general-homo}$$ then by simple computations it is possible to observe that the set $M$ defined in $\left( \ref{Def-M}\right) $ indeed coincides with the following set:$$\left\{ (\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}:\text{ }\exists i\in \left\{ 0,1\right\} ,r\in \left\{ 0,...,\mathsf{N}-1\right\} \text{ such that }\left( \ref{Cond-general-homo}\right) \text{ is satisfied}\right\} ,$$from which the theorem clearly follows. Homogeneous Baxter equation --------------------------- The discrete symmetries of the transfer matrix allow also to define the general conditions on the boundary parameters for which the spectrum can be characterized by a homogeneous Baxter equation. In particular the following corollary holds: Let $(\tau _{+},\alpha _{+},\beta _{+},\tau _{-},\alpha _{-},\beta _{-})\in \mathbb{C}^{6}\backslash N_{SOV}$ satisfy the condition:$$\begin{aligned} &\kappa _{+}\neq 0,\kappa _{-}\neq 0,\nonumber\\ &\exists i\in \left\{ 0,1\right\} ,\text{\ }\exists (\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })\in \{-1,1\}\times \{-1,1\}\times \{-1,1\}:Y^{(i,2\mathsf{N})}(\epsilon _{\tau }\tau _{\pm },,\epsilon _{\alpha }\alpha _{\pm },\epsilon _{\beta }\beta _{\pm })=0\end{aligned}$$and let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C} $ $^{\mathsf{N}}$ be generic [(\[xi-conditions\])]{}, then $\tau (\lambda )\in \Sigma _{\mathcal{T}}$ if and only if $\exists !Q(\lambda )\in \Sigma _{\mathcal{Q}}$ such that $$\tau (\lambda )Q(\lambda )=\mathbf{A}_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda )(\lambda )Q(\lambda -\eta )+\mathbf{A}_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(-\lambda )Q(\lambda +\eta ).$$Or equivalently we can define the set of all the solutions of the Bethe equations $$\Sigma _{BAE}=\left\{ \{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in \mathbb{C}^{\mathsf{N}}:\mathbf{A}(\lambda _{a})Q_{\mathbf{\lambda }}(\lambda _{a}-\eta )+\mathbf{A}(-\lambda _{a})Q_{\mathbf{\lambda }}(\lambda _{a}+\eta )=0,\text{ \ }\forall a\in \{1,...,\mathsf{N}\}\right\} .$$Then $\tau (\lambda )\in \Sigma _{\mathcal{T}}$  if and only if $\exists !\{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in \Sigma _{BAE}$ such that:$$\tau (\lambda )=\frac{\mathbf{A}_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(\lambda )Q(\lambda -\eta )+\mathbf{A}_{(\epsilon _{\tau },\epsilon _{\alpha },\epsilon _{\beta })}(-\lambda )Q(\lambda +\eta )}{Q(\lambda )},$$with $$Q(\lambda )=2^{\mathsf{N}}\prod_{a=1}^{\mathsf{N}}\left( \cosh 2\lambda -\cosh 2\lambda _{a}\right).$$ Moreover, under the condition of normality defined in Proposition [normality]{}, the set $\Sigma _{BAE}$ of the solutions to the homogeneous system of Bethe ansatz type equations $\left( \ref{BAE}\right) $ contains $2^{\mathsf{N}}$ elements. XXX chain by SOV and Baxter equation ==================================== The construction of the SOV characterization can be naturally applied in the case of the rational 6-vertex $R$-matrix, which in the homogeneous limit reproduces the XXX open quantum spin-1/2 chain with general integrable boundary conditions[^4]. Let us define:$$R_{12}(\lambda )=\left( \begin{array}{cccc} \lambda +\eta & 0 & 0 & 0 \\ 0 & \lambda & \eta & 0 \\ 0 & \eta & \lambda & 0 \\ 0 & 0 & 0 & \lambda +\eta\end{array}\right) \in \text{End}(\mathcal{H}_{1}\otimes \mathcal{H}_{2}).$$Due to the $SU(2)$ invariance of the bulk monodromy matrix the boundary matrices defining the most general integrable boundary conditions can be always recasted in the following form:$$K_{-}(\lambda ;p)=\left( \begin{array}{cc} \lambda -\eta /2+p & 0 \\ 0 & p-\lambda +\eta /2\end{array}\right) ,\text{ \ \ \ }K_{+}(\lambda ;q,\xi )=\left( \begin{array}{cc} \lambda +\eta /2+q & \xi (\lambda +\eta /2) \\ \xi (\lambda +\eta /2) & q-(\lambda +\eta /2)\end{array}\right) ,$$leaving only three arbitrary complex parameters here denoted with $\xi ,$ $p$ and $q$. Then the one parameter family of commuting transfer matrices:$$\mathcal{T}(\lambda )=\text{tr}_{0}\{K_{+}(\lambda )\,M(\lambda )\,K_{-}(\lambda )\hat{M}(\lambda )\}\in \text{\thinspace End}(\mathcal{H}),$$in the homogeneous limit leads to the following Hamiltonian:$$H=\sum_{n=1}^{\mathsf{N}}\left( \sigma _{n}^{x}\sigma _{n+1}^{x}+\sigma _{n}^{y}\sigma _{n+1}^{y}+\sigma _{n}^{z}\sigma _{n+1}^{z}\right) +\frac{\sigma _{\mathsf{N}}^{z}}{p}+\frac{\sigma _{1}^{z}+\xi \sigma _{1}^{x}}{q}.$$It is simple to show that the following identities hold:$$\mathrm{det}_{q}K_{+}(\lambda )\mathrm{det}_{q}\,\mathcal{U}_{-}(\lambda )=4(\lambda ^{2}-\eta ^{2})(\lambda ^{2}-p^{2})((1+\xi ^{2})\lambda ^{2}-q^{2})\prod_{b=1}^{\mathsf{N}}(\lambda ^{2}-(\xi _{n}+\eta )^{2})(\lambda ^{2}-(\xi _{n}-\eta )^{2}).$$We define:$$\mathbf{A}(\lambda )=(-1)^{\mathsf{N}}\frac{2\lambda +\eta }{2\lambda }(\lambda -\eta /2+p)(\sqrt{(1+\xi ^{2})}(\lambda -\eta /2)+q)\prod_{b=1}^{\mathsf{N}}(\lambda -\zeta _{b}^{(0)})(\lambda +\zeta _{b}^{(1)}),$$then it is easy to derive the following quantum determinant identity:$$\frac{\mathrm{det}_{q}K_{+}(\lambda )\mathrm{det}_{q}\,\mathcal{U}_{-}(\lambda )}{(4\lambda ^{2}-\eta ^{2})}=\mathbf{A}(\lambda +\eta /2)\mathbf{A}(-\lambda +\eta /2).$$ From the form of the boundary matrices it is clear that for the rational 6-vertex case one can directly derive the SOV representations using the method developed in [@Nic12b] without any need to introduce Baxter’s gauge transformations. Some results in this case also appeared in [@FraSW08; @FraGSW11] based on a functional version of the separation of variables of Sklyanin, a method which allows to define the eigenvalues and wave-functions but which does not allow to construct in the original Hilbert space of the quantum chain the transfer matrix eigenstates. The separation of variable description in this rational 6-vertex case reads: \[C:T-eigenstates- copy(1)\]Let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C}$ $^{\mathsf{N}}$ be generic: $$\xi _{a}\neq \pm\xi _{b}+r\eta \text{ \ }\forall a\neq b\in \{1,...,\mathsf{N}\}\,\,\text{and\thinspace \thinspace }r\in \{-1,0,1\}, \label{xi-conditions-xxx}$$ then $\mathcal{T}(\lambda )$ has simple spectrum and $\Sigma _{\mathcal{T}}$ is characterized by:$$\Sigma _{\mathcal{T}}=\left\{ \tau (\lambda ):\tau (\lambda )=f(\lambda )+\sum_{a=1}^{\mathsf{N}}g_{a}(\lambda )x_{a},\text{ \ \ }\forall \{x_{1},...,x_{\mathsf{N}}\}\in \Sigma _{T}\right\} ,$$where:$$g_{a}(\lambda )=\frac{4\lambda ^{2}-\eta ^{2}}{4{\zeta _{a}^{(0)}} ^{2}-\eta ^{2}}\,\prod_{\substack{ b=1 \\ b\neq a}}^{\mathsf{N}}\frac{\lambda ^{2}-{ \zeta _{b}^{(0)}} ^{2}}{{ \zeta _{a}^{(0)}} ^{2}-{ \zeta _{b}^{(0)}} ^{2}}\quad \text{ \ for }a\in \{1,...,\mathsf{N}\},$$and$$f(\lambda )=\prod_{b=1}^{\mathsf{N}}\frac{\lambda ^{2}-{ \zeta _{b}^{(0)}}^{2}}{{ \zeta _{a}^{(0)}} ^{2}-{\zeta _{b}^{(0)}} ^{2}}\mathbf{A}(\eta /2)+2\left( 4\lambda ^{2}-\eta ^{2}\right) \,\prod_{b=1}^{\mathsf{N}}\lambda ^{2}-{ \zeta _{b}^{(0)}} ^{2},$$$\Sigma _{T}$ is the set of solutions to the following inhomogeneous system of $\mathsf{N}$ quadratic equations:$$x_{n}\sum_{a=1}^{\mathsf{N}}g_{a}(\zeta _{n}^{(1)})x_{a}+x_{n}f(\zeta _{n}^{(1)})=q_{n},\text{ \ \ \ }q_{n}=\frac{\mathrm{det}_{q}K_{+}(\xi _{n})\mathrm{det}_{q}\,\mathcal{U}_{-}(\xi _{n})}{\eta -4\xi _{n}^{2}},\text{ \ \ }\forall n\in \{1,...,\mathsf{N}\},$$in $\mathsf{N}$ unknowns $\{x_{1},...,x_{\mathsf{N}}\}$. We are now ready to present the following equivalent characterization of the transfer matrix spectrum: \[T-eigenvalue-F-eq copy(1)\]Let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C}$ $^{\mathsf{N}}$ be generic [(\[xi-conditions-xxx\])]{}, then for $\xi \neq 0$ the set of transfer matrix eigenvalue functions $\Sigma _{\mathcal{T}}$ is characterized by:$$\tau (\lambda )\in \Sigma _{\mathcal{T}}\text{ \ if and only if }\exists !Q(\lambda )=\prod_{b=1}^{\mathsf{N}}\left( \lambda ^{2}-\lambda _{b}^{2}\right) \text{ such that }\tau (\lambda )Q(\lambda )=Z_{Q}(\lambda )+F(\lambda ),$$with$$F(\lambda )=2(1-\sqrt{(1+\xi ^{2})})\left( 4\lambda ^{2}-\eta ^{2}\right) \,\prod_{b=1}^{\mathsf{N}}\prod_{i=0}^{1}\left( \lambda ^{2}- {�\zeta _{b}^{(i)}} ^{2}\right) .$$ The proof presented in Theorem \[T-eigenvalue-F-eq\] applies with small modifications also to present rational case. The previous characterization of the transfer matrix spectrum allows to prove that the set $\Sigma _{InBAE}\subset\mathbb{C}^\mathsf{N}$ of all the solutions of the Bethe equations$$\{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in\Sigma _{InBAE}$$ if $$\mathbf{A}(\lambda _{a})Q_{\mathbf{\lambda }}(\lambda _{a}-\eta )+\mathbf{A}(-\lambda _{a})Q_{\mathbf{\lambda }}(\lambda _{a}+\eta )=-F(\lambda _{a}),\text{ \ }\forall a\in \{1,...,\mathsf{N}\} , \label{I-BAE-XXX}$$ define the complete set of transfer matrix eigenvalues. In particular, the following corollary can be proved: \[Theo-InBAE-XXX\]Let the inhomogeneities $\{\xi _{1},...,\xi _{\mathsf{N}}\}\in \mathbb{C}$ $^{\mathsf{N}}$ satisfy the following conditions [(\[xi-conditions\])]{}, then $\mathcal{T}(\lambda )$ has simple spectrum and for $\xi \neq 0$ then $\tau (\lambda )\in \Sigma _{\mathcal{T}}$  if and only if $\exists !\{\lambda _{1},...,\lambda _{\mathsf{N}}\}\in \Sigma _{InBAE}$ such that:$$\tau (\lambda )=\frac{Z_{Q}(\lambda )+F(\lambda )}{Q(\lambda )}\text{ \ \ with \ \ }Q(\lambda )=\prod_{b=1}^{\mathsf{N}}\left( \lambda ^{2}-\lambda _{b}^{2}\right) .$$ Homogeneous chains and existing numerical analysis ================================================== It is important to stress that the spectrum construction together with the corresponding statements of completeness presented in this paper strictly work for the most general spin 1/2 representations of the 6-vertex reflection algebra only for generic inhomogeneous chains. However, it is worth mentioning that the transfer matrix as well as the coefficients and the inhomogeneous term in our functional equation characterization of the SOV spectrum are analytic functions of the inhomogeneities $\{\xi _{j}\}$ so we can take without any problem the homogeneous limit ($\xi _{a}\rightarrow 0$ $\forall a\in \{1,...,\mathsf{N}\}$) in the functional equations. The main problem to be addressed then is the completeness of the description by this functional equations. Some first understanding of this central question can be derived looking at the numerical analysis [@Nep-R-2003; @Nep-R-2003add] of the completeness of Bethe Ansatz equations when the boundary constraints are satisfied and for the open XXX chain with general boundary terms [@Nep-2013]. Comparison with numerical results for the XXZ chain --------------------------------------------------- The numerical checks of the completeness of Bethe Ansatz equations for the open XXZ quantum spin 1/2 chains were first done in [@Nep-R-2003] for the chains with non-diagonal boundaries satisfying boundary constraints: $$\kappa _{+}\neq 0,\kappa _{-}\neq 0,\text{ \ }\exists i\in \left\{ 0,1\right\} ,\mathsf{M}\in \mathbb{N}\text{\ }:Y^{(i,2\mathsf{M})}(\tau _{\pm },\alpha _{\pm },\beta _{\pm })=0.$$Indeed, under these conditions some generalizations of algebraic Bethe Ansatz can be used and so the corresponding Bethe equations can be defined. In particular, the Nepomechie-Ravanini’s numerical results reported in [Nep-R-2003,Nep-R-2003add]{} suggest that the Bethe ansatz equations $\left( \ref{BAE}\right) $ in the homogeneous limit for the roots of the $Q$ function:$$Q(\lambda )=2^{\mathsf{M}}\prod_{a=1}^{\mathsf{M}}\left( \cosh 2\lambda -\cosh 2\lambda _{a}\right) ,$$with the degree $\mathsf{M}$ obtained from the boundary constraint - for $\mathsf{M}=\mathsf{N}$ they define the complete transfer matrix spectrum. - for $\mathsf{M}<\mathsf{N}$ the complete spectrum of the transfer matrix contains two parts described by different Baxter equations. The first one has trigonometric polynomial solutions of degree $2\mathsf{M}$ the second one has a trigonometric polynomial solutions of degree $2\mathsf{N}-2-2\mathsf{M}$. - for $\mathsf{M}>\mathsf{N}$ the complete spectrum of the transfer matrix spectrum plus $\tau (\lambda )$ functions which do not belong to the spectrum of the transfer matrix. These results seem to be compatible with our characterization for the inhomogeneous chains. Indeed, the case $\mathsf{M}=\mathsf{N}$ coincides with the case in which our Baxter functional equation becomes homogeneous. Theorem \[homogeneousBE\_N\] states that in this case for generic inhomogeneities the Bethe ansatz is complete so we can expect (from the numerical analysis) that completeness will survive in the homogeneous limit. In the case $\mathsf{M}<$, our description of the spectrum by Lemma [mixed-condition]{} separates the spectrum in two parts. A first part of the spectrum is described by trigonometric polynomial solutions of degree $2\mathsf{M}$ to the homogeneous Baxter equation [(\[homogen-Bax-eq-M\])]{} and a second part is instead described by trigonometric polynomial solutions of degree 2 of the inhomogeneous Baxter equation [(\[inhomogen-Bax-eq-M\])]{}. However, by implementing the following discrete symmetry transformations $\alpha _{\pm }\rightarrow -\alpha _{\pm }$, $\beta _{\pm }\rightarrow -\beta _{\pm }$, $\tau _{\pm }\rightarrow -\tau _{\pm }$ and applying the same Lemma [mixed-condition]{} w.r.t. the Baxter equations with coefficients $\mathbf{A}_{(-,-,-)}(\lambda )$ we get an equivalent description of the spectrum separated in two parts. One part of the spectrum is described in terms of the solutions of the transformed homogeneous Baxter equation which should be trigonometric polynomials of degree $2\mathsf{M}^{\prime }$, with $\mathsf{M}^{\prime }=\mathsf{N}-1-\mathsf{M}$ and the second part by the inhomogeneous Baxter equation. The comparison with the numerical results then suggests that, at least in the limit of homogeneous chains, the part of the spectrum generated by the trigonometric polynomial solutions of degree 2 of the inhomogeneous Baxter equation [(\[inhomogen-Bax-eq-M\])]{} coincides with the part generated by the trigonometric polynomial solutions of degree $2\mathsf{M}^{\prime }$ of the transformed homogeneous Baxter equation. Finally, in the case $\mathsf{M}>\mathsf{N}$ we have a complete characterization of the spectrum given by an inhomogeneous Baxter functional equation however nothing prevent to consider solutions to the homogeneous Baxter equation once we take the appropriate $Q$-function with $\mathsf{M}>\mathsf{N}$ Bethe roots. The numerical results however seem to suggest that considering the homogeneous Baxter equations is not the proper thing to do in the homogeneous limit. The previous analysis seems to support the idea that in the limit of homogeneous chain our complete characterization still describe the complete spectrum of the homogeneous transfer matrix. Comparison with numerical results for the XXX chain --------------------------------------------------- In the case of the open spin 1/2 XXX chain an ansatz based on two $Q$-functions and an inhomogeneous Baxter functional equation has been first introduced in [@CaoYSW13-2], the completeness of the spectrum obtained by that ansatz has been later verified numerically for small chains [@CaoJYW2013]. Using these results Nepomechie has introduced a simpler ansatz and developed some further numerical analysis in [@Nep-2013] confirming once again that the ansatz defines the complete spectrum for small chains. Here, we would like to point out that our complete description of the transfer matrix spectrum in terms of a inhomogeneous Baxter functional equation obtained for the inhomogeneous chains has the following well defined homogeneous limit: $$\tau (\lambda )Q(\lambda )=\mathbf{A}(\lambda )Q(\lambda -\eta )+\mathbf{A}(-\lambda )Q(\lambda +\eta )+F(\lambda )$$where:$$\begin{aligned} F(\lambda ) &=&8(1-\sqrt{(1+\xi ^{2})})\left( \lambda ^{2}-\left( \eta /2\right) ^{2}\right) ^{2\mathsf{N}+1}, \\ \mathbf{A}(\lambda ) &=&(-1)^{\mathsf{N}}\frac{2\lambda +\eta }{2\lambda }\left(\vphantom{\sqrt{(1+\xi ^{2})}}\lambda -\eta /2+p\right)\left(\sqrt{(1+\xi ^{2})}(\lambda -\eta /2)+q\right)\left(\lambda ^{2}-\left( \eta /2\right) ^{2}\right)^{\mathsf{N}}.\end{aligned}$$Taking into account the shift in our definition of the monodromy matrix which insures that the transfer matrix is an even function of the spectral parameter, the limit of our inhomogeneous Baxter functional equation coincides with the ansatz proposed by Nepomechie in [@Nep-2013]. Then the numerical evidences of completeness derived by Nepomechie in [@Nep-2013] suggest that the exact and complete characterization that we get for the inhomogeneous chain is still valid and complete in the homogeneous limit. Conclusion and outlook {#conclusion-and-outlook .unnumbered} ====================== In this paper we have shown that the transfer matrix spectrum associated to the most general spin-1/2 representations of the 6-vertex reflection algebras (rational and trigonometric), on general inhomogeneous chains is completely characterized in terms of a second order difference functional equations of Baxter $T$-$Q$ type with an inhomogeneous term depending only on the inhomogeneities of the chain and the boundary parameters. This functional $T$-$Q$ equation has been shown to be equivalent to the SOV complete characterization of the spectrum when the $Q$-functions belong to a well defined set of polynomials. The polynomial character of the $Q$-function is a central feature of our characterization which allows to introduce an equivalent finite system of generalized Bethe ansatz equations. Moreover, we have explicitly proven that our functional characterization holds for all the values of the boundary parameters for which SOV works, clearly identifying the only 3-dimensional hyperplanes in the 6-dimensional space of the boundary parameters where our description cannot be applied. We have also clearly identified the 5-dimensional hyperplanes in the space of the boundary parameters where the spectrum (or a part of the spectrum) can be characterized in terms of a homogeneous $T$-$Q$ equation and the polynomial character of the $Q$-functions is then equivalent to a standard system of Bethe equations. Completeness of this description is a built in feature due to the equivalence to the SOV characterization. The equivalence between our functional $T$-$Q$ equation and the SOV characterization holds for generic values of the $\xi _{a}$ in the $\mathsf{N}$-dimensional space of the inhomogeneity parameters however there exist hyperplanes for which the conditions [(\[xi-conditions\])]{} are not satisfied and so a direct application of the SOV approach is not possible (at least for the separate variables described in [@Fald-KN13]) and the limit of homogeneous chains ($\xi _{a}\rightarrow 0$ $\forall a\in \{1,...,\mathsf{N}\}$) clearly belong to these hyperplanes. From the analyticity of the transfer matrix eigenvalues, of the coefficients of the functional $T$-$Q$ equation and of the inhomogeneous term in it w.r.t. the inhomogeneity parameters it is possible to argue that these functional equations still describes transfer matrix eigenvalues on the hyperplanes where SOV method cannot be applied and, in particular, in the homogeneous limit. However, in all these cases the statements about the simplicity of the transfer matrix spectrum and the completeness of the description by our functional $T$-$Q$ equation are not anymore granted and they require independent proofs. These fundamental issues will be addressed in a future publication. Here we want just to recall that the comparison with the few existing numerical results on the subject seems to suggests that the statement of completeness should be satisfied even in the homogeneous limit of special interest as it allows to reproduce the spectrum of the Hamiltonian of the spin-1/2 open XXZ quantum chains under the most general integrable boundary conditions. Finally, it is important to note that the form of the Baxter functional equation for the most general spin-1/2 representations of the 6-vertex reflection algebras and in particular the necessity of an inhomogeneous term are mainly imposed by the requirement that the set of solutions is restricted to polynomials. Then the problem to get homogeneous Baxter equations relaxing this last requirement remains an interesting open problem. Acknowledgements {#acknowledgements .unnumbered} ================ The authors would like to thank E. Sklyanin and V. Terras for discussions. J.M.M. and G. N. are supported by CNRS. N.K and J.M.M. are supported by ANR grant “DIADEMS”. N. K. would like to thank LPTHE, University Paris VI and Laboratoire de Physique, ENS-Lyon for hospitality. [99]{} P. Baseilhac. The $q$-deformed analogue of the Onsager algebra: Beyond the Bethe ansatz approach 754 (2006) 309. P. Baseilhac, K. Koizumi. A deformed analogue of Onsager’s symmetry in the XXZ open spin chain (2005) P10005. R. Baxter. Partition function of the eight-vertex lattice model. , 70:193–228, 1972. R. J. Baxter. One-dimensional anisotropic [H]{}eisenberg chain. , 70:323–37, 1972. H. Bethe. Zür [T]{}heorie der [M]{}etalle [I]{}. [E]{}igenwerte und [E]{}igenfunktionen [A]{}tomkete. , 71:205–226, 1931. J. Cao, H.-Q. Lin, K.-J. Shi, and Y. Wang. Exact solution of [$XXZ$]{} spin chain with unparallel boundary fields. , 663(3):487–519, 2003. J. Cao, W. Yang, K. Shi, and Y. Wang. Off-diagonal Bethe ansatz solutions of the anisotropic spin-1/2 chains with arbitrary boundary fields. , 887:152-175, 2013. J. Cao, W. Yang, K. Shi, and Y. Wang. Off-diagonal Bethe ansatz and exact solution a topological spin ring. 111, 137201, 2013. J. Cao, W. Yang, K. Shi, and Y. Wang. Off-diagonal bethe ansatz solution of the XXX spin-chain with arbitrary boundary conditions. , 875:152–165, 2013. J. Cao, W. Yang, K. Shi, and Y. Wang. Spin-1/2 XYZ model revisit: general solutions via off-diagonal Bethe ansatz. arXiv preprint arXiv:1307.0280, 2013. I. V. Cherednik. Factorizing particles on a half-line and root systems. , 61:977–983, 1984. N. Crampé, E. Ragoucy, D. Simon. Eigenvectors of open XXZ and ASEP models for a class of non-diagonal boundary conditions P11038 (2010). N. Crampé, E. Ragoucy Generalized coordinate Bethe ansatz for non-diagonal boundaries 858 (2012) 502. J. de Gier and F. H. L. Essler. Bethe ansatz solution of the asymmetric exclusion process with open boundaries. , 95(24):240601, 4, 2005. J. de Gier and F. H. L. Essler. Exact spectral gaps of the asymmetric exclusion process with open boundaries. , 2006(12):P12011, 2006. W. Galleas. Functional relations from the Yang-Baxter algebra: Eigenvalues of the XXZ model with non-diagonal twisted and open boundary conditions 790 (2008) 524. J. de Gier and P. Pyatov. Bethe ansatz for the Temperley-Lieb loop model with open boundaries (2004). N. Grosjean, J. M. Maillet, and G. Niccoli. On the form factors of local operators in the lattice sine-Gordon model. , P10006, 2012. N. Grosjean, J.-M. Maillet, G. Niccoli, On the form factors of local operators in the Bazhanov-Stroganov and chiral Potts models. arXiv:1309.4701. N. Grosjean, G. Niccoli, The $\tau_2$-model and the chiral Potts model revisited: completeness of Bethe equations from Sklyanin’s SOV method. P11005 (2012). L. D. Faddeev, E. K. Sklyanin, and L. A. Takhtajan. Quantum inverse problem method [I]{}. , 40:688–706, 1979. S. Faldella, N. Kitanine, G. Niccoli. Complete spectrum and scalar products for the open spin-1/2 XXZ quantum chains with non-diagonal boundary terms. Accepted for publication on *J. Stat. Mech.: Theory Exp.* arXiv:1307.3960. S. Faldella, G. Niccoli. SOV approach for integrable quantum models associated to the most general representations on spin-1/2 chains of the 8-vertex reflection algebra. arXiv:1307.5531. G. Filali and N. Kitanine. Spin chains with non-diagonal boundaries and trigonometric [SOS]{} model with reflecting end. , 7:Paper 012, 22, 2011. H. Frahm, A. Seel, T. Wirth. Separation of variables in the open XXX chain. (2008) 351. H. Frahm, J. H. Grelik, A. Seel, T. Wirth. Functional Bethe ansatz methods for the open XXX chain. (2011) 015001. Y. Jiang, S. Cui, J. Cao, Wen-Li Yang and Y. Wang, Completeness and Bethe root distribution of the spin- 1/2 Heisenberg chain with arbitrary boundary fields. arXiv:1309.6456v1. N. Kitanine, K. Kozlowski, J. Maillet, G. Niccoli, N. Slavnov, and V. Terras. On correlation functions of the open [$XXZ$]{} chain [I]{}. , pages P10009, 37 pp. (electronic), 2007. N. Kitanine, K. Kozlowski, J. Maillet, G. Niccoli, N. Slavnov, and V. Terras. On correlation functions of the open [$XXZ$]{} chain [II]{}. , page P07010, 2008. arXiv:0803.3305. N. Kitanine, J. M. Maillet, and V. Terras. Form factors of the [$XXZ$]{} [H]{}eisenberg spin-1/2 finite chain. , 554 \[FS\]:647–678, 1999. N. Kitanine, J. M. Maillet, and V. Terras. Correlation functions of the [$XXZ$]{} heisenberg spin-1/2 chain in a magnetic field. , 567 \[FS\]:554–582, 2000. J. M. Maillet and V. Terras. On the quantum inverse scattering problem. , 575:627, 2000. Preprint LPENSL-TH-19/99, hep-th/9911030. R. Murgan and R. I. Nepomechie. Bethe ansatz derived from the functional relations of the open XXZ chain for new special cases. , (2005) P08002. R. I. Nepomechie. Solving the open [XXZ]{} spin chain with nondiagonal boundary terms at roots of unity. , 622(3):615 – 632, 2002. R. I. Nepomechie. Bethe ansatz solution of the open [$XXZ$]{} chain with nondiagonal boundary terms. , 37(2):433–440, 2004. Special issue on recent advances in the theory of quantum integrable systems. R. I. Nepomechie, Inhomogeneous T-Q equation for the open XXX chain with general boundary terms: completeness and arbitrary spin. arXiv:1307.5049. R. I. Nepomechie and F. Ravanini, Completeness of the Bethe Ansatz solution of the open XXZ chain with nondiagonal boundary terms , 11391-11402, 2003. R. I. Nepomechie and F. Ravanini, Addendum to ‘Completeness of the Bethe Ansatz solution of the open XXZ chain with nondiagonal boundary terms’ , 1945-1946, 2004. R. I. Nepomechie and C. Wang, Boundary energy of the open XXX chain with a non-diagonal boundary term. arXiv:1310.6305. G. Niccoli. Reconstruction of Baxter Q-operator from Sklyanin SOV for cyclic representations of integrable quantum models. , 835: 263-283, 2010. G. Niccoli. Completeness of Bethe Ansatz by Sklyanin SOV for Cyclic Representations of Integrable Quantum Models. , 1103:123, 2011. G. Niccoli. Non-diagonal open spin 1/2 [X]{}[X]{}[Z]{} quantum chains by separation of variables: complete spectrum and matrix elements of some quasi-local operators. , 2012(10):P10025, 2012. G. Niccoli, On the form factors of local operators in the Bazhanov-Stroganov and chiral Potts models. ICMP12 Proceedings by World Scientific. arXiv:1301.4924. G. Niccoli. Antiperiodic spin-1/2 XXZ quantum chains by separation of variables: Complete spectrum and form factors. , 870: 397 – 420, 2013. G. Niccoli. Form factors and complete spectrum of XXX antiperiodic higher spin chains by quantum separation of variables. 54, 053516 (2013). G. Niccoli. Antiperiodic dynamical 6-vertex model I: Complete spectrum by SOV, matrix elements of the identity on separate states and connections to the periodic 8-vertex model. 46 075003, 2013. G. Niccoli and J. Teschner. The sine-Gordon model revisited: I. , P09014 (2010). A. Nichols, V. Rittenberg and J. de Gier. One-boundary Temperley-Lieb algebras in the XXZ and loop models. P03003 (2005). T. Prosen. Open [X]{}[X]{}[Z]{} spin chain: Nonequilibrium steady state and a strict bound on ballistic transport. , 106:217206, May 2011. J. Sirker, R. G. Pereira, and I. Affleck. Diffusion and ballistic transport in one-dimensional quantum systems. , 103:216602, Nov 2009. E. K. Sklyanin. The quantum [T]{}oda chain. In *Nonlinear equations in classical and quantum field theory ([M]{}eudon/[P]{}aris, 1983/1984)*, volume 226 of *Lecture Notes in Phys.*, pages 196–233. Springer, Berlin, 1985. E. Sklyanin. Boundary conditions for integrable quantum systems. , 21:2375–2389, 1988. E. K. Sklyanin. Poisson structure of a periodic classical XYZ chain. [*J. Soviet Math.*]{}, 1989, v.46, n.1., p.1664-1683. E. K. Sklyanin. Poisson structure of classical XXZ chain. [*J. Soviet Math.*]{}, 1989, v.46, n.5, p.2104-2111. E. K. Sklyanin. Quantum inverse scattering method. selected topics. In M.-L. Ge, editor, *Quantum group and Quantum Integrable Systems*, pages 63–97. Nankai Lectures in Mathematical Physics, World Scientific, 1992. L. A. Takhtajan and L. D. Faddeev. The quantum method of the inverse problem and the Heisenberg [XYZ]{} model. , 34(5):11–68, 1979. W.-L. Yang and Y.-Z. Zhang. On the second reference state and complete eigenstates of the open [$XXZ$]{} chain. , pages 044, 11 pp. (electronic), 2007. [^1]: Different methods leading to Bethe ansatz equations have been also proposed under the same boundary conditions by using the framework of the Temperley-Lieb algebra in [@deGP04; @NicRd05] and by making a combined use of coordinate Bethe ansatz and matrix ansatz in [@CraRS10; @CraR12]. [^2]: Some partial results in this direction were achieved in [@FilK11] but only in the special case of double boundary constrains introduced in [YanZ07]{}. [^3]: See also the papers [@CaoYSW13-1; @CaoYSW13-2; @CaoYSW13-3] for the application of the same method to different models. [^4]: Here we use notations similar to those introduced in the papers [@CaoJYW2013] and [@Nep-2013] where some inhomogeneous Baxter equation ansatzs appear with the aim to make simpler for the reader a comparison when the limit of homogeneous chain is implemented.

ArXiv

--- abstract: | We study non-topological solitons, so called $Q$-balls, which carry a non-vanishing Noether charge and arise as lump solutions of self-interacting complex scalar field models. Explicit examples of new axially symmetric non-spinning $Q$-ball solutions that have not been studied so far are constructed numerically. These solutions can be interpreted as angular excitations of the fundamental $Q$-balls and are related to the spherical harmonics. Correspondingly, they have higher energy and their energy densities possess two local maxima on the positive $z$-axis. We also study two $Q$-balls interacting via a potential term in $3+1$ dimensions and construct examples of stationary, solitonic-like objects in $(3+1)$-dimensional flat space-time that consist of two interacting global scalar fields. We concentrate on configurations composed of one spinning and one non-spinning $Q$-ball and study the parameter-dependence of the energy and charges of the configuration. In addition, we present numerical evidence that for fixed values of the coupling constants two different types of 2-$Q$-ball solutions exist: solutions with defined parity, but also solutions which are asymmetric with respect to reflexion through the $x$-$y$-plane. --- [**Interacting $Q$-balls**]{} ${}^{a)}$[*Faculté des Sciences, Université de Mons-Hainaut, 7000 Mons, Belgium*]{}\ ${}^{b)}$[*School of Engineering and Science, Jacobs University Bremen, 28759 Bremen, Germany*]{}\ Introduction ============ Solitons play an important role in many areas of physics. As classical solutions of non-linear field theories, they are localised structures with finite energy, which are globally regular. In general, one can distinguish topological and non-topological solitons. While topological solitons [@ms] possess a conserved quantity, the topological charge, that stems (in most cases) from the spontaneous symmetry breaking of the theory, non-topological solitons [@fls; @lp] have a conserved Noether charge that results from a symmetry of the Lagrangian. The standard example of non-topological solitons are $Q$-balls [@coleman], which are solutions of theories with self-interacting complex scalar fields. These objects are stationary with an explicitely time-dependent phase. The conserved Noether charge $Q$ is then related to the global phase invariance of the theory and is directly proportional to the frequency. $Q$ can e.g. be interpreted as particle number [@fls]. While in standard scalar field theories, it was shown that a non-renormalisable $\Phi^6$-potential is necessary [@vw], supersymmetric extensions of the Standard Model (SM) also possess $Q$-ball solutions [@kusenko]. In the latter case, several scalar fields interact via complicated potentials. It was shown that cubic interaction terms that result from Yukawa couplings in the superpotential and supersymmetry breaking terms lead to the existence of $Q$-balls with non-vanishing baryon or lepton number or electric charge. These supersymmetric $Q$-balls have been considered recently as possible candidates for baryonic dark matter [@dm] and their astrophysical implications have been discussed [@implications]. Two interacting scalar fields are also interesting from another point of view. Up until now, the number of explicit examples of stationary solitonic-like solutions that involve two interacting global scalar fields is small. An important example are superconducting strings, which are axially symmetric in $2+1$ dimensions extended trivially into the $z$-direction [@witten]. Axially symmetric generalisations in $3+1$ dimensions, so-called vortons, have been constructed in [@ls]. Note that all these solutions have been constructed in models which have a renormalisable $\Phi^4$-potential. Here, we study two interacting scalar fields in $3+1$ dimensions and construct explicit examples of stationary solitonic-like axially symmetric solutions consisting of two global scalar fields. While vortons possess one scalar field with an unbroken U(1) symmetry (the condensate field) and a scalar field whose U(1) is spontaneously broken (the string field), we here consider two scalar fields with unbroken U(1) symmetries. One can thus see our model as the limit of vanishing vacuum expectation value for the second scalar field. Then, stationary solitonic like objects can be constructed explicitely. Note that the model in [@ls] contains a renormalisable $\Phi^4$-potential, while we need a non-renormalisable $\Phi^6$-potential here. However, as stated in [@ls], the explicit construction of vortons was done using also a non-renormalisable potential which contains an interaction term of the form $\Phi_1^6 \Phi_2^2$. $Q$-ball solutions in $3+1$ dimensions have been first studied in detail in [@vw]. It was realised that next to non-spinning $Q$-balls, which are spherically symmetric, spinning solutions exist. These are axially symmetric with energy density of toroidal shape and angular momentum $J=kQ$, where $Q$ is the Noether charge of the solution and $k\in \mathbb{Z}$ corresponds to the winding around the $z$-axis. Approximated solutions of the non-linear partial differential equations were constructed in [@vw] by means of a truncated series in the spherical harmonics to describe the angular part of the solutions. The full partial differential equation was solved numerically in [@kk]. It was also realised in [@vw] that in each $k$-sector, parity-even ($P=+1$) and parity-odd ($P=-1$) solutions exist. Parity-even and parity-odd refers to the fact that the solution is symmetric and anti-symmetric, respectively with respect to a reflection through the $x$-$y$-plane, i.e. under $\theta\rightarrow \pi-\theta$. These two types of solutions are closely related to the fact that the angular part of the solutions constructed in [@vw; @kk] is connected to the spherical harmonic $Y_0^0(\theta,\varphi)$ for the spherically symmetric $Q$-ball, to the spherical harmonic $Y_1^{1}(\theta,\varphi)$ for the spinning parity even ($P=+1$) solution and to the spherical harmonic $Y_2^{1}(\theta,\varphi)$ for the parity odd ($P=-1$) solution, respectively. Radially excited solutions of the spherically symmetric, non-spinning solution were also obtained. These solutions are still spherically symmetric but the scalar field develops one or several nodes for $r\in ]0,\infty[$. In relation to the apparent connection of the angular part of the known solutions to the spherical harmonics, it is natural to investigate whether “$\theta$-angular excitations” of the $Q$-balls exist in correspondence to the whole family of spherical harmonics $Y_L^k(\theta,\varphi)$, $-L \leq k \leq L$. This can further be motivated by the fact that, in the small field limit where a linear approximation can be used, the field equation describing the $Q$-ball becomes a standard harmonic equation that can be solved by separation of variables and whose fundamental solutions are given in terms of spherical harmonics for the angular part. Of course, it has to be checked whether this correspondence, expected from the linear limit, still holds for the full, i.e. non-linear equation. In the present paper, we present strong numerical arguments that new angularly excited solutions of the non-linear field equations exist and that the correspondence between angular excitations of the $Q$-balls and spherical harmonics indeed holds. In addition to the solutions corresponding to $Y_k^{ k}$ and $Y_k^{k-1}$ for $k=1,2,3$ presented in [@vw] we have constructed solutions with angular dependence and symmetries corresponding to the spherical harmonics $Y_1^0$ and $Y_2^0$. These solutions are non-spinning but constitute axially symmetric excitations with respect to the angular coordinate $\theta$. As expected, these new solutions have higher energies and charges than the spherically symmetric solutions and we would thus expect them to be unstable. These solutions thus complete the already known spectrum of $Q$-ball solutions and show that not only radial excitations of fundamental soliton solutions, but also angular excitations exist. We also study two interacting $Q$-balls and put the emphasis on the interaction between a non-spinning and a spinning $Q$-ball. In particular, we investigate the dependence of the energy and the charges of the solution on the interaction parameter and the frequencies, respectively. Next to parity-even and parity-odd solutions, we also construct solutions that have no defined parity with respect to reflexion through the $x$-$y$-plane. The explicit construction of solutions with two interacting complex scalar fields is surely of interest for the astrophysical implications of such objects, especially for the construction of such objects in supersymmetric theories. Moreover, it adds to the spectrum of soliton solutions that e.g. possess no definite parity. The differential equations describing both excited as well as interacting $Q$-balls are non-linear partial differential equations, which -to our knowledge- cannot be solved analytically. We thus solve these equations numerically using an appropriate PDE solver [@fidi]. Our paper is organised as follows: in Section 2, we discuss the model and give the equations and boundary conditions. In Section 3, we discuss the new $Q$-ball solutions for $k=0$, while in Section 4, we present our results for two interacting $Q$-balls. Section 5 contains our conclusions. The model ========= In the following, we study a scalar field model in $3+1$ dimensions describing two $Q$-balls interacting via a potential term. The Lagrangian reads: $$\label{lag} {\cal L}=\frac{1}{2}\partial_{\mu} \Phi_1 \partial^{\mu} \Phi_1^*+ \frac{1}{2}\partial_{\mu} \Phi_2 \partial^{\mu} \Phi_2^* - V(\Phi_1,\Phi_2)$$ where both $\Phi_1$ and $\Phi_2$ are complex scalar fields. The potential reads: $$U(\Phi_1,\Phi_2)=\sum_{i=1}^2\left( \alpha_i \vert\Phi_i\vert^6 - \beta_i \vert\Phi_i\vert^4 + \gamma_i \vert\Phi_i\vert^2 \right) +\lambda \vert\Phi_1\vert^2 \vert\Phi_2\vert^2$$ where $\alpha_i$, $\beta_i$, $\gamma_i$, $i=1,2$ are the standard potential parameters for each $Q$-ball, while $\lambda$ denotes the interaction parameter. In [@vw] it was argued that a $\Phi^6$-potential is necessary in order to have classical $Q$-ball solutions. This is still necessary for the model we have defined here, since we want $\Phi_1=0$ and $\Phi_2=0$ to be a local minimum of the potential. A pure $\Phi^4$-potential which is bounded from below wouldn’t fulfill these criteria. The Lagrangian (\[lag\]) is invariant under the two global U(1) transformations $$\Phi_1 \rightarrow \Phi_1 e^{i\alpha_1} \ \ \ , \ \ \ \Phi_2 \rightarrow \Phi_2 e^{i\alpha_2}$$ which can be applied separately or together. As such the total conserved Noether current $j^{\mu}_{(tot)}$, $\mu=0,1,2,3$, associated to these symmetries is just the sum of the two individually conserved currents $j^{\mu}_{1}$ and $j^{\mu}_2$ with $$j^{\mu}_{(tot)}= j^{\mu}_1 +j^{\mu}_2 = -i \left(\Phi_1^* \partial^{\mu} \Phi_1 - \Phi_1 \partial^{\mu} \Phi_1^*\right) -i \left(\Phi_2^* \partial^{\mu} \Phi_2 - \Phi_2 \partial^{\mu} \Phi_2^*\right)\ \ .$$ with $\partial_{\mu} j^{\mu}_{1}=0$, $\partial_{\mu} j^{\mu}_{2}=0$ and $\partial_{\mu} j^{\mu}_{(tot)}=0$. The total Noether charge $Q_{(tot)}$ of the system is then the sum of the two individual Noether charges $Q_1$ and $Q_2$: $$Q_{(tot)}=Q_1+Q_2= -\int j_1^0 d^3 x -\int j_2^0 d^3 x$$ Finally, the energy-momentum tensor reads: $$T_{\mu\nu}=\sum_{i=1}^2 \left(\partial_{\mu} \Phi_i \partial_{\nu} \Phi_i^* +\partial_{\nu} \Phi_i \partial_{\mu} \Phi_i^*\right) -g_{\mu\nu} {\cal L}$$ Ansatz ------ We choose as Ansatz for the fields in spherical coordinates: $$\label{ansatz1} \Phi_i(t,r,\theta,\varphi)=e^{i\omega_i t+ik_i\varphi} \phi_i(r,\theta) \ \ , \ i=1,2$$ where the $\omega_i$ and the $k_i$ are constants. Since we require $\Phi_i(\varphi)=\Phi_i(\varphi+2\pi)$, $i=1,2$, we have that $k_i\in \mathbb{Z}$. It was moreover demonstrated in [@vw; @kk] that $Q$-balls exist only in a specific parameter range $\omega_{min} < \omega < \omega_{max}$ and that the charge $Q$ tends to infinity when either $\omega \rightarrow \omega_{min}$ or $\omega \rightarrow \omega_{max}$. We discuss the limits in the 2 $Q$-ball system in the following section. The Noether charges of the solution then read: $$Q_i = 2\omega_i \int \vert \Phi_i \vert^2 \ d^3 x = 4\pi \omega_i \int_{0}^{\pi} \int_0^{\infty} r^2 \sin\theta \ dr \ d\theta \ \phi_i^2 \ \ \ , \ \ i=1,2$$ while the energy is given by the volume integral of the $tt$-component of the energy-momentum tensor: $$\begin{aligned} E=\int T_{00} \ d^3 x &=&2\pi \int\limits_0^{\pi}\int\limits_0^{\infty}\left[ r^2 \ \sin\theta \ dr \ d\theta \sum_{i=1}^2\left(\omega_i \phi_i + (\phi_i')^2 + \frac{(\dot{\phi}_i)^2 }{r^2} \right.\right. \nonumber \\ &+& \left. \left. \frac{k_i^2\phi_i^2}{r^2\sin^2\theta} + \alpha_i \vert\phi_i\vert^6 -\beta_i \vert\phi_i\vert^4 + \gamma_i \vert\phi_i\vert^2 \right) + \lambda \vert\phi_1\vert^2 \vert\phi_2\vert^2 \right]\end{aligned}$$ where the prime and dot denote the derivative with respect to $r$ and $\theta$, respectively. For $k_i\neq 0$, the solutions have non-vanishing angular momentum that is quantised. The total angular momentum $J$ is the sum of the angular momenta of the two individual $Q$-balls: $$J=\int T_{0\varphi} d^3 x =J_1+J_2= k_1 Q_1 + k_2 Q_2$$ We will thus in the following refer to solutions with $k_i=0$ as non-spinning and to solutions with $k_i\neq 0$ as spinning. The Euler-Lagrange equations read: $$\label{eom} \phi_i''+ \frac{2}{r} \phi_i' + \frac{1}{r^2} \ddot{\phi}_i + \frac{1}{r^2} \cot\theta \dot{\phi}_i - \frac{k_i^2}{r^2\sin^2\theta}\phi_i + \omega_i^2 \phi_i = 3\alpha_i \phi_i^5 - 2\beta_i \phi_i^3 + \gamma_i \phi_i + \lambda \phi_i \phi_k^2$$ with $i=1,2$ and $k\neq i$. The boundary conditions, which result from requirements of regularity, finiteness of the energy and the symmetry of the solution, are: $$\label{bc1} \partial_r\phi_i(r=0,\theta)= 0 \ , \ \partial_{\theta}\phi_i(r=\infty,\theta)=0 \ , \ \partial_{\theta}\phi_i(r,\theta=0,\pi)=0 \ , \ i=1,2 \ .$$ for non-spinning solutions with $k_i=0$ and $$\label{bc2} \phi_i(r=0,\theta)= 0 \ , \ \phi_i(r=\infty,\theta)=0 \ , \ \phi_i(r,\theta=0,\pi)=0 \ , \ i=1,2 \ .$$ for spinning solutions $k_i\neq 0$. Bounds on $\omega_1$ and $\omega_2$ in the 2-$Q$-ball system ------------------------------------------------------------ In [@vw; @kk] the bounds on the frequency $\omega$ have been discussed in the case of one $Q$-ball. Here, we note that these bounds have to be modified if one considers two interacting $Q$-balls. The set of equations (\[eom\]) can be interpreted as the mechanical equations describing the frictional motion of a particle in two dimensions. The effective potential in this case reads: $$V(\phi_1,\phi_2)= \frac{1}{2} (\omega_1^2\phi_1^2 + \omega_2^2\phi_2^2) - \frac{1}{2} U(\phi_1,\phi_2)$$ $Q$-ball solutions exist provided the configuration $(\phi_1=0$,$\phi_2=0)$ corresponds to a local maximum of the effective potential and provided the effective potential has positive values in any radial direction from the origin in the $\phi_1-\phi_2-$plane. This leads to non-trivial bounds for the parameters $\omega_1$ and $\omega_2$. The former condition leads to the requirement that $$\omega_1^2 < \omega_{1,max}^2 = \gamma_1 \ \ , \ \ \omega_2^2 < \omega_{2,max}^2 = \gamma_2 \ \ .$$ The latter condition leads to a more complicated domain of existence in the $\omega_1$-$\omega_2$-plane. To describe this condition, we introduce the polar decomposition of $\phi_1$ and $\phi_2$ as follows: $$\phi_1=\rho\cos\chi \ \ , \ \ \phi_2=\rho\sin\chi \ .$$ where $0 \le \chi < 2\pi$ and $0 \le \rho < \infty$. The condition on the frequencies $\omega_1$ and $\omega_2$ then read: $$\omega_1^2 \cos^2\chi + \omega_2^2 \sin^2\chi > (\omega_1^2 \cos^2\chi + \omega_2^2 \sin^2\chi)_{min} = {\rm min}_{\rho}[U(\rho,\chi)/\rho^2] \ \ , \ \ \forall \ \chi$$ In the particular case that we have studied throughout this paper, namely $\alpha_1=\alpha_2=1$, $\beta_1=\beta_2=2$ and $\gamma_1=\gamma_2=1.1$ this inequality takes the form: $$\begin{aligned} \omega_1^2 \cos^2\chi + \omega_2^2 \sin^2\chi & > & \left[-5\lambda^2 \cos^4\chi\sin^4\chi+ 20\lambda\cos^2\chi\sin^2\chi (\cos^4\chi + \sin^4\chi) \nonumber \right. \\ &+& \left. 2(\cos^8\chi + \sin^8\chi + 11 \cos^6\chi\sin^2\chi + 11 \sin^6\chi\cos^2\chi \nonumber \right. \\ &-& \left. 20 \cos^4\chi\sin^4\chi)\right]/(\cos^4\chi +\sin^4\chi -\cos^2\chi\sin^2\chi) \end{aligned}$$ For $\chi= n \pi/2$, $n=0,1,2,...$, we recover the results of the one $Q$-ball system discussed in [@vw; @kk]. For all other values of $\chi$, the limiting values for $\omega_1$ and $\omega_2$ will depend on the value of the interaction coupling $\lambda$. E.g. for $\phi_1=\phi_2$, i.e. $\chi=\pi/4$, we find : $$\omega_1^2 + \omega_2^2 > 1/5 + \lambda - 1/8 \lambda^2$$ Thus, for small $\lambda$, the lower bound on the value of $\omega_1^2 + \omega_2^2$ will be larger than in the non-interacting limit. New non-spinning $Q$-ball solutions for $\alpha_2=\beta_2=\gamma_2=\lambda=0$ ============================================================================= In order to be able to understand the structure of a system of two $Q$-balls, we have reconsidered the one $Q$-ball system. We set all quantities with index “2” to zero in the following and omit the index “1” for the remaining quantities. In this section, we would like to point out that more than the previously in the literature discussed solutions exist. For this, we first consider the equation for one $Q$-ball with vanishing potential. This reads: $$\phi''+ \frac{2}{r}\phi' + \frac{1}{r^2} \ddot{\phi} + \frac{1}{r^2} \cot\theta \dot{\phi} - \frac{k^2}{r^2\sin^2\theta}\phi + \omega^2 \phi=0$$ Although the solutions of the above equation are well known, it will be useful for the following to recall their properties. Using the standard separation of variables, the solutions read: $$\phi(r,\theta,\varphi)\propto \frac{J_{L+1/2}(\omega r)}{\sqrt{r}} Y_L^k(\theta,\varphi)$$ where $J$ denotes the Bessel function, while $Y_L^k$ are the standard spherical harmonics with $-L \le k \le L$. One may hope that solutions of the full equations with the discrete symmetries corresponding to the ones of the spherical harmonics will exist. Of course, the non-linear potential interaction will deform the radial part of the solutions of the linear equation in a highly non-trivial manner. The solutions of the full equation constructed so far for $k=0$ have been spherically symmetric. With the above arguments, axially symmetric solutions should equally exist with an angular dependence of the form $Y_L^0$, e.g. for $L=1$, the angular dependence should be of the form $\cos\theta$. In the following, we will denote the solutions of the full non-linear equations with angular symmetries corresponding to the symmetries of the spherical harmonic $Y_L^k$ by $\phi_L^k$. Numerical results ----------------- ![\[fig1a\] The profile of the function $\phi_1^0$ is shown for $\omega = 0.8$, $\alpha=1$, $\beta=2$, $\gamma=1.1$ (left). The corresponding energy density $T_{00}$ is also given (right).](newsol1a.eps "fig:"){width="8cm"} ![\[fig1a\] The profile of the function $\phi_1^0$ is shown for $\omega = 0.8$, $\alpha=1$, $\beta=2$, $\gamma=1.1$ (left). The corresponding energy density $T_{00}$ is also given (right).](newsol1b.eps "fig:"){width="8cm"} The partial differential equation has been solved numerically subject to the boundary conditions (\[bc1\]) or (\[bc2\]) using the finite difference solver FIDISOL [@fidi]. We have mapped the infinite interval of the $r$ coordinate $[0:\infty]$ to the finite compact interval $[0:1]$ using the new coordinate $z:=r/(r+1)$. We have typically used grid sizes of $150$ points in $r$-direction and $50$ points in $\theta$ direction. The solutions have relative errors of $10^{-3}$ or smaller. Throughout this section, we choose $\alpha_1\equiv\alpha=1$, $\beta_1\equiv\beta=2$, $\gamma_1\equiv\gamma=1.1$. In Fig.\[fig1a\] (left), we show the profile of a new solution that we obtained for $k=0$ and $\omega=0.8$. This solution looks like a deformation of the spherical harmonic $Y_1^0$ with the appropriate symmetry with respect to $\theta=\pi/2$ and is clearly axially symmetric. In particular it fulfills $\phi_1^0(r,\pi/2)=0$. The field $\phi_1^0(r,\theta)$ is maximal at a finite distance from the origin on the positive $z$-axis. Moreover, the configuration is anti-symmetric under reflexion through the $x-y-$plane, i.e. under $\theta \rightarrow \pi-\theta$. Thus the solution is parity-odd: $P=-1$. Note that we have only plotted the function for $\theta\in [0:\pi/2]$, but that we have verified the symmetry of the solution. =11.0cm\ We also present the corresponding energy density $T_{00}$ in Fig.\[fig1a\] (right). It shows that the density of the solution is mainly concentrated within two small “balls” situated around the positive $z$-axis (at $z\approx 2.4$ and $z\approx 7.6$) and separated by a minimum (at $z\approx 5$). The position of this minimum coincides with the maximum of the scalar field $(\phi_1^0)_{max} \approx \phi_1^0(5,0) \approx 1.2$. It can be checked that this value corresponds roughly to a local minimum of the potential while the partial derivatives are evidently small in this region, explaining the occurrence of a minimal value of the energy density at $(x,y,z) \approx (0,0,5)$. Of course, due to the anti-symmetry of the solution this pattern is equally given on the negative $z$-axis. =10.0cm\ The classical energy and charge of this new solution is higher than that of the spherically symmetric $k=0$ solution (see Table 1 and Table 2 below), however lower than that of the $k=1$ spinning $Q$-ball. In order to investigate further our idea of constructing new solutions as deformations of the spherical harmonics, we have also investigated solutions with higher value of $L$ and we managed to construct solutions $\phi_2^0$ and $\phi_3^0$ corresponding in their angular symmetries to those of the spherical harmonics $Y_2^0 \propto 3\cos^2\theta -1$ and $Y_3^0 \propto 5\cos^3\theta -3\cos\theta$, respectively. In Fig.\[fignew\], we plot $\phi_1^0$, $\phi_2^0$ and $\phi_3^0$ as functions of $\theta$ for a fixed value of $r$ together with the corresponding spherical harmonics $Y_1^0$, $Y_2^0$ and $Y_3^0$. Here, we have chosen $r\sim 5$ for $\phi_1^0$, $r\sim 2$ for $\phi_2^0$ and $r\sim 6$ for $\phi_3^0$. The first thing to notice is that the symmetries of the solutions $\phi_1^0$, $\phi_2^0$ and $\phi_3^0$ with respect to reflection at $\theta=\pi/2$ are exactly equal to those of the corresponding spherical harmonics. The actual solutions are, of course, deformed with respect to the spherical harmonics, but the correspondence is apparent. E.g. the solution $\phi_2^0$ has $\partial_\theta \phi_2^0(r, \pi/2)=0$ (in contrast to the solution $\phi_1^0$ which has $\phi_1^0(r, \pi/2)=0$). We don’t show the energy density of $\phi_2^0$ and $\phi_3^0$ here, since it resembles that shown in Fig. 1. We believe that the correspondence also holds for higher spherical harmonics. $\phi_0^0$ $\phi_1^0$ $\phi_2^0$ $\phi_1^1$ ------------ ------------ ------------ ------------ ------------ $E$ $73.6$ $141.6$ $170.9$ $223.5$ $Q/\omega$ $75.2$ $146.9$ $176.1$ $220.4$ $P$ $+1$ $-1$ $+1$ $+1$ symmetry spherical axial axial axial : The energy $E$, the charge per frequency $Q/\omega$, the parity $P$ and the symmetry of the first few $Q$-ball solutions is given for $\omega=0.8$. $\phi_0^0$ $\phi_1^0$ $\phi_2^0$ $\phi_1^1$ ------------ ------------ ------------ ------------ ------------ $E$ $61.2$ $115.8$ $179$ $195.6$ $Q/\omega$ $60.0$ $114.7$ $192.1$ $186.3$ $P$ $+1$ $-1$ $+1$ $+1$ symmetry spherical axial axial axial : The energy $E$, the charge per frequency $Q/\omega$, the parity $P$ and the symmetry of the first few $Q$-ball solutions is given for $\omega=0.84$. Since we have presented strong numerical evidence that the correspondence with the spherical harmonics holds, it is justified to label the different solutions of the field equation by means of the quantum numbers of the corresponding spherical harmonic, i.e. by $L$ and $k$ referring to $Y_L^k$, with $L,k$ integers and $-L \le k \le L $. Needless to say that the numerical construction becomes more involved when the difference $L-|k|$ increases. Adopting these notations and fixing the potential according to $\alpha=1$, $\beta=2$, $\gamma=1.1$, we find for the solutions corresponding to $\omega = 0.8$ and $\omega=0.84$ the values for the energy $E$ and charge per frequency $\frac{Q}{\omega}$ given in Table 1 and Table 2, respectively. The first three solutions $\phi_L^0$, $L=0,1,2$ in this list are static (i.e. non spinning) while the last, $\phi_1^1$, is stationary (i.e. spinning). For all the solutions we constructed, the energy of the non-spinning solutions is lower that the energy of the spinning ones. =12.0cm\ Interacting $Q$-balls ===================== Since in supersymmetric extensions of the Standard Model, $Q$-balls exist that result from the interaction of several scalar fields, we here investigate the interaction of two classical $Q$-balls as toy-model for these systems. For two spherically symmetric $Q$-balls ($k_1=k_2=0$) in interaction, the 2-$Q$-ball solution is still spherically symmetric and the domain of existence in the $\omega_1$-$\omega_2$-plane can be determined by using the reasoning given in Section 2.2. Here, we put the emphasis on solutions where the two $Q$-balls have different symmetries and study the effect of the direct interaction parameterized by the coupling constant $\lambda$. We believe that a particularly interesting case is the interaction between a spherically symmetric, non-spinning $Q$-ball ($k_1=0$) and a spinning $Q$-ball ($k_2=0$). We have thus restricted our analysis to this case and set $k_1=0$ and $k_2=1$ in the following. Note that we will index all quantities related to the spherical $Q$-ball in the following with “1”, while all quantities related to the axially symmetric $Q$-ball will be indexed with “2”. For later use, we define the “binding energy” of the solution according to E = E- E\_[k\_1=0]{}- E\_[k\_2=1]{}  . It represents the difference between the energy $E$ of the 2-$Q$-ball configuration and the sum of the energies of the two single (i.e. non-interacting) $Q$-balls $E_{k_1=0}$, $E_{k_2=1}$ with the same frequency. We expect that those solutions which have $\Delta E <0$ to be stable, while those with $\Delta E > 0$ would be unstable. Numerical results ----------------- We have solved the two coupled partial differential equations using the solver FIDISOL [@fidi] for several values of $\omega_1$, $\omega_2$ and $\lambda$ and fixing $\alpha_1=\alpha_2=1$, $\beta_1=\beta_2=2$ and $\gamma_1=\gamma_2=1.1$. As starting profiles, we have used the corresponding non-interacting $Q$-ball solutions. For $\lambda=0$, these solve the two decoupled partial differential equations. We have then slowly increases the parameter $\lambda$ to obtain the interacting solutions. ### $\omega_1=\omega_2$ In order to understand the influence of the interaction parameter $\lambda$, we show the energy density $T_{00}$ for $\omega_1=\omega_2=0.8$ and three different values of $\lambda$ in Fig.\[t00\]. For $\lambda=0$, the two $Q$-balls are non-interacting and the energy density is just a simple superposition of the energy densities of the two individual $Q$-balls. For $\lambda \neq 0$ the $Q$-balls interact. For $\lambda > 0$, it is energetically favourable to have the two $Q$-balls’ cores in different regions of space. As seen in Fig.\[t00\] for $\lambda=1$, the spinning $Q$-ball seems to be “pushed away” from the non-spinning, spherically symmetric one. For $\lambda < 0$, it is energetically favourable to have two $Q$-balls sitting “on top of each other”. This is shown in Fig.\[t00\] for $\lambda=-0.5$, where the two $Q$-balls seem to be localised at the same place. We have also studied the dependence of the energy $E$, the binding energy $\Delta E$ and the two charges $Q_1$ and $Q_2$ on the interaction parameter $\lambda$. The results are shown in Fig.\[lambdavary\] for $\omega_1=\omega_2=0.8$. All quantities increase with the increase of $\lambda$, specifically it is evident that the 2-$Q$-ball configuration is energetically more stable for $\lambda <0$ than for $\lambda > 0$. Specifically, we would thus expect the solution to be stable for $\lambda <0$ and unstable for $\lambda > 0$. Following our discussion in Section 2.2, we have also studied the dependence of the energy $E$ and of the charges $Q_1$ and $Q_2$ on the frequencies $\omega_1$ and $\omega_2$. Our results for $\omega_1=\omega_2$ are shown in Fig.\[fig\_qq2\] for $\lambda=-0.5$, $0$ and $0.5$, respectively. =10.0cm\ As expected, the energy $E$ for a given frequency $\omega_1=\omega_2$ is higher (resp. lower) than in the non-interacting case for positive (resp. negative) values of $\lambda$. ![\[fig\_qq2\] The energy $E$ (left) and the charges $Q_1$ and $Q_2$ (right) are shown as functions of the frequency $\omega_1=\omega_2$ for $\lambda= -0.5,0,0.5$.](eqq_om12a.eps "fig:"){width="8cm"} ![\[fig\_qq2\] The energy $E$ (left) and the charges $Q_1$ and $Q_2$ (right) are shown as functions of the frequency $\omega_1=\omega_2$ for $\lambda= -0.5,0,0.5$.](eqq_om12b.eps "fig:"){width="8cm"} As before, we find that the solutions exist in a given interval of the frequency: $\omega_{1,min}(\lambda) \leq \omega \leq \omega_{1,max}(\lambda)$ (and equally for $\omega_2$ since $\omega_1=\omega_2$). We have determined the bounds on $\omega_1$ and $\omega_2$ in Section 2.2 for two spherically symmetric $Q$-balls. Here, we would expect that these values change slightly since we have a system of one spherically symmetric and one axially symmetric $Q$-ball. However, we see that the qualitative results are similar here. We observe that for $\lambda \ge 0$, the values of the energy $E$ and of the charges $Q_1$, $Q_2$ diverge at $\omega_1=\omega_{1,min}$ and $\omega_1=\omega_{1,max}$. Following the discussion of Section 2.2 we find that the maximal value of $\omega_1$ is independent of $\lambda$. This can be clearly seen in Fig. \[fig\_qq2\] where the energy $E$ and the charges $Q_1$ and $Q_2$ diverge at $\omega_1=\omega_{1,max}\approx 1.035$ for all three values of $\lambda$. Note that this maximal value is only slightly lower than the bound given in Section 2.2: $\omega^2_{1,max}=1.1$. The reason why the bound is not equal is that here we are dealing with an axially symmetric solution interacting with a spherically symmetric one. Analytic arguments of the type done in Section 2.2 are, however, only possible if the Euler-Lagrange equations are ordinary differential equations, i.e. only in the case where the solutions are spherically symmetric. So, it is not surprising that the analytic values differ from the numerical ones. On the other hand, the minimal value of $\omega_1$ is $\lambda$-dependent. This can be seen in Fig. \[fig\_qq2\]. We have given our results only for $\omega \ge 0.6$ in this figure since the construction of solutions becomes increasingly difficult for $\omega < 0.6$. However, it can be clearly seen that the energy $E$ and $Q_2$ diverge at different values of $\omega_1=\omega_{1,min}$. In agreement with Section 2.2., we find that $\omega_{1,min}$ is increasing for increasing (and small) $\lambda$. For $\lambda < 0$ the behaviour at the lower bound of $\omega_1$ changes. We observe that $Q_1$ corresponding to the spherically symmetric field $\phi_1$ decreases when $\omega_1$ decreases. The analysis of the profile of the solution reveals that the field $\phi_1$ deviates only slightly from the spherically symmetric configuration for frequencies close to $\omega_{1,max}$. However, it gets more and more deformed in the equatorial plane when $\omega_1$ decreases. At the same time, the field $\phi_2$ increases in the equatorial plane. This phenomenon is illustrated in Fig.\[coupe\] for $\lambda = - 0.5$, $\omega = 0.6$ and $\omega= \omega_{1,max}$, respectively. In this figure, the fields $\phi_1$ and $\phi_2$ as well as the energy density $T_{00}$ are shown as function of $r$ for two angles $\theta= 0$ and $\theta = \pi/2$, respectively. =10.0cm\ ### Solutions with $\omega_1 \neq \omega_2$ =10.0cm\ We have also constructed 2-$Q$-ball solutions for $\omega_1 \neq \omega_2$. The energy density $T_{00}$ of a 2-$Q$-ball solution corresponding to $\omega_1=0.65$ and $\omega_2=1$ is shown in Fig.\[t00bis\] for four different values of $\lambda$. The result is qualitatively similar to the case $\omega_1=\omega_2$. This figure however suggests very clearly that for $\lambda < 0$ the $k_2=1$ $Q$-ball has tendency to disappear from the 2-$Q$-ball system. For instance the maximal value of the $\phi_2$ field, $\vert \phi_{2,max}\vert$, decreases for decreasing $\lambda$. We have also studied the dependence of the solution’s conserved quantities on $\omega_2 = \omega_1/0.65$ for $\lambda = \pm 0.5$. The dependence of the energy $E$ and the charges $Q_1$, $Q_2$ is shown in Fig.\[eqq\_mix\]. These results strongly suggest that for $\lambda<0$ and in the region of the parameter space chosen, the field $\phi_2$ corresponding to the $k_2=1$ $Q$-ball tends uniformly to zero for a critical value of $\omega_2=\omega_2^{(cr)}$ such that $Q_2\rightarrow 0$ for $\omega_2\rightarrow \omega_2^{(cr)}$. Only the field $\phi_1$ remains non-trivial when $\omega_2\leq \omega_2^{(cr)}$. This effect can also be observed in Fig.\[t00bis\], where the solution for $\lambda=-0.5$ has nearly lost all its axially symmetric character. We observe the inverse phenomenon for $\omega_1=c \omega_2$ with a constant $c > 0$. We don’t present our detailed results here since they are qualitatively equivalent to the case discussed above. We find that $Q_1\rightarrow 0$ for $\omega_1\rightarrow \omega_1^{(cr)}$. Thus, the spherically symmetric solution disappears from the system, while $\phi_2$ remains non-trivial for $\omega_1 \leq \omega_1^{(cr)}$. Apparently, while in the case $\omega_1=\omega_2$ and $\lambda < 0$, the charge $Q_1$ associated to the spherical $Q$-ball tends to zero for $\omega_1\rightarrow \omega_1^{(cr)}$, it is the charge $Q_i$, $i=1,2$ of the $Q$-ball with the higher frequency that tends to zero for $\omega_i\rightarrow \omega_i^{(cr)}$, $i=1,2$ when $\omega_1\neq \omega_2$ and $\lambda < 0$. Note that nothing similar is observed when $\lambda \ge 0$. =12.0cm\ While 1-$Q$-ball solutions known so far are always either parity-even or parity-odd with respect to $\theta\rightarrow \pi-\theta$, we have constructed several examples of 2-$Q$-ball solutions that do not have a defined parity. One such solution is shown in Fig.\[asym\] (lower part) together with a parity-even solution (upper part). These solutions exist for exactly the same values of the coupling constants. Both functions $\phi_1, \phi_2$ are clearly neither parity-even nor parity-odd and the field $\phi_2$ possesses in addition nodes in the radial direction. This solution is thus an asymmetric, radially excited 2-$Q$-ball solution. As expected, we observe that this asymmetric solution has much higher energy and charges than the corresponding parity-even solution. The investigation of solutions of this type and their eventual bifurcation into branches of solutions with defined parity is currently underway. =12.0cm\ Concluding remarks ================== In this paper, we have presented numerical evidence that $Q$-balls solutions admit several types of excitations labelled by integers. So far, it was known that the static, spherically symmetric solution is the “ground state” of a series of radially excited solutions. Families of spinning solutions are also known, they are axially symmetric and can be labelled according to the winding $k$ around the axis of symmetry. Here we present evidence that excitation with respect to $\theta$ can be constructed as well. Generally, the previous results and the present analysis suggest that families of elementary solutions of the field equations exist and are labelled by $n$, $L$, $k$, where $n$ refers to the number of nodes in radial direction, while $L$, $k$ refer to the “quantum numbers” related to the spherical harmonics. At the moment, the only analytic argument we have for this property is its analogy to the linearized version (i.e. small field limit) of the equation where this result holds true by standard harmonic analysis. It is likely that the qualitative properties of the solutions exist also in the case of the full non-linear equations. We have also studied a system of two interacting $Q$-balls and have constructed several examples of axially symmetric, stationary solutions that carry conserved currents and charges. We observe that the 2-$Q$-ball solutions exist in a finite range of the frequency $\omega_{i,min}\le \omega_i \le \omega_{i,max}$, $i=1,2$, where $\omega_{i,max}$ is independent of the interaction coupling, while $\omega_{i,min}$ does dependent of the interaction coupling in a highly non-trivial manner. We find that the charges $Q_i$, $i=1,2$ of the 2 $Q$-balls in interaction tend to infinity when $\omega_i\rightarrow \omega_{i,max}$ or $\omega_i\rightarrow \omega_{i,min}$ as long as $\lambda \ge 0$. For $\lambda < 0$, however, we observe that the charges $Q_i$ associated to the $Q$-ball with the higher frequency $\omega_i$ tends to zero for $\omega_i\rightarrow \omega_{i}^{(cr)} < \omega_{i,max}$. For $\omega_{i,min} \le \omega_i \le \omega_{i}^{(cr)}$ only the remaining field $\phi_j$, $j\neq i$ is non-zero. In a future publication, we intend to construct solutions with the more realistic potential available from supersymmetry [@kusenko] and put the emphasis on the possibility of constructing $Q$-balls and their excited and/or spinning versions with potentials involving only quartic terms in the scalar fields.\ \ \ [**Acknowledgments**]{} We thank Y. Verbin for discussions at the first stages of this paper. Y.B. thanks the Belgian FNRS for financial support. [99]{} N.S. Manton and P.M. Sutcliffe, [*Topological solitons*]{}, Cambridge University Press, 2004. R. Friedberg, T. D. Lee and A. Sirlin, Phys. Rev. D [**13**]{} (1976) 2739. T. D. Lee and Y. Pang, Phys. Rep. [**221**]{} (1992), 251. S. R. Coleman, Nucl. Phys. B [**262**]{} (1985), 263. M.S. Volkov and E. Wöhnert, Phys. Rev. D [**66**]{} (2002), 085003. A. Kusenko, Phys. Lett. B [**404**]{} (1997),285. A. Kusenko, hep-ph/0009089. K. Enqvist and J. McDonald, Phys. Lett. B [**425**]{}, 309 (1998); S. Kasuya and M. Kawasaki, Phys. Rev.  D [**61**]{}, 041301 (2000); A. Kusenko and P. J. Steinhardt, Phys. Rev. Lett.  [**87**]{}, 141301 (2001); T. Multamaki and I. Vilja, Phys. Lett.  B [**535**]{}, 170 (2002); M. Fujii and K. Hamaguchi, Phys. Lett.  B [**525**]{}, 143 (2002); M. Postma, Phys. Rev.  D [**65**]{}, 085035 (2002); K. Enqvist, [*et al.*]{}, Phys. Lett.  B [**526**]{}, 9 (2002); M. Kawasaki, F. Takahashi and M. Yamaguchi, Phys. Rev.  D [**66**]{}, 043516 (2002); A. Kusenko, L. Loveridge and M. Shaposhnikov, Phys. Rev.  D [**72**]{}, 025015 (2005); Y. Takenaga [*et al.*]{} \[Super-Kamiokande Collaboration\], Phys. Lett.  B [**647**]{}, 18 (2007); S. Kasuya and F. Takahashi, arXiv:0709.2634 \[hep-ph\]. E. Witten, Nucl. Phys. B [**249**]{} (1985), 557. Y. Lemperiere and E.S. Shellard, Phys. Rev. Lett. [**91**]{} (2003), 141601. B. Kleihaus, J. Kunz and M. List, Phys. Rev. D [**72**]{} (2005), 064002. W. Schönauer and R. Weiß, J. Comput. Appl. Math. [**27**]{} (1989) 279; M. Schauder, R. Weißand W. Schönauer, “The CADSOL Program Package”, Universität Karlsruhe, Interner Bericht Nr. 46/92 (1992); W. Schönauer and E. Schnepf, ACM Trans. Math. Softw. [**13**]{} (1987) 333.

ArXiv

--- author: - | José L. Flores\ [*Departamento de Álgebra, Geometría y Topología*]{}\ [*Facultad de Ciencias, Universidad de Málaga*]{}\ [*Campus Teatinos, 29071 Málaga, Spain*]{}\ [*e-mail: [email protected]*]{} --- > [**Abstract.**]{} > > [We present a new development of the causal boundary of spacetimes, originally introduced by Geroch, Kronheimer and Penrose. Given a strongly causal spacetime (or, more generally, a chronological set), we reconsider the GKP ideas to construct a family of completions with a chronology and topology extending the original ones. Many of these completions present undesirable features, like those which appeared in previous approaches by other authors. However, we show that all these deficiencies are due to the attachment of an “excessively big” boundary. In fact, a notion of “completion with minimal boundary” is then introduced in our family such that, when we restrict to these minimal completions, which always exist, all previous objections disappear. The optimal character of our construction is illustrated by a number of satisfactory properties and examples.]{} > [*Keywords:*]{} [boundary of spacetimes, causal boundary, causality theory, chronology.]{}\ > [*2000 MSC:*]{} [53C50, 83C75]{} Introduction ============ Many properties in Mathematics are usually best understood by attaching an ideal boundary to the target space. This situation is also common in General Relativity, where important physical questions about spacetimes are closely related with properties of their boundaries. In the construction of such boundaries, the causal structure of the spacetime plays a decisive role. The most common method to place an ideal boundary on a spacetime is by embedding it conformally into a larger spacetime and, then, by taking the boundary of the image. The [*conformal boundary*]{} was firstly introduced for asymptotically simple spacetimes in [@P] and, since then, it has provided a number of interesting insights in specific examples. However, this approach presents several important handicaps. The construction imposes strong mathematical restrictions on the spacetime, even though such restrictions are satisfied by many spacetimes of physical interest. On the other hand, there is no a systematic and totally general way to carry out the embedding such that the standard character of the conformal boundary is ensured. An alternative, but similar, construction based on the new concept of [*isocausality*]{} has been introduced recently in [@GS]. The more accurate character of the isocausality with respect to the causal structure of the spacetime allows the authors to generalize the conformal approach. As a consequence, this new construction is applicable to larger classes of spacetimes. However, it is unclear if this method will overcome the remaining problems of the conformal method. The existence of a systematic and intrinsic procedure to construct an ideal boundary for general spacetimes was first envisioned by Geroch, Kronheimer and Penrose in [@GKP]. They suggested a construction totally based on the causal structure of the spacetime. In particular, it is invariant under conformal changes. Roughly speaking, they placed a future (past) ideal endpoint for every inextensible future (past) timelike curve, in such a way that it only depends on the curve’s past (future). Then, they characterized these ideal endpoints by means of [*terminal indecomposable past (future) sets*]{} TIPs (TIFs) (see Section \[preliminaries\] for definitions). The GKP approach, also called [*causal completion*]{} or $c$-[*completion*]{}, overcomes the handicaps of the conformal construction. In fact, this method can be applied successfully to any strongly causal spacetime (see, however, [@S1]), and yields a systematic procedure to construct an unique boundary. However, this method presents an important technical difficulty: in [@GKP] the authors remarked that some TIPs and TIFs must act as the same ideal endpoint. In particular, this makes it necessary to define [*non-trivial*]{} identifications between these sets. There are a number of papers written in order to solve this question, which is closely related to the introduction of a satisfactory topology for the causal completion (see [@GS2] for a detailed review on the subject). The story just begins in [@GKP]. The authors introduced a generalized Alexandrov topology on the initial construction, and then, they suggested the minimum set of identifications necessary to obtain a Hausdorff ($T_{2}$) space, i.e. any two points can be separated by neighborhoods. However, this method fails to produce the “expected” completion in some examples [@KLL], [@KL1], [@MR Section 5]. On the other hand, although strong separation properties as $T_{2}$ are desirable, there are no physical reasons to impose it a priori. More annoying, from a topological viewpoint the causal boundary of Minkowski space does not match all the common expectation of a cone [@H2]. Afterwards, other more accurate attempts have been suggested. The procedure proposed in [@R], very close to the GKP approach, also fails in simple examples (see [@KL2]). Another approach proposed in [@BS], and improved later in [@S1; @S2] via the [*Szabados relation*]{} (Definition \[Szab\]; see also Section \[comparing\]), again presents undesirable properties (see [@KL1], [@KL2], [@MR Sections 2.2, 5]). Recently, Marolf and Ross have introduced in [@MR] an entirely new use of the Szabados relation, including a new topology for the completion, which overcomes important difficulties in previous attempts (see Section \[comparing\]). In fact, the MR construction satisfies essential requirements to be considered a reasonable completion: (i) the original spacetime becomes densely, chronologically and topologically embedded into the completion and, (ii) any timelike curve in the spacetime has some limit in the completion. However, apart from certain “anomalous” limit behaviors in some examples (see [@MR]), they also admit an annoying failing: not only is their topology not necessarily Hausdorff (which cannot be regarded as unsatisfactory, as we will see later), but it might not even be $T_{1}$, i.e. some point might not be closed (Example \[5\]). On the other hand, the MR completion sometimes includes too many ideal points (Example \[yo\]). Another viewpoint in the study of causal completions was previously inaugurated by Harris in [@H1]. Prevented by the necessity of non-trivial identifications when considering the past and future completions simultaneously, he only focused on the [*future chronological completion*]{} $\hat{X}$ (the same study also works for the past). In [@H1], he showed the universality of this partial completion. In [@H2], he introduced a topology for $\hat{X}$ based on a limit operator $\hat{L}$, the so called [*future chronological topology*]{} (see Section \[preliminaries\]). Then, a series of satisfactory properties for this topology were shown, including universality when the boundary is spacelike. The specific utility of this approach is checked in some wide classes of spacetimes, as static and multiwarped spacetimes (see [@H3], [@H4], [@FH]; see also [@H5] for a review). However, the lack of causal information by considering only a partial boundary also implies anomalous limit behaviors in simple cases (Example \[3\]). In this article we present a whole revision of the causal boundary of spacetimes by combining in a totally new way the GKP ideas. Our approach is very general, and, indeed, it includes not one, but a family of different completions. By imposing a minimality condition, we will choose between them those completions which are optimal, showing that these minimal completions overcome all the deficiencies which appeared in previous constructions by other authors. We have included in Section \[preliminaries\] some basic concepts and preliminary results useful for the next sections. In order to gain in generality, our paper does not treat just with spacetimes, but also [*chronological sets*]{} $(X,\ll)$, Definition \[chronrel\], a mathematical object which abstracts the chronological structure of the spacetime. Our approach essentially begins in Section \[chron\]. In Definition \[eend\] we extend the usual notion of endpoint of a timelike curve in a spacetime, to that of [*endpoint*]{} of an arbitrary chain (totally chronologically related sequence) in a chronological set. Then, we use this definition to introduce a general notion of [*completion*]{} $\overline{X}$ for a chronological set $X$, by imposing that any chain in $X$ admits some endpoint in $\overline{X}$, Definition \[completion\]. According to this definition, now a chronological set may admit [*many different*]{} completions, including the GKP [*pre-completion*]{} and the [*Marolf-Ross construction*]{} as particular cases. In order to go further, our completions need also to be made chronological sets. This is done in Section \[chronol\], where any completion $\overline{X}$ is endowed with a chronological relation $\overline{\ll}$ such that the natural inclusion ${\bf i}$ from $(X,\ll)$ to $(\overline{X},\overline{\ll})$ becomes a [*dense chronological embedding*]{}, Theorem \[tt8\]. With this structure of a chronological set, we can verify the consistency of our notion of completion. This is checked in Section \[joser\] by showing that every completion is indeed a [*complete*]{} chronological set, Definition \[def-compl\] and Theorem \[comp\]. In order to get a deeper analysis of our construction, in Section \[topology\] we have endowed any chronological set with a topology, the so-called [*chronological topology*]{}, Definitions \[overline\], \[closed\], which is inspired by the ideas in [@H2]: first, we have introduced a (sort of theoretic-set) [*limit operator*]{} $L$ on $X$, and then, defined the [*closed sets*]{} as those subsets $C\subset X$ such that $L(\sigma)\subset C$ for any sequence $\sigma\subset C$. In particular, every completion now becomes a topological space. Then, a number of very satisfactory properties are shown. With this topology, every chronological set $X$ becomes [*topologically embedded*]{} into $\overline{X}$ via the natural inclusion ${\bf i}$, Theorem \[artu\]. Therefore, as the manifold topology of a strongly causal spacetime $V$ [*coincides*]{} with its chronological topology, Theorem \[princ\], the manifold topology [*is just*]{} the restriction to $V$ of the chronological topology of $\overline{V}$. Moreover, the notion of endpoint previously introduced becomes now [*compatible*]{} (even though, non necessarily equivalent) with the notion of [*limit*]{} of a chain provided by this topology, Theorem \[compatible\]. As a consequence, $X$ is [*topologically dense*]{} in $\overline{X}$, and any timelike curve in $V$ [*has some limit*]{} in $\overline{V}$, Corollary \[thchar\]. All these properties show that these constructions verify essential requirements to be considered reasonable completions. However, many of these completions are still non-optimal, in the sense that they have boundaries formed by “too many” ideal points: clearly, this is the case of the GKP pre-completion in [@GKP], which sometimes attaches two ideal points where we would expect only one. The existence of these spurious ideal points implicitly leads to other undesirable features: non-equivalence between the notions of endpoint and limit of a chain; non-closed boundaries; completions with bad separation properties... In order to overcome these deficiencies, in Section \[universality\] we have restricted our attention to those completions $\overline{X}$ with “the smallest boundary”: that is, those completions which are [*minimal*]{} for a certain order relation based on the “size” of the boundary, Definition \[gh\]. These minimal completions, which [*always*]{} exist, Theorem \[ghh\], and, indeed, may be [*non-unique*]{} in certain cases (Example \[yo\]), are called [*chronological completions*]{}, Definition \[ghhh\]. In Theorem \[7.3\] these completions are characterized in terms of some nice properties (indeed some of them were axiomatically imposed in previous approaches). The rest of Section \[universality\] is devoted to show the very satisfactory properties of these completions for strongly causal spacetimes $V$: the notion of endpoint is now totally [*equivalent*]{} to that of the limit of a chain, Theorem \[compatible’\]; the chronological boundaries are [*closed*]{} in the completions, Theorem \[boundaryclosed\]; the chronological completions $\overline{V}$ are always $T_{1}$, Theorem \[6.4\]; and, even though non-Hausdorffness is possible, violation of $T_{2}$ is exclusively restricted to points at the boundary, Theorem \[non-hausdorff\]. Section \[comparing\] has been devoted to emphasize the optimal character of our approach by comparing it with some previous approaches, see Theorem \[optm\] and discussion below. In Section \[applications\] we have shown the utility of our approach in Causality Theory by characterizing two levels of the causal ladder in terms of the chronological completion: [*global hyperbolicity*]{}, Theorem \[globhyp\]; and [*causal simplicity*]{}, Theorem \[causallysimple\]. Finally, in Section \[examples\] we have described some simple examples illustrating the results and properties stated in previous sections. We have also checked our construction in two important classes of physical spacetimes, [*standard static spacetimes*]{} and [*plane wave solutions*]{}. Before concluding this introduction, we remark that our approach does not include considerations about causal, but non-chronological, relations. We have omitted them because the essential part of the causal structure is exclusively carried out by the chronology of the spacetime. Thus, only chronological relations have been considered here, and so, we have gained in simplicity. Nevertheless, the inclusion of purely causal relations into our framework may be the subject of future investigation. Preliminaries ============= Our construction is intended to be exclusively based on the chronological structure of the spacetime. In order to stress this idea, throughout this paper we will work on the simplest mathematical object carrying this structure; the so called [*chronological set*]{} (first introduced in [@H1]). \[chronrel\] A [*chronological set*]{} is a set $X$ with a relation $\ll$ (called [*chronological relation*]{}) such that $\ll$ is transitive and non-reflexive ($x\not\ll x$), there are no isolates (everything is related chronologically to something), and $X$ has a countable set $D$ which is dense: if $x\ll y$ then for some $d\in D$, $x\ll d\ll y$. When working on a chronological set $X$, the role of future timelike curves in a spacetime is now played by [*future chains*]{}: sequence $\varsigma=\{x_{n}\}\subset X$ obeying $x_{1}\ll \cdots\ll x_{n}\ll x_{n+1}\ll\cdots$. As in spacetimes, a subset $P\subset X$ is called a [*past*]{} set if it coincides with its past, that is, $P=I^{-}[P]:=\{x\in X: x\ll x'\;\hbox{for some}\; x'\in P\}$[^1]. Given a subset $S\subset X$, we define the [*common past*]{} of $S$ as $\downarrow S:=I^{-}[\{x\in X:\;\; x\ll x'\;\;\forall x'\in S\}]$. A past set that cannot be written as the union of two proper subsets both of which are also past sets is called an [*indecomposable past*]{} set IP. (Here, the emptyset $\emptyset$ will be assumed to be a past set which is not indecomposable.) An IP which does not coincide with the past of any point in $X$ is called a [*terminal indecomposable past set*]{} TIP. Otherwise, it is called a [*proper indecomposable past set*]{} PIP. In a spacetime the past of a point is always indecomposable, however it is easy to give examples of chronological sets where this does not happen (Example \[yo\]). See Figure 1 for an illustration of these definitions. {width="6cm"} The following adaptation of [@H2 Prop. 4.1] shows that any past set admits an useful decomposition in terms of IPs: \[maximality\] Let $X$ be a chronological set. Every past set $\emptyset\neq P\subset X$ can be written as $P=\cup_{\alpha}P_{\alpha}$, where $\{P_{\alpha}\}_{\alpha}$ is the set of all maximal (under the inclusion relation) IPs included in P. [*Proof.*]{} Consider an arbitrary point $x\in P\neq\emptyset$. Let ${\cal A}_{x}$ be the set of all IPs included in $P$ which contain $x$, endowed with the partial order of inclusion. Since $P$ is a past set, we can construct inductively a future chain $c$ starting at $x$ and entirely contained in $P$. In particular, the past of $c$ is an IP in ${\cal A}_{x}\neq\emptyset$. On the other hand, consider $\{P_{i}\}_{i\in I}\subset {\cal A}_{x}$, $I$ a well-ordered index set with $P_{i}\subset P_{k}$ for $i\leq k$. Then $\cup_{i}P_{i}$ is clearly an IP into $P$ which also contains $x$. Whence, $\cup_{i}P_{i}$ is an upper bound in ${\cal A}_{x}$ for $\{P_{i}\}_{i\in I}$. Therefore, Zorn’s Lemma[^2] ensures the existence of a maximal IP $P_{x}$ into $P$ which contains $x$.   ------------------------------------------------------------------------ Proposition \[maximality\] justifies now the following definition: \[decomposition\] Given a past set $\emptyset\neq P$ in a chronological set $X$, the set ${\rm dec}(P):=\{P_{\alpha}\}_{\alpha}$ of all maximal IPs included in $P$ is called the [*decomposition*]{} of $P$. By convention, we will assume $dec(\emptyset)=\emptyset$. Denote by $\hat{X}$ the set of all IPs of $X$. If $X$ is [*past-regular*]{} (i.e. $I^{-}(x)$ is IP for all $x\in X$) and [*past-distinguishing*]{} (i.e. $I^{-}(x)=I^{-}(x')$ implies $x=x'$), then $\hat{X}$ is called the [*future chronological completion*]{} of $X$. In this case, $\hat{X}$ can be endowed with a structure of chronological set and the map $x\mapsto I^{-}(x)$ injects $X$ into $\hat{X}$. Therefore, we can write $\hat{X}=X\cup \hat{\partial}(X)$, $\hat{\partial}(X)$ being the set of all TIPs of $X$, which is called the [*future chronological boundary*]{} of $X$. More details about the future chronological completion can be found in [@H1]. It is possible to endow a chronological set with a topology. The heart of the [*future chronological topology*]{}, firstly introduced in [@H2], is the following limit operator $\hat{L}$: \[hhat\] Given a sequence $\sigma=\{P_{n}\}$ of past sets in $X$, an IP $P\subset X$ satisfies $P\in\hat{L}(\sigma)$ if - $P\subset {\rm LI}(P_{n})$ and - $P$ is maximal $IP$ within ${\rm LS}(P_{n})$, where [*LI*]{} and [*LS*]{} denote the standard inferior and superior limits of sets: $$\begin{array}{l} {\rm LI}(P_{n})=\liminf_{n\rightarrow\infty}(P_{n})=\cup_{n=1}^{\infty}\cap_{k=n}^{\infty}P_{k} \\ {\rm LS}(P_{n})=\limsup_{n\rightarrow\infty}(P_{n})=\cap_{n=1}^{\infty}\cup_{k=n}^{\infty}P_{k}. \end{array}$$ The limit operator $\hat{L}$ was introduced in [@H2] in a different way. However, it is not hard to show that both definitions are equivalent (see [@H5]). Then, the [*future chronological topology*]{} ($\;\widehat{}\;$-[*topology*]{}) of $X$ is introduced by defining the [*closed sets*]{} as those subsets $C\subset X$ such that $\hat{L}(\sigma)\subset C$ for any sequence $\sigma\subset C$[^3]. With this definition, $\hat{L}(\sigma)$ is to be thought of as first-order limits of $\sigma$. In the particular case of $X=V$ being a strongly causal spacetime, the $\hat{L}$-limit of a sequence coincides with the limit with respect to the manifold topology: \[2.3\] Let $V$ be a strongly causal spacetime. For any sequence $\sigma=\{p_{n}\}\subset V$, a point $p$ is $\hat{L}$-limit of $\sigma$ if and only if it is the limit of $\sigma$ with the topology of the manifold. [*Proof.*]{} See [@H2 Theorem 2.3]. Of course, the dual notions of the concepts introduced here (past chain, future set, $\uparrow S$, IF, TIF, PIF, $\check{X}$, $\check{L}$...), and the corresponding results, can be defined and proved just by interchanging the roles of past and future. We finish this section with a remarkable result coming from [@S1 Prop. 5.1]. Previously, we recall the following definition: \[Szab\] If $P$ is maximal IP into $\downarrow F$ and $F$ is maximal IF into $\uparrow P$ then we say that $P$, $F$ are [*S-related*]{}, $P\sim_{S} F$ (see Figure 1). \[propS\] Let $V$ be a strongly causal spacetime. The unique S-relations involving proper indecomposable sets in $V$ are $I^{-}(p)\sim_{S} I^{+}(p)$ for all $p\in V$. Completing Chronological Sets {#chron} ============================= A central property to be satisfied by any space $\overline{X}$ intended to be a completion of $X$ is that any chain in $X$ admits some “limit” in $\overline{X}$. So, a natural strategy for completing a chronological set consists of adding to $X$ “ideal endpoints” associated to every “endless” chain in $X$. In order to develop this idea, we are going to restrict our attention to [*weakly distinguishing*]{} chronological sets; that is, those chronological sets satisfying that any two points with the same past and future must coincide. Observe that this condition is not very restrictive at all, since it is satisfied by any strongly causal spacetime. Denote by $X_{p}$, $X_{f}$ the sets of all past and future sets of $X$, resp. Then, the map $$\label{injection} \begin{array}{rl} {\bf i}: & X\rightarrow X_{p}\times X_{f} \\ & x\mapsto (I^{-}(x),I^{+}(x)) \end{array}$$ injects $X$ into $X_{p}\times X_{f}$ in a natural way. This injection, joined to the fact that our construction must be exclusively based on the chronological structure of $X$, makes natural to conceive $\overline{X}$ as verifying $${\bf i}[X]\subset\overline{X}\subset X_{p}\times X_{f}.$$ So, if we want to completely determine $\overline{X}$, we need to establish which elements of $X_{p}\times X_{f}$ belong to the completion, or, equivalently, which past and future sets must be paired to form every element of $\overline{X}$. According to the central idea suggested at the beginning of this section, this will be done by formalizing the notion of the “endpoint” of a chain. To this aim, we are not going to define a topology on $X_{p}\times X_{f}$. Instead, we will directly deduce a reasonable notion of “endpoint”, which will be justified a posteriori by showing that it is compatible with the topology for $\overline{X}$ suggested in Section \[topology\]. Consider a future chain $\varsigma=\{x_{n}\}\subset X$ and assume that it “approaches” to some $(P,F)\in X_{p}\times X_{f}$, where $P$ and $F$ represent the past and future (computed in $X$) of the limit point. If a sequence $\{p_{n}\}$ converges to some point $p$ in a spacetime, then every point in the past of $p$ is eventually in the past of $p_{n}$. Therefore, if we translate this property to our situation, we should obtain $P\subset I^{-}[\varsigma]$. Moreover, since $\varsigma$ is a future chain “approaching” to $(P,F)$, it becomes natural to assume $\varsigma\subset P$. In particular, $I^{-}[\varsigma]\subset P$, and thus, $P=I^{-}[\varsigma]$. On the other hand, by transitivity with respect to $(P,F)$, we should also expect $F\subset \uparrow P$. Of course, there is no reason to impose $F =\uparrow P$; however, arguing by analogy to what happens in spacetimes, the fact that $\varsigma$ is “approaching” to $(P,F)$ also leads to strengthen the inclusion $F\subset \uparrow P$ by assuming that any element in ${\rm dec}(F)$ is maximal in $\uparrow P$, or, equivalently, ${\rm dec}(F)\subset \check{L}(\varsigma)$. Obviously, dual conditions are deduced in the case of $\varsigma$ being a past chain approaching to $(P,F)$. Summarizing, we suggest the following definition: \[eend\] A pair $(P,F)\in X_{p}\times X_{f}$ is [*endpoint*]{} of a future (resp. past) chain $\varsigma\subset X$ if $$\label{c1} P=I^{-}[\varsigma],\quad {\rm dec}(F)\subset\check{L}(\varsigma)\quad\qquad (\hbox{resp.}\quad {\rm dec}(P)\subset\hat{L}(\varsigma),\quad F=I^{+}[\varsigma]).$$ We will denote by $X^{end}$ the subset of $X_{p}\times X_{f}$ formed by the union of ${\bf i}[X]$ with all the endpoints of every chain in $X$. Now, we are in a position to formulate the notion of [*completion*]{} for a chronological set: \[completion\] Let $X$ be a weakly distinguished chronological set. A set $\overline{X}$ satisfying $${\bf i}[X]\subset\overline{X}\subset X^{end}(\subset X_{p}\times X_{f})$$ is called a [*completion*]{} of $X$ if any chain in $X$ has some endpoint in $\overline{X}$. Then, the [*boundary*]{} of $X$ in $\overline{X}$ is defined as $\partial(X):=\overline{X}\setminus {\bf i}[X]$. According to this definition, a chronological set will admit different completions (see Example \[ñ\]). We will denote by ${\cal C}_{X}$ the set of all these completions. A well-known example of completion covered by Definition \[completion\] is the GKP [*pre-completion*]{} of spacetimes, first introduced in [@GKP]. In fact, the pre-completion $V^{\sharp}$ of a strongly causal spacetime $V$ can be seen as formed by adding to ${\bf i}[V]$ the endpoints $(I^{-}[\varsigma],\emptyset)$ (resp. $(\emptyset,I^{+}[\varsigma])$) for every endless future (resp. past) chain $\varsigma$[^4]. An alternative completion is formed by adding to ${\bf i}[V]$ the endpoints $(P,F)$ given by $$\label{kkk} P=I^{-}[{\rm LI}(I^{-}(x_{n}))]\quad\hbox{and}\quad F=I^{+}[{\rm LI}(I^{+}(x_{n}))]$$ for every endless chain $\varsigma=\{x_{n}\}$. In this case, the resulting space $V^{\flat}$, which, in general, is different from $V^{\sharp}$, may contain pairs (P,F) with some component $P$ or $F$ not necessarily indecomposable (see Example \[1\]). Notice also that $X^{end}$ is another example of completion, which, indeed, contains any other completion of the chronological set $X$. Observe that Definition \[completion\] is far from being accurate: for example, it does not avoid the possibility of having completions which remain as completions when some boundary point is removed (see Example \[ñ\]). Of course, we can eliminate this possibility by hand in Definition \[completion\]; however, this is not sufficient for ensuring that there are no spurious ideal points at the boundary: for example, the GKP pre-completion $V^{\sharp}$, which does not fall under the previous possibility, sometimes attaches two [*different*]{} ideal points where we would intuitively expect only one (see Example \[ñ\]). We will postpone to Section \[universality\] the introduction of a non-trivial notion of minimal completion, the so-called [*chronological completion*]{}. Even though many completions included in Definition \[completion\] are not optimal, they still satisfy enough properties to justify the name of “completions” for these constructions. The next three sections are devoted to analyze these properties in depth. To this aim, the definition and characterizations below will be very useful: \[gener\] Let $X$ be a chronological set. A pair $(P,F)\in X_{p}\times X_{f}$ is [*generated*]{} by a chain $\varsigma=\{x_{n}\}\subset X$ if equalities (\[kkk\]) hold. Of course, every chain $\varsigma$ in $X$ generates an unique pair $(P,F)\in X_{p}\times X_{f}$. \[carutil\] Let $X$ be a chronological set and consider a chain $\varsigma=\{x_{n}\}\subset X$ and a pair $(P,F)\in X_{p}\times X_{f}$. Then, the following statements are equivalent: - $(P,F)$ is generated by $\varsigma$; - there exists a countable dense set $D\subset X$ such that $$\label{conan} P\cap D={\rm LI}(I^{-}(x_{n}))\cap D\qquad\hbox{and}\qquad F\cap D={\rm LI}(I^{+}(x_{n}))\cap D;$$ - $(P,F)$ satisfies $$\begin{array}{c} \hat{L}(\varsigma)={\rm dec}(P)\qquad\hbox{and}\qquad \check{L}(\varsigma)={\rm dec}(F). \end{array}$$ [*Proof.*]{} We will follow this scheme: first, we will prove [*(iii)*]{}$\Rightarrow$[*(ii)*]{}; then, [*(i)*]{}$\Rightarrow$[*(iii)*]{}; and, finally, [*(ii)*]{}$\Rightarrow$[*(i)*]{}. In order to prove [*(iii)*]{}$\Rightarrow$[*(ii)*]{}, first observe that $\hat{L}(\varsigma)={\rm dec}(P)$ and $\check{L}(\varsigma)={\rm dec}(F)$ imply $$P\subset {\rm LI}(I^{-}(x_{n}))\quad\hbox{and}\quad F\subset {\rm LI}(I^{+}(x_{n})).$$ Therefore, we only need to prove that $$\label{already} P\cap D\supset {\rm LI}(I^{-}(x_{n}))\cap D\qquad\hbox{and}\qquad F\cap D\supset {\rm LI}(I^{+}(x_{n}))\cap D$$ for some countable dense set $D\subset X$. To this aim, take any countable dense set $D'\subset X$ and define $$D:=D'\setminus D_{0},$$ with $D_{0}=\{d\in D':\; I^{-}(d)\subset P\;\;\hbox{but}\;\; d\not\in P\}$. In order to prove the density of $D$, consider $y\ll y'\in X$. Since $D'$ is dense, there exists $d\in D'$ such that $y\ll d\ll y'$. If $d\not\in D_{0}$, necessarily $d\in D$ and we finish. Otherwise, consider $y\ll d$ and take $d'\in D'$ such that $y\ll d'\ll d$. Then, necessarily $d'\not\in D_{0}$ since $d'\in I^{-}(d)\subset P$. Therefore, $d'\in D$, and thus, $D$ is dense in $X$. For the first inclusion in (\[already\]), assume by contradiction the existence of $d\in {\rm LI}(I^{-}(x_{n}))\cap D$ such that $d\not\in P\cap D$. From the definition of $D$, necessarily $I^{-}(d)\not\subset P$. Therefore, there exists $x\in I^{-}(d)\setminus P$. In particular, $x\in I^{-}[{\rm LI}(I^{-}(x_{n}))]\neq\emptyset$, and thus, Proposition \[maximality\] ensures the existence of a maximal IP $P_{x}$ in $I^{-}[{\rm LI}(I^{-}(x_{n}))]$ containing $x$. Taking into account that $\varsigma$ is a chain, $P_{x}$ is also maximal in ${\rm LS}(I^{-}(x_{n}))$. Therefore, $P_{x}\in \hat{L}(\varsigma)$, which contradicts the equality $\hat{L}(\varsigma)={\rm dec}(P)$. In conclusion, the first inclusion in (\[already\]) holds. We can repeat the same reasoning for the future, but taking the set $D$ instead of $D'$ as an initial countable dense set, and removing the elements $d\in D$ such that $I^{+}(d)\subset F$ but $d\not\in F$. Then, the resulting set, which we also denote by $D$, clearly satisfies both inclusions in (\[already\]). In order to prove [*(i)*]{}$\Rightarrow$[*(iii)*]{}, it is clear that $P\subset {\rm LI}(I^{-}(x_{n}))$. So, assume by contradiction that $P_{\alpha_{0}}\in {\rm dec}(P)$ is not maximal in ${\rm LS}(I^{-}(x_{n}))={\rm LI}(I^{-}(x_{n}))$. Then, there exists an IP $P'$ with $P_{\alpha_{0}}\varsubsetneq P'\subset {\rm LI}(I^{-}(x_{n}))$. In particular, $P'\not\subset P$, and thus, there exist $x,x'\in P'\setminus P$ such that $x\ll x'$. As $x'\in {\rm LI}(I^{-}(x_{n}))$, necessarily $x\in I^{-}[{\rm LI}(I^{-}(x_{n}))]\setminus P$, which contradicts that $P=I^{-}[{\rm LI}(I^{-}(x_{n}))]$. Therefore, any $P_{\alpha}\in {\rm dec}(P)$ is maximal in ${\rm LS}(I^{-}(x_{n}))$, and thus, ${\rm dec}(P)\subset \hat{L}(\varsigma)$. To prove $\hat{L}(\varsigma)\subset {\rm dec}(P)$, assume by contradiction the existence of an IP $P'\in\hat{L}(\varsigma)$ such that $P'\not\in {\rm dec}(P)$. Then, necessarily $P'\not\subset P$, since otherwise $P'$ would be maximal in $P$ (recall that $P=I^{-}[{\rm LI}(I^{-}(x_{n}))]$), and thus, $P'\in {\rm dec}(P)$. Reasoning as in the previous paragraph, there exist $x,x'\in P'\setminus P$ such that $x\ll x'$. In particular, $x'\in {\rm LI}(I^{-}(x_{n}))$. Therefore, $x\in I^{-}[{\rm LI}(I^{-}(x_{n}))]\setminus P$, which contradicts that $P=I^{-}[{\rm LI}(I^{-}(x_{n}))]$. In conclusion, $\hat{L}(\varsigma)={\rm dec}(P)$. Finally, an analogous reasoning proves that $\check{L}(\varsigma)={\rm dec}(F)$. In order to prove [*(ii)*]{}$\Rightarrow$[*(i)*]{}, assume by contradiction that $P\not\subset I^{-}[{\rm LI}(I^{-}(x_{n}))]$. Since $P$ is a past set, necessarily $P\not\subset {\rm LI}(I^{-}(x_{n}))$. Therefore, there exist $x,x'\in P\setminus {\rm LI}(I^{-}(x_{n}))$ and $d\in D$ such that $x\ll d\ll x'$. In particular, $d\in P$ but $d\not\in {\rm LI}(I^{-}(x_{n}))$. Whence, $d$ contradicts the first equality in (\[conan\]), and thus, $P\subset I^{-}[{\rm LI}(I^{-}(x_{n}))]$. For the other inclusion, assume that $x\in I^{-}[{\rm LI}(I^{-}(x_{n}))]$. This means $x\ll x'$ for certain $x'\in {\rm LI}(I^{-}(x_{n}))$. Therefore, there exists $d\in D$ with $x\ll d\ll x'$. In particular, $d\in {\rm LI}(I^{-}(x_{n}))\cap D$, which, joined to the first equality in (\[conan\]), implies $x\ll d\in P$. Whence, $I^{-}[{\rm LI}(I^{-}(x_{n}))]\subset P$, and the equality follows. Finally, an analogous reasoning also shows $F=I^{+}[{\rm LI}(I^{+}(x_{n}))]$.   ------------------------------------------------------------------------ In the next Sections \[chronol\]–\[topology\] by $\overline{X}$ we will understand [*any*]{} completion of $X$, according to Definition \[completion\]. The Completions as Chronological Sets {#chronol} ===================================== Now that we have introduced a family of completions ${\cal C}_{X}$ for any weakly distinguishing chronological set $(X,\ll)$, the next step consists of endowing any completion $\overline{X}\in {\cal C}_{X}$ with a structure of weakly distinguishing chronological set, such that $X$ becomes densely and chronologically embedded into $\overline{X}$ via the injection [**i**]{} (see (\[injection\])). To be more precise, let us introduce some definitions: A bijection $f:X\rightarrow X'$ between two chronological sets $(X,\ll)$, $(X',\ll')$ is a [*(chronological) isomorphism*]{} if $f$ and $f^{-1}$ preserve the chronological relations. When $f$ is only injective but the image $f(X)\subset X'$ endowed with $\ll'$ is still isomorph to $(X,\ll)$ via $f$, we say that $f$ is a [*(chronological) embedding*]{} of $(X,\ll)$ into $(X',\ll')$. Consider the relation $$(P,F)\overline{\ll} (P',F')\quad\hbox{iff}\quad F\cap P'\neq\emptyset,\qquad\quad \forall (P,F),(P',F')\in \overline{X}$$ (first introduced in [@S1], and used later in [@MR]). Then, the following results hold: \[tt8\] If $(X,\ll)$ is a weakly distinguishing chronological set then $(\overline{X},\overline{\ll})$ is also a chronological set. Moreover, ${\bf i}$ chronologically embeds $(X,\ll)$ into $(\overline{X},\overline{\ll})$ in such a way that ${\bf i}[X]$ is dense in $\overline{X}$. [*Proof.*]{} To prove transitivity, assume $(P,F)\overline{\ll}(P',F')$ and $(P',F')\overline{\ll}(P'',F'')$. Then, there exist $x\in F\cap P'$ and $x'\in F'\cap P''$. Let $\varsigma=\{x_{n}\}\subset X$ be a chain with endpoint $(P',F')$ (if $(P',F')={\bf i}(x_{0})$ for some $x_{0}\in X$, take instead $\varsigma=\{x_{n}\}\equiv \{x_{0}\}\subset X$). Then, $P'\subset {\rm LI}(I^{-}(x_{n}))$, $F'\subset {\rm LI}(I^{+}(x_{n}))$. In particular, for all $n$ big enough $x\ll x_{n}\ll x'$. But $x\in F$ and $x'\in P''$. Hence, $x_{n}\in F\cap P''\neq\emptyset$ for all $n$ big enough, and thus, $(P,F)\overline{\ll} (P'',F'')$. To show that $\overline{\ll}$ is non-reflexive, assume by contradiction that $(P,F)\overline{\ll} (P,F)$. Then, there exists $x\in F\cap P\neq \emptyset$. As before, let $\varsigma=\{x_{n}\}\subset X$ be a chain with endpoint $(P,F)$ (again, if $(P,F)={\bf i}(x_{0})$ for some $x_{0}\in X$, take instead $\varsigma=\{x_{n}\}\equiv \{x_{0}\}\subset X$). Then, $P\subset {\rm LI}(I^{-}(x_{n}))$ and $F\subset {\rm LI}(I^{+}(x_{n}))$. In particular, for all $n$ big enough $x_{n}\ll x\ll x_{n}$. This contradicts that $\ll$ is non-reflexive. To prove that there are no isolates, consider $(P,F)\in \overline{X}$. Assume for example that $x\in P\neq\emptyset$ (if $x\in F\neq\emptyset$, the argument is analogous). As $P$ is a past set, there exists $x'\in P$ such that $x\ll x'$. Then, ${\bf i}(x)\in\overline{X}$ satisfies ${\bf i}(x)\overline{\ll} (P,F)$, since $x'\in I^{+}(x)\cap P\neq\emptyset$. The set $${\bf i}[D]=\{{\bf i}(d): d\in D,\; D\;\hbox{countable dense set of}\; (X,\ll)\}\subset \overline{X}$$ is a countable dense set of $(\overline{X},\overline{\ll})$. In fact, if $(P,F)\overline{\ll}(P',F')$ then $F\cap P'\neq\emptyset$. Therefore, there exist $x,x'\in F\cap P'$ with $x\ll x'$. Let $d\in D$ be such that $x\ll d\ll x'$. Then, $(P,F)\overline{\ll}{\bf i}(d)\overline{\ll} (P',F')$, since $x\in F\cap I^{-}(d)\neq\emptyset$ and $x'\in I^{+}(d)\cap P'\neq\emptyset$. In particular, this also shows that ${\bf i}[X]$ is dense in $\overline{X}$. Finally, we show that $\overline{\ll}$ extends $\ll$ without introducing new chronological relations in ${\bf i}[X]$. Assume first that $x,x'\in X$ satisfy $x\ll x'$. Then, there exists $d\in D$ such that $x\ll d\ll x'$. Therefore, $d\in I^{+}(x)\cap I^{-}(x')\neq \emptyset$ and, thus, ${\bf i}(x)\overline{\ll} {\bf i}(x')$. Assume now that ${\bf i}(x)\overline{\ll} {\bf i}(x')$. Then, there exists $y\in I^{+}(x)\cap I^{-}(x')\neq\emptyset$. Therefore, $x\ll y\ll x'$, and thus, $x\ll x'$.   ------------------------------------------------------------------------ \[tech\] Let $(X,\ll)$ be a weakly distinguishing chronological set. Then,[^5] $$\label{expected} {\bf i}^{-1}[I^{-}((P,F))\cap {\bf i}[X]]=P,\quad {\bf i}^{-1}[I^{+}((P,F))\cap {\bf i}[X]]=F\qquad\forall\, (P,F)\in\overline{X}.$$ In particular, $(\overline{X},\overline{\ll})$ is a weakly distinguishing chronological set. [*Proof.*]{} For the first equality in (\[expected\]) consider $x\in {\bf i}^{-1}[I^{-}((P,F))\cap {\bf i}[X]]$. This means that ${\bf i}(x)\overline{\ll}(P,F)\in\overline{X}$, $x\in X$. Therefore, there exists $x'\in I^{+}(x)\cap P\neq \emptyset$, and thus, $x\ll x'\in P$. In particular, $x\in P$. Conversely, consider $x\in P$. Since $P$ is a past set, there exists $x'\in X$ with $x'\in I^{+}(x)\cap P\neq\emptyset$. Therefore, ${\bf i}(x)\overline{\ll} (P,F)$, and thus, $x\in {\bf i}^{-1}[I^{-}((P,F))\cap {\bf i}[X]]$. The second equality in (\[expected\]) is proved analogously. In order to prove that $(\overline{X},\overline{\ll})$ is weakly distinguishing, assume $I^{-}((P,F))=I^{-}((P',F'))$ and $I^{+}((P,F))=I^{+}((P',F'))$ for some $(P,F), (P',F')\in\overline{X}$. From (\[expected\]) $$\begin{array}{l} P={\bf i}^{-1}[I^{-}((P,F))\cap {\bf i}[X]]={\bf i}^{-1}[I^{-}((P',F'))\cap {\bf i}[X]]=P' \\ F={\bf i}^{-1}[I^{+}((P,F))\cap {\bf i}[X]]={\bf i}^{-1}[I^{+}((P',F'))\cap {\bf i}[X]]=F'. \end{array}$$ Whence, $(P,F)=(P',F')$.   ------------------------------------------------------------------------ The “Complete Character” of the Completions {#joser} =========================================== In previous section we have showed that given any weakly distinguishing chronological set $(X,\ll)$, the pair $(\overline{X},\overline{\ll})$ is also a weakly distinguishing chronological set. So, we can be tempted to repeat the process on $(\overline{X},\overline{\ll})$, and construct a new pair $(\overline{\overline{X}},\overline{\overline{\ll}})$ with $\overline{\overline{\ll}}$ defined by $$({\cal P},{\cal F})\overline{\overline{\ll}}({\cal P'},{\cal F'})\Longleftrightarrow {\cal F}\cap {\cal P'}\neq \emptyset,\qquad\quad\forall\; ({\cal P},{\cal F}), ({\cal P'},{\cal F'})\in \overline{\overline{X}}.$$ In this section we are going to justify that completing a completion is unnecessary, in the sense that any completion is already a “complete” chronological set. To this aim, of course we previously need to introduce a reasonable notion of [*complete*]{} chronological set: \[def-compl\] A weakly distinguishing chronological set $(Y,\ll)$ is [*(chronologically) complete*]{} if ${\bf i}[Y]$ itself is a completion of $Y$, that is, if any chain in $Y$ has some endpoint in ${\bf i}[Y]$. Next, we are going to establish a suitable correspondence between the pairs in $X_{p}\times X_{f}$ and those in $(\overline{X})_{p}\times (\overline{X})_{f}$, for any completion $\overline{X}$ of $X$. Consider the map $$\begin{array}{rl} {\bf j}: & X_{p}\times X_{f}\rightarrow (\overline{X})_{p}\times (\overline{X})_{f} \\ & (P,F)\mapsto (j(P),j(F)), \end{array}$$ where $$j(P):=I^{-}[{\bf i}[P]] \qquad\hbox{and}\qquad j(F):=I^{+}[{\bf i}[F]]$$ (here, $I^{\pm}[\cdot]$ are computed in $(\overline{X},\overline{\ll})$). From (\[expected\]) and the density of ${\bf i}[X]$ into $\overline{X}$, it follows $$\label{nueva} (j(P),j(F))=(I^{-}((P,F)),I^{+}((P,F)))\quad\;\;\forall\; (P,F)\in \overline{X}.$$ Therefore, the map ${\bf j}$ restricted to $\overline{X}\subset X_{p}\times X_{f}$ coincides with the injection ${\bf i}$ for the chronological set $Y=\overline{X}$. Notice also that the inverse map of [**j**]{} is given by: $$\begin{array}{rl} {\bf k}: & (\overline{X})_{p}\times (\overline{X})_{f}\rightarrow X_{p}\times X_{f} \\ & ({\cal P},{\cal F})\mapsto (k({\cal P}),k({\cal F})), \end{array}$$ where $$k({\cal P}):={\bf i}^{-1}[{\cal P}\cap {\bf i}[X]] \qquad\hbox{and}\qquad k({\cal F}):={\bf i}^{-1}[{\cal F}\cap {\bf i}[X]].$$ In fact, first notice that ${\bf k}$ is well-defined. For example, to check that $k({\cal P})$ is a past set, consider $x\in k({\cal P})$. This means ${\bf i}(x)\in {\cal P}$. Since ${\cal P}$ is a past set of $\overline{X}$ and ${\bf i}[X]$ is dense in $\overline{X}$, there exists $x'\in X$ with ${\bf i}(x)\overline{\ll}{\bf i}(x')\in {\cal P}$. In particular, $x'\in {\bf i}^{-1}[{\cal P}\cap {\bf i}[X]]$. Therefore, $x\ll x'\in k({\cal P})$, showing that $k({\cal P})\subset I^{-}[k({\cal P})]$. Conversely, assume now that $x\in I^{-}[k({\cal P})]$. Then, $x\ll x'\in k({\cal P})$, which implies ${\bf i}(x)\ll {\bf i}(x')\in {\cal P}$. As ${\cal P}$ is a past set, it follows ${\bf i}(x)\in {\cal P}$, and thus, $x\in k({\cal P})$. Therefore, $I^{-}[k({\cal P})]\subset k({\cal P})$. It rests to show that ${\bf j}$ and ${\bf k}$ satisfy the identities $$\label{otronumero} {\bf k}\circ {\bf j}=Id_{X_{p}\times X_{f}}, \qquad\qquad {\bf j}\circ {\bf k}=Id_{(\overline{X})_{p}\times(\overline{X})_{f}}.$$ The first identity is clearly equivalent to the equalities: $$\label{*1} P=k(j(P)),\qquad F=k(j(F))\qquad\forall\; (P,F)\in X_{p}\times X_{f}.$$ To prove the first equality in (\[\*1\]), recall that $P$ is a past set and ${\bf i}$ a chronological embedding. Therefore, $x\in P$ iff ${\bf i}(x)\in {\bf i}[P]\subset I^{-}[{\bf i}[P]]=j(P)$. But, ${\bf i}(x)\in {\bf i}[X]$. Whence, $x\in P$ iff ${\bf i}(x)\in j(P)\cap {\bf i}[X]$. In conclusion, $x\in P$ iff $x\in {\bf i}^{-1}[j(P)\cap {\bf i}[X]]=k(j(P))$. The second equality in (\[\*1\]) can be proved analogously. On the other hand, the second identity in (\[otronumero\]) is equivalent to these other equalities: $$\label{*2} {\cal P}=j(k({\cal P})),\qquad {\cal F}=j(k({\cal F}))\qquad\forall\; ({\cal P},{\cal F})\in (\overline{X})_{p}\times (\overline{X})_{f}.$$ To prove the first equality, recall that ${\cal P}$ is a past set and ${\bf i}[X]$ is dense in $\overline{X}$. Therefore, $(P,F)\in {\cal P}$ iff there exists $x\in X$ with $(P,F)\overline{\ll}{\bf i}(x)\in {\cal P}$. In particular, $(P,F)\in {\cal P}$ iff $(P,F)\overline{\ll}{\bf i}(x)\in {\bf i}[k({\cal P})]$. Therefore, $(P,F)\in {\cal P}$ iff $(P,F)\in j(k({\cal P}))$. The second equality in (\[\*2\]) can be proved analogously. With these tools, we are now ready to prove Theorem \[comp\] below. The hard part of this proof has been extracted in the following lemma: \[completeness\] Let $\overline{X}$ be a completion of a weakly distinguishing chronological set $(X,\ll)$. - If $(P,F)\in X_{p}\times X_{f}$ is an endpoint of a chain $\delta=\{x_{i}\}\subset X$, then the pair $(j(P),j(F))\in (\overline{X})_{p}\times (\overline{X})_{f}$ is an endpoint of the chain ${\bf i}[\delta]=\{{\bf i}(x_{i})\}_{i}\subset \overline{X}$. - Given a chain $\varsigma\subset\overline{X}$ there exists another chain $\delta\subset X$ such that ${\bf i}[\delta]\subset\overline{X}$ and $\varsigma$ have the same endpoints. [*Proof.*]{} For [*(i)*]{}, we assume without restriction that $(P,F)$ is an endpoint of a future chain $\delta=\{x_{i}\}\subset X$. From Definitions \[eend\], \[gener\], this implies $$\label{5.0} P=P'\quad\hbox{and}\quad {\rm dec}(F)\subset {\rm dec}(F'),$$ for $(P',F')$ being the pair generated by $\delta$. From Proposition \[carutil\] there exists a countable dense set $D\subset X$ such that $$\label{5.4} \begin{array}{l} P'\cap D={\rm LI}(I^{-}(x_{i}))\cap D \\ F'\cap D={\rm LI}(I^{+}(x_{i}))\cap D. \end{array}$$ Taking into account that ${\bf i}:X\hookrightarrow\overline{X}$ is a dense and chronological embedding, from (\[5.0\]) and (\[5.4\]) we easily obtain $$\label{k1} j(P)=j(P')\quad\hbox{and}\quad {\rm dec}(j(F))\subset {\rm dec}(j(F')),$$ with $$\label{k2} \begin{array}{l} j(P')\cap {\bf i}[D]={\rm LI}(I^{-}({\bf i}(x_{i})))\cap {\bf i}[D] \\ j(F')\cap {\bf i}[D]={\rm LI}(I^{+}({\bf i}(x_{i})))\cap {\bf i}[D]. \end{array}$$ From (\[k2\]) and Proposition \[carutil\], we deduce that $(j(P'),j(F'))$ is generated by ${\bf i}[\delta]\subset\overline{X}$. This joined to (\[k1\]) proves that $(j(P),j(F))$ is an endpoint of ${\bf i}[\delta]$ (recall again Definitions \[eend\], \[gener\]). In order to prove [*(ii)*]{}, we can assume without restriction that $\varsigma=\{(P_{n},F_{n})\}_{n}\subset\partial(X)$ is a future chain. Let $\varsigma^{o}=\{(P_{n}^{o},F_{n}^{o})\}\subset X^{end}$ be a future chain formed by pairs $(P_{n}^{o},F_{n}^{o})$ generated by some chain $\varsigma^{n}=\{x^{n}_{m}\}_{m}\subset X$ admitting some endpoint equal to $(P_{n},F_{n})$. In particular, $$\label{true} P_{n}\subset P_{n}^{o}\quad\hbox{and}\quad F_{n}\subset F_{n}^{o}\quad \hbox{for all}\;\; n.$$ Denote by $({\cal P}^{o},{\cal F}^{o})\in (X^{end})_{p}\times (X^{end})_{f}$ the pair generated by $\varsigma^{o}$. First we are going to prove the existence of some future chain $\delta=\{x_{i}\}\subset X$ generating $(k({\cal P}^{o}),k({\cal F}^{o}))$. From Proposition \[carutil\] and (\[nueva\]) we have $$\label{005} \begin{array}{l} {\cal P}^{o}\cap {\bf i}[D]={\rm LI}(I^{-}((P^{o}_{n},F^{o}_{n})))\cap {\bf i}[D]={\rm LI}(j(P^{o}_{n}))\cap {\bf i}[D] \\ {\cal F}^{o}\cap {\bf i}[D]={\rm LI}(I^{+}((P^{o}_{n},F^{o}_{n})))\cap {\bf i}[D]={\rm LI}(j(F^{o}_{n}))\cap {\bf i}[D]. \end{array}$$ Applying ${\bf i}^{-1}$ to (\[005\]), from (\[\*1\]) we deduce $$\label{007} \begin{array}{l} k({\cal P}^{o})\cap D={\rm LI}(k(j(P^{o}_{n})))\cap D={\rm LI}(P^{o}_{n})\cap D \\ k({\cal F}^{o})\cap D={\rm LI}(k(j(F^{o}_{n})))\cap D={\rm LI}(F^{o}_{n})\cap D. \end{array}$$ Therefore, if $\varsigma^{o}\subset {\bf i}[X]$ then $\delta={\bf i}^{-1}[\varsigma^{o}]$ is the required sequence. Otherwise, observe that chains $\varsigma^{n}=\{x^{n}_{m}\}_{m}\subset X$ satisfy $$\label{gota} P^{o}_{n}=I^{-}[{\rm LI}(I^{-}(x^{n}_{m}))]\;\;\; \hbox{and}\;\;\; F^{o}_{n}=I^{+}[{\rm LI}(I^{+}(x^{n}_{m}))]$$ (recall Definition \[gener\]). In order to construct the announced chain $\delta=\{x_{i}\}_{i}\subset X$, we will argue inductively: [*Step 1.*]{} Consider $d_{1}\in D$. If $d_{1}\in k({\cal P}^{o})$ (resp. $d_{1}\in k({\cal F}^{o})$), from (\[007\]) we can define a sequence $\{n^{1}_{k}\}_{k}\subset {{\mathbb N}}$ by removing from $\{n\}_{n}$ those elements $n$ with $d_{1}\not\in P^{o}_{n}$ (resp. $d_{1}\not\in F^{o}_{n}$). Moreover, from (\[gota\]) we can construct a sequence $\{m^{1}_{k,l}\}_{l}\subset {{\mathbb N}}$ by removing from $\{m\}_{m}$ those elements $m$ with $d_{1}\not\ll x^{n^{1}_{k}}_{m}$ (resp. $x^{n^{1}_{k}}_{m}\not\ll d_{1}$). With these definitions, $d_{1}\in I^{-}(x^{n^{1}_{k}}_{m^{1}_{k,l}})$ (resp. $d_{1}\in I^{+}(x^{n^{1}_{k}}_{m^{1}_{k,l}})$) for all $k,l$. If $d_{1}\not\in k({\cal P}^{o})\cup k({\cal F}^{o})$ define $\{n^{1}_{k}\}_{k}\equiv \{n\}_{n}$, $\{m^{1}_{k,l}\}_{l}\equiv \{m\}_{m}$. [*Step 2:*]{} Assume now that $\{n^{i}_{k}\}_{k}, \{m^{i}_{k,l}\}_{l}\subset{{\mathbb N}}$ have been defined for certain $i$. Consider $d_{i+1}\in D$. If $d_{i+1}\in k({\cal P}^{o})$ (resp. $d_{i+1}\in k({\cal F}^{o})$), from (\[007\]) we can define a sequence $\{n^{i+1}_{k}\}_{k}\subset {{\mathbb N}}$ by removing from $\{n^{i}_{k}\}_{k}$ those elements $n^{i}_{k}$ with $d_{i+1}\not\in P^{o}_{n^{i}_{k}}$ (resp. $d_{i+1}\not\in F^{o}_{n^{i}_{k}}$). Moreover, from (\[gota\]) we can construct a sequence $\{m^{i+1}_{k,l}\}_{l}\subset {{\mathbb N}}$ by removing from $\{m^{i}_{k,l}\}_{l}$ those elements $m^{i}_{k,l}$ with $d_{i+1}\not\ll x^{n^{i+1}_{k}}_{m^{i}_{k,l}}$ (resp. $x^{n^{i+1}_{k}}_{m^{i}_{k,l}}\not\ll d_{i+1}$). With these definitions, $d_{i+1}\in I^{-}(x^{n^{i+1}_{k}}_{m^{i+1}_{k,l}})$ (resp. $d_{i+1}\in I^{+}(x^{n^{i+1}_{k}}_{m^{i+1}_{k,l}})$) for all $k,l$. If $d_{i+1}\not\in k({\cal P}^{o})\cup k({\cal F}^{o})$ define $\{n^{i+1}_{k}\}_{k}\equiv \{n^{i}_{k}\}_{k}$, $\{m^{i+1}_{k,l}\}_{l}\equiv \{m^{i}_{k,l}\}_{l}$. Therefore, we can construct by induction $\{n^{i}_{k}\}_{k}$, $\{m^{i}_{k,l}\}_{l}$ for all $i\in{{\mathbb N}}$. Moreover, it is possible to choose $l(i)$ in such a way that $\delta=\{x_{i}\}_{i}\equiv \{x^{n^{i}_{i}}_{m^{i}_{i,l(i)}}\}_{i}$ is a future chain contained in $k({\cal P}^{o})$. With this definition of $\delta$, the following inclusions hold: $$\label{assumption2} \begin{array}{c} {\rm LI}(P^{o}_{n})\cap D\subset {\rm LI}(I^{-}(x_{i}))\cap D\subset {\rm LI}(P^{o}_{n})\cap D \\ {\rm LI}(F^{o}_{n})\cap D\subset {\rm LI}(I^{+}(x_{i}))\cap D\subset {\rm LI}(F^{o}_{n})\cap D. \end{array}$$ In fact, assume $d\in {\rm LI}(P^{o}_{n})\cap D$. From (\[007\]) and previous construction, necessarily $d\ll x^{n^{i}_{k}}_{m^{i}_{k,l}}$ for all $i$ big enough and all $k,l$. In particular, $d\ll x_{i}$ for all $i$ big enough, and thus, $d\in {\rm LI}(I^{-}(x_{i}))\cap D$. This proves the first inclusion in (\[assumption2\]). Assume now $d\in {\rm LI}(I^{-}(x_{i}))\cap D$. As $\delta\subset k({\cal P}^{o})$, necessarily $d\in k({\cal P}^{o})$. Therefore, from (\[007\]) we have $d\in {\rm LI}(P^{o}_{n})\cap D$. For the inclusions in the second line of (\[assumption2\]), assume first $d\in {\rm LI}(F^{o}_{n})\cap D$. Reasoning as before, we deduce $d\gg x^{n^{i}_{k}}_{m^{i}_{k,l}}$ for all $i$ big enough and all $k,l$, and thus, $d\in {\rm LI}(I^{+}(x_{i}))\cap D$. Assume now $d\in {\rm LI}(I^{+}(x_{i}))\cap D$. There exists $d_{i_{n}}\in k({\cal P}^{o})\cap F^{o}_{n}\neq\emptyset$ for all $n$. But, $d_{i_{n}}\ll x_{i}$ for all $i\geq i_{n}$. Whence, $d\gg d_{i_{n}}\in F^{o}_{n}$ for all $n$. This proves the last inclusion in (\[assumption2\]). In conclusion, from (\[007\]) and (\[assumption2\]) we have $$\label{isstar} \begin{array}{c} k({\cal P}^{o})\cap D={\rm LI}(I^{-}(x_{i}))\cap D \\ k({\cal F}^{o})\cap D={\rm LI}(I^{+}(x_{i}))\cap D. \end{array}$$ Therefore, from Proposition \[carutil\] we conclude that $(k({\cal P}^{o}),k({\cal F}^{o}))$ is generated by $\delta$. Next, it remains to show that ${\bf i}[\delta]\subset\overline{X}$ and $\varsigma$ have the same endpoints. Taking into account that ${\bf i}$ is a chronological embedding, from (\[\*2\]) and (\[isstar\]) we deduce $$\begin{array}{l} {\cal P}^{o}\cap {\bf i}[D]=jk({\cal P}^{o})\cap {\bf i}[D]={\rm LI}(I^{-}({\bf i}(x_{i})))\cap {\bf i}[D] \\ {\cal F}^{o}\cap {\bf i}[D]=jk({\cal F}^{o})\cap {\bf i}[D]={\rm LI}(I^{+}({\bf i}(x_{i})))\cap {\bf i}[D]. \end{array}$$ Whence, Proposition \[carutil\] ensures that $({\cal P}^{o},{\cal F}^{o})$ is generated by ${\bf i}[\delta]$. Since $({\cal P}^{o},{\cal F}^{o})$ is also generated by $\varsigma^{o}$, for every $i_{0}$, $n_{0}$ there exists $n$, $i$ big enough such that $$\label{112} {\bf i}(x_{i_{0}})\overline{\ll} (P^{o}_{n-1},F^{o}_{n-1})\qquad\hbox{and}\qquad (P^{o}_{n_{0}+1},F^{o}_{n_{0}+1})\overline{\ll} {\bf i}(x_{i}).$$ Moreover, from (\[true\]) and the fact that $\varsigma$ is a future chain, necessarily $$\label{111} (P^{o}_{n-1},F^{o}_{n-1})\overline{\ll} (P_{n},F_{n})\qquad\hbox{and}\qquad (P_{n_{0}},F_{n_{0}})\overline{\ll} (P^{o}_{n_{0}+1},F^{o}_{n_{0}+1}).$$ Therefore, taking into account (\[112\]) and (\[111\]) we have proved that fixed for $i_{0}$, $n_{0}$ there exists $n$, $i$ big enough such that $${\bf i}(x_{i_{0}})\overline{\ll} (P_{n},F_{n})\qquad\hbox{and}\qquad (P_{n_{0}},F_{n_{0}})\overline{\ll} {\bf i}(x_{i}).$$ In conclusion, $\varsigma$ and ${\bf i}[\delta]$ have the same endpoints.   ------------------------------------------------------------------------ Now, the main result of this section can be proved easily: \[comp\] (Completeness). If $(X,\ll)$ is a weakly distinguishing chronological set then $(\overline{X},\overline{\ll})$ is complete. [*Proof.*]{} Given any chain $\varsigma=\{(P_{n},F_{n})\}\subset\overline{X}$ we need to prove that $\varsigma$ has some endpoint in ${\bf j}[\overline{X}]$. From Lemma \[completeness\] [*(ii)*]{} there exists some chain $\delta\subset X$ such that ${\bf i}[\delta]\subset \overline{X}$ and $\varsigma$ have the same endpoints. Let $(P,F)$ be some endpoint of $\delta$ in $\overline{X}$. From Lemma \[completeness\] [*(i)*]{}, $(j(P),j(F))$ is an endpoint of ${\bf i}[\delta]$. Therefore, $(j(P),j(F))\in {\bf j}[\overline{X}]$ is also endpoint of $\varsigma$, as required.   ------------------------------------------------------------------------ The Chronological Topology {#topology} ========================== In order to provide a deeper description of the relation between a chronological set and its boundary, in this section we are going to introduce a topological structure. More precisely, we are going to endow any chronological set with the [*chronological topology*]{}, a non-trivial generalization of the $\;\widehat{}\;$-topology in [@H2]. To this aim, first we are going to define a limit operator $L$ for any chronological set $Y$. We take as a guide property the fact that our topology must turn the endpoints of chains (Definition \[eend\]) into topological limits. So, the following definition, based on a simple generalization of conditions (\[c1\]), becomes natural: \[overline\] Given a sequence $\sigma\subset Y$, we say that $x\in L(\sigma)$ if $${\rm dec}(I^{-}(x))\subset \hat{L}(\sigma)\quad\hbox{and}\quad {\rm dec}(I^{+}(x))\subset \check{L}(\sigma).$$ With this limit operator in hand we can now define the [*closed*]{} sets of $Y$, which determine the [*chronological topology*]{} ($chr$-[*topology*]{}): \[closed\] The [*closed sets*]{} of $Y$ with the $chr$-[*topology*]{} are those subsets $C\subset Y$ such that $L(\sigma)\subset C$ for any sequence $\sigma\subset C$. It is worth noting that the $chr$-topology has been defined for [*any*]{} chronological set. In particular, it is applicable to both strongly causal spacetimes $V$ and their completions $\overline{V}$. So, two natural questions arise: does the manifold topology of a strongly causal spacetime coincide with the $chr$-topology it inherits when considered as a chronological set?; does the manifold topology of a strongly causal spacetime $V$ coincide with the restriction to $V$ of the $chr$-topology of $\overline{V}$? The following two theorems will answer positively to these questions. For the second one, the hypothesis of strong causality (further than weakly distinguishing) becomes essential. \[artu\] Any weakly-distinguishing chronological set $(X,\ll)$ is topologically embedded into $(\overline{X},\overline{\ll})$ via the injection [**i**]{} if both spaces are endowed with the $chr$-topology. [*Proof.*]{} It suffices to show that a point $x\in X$ satisfies $x\in L(\sigma)$ for some sequence $\sigma=\{x_{n}\}\subset X$ if and only if ${\bf i}(x)\in L(\rho)$ for $\rho=\{{\bf i}(x_{n})\}\subset\overline{X}$. Taking into account that ${\bf i}:X\hookrightarrow\overline{X}$ is a dense and chronological embedding, and equalities (\[\*1\]), (\[\*2\]), we deduce: $$\begin{array}{l} P\in {\rm dec}(I^{-}(x)),\;\; P\subset {\rm LI}(I^{-}(x_{n})) \\ P\;\;\hbox{maximal in}\;\; {\rm LS}(I^{-}(x_{n})) \end{array} \Rightarrow \begin{array}{l} j(P)\in {\rm dec}(I^{-}({\bf i}(x))),\;\; j(P)\subset {\rm LI}(I^{-}({\bf i}(x_{n}))) \\ j(P)\;\;\hbox{maximal in}\;\; {\rm LS}(I^{-}({\bf i}(x_{n}))), \end{array}$$ $$\begin{array}{l} {\cal P}\in {\rm dec}(I^{-}({\bf i}(x))),\;\; {\cal P}\subset {\rm LI}(I^{-}({\bf i}(x_{n}))) \\ {\cal P}\;\;\hbox{maximal in}\;\; {\rm LS}(I^{-}({\bf i}(x_{n}))) \end{array} \Rightarrow \begin{array}{l} k({\cal P})\in {\rm dec}(I^{-}(x)),\;\; k({\cal P})\subset {\rm LI}(I^{-}(x_{n})) \\ k({\cal P})\;\;\hbox{maximal in}\;\; {\rm LS}(I^{-}(x_{n})). \end{array}$$ Analogously, we deduce the corresponding implications for the future. Therefore, the thesis follows from (\[nueva\]) and Definitions \[hhat\], \[overline\].   ------------------------------------------------------------------------ \[princ\] The topology of a strongly causal spacetime $V$ as a manifold coincides with the corresponding $chr$-topology. [*Proof.*]{} From Proposition \[2.3\] (and its dual), a point $p\in V$ is the limit of a sequence $\sigma=\{p_{n}\}\subset V$ with the topology of the manifold if and only if $$\label{uyy} I^{-}(p)\in\hat{L}(\sigma)\qquad\hbox{and}\qquad I^{+}(p)\in\check{L}(\sigma).$$ Taking into account that ${\rm dec}(I^{-}(p))=\{I^{-}(p)\}$ and ${\rm dec}(I^{+}(p))=\{I^{+}(p)\}$, conditions (\[uyy\]) can be written as $${\rm dec}(I^{-}(p))\subset \hat{L}(\sigma)\quad\hbox{and}\quad {\rm dec}(I^{+}(p))\subset \check{L}(\sigma).$$ Therefore, from Definition \[overline\], $\sigma$ converges to $p$ with the manifold topology if and only if $p$ is the $L$-limit of $\sigma$.   ------------------------------------------------------------------------ With this topology, we can also prove that the concept of [*endpoint*]{} is compatible with the notion of [*limit*]{} of a chain: \[compatible\] Let $\varsigma=\{x_{n}\}$ be a chain in a weakly distinguishing chronological set $X$. Then, the following statements hold: - If ${\bf i}(x)\in {\bf i}[X]$ is endpoint of $\varsigma$ then $x\in L(\varsigma)$. Moreover, the reciprocal is true if, in addition, $X$ is regular (i.e., past- and future-regular). - If $(P,F)\in\overline{X}$ is an endpoint of $\varsigma$ then $(P,F)\in L(\rho')$ for any subsequence $\rho'\subset \rho=\{{\bf i}(x_{n})\}_{n}\subset {\bf i}[X]$. In particular, ${\bf i}[X]$ is topologically dense in $\overline{X}$. [*Proof.*]{} Statement [*(1)*]{} is a direct consequence of Definitions \[eend\], \[overline\]. For [*(2)*]{}, assume without restriction that $\varsigma$ is a future chain. From Definitions \[eend\], \[gener\], $$\label{?} P=P^{o}\quad\hbox{and}\quad {\rm dec}(F)\subset {\rm dec}(F^{o}),$$ where $(P^{o},F^{o})$ is the pair generated by $\varsigma$. From Proposition \[carutil\] there exists a countable dense set $D\subset X$ such that $$\label{lla} P^{o}\cap D={\rm LI}(I^{-}(x_{n}))\cap D,\qquad F^{o}\cap D={\rm LI}(I^{+}(x_{n}))\cap D.$$ Taking into account that ${\bf i}$ is a chronological embedding, from (\[lla\]) we obtain: $$\label{virt} \begin{array}{l} j(P^{o})\cap {\bf i}[D]={\rm LI}(I^{-}({\bf i}(x_{n})))\cap {\bf i}[D] \\ j(F^{o})\cap {\bf i}[D]={\rm LI}(I^{+}({\bf i}(x_{n})))\cap {\bf i}[D]. \end{array}$$ Moreover, equalities (\[virt\]) also hold for any subsequence $\rho'\subset \rho=\{{\bf i}(x_{n})\}$ since $\rho$ is a chain in $\overline{X}$. This joined to (\[?\]) and Proposition \[carutil\] imply $$\{j(P)\}=\{j(P^{o})\}=\hat{L}(\rho') ,\quad {\rm dec}(j(F))\subset{\rm dec}(j(F^{o}))=\check{L}(\rho')\quad\hbox{for any}\;\; \rho'\subset\rho,$$ and thus, $(P,F)\in L(\rho')$ for any $\rho'\subset\rho$.   ------------------------------------------------------------------------ Finally, as a direct consequence of Theorem \[compatible\] [*(1)*]{} and Theorem \[comp\] we obtain the following result: \[thchar\] If $\varsigma$ is a chain in a complete chronological set $Y$ then $\varsigma$ has some limit in $Y$. In particular, if $Y=\overline{V}$ is a completion for some strongly causal spacetime $V$, then any timelike curve $\gamma(\equiv {\bf i}[\gamma])$ in $V(\equiv {\bf i}[V])$ has some limit in $\overline{V}$. The results obtained so far in this paper show that our construction satisfies essential requirements to provide reasonable completions for any strongly causal spacetime. However, many of these completions are not optimal, in the sense that they include spurious ideal points. As a consequence: the notions of endpoint and limit of chains, even if compatible, are not totally equivalent (Example \[ñ\]), the boundaries may not be closed in the completions (Example \[3\]), the completions may present bad separation properties (Example \[ñ\])... In the next section we are going to show that all these deficiencies disappear when only completions with minimal boundaries are considered. The Chronological Completions {#universality} ============================= In order to look for completions with minimal boundaries, first we delete from ${\cal C}_{X}$ those completions which are still completions when some point of its boundary is removed. Denote by ${\cal C}_{X}^{*}$ the resulting set, which is always non-empty: for example, ${\bf i}[X]$ joined to those pairs of the form $(I^{-}[\varsigma],\emptyset)$ or $(\emptyset,I^{+}[\varsigma])$ for any future or past chain $\varsigma$ without endpoints in ${\bf i}[X]$, is a completion in ${\cal C}_{X}^{*}$. Then, we introduce a partial order relation $\leq$ in ${\cal C}_{X}^{*}$. Roughly speaking, we will say that $\overline{X}^{\imath}$ [*precedes*]{} $\overline{X}^{\jmath}$ if there exists a suitable partition of $\partial^{\jmath}(X)$ by $\partial^{\imath}(X)$. More precisely: \[gh\] Let $X$ be a weakly distinguishing chronological set and consider two completions $\overline{X}^{\imath}, \overline{X}^{\jmath}\in {\cal C}_{X}^{*}$ with $$\partial^{\imath}(X)=\{(P_{i},F_{i}): i\in I\}\qquad\hbox{and}\qquad \partial^{\jmath}(X)=\{(P_{j},F_{j}): j\in J\}.$$ Then, we say $\overline{X}^{\imath}\leq\overline{X}^{\jmath}$ if there exists some partition $\partial^{\jmath}(X)=\cup_{i\in I}S_{i}$, $S_{i}\cap S_{i'}=\emptyset$ if $i\neq i'$, satisfying the following conditions: - if a chain in $X$ has some endpoint in $S_{i}$ then $(P_{i},F_{i})$ is also an endpoint of that chain,[^6] and - for every $i\in I$ such that $S_{i}=\{(P,F)\}$ with $(P,F)$, $(P_{i},F_{i})$ endpoints of the same chains, it is ${\rm dec}(P_{i})\subset {\rm dec}(P)$ and ${\rm dec}(F_{i})\subset {\rm dec}(F)$. With this definition the pair $({\cal C}_{X}^{*},\leq)$ becomes a partially ordered set ([*reflexivity*]{} and [*transitivity*]{} are direct; [*antisymmetry*]{} needs a simple discussion involving several cases). Furthermore, $({\cal C}_{X}^{*},\leq)$ always admits some minimal element: \[ghh\] If $X$ is a weakly distinguishing chronological set then $({\cal C}_{X}^{*},\leq)$ has some minimal element. [*Proof.*]{} We can assume without restriction that any completion $\overline{X}\in {\cal C}_{X}^{*}$ satisfies that any pair $(P,F)\in \partial(X)$ has ${\rm dec}(P)$ and ${\rm dec}(F)$ finite (otherwise, remove from ${\cal C}_{X}^{*}$ those completions which do not verify this property; the minimal elements of the resulting set, which is non-empty, are still minimal elements of ${\cal C}_{X}^{*}$). Consider $\{\overline{X}^{\imath_{\alpha}}\}_{\alpha\in \Lambda}\subset {\cal C}_{X}^{*}$, $\Lambda$ a well-ordered set with $\overline{X}^{\imath_{\alpha}}\leq\overline{X}^{\imath_{\beta}}$ for $\alpha\geq\beta$. Fix any $\alpha_{0}\in\Lambda$, let ${\cal S}$ be a set of chains in $X$ such that $$\partial^{\imath_{\alpha_{0}}}(X)=\{(P^{\varsigma}_{\alpha_{0}},F^{\varsigma}_{\alpha_{0}}): \varsigma\in {\cal S}\},$$ with $(P_{\alpha_{0}}^{\varsigma},F_{\alpha_{0}}^{\varsigma})$ being some endpoint of $\varsigma$ in $\overline{X}^{\imath_{\alpha_{0}}}$. Then, ${\cal S}$ also satisfies $$\partial^{\imath_{\alpha}}(X)=\{(P^{\varsigma}_{\alpha},F^{\varsigma}_{\alpha}): \varsigma\in {\cal S}\},\qquad\hbox{for all}\;\; \alpha\geq\alpha_{0},$$ with $(P^{\varsigma}_{\alpha},F^{\varsigma}_{\alpha})$ being the pre-image in $\partial^{\imath_{\alpha}}(X)$ of $(P^{\varsigma}_{\alpha_{0}},F^{\varsigma}_{\alpha_{0}})$ via some partition of $\partial^{\imath_{\alpha_{0}}}(X)$ by $\partial^{\imath_{\alpha}}(X)$ according to Definition \[gh\]. With these definitions, $\{P^{\varsigma}_{\alpha}\}_{\alpha\geq\alpha_{0}}$, $\{F^{\varsigma}_{\alpha}\}_{\alpha\geq\alpha_{0}}$ are necessarily constants for all $\alpha\geq\alpha^{*}$, for some $\alpha^{*}\in\Lambda$ depending on $\varsigma$. Therefore, the set $$\overline{X}:={\bf i}[X]\cup \{(P^{\varsigma}_{\alpha^{*}},F^{\varsigma}_{\alpha^{*}}): \varsigma\in {\cal S}\}\in {\cal C}_{X}^{*}$$ is a lower bound for $\{\overline{X}^{\imath_{\alpha}}\}_{\alpha\in \Lambda}$, and thus, Zorn’s Lemma ensures the existence of some minimal completion in ${\cal C}_{X}^{*}$.   ------------------------------------------------------------------------ We are now ready to introduce the notion of [*chronological completion*]{}: \[ghhh\] A completion $\overline{X}$ in ${\cal C}_{X}$ is a [*chronological completion*]{} if it is a minimal element of $({\cal C}_{X}^{*},\leq)$. Then, the chronological boundary of $X$ in $\overline{X}$ is defined as $\partial(X):=\overline{X}\setminus {\bf i}[X]$. Even if it is not very common, there are spacetimes admitting different chronological completions (see Example \[yo\]). However, if $X$ is complete then ${\cal C}_{X}^{*}=\{{\bf i}[X]\}$, and thus, ${\bf i}[X]$ is the unique chronological completion of $X$. The next result establishes a series of nice properties (some of them axiomatically imposed in previous approaches) which totally characterize these constructions: \[7.3\] Let $V$ be a strongly causal spacetime. Then, a subset $\partial(V)\subset V_{p}\times V_{f}$ is the chronological boundary associated to some chronological completion $\overline{V}$ of $V$ if and only if the following properties hold: - Every terminal indecomposable set in $V$ is the component of some pair in $\partial(V)$. Moreover, if $(P,F)\in\partial(V)$ then $P$, $F$ are both indecomposable sets if non-empty. - If $P,F\neq\emptyset$ satisfy $(P,F)\in\partial(V)$ then $P\sim_{S} F$. - If $(P,\emptyset)\in\partial(V)$ (resp. $(\emptyset,F)\in\partial(V)$) then $P$ (resp. $F$) is not S-related to anything. - If $(P,F)\in\partial(V)$ then $P$, $F$ are both terminal sets if non-empty. - If $(P,F_{1}), (P,F_{2})\in\partial(V)$, $F_{1}\neq F_{2}$ (resp. $(P_{1},F)$, $(P_{2},F)\in\partial(V)$, $P_{1}\neq P_{2}$) then $F_{i}$ (resp. $P_{i}$), $i=1,2$, do not appear in any other pair of $\partial(V)$. [*Proof.*]{} First, we will prove the implication to the right. [*(1)*]{} If, for example, some TIP $P\neq\emptyset$ is not the component of any pair in $\partial(V)$, then any future chain $\varsigma\subset V$ with $I^{-}[\varsigma]=P$ has no endpoint in $\overline{V}$, which contradicts that $\overline{V}$ is a completion. For the second assertion, assume that $(P,F)\in\partial(V)$ is an endpoint of some future chain $\varsigma\subset V$. From Definition \[eend\], it is $P=I^{-}[\varsigma]$, and thus, $P$ is IP. Assume by contradiction that $F\neq\emptyset$ is not IF. Then, if we replace the pair $(P,F)$ in $\overline{V}$ by $(P,\emptyset)$, the resulting set is still a completion which contradicts the minimal character of $\overline{V}$. [*(2)*]{} By contradiction, assume for example that $(P,F)\in\partial(V)$ is an endpoint of some future chain, but $P$ is not maximal IP into $\downarrow F$. Since $\overline{V}$ is a completion, there exists some past set $P'\neq P$ such that $(P',F)\in\overline{V}$. Therefore, if we replace $(P,F)$ in $\overline{V}$ by $(P,\emptyset)$, the resulting set is still a completion which contradicts the minimal character of $\overline{V}$. [*(3)*]{} By contradiction, assume that $(P,\emptyset)\in\partial(V)$ but $P\sim_{S} F$ for some IF $F$. Then, if we replace $(P,\emptyset)$ in $\overline{V}$ by $(P,F)$, the resulting set is still a completion which contradicts the minimal character of $\overline{V}$. [*(4)*]{} It directly follows from [*(2)*]{}, [*(3)*]{} and Proposition \[propS\]. [*(5)*]{} Assume by contradiction that $(P,F_{1}), (P,F_{2}), (P',F_{1})\in\partial(V)$, with $F_{1}\neq F_{2}$ and $P\neq P'$. Then, if we remove the pair $(P,F_{1})$ from $\overline{V}$, the resulting set is still a completion, and thus, contradicts that $\overline{V}\in {\cal C}_{V}^{*}$. Conversely, consider $\overline{V}:={\bf i}[V]\cup \partial(V)$ with $\partial(V)$ satisfying conditions [*(1)*]{}–[*(5)*]{}. From [*(1)*]{} and [*(2)*]{}, $\overline{V}$ is a completion. From [*(1)*]{}, [*(2)*]{}, [*(3)*]{} and [*(5)*]{}, it is $\overline{V}\in {\cal C}_{V}^{*}$. In order to prove that $\overline{V}$ is minimal in $({\cal C}_{V}^{*},\leq)$, assume that $\overline{V}^{\imath}\leq\overline{V}$ for some completion $\overline{V}^{\imath}\in {\cal C}_{V}^{*}$. Then, there exists some partition $\partial(V)=\cup_{i\in I}S_{i}$, $S_{i}\cap S_{i'}=\emptyset$ if $i\neq i'$, satisfying condition [*(i)*]{}, [*(ii)*]{} in Definition \[gh\]. From [*(1)*]{}, [*(2)*]{} and [*(3)*]{}, if $(P,F)\in S_{i}$ then $P=P_{i}$, $F=F_{i}$. Whence, $S_{i}=\{(P_{i},F_{i})\}$ for all $i\in I$, and thus, $\overline{V}^{\imath}=\overline{V}$.   ------------------------------------------------------------------------ When $X=V$ is a strongly causal spacetime, the chronological completions verify a number of very satisfactory properties. We begin by showing the equivalence between the notions of [*endpoint*]{} and [*limit*]{} of a chain: \[compatible’\] Let $\varsigma=\{p_{n}\}$ be a chain in a strongly causal spacetime $V$. A pair $(P,F)$ is an endpoint of $\varsigma$ in some chronological completion $\overline{V}$ if and only if $(P,F)\in L(\rho')$ for any subsequence $\rho'\subset \rho=\{{\bf i}(p_{n})\}_{n}\subset {\bf i}[V]$. [*Proof.*]{} From Theorem \[compatible\] [*(2)*]{}, we only need to prove the implication to the left. So, assume $(P,F)\in L(\rho)$ for $\rho=\{{\bf i}(p_{n})\}_{n}\subset {\bf i}[V]$. If $P\neq\emptyset$ (resp. $F\neq\emptyset$) and $\varsigma$ is a future (resp. past) chain then $P=I^{-}[\varsigma]$ (resp. $F=I^{+}[\varsigma]$), and thus, the implication directly follows from Definitions \[eend\], \[overline\]. So, assume for example that $(\emptyset,F)\in\partial(V)$ is a $L$-limit of the future chain $\rho$. From Definition \[overline\] and Theorem \[7.3\] [*(1)*]{}, $F$ is maximal IF into $\uparrow I^{-}[\varsigma]$. Let $P'$ be a maximal IP into $\downarrow F$ containing $I^{-}[\varsigma]$. Then, $P'\sim_{S} F$, and thus, Theorem \[7.3\] [*(3)*]{} implies $(\emptyset,F)\not\in \partial(V)$, a contradiction. Whence, this last possibility cannot happen.   ------------------------------------------------------------------------ Moreover, the chronological boundary is always closed in the corresponding chronological completion: \[boundaryclosed\] If $V$ is a strongly causal spacetime then $\partial(V)$ is closed in $\overline{V}$. [*Proof.*]{} By contradiction, assume the existence of a sequence $\sigma=\{(P_{n},F_{n})\}\subset \partial(V)$ such that ${\bf i}(p)\in L(\sigma)$ for some $p\in V$. For every $n$, consider a chain $\varsigma^{n}=\{p^{n}_{m}\}_{m}\subset V$ with endpoint $(P_{n},F_{n})$. Then, $$\label{peli} \hbox{either}\quad \hat{L}(\varsigma^{n})=\{I^{-}[\varsigma^{n}]\}=\{P_{n}\}\quad\hbox{or}\quad \check{L}(\varsigma^{n})=\{I^{+}[\varsigma^{n}]\}=\{F_{n}\},$$ depending on if $\varsigma^{n}$ is either future or past chain, resp. Let $U\subset V$ be a precompact neighborhood of $p$. For every $n$, necessarily $\{p^{n}_{m}\}_{m}\subset V\setminus\overline{U}$ eventually for all $m$. In fact, otherwise $\varsigma^{n}$ converges (up to a subsequence) to certain $r_{n}\in \overline{U}$ with the topology of the manifold, and thus, $I^{-}(r_{n})\in \hat{L}(\varsigma^{n})$ and $I^{+}(r_{n})\in \check{L}(\varsigma^{n})$ (Proposition \[2.3\] and its dual). Therefore, from (\[peli\]), either $P_{n}=I^{-}(r_{n})$ or $F_{n}=I^{+}(r_{n})$, in contradiction with $(P_{n},F_{n})\in\partial(V)$ (recall Theorem \[7.3\] [*(4)*]{}). In conclusion, fixed future and past chains $\varsigma=\{q_{k}\}$, $\varsigma'=\{q'_{k}\}$ such that $I^{-}(p)=I^{-}[\varsigma]$ and $I^{+}(p)=I^{+}[\varsigma']$, we can choose $\{n_{k}\}_{k}$, $\{m_{k}\}_{k}$ satisfying $q_{k}\ll p^{n_{k}}_{m_{k}}\ll q'_{k}$ and $p^{n_{k}}_{m_{k}}\in V\setminus \overline{U}$ for all $k$. Therefore, taking into account that $\varsigma$, $\varsigma'$ converge to $p$ with the topology of the manifold, any sequence of future-directed timelike curves joining $q_{k}$ with $p^{n_{k}}_{m_{k}}$ and then with $q'_{k}$ contradicts the strong causality of $V$.   ------------------------------------------------------------------------ The chronological completions also satisfy reasonably good separation properties. In fact, from Definition \[overline\], Proposition \[propS\] and Theorem \[7.3\] [*(1)*]{}, [*(2)*]{}, [*(3)*]{} every element $(P,F)\in\overline{V}$ is the unique limit in $\overline{V}$ of the sequence constantly equal to $(P,F)$, thus: \[6.4\] If $V$ is a strongly causal spacetime then $\overline{V}$ is $T_{1}$. Notice however that $\overline{V}$ is not always $T_{2}$ (Example \[1\]). The lack of Hausdorffness in chronological completions cannot be attributed to a defect of our particular approach. On the contrary, it seems a remarkable property intrinsic to the causal boundary approach itself (see [@MR] for an interesting discussion on this question). Even if the chronological completions may be non-Hausdorff, there are still some restrictions to the elements of $\overline{V}$ which can be non-Hausdorff related, Theorem \[non-hausdorff\]. In order to prove this result, we need the following proposition: \[intermedio\] Let $\overline{V}$ be a chronological completion of a strongly causal spacetime $V$. If $K\subset V$ is compact in $V$ then ${\bf i}[K]$ is closed in $\overline{V}$. [*Proof.*]{} From Theorems \[artu\], \[princ\], ${\bf i}[K]$ is closed in ${\bf i}[V]$. So, by contradiction, we will assume the existence of $(P,F)\in \partial(V)$, with $(P,F)\in L(\rho)$ for a certain sequence $\rho=\{{\bf i}(p_{n})\}$ such that $\sigma=\{p_{n}\}\subset K$. Since $K$ is compact, we can also assume that $\sigma=\{p_{n}\}\subset K$ converges to some $p\in K$. First, observe that $P,F\neq \emptyset$. In fact, by contradiction, assume for example that $F=\emptyset$. Let $F'$ be a maximal IF in $\uparrow P$ containing $I^{+}(p)$. Then, necessarily $P$ is maximal IP into $\downarrow F'$, and thus, $P\sim_{S}F'$, which contradicts Theorem \[7.3\] [*(3)*]{}. Moreover, from Theorem \[7.3\] [*(2)*]{} it is also $P\sim_{S} F$. Whence, it cannot happen that $P\subset I^{-}(p_{n})$ and $F\subset I^{+}(p_{n})$ for some $n$; so, assume for example that $$\label{u1} P\not\subset I^{-}(p_{n})\quad\hbox{for infinitely many}\;\; n.$$ Let $\gamma$ be a future-directed timelike curve with $P=I^{-}[\gamma]$. Since $(P,F)\in L(\rho)$, necessarily $$\label{u2} I^{-}[\gamma]=P\subset {\rm LI}(I^{-}(p_{n})).$$ Up to a subsequence of $\sigma$, from (\[u1\]) and (\[u2\]) we can choose a future chain $\varsigma=\{r_{n}\}_{n}\subset\gamma$ with $I^{-}[\varsigma]=I^{-}[\gamma]$ and $r_{n}\in\dot{I}^{-}(p_{n})$ for all $n$ ($\dot{I}^{-}(p_{n})$ denotes the topological boundary of $I^{-}(p_{n})$) such that $q\ll p_{n}$ for all $q\ll r_{n}$. These conditions necessarily imply $P\varsubsetneq I^{-}(p)\subset \downarrow F$, which contradicts that $P$ is maximal IP into $\downarrow F$.   ------------------------------------------------------------------------ \[non-hausdorff\] Let $\overline{V}$ be a chronological completion of a strongly causal spacetime $V$. If two elements of $\overline{V}$ are non-Hausdorff related then they are both in $\partial(V)$. [*Proof.*]{} Let $\sigma=\{(P_{n},F_{n})\}\subset\overline{V}$ be a sequence such that $(P,F),(P',F')\in L(\sigma)$ with $(P,F)\neq (P',F')$. Assume by contradiction that $(P,F)= (I^{-}(p),I^{+}(p))$ for some $p\in V$. From Theorems \[boundaryclosed\], \[artu\] and \[princ\], it is not a restriction to assume $(P_{n},F_{n})=(I^{-}(p_{n}),I^{+}(p_{n}))$ for all $n$, with $\{p_{n}\}_{n}\subset V$ converging to $p$ with the topology of the manifold. In particular, $K=\{p_{n}\}_{n}\cup \{p\}$ is a compact set in $V$. Therefore, from Proposition \[intermedio\], [**i**]{}\[K\] is closed in $\overline{V}$, and thus, $(P',F')\in {\bf i}[K]$. Again from Theorems \[artu\], \[princ\], this contradicts that $(P,F)\neq (P',F')$.   ------------------------------------------------------------------------ We finish this section by remarking that all these satisfactory results do not avoid the existence of some “contra-intuitive” limit behaviors for the $chr$-topology. Consider the situation described in Example \[3\]. If we appeal to our intuition, inherited from the natural embedding of this space into Minkowski, we would expect that the sequence $\sigma$ converges to $(P_{0},F_{0})\in\partial(V)$, which is not the case for the $chr$-topology. Notice however that this intuition is using additional information not exclusively contained in the causal structure of the spacetime. More precisely, if we analyze the chronology of the elements of $\sigma$, we observe that all of them have empty future. But the future in $V$ of $(P_{0},F_{0})$ is $F_{0}\neq \emptyset$. Therefore, any topology exclusively based on the chronology must conclude that $\sigma$ does [*not*]{} converge to $(P_{0},F_{0})$, as the $chr$-topology does (notice that this situation is totally different from that showed by the examples in [@KL1 Sect. III], [@KL2 Sect. II, III], where the chronology of the elements of the sequences have a good limit behavior, but, still, there is no convergence with the topologies involved there). This discussion shows that the causal boundary approach should not be considered an innocent variation of the conformal boundary. On the contrary, it provides a genuine insight on the asymptotic causal structure of the spacetime. In this sense, Examples \[1\] and \[3\] tell us that the asymptotic causal structure of $\L^{2}$ is modified in a very different way if we remove a vertical segment than if we remove a horizontal one (this is reasonable since time evolves just in the vertical direction), in contraposition with the conformal boundary approach, which does not reflect this asymmetry. Therefore, even though this limit behavior does not reproduce the situation in the conformal boundary, we consider this difference very satisfactory. Comparison with Other Approaches {#comparing} ================================ A natural question still needs to be investigated in order to emphasize the optimal character of our construction: what is the relation between the chronological completions and the completions suggested by other authors? In order to fix ideas we have chosen what perhaps are the most accurate approaches to the (total) causal boundary, up to date: the [*Marolf-Ross*]{} and the [*Szabados*]{} completions. The Marolf-Ross completion $\overline{V}_{MR}$ of a strongly causal spacetime $V$ is formed by all the pairs $(P,F)$ composed by an IP $P$ and an IF $F$, such that: (i) $P\sim_{S} F$, or (ii) $P=\emptyset$ and $P'\not\sim_{S}F$ for any IP $P'$, or (iii) $F=\emptyset$ and $P\not\sim_{S}F'$ for any IF $F'$. The chronology adopted here is also $$(P,F)\overline{\ll} (P',F')\quad\hbox{iff}\quad F\cap P'\neq\emptyset.$$ So, taking into account Proposition \[propS\], the Marolf-Ross construction becomes the biggest completion (according to Definition \[completion\]) which satisfies properties [*(1)*]{}–[*(4)*]{} in Theorem \[7.3\]. In particular, if $V$ admits more than one chronological completion then $\overline{V}_{MR}$ is [*strictly greater*]{} than any of them, since it includes the union of all of them, as illustrated in the first spacetime of Example \[yo\]. Remarkably, the strict inclusion $\overline{V}\varsubsetneq \overline{V}_{MR}$ may also hold even when $V$ admits just one chronological completion, as illustrated in the second spacetime of Example \[yo\]. The authors also adopt a topology for $\overline{V}_{MR}$: the topology generated by the subbasis $\overline{V}_{MR}\setminus L^{\pm}(S)$ for any $S\subset \overline{V}_{MR}$, where $$\begin{array}{c} L^{+}(S)=Cl_{FB}[S\cup L^{+}_{IF}(S)] \\ L^{-}(S)=Cl_{PB}[S\cup L^{-}_{IP}(S)] \end{array}$$ with $$\begin{array}{c} Cl_{FB}(S)=S\cup\{(P,F)\in\overline{V}_{MR}: F=\emptyset, P\in\hat{L}(P_{n})\;\;\hbox{for}\;\; (P_{n},F'_{n})\in S\} \\ Cl_{PB}(S)=S\cup\{(P,F)\in\overline{V}_{MR}: P=\emptyset, F\in\check{L}(F_{n})\;\;\hbox{for}\;\; (P'_{n},F_{n})\in S\} \end{array}$$ and $$\begin{array}{c} L^{+}_{IF}(S)=\{(P,F)\in\overline{V}_{MR}: F\neq\emptyset, F\subset\cup_{(P',F')\in S}F'\} \\ L^{-}_{IP}(S)=\{(P,F)\in\overline{V}_{MR}: P\neq\emptyset, P\subset\cup_{(P',F')\in S}P'\}. \end{array}$$ With these structures, the following remarkable comparison result can be stated: \[optm\] Given a strongly causal spacetime $V$, the inclusion defines a continuous and chronological map from every chronological completion $\overline{V}$ into the MR completion $\overline{V}_{MR}$. [*Proof.*]{} Let $\overline{V}$ be any chronological completion of $V$. As $\overline{V}\subset\overline{V}_{MR}$, the inclusion $i:\overline{V}\hookrightarrow\overline{V}_{MR}$ is well-defined and is always chronological. In order to prove that $i$ is also continuous, suppose that $\{(P_{n},F_{n})\}_{n}\subset \overline{V}$ converges to some $(P,F)\in\overline{V}$ with the $chr$-topology. We wish to prove that any open set $U=\overline{V}_{MR}\setminus L^{\pm}(S)$, $S\subset\overline{V}_{MR}$, of the subbasis which generates the MR topology such that $(P,F)\in U$ necessarily contains $(P_{n},F_{n})$ for all $n$ big enough. First, notice that $(P_{n},F_{n})\not\in S\cup L^{+}_{IF}(S)$ for all $n$ big enough. In fact, otherwise, $F\subset {\rm LI}(F_{n})\subset\cup_{(P',F')\in S}F'$. So, if $F\neq\emptyset$ then $(P,F)\in L^{+}_{IF}(S)\subset L^{+}(S)$, a contradiction. If, instead, $F=\emptyset$, taking into account that $P\in \hat{L}(P_{n})$, necessarily $(P,F)\in Cl_{FB}[S\cup L^{+}_{IF}(S)]=L^{+}(S)$, which is again a contradiction. Therefore, $(P_{n},F_{n})\not\in S\cup L^{+}_{IF}(S)$ for all $n$ big enough. Furthermore, $(P_{n},F_{n})$ cannot be $(P_{n},\emptyset)$, $P_{n}\in\hat{L}(P_{k}^{n})$, $(P_{k}^{n},F_{k}^{n})\in S\cup L^{+}_{IF}(S)$ for all $n$ big enough. In fact, otherwise, it must be $F=\emptyset$ and $P\in\hat{L}(P^{n}_{k_{n}})$ for some subsequence $\{k_{n}\}_{n}\subset \{k\}_{k}$, with $(P^{n}_{k_{n}},F^{n}_{k_{n}})\in S\cup L^{+}_{IF}(S)$, and thus, $(P,F)\in Cl_{FB}[S\cup L^{+}_{IF}(S)]=L^{+}(S)$, a contradiction. In conclusion, $(P_{n},F_{n})\not\in Cl_{FB}[S\cup L^{+}_{IF}(S)]=L^{+}(S)$ for all $n$ big enough, as required.   ------------------------------------------------------------------------ A conceptually different approach to the causal boundary of spacetimes consists of using identifications instead of pairs to form the ideal points of the boundary (see [@GKP]; and the subsequent papers [@R; @BS; @S1; @S2]). This approach presents important objections (see for example [@MR Section 2.2] for an interesting discussion); however, sometimes some identifications may be useful to emphasize certain aspects of the original spacetime. Our purpose here is to provide some evidence that chronological completions are the optimal spaces on which to establish eventual identifications. To this aim, we are going to give an improved version of the Szabados construction just by establishing some natural identifications on any chronological completion. The Szabados completion $\overline{V}_{S}$ of a strongly causal spacetime $V$ is formed by taking the quotient of $V^{\sharp}$ (as defined in footnote 4) by the minimum equivalence relation ${\cal R}$ containing $\sim_{S}$[^7]. Whence, each point $m\in\overline{V}_{S}$ is a class $[P_{1},P_{2},\ldots;F_{1},F_{2},\ldots]$ of ${\cal R}$-equivalent IPs and IFs. Szabados writes $m\ll m'$ if, for some $F_{\alpha}\in \pi^{-1}(m)$ and $P'_{\mu}\in\pi^{-1}(m')$, $F_{\alpha}\cap P'_{\mu}\neq\emptyset$. He also endows $\overline{V}_{S}$ with the quotient topology of ${\cal T}^{\sharp}$ under ${\cal R}$, where ${\cal T}^{\sharp}$ is the [*extended Alexandrov topology*]{} defined on $V^{\sharp}$, that is, the coarsest topology such that for each $A\in\check{V}$, $B\in \hat{V}$ the four sets $A^{{\rm int}}$, $B^{{\rm ext}}$, $B^{{\rm int}}$, $A^{{\rm ext}}$ are open sets, where $$\begin{array}{l} A^{\rm int}=\{P^{*}\in V^{\sharp}: P\in \hat{V}\;\hbox{and}\; P\cap A\neq\emptyset\}, \\ A^{\rm ext}=\{P^{*}\in V^{\sharp}: P\in \hat{V}\;\hbox{and}\; \forall S\subset V\; P=I^{-}[S]\Rightarrow I^{+}[S]\not\subset A \} \end{array}$$ (the sets $B^{{\rm int}}$ and $B^{{\rm ext}}$ have similar definitions with the roles of past and future interchanged). In order to compare any chronological completion $\overline{V}$ with the Szabados completion $\overline{V}_{S}$, the following identifications on $\overline{V}$ become natural: two pairs in $\overline{V}$ are $R$-related iff some of its respective components are ${\cal R}$-related. Then, we endow the resulting quotient space $\overline{V}/R$ with the corresponding quotient structures; that is, the quotient chronology $$[(P_{1},F_{1})]\overline{\ll} [(P_{2},F_{2})]\quad\hbox{iff}\quad (P_{1}',F_{1}')\overline{\ll} (P_{2}',F_{2}')\;\;\hbox{for some}\;\;\left\{\begin{array}{l} (P_{1}',F_{1}')\,R\,(P_{1},F_{1}) \\ (P_{2}',F_{2}')\,R\,(P_{2},F_{2}); \end{array}\right.$$ and the quotient of the $chr$-topology. There is an obvious bijection $b:\overline{V}/R\rightarrow\overline{V}_{S}$ which maps every class $[(P,F)]\in\overline{V}/R$ to the class $m\in \overline{V}_{S}$ formed by all the IPs and IFs appearing in some pair in $[(P,F)]$. With this definition, $b$ is obviously a chronological isomorphism. Furthermore, the only examples where $b$ is not continuous seem to be exclusively caused by “pathologies” of the Szabados topology, and thus, cannot be regarded as an anomaly of our construction. An illustrative example of this situation is showed in [@KL2 Sect. III]. Causal Ladder and the Boundary of Spacetimes {#applications} ============================================ The causal boundary approach constitutes an useful tool for examining the causal nature of a spacetime “at infinity”, which usually reflects important global aspects of the causal structure. Therefore, it may be interesting to analyze the causality of a spacetime just by looking at the boundary. An illustrative example of this situation is the following characterization of global hyperbolicity: [*a spacetime is globally hyperbolic iff there are no elements at the boundary whose past and future are both non-empty.*]{} This result, proposed in [@Se] in a slightly different context, was proved by Budic and Sachs in [@BS Th. 6.2]. However, their proof lies on the particular approach of the authors to the causal boundary (developed also in [@BS]), and thus, it suffers from the same important restriction: it is only valid for [*causally continuous*]{} spacetimes. The main aim of this section consists of extending this characterization to any strongly causal spacetime by using the chronological boundary of spacetimes. More precisely, we prove: \[globhyp\] Let $V$ be a strongly causal spacetime. Then, $V$ is globally hyperbolic if and only if there are no elements $(P,F)\in\partial(V)$ with $P,F\neq\emptyset$. [*Proof.*]{} First, recall that a strongly causal spacetime is globally hyperbolic if and only if the causal diamond $J(p,q):=J^{+}(p)\cap J^{-}(q)$ is compact for any $p,q\in V$. As a direct consequence of [@O p. 409, Lemma 14], this holds if $J(p,q)$ is included in a compact set $K\subset V$. For the implication to the left, assume that there are no elements $(P,F)\in\partial(V)$ with $P,F\neq\emptyset$. Take $K=\overline{I(p,q)}\supset J(p,q)$ with $I(p,q):=I^{+}(p)\cap I^{-}(q)$. Therefore, in order to prove that $K$ is compact we only need to show that any sequence in $I(p,q)$ admits a subsequence with some limit in $V$. By contradiction, assume that some $\sigma\subset I(p,q)$ does not satisfy this assertion. From [@FH Theorem 5.11] applied to $\hat{V}$, there exists some subsequence $\sigma'\subset\sigma$ and some TIP $P$ such that $\emptyset\neq I^{-}(p)\subset P$ and $P\in\hat{L}(\sigma')$. Let $\emptyset\neq F$ be a maximal TIF into $\uparrow P$. Let $\emptyset\neq P\subset P'$ be some maximal TIP into $\downarrow F$. Then, necessarily $P'\sim_{S} F$. Therefore, from Theorem \[7.3\] ([*1*]{}), ([*3*]{}), there exists some IF $F'\neq\emptyset$, such that $(P',F')\in\partial(V)$. This contradicts the hypothesis on the boundary. Whence, $K$ is compact. For the implication to the right, assume the existence of $(P,F)\in\partial(V)$ with $P,F\neq\emptyset$. Take points $p\in P$, $q\in F$. Take a chain $\varsigma\subset V$ with endpoint $(P,F)$. Then, the elements of $\varsigma$ are eventually contained in $I(p,q)$. However, from Theorem \[compatible\] [*(2)*]{} and Theorem \[non-hausdorff\], there cannot exist subsequences of $\varsigma$ converging in $V$. Whence, $J(p,q)\subset V$ is not compact.   ------------------------------------------------------------------------ A simple illustration of this result is provided by the Minkowski plane $\L^{2}$: any element $(P,F)\in \partial(\L^{2})$ satisfies either $P=\emptyset$ or $F=\emptyset$ (Example \[0\]); however, when a point is removed, and thus, the spacetime is no longer globally hyperbolic, a pair $(P,F)$ with $P,F\neq\emptyset$ immediately appears (Example \[ñ\]). The causal boundary approach also becomes useful to characterize other levels of the causal ladder as [*causal simplicity*]{}, [@BS Cor. 5.2]. For completeness, we are going to prove an extension of this result, again valid for any strongly causal spacetime. To this aim, consider any causal relation $\prec$ on any chronological completion $\overline{V}$ such that over ${\bf i}[V]$ it satisfies:[^8] $${\bf i}(p)\prec {\bf i}(q)\quad\hbox{iff}\quad\hbox{either}\;\; I^{+}(q)\subset I^{+}(p)\;\;\hbox{or}\;\; I^{-}(p)\subset I^{-}(q).$$ Then, the following results hold: \[las\] For any two points $p,q$ in a strongly causal spacetime $V$, ${\bf i}(p)\prec {\bf i}(q)$ if and only if, either $q\in \overline{J^{+}(p)}$ or $p\in \overline{J^{-}(q)}$. [*Proof.*]{} For the implication to the right, assume that ${\bf i}(p)\prec {\bf i}(q)$. If $I^{+}(q)\subset I^{+}(p)$, take any sequence $\{q_{n}\}\subset I^{+}(q)$ such that $q_{n}\rightarrow q$. Then, we obtain $q\in\overline{I^{+}(q)}\subset\overline{I^{+}(p)}\subset\overline{J^{+}(p)}$. If, instead, $I^{-}(p)\subset I^{-}(q)$, just reason analogously to obtain $p\in\overline{I^{-}(p)}\subset\overline{I^{-}(q)}\subset\overline{J^{-}(q)}$. For the implication to the left, first assume that $q\in \overline{J^{+}(p)}$. Take $q'\in I^{+}(q)$ and a sequence $\{q_{n}\}\subset J^{+}(p)$ with $q_{n}\rightarrow q$. For all $n$ big enough, $q_{n}\ll q'$. Whence, $q'\in I^{+}(p)$, and thus, $I^{+}(q)\subset I^{+}(p)$. Therefore, ${\bf i}(p)\prec {\bf i}(q)$. Assume now that $p\in\overline{J^{-}(q)}$. Reasoning analogously we deduce $I^{-}(p)\subset I^{-}(q)$, and thus, ${\bf i}(p)\prec {\bf i}(q)$ holds again.   ------------------------------------------------------------------------ \[causallysimple\] A strongly causal spacetime $V$ is causally simple if and only if the causality $\prec$ of $\overline{V}$ restricted to ${\bf i}[V]$ coincides with that of $V$. [*Proof.*]{} For the implication to the right, Lemma \[las\] implies that relation ${\bf i}(p)\prec {\bf i}(q)$ holds if and only if either $q\in\overline{J^{+}(p)}$ or $p\in\overline{J^{-}(q)}$, and, from the hypothesis, this holds if and only if $q\in J^{+}(p)$. Therefore, the causality of $\overline{V}$ restricted to ${\bf i}[V]$ coincides with that of $V$. For the implication to the left, first assume that $q\in \overline{J^{+}(p)}$. From Lemma \[las\], it is ${\bf i}(p)\prec {\bf i}(q)$, and, from the hypothesis, this implies $q\in J^{+}(p)$. Therefore, $J^{+}(p)$ is closed. Analogously, if $p\in \overline{J^{-}(q)}$, necessarily $p\in J^{-}(q)$, and thus, $J^{-}(q)$ is closed too. Therefore, $V$ is causally simple.   ------------------------------------------------------------------------ Finally, let us remark that the link between the causal ladder and the boundary of spacetimes may also arise in a deeper way. In Example \[10\] we have indicated how the dramatic change in the level of causality of generalized wave type spacetimes (\[genwav\]), when metric coefficient $-H$ leaves the quadratic growth, translates into a low dimensionality of the chronological boundary for these spacetimes. It would be interesting to explore if this intriguing relation between the critical behavior of the causality of a spacetime and the dimensionality of its boundary is generalizable to further classes of spacetimes. Examples ======== In this section we briefly examine our construction in some examples and compare it with previous approaches, putting special emphasis in the differences between them. \[0\] *Consider Minkowski space $V=\L^{n+1}(\equiv {{\mathbb R}}^{n}\times\L^{1})$. In order to construct the chronological boundary $\partial(V)$, in this case it suffices to consider the set of pairs $(I^{-}[\gamma],\emptyset)$ (resp. $(\emptyset,I^{+}[\gamma])$), for every inextensible future-directed (resp. past-directed) lightlike geodesic $\gamma$, in addition to the pairs $(V,\emptyset)$ and $(\emptyset,V)$. Therefore, $\partial(V)$ can be identified with a pair of cones on $S^{n-1}$ with apexes $i^{+}$, $i^{-}$ (Figure 2). This is in total agreement with the image of the (standard) conformal embedding of Minkowski space into the Einstein Static Universe.* {width="5cm"} On the other hand, the limit of every sequence in $\overline{V}$ with the $chr$-topology coincides with the set-theoretic limit of the elements of the sequence. Again, this provides just the same topology as that inherited from the (standard) conformal embedding of $\L^{n+1}$ into ESU. \[ñ\] [*Let $V$ be $\L^{2}$ with the origin point removed (Figure 3): $$V=\L^{2}\setminus \{(0,0)\}.$$ The GKP pre-completion $V^{\sharp}$ of this spacetime attaches at the origin two ideal endpoints given by the pairs $(P,\emptyset)$, $(\emptyset,F)$. This provides a simple example of the non-equivalence between the notions of [*limit*]{} and [*endpoint*]{} of a chain: the pair $(\emptyset,F)\in V^{\sharp}$ is a limit of the chain $\{{\bf i}(p_{n})\}\subset {\bf i}[V]$, however, the unique endpoint of $\{p_{n}\}$ in $V^{\sharp}$ is instead the pair $(P,\emptyset)$. Consider now the completion resulting from replacing in $V^{\sharp}$ the pair $(P,\emptyset)$ by $(P,F)$. In this case, the sequence constantly equal to $(P,F)$ converges with the $chr$-topology to both, $(\emptyset,F)$, $(P,F)$, showing that this topology is not $T_{1}$ for this completion. The (unique) chronological completion $\overline{V}$ of $V$ only attaches at the origin the ideal point $(P,F)$, showing in particular that $V$ is not globally hyperbolic (Theorem \[globhyp\]). On the other hand, this example shows that “boundary” need not be metrically infinitely far along geodesics, as illustrated by the curve $\gamma(t)=(0,t)$, $t<0$.*]{} {width="4cm"} \[1\] *Let $V$ be $\L^{2}$ with the vertical segment ${\cal V}_{+}=\{(0,t): t\geq 0\}$ removed (Figure 4): $$V=\L^{2}\setminus \{(0,t): t\geq 0\}.$$ Let $\overline{V}$ be the (unique) chronological completion of $V$. The pairs $(P,F_{l})$, $(P,F_{r})$ are the unique endpoints in $\overline{V}$ of the chains $\{q_{n}\}$, $\{p_{n}\}$, resp. They represent two ideal endpoints attached at the extreme of ${\cal V}_{+}$. These pairs are also endpoints of the future chain $\{r_{n}\}$, and thus, limits of $\{{\bf i}(r_{n})\}$ (recall Theorem \[compatible\] [*(2)*]{}). Therefore, $\overline{V}$ is non-Hausdorff with the $chr$-topology. If we extend this analysis to the whole line ${\cal V}_{+}$, we obtain that $\partial(V)$ contains two copies of ${\cal V}_{+}$, with only the extreme ideal points $(P,F_{l})$, $(P,F_{r})$ being non-Hausdorff related.* On the other hand, observe that the chain $\{r_{n}\}$ generates the pair $(P,F_{l}\cup F_{r})\in V_{p}\times V_{f}$. This pair belongs to the completion $V^{\flat}$ (see Section \[chron\]), showing in particular that some completions may contain pairs whose components are not necessarily indecomposable. {width="4cm"} \[3\] *Let $V$ be $\L^{2}$ with the horizontal segment ${\cal H}_{-}=\{(x,0): x\leq 0\}$ removed (Figure 5): $$V=\L^{2}\setminus \{(x,0): x\leq 0\}.$$ Let $\overline{V}$ be the (unique) chronological completion of $V$. For $x<0$, the unique endpoints in $\overline{V}$ of the chains $\{(x,-1/n)\}$ and $\{(x,1/n)\}$ are the pairs $(P_{x},\emptyset)$ and $(\emptyset,F_{x})$, resp. However, for $x=0$ the unique endpoint in $\overline{V}$ of the chains $\{(0,-1/n)\}$ and $\{(0,1/n)\}$ is the pair $(P_{0},F_{0})$. Therefore, in this case the chronological completion $\overline{V}$ contains two copies of ${\cal H}_{-}$ with the right extreme points of the copies identified via $(P_{0},F_{0})$. On the other hand, we can ask for the limit of the sequence $\sigma=\{(P_{x_{n}},\emptyset)\}_{n}\subset \overline{V}$, with $x_{n}=-1/n$ for all $n$. Surprisingly, $\sigma$ does [*not*]{} converge to $(P_{0},F_{0})$ with the $chr$-topology, violating the common intuition inherited from the natural embedding of this space into Minkowski.* When we consider a completion different from the chronological one, the corresponding boundary may be [*non-closed*]{}. In fact, take for example the completion $V^{end}$, which contains in particular the endpoint $(P,F)$, with $P=P_{0}\cup I^{-}(p)$ and $F=I^{+}(p)$, $p=(-1,1)$. Then, the sequence constantly equal to $(P,F)$, which is obviously contained in the boundary, converges to ${\bf i}(p)\in {\bf i}[V]$ with the $chr$-topology. Finally, consider the future chronological completion $\hat{V}$ of this spacetime. Apart from the obvious limit $I^{-}(p)$, the sequence of PIPs $\delta=\{I^{-}(p_{n})\}\subset \hat{V}$, $p_{n}=(-1+1/n,1)$ for all $n$, also converges to $P_{0}$ with the $\;\widehat{}\;$-topology. This anomalous limit is due to the fact that $\hat{V}$ and $\;\widehat{}\;$-topology only retain partial information about the chronology of $V$. This situation contrasts with our construction, where the full chronology is taken into consideration. In fact, under our approach, Theorems \[artu\], \[princ\] and \[non-hausdorff\] imply that $\{{\bf i}(p_{n})\}_{n}\subset \overline{V}$ only converges to ${\bf i}(p)$, as expected. {width="11cm"} \[5\] [*Consider the spacetime $V$ represented in Figure 6 (this example comes from [@S1 Figure 2]; see also [@MR Figure 7]). Here, infinite null segments $\{L_{n}\}_{n}$ and the point $r$ have been removed from the Minkowski plane, resulting in a spacetime such that $I^{-}[\gamma']\varsubsetneq I^{-}[\gamma]$. In this case the (unique) chronological boundary $\partial(V)$ coincides with the MR boundary (see [@MR]), including the pairs $(I^{-}[\gamma],F)$ and $(I^{-}[\gamma'],\emptyset)$ as endpoints of $\gamma$ and $\gamma'$, resp. However, the MR topology is different from the $chr$-topology. In fact, as showed in [@MR], the sequence constantly equal to $(I^{-}[\gamma],F)$ also converges to $(I^{-}[\gamma'],\emptyset)$ with the MR topology, and thus, it is not $T_{1}$. This is not the case for the $chr$-topology, which is always $T_{1}$ (Theorem \[6.4\]).*]{} {width="4cm"} \[yo\] *Let $X$ be the disjoint union of $I_{i}=(-\infty,+\infty)_{i}$, $i=1,2$, under the equivalence relation $0_{1}\sim 0_{2}$ (Figure 7). Endow $X$ with the (quotient of the) chronology relation given by $$x\ll x'\quad\hbox{iff}\quad\left\{\begin{array}{l} \hbox{either}\quad x<x'\;\;\hbox{and}\;\; x,x'\in I_{i}\;\; \hbox{for some}\;\; i; \\ \hbox{or}\quad x<x'\;\;\hbox{and}\;\; 0\geq x\in I_{i},\;\, 0\leq x'\in I_{j}\;\;\hbox{for}\;\; i\neq j. \end{array}\right.$$ Then, $X$ becomes a chronological set satisfying $$I^{\pm}([0])=I_{1}^{\pm}\cup I_{2}^{\pm},\qquad\hbox{with}\quad\begin{array}{l} I_{i}^{-}:=(-\infty,0)_{i} \\ I_{i}^{+}:=(0,+\infty)_{i} \end{array}.$$ Therefore, the past and future of $[0]\in X$ are [*not*]{} indecomposable.* {width="6cm"} Motivated by this example, consider now the $(2+1)$-spacetime $V$ constructed by deleting from $\L^{3}$ the subsets $\{x=0, -1\leq t\leq 0\}$ and $\{y=0, 0\leq t\leq 1\}$. In fact, $V$ has two TIPs $P_{1}$, $P_{2}$ and two TIFs $F_{1}$, $F_{2}$ associated to the removed origin (Figure 8), which verify $P_{1}\sim_{S} F_{1},F_{2}$ and $P_{2}\sim_{S} F_{1},F_{2}$. Therefore, there are [*two*]{} different chronological completions for $V$: one of them attaches at the origin the ideal endpoints $(P_{1},F_{1})$, $(P_{2},F_{2})$, while the other one attaches the ideal endpoints $(P_{1},F_{2})$, $(P_{2},F_{1})$. Both of these chronological completions are [*different*]{} from the MR completion, which attaches at the origin “more” ideal endpoints: $(P_{1},F_{1})$, $(P_{1},F_{2})$, $(P_{2},F_{1})$, $(P_{2},F_{2})$. If we additionally delete from $\L^{3}$ the subset $\{x>0, y>0, t=0\}$ then the MR completion attaches at the origin the pairs $(P_{1},F_{1})$, $(P_{1},F_{2})$, $(P_{2},F_{2})$. In this case the spacetime admits an unique chronological completion, which attaches at the origin the pairs $(P_{1},F_{1})$, $(P_{2},F_{2})$. Another example of spacetime whose MR completion also includes spurious ideal points appears in [@MR Appendix A]. {width="7cm"} \[9.5\] (Standard Static Spacetimes) *The causal boundary for standard static spacetimes has been studied in [@H3; @FH; @AF]. From the conformal invariance of the causal boundary approach, it is not a restriction to assume that these spacetimes can be written as $$V=M\times{{\mathbb R}},\qquad g=h-dt^{2},$$ where $(M,h)$ is an arbitrary Riemannian $3$-manifold. In these spacetimes, the spatial projection $c$ of every inextensible future-directed timelike curve $\gamma(t)=(c(t),t)$ whose past is different from the whole spacetime is an [*asymptotically ray-like curve*]{}, i.e. an inextensible curve with domain $[w,\Omega)$, $\Omega\leq\infty$, velocity $|\dot{c}|\leq 1$ and finite-valued [*Busemann function*]{} $b_{c}$; that is, $$b_{c}:M\rightarrow{{\mathbb R}}^{*}\equiv{{\mathbb R}}\cup\{\infty\},\qquad b_{c}(\cdot):=\lim_{t\rightarrow\Omega}(t-d(\cdot,c(t)))<\infty.$$ Moreover, the pasts of these curves are totally characterized by the corresponding Busemann functions $b_{c}$. Therefore, if we denote by ${\cal B}(M)$ the set of all Busemann functions associated to asymptotically ray-like curves in $(M,h)$, we obtain $$\hat{\partial}(V)={\cal B}(M)\cup\{\infty\}.$$ In other words, if we define the [*Busemann boundary*]{} of $(M,h)$ as the quotient space $$\partial_{{\cal B}}(M):={\cal B}(M)/{{\mathbb R}},$$ which, in particular, includes the Cauchy boundary $\partial_{c}(M)$, then the future chronological boundary of $V$ is a cone with apex $i^{+}$ and base $\partial_{{\cal B}}(M)$; i.e. $$\hat{\partial}(V)={\cal B}(M)\cup\{\infty\}\equiv(\partial_{{\cal B}}(M)\times{{\mathbb R}})\cup \{i^{+}\}$$ Analogously, the past chronological boundary of $V$ is a cone with apex $i^{-}$ and base $\partial_{{\cal B}}(M)$; i.e. $$\check{\partial}(V)=(\partial_{{\cal B}}(M)\times{{\mathbb R}})\cup \{i^{-}\}.$$* But, what about the (total) chronological boundary? The common future (resp. common past) of any inextensible future-directed (resp. past-directed) timelike curve $\gamma$ is non-empty iff $\gamma$ approaches to some $(t,p)\in\hat{\partial}(V)$ with $p\in\partial_{c}(M)$; moreover, in this case the common future (resp. common past) of $\gamma$ coincides with the future (resp. past) of any past-directed (resp. future-directed) timelike curve approaching also to $(t,p)\in\check{\partial}(V)$. So, the (total) chronological boundary is a double cone with base $\partial_{{\cal B}}(M)$, apexes $i^{+}$, $i^{-}$, and future and past copies of lines over the same point $p$ of the Cauchy boundary $\partial_{c}(M)$ identified. Summarizing: $$\partial(V)=(\hat{\partial}(V)\cup \check{\partial}(V))/\sim,\qquad\hbox{with}\;\; (p^{+},t^{+})\sim (p^{-},t^{-})\;\;\hbox{iff}\;\;\left\{\begin{array}{l} (p^{+},t^{+})\in \hat{\partial}(V) \\ (p^{-},t^{-})\in \check{\partial}(V) \\ p^{+}=p^{-}\in\partial_{c}(M) \\ t^{+}=t^{-}\in{{\mathbb R}}\end{array}\right.$$ (see Figure 9). On the other hand, the $chr$-topology on $\overline{V}$ coincides with the quotient topology over $\sim$ of the topology generated by the limits operators $\hat{L}$ and $\check{L}$ on $\hat{V}\cup \check{V}$. {width="10cm"} \[10\] (Locally Symmetric Plane Waves) *Consider $V={{\mathbb R}}^{n+2}$ with metric $$\langle\cdot,\cdot\rangle =\langle\cdot,\cdot\rangle_{0}+2du\,dv+H(x)du^{2},$$ where $\langle\cdot,\cdot\rangle_{0}$ is the canonical metric of ${{\mathbb R}}^{n}$ and $$H(x)=-\mu_{1}^{2}x_{1}^{2}-\cdots - \mu_{j}^{2}x_{j}^{2}+\mu_{j+1}^{2}x_{j+1}^{2}+\cdots + \mu_{n}^{2}x_{n}^{2},\qquad\hbox{with}\;\; j>0$$ and $\mu_{1}\geq \mu_{2}\geq\cdots\geq \mu_{j}$. The causal boundary of these spacetimes was first analyzed in [@MR0; @MR] by using the MR approach. If we denote by $L^{+}$, $L^{-}$ two copies of the line $u\in (-\infty,\infty)$, and define $$L=(L^{+}\cup L^{-})/\,R,\quad\qquad u\; R\; u'\Longleftrightarrow u\in L^{+},\; u'\in L^{-},\; u=u'-\pi/\mu_{1},$$ the authors found that the MR boundary of $V$ can be represented by the single line $\partial(V)=L\cup\{i^{+},i^{-}\}$. Moreover, in this case the MR construction coincides with our construction (see [@FS]). Therefore, the chronological boundary of $V$ can be also represented by this single line (see Figure 10). This picture agrees with the result previously obtained in [@BN] for the maximally symmetric case by using the conformal approach (see [@MR Section 5] for a brief discussion).* ![Chronological boundary for Locally Symmetric Plane Waves as pictured in [@MR].](Wdib9.eps){width="5cm"} This $1$-dimensional character of the boundary admits an intriguing interpretation in terms of the global causal behavior of the wave. In fact, in [@FS0] the authors found that the causality of [*generalized wave type spacetimes*]{} $$\label{genwav} V=M\times{{\mathbb R}}^{2},\qquad \langle\cdot,\cdot\rangle_{z}=\langle\cdot,\cdot\rangle_{x}+2du\,dv+H(x,u)du^{2},$$ where $(M,\langle\cdot,\cdot\rangle_{x})$ is a Riemannian $n$-manifold and $H:M\times{{\mathbb R}}\rightarrow{{\mathbb R}}$ a smooth function, presents a critical behavior with respect to the metric coefficient $-H$. More precisely, these waves pass from being globally hyperbolic for $-H$ subquadratic (and $M$ complete) to being non-distinguishing for $-H$ superquadratic. This gap in the causal ladder has to do with a sort of degeneracy for the chronology of the wave “at infinity”, which, in this case, translates into a [*low dimensionality*]{} of the boundary. We refer the reader to [@MR0; @HR; @FS] for a systematic study (of increasing generality) of the causal boundary for spacetimes (\[genwav\]), including plane waves and pp-waves. Acknowledgments =============== I would like to thank Professors Steven Harris and Miguel Sánchez for useful discussions and comments. Part of this work was completed during a research stay at Department of Mathematics, Stony Brook University. Partially supported by MECyD Grant EX-2002-0612, JA regional Grant P06-FQM-01951 and MEC Grants RyC-2004-382 and MTM2007-60731. [99]{} V. Alaña, J.L. Flores, [*The causal boundary of product spacetimes*]{}, Gen. Relat. Grav., in press. Available at arXiv:0704.3148. D. Berenstein, H. Nastase, [*On lightcone string field theory from super Yang-Mills and holography*]{}. E-print available at hep-th/0205048. R. Budic, R.K. Sachs, [*Causal boundaries for general relativistic spacetimes*]{}, J. Math. Phys. [**15**]{} (1974), 1302-09. J.L. Flores, S.G. Harris, [*Topology of causal boundary for Standard Static spacetimes*]{}, Class. Quantum Grav. [**24**]{} (2007), 1211-60. Available at gr-qc/0607049. J.L. Flores, M. Sánchez, [*Causality and conjugate points in general planes waves*]{}, Class. Quantum Grav. [**20**]{} (2003), 2275-91. J.L. Flores, M. Sánchez, [*The causal boundary of wave-type spacetimes.*]{} Preprint (2007). A. García-Parrado, J.M. Senovilla, [*Causal relationship: A new tool for the causal characterization of Lorentzian manifolds*]{}, Class. Quantum Grav. [**20**]{} (2003), 625-64. A. García-Parrado, J.M. Senovilla, [*Causal structures and causal boundaries*]{}, Class. Quantum Grav. [**22**]{} (2005), R1-R84. R.P. Geroch, E.H. Kronheimer and R. Penrose, [*Ideal points in spacetime,*]{} Proc. Roy. Soc. Lond. A [**237**]{} (1972), 545-67. S.G. Harris, [*Universality of the future chronological boundary,*]{} J. Math. Phys. [**39**]{} (1998), 5427-45. S.G. Harris, [*Topology of the future chronological boundary: universality for spacelike boundaries,*]{} Class. Quantum Grav. [**17**]{} (2000), 551-603. S.G. Harris, [*Causal boundary for Standard Static spacetimes,*]{} Nonlinear analysis [**47**]{} (2001), 2971-81. S.G. Harris, [*Discrete group actions on spacetimes: causality conditions and the causal boundary,*]{} Class. Quantum Grav. [**21**]{} (2004), 1209-36. S.G. Harris, [*Boundaries on spacetimes: an outline*]{}, Contemp. Math. [**359**]{} (2004), 65-85. V. Hubeny, M. Rangamany, [*Causal structures of pp-waves*]{}, J. High Energy Phys. [**12**]{} (2002), 043. Kuang zhi-quan, Li jian-zeng and Liang can-bi, [*$c$-boundary of Taub’s plane-symmetric static vacuum spacetime*]{}, Phys. Rev. D [**33**]{} (1986), 1533-37. Kuang zhi-quan, Liang can-bi, [*On the GKP and BS constructions of the $c$-boundary*]{}, J. Math. Phys. [**29**]{} (1988), 433-5. Kuang zhi-quan, Liang can-bi, [*On the Racz and Szabados constructions of the $c$-boundary*]{}, Phys. Rev. D [**46**]{} (1992), 4253-6. D. Marolf, S. Ross, [*Plane Waves: To infinity and beyond!*]{}, Class. Quant. Grav. [**19**]{} (2002), 6289-302. D. Marolf, S.F. Ross, [*A new recipe for causal completions*]{}, Class. Quantum Grav. [**20**]{} (2003), 4085-117. B. O’Neill, [*Semi-Riemannian Geometry with applications to Relativity,*]{} Series in Pure and Applied Math. [**103**]{} Academic Press, N.Y. (1983). R. Penrose, [*Conformal treatment of infinity,*]{} Relativity, Groups and Topology, ed. C.M. de Witt and B. de Witt, (1964, Gordon and Breach, New York); ibid, [*Zero rest-mass fields including gravitation: asymptotic behavior*]{}, Proc. Roy. Soc. Lond. A [**284**]{} (1965), 159-203. I. Racz. [*Causal boundary of space-times*]{}, Phys. Rev. D [**36**]{} (1987), 1673-5: ibid, [*Causal boundary for stably causal space-times*]{}, Gen. Relat. Grav. [**20**]{} (1988), 893-904. H. Seifert, [*The Causal Boundary of Space-Times*]{}, Gen. Rel. Grav. [**1**]{} (1971), 247-59. L.B. Szabados, [*Causal boundary for strongly causal spaces,*]{} Class. Quantum Grav. [**5**]{} (1988), 121-34; L.B. Szabados, [*Causal boundary for strongly causal spacetimes: II,*]{} Class. Quantum Grav. [**6**]{} (1989), 77-91. [^1]: Here we are following standard notation: that is, $I^{-}[\cdot]$ denotes the past of a set of points, while $I^{-}(\cdot)$ is reserved for past of a point. [^2]: Zorn’s Lemma: Every non-empty partially ordered set in which every totally ordered subset has an upper bound contains at least one maximal element. [^3]: When using the limit operator $\hat{L}$, it is common to implicitly identify the past or future of a point with the point itself. [^4]: Actually, $V^{\sharp}$ was introduced in [@GKP] by using identifications instead of pairs: that is, $V^{\sharp}:=\hat{V}\cup\check{V}/\sim$, where $P\sim F$ iff $P=I^{-}(p)$, $F=I^{+}(p)$, for some $p\in V$. [^5]: Of course, symbols $I^{-}(\cdot)$, $I^{+}(\cdot)$ in (\[expected\]) refer to the chronological relation $\overline{\ll}$ instead of $\ll$. [^6]: In particular, observe that every $S_{i}$ has at most two elements. [^7]: According to [@S2] further identifications between pairs of TIPs or TIFs must be considered; however, they will be omitted in our discussion. [^8]: We have introduced here causal relations just to establish Theorem \[causallysimple\], but with no further pretension. We postpone to a future paper a more precise definition of causal relations into our framework.

ArXiv

--- abstract: 'The approximate frequency shift of a radar echo from a meteoroid is derived. The origin of head echos are discussed by considering schematic models.' author: - 'Hans Wilschut[^1]' title: | Frequency shifts of head echos\ in meteoroid trail formation --- Introduction ============ Many aspects concerning radio echos from meteoroid trails can be understood in terms of the line-oscillator model. The standard work for this is McKinley’s book [@McK], chapter 8 and 9. However, McKinley does not show how frequency shifts near the optimal reflection point can be calculated. In this paper the formalism of [@McK] is extended to include frequency shifts. In doing so a close relation appears between the line-oscillator and what is referred to as the “moving ball” model. In the latter model frequency shifts occur because the radio echo will be Doppler shifted. Inspecting measured meteoroid radio echo’s such as shown in figures \[ex1\] and \[ex2\] one can observe two distinct parts. The first part is a signal with a frequency deviating from the transmitter frequency and the second part,which extends over longer time, is centered around the transmitter frequency. This first part is often referred to as the “head echo”. Many authors interpret this signal in terms of Doppler shifts. The larger second part is well understood as a reflection from the trail left behind after a meteoroid passed. The free thermalized electrons in the trail can cause strong reflections of the radio signal when their individual amplitudes add coherently. The head echo is produced during trail formation. In the line-oscillator description the scattering electrons, created as the trail is formed, will have relative phases such that it appears as a frequency shift away from the emitter frequency. This is not the Doppler shift due to reflections on a co-moving plasma. The latter is assumed in the moving ball description. Some recent work on Doppler shifts in head echos can be found in this journal. See for example [@Kauf; @Verb; @German]. One aim of this paper is to derive an approximate value of the frequency shift using the line-oscillator model of [@McK]. The notation of that work will be followed unless otherwise indicated. For the convenience of the reader some of the formalism in ref. [@McK] will be repeated here. The derivation of the frequency shift in this model will be given in the next section. The third section considers the case of Doppler shifts, showing the close connection with the line-oscillator model. In section 4 example calculations are made for a qualitative comparison with observations. Section 5 gives a suggestion why spectra like the one shown in fig. \[ex1\] with only a half head echo are seen more frequently than the one in fig. \[ex2\] with a complete head echo. The final section contains some concluding remarks. Derivation of the frequency shift ================================= When an ionized trail is made by a meteoroid (see fig. \[fig1\]), the created free electrons act as individual scatterers, they re-emit the signal of a transmitter in all directions. These can be observed in a receiver contributing an amplitude $dA_R$ $$\label{dAr} dA_R \propto \sin\left( 2\pi f t - \frac{2\pi(R'_1+R'_2)}{\lambda} \right) \ ,$$ where $\lambda$ is the transmitters wavelength and $f$ its frequency. ()All other parameters in this work are defined in fig. \[fig1\].) Each scattering contribution has a different phase depending on the distance $(R'_1+R'_2)$ the signal travels. Going along the trail the addition becomes coherent when $~{d (R'_1+R'_2)/dt=0}$. At this point the path followed is a reflection, the specular condition. The length of this path is $R_1+R_2$ and is the shortest path between transmitter and receiver via a point on the meteoroid trajectory. This point will be referred to as the specular point. In fig. \[fig1\] also shows signal paths corresponding to back scattering. This is the radar setup, where emitter and transmitter are at nearly the same location. An arbitrary path has length $2R$ and the shortest path is $2R_0$. Near the specular point where $s\ll R_0$ one finds that ![\[color on-line\] TX and RX are the locations of the transmitter and receiver, respectively. The signal path for back scattering (red) and for forward scattering (blue) is shown for the optimal path corresponding to the reflection (specular) point, $t_0$, (full lines) and the path of a contributing neighboring point (dashed lines) separated by a distance $s$. []{data-label="fig1"}](fig188.pdf){width="0.8\linewidth"} $$\begin{aligned} R_1'+R_2'&\approx &R_1 +R_2 +\frac{s^2 \sin^2 \omega}{2}\left (\frac{1}{R_1-s\cos\omega} +\frac{1}{R_2+s\cos\omega}\right )\\ &\approx& R_1 +R_2 +\frac{s^2 \sin^2 \omega}{2}\ \frac{R_1+R_2}{R_1 R_2} \ .\label{aberhallo}\end{aligned}$$ This reduces to $$\label{simple} R \approx R_0 +\frac{s^2}{2R_0}$$ for back scattering. Also note that in chapter 9 of [@McK] the notation was changed: $s \rightarrow f$ or **. Here we will use $s$ consistently for the path of the meteoroid. Further note that, in general, $\omega \ne \pi/2-\phi$ because the trail may make an angle $\beta$ with the plane where forward scattering takes places, in which case $ \sin^2\omega=1- \sin^2\phi\cos^2\beta$. ![Typical example, of a strong head echo. The horizontal scale is one second long, the vertical scale 1.5 kHz. The slope corresponds to -9.7 kHz/s Data were taken 05 May 2020 in Kampenhout (BE) using the 49990 kHz VVS beacon near Ieper (BE)[@Felix][]{data-label="ex1"}](ex1.jpg){width="0.8\linewidth"} . Central to the problem is the summation over individual scatterers along the trail. The model assumes a constant ionization density along the trail, the amplitude is then given by $$\label{Integral} A_R\propto\int_{x_1}^{x}\sin(\chi - \frac{\pi x^2}{2} )dx \ .$$ Following [@McK], we first evaluate the integral for back scattering where $2s=x(R_0 \lambda)^{1/2}$ and $\chi=2\pi f t\ + $ a time independent phase. Thus one sums from the beginning of the trail until a point $s(t)$, i.e. the length of the trail at time t. Without loss of generality $x_1 \rightarrow -\infty$ can be assumed, as shown explicitly in section 3. One obtains $$A_R \propto \left(C(x)+\frac{1}{2}\right) \sin \chi -\left(S(x)+\frac{1}{2}\right) \cos \chi \ ,$$ where $C$ and $S$ denote Fresnel integrals[^2]. To see what this means for the frequency we look at times $t<t_0$ and $t>t_0$ avoiding the complex behavior near $t_0$ by considering $|x|\gtrapprox1$. With this approximation $$\label{aleft} A_{<} \propto \frac{\cos (\chi- \frac{\pi x^2}{2})}{\pi x}\ \mathrm{and} \ A_{>}\propto\sin (\chi )-\cos (\chi )+\frac{\cos (\chi -\frac{\pi x^2}{2})}{\pi x}\ .$$ As $t$ approaches $t_0$ the amplitude $A_<$ increases slowly compared with the frequency $f$. At each $t$ the instantaneous frequency $f_i$ can be obtained by determining the phase $\Phi$ of $A_<(t)$ $$\label{aleft2} \Phi=\chi - \frac{\pi x^2}{2} =2\pi f t - 2\pi \frac{s^2}{R_0\lambda}\$$ and taking its derivative with respect to $t$, giving $$\label{fi} f_i=\frac{1}{2\pi}\frac{d \Phi}{dt}=f-\frac{2s\ ds/dt}{R_0 \lambda} \ .$$ Note that here $s>0$ and $ds/dt <0$, the shift is thus positive. In practice one analyses the change in the instantaneous frequency $f_i$, $$\label{timedepshift} \frac{df_i}{dt}=-2\left( \left[ \frac{ds}{dt} \right]^2 + s\frac{d^2s}{dt^2} \right)\frac{1}{R_0 \lambda} \ .$$ A simple model choice is a constant velocity where $s=|V (t_0- t)|$, so that $$\label{8-15r} \frac{df_i}{dt}=-\frac{2V^2}{R_0\lambda}\ .$$ To obtain the shift in forward scattering one simply replaces eq. \[simple\] wit eq. \[aberhallo\] to find $$\label{result} \frac{df_i}{dt}\approx-\frac{V^2 \sin^2\omega}{\lambda}\frac{R_1+R_2}{R_1R_2}\ .$$ For $t>t_0$ there is no such simple derivation for the phase of $A_>$ possible. In order to get insight numerical calculations were done. These will be discussed in section \[calcs\]. Derivation of the frequency shift in terms of Doppler shifts ============================================================ ![Less common head echo showing a shift that extends to negative shifts. The slope is -4.9 kHz/s. Data taken on 09 June 2011 in Kampenhout using the 49970 kHz BRAMS beacon at Dourbes (BE)[@Felix][]{data-label="ex2"}](ex2.jpg){width="0.8\linewidth"} . In this section we consider the possibility that only a short part of the trail survives, so that it has a length $x <1$. Thus where at the front free electrons are created they disappear at the back. Therefore, the trail has a constant length $\Delta x$. In fact for the observer it appears as a passing object, which maybe as well be the meteoroid itself. When this object passes point $t_0$ it will give an echo. In this case one can approximate eq. \[Integral\] by $$\label{delta} A_R\propto\sin(\chi - \frac{\pi x^2}{2} )\Delta x \ .$$ The phase of this amplitude is identical to that in eq. \[aleft2\] and the same relations hold for the shifts. The shift continues for $t>0$ where the instantaneous frequency, $f_i$, is lower than the emitter frequency $f$. The bistatic Doppler shift is given by $-\frac{1}{\lambda}{d (R'_1+R'_2)/dt}$. Using the same approximations as eq. \[aberhallo\]ione arrives at the identical expression as in eq. \[fi\]. Thus the shifts are the same as they must be, because of Galilean invariance. A more informative evaluation will come from a numerical calculation in the next section. Example Calculations {#calcs} ==================== At this point it may be interesting to consider in more detail how the signal appears in observation. The result of section 2 and 3 can be generalized into one expression $$\label{ARl} A_R \propto \left(C(x)-C(x-\Delta x)\right) \sin \chi -\left(S(x)-S(x-\Delta x )\right) \cos \chi \ ,$$ where $\Delta x$ is the length of the object or trail as defined above. In the following calculation we use for $s$ the backscatter configuration with $V=30$ km/s, $R_0=100$ km, and an emitter frequency $f=50$ MHz. First we consider the signal power, $\overline{ A_R^2(t)}$, which is obtained by averaging over a time long with respect to the frequency but small with respect to $x(t)$. In fig. \[powerplot\] this quantity is shown, evaluated for various values of $\Delta x$. ![$\overline{ A_R^2(x(t))}$ for $\Delta x= 0.5,1,2,4,8,\infty$ (black, red, blue, green, cyan, magenta)[]{data-label="powerplot"}](signalstrength.pdf){width="0.7\linewidth"} For small $\Delta x$ the signal has a nearly Gaussian dependence, where it should be noted that $x=1$ corresponds to about one Fresnel zone (here it corresponds to $t=\sqrt{\lambda R_0}/2V=13\ \mathrm{ms}$ or a distance of $387\ \mathrm{m}$). At this small distance the signal has the characteristics of Fraunhofer slit scattering. For $\Delta x>2 $ and $x>1$ one observes what are called Fresnel oscillations. $\Delta x=\infty$ refers to the situation discussed in section 2; the pattern seen here corresponds to Fresnel edge scattering. The oscillation frequency in the Fresnel pattern (see chapter 8 eq. (8-15) in [@McK]) is identical to the frequency shift calculated in section 2. In fact, historically, it has been a main tool for determining meteoroid velocities instead of the frequency shift. ![\[Color online\] Time versus frequency of $\log(A_R)$ obtained with a STFT procedure. Left: passing trail or object. Right: Dynamic meteoroid trail. See text[]{data-label="spectra"}](exnew1.pdf "fig:"){width="0.4\linewidth"}![\[Color online\] Time versus frequency of $\log(A_R)$ obtained with a STFT procedure. Left: passing trail or object. Right: Dynamic meteoroid trail. See text[]{data-label="spectra"}](exnew2.pdf "fig:"){width="0.4\linewidth"} To observe the frequency dependence of eq. \[ARl\], it should be displayed as a time dependent frequency spectrum similar to the observed data. This can be done by a Short Time Fourier Transform (STFT) of the theoretical expressions for $A_R$ in eq. \[ARl\] (see appendix \[appa\]). The results are shown in fig. \[spectra\]. First we discuss the spectrum (left) which is close to what was discussed in the previous section. The size of the object or trail passing through the specular point is taken to be $\Delta x=1$. The signal has the characteristic Doppler shift dependence with a maximum at the specular point. Superimposed is a Fraunhofer diffraction spectrum. This pattern has to occur independent of the interpretation of the object in terms of a a short-lived trail or as moving object. The spectrum on the right is for $\Delta x=\infty$, the line oscillator model of section 2. It is dominated by the ridge at $f_i - f=0$ at $t>t_0$. Its origin is the stationary trail. The diagonal in the spectrum has the same time-frequency dependence as the left-hand spectrum but without the diffraction pattern. This is the head echo from the trail as it is being formed. There are several observations to be made at this point: The Fresnel oscillations found in fig. \[powerplot\] are not seen in the spectrum. They appear by reducing the Hann window in the STFT (see appendix \[appa\]). Calculations were made for a window width of 0.5, 1 and 2 Fresnel zones. For the short window the oscillations can be seen, but at the cost of resolution in frequency. This is characteristic for a Fourier transform where time and frequency resolution exclude each other mutually. In fig. \[three\] the frequency spectrum is shown at $t=0.2$ s for the three Hann windows. The spectrum with the largest Hann window was also used to obtain fig. \[spectra\]. The branch of the head echo is clearly seen for that window, while for the smallest window it has all but disappeared due to the reduced frequency resolution. Notice that the head echo signal is more than 100 times smaller compared to the signal of the stationary trail. This indicates that measuring the head echo at $t>t_0$ requires that the receiver settings and the associated analysis programs need to take the time average in consideration. The observation of the frequency shift at $t<t_0$ is rather robust and thus also in an actual measurement. If possible, one could determine both the Fresnel oscillations and the frequency shift by analyzing the data in two different ways by good time and poor frequency resolution and vice versa, respectively. In this way one has two ways to measure the meteoroid velocity. One might expect measured events as in fig. \[ex2\] to be the most common. However, most observations are as in fig. \[ex1\]. This is either due to problems associated with measuring the head exho at $t>t_0$ for the reasons mentioned above or because the line-scattering model is not describing the meteoroid events adequately. Another reason for this can be the observational bias for strong echo’s with a short trail below the specular point. ![\[Color online\] Frequency spectrum of $\log(\mathcal{A_R})$ at $t=0.2$ s the green, orange and blue curves refer to an average over 0.5, 1, and 2 Fresnel zones, respectively. The branch of the head echo is at -700 Hz. See also text[]{data-label="three"}](logplot.pdf "fig:"){width="0.8\linewidth"}  The half-ball model =================== The characteristic asymmetric behavior of the head echo noticed at the end of the previous section can be resolved assuming reflections from the plasma in front of – and moving with – the meteoroid. Research around 2000, for example with Arecibo [@Mathews] and ALTAIR [@NASA] shows the importance of the head echo plasma. Such a plasma reflects radio waves like a metal mirror. The specular condition is then irrelevant since there will always be a spot on a sphere allowing reflection from transmitter to receiver; also in forward scattering. But in forward scattering it is important to realize that only the front half of the meteoroid has this property. The situation is sketched in fig. \[halfcanon\]. Only in the approaching phase reflections are possible up to the specular point, after that reflections would have to come from the back of the meteoroid. Therefore, in this scenario a true Doppler shift signal is observed until $t=t_0$, and after that only the trail left by the meteoroid at the specular point contributes to the signal. This scenario, which could be called the “half-ball model”, would explain experimental data as in fig. \[ex1\]. Reflections at the specular point proceed via the shortest path between transmitter and receiver and therefore also give the strongest reflection in this model, as it does in the line-oscillator model. It will be interesting to combine line-oscillator model (observing the back tail) with the “half-ball model” as the relevant trail parts do not pass the specular point at exactly the same time. In any case it is somewhat surprising that the head plasma reflection is strong enough to be seen in a comparatively modest amateur setup. ![\[Color online\] Half-ball model. An echo can be seen at any point along the meteoroid trajectory. But only from the front plasma of the meteoroid. After passing the specular point ($t=t_0$), there will be no echo any longer except from the stationary trail. The right hand side shows a calculation where this was implemented in the line-oscillator model although this is not the right model for such a scenario. []{data-label="halfcanon"}](halfcanon.pdf "fig:"){width="0.7\linewidth"}![\[Color online\] Half-ball model. An echo can be seen at any point along the meteoroid trajectory. But only from the front plasma of the meteoroid. After passing the specular point ($t=t_0$), there will be no echo any longer except from the stationary trail. The right hand side shows a calculation where this was implemented in the line-oscillator model although this is not the right model for such a scenario. []{data-label="halfcanon"}](halfballplot.pdf "fig:"){width="0.3\linewidth"} Conclusions =========== Two very different approaches lead to the same shifts in the radio-echo frequency when a meteoroid passes through a point where the specular condition is fulfilled. One assumes either thermalized electrons in the local atmosphere \[key\]or electrons co-moving with the meteoroid as the scatterers. They lead to the same frequency shifts because they are both based on the time derivative of eq. \[aberhallo\]. The model discussed in section 2 describes head echos and the resulting stationary trail in a unified way. More explicitly, further calculations in section \[calcs\] show that frequency shift and the Fresnel diffraction pattern relates to the same observable eq. \[fi\]. In section 3 it was shown that a short-lived trail and a moving object give the same frequency shift of the head echo as in section 2. The size of the object must have $|x|\lessapprox1$, i.e one Fresnel zone. The calculations show that in these cases a diffraction pattern should be visible on the head echo. In practice this is not observed. This may be because of the simplicity of the model. The calculations in section \[calcs\] allow for a qualitative comparison with actual observations. The contradicting requirements for frequency and time resolution were pointed out and it was also shown that this is reflected in the duality observing either the Fresnel diffraction or the frequency shifts. This duality may also play a role in the absence of a head echo at $t>0$ in most observations. Another reason for this can be the observational bias for strong echo’s with a short trail below the specular point. However, assuming that a co-moving plasma in front of the meteoroid is important, there is a simple geometric argument why there is no head echo after the meteoroid passes the specular point $t_0$. As a final remark and recommendation: The measurements of meteoroid echos contain much more information than the frequency shifts. In particular the rapid decay of the signal after the head echo, occurring within a second, may contain more hints as to the precise nature of the meteoroid event. Modern equipment allows a quantitative measurement of the signal strength and the classic text of [@McK] provides several methods for its analysis. Appendix STFT {#appa} ============= The Short Time Fourier Transform has here the following form $$\begin{aligned} \label{STFT} \mathcal{A}_R(f_i-f,t)&=&\int_{-\infty}^{\infty}A_R((2\pi f (t+t'), x(t+t'))e^{i2\pi(f_i-f)(t+t')}\mathrm{HannWindow}(\alpha t')dt'\ ,\\ \mathrm{HannWindow}(w)&=& \begin{array}{cc} \left\{ \begin{array}{cc} \frac{1}{2}+\frac{1}{2} \cos (2 \pi w)& -\frac{1}{2}\leq w\leq \frac{1}{2} \\ & \\ 0 & | w| >\frac{1}{2} \\ \end{array}\right . \end{array} \ ,\end{aligned}$$ where $\alpha$ determines the width of the window. Acknowledgement {#acknowledgement .unnumbered} =============== The author thanks F. Verbelen, W. Kaufmann, and M.T. German for clarifying some of the concepts in radio meteoroid measurements, for sharing their data, and for discussions. [99]{} D.W.R. McKinley, “Meteor Science and Engineering”, McGraw-Hill New York,1961. W. Kaufmann, WGN, JIMO 48 (1998)12-16 F. Verbelen, WGN,JIMO 47 (2019)49-54 M.T. German, WGN, JIMO 48(2020)4-11 F. Verbelen, private communication J.D. Mathews et al., Icarus 126(1997)157-169 R. M. Suggs et al., NASA report SEE/TP-2004-400 [^1]: University of Groningen\ Email: [email protected]\ home address: Sankt Augustin, Germany [^2]: Here $S(z)=\int_{0}^{z} \sin\frac{\pi t^2}{2} dt$ and $C(z)=\int_{0}^{z} \cos\frac{\pi t^2}{2} dt$

ArXiv

--- abstract: 'Data science is creating very exciting trends as well as significant controversy. A critical matter for the healthy development of data science in its early stages is to deeply understand the nature of data and data science, and to discuss the various pitfalls. These important issues motivate the discussions in this article.' author: - | Longbing Cao\ Advanced Analytics Institute, University of Technology Sydney, Australia\ bibliography: - 'ds-ieeeis.bib' title: 'Data Science: Nature and Pitfalls' --- PS. @headings[oddheadevenheadoddfootevenfoot]{} Data science, big data, analytics, advanced analytics, big data analytics Introduction {#sec:intro} ============ The era of analytics [@Tukey62], data science [@Tukey77], and big data [@Mckinsey11] has driven substantial governmental, industrial, and disciplinary interest, goal and strategy transformation as well as a paradigm shift in research and innovation. This has resulted in significant new opportunities and prospects becoming available which were not previously possible, while an increasing and overwhelming amount of fanfare and hype has spread across multiple domains, areas and events. In reviewing the related initiatives, progress, and status of data science, analytics and big data [@cao16-2], and the diversified discussions about prospects, challenges and directions [@Cao16ds], the controversy caused by the potential conflict of these various elements becomes clear. This review discloses the need for deep discussions about the nature and pitfalls of data science, the clarification of fundamental concepts and myths, and the demonstration of the intrinsic characteristics and opportunities of data science. This paper thus focuses on discussing the two fundamental issues in data science: the nature and the pitfalls of data science. Addressing the first highlights the status, intrinsic factors, characteristics, and features of the era of data science and analytics, as well as the challenges and opportunities for new generation innovation, research and disciplinary development. The second summarizes common pitfalls about the concepts of data science, data volume, infrastructure, analytics, and capabilities and roles. Building on the above discussions, the concepts and possible future directions of data science are then presented. Features of the Data Science Era {#sec:features} ================================ Drawing a picture of the features and characteristics of the era of data science is critical and challenging. We explore this from the perspective of the transformation and paradigm shift caused by data science, and discuss the core driving forces, as well as the status of a number of typical issues confronting the data science field. The Era’s Transformation and Paradigm Shift {#subsec:transformation} ------------------------------------------- The emergence of the era of data science and analytics can be highlighted by three key indicators: a significant *disciplinary paradigm shift*, *technological transformation*, and *innovative production*. - Disciplinary paradigm shifting: The shifting of data-centric disciplinary paradigms from one to another. - Technological transformation: The upgrading of data technology from one generation to another. - Innovative production: The innovation of technical and practical data products. These are discussed below. The disciplinary paradigm shift of data-oriented and data-centric research, innovation and profession can be embodied by such aspects as: - From data analysis to data analytics; - From descriptive analytics to deep analytics; - From data analytics to data science; The disciplinary paradigm shift promotes data-related technological transformation by such means as follows: - From large-scale data to big data; - From business operational systems to business analytical systems; - From World Wide Web to Wisdom web; - From Internet to Internet of Everything (incl. Internet of Things, mobile network, social network); Innovative production in the data and analytics areas can be represented by typical indicators such as the following: - From digital economy to data economy; - From closed government to open government; - From e-commerce to online business; - From telephone to smart phone; - From Internet to mobile and social network. Data-centric Driving Forces {#subsec:drivers} --------------------------- The transformation and paradigm shift of data-oriented discipline, technologies and production are driven by core forces including data-enabled opportunities, data-related ubiquitous factors, and various complexities and intelligences embedded in data-oriented production and products. Ubiquitous data-oriented factors include data, behavior, complexity, intelligence, service, and opportunities. - Data is ubiquitous: Involving historical, real-time and future data; - Behavior is ubiquitous: Bridging the gaps between the physical world and the data world; - Complexity is ubiquitous: The types and extent of complexity differentiate one data system from another; - Intelligence is ubiquitous: Diversified intelligences are embedded in a data system; - Service is ubiquitous: Data services present in various forms and domains; - Opportunities are ubiquitous: Data enables enormous opportunities. Data-enabled opportunities, also called *X-opportunities*, are overwhelming and extend from research, innovation, education, and government to economy. We briefly elaborate on them below. - Research opportunities: Inventing data-focused breakthrough theories and technologies; - Innovation opportunities: Developing data-based cutting-edge services, systems and tools; - Education opportunities: Innovating data-oriented courses and training; - Government opportunities: Enabling data-driven government decision-making and objectives; - Economic opportunities: Fostering data economy, services, and industrialization; - Lifestyle opportunities: Promoting data-enabled smarter living and smarter cities; - Entertainment opportunities: Creating data-driven entertainment activities, networks, and societies. A data science problem is a complex system [@Mitchel2011; @Metasynthetic15] in which comprehensive system complexities, also called *X-complexities* [@Cao16ds], are embedded. These comprise complexities of data characteristics, behavior, domain, society (social complexity), environment, learning, and decision-making. - Data complexity: Embodied by such factors as comprehensive data circumstances and characteristics; - Behavior complexity: Demonstrated by such aspects as individual and group activities, evolution, utility, impact and change; - Domain complexity: Represented by such aspects as domain factors, processes, norms, policies, knowledge and domain expert engagement in problem-solving; - Social complexity: Indicated by such aspects as social networking, community formation and divergence, sentiment, the dissemination of opinion and influence, and other social issues such as trust and security; - Environment complexity: Capturing such aspects as contextual factors, interactions with systems, changes, and uncertainty. - Learning complexity: Including the development of appropriate methodologies, frameworks, processes, models and algorithms, and theoretical foundation and explanation; - Decision complexity: Involving such issues as the methods and forms of deliverables, communications and decision-making actions. In a complex data science problem, ubiquitous intelligence, also called *X-intelligence* [@Cao16ds], is often demonstrated and has to be incorporated and synergized [@Metasynthetic15] in problem-solving processes and systems. - Data intelligence: Highlighting the interesting information, insights and stories hidden in data about business problems and driving forces. - Behavior intelligence: Demonstrating the insights of activities, processes, dynamics, impact and trust of individual and group behaviors by human and action-oriented organisms. - Domain intelligence: Domain values and insights emerge from involving domain factors, knowledge, meta-knowledge, and other domain-specific resources. - Human intelligence: Contributions made by the empirical knowledge, beliefs, intentions, expectations, critical thinking and imaginary thinking of human individual and group actors. - Network intelligence: Intelligence created by the involvement of networks, web, and networking mechanisms in problem comprehension and problem-solving. - Organizational intelligence: Insights and contributions created by the involvement of organization-oriented factors, resources, competency and capabilities, maturity, evaluation and dynamics. - Social intelligence: Contributions and values generated by the inclusion of social, cultural, and economic factors, norms and regulation. - Environmental intelligence: This may be embodied through other intelligences specific to the underlying domain, organization, society, and actors. The above data-oriented and data-driven factors, complexities, intelligences and opportunities constitute the nature and characteristics of data science, and drive the evolution and dynamics of data science problems. Data DNA {#subsec:dna} -------- In the biological domain, DNA is a molecule that carries genetic instructions that are uniquely valuable to the biological development, functioning and reproduction of humans and all known living organisms. As a result of data quantification, data is everywhere, and is present in the public Internet, Internet of Things, sensor networks, socio-cultural, economic and geographical repositories and quantified personalized sensors, including mobile, social, living, entertaining and emotional sources. This forms the “datalogical” constituent: “data DNA”, which plays a critical role in data organisms and performs a similar function to biological DNA in living organisms. *Data DNA* is the datalogical “molecule” of data, consisting of fundamental and generic constituents: entity (E), property (P) and relationship (R). Here “datalogical” means that data DNA plays a similar role in data organisms as biological DNA plays in living organisms. *Entity* can be an object, an instance, a human, an organization, a system, or a part of a sub-system of a system. *Property* refers to the attributes that describe an entity. *Relationship* corresponds to (1) entity interactions, and (2) property interactions, including property value interactions. Entity, property and relationship present different characteristics in terms of quantity, type, hierarchy, structure, distribution and organization. A data-intensive application or system is often composed of a large number of diverse entities, each of which has specific properties, and different relationships are embedded within and between properties and entities. From the very lowest level to the very highest level, data DNA presents heterogeneity and hierarchical couplings across levels. On each level, it maintains *consistency* (inheritance of properties and relationships) as well as *variations* (mutations) across entities, properties and relationships, while *personalized characteristics* are supported for each individual entity, property and relationship. For a given data, its entities, properties and relationships are instantiated into diverse and domain-specific forms, which carry most of the data ecological and genetic information in data generation, development, functioning, reproduction, and evolution. In the data world, *data DNA* is embedded in the whole body of personal [@personaldata11] and non-personal data organisms, and in the generation, development, functioning, management, analysis and use of all data-based applications and systems. Data DNA drives the evolution of a data-intensive organism. For example, university data DNA connects the data of students, lecturers, administrative systems, corporate services and operations. The student data DNA further consists of academic, pathway, library access, online access, social media, mobile service, GPS, and Wifi usage data. Such student data DNA is both steady and evolving. In complex data, data DNA is embedded within various X-complexities (see detailed discussions in [@Cao16ds] and [@Metasynthetic15]) and ubiquitous X-intelligence (more details in [@Cao16ds] and [@Metasynthetic15]) in a data organism. This makes data rich in content, characteristics, semantics and value, but challenging in acquisition, preparation, presentation, analysis and interpretation. Data Quality {#subsec:quality} ------------ Data science tasks involve roles and follow processes different from more generalized IT projects, since data science and analytics works tend to be creative, intelligent, exploratory, non-standard, unautomated and personalized, and have the objective of discovering evidence and indicators for decision-making actions. They inevitably involve quality issues such as data validity, veracity, variability and reliability, and social issues such as privacy, security, accountability and trust, which need to be taken into account in data science and analytics. *Data quality* is a critical problem in data science and engineering. Given a data science problem, we should not assume that - The data available or given is perfect, - The data always generates good outcomes, - The outputs (findings) generated are always good and meaningful, and - The outcomes can always inform better decisions. These assumption myths involve the quality of the data (input), the model, and the outcomes (output), in particular, validity, veracity, variability and reliability. We briefly discuss these aspects below. Data and analytics *validity* determines whether a data model, concept, conclusion or measurement is well-founded and corresponds accurately to the data characteristics and real-world facts, making it capable of giving the right answer. Similarly, data and analytics *veracity* determines the correctness and accuracy of data and analytics outcomes. Both validity and veracity also need to be checked from the perspectives of data content, representation, design, modeling, experiments, and evaluation. Data and analytics *variability* is determined by the changing and uncertain nature of data, reflecting business dynamics (including the problem context and problem-solving purposes), and thus requires the corresponding analytics to adapt to the dynamic nature of the data. Due to the changing nature of data, the need to check the validity, veracity and reliability of the data used and analytics undertaken is thus highly important. Data and analytics *reliability* refers to the consistency, redundancy, repeatability and trust properties of the data used, the analytic models generated, and the outcomes delivered on the data. Reliable data and analytics are not necessarily static. Making data analytics adaptive to the evolving, streaming and dynamic nature of data, business and decision requests is a critical challenge in data science and analytics. {width="65.00000%"} Social Issues {#subsec:socialissues} ------------- Domain-specific data and business are embedded in social contexts and incorporated with social issues. Data science tasks typically involve such social issues as *privacy*, *security*, *accountability* and *trust* on data, modeling and deliverables, which we discuss below. Data and analytics *privacy* addresses the challenge of collecting, analyzing, disseminating and sharing data and analytics while protecting personally identifiable or other sensitive information and analytics from improper disclosure. Protection technology, regulation and policies are required to balance protection and appropriate disclosure in the process of data manipulation. Data and analytics *security* protects target objects from destructive forces and from the unwanted actions of unauthorized users, including improper use or disclosure, and not only addresses privacy issues but also other aspects beyond privacy, such as software and hardware backup and recovery. Data and analytics security also involves the development of regulating, political or legal mechanisms and systems to address such issues. Data and analytics *accountability* refers to an obligation to comply with data privacy and security legislation, and to report, explain, trace and identify the data manipulated and analytics conducted to maintain the transparency and traceability, liability and warranty of both measurement and results, as well as the efficacy and verifiability of analytics and protection. Data and analytics *trust* refers to the belief in the reliability, truth or ability of data and analytics to achieve the relevant goals. This involves the development of appropriate technology, social norms, ethical rules, or legislation to ensure, measure and protect trust in the data and analytics used and confidence in the corresponding outcomes and evaluation of analytics. The Extreme Challenge {#subsec:extrechallenge} --------------------- Different types and levels of analytical problems trouble the existing knowledge base, and we are especially challenged by the complex problems in complex data and in complex environments. Our particular focus on data science research and innovation concerns the type of scenario we call an *extreme data challenge* in data science and analytics. The *extreme data challenge as illustrated in Fig. \[fig:five\], seeks to discover and deliver complex knowledge in complex data taking into account complex behavior within a complex environment to achieve actionable insights that will inform and enable decision action-taking in complex business problems* that cannot be better handled by other methods. The critical future directions of data science research and innovation in this case are focused on - *Complex data* with complex characteristics and various complexities (as discussed above on data complexity and intelligence; for more, see [@Cao16ds] and [@Metasynthetic15]), - *Complex behaviors* with complex relationships and dynamics (as discussed above on behavior complexity and intelligence; for more, see [@Cao16ds] and [@Metasynthetic15]), - *Complex environment* in which complex data and behaviors are embedded and interacted with (as discussed above on domain-specific, organizational, social and environmental complexities and intelligence; for more, see [@Cao16ds] and [@Metasynthetic15]), - *Complex models* to address the data and behavior complexities in complex environment (as discussed about learning complexities and decision complexities), - *Complex findings* to uncover hidden but technically interesting and business-friendly observations, indicators or evidence, statements or presentations, and - *Actionable insights* to evidence the next best or worst situation and inform the optimal strategies that should be taken to support effective business decision-making (see more discussion on actionability in [@Cao10dddm; @dddm10]. Many real-life problems fall into this level of complexities and challenges, as shown in the extreme data challenge, and they have not been addressed well; for example, - Understanding group behaviors by multiple actors where there are complex interactions and relationships, such as in the manipulation of large-scale cross-capital markets pool by internationally collaborative investors [@Cao2012cba], each of whom plays a specific role by connecting information from the underlying markets, social media, other financial markets, socio-economic data and policies [@noniid14]; - Predicting local climate change and effect by connecting local, regional and global climate data, geographical data, agricultural data and other information [@Faghmous14]. Disciplinary Development of Data Science {#sec:opp} ======================================== In this section, we present a status summary of the disciplinary development of data science by reviewing the development gaps between the potential that data may have and the state-of-the-art capabilities to fulfill such potential, the research map of data science, and the course framework of data science. Data-to-Capability Development Gaps {#subsec:gaps} ----------------------------------- The rapid increase in big data unfortunately does not only present opportunities. As discussed in [@Cao10dddm] and [@dddm10], there are significant gaps between what we are of and what we are capable of understanding. An empirical observation of the *data development gaps* between (1) the growth of *data potentials* and (2) the *state-of-the-art capabilities* is shown in Fig. \[fig:four\]. Such gaps have increased in the last 10 years and especially recently, due to the imbalance between potential exponential increase and progressive state-of-the-art capability development. We illustrate several such gaps below. - Gap between data availability and currently understandable data level, scale and degree; - Gap between data complexities and currently available analytics theories and tools; - Gap between data complexities and currently available technical capabilities; - Gap between possible values and impact and currently achievable outcomes and benefits; - Gap between organizational needs and currently available talents/data scientists; - Gap between potential opportunities and current outcomes and benefits achievable. Such growth gaps are driven by critical challenges for which there is a shortage of effective theories and tools. For example, a typical challenge in complex data concerns intrinsic complex coupling relationships and heterogeneity, forming non-IID data [@noniid14], which cannot be simplified in such a way that they can be handled by classic IID learning theories and systems. Other examples include the real-time learning of large-scale online data, such as learning shopping manipulation and making real-time recommendations on extremely high frequency data in the “11-11” shopping seasons launched by Alibaba, or identifying suspects in an extremely imbalanced and multi-source data and environment such as fraud detection in high frequency marketing trading. Other challenges are high invisibility, high frequency, high uncertainty, high dimensionality, dynamic nature, mixed sources, online learning at the web scale, and developing human-like thinking. {width="85.00000%"} Research Map of Data Science {#subsec:advanalytics} ---------------------------- The way to explore the fundamental challenges and innovative opportunities facing big data and data science is to conduct problem-, data-, and goal-driven discovery. - *Problem-driven discovery*: This requires understanding the intrinsic nature, characteristics, complexities, and boundaries of the problem, and then analyzing the gaps between the problem complexities and the existing capability set. This gap analysis is critical for original research and breakthrough scientific discovery. - *Goal-driven discovery*: This requires understanding the business, technical and decision goals to be achieved by understanding the problem, and conducting gap analysis of what has been implemented and achieved and what is expected to be achieved. - *Data-driven discovery*: This requires understanding the data characteristics, complexities and challenges in data, and the gaps between the nature of a problem and the data capabilities. Due to the limitations of existing data systems, projection from the underlying physical world where the problem sits to the data world where the problem is datafied may be biased, dishonest, or highly manipulated. As a result, the data does not completely capture the problem and thus cannot create a full picture of of the through any type of data exploration. There are two ways to explore major research challenges: one is to summarize what concerns the relevant communities, and the other is to scrutinize the potential issues arising from the intrinsic complexities and nature of data science problems as complex systems [@Metasynthetic15], [@Metasynthetic15]. Taking the first approach, we can obtain a picture of the main research challenges by summarizing the main topics and issues in the statistics communities [@Chambers93], [@wu97], [@amstatnews15], informatics and computing communities [@Rudin14], [@Cao16ds], vendors [@Stonebraker13], government initiatives [@USNSF], [@UNpulse], [@CNbd], [@ECbd14], [@UKbd] and research institutions [@UTSAAI], [@IAA], which focus on data science and analytics. The second approach is much more challenging, as it requires us to explore the unknown space of the complexities and comprehensive intelligence in complex data problems. Below, we list some of the main challenges confronting the data science community in addressing the big data complexities represented by key topics in the data A-Z (see Section \[subsec:dataa-z\]). We categorize the challenges facing domain-specific data applications and problems in terms of the following major areas: - Challenges in data/business understanding, - Challenges in mathematical and statistical foundations, - Challenges in X-analytics and data/knowledge engineering, - Challenges in data quality and social issues, - Challenges in data value, impact and usability, - Challenges in data-to-decision and actions. X-analytics and data/knowledge engineering encompass many specific research issues that have not been addressed properly; for example: - Behavior and event processing, - Data storage and management systems, - Data quality enhancement, - Data modeling, learning and mining, - Deep analytics, learning and discovery, - Simulation and experimental design, - High-performance processing and analytics, - Analytics and computing architectures and infrastructure, - Networking, communication and interoperation. Course Framework of Data Science {#subsec:courseframework} -------------------------------- The goal of data science and analytics education is to train and generate the data and analytics knowledge and proficiency required to manage the capability and capacity gaps in the creation of a data science profession [@Walker15; @Manieri15-1], and to achieve the goals of data science innovation and the data economy. Accordingly, different levels of education and training are necessary, from attending public courses, corporate training, and undergraduate courses, to joining a master of data science and/or PhD in data science program. *Public courses* are designed for the general community, to lift their understanding, skills, profession and specialism in data science through multi-level short courses. They range from basic courses to intermediary and advanced courses. The knowledge map consists of such components as data science, data mining, machine learning, statistics, data management, computing, programming, system analysis and design, and modules related to case studies, hands-on practices, project management, communication, and decision support. *Corporate training* and workshops are customized to upgrade and foster corporate thinking, knowledge, capability and practices for entire enterprise innovation and raising productivity. This involves offering courses and workshops for the workforce, from senior corporate executives to business owners, business analysts, data modelers, data scientists, data engineers, and deployment and enterprise strategists. Such courses cover scope and topics such as data science, data engineering, analytics science, decision science, data and analytics software engineering, project management, communications, and case management. *Undergraduate courses* may be offered on either a general data science basis that focuses on building data science foundations and data and analytics computing, or specific areas such as data engineering, predictive modeling, and visualization. Double degrees or majors may be offered to train professionals who will gain knowledge and abilities across disciplines such as business and analytics, or statistics and computing. *Master of data science and analytics* aims to train specialists and foster the talent of those who have the capacity to conduct a deep understanding of data and undertake analytics tasks in data mining, knowledge discovery and machine learning-based advanced analytics. Interdisciplinary experts may be trained from those who have a solid foundation in statistics, business, social science or other specific disciplines and are able to integrate data-driven exploration technologies with disciplinary expertise and techniques. A critical area in which data science and analytics should be incorporated is the classic master of business administration course. This is where new generation business leaders can be trained for the new economy and a global view of economic growth. *PhD in data science and analytics* aims to train high level talent and specialists who have independent thinking, leadership, research, innovation and better practices for theoretical innovation to manage the significant knowledge and capability gaps, and for substantial economic innovation and raising productivity. Interdisciplinary research is encouraged to train leaders who have a systematic and strategic understanding of the what, how and why of data and economic innovation. Fig. \[fig:12\] shows the level, objective, capability set and outcomes of hierarchical data science and analytics education and training. {width="85.00000%"} Data Science as A New Science {#sec:concepts} ============================= So, what makes data science a new science? In this section, we discuss *data A-Z*, which may be used to capture every aspect of data science to form a data science ontological system, *the concept of data science*, which is built on the above discussions about the features and disciplinary development of data science, and *the future of data science*. Data A-Z {#subsec:dataa-z} -------- In the big data community, multiple Vs are typically used to describe what constitutes big data, i.e., the characteristics, challenges and opportunities of big data. They include the volume (size), velocity (speed), variety (diversity), veracity (quality and trust), value (insight), visualization, and variability (formality) of data. In fact, these big Vs cannot describe a complete picture of big data, and cannot capture the field of data science. Therefore, it is very valuable to build a *data A-Z dictionary* to represent and capture the intrinsic comprehensive but diverse aspects, characteristics, challenges, domains, tasks, processes, purposes, applications and outcomes of, or related to, data. To this end, we list a sample sequence of data science keywords: $$\begin{aligned} Actionability/Adaptation, Behavior/Boosting, \nonumber \\ Causality/Change, Dimensionality/Divergence, \nonumber \\ Embedding/Ethics, Fusion/Forecasting, \nonumber \\ Governance/Generalization, Heterogeneity/ \nonumber \\ Hashing, Integrity/Inference, Join/Jungle, \nonumber \\ Kernelization/Knowledge, Linkage/Learning, \nonumber \\ Metrology/Migration, Normalization/Novelty, \nonumber \\ Optimization/Outlier, Privacy/Provenance, \nonumber \\ Quality/Quantity, Relation/Regularization, \nonumber \\ Scalability/Sparsity, Transformation/Transfer, \nonumber \\ Utility/Uncertainty, Variation/Visibility, \nonumber \\ Wrangling/Weighting, X-analytics/ \nonumber \\ X-informatics, Yield/Yipit, Zettabyte/Zenit. \end{aligned}$$ It is notable that such a data A-Z ontology probably covers most of the topics of interest to major data science communities. The exercise of constructing data A-Z can substantially deepen and broaden the understanding of intrinsic data characteristics, complexities, challenges, prospects and opportunities [@Geczy14]. What Is Data Science {#subsec:ds} -------------------- Generally speaking, *data science is the science of data*, which concerns the study of data. There are different ways to define what data science is: it may be object-focused, process-based, and/or discipline-oriented [@Cao16ds]. From the *process* perspective, data science is a systematic approach to ‘think with wisdom’, ‘understand domain’, ‘manage data’, ‘compute with data’, ‘mine on knowledge’, ‘communicate with stakeholders’, ‘deliver products’, and ‘act on insights’. A *process-based data science formula* is accordingly given below: $$\begin{aligned} data~science = think + understand + manage + \nonumber \\ compute + mine + communicate + deliver + act \end{aligned}$$ In contrast, *data analytics understands data and its underlying business, discovers knowledge, delivers actionable insights, and enable decision-making*. From this perspective, we can say that analytics is a keystone of data science. From the *disciplinary* perspective, *data science is a new interdisciplinary field* in which to study data and its domain in terms of a data-to-knowledge-to-wisdom thinking for generating data products [@Cao16ds]. Data science integrates *traditionally data-oriented disciplines* such as statistics, informatics and computing with *traditionally data-independent fields* such as communication, management and sociology. The Future of Data Science {#subsec:dsfuture} -------------------------- It is difficult at this very early stage of data science to predict specific future data science innovation and research, thus the next-generation data science will need to address the unknown space that is currently invisible to existing science and create new data products. We will need to: - Deepen our *understanding of data invisibility* (i.e., *invisible data characteristics*) in the hidden and blind spaces (Spaces B and D in Fig. 1 [@Cao16ds]), to understand their X-complexities (see [@Cao16ds]) and X-intelligence (see [@Cao16ds]), and strengthen our capabilities; - Invent *new data representation capabilities*, including designs, structures, schemas and algorithms to make invisible data in Spaces B and D in Fig. 1 [@Cao16ds] more visible and explicit; - Create *new analytical and learning capabilities*, including original theories, algorithms and models, to disclose the unknown knowledge in unknown Space D in Fig. 1 [@Cao16ds]; - Build new intelligent systems and services, including corporate and Internet-based collaborative platforms and services, to support collaborative and collective exploration of invisible and unknown challenges in the fully unknown space D in Fig. 1 [@Cao16ds]. - Train a generation of qualified data science professionals in data literacy, thinking, competency, consciousness and cognitive intelligence to work on the above data science agenda. Pitfalls in Data Science {#sec:pitfall} ======================== A typical feature of data science being at this very early stage is that different and sometimes contradictory views appear in various communities. It is essential to share and discuss the myths and reality [@Jagadish15], memes [@Donoho15], and pitfalls to ensure the healthy development of the field. Based on our observations about the relevant communities, and our experience and lessons learned in conducting data science and analytics research, education and services, we list the following myths and pitfalls for discussion. About Data Science Concepts {#subsec:dsc} --------------------------- Typically, data science has been defined in terms of specific disciplinary foundations, principles, goals, inputs, algorithms and models, processes, tools, outputs, applications, and/or professions. Often, a fragmented statement is given, which may cause debate and result in the phenomenon of “how does a bind person recognize an elephant?” In this section, we discuss some common arguments and observations. - Data science is statistics [@Broman13; @Diggle15]; “why do we need data science when we’ve had statistics for centuries” [@Wladawsky14]? How does data science really differ from statistics [@Donoho15]? (Comments: Data science provides systematic, holistic and multi-disciplinary solutions for learning explicit and implicit insights and intelligence from complex and large-scale data and generates evidence or indicators from data by undertaking diagnostic, descriptive, predictive and/or prescriptive analytics, in addition to supporting other tasks on data such as computing and management.) - Why do we need data science when information science and data engineering have been explored for many years? (Comments: Consider the issues faced in related areas by the enormity of the task and the parallel example of enabling a blind person to recognize an animal as large as an elephant. Information science and data engineering alone cannot achieve this. Other aspects may be learned from the discussion about greater or fewer statistics; more in [@Chambers93].) - I have been doing data analysis for dozens of years; data science has nothing new to offer me. (Comments: Classic data analysis and technologies focus mostly on explicit observation analysis and hypothesis testing on small and simpler data.) - Is data science old wine in a new bottle? What are the new grand challenges foregrounded by data science? (Comments: The analysis of the gaps between existing developments and the potential of data science (see Fig. \[fig:four\]) shows that many opportunities can be found to fill the theoretical gaps when data complexities extend significantly beyond the level that can be handled by the state-of-the-art theories and systems, e.g., classic statistical and analytical theories and systems were not designed to handle the non-IIDness [@noniid14] in complex real-life systems.) - Data science mixes statistics, data engineering and computing, and does not contribute to breakthrough research. (Comments: Data science attracts attention because of the significant complexities in handling complex real-world data, applications and problems that cannot be addressed well by existing statistics, data engineering and computing theories and systems. This drives significant innovation and produces unique opportunities to generate breakthrough theories.) - Data science is also referred to as data analytics and big data [@Anderson14]. (Comments: This confuses the main objectives, features, scopes of the three concepts and areas. Data science needs to be clearly distinguished from both data analytics and big data.) - Other definitions ascribed to data science are that it is big data discovery [@Davenport12], prediction [@Dhar13], or the combination of principle and process with technique [@Provost13]. It is also worth noting attention that the terms “big data”, “data science” and “advanced analytics” are often extensively misused, over-used or improperly used by diverse communities and for various purposes, particularly given the influence of media hype and buzz. A large proportion of Google searches on these keywords returns results that are irrelevant to their intrinsic semantics and scope, or simply repeat familiar arguments about the needs of data science and existing phenomena. In many such findings [@Brown; @Stanton12; @Chawla14; @Faris11; @Anya15; @Dierick15; @Priebe15; @Kanter15; @Kirkpatrick15; @Clancy14; @Stevens14; @Miller13; @Pal15; @Moraes15; @Horton15; @Loukides12; @Neil13; @Hand15; @Hazena14; @Manieri15-2; @Cuzzocrea13; @Casey15; @Siart15; @Gold13; @Gupta15], big data is described as being simple, data science has nothing to do with the science of data, and advanced analytics is the same as classic data analysis and information processing. There is a lack of deep thinking and exploration of why, what and how these new terms should be defined, developed and applied. The above observations strongly illustrate that data science is still in a very early stage. They also justify the urgent need to develop sound terminology, standards, a code of conduct, statement and definitions, theoretical frameworks, and better practices that will exemplify typical data science professional practices and profiles. About Data Volume {#subsec:dv} ----------------- - What makes data “big”? (Comments: It is usually not the volume but the complexities (as discussed in [@Cao16ds; @Metasynthetic15]) and large values that make data big.) - Why is the bigness of data important? (Comments: The bigness (referring to data science complexities) of data heralds new opportunities for theoretical, technological, practical, economic and other development or revolution.) - Big data refers to massive volumes of data. (Comments: Here, “big” refers mainly to significant data complexities. From the volume perspective, a data set is big when the size of the data itself becomes a quintessential part of the problem.) - Data science is big data analytics. (Comments: Data science is a comprehensive field centered on manipulating data complexities and extracting intelligence, in which data can be big or small and analytics is a core component and task.) - I do not have big data so I cannot do big data research. (Comments: Most researchers and practitioners do not have sizeable amounts of data and do not have access to big infrastructure either. However, significant research opportunities still exist to create fundamentally new theories and tools to address respective X-complexities and X-intelligence.) - The data I can find is small and too simple to be explored. (Comments: While scale is a critical issue in data science, small data, which is widely available, may still incorporate interesting data complexities that have not been well addressed. Often, we see experimental data, which is usually small, neat and clean. Observational data from real business is live, complex, large and frequently messy.) - I am collecting data from all sources in order to conduct big data analytics. (Comments: Only relevant data is required to achieve a specific analytical goal.) - It is better to have too much data than too little. (Comments: While more data generally tends to present more opportunities, the data amount needs to be relevant to the data needed and the data manipulation goals. Whether bigger is better depends on many aspects.) About Data Infrastructure {#subsec:di} ------------------------- - I do not have big infrastructure, so I cannot do big data research. (Comments: While big infrastructure is useful or necessary for some big data tasks, theoretical research on significant challenges may not require big infrastructure.) - My organization will purchase a high performance computer to support big data analytics (Comments: Many big data analytics tasks can be done without a high performance computer. It is also essential to differentiate between distributed/parallel computing and high performance computing.) About Analytics {#subsec:ana} --------------- - Thinking data-analytically is crucial for data science. (Comments: Data-analytic thinking is not only important for a specific problem-solving, but is essential for obtaining a systematic solution and for a data-rich organization. Converting an organization to think data analytically is a critical competitive advantage in the data era.) - The task of an analyst is mainly to develop common task frameworks and conduct inference [@Breiman01] from the particular to the general. (Comments: Analytics in the real world is often specific. Focusing on certain common task frameworks may trigger incomplete or even misleading outcomes. As discussed in Section \[subsec:ds\], an analyst may take other roles, e.g., predictive modeling is typically problem-specific.) - I only trust the quality of models built in commercial analytical tools. (Comments: Such tools may produce misleading or even incorrect outcomes if the assumption of their theoretical foundation does not fit the data, e.g., if they only suit imbalanced data, normal distribution-based data, or IID data.) - Most published models and algorithms and their experimental outcomes are not repeatable. (Comments: Such works seem to be more hand-crafted rather than manufactured. Repeatability, reproducibility, open data and data sharing are critical to the healthy field development.) - I want to do big data analytics, can you tell me which algorithms and program language I should learn? (Comments: Public survey outcomes (see examples in [@cao16-2]) give responses to such questions. Which algorithms, language and platform should be chosen also depends on organizational maturity and needs. For long-term purposes, big data analytics is about building competencies rather than specific functions). - My organization’s data is private and thus you cannot be involved in our analytics. (Comments: Private data can still be explored by external parties by implementing proper privacy protection and setting up appropriate policies for onsite exploration.) - Let me (an analyst) show you (business people) some of my findings which are statistically significant. (Comments: As domain-driven data mining [@dddm10] shows, many outcomes are often statistically significant but are not actionable. An evaluation of those findings needs to be conducted to discover what business impact [@cao08-3] might be generated if the findings they may generate are operationalized.) - Strange, why can I not understand and interpret the outcomes? (Comments: This may be because the problem has been misstated, the model may be invalid for the data, or the data used is not relevant or correct.) - Your outcomes are too empirical without theoretical proof and foundation. (Comments: While it would be ideal if questions about the outcomes could be addressed from theoretical, optimization and evaluation perspectives, real-life complex data analytics may often be more exploratory and it may initially be difficult to optimize empirical performance.) - My analysis shows what you delivered is not the best for our organization. (Comments: It may be challenging to claim “the best” when a variety of models, workflows and data features are used in analytics. It is not unusual for analysts to obtain different or contradictory outcomes on the same data as a result of the application of different theories, settings and models. It may turn out to be a very challenging job to find a solid model that perfectly and stably fits the invisible aspect of data characteristics. It is important to appropriately check the relevance and validity of the data, models, frameworks and workflows available and used. Doing the right thing at the right time for the right purpose is a very difficult task when attempting to understand complex real-life data and problems.) - Can your model address all of my business problems? (Comments: Different models are often are required to address diverse business problems, as a single model cannot handle a problem sufficiently well.) - This model is very advanced with solid theoretical foundation, let us try it in your business. (Comments: While having solid scientific understanding of a model is important, it is data-driven discovery may better capture the actual data characteristics in real-life problem solving. A model may be improperly used without a deep understanding of model and data suitability. Combining data driven approaches with model driven approaches may be more practical.) - My analytical reports consist of lots of figures and tables that summarize the data mining outcomes, but my boss seems not so interested in them. (Comments: Analytics is not just about producing meaningful analytical outcomes and reports; rather, it concerns insights, recommendations and communication with upper management for decision-making and action.) - It is better to have advanced rather than simple models. (Comments: Generally, simpler is better. The key to deploying a model is to fit the model to the data while following the same assumption taken by the model.) - We just tuned the models last month, but again they do not work well. (Comments: Monitoring a model’s performance by watching the dynamics and significant change that may take place in the data and business is critical. Real-time analytics requires adaptive and automated re-learning and adjustment.) - I designed the model, so I trust the outcomes. (Comments: The reproducibility of model outcomes relies on many factors. A model that is properly constructed may fall short in other aspects such as data leakage, overfitting, insufficient data cleaning, and poor understanding of data characteristics and business. Similarly, a lack of communication with the business may cause serious problems in the quality of the outcome.) - Data science and analytics projects are just other kinds of IT projects. (Comments: While data projects share many similar aspects to mainstream IT projects, certain distinctive features in data, the manipulation process, delivery, and especially the exploratory nature of data science and analytics projects require different strategies, procedures and treatments. Data science projects are more exploratory, ad hoc, decision-oriented and intelligence-driven.) About Capabilities and Roles {#subsec:cap} ---------------------------- - I am a data scientist. (Comments: Lately, it seems that everyone has suddenly become a data scientist. Most data scientists simply conduct normal data engineering and descriptive analytics. Do not expect omnipotence from data scientists.) - “A human investigative journalist can look at the facts, identify what’s wrong with the situation, uncover the truth, and write a story that places the facts in context. A computer can’t.” [@Kirkpatrick15] (Comments: The success of AlphaGo [@deepmind] may show the potential that a data science-enabled computer has to undertake a large proportion of the job a journalist does.) - My organization wants to do big data analytics, can you recommend some of your PhD graduates to us? (Comments: While data science and advanced analytics tasks usually benefit from the input of PhDs, an organization requires different roles and competencies according to the maturity level of the analytics and the organization.) - Our data science team consists of a group of data scientists. (Comments: An effective data science team may consist of statisticians, programmers, physicists, artists, social scientists, decision-makers, or even entrepreneurs.) - A data scientist is a statistical programmer. (Comments: In addition to the core skills of coding and statistics, a data scientist needs to handle many other matters; see discussions in [@cao16-2].) Other Matters ------------- In addition to the above aspects, there are other matters that require careful consideration in conducting data science and analytics. We list some here. - Garbage in, garbage out. (Comments: The quality of data determines the quality of output.) - More complex data, a more advanced model, and better outcomes. (Comments: Good data does not necessarily lead to good outcomes; A good model also does not guarantee good outcomes.) - More general models, better applicability. (Comments: General models may lead to weaker outcomes on a specific problem. It is not reasonable or practical to expect a single tool for all tasks.) - More frequent patterns, more interesting. (Comments: It has been shown that frequent patterns mined by existing theories are generally not useful and actionable.) - We’re interested in outcomes, not theories. (Comments: Actionable outcomes may need to satisfy both technical and business significance [@dddm10].) - The goal of analytics is to support decision-making actions, not just to present outcomes about data understanding and analytical results. (Comments: This addresses the need for actionable knowledge delivery [@Cao10dddm] to recommend actions from data analytics for decision-support.) - Whatever you do, you can at least get some values. (Comments: This is true, but it may be risky or misleading. Informed data manipulation and analytics requires a foundation for interpreting why the outcomes look the way they do.) - Many end users are investing in big data infrastructure without project management. (Comments: Do not rush into data infrastructure investment without a solid strategic plan of your data science initiatives, which requires the identification of business needs and requirements, the definition of reasonable objectives, the specification of timelines, and the allocation of resources.) - Pushing data science forward without suitable talent. (Comments: On one hand, you should not simply wait for the right candidate to come along, but should, actively plan and specify the skills needed for your organization’s initiatives and assemble a team according to the skill-sets required. On the other hand, getting the right people on board is critical, as data science is essentially about intelligence and talent.) - No culture for converting data science insights into actionable outcomes. (Comments: This may be common in business intelligence and technically focused teams. Fostering a data science-friendly culture requires a top-down approach driven by business needs, and making data-driven decisions that enable data science specialists and project managers to be part of the business process, and to conduct change management.) - Correct evaluation of outcomes. (Comments: This goes far beyond such technical metrics as Area Under the ROC Curve and Normalized Mutual Information. Business performance after adopting the recommended outcomes needs to be evaluated [@Cao10dddm]. For example, recent work on high utility analysis [@yin12] and high impact behavior analysis [@cao08-3] study how business performance can be taken into data modeling and evaluation account. Solutions that lack business viability are not actionable.) - Apply a model in a consistent way. (Comments: It is essential to understand the hypothesis behind a model and to apply a model consistent with its hypothesis.) - Overthinking and overusing models. (Comments: All models and methods are specific to certain hypotheses and scenarios. No models are universal and sufficiently “advanced” to suit everything. Do not assume that if the data is tortured long enough, it will confess to anything.) - Know nothing about the data before applying a model. (Comments: Data understanding is a must-do step before a model is applied.) - Analyze data for the sake of analysis only. (Comments: This involves the common bad practice of overusing analytics.) - What makes an insight (knowledge) actionable? (Comments: This is dependent on not only the statistical and practical values of the insight, but also predictive power and business impact.) - Do not assume the data you are given is perfect. (Comments: Data quality forms the basis of obtaining good models, outcomes and decisions. Poor quality data, the same as poor quality models, can lead to misleading or damaging decisions. Real-life data often contains imperfect features such as incompleteness, uncertainty, bias, rareness, imbalance and non-IIDness.) Conclusions {#sec:concl} =========== In the era of data science, big data and advanced analytics, numerous debates have emerged from a wide range of backgrounds, domains, areas and perspectives and for diversified reasons and purposes. It is difficult but critical to explore the nature of data science. To do so, a fundamental perspective is to explore the intrinsic characteristics, challenges, working mechanisms, and dynamics of data and the science about data. As part of our comprehensive review of data science [@cao16-1; @cao16-2; @Cao16ds], the discussions about the nature and pitfalls of data science in this work will hopefully stimulate deep and intrinsic discussions about what makes data science a new science, and what makes data science valuable for research, innovation, the economy, services and professionals. Acknowledgment {#acknowledgment .unnumbered} ============== This work is partially sponsored by the Australian Research Council Discovery Grant (DP130102691).

ArXiv

--- abstract: 'Wintgen ideal submanifolds in space forms are those ones attaining equality at every point in the so-called DDVV inequality which relates the scalar curvature, the mean curvature and the normal scalar curvature. This property is conformal invariant; hence we study them in the framework of Möbius geometry, and restrict to three dimensional Wintgen ideal submanifolds in $\mathbb{S}^5$. In particular we give Möbius characterizations for minimal ones among them, which are also known as (3-dimensional) austere submanifolds (in 5-dimensional space forms).' author: - 'Zhenxiao Xie, Tongzhu Li, Xiang Ma, Changping Wang' title: '**Möbius geometry of three dimensional Wintgen ideal submanifolds in $\mathbb{S}^{5}$**' --- [**2000 Mathematics Subject Classification:**]{} 53A30, 53A55, 53C42. [**Key words:**]{} Wintgen ideal submanifolds, DDVV inequality, Möbius geometry, austere submanifolds, complex curves\ 1 cm Introduction ============ The so-called DDVV inequality says that, given a $m$-dimensional submanifold $x:M^m\longrightarrow \mathbb{Q}^{m+p}(c)$ immersed in a real space form of dimension $m+p$ with constant sectional curvature $c$, at any point of $M$ we have $$\label{1.1} s\leq c+||H||^2-s_N.$$ Here $s=\frac{2}{m(m-1)}\sum_{1\leq i<j\leq n}\langle R(e_i,e_j)e_j,e_i\rangle$ is the normalized scalar curvature with respect to the induced metric on $M$, $H$ is the mean curvature, and $s_N=\frac{2}{m(m-1)}||\mathbb{R}^{\perp}||$ is the normal scalar curvature. This remarkable inequality was first a conjecture due to De Smet, Dillen, Verstraelen and Vrancken [@Smet] in 1999, and proved by J. Ge, Z. Tang [@Ge] and Z. Lu [@Lu1] in 2008 independently. As pointed out in [@Dajczer2][@Smet][@Lu1][@Lu3], it is a natural and important problem to characterize the extremal case, i.e., those submanifolds attaining the equality at every point, called *Wintgen ideal submanifolds*. In [@Ge] it was shown that the equality holds at $x\in M^m$ if and only if there exist an orthonormal basis $\{e_1,\cdots,e_m\}$ of $T_xM^m$ and an orthonormal basis $\{n_1,\cdots,n_p\}$ of $T_x^{\bot}M^m$ such that the shape operators $\{A_{n_i},i=1,\cdots,m\}$ have the form $$\label{form1} A_{n_1}= \begin{pmatrix} \lambda_1 & \mu_0 & 0 & \cdots & 0\\ \mu_0 & \lambda_1 & 0 & \cdots & 0\\ 0 & 0 & \lambda_1 & \cdots & 0\\ \vdots & \vdots & \vdots & \ddots & \vdots\\ 0 & 0 & 0 & \cdots & \lambda_1 \end{pmatrix},~~ A_{n_2}= \begin{pmatrix} \lambda_2\!+\!\mu_0 & 0 & 0 & \cdots & 0\\ 0 & \lambda_2\!-\!\mu_0 & 0 & \cdots & 0\\ 0 & 0 & \lambda_2 & \cdots & 0\\ \vdots & \vdots & \vdots & \ddots & \vdots\\ 0 & 0 & 0 & \cdots & \lambda_2 \end{pmatrix},$$ and $$A_{n_3}=\lambda_3I_p,~~~~ A_{n_r}=0, r\ge 4.$$ This is the first step towards a complete classification. Wintgen [@wint] first proved the inequality for surfaces $M^2$ in $\mathbb{R}^4$, and that the equality holds if and only if the curvature ellipse of $M^2$ in $\mathbb{R}^4$ is a circle. Such surfaces are called *super-conformal* surfaces. They come from projection of complex curves in the twistor space $\mathbb{C}P^3$ of $\mathbb{S}^4$ [@fb]. Together with totally umbilic submanifolds (spheres and planes), they provide the first examples of Wintgen ideal submanifolds. Note that they are not necessarily minimal surfaces in space forms. In particular, being super-conformal is a conformal invariant property, whereas being minimal is not. The conformal invariance of Wintgen ideal property in the general case was pointed out in [@Dajczer1]. Thus it is appropriate to investigate and classify Wintgen ideal submanifolds under the framework of Möbius geometry. For this purpose, the submanifold theory in Möbius geometry established by the fourth author will be briefly reviewed in Section 2. We will always assume that the Wintgen ideal submanifolds in consideration are not totally umbilic. Note that to have the shape operators taking the form in , the distribution $\mathbb{D}=\mathrm{Span}\{e_1,e_2\}$ is well-defined. We call it *the canonical distribution*. The first Möbius classification result was obtained by us in [@Li1].\ [**Theorem A(Li-Ma-Wang[@Li1]):**]{} *Let $x:M^m\to\mathbb{S}^{m+p}(m\geq3)$ be a Wintgen ideal submanifold and it is not totally umbilic. If the canonical distribution $\mathbb{D}=\mathrm{Span}\{e_1,e_2\}$ is integrable, then locally $x$ is Möbius equivalent to either one of the following three kinds of examples described in $\mathbb{R}^{m+p}$:\ (i) a cone over a minimal Wintgen ideal surface in ${\mathbb S}^{2+p}$;\ (ii) a cylinder over a minimal Wintgen ideal surface in ${\mathbb R}^{2+p}$;\ (iii) a rotational submanifold over a minimal Wintgen ideal surface in ${\mathbb H}^{2+p}$.*\ In this paper we consider three dimensional Wintgen ideal submanifolds $x:M^3\to\mathbb{S}^5$ whose canonical distribution $\mathbb{D}$ is not integrable. There is a Möbius invariant 1-form $\omega$ associated with $x$. For its definition as well as other basic equations and invariants, see Section 3. Our main result is stated as below, which is proved in Section 4.\ [**Theorem B:**]{} *Suppose $x:M^3\to\mathbb{S}^5$ is a Wintgen ideal submanifold whose canonical distribution $\mathbb{D}$ is not integrable. It is Möbius equivalent to a minimal Wintgen ideal submanifold in a five dimensional space form $\mathbb{Q}^5(c)$ if and only if the 1-form $\omega$ is closed.*\ Under some further conditions, in Section 5 we characterize minimal Wintgen ideal submanifolds coming from Hopf bundle over complex curves in $\mathbb{C}P^2$. We also discuss the classification of Möbius homogeneous ones among Wintgen ideal 3-dimensional submanifolds in $\mathbb{S}^5$, which include the following example: $$x: \mathrm{SO}(3)~ \longrightarrow~\mathbb{S}^5,~~~ (u, v, u\times v) \mapsto \frac{1}{\sqrt{2}}(u, v).$$ As to the geometric meaning of the 1-form $\omega$, we just mention that it could still be defined for Wintgen ideal submanifolds with dimension $m\ge 4$. In a forthcoming paper [@Li2] we will show that $d\omega=0$ is equivalent to the property that $\mathbb{D}=\mathrm{Span}\{e_1,e_2\}$ generates a 3-dimensional integrable distribution on $M^3$. Assume this is the case; then we will obtain a similar classification [@Li2] as in Theorem A. These results again demonstrate the phenomenon described by our reduction theorem [@Li0]. To understand the classification result, it is necessary to note that among Wintgen ideal submanifolds, there are a lot of minimal examples in space forms. Although they do not exhaust all possible examples, our classification demonstrates their importance as being representatives in a Möbius equivalence class of submanifolds, or as building blocks of generic examples. Those minimal Wintgen ideal surfaces are called *super-minimal* in the previous literature, including examples like complex curves in $\mathbb{C}^n$ and minimal 2-spheres in $\mathbb{S}^n$. For three dimensional submanifolds in 5-dimensional space forms $\mathbb{S}^5,\mathbb{R}^5,\mathbb{H}^5$, being minimal and Wintgen ideal is equivalent to being *austere submanifolds*, i.e. the eigenvalues of the second fundamental form with respect to any normal direction occur in oppositely signed pairs. Such submanifolds have been classified locally by Bryant [@br] for $M^3\to\mathbb{R}^5$ , by Dajczer and Florit [@Dajczer3] for $M^3\to\mathbb{S}^5$, and by Choi and Lu [@Lu] for $M^3\to\mathbb{H}^5$. Finally we note that in [@Dajczer1], Dajczer and Florit have provided a parametric construction of Wintgen ideal submanifolds of codimension two and arbitrary dimension in terms of minimal surfaces in $\mathbb{R}^{m+2}$. Compared to our work, they had no restriction on the dimension of $M$, and the construction is explicit and valid for generic examples. On the other hand, their descriptions were not in a Möbius invariant language. In another paper [@Li3], we will give a construction of all Wintgen ideal submanifolds of codimension two and arbitrary dimension $m$ in terms of holomorphic, isotropic curves in a complex quadric $Q^{m+2}$.\ **Acknowledgement** This work is funded by the Project 10901006 and 11171004 of National Natural Science Foundation of China. We thank Professor Zizhou Tang for pointing out the homogeneous embedding of $\mathrm{SO}(3)$ in $\mathbb{S}^5$ to us. We are grateful to the referees for their helpful suggestions. 1 cm Submanifold theory in Möbius geometry ===================================== In this section we briefly review the theory of submanifolds in Möbius geometry. For details we refer to [@CPWang], [@liu]. Recall that in the classical light-cone model, the light-like (space-like) directions in the Lorentz space $\mathbb{R}^{m+p+2}_1$ correspond to points (hyperspheres) in the round sphere $\mathbb{S}^{m+p}$, and the Lorentz orthogonal group correspond to conformal transformation group of $\mathbb{S}^{m+p}$. The Lorentz metric is written out explicitly as $$\langle Y,Z\rangle=-Y_0Z_0+Y_1Z_1+\cdots+Y_{m+p+1}Z_{m+p+1},$$ for $Y=(Y_0,Y_1,\cdots,Y_{m+p+1}), Z=(Z_0,Z_1,\cdots,Z_{m+p+1})\in \mathbb{R}^{m+p+2}_1$. Let $x:M^m\rightarrow \mathbb{S}^{m+p}\subset \mathbb{R}^{m+p+1}$ be a submanifold without umbilics. Take $\{e_i|1\le i\le m\}$ as the tangent frame with respect to the induced metric $I=dx\cdot dx$, and $\{\theta_i\}$ as the dual 1-forms. Let $\{n_{r}|1\le r\le p\}$ be an orthonormal frame for the normal bundle. The second fundamental form and the mean curvature of $x$ are $$\label{2.1} II=\sum_{ij,r}h^{r}_{ij}\theta_i\otimes\theta_j n_{r},~~H=\frac{1}{m}\sum_{j,r}h^{r}_{jj}n_{r}=\sum_{r}H^{r}n_{r},$$ respectively. We define the Möbius position vector $Y: M^m\rightarrow \mathbb{R}^{m+p+2}_1$ of $x$ by $$\label{2.2} Y=\rho(1,x),~~~ ~~\rho^2=\frac{m}{m-1}\left|II-\frac{1}{m} tr(II)I\right|^2~.$$ $Y$ is called *the canonical lift* of $x$ [@CPWang]. Two submanifolds $x,\bar{x}: M^m\rightarrow \mathbb{S}^{m+p}$ are Möbius equivalent if there exists $T$ in the Lorentz group $O(m+p+1,1)$ such that $\bar{Y}=YT.$ It follows immediately that $$\mathrm{g}=\langle dY,dY\rangle=\rho^2 dx\cdot dx$$ is a Möbius invariant, called the Möbius metric of $x$. Let $\Delta$ be the Laplacian with respect to $\mathrm{g}$. Define $$N=-\frac{1}{m}\Delta Y-\frac{1}{2m^2} \langle \Delta Y,\Delta Y\rangle Y,$$ which satisfies $$\langle Y,Y\rangle=0=\langle N,N\rangle, ~~ \langle N,Y\rangle=1~.$$ Let $\{E_1,\cdots,E_m\}$ be a local orthonormal frame for $(M^m,\mathrm{g})$ with dual 1-forms $\{\omega_1,\cdots,\omega_m\}$. Write $Y_j=E_j(Y)$. Then we have $$\langle Y_j,Y\rangle =\langle Y_j,N\rangle =0, ~\langle Y_j,Y_k\rangle =\delta_{jk}, ~~1\leq j,k\leq m.$$ We define $$\xi_r=(H^r,n_r+H^r x),~~~1\le r\le p.$$ Then $\{\xi_{1},\cdots,\xi_p\}$ form the orthonormal frame of the orthogonal complement of $\mathrm{Span}\{Y,N,Y_j|1\le j\le m\}$. And $\{Y,N,Y_j,\xi_{r}\}$ is a moving frame in $\mathbb{R}^{m+p+2}_1$ along $M^m$. \[rem-xi\] Geometrically, at one point $x$, $\xi_r$ corresponds to the unique sphere tangent to $M^m$ with normal vector $n_r$ and the same mean curvature $H^r=\langle \xi_r,g\rangle$ where $g=(1,\vec{0})$ is a constant time-like vector. We call $\{\xi_r\}_{r=1}^p$ *the mean curvature spheres* of $M^m$. We fix the range of indices in this section as below: $1\leq i,j,k\leq m; 1\leq r,s\leq p$. The structure equations are: $$\label{equation} \begin{split} &dY=\sum_i \omega_i Y_i,\\ &dN=\sum_{ij}A_{ij}\omega_i Y_j+\sum_{i,r} C^r_i\omega_i \xi_{r},\\ &d Y_i=-\sum_j A_{ij}\omega_j Y-\omega_i N+\sum_j\omega_{ij}Y_j +\sum_{j} B^{r}_{ij}\omega_j \xi_{r},\\ &d \xi_{r}=-\sum_i C^{r}_i\omega_i Y-\sum_{i,j}\omega_i B^{r}_{ij}Y_j +\sum_{s} \theta_{rs}\xi_{s}, \end{split}$$ where $\omega_{ij}$ are the connection $1$-forms of the Möbius metric $\mathrm{g}$, and $\theta_{rs}$ are the normal connection $1$-forms. The tensors $${\bf A}=\sum_{i,j}A_{ij}\omega_i\otimes\omega_j,~~ {\bf B}=\sum_{i,j,r}B^{r}_{ij}\omega_i\otimes\omega_j \xi_{r},~~ \Phi=\sum_{j,r}C^{r}_j\omega_j \xi_{r}$$ are called the Blaschke tensor, the Möbius second fundamental form and the Möbius form of $x$, respectively. The covariant derivatives $A_{ij,k}, B^{r}_{ij,k}, C^{r}_{i,j}$ are defined as usual. For example, $$\begin{aligned} &&\sum_j C^{r}_{i,j}\omega_j=d C^{r}_i+\sum_j C^{r}_j\omega_{ji} +\sum_{s} C^{s}_i\theta_{sr},\\ &&\sum_k B^{r}_{ij,k}\omega_k=d B^{r}_{ij}+\sum_k B^{r}_{ik}\omega_{kj} +\sum_k B^{r}_{kj}\omega_{ki}+\sum_{s} B^{s}_{ij}\theta_{sr}.\end{aligned}$$ The integrability conditions are given as below: $$\begin{aligned} &&A_{ij,k}-A_{ik,j}=\sum_{r}(B^{r}_{ik}C^{r}_j -B^{r}_{ij}C^{r}_k),\label{equa1}\\ &&C^{r}_{i,j}-C^{r}_{j,i}=\sum_k(B^{r}_{ik}A_{kj} -B^{r}_{jk}A_{ki}),\label{equa2}\\ &&B^{r}_{ij,k}-B^{r}_{ik,j}=\delta_{ij}C^{r}_k -\delta_{ik}C^{r}_j,\label{equa3}\\ &&R_{ijkl}=\sum_{r}(B^{r}_{ik}B^{r}_{jl}-B^{r}_{il}B^{r}_{jk} +\delta_{ik}A_{jl}+\delta_{jl}A_{ik} -\delta_{il}A_{jk}-\delta_{jk}A_{il}),\label{equa4}\\ &&R^{\perp}_{rs ij}=\sum_k (B^{r}_{ik}B^{s}_{kj}-B^{s}_{ik}B^{r}_{kj}). \label{equa5}\end{aligned}$$ Here $R_{ijkl}$ denote the curvature tensor of $\mathrm{g}$. Other restrictions on tensor $\bf B$ are $$\sum_j B^{r}_{jj}=0, ~~~\sum_{i,j,r}(B^{r}_{ij})^2=\frac{m-1}{m}. \label{equa7}$$ All coefficients in the structure equations are determined by $\{\mathrm{g}, {\bf B}\}$ and the normal connection $\{\theta_{rs}\}$. Coefficients of Möbius invariants and the isometric invariants are related as below. (We omit the formula for $A_{ij}$ since it will not be used later.) $$\begin{aligned} B^{r}_{ij}&=\rho^{-1}(h^{r}_{ij}-H^{r}\delta_{ij}),\label{2.22}\\ C^{r}_i&=-\rho^{-2}[H^{r}_{,i}+\sum_j(h^{r}_{ij} -H^{r}\delta_{ij})e_j(\ln\rho)]. \label{2.23}\end{aligned}$$ \[rem-xi2\] For $x: M^3 \rightarrow \mathbb{R}^5$, the Möbius position vector $Y: M^3\rightarrow \mathbb{R}^7_1$ and the mean curvature sphere $\{\xi_{1},\cdots,\xi_p\}$ are given by $$Y=\rho(\frac{1+|x|^2}{2}, \frac{1-|x|^2}{2}, x),$$ $$\xi_r=\left(\frac{1+|x|^2}{2}, \frac{1-|x|^2}{2}, x\right)H^r+(x\cdot n_r,-x\cdot n_r,n_r).$$ Note that $H^r=\langle \xi_r,g\rangle$ where $g=(-1,1,\vec{0})$ is a constant light-like vector. For $x: M^3 \rightarrow \mathbb{H}^5 \subset \mathbb{R}^6_1$ (the hyperboloid model of the hyperbolic space), the corresponding formulae are $$Y=\rho(x,1),~~~\xi_r=(n_r+H^rx,H^r),~~r=1,\cdots,p.$$ In this case $H^r=\langle \xi_r,g\rangle$ where $g=(\vec{0},1)$ is a constant space-like vector. The Möbius invariants are related to the isometric invariants still by $\sim$ . Three dimensional Wintgen ideal submanifolds in $\mathbb{S}^5$ ============================================================== From now on, we assume $x: M^{3}\to \mathbb{S}^{5}$ to be a three dimensional Wintgen ideal submanifold without umbilic points. According to and ,, that means we can choose a suitable tangent and normal frame ($\{E_1,E_2,E_3\}$ and $\{\xi_1,\xi_2\}$) such that the Möbius second fundamental form $\bf{B}$ takes the form $$\label{3.1} B^{1}= \begin{pmatrix} 0 & \mu & 0\\ \mu & 0 & 0\\ 0 & 0 & 0 \end{pmatrix},~~~~~~ B^{2}= \begin{pmatrix} \mu & 0 & 0\\ 0 & -\mu & 0\\ 0 & 0 & 0 \end{pmatrix}, ~~~~\mu=\frac{1}{\sqrt{6}}.$$ \[rem-transform\] The distribution $\mathbb{D}=\mathrm{Span}\{E_1,E_2\}$ is well-defined. The same is true for the vector field $E_3$ up to a sign, and this sign is fixed on a connected and orientable subset of $M^3$. Notice that the tangent and normal frames still allow a simultaneous transformation $$\label{transform} (\widetilde{E}_1,\widetilde{E}_2)=(E_1,E_2) \begin{pmatrix} ~~\cos t &\sin t \\ -\sin t & \cos t\end{pmatrix}, ~~(\widetilde{\xi}_1,\widetilde{\xi}_2)=(\xi_1,\xi_2) \begin{pmatrix} \cos 2t &-\sin 2t \\ \sin 2t & ~~\cos 2t\end{pmatrix}$$ if we fix the induced orientation on the tangent and normal bundles and require that $B^1,B^2$ still take the form . First we compute the covariant derivatives of $B^{r}_{ij}$. From (\[3.1\]) we get $$\label{bb1} B^1_{33,i}=B^2_{33,i}=B^1_{12,i}=B^2_{11,i}=B^2_{22,i}=0,~~1\leq i\leq 3.$$ Other derivatives are related with the connection 1-forms $\o_{ij}$ as below: $$\label{bb2} \begin{aligned} \omega_{23}&=\sum_i\frac{B^1_{13,i}}{\mu}\omega_i =-\sum_i\frac{B^2_{23,i}}{\mu}\omega_i;\\ \omega_{13}&=\sum_i\frac{B^1_{23,i}}{\mu}\omega_i =~~\sum_i\frac{B^2_{13,i}}{\mu}\omega_i;\\ 2\omega_{12}+\theta_{12}&=\sum_i\frac{-B^1_{11,i}}{\mu}\omega_i =\sum_i\frac{B^1_{22,i}}{\mu}\omega_i =\sum_i\frac{B^2_{12,i}}{\mu}\omega_i. \end{aligned}$$ By and we know the symmetry property below, $$\label{bb3} B^1_{13,2}=B^1_{23,1}=B^1_{12,3}=0, ~B^2_{13,2}=B^2_{23,1}=B^2_{12,3},$$ where we have used $B^1_{12,3}=0$ by . From this fact and comparing coefficients in , we obtain $$\label{bb4} \mu\o_{13}(E_1)=B^2_{13,1}=B^1_{23,1}=0,~ \mu\o_{23}(E_2)=-B^2_{23,2}=B^1_{13,2}=0.$$ Similarly we know the coefficients in the following three equalities are equal to each other: $$\label{bb5} \begin{aligned} -\mu\o_{23}(E_1)&=-B^1_{13,1}=B^2_{23,1},\\ \mu\o_{13}(E_2)&=~~B^1_{23,2}=B^2_{13,2},\\ \mu(2\o_{12}+\theta_{12})(E_3)&=-B^1_{11,3}=B^1_{22,3}=B^2_{12,3}. \end{aligned}$$ Next we derive the Möbius form, using and the information on $B^r_{ij,k}$: $$\label{cc1} C^1_3=B^1_{22,3}-B^1_{23,2}=0,~C^2_3=B^2_{11,3}-B^2_{13,1}=0.$$ The other coefficients $\{C^{r}_j\}$ are obtained similarly as below: $$\label{cc2} \begin{aligned} &C^{1}_1=-B^{1}_{13,3}=-\mu\omega_{23}(E_3), ~~~~C^{2}_2=-B^{2}_{23,3}=~~\mu\omega_{23}(E_3), \\ &C^{1}_2=-B^{1}_{23,3}=-\mu\omega_{13}(E_3), ~~~~C^{2}_1=-B^{2}_{13,3}=-\mu\omega_{13}(E_3),\\ &C^{1}_1=B^{1}_{22,1}=~\mu(2\o_{12}+\theta_{12})(E_1) =~B^{2}_{12,1}=-C^{2}_2, \\ &C^{1}_2=B^{1}_{11,2}=-\mu(2\o_{12}+\theta_{12})(E_2) =-B^{2}_{12,2}=~C^{2}_1. \end{aligned}$$ For simplicity we introduce the following notations: $$\label{UVL} \begin{aligned} &U=~~\o_{23}(E_3)=-\frac{C^1_1}{\mu}=\frac{C^2_2}{\mu},\\ &V=-\o_{13}(E_3)=~~\frac{C^1_2}{\mu}=\frac{C^2_1}{\mu},\\ &L=\o_{13}(E_2)=-\o_{23}(E_1)=-\frac{B^1_{11,3}}{\mu}. \end{aligned}$$ Then we summarize what we know about the connection 1-forms and the covariant derivatives $B^r_{ij,k}$ as below: $$\begin{gathered} \o_{13}=L\o_2-V\o_3,~~~\o_{23}=-L\o_1+U\o_3;\notag\\ 2\o_{12}+\theta_{12}=-U \o_1-V\o_2+L\o_3.\label{omega}\end{gathered}$$ By we have $$\label{L} d\o_3=\o_{31}\wedge \o_1+\o_{32}\wedge\o_2\equiv 2L\o_1\wedge \o_2 ~~~mod(\o_3).$$ So the distribution $\mathbb{D}=\mathrm{Span}\{E_1, E_2\}$ is integrable if and only if $L=0$ identically. For the information on the Blaschke tensor $\bf A$, we use . It requires to compute the covariant derivatives of $C^r_j$, which is quite straightforward: $$C^1_{1,i}=-C^2_{2,i}, ~C^1_{2,i}=C^2_{1,i}, ~C^1_{3,2}=C^1_1 L=-\mu UL, ~C^1_{3,1}=-C^1_2 L=-\mu VL.$$ Now it follows that $$\begin{aligned} \mu (A_{11}-A_{22})&=\sum_k(B^1_{2k}A_{k1}-B^1_{1k}A_{k2}) =C^1_{2,1}-C^1_{1,2},\label{A11}\\ 2\mu A_{12}&=\sum_k(B^2_{1k}A_{k2}-B^2_{2k}A_{k1}) =C^2_{1,2}-C^2_{2,1}=C^1_{1,1}+C^1_{2,2},\label{A12}\\ \mu A_{13}&=\sum_k(B^1_{2k}A_{k3}-B^1_{3k}A_{k2}) =C^1_{2,3}-C^1_{3,2}=C^1_{2,3}+\mu UL,\label{A13}\\ \mu A_{23}&=\sum_k(B^1_{1k}A_{k3}-B^1_{3k}A_{k1}) =C^1_{1,3}-C^1_{3,1}=C^1_{1,3}+\mu VL.\label{A23}\end{aligned}$$ Consider a new frame $\{Y,{\hat Y},\eta_1,\eta_2,\eta_3, \xi_1,\xi_2\}$ in ${\mathbb R}^{7}_1$ along $M^3$ as below, whose geometric meaning will be clear later (see Theorem \[thm-envelop\] and Remark \[rem-hatY\]). $$\begin{gathered} \eta_1=Y_1+\frac{C_2^1}{\mu}Y=Y_1+VY,~~ \eta_2=Y_2+\frac{C^1_1}{\mu}Y=Y_2-UY,~~ \eta_3=Y_3-\lambda Y;\label{eta}\\ {\hat Y}=N-\frac{1}{2}(U^2+V^2+\lambda^2)Y-VY_1+U Y_2+\lambda Y_3.\label{hatY}\end{gathered}$$ Here $\lambda\in C^{\infty}(M^3)$ is an arbitrarily given smooth function at the beginning. The new frame is orthonormal except that $$\langle Y,Y\rangle=0=\langle {\hat Y},{\hat Y}\rangle,~ \langle Y,{\hat Y}\rangle=1.$$ By the original structure equations we get $$\begin{aligned} d\xi_{1}&=-\mu\o_2\eta_1-\mu\o_1\eta_2+\theta_{12}\xi_2,\label{3.7}\\ d\xi_{2}&=-\mu \o_1\eta_1+\mu\o_2\eta_2-\theta_{12}\xi_1,\label{3.8}\\ d\eta_1&=-{\hat\o}_1Y-\o_1{\hat Y}+\sum_k\Omega_{1k}\eta_k+\mu\o_2\xi_1+\mu\o_1\xi_2,\label{3.3}\\ d\eta_2&=-{\hat\o}_2Y-\o_2{\hat Y}+\sum_k\Omega_{2k}\eta_k+\mu\o_1\xi_1-\mu\o_2\xi_2,\label{3.4}\\ d\eta_3&=-{\hat\o}_3Y-\o_3{\hat Y}+\sum_k\Omega_{3k}\eta_k,\label{3.5}\\ dY&=\o Y+\o_1\eta_1+\o_2\eta_2+\o_3\eta_3,\label{3.9}\\ d{\hat Y}&=-\o {\hat Y}+{\hat\o}_1\eta_1+{\hat\o}_2\eta_2+{\hat\o}_3\eta_3, \label{3.10}\end{aligned}$$ Note that and give the first motivation for the definition of $\eta_1,\eta_2$ in . Differentiate $\sim$ . We get the following integrability equations: $$\begin{aligned} &d\o_1=\o\wedge\o_1+\Omega_{12}\wedge\o_2+\Omega_{13}\wedge\o_3; \label{3.12}\\ &d\o_2=\o\wedge\o_2-\Omega_{12}\wedge\o_1+\Omega_{23}\wedge\o_3; \label{3.13}\\ &d\o_3=\o\wedge\o_3+\Omega_{31}\wedge\o_1+\Omega_{32}\wedge\o_2; \label{3.21}\\ &d\o_1=-(\theta_{12}+\Omega_{12})\wedge\o_2; \label{3.14}\\ &d\o_2=~~(\theta_{12}+\Omega_{12})\wedge\o_1; \label{3.15}\\ &d{\hat\o}_1=-\o\wedge{\hat\o}_1+\Omega_{12}\wedge{\hat\o}_2 +\Omega_{13}\wedge{\hat\o}_3; \label{3.16}\\ &d{\hat\o}_2=-\o\wedge{\hat\o}_2-\Omega_{12}\wedge{\hat\o}_1 +\Omega_{23}\wedge{\hat\o}_3; \label{3.17}\\ &d{\hat\o}_3=-\o\wedge{\hat \o}_3+\Omega_{31}\wedge{\hat\o}_1 +\Omega_{32}\wedge{\hat\o}_2; \label{3.22}\\ &d\Omega_{12}=\Omega_{13}\wedge\Omega_{32}-\o_1\wedge{\hat\o}_2 -{\hat\o_1}\wedge\o_2+2\mu^2\o_1\wedge\o_2; \label{3.18}\\ &d\Omega_{13}=\Omega_{12}\wedge\Omega_{23}-\o_1\wedge{\hat\o}_3 -{\hat\o}_1\wedge\o_3; \label{3.19}\\ &d\Omega_{23}=-\Omega_{12}\wedge\Omega_{13}-\o_2\wedge{\hat\o}_3 -{\hat\o}_2\wedge\o_3; \label{3.20}\\ &\Omega_{13}\wedge\o_1=\Omega_{23}\wedge\o_2;\hskip 5pt \Omega_{13}\wedge\o_2=-\Omega_{23}\wedge\o_1;\label{3.23}\\ &d\theta_{12}=2\mu^2\o_1\wedge\o_2;\label{3.24}\\ &\o_1\wedge{\hat\o_2}=-\o_2\wedge{\hat\o}_1;\hskip 5pt \o_1\wedge{\hat\o_1}=\o_2\wedge{\hat\o}_2;\label{3.25}\\ &d\o=-\o_1\wedge{\hat\o}_1-\o_2\wedge{\hat\o}_2 -\o_3\wedge{\hat\o}_3.\label{3.26}\end{aligned}$$ The 1-forms $\omega,{\hat\o}_i,\Omega_{ij}=-\Omega_{ji}$ are determined by and ,: $$\begin{aligned} \Omega_{12}&=\langle d\eta_1,\eta_2\rangle =\o_{12}+U\o_1+V\o_2,\label{Omega12}\\ \Omega_{13}&=\langle d\eta_1,\eta_3\rangle =\lambda\o_1+L\o_2,\label{Omega13}\\ \Omega_{23}&=\langle d\eta_2,\eta_3\rangle =-L\o_1+\lambda\o_2,\label{Omega23}\\ \o&=\langle dY,{\hat Y} \rangle=-V\o_1+U\o_2+\lambda\o_3.\end{aligned}$$ It follows from that there exist some functions $\hat{F},\hat{G}$ such that $${\hat\o_1}=\hat{F}\o_1+\hat{G}\o_2,~~ {\hat\o_2}=-\hat{G}\o_1+\hat{F}\o_2.\label{FG}$$ A straightforward yet lengthy computation find $$\begin{aligned} {\hat\o}_1&=\langle d{\hat Y},\eta_1\rangle \\ &= d\left(N-\frac{1}{2}(U^2+V^2+\lambda^2)Y-VY_1+U Y_2+\lambda Y_3\right)\cdot (Y_1+VY)\\ &=\sum_i A_{i1}\o_i-\frac{1}{2}(U^2+V^2+\lambda^2)\o_1\\ &~~~~-dV +V^2\o_1+U\o_{21}-UV\o_2+\lambda\o_{31}-\lambda V\o_3\\ &=\left(A_{11}+\frac{1}{2}(U^2+V^2-\lambda^2) -\frac{1}{\mu}C^1_{2,1}\right)\o_1 +\left(A_{12}-\frac{1}{\mu}C^1_{2,2}-\lambda L\right)\o_2,\end{aligned}$$ where we have used $d C^1_2=\sum_i C^1_{2,i}\o_i-C^1_1\o_{12}-C^2_2\theta_{21}$ and , . For ${\hat\o}_2$ we compute in a similar manner, using , and to verify , with the result as below: $$\begin{aligned} \hat{F}&=A_{11}+\frac{1}{2}(U^2+V^2-\lambda^2) -\frac{1}{\mu}C^1_{2,1},\label{hatF}\\ \hat{G}&=A_{12}-\frac{1}{\mu}C^1_{2,2}-\lambda L =\frac{C^1_{1,1}-C^1_{2,2}}{2\mu}-\lambda L.\label{hatG}\end{aligned}$$ In particular we have the following observation. \[lem-G\] $\hat{G}$=0 if and only if $\lambda=\frac{G}{L}$ with $L=-\frac{B^1_{11,3}}{\mu}, G=\frac{C^1_{1,1}-C^1_{2,2}}{2\mu}$. In the end of this section we make the following important geometric observation.\ The spacelike 2-plane $\mathrm{Span}_{\mathbb{R}}\{\xi_1,\xi_2\}$ at $p\in M^3$ is well-defined, and we call it *the 3-dimensional mean curvature sphere*, because it defines a 3-sphere tangent to $M^3$ at $p$ with the same mean curvature vector. (Please compare to Remark \[rem-xi\].) The first key observation is that in the codimension two case, $\mathrm{Span}_{\mathbb{R}}\{\xi_1,\xi_2\}$ is determined by the complex line $\mathrm{Span}_{\mathbb{C}}\{\xi_{1}-i\xi_{2}\}$ and vice versa. So $\xi_1-i\xi_2\in \mathbb{C}^7$ represents the same geometric object. It is a null vector with respect to the $\mathbb{C}$-linear extension of the Lorentz metric to $\mathbb{R}^7_1\otimes\mathbb{C}$. The complex line spanned by it corresponds to a point $$[\xi_{1}-i\xi_{2}] \in Q^5=\{[Z]\in \mathbb{C}P^6| \langle Z,Z\rangle =0\}.$$ It is similar to the conformal Gauss map of a (Willmore) surface [@br0] and to the Gauss map of a hypersurface in $\mathbb{S}^n$ [@MaOhnita]. The second key observation is that under the hypothesis of being Wintgen ideal, this 3-sphere congruence is indeed a 2-parameter family, and its envelope not only recovers $M^3$, but also extends it to a 3-manifold as a circle bundle over a Riemann surface $\overline{M}$ (a holomorphic curve). The underlying surface $\overline{M}$ comes from the quotient surface $\overline{M}=M^3/\Gamma$ (at least locally) where $\Gamma$ is the foliation of $M^3$ by the integral curves of the vector fields $E_3$. More precise statement is as below. \[thm-envelop\] For a Wintgen ideal submanifold $x:M^3\to \mathbb{S}^5$ we have: \(1) The complex vector-valued function $\xi_{1}-i\xi_{2}$ locally defines a complex curve $$[\xi_1-i\xi_2]:\overline{M}=M^3/\Gamma\to Q^5\subset\mathbb{C}P^6.$$ (2) *The 3-dimensional mean curvature spheres* $\mathrm{Span}\{\xi_1(p),\xi_2(p)\}$ is a two-parameter family of 3-spheres in $\mathbb{S}^5$ when $p$ runs through $M^3$. \(3) The Lorentz 3-space $\mathrm{Span}\{Y,{\hat Y},\eta_3\}$ and the two light-like directions $[Y],[{\hat Y}]\in \mathbb{R}P^6$ correspond to a circle and two points on it. These circles are a two-parameter family. They foliate a three dimensional submanifold $\widehat{M}^3$ enveloped by the 3-dimensional mean curvature spheres $\mathrm{Span}\{\xi_{1},\xi_{2}\}$, which includes $M^3$ as part of it. On $M^3$ these circular arcs are indeed the integral curves of the vector field $E_3$. \(4) This envelope $\widehat{M}^3\subset\mathbb{S}^5$, as a natural extension of $x:M^3\to\mathbb{S}^5$, is still a Wintgen ideal submanifold (at its regular points). The structure equations and imply $$\label{J} d(\xi_{1}-i\xi_{2})=i\mu(\o_1+i\o_2)(\eta_1+i\eta_2) +i\theta_{12}(\xi_1-i\xi_2).$$ Geometrically that means $\xi=[\xi_{1}-i\xi_{2}]:M^3\to \mathbb{C}P^6$ decomposes as a quotient map $\pi:M^3\to \overline{M}=M^3/\Gamma$ composed with a holomorphic immersion $\bar\xi:\overline{M}\to \mathbb{C}P^6$. Thus conclusion (1) is proved, and (2) follows directly. To prove (3), notice that the light-like directions in $\mathrm{Span}\{Y,{\hat Y},\eta_3\}$ represent points on a circle. Since $\{\xi_1,\xi_2,d\xi_1,d\xi_2\}$ span a 4-dimensional spacelike subspace by , the corresponding 2-parameter family of 3-dimensional mean curvature sphere congruence has an envelop $\widehat{M}$, whose points correspond to the light-like directions in the orthogonal complement $\mathrm{Span}\{Y,{\hat Y},\eta_3\}$. In particular $[Y],[{\hat Y}]$ are two points on this circle. Such circles form a 2-parameter family, with $\overline{M}$ as the parameter space. They give a foliation of $\widehat{M}^3$ which is also a circle fibration. We assert that every integral curve $\gamma$ of $E_3$ is contained in such a circle. Because $Span\{\xi_1,\xi_2,\eta_1,\eta_2\}(p)$ is a fixed subspace along an integral curve of $E_3$ passing $p\in M^3$ by $\sim$ and ,,. Thus the integration of $Y$ along $E_3$ direction is always located in the orthogonal complement $\mathrm{Span}\{Y(p),{\hat Y}(p),\eta_3(p)\}$, which describes a circle as above. Thus $\widehat{M}^3\supset M^3$, and each circle fiber cover an integral curve of $E_3$. This verifies (3). To prove (4), we need only to show that for arbitrarily chosen smooth function $\lambda:M^3\to \mathbb{R}$, the corresponding $[\hat{Y}]:M^3\to \mathbb{S}^5$ is a Wintgen ideal submanifold. This is because at one point $p\in M^3$, when $\lambda$ runs over $(-\infty,\infty)$, $[\hat{Y}(p)]$ given by will cover the circle fiber except $[Y(p)]$ itself; and when $\lambda: M^3\to \mathbb{R}$ is arbitrary, all such local mappings will cover $\widehat{M}$ by their images. We have to compute out the Laplacian of $\hat{Y}$ with respect to its induced metric $\hat\o_1^2+\hat\o_2^2+\hat\o_3^2$ which is necessary to determine the normal frames (the mean curvature spheres) $\{\hat\xi_r\}$ of $\hat{Y}$. The main difficulty is that the map $\hat{Y}:M^3\to \mathbb{R}^7_1$ is generally not conformal to $Y$ . Fortunately we need only to find an orthogonal frame $\{\hat{E}_j\}_{j=1}^3$ with the same length for $\hat{Y}$, and then using the fact $ \mathrm{Span}_{\mathbb{R}}\{\hat{Y},\hat{Y}_j,\sum_{j=1}^3 \hat{E}_j\hat{E}_j(\hat{Y})\} =\mathrm{Span}_{\mathbb{R}}\{\hat{Y},\hat{Y}_j, \hat{\Delta}\hat{Y}\}.$\ **Claim:** $\hat{Y}$ shares the same mean curvature spheres $\{\xi_1,\xi_2\}$ as $Y$.\ This requires to show $\langle\sum_{j=1}^3 \hat{E}_j\hat{E}_j(\hat{Y}),\xi_r\rangle=0.$ Since $0=\langle\hat{Y},\xi_r\rangle=\langle d\hat{Y},\xi_r\rangle=\langle \hat{Y},d\xi_r\rangle$ by , and , we need only to verify $$\langle \hat{Y},\sum_{j=1}^3 \hat{E}_j\hat{E}_j(\xi_r)\rangle=0,~~r=1,2.$$ For this purpose, suppose (keeping in mind): $${\hat\o_1}=\hat{F}\o_1+\hat{G}\o_2,~~ {\hat\o_2}=-\hat{G}\o_1+\hat{F}\o_2,~~ \hat\o_3=a\o_1+b\o_2+c\o_3.$$ Notice that we can always assume $\hat{Y}$ to be an immersion at the points where $\widehat{M}$ is regular, hence $c\ne 0$. Then one can take $$\hat{E}_1 = \hat{F} E_1+\hat{G} E_2 + a_{13}E_3,~ \hat{E}_2 = -\hat{G} E_1+\hat{F} E_2 +a_{23}E_3,~ \hat{E}_3 = a_{33}E_3,$$ where $a_{13},a_{23},a_{33}$ are uniquely determined by $\hat\o_i(\hat{E}_j)=(\hat{F}^2+\hat{G}^2)\delta_{ij}$. The explicit form of $a_{13},a_{23},a_{33}$ is not important, because when we insert the formulae above to $\sum_{j=1}^3 \hat{E}_j\hat{E}_j(\xi_r)$, the terms involving $E_3$ will always be orthogonal to $\hat{Y}$. For example, $\langle E_3(\eta_1),\hat{Y}\rangle=-\langle \eta_1,E_3(\hat{Y})\rangle=\hat\o_1(E_3)=0$ by . Thus we need only to compute the effect on $\xi_r$ of the operator below: $$(\hat{F} E_1+\hat{G} E_2)^2+(-\hat{G} E_1+\hat{F} E_2)^2 \thickapprox(\hat{F}^2+\hat{G}^2)\partial\bar{\partial}.$$ The two sides are equal up to first order differential operators like $[E_1,E_2],E_1,E_2$, whose action on $\xi_r$ must be orthogonal to $\hat{Y}$; the complex differential operators are defined as usual: $$\partial=E_1-iE_2,~\bar\partial=E_1+iE_2.$$ Since $(\o_1+i\o_2)(\bar\partial)=0$, it follows from that $$\begin{aligned} \bar{\partial}(\xi_1-i\xi_2)&=i\theta_{12}(\bar{\partial}) (\xi_1-i\xi_2),\\ \partial\bar{\partial}(\xi_1-i\xi_2)&\in \mathrm{Span}_{\mathbb{C}}\{\xi_1-i\xi_2,\eta_1+i\eta_2\}~~\bot~~ \hat{Y}.\end{aligned}$$ This completes the proof of the previous claim. For $\hat{Y}$ we still take its canonical lift, whose derivatives are combinations of $\hat{Y},\eta_1,\eta_2,\eta_3$; its normal frame is still $\{\xi_1,\xi_2\}$. We read from that its Möbius second fundamental form still take the same form as . Thus conclusion (4) and the whole theorem is proved. In the proof above, among the integrability equations from $\sim$, only and are necessary for us (to deduce the algebraic form and ,, where the explicit coefficients are not important). This is somewhat striking to the authors that the strong conclusion (4) follows from so few conditions. In a forthcoming paper [@Li3] we will give a general treatment of codimension-two Wintgen ideal submanifolds based on the observations in this theorem. Another interesting feature is the resemblance between conclusion (4) and the duality theorem for Willmore surfaces [@br0]. Minimal Wintgen ideal submanifolds ================================== Since we have classified all Wintgen ideal submanifolds in [@Li1] whose canonical distribution $\mathbb{D}=\mathrm{Span}\{E_1,E_2\}$ is integrable (that means $L=0$), in the rest of this paper we will only consider the case $$L\neq0.$$ From now on we take the frame and , and make the following $$\label{lambda} \textbf{Assumption:}~~~~ \lambda=\frac{G}{L}=\frac{C^1_{1,1}-C^1_{2,2}}{-2 B^1_{11,3}}~.\qquad\qquad\qquad\qquad\qquad$$ By Lemma \[lem-G\] we have $$\label{3.27} {\hat\o_1}=\hat{F}\o_1,~~ {\hat\o_2}=\hat{F}\o_2.$$ \[rem-hatY\] The correspondence $Y\to \hat{Y}$ describes a self-mapping of the enveloping submanifold $\widehat{M}^3$ where $\hat{Y}$ and $Y$ are located on the same circle fiber. At corresponding points they share the same normal vector fields $\{\xi_r\}$, with respect to which we can talk about principal directions. According to Theorem \[thm-envelop\] and Lemma \[lem-G\], the correspondence $Y\to \hat{Y}$ preserves the principal directions for any $\xi_r$ if and only if $\lambda=G/L$. This explains the geometric significance of the condition . Under the assumption , $$\label{o} \o=-V\o_1+U\o_2+\frac{G}{L}\o_3.$$ Together with and there must be $$d\o=-\o_3\wedge{\hat\o}_3.\label{3.28}$$ On the other hand, it follows from and $\lambda=G/L$ that , and now take the form $$\begin{gathered} 2\Omega_{12}+\theta_{12}=U\o_1+V\o_2+L\o_3,\label{Omega12+}\\ \Omega_{13}=\frac{G}{L}\o_1+L\o_2,~~ \Omega_{23}=-L\o_1+\frac{G}{L}\o_2.\label{Omega13+}\end{gathered}$$ Insert these into structure equations and , and simplify by , together with . We have $$\begin{aligned} dL\wedge\o_2&= (U\o_1+V\o_2+L\o_3)\wedge(-L\o_1+\frac{G}{L}\o_2) -\o_1\wedge({\hat \o}_3-d\frac{G}{L})-\hat{F}\o_1\wedge\o_3,\\ dL\wedge\o_1&= (U\o_1+V\o_2+L\o_3)\wedge(\frac{G}{L}\o_1+L\o_2) +\o_2\wedge({\hat \o}_3-d\frac{G}{L})+\hat{F}\o_2\wedge\o_3.\end{aligned}$$ Comparing the coefficients of $\o_1\wedge\o_2,\o_1\wedge\o_3,\o_2\wedge\o_3$ in these two equations separately, we obtain $$\begin{aligned} \hat\o_3(E_1)&=E_2(L)+UL+E_1(\frac{G}{L})-V\frac{G}{L},\label{compare1}\\ \hat\o_3(E_2)&=-E_1(L)+VL+E_2(\frac{G}{L})+U\frac{G}{L},\label{compare2}\\ \hat\o_3(E_3)&=E_3(\frac{G}{L})+L^2-\hat{F},\label{compare3}\\ E_3(L)&=G.\label{compare4}\end{aligned}$$ Similarly, inserting into and yields $$\begin{aligned} d\hat{F}\wedge\o_1&= (2\Omega_{12}+\theta_{12})\wedge \hat{F}\o_2 -\o\wedge \hat{F}\o_1+\Omega_{13}\wedge\hat\o_3,\\ d\hat{F}\wedge\o_2&= -(2\Omega_{12}+\theta_{12})\wedge \hat{F}\o_1 -\o\wedge \hat{F}\o_2+\Omega_{23}\wedge\hat\o_3.\end{aligned}$$ Invoking ,, and comparing the coefficients of $\o_1\wedge\o_2,\o_1\wedge\o_3,\o_2\wedge\o_3$ in these two equations separately, one gets $$\begin{aligned} \hat{F}&=\hat\o_3(E_3),\label{compare5}\\ E_1(\hat{F}) &=2V\hat{F}-\frac{G}{L}\hat\o_3(E_1)-L\hat\o_3(E_2),\label{compare6}\\ E_2(\hat{F}) &=-2U\hat{F}+L\hat\o_3(E_1)-\frac{G}{L}\hat\o_3(E_2),\label{compare7}\\ E_3(\hat{F}) &=-\frac{G}{L}(\hat{F}+\hat\o_3(E_3))=-2\frac{G}{L}\cdot\hat{F}.\label{compare8}\end{aligned}$$ \[rem-invariant\] We point out that $L,G,\o,\hat{F}$ are well-defined Möbius invariants. It is necessary and sufficient to verify that they are independent to the choice of the frames, or equivalently, that they are invariant under the transformation . This is obvious for $L$ by , for $G$ by , and for $\o$ by $\o=\langle dY,{\hat Y}\rangle$ where the frame vectors $\hat{Y},\eta_3$ are now canonically chosen after taking $\lambda=\frac{G}{L}$. For $\hat{F}$ we can verify the invariance under the transformation by $\hat{F}=\langle E_1({\hat Y}),\eta_1\rangle$, or just using . On the other hand, $U,V$ correspond to $\{C^r_i\}$, components of the Möbius form, which depend on the choice of $\{E_1,E_2\}$ and $\{\xi_1,\xi_2\}$. Yet in that case we can choose the angle $t$ in suitably such that $V=0$ identically. Then the new function $U$ is well-defined and Möbius invariant. (There are other choice of the frame in a canonical way, and any of them works in the proof to Theorem \[thm-homog\] later.) Now we can state our Möbius characterization theorem for minimal Wintgen ideal submanifolds of dimension three in five dimensional space form. \[thm-minimal\] Let $x: M^3\to {\mathbb S}^{5}$ be a Wintgen ideal submanifold. Assume the distribution $\mathbb{D}=\mathrm{Span}\{E_1, E_2\}$ to be non-integrable, i.e., $L\ne 0$. Then the following conditions are equivalent: \(1) $ d\o=0,$ where $\o=-V\o_1+U\o_2+\frac{G}{L}\o_3$. \(2) The correspondence $Y\to \hat{Y}$ of the enveloping submanifold $\widehat{M}^3$ is a conformal map. ($\hat{Y}$ might be degenerate.) \(3) $x: M^3\to {\mathbb S}^{5}$ is Möbius equivalent to a minimal Wintgen ideal submanifold in a space form. In particular, the space form is $\mathbb{S}^5$,$\mathbb{R}^5$ or $\mathbb{H}^5$ depending on whether $\hat{F}$ is positive, zero or negative. If $Y$ and $\hat Y$ are conformal, then there exists a non-negative function $a$ so that $$\hat\o_1^2+\hat\o_2^2+\hat\o_3^2=a(\o_1^2+\o_2^2+\o_3^2).\label{3.51}$$ Combining with and we have $d\o=0$. Thus (2) implies (1). Next we show (1) implies (2) and (3). If $d\o=0$, it follows from that $\hat \o_3=\hat{F}\o_3$. Together with , $\hat \o_j=\hat{F}\o_j$ for $j=1,2,3$. So $Y$ and $\hat Y$ are conformal, and (2) is proved. Using $\sim$ we get $$d\hat{F}+2\hat{F}\omega = 0.\label{3.50}$$ Now the structure equations can be rewritten as below: $$\begin{aligned} d(\hat{F}Y+{\hat Y})&=-\o (\hat{F}Y+{\hat Y})+2\hat{F}(\o_1\eta_1+\o_2\eta_2+\o_3\eta_3);\label{3.57}\\ d\eta_1&=-{\o}_1(\hat{F}Y+{\hat Y})+\Omega_{12}\eta_2+L\o_2\eta_3+\mu\o_2\xi_1+\mu\o_1\xi_2;\\ d\eta_2&=-{\o}_2(\hat{F}Y+{\hat Y})-\Omega_{12}\eta_1-L\o_1\eta_3+\mu\o_1\xi_1-\mu\o_2\xi_2;\\ d\eta_3&=-{\o}_3(\hat{F}Y+{\hat Y})-L\o_2\eta_1+L\o_1\eta_2;\\ d\xi_{1}&=-\mu\o_2\eta_1-\mu\o_1\eta_2+\theta_{12}\xi_2;\\ d\xi_{2}&=-\mu \o_1\eta_1+\mu\o_2\eta_2-\theta_{12}\xi_1;\label{3.62}\\ d({\hat{F}Y-\hat Y})&=-\o (\hat{F}Y-{\hat Y}). \label{3.63}\end{aligned}$$ So $\mathrm{Span}\{\hat{F}Y-\hat Y\}$ is parallel along $M^3$, as well as its orthogonal complement $$\mathbb{V}^6=\mathrm{Span}\{\hat{F}Y+\hat Y, \eta_1, \eta_2, \eta_3, \xi_1, \xi_2\}~.$$ That means both of them are fixed subspaces of $\mathbb{R}^7_1$. The type of the inner product restricted on these subspaces depends on the sign of $\hat{F}$, which will not change on a connected open set, because $\hat{F}$ satisfies a linear PDE . We discuss them case by case. [**Case 1:**]{} $\hat{F}>0$. This case $\mathbb{V}^6$ is a fixed space-like subspace orthogonal to a fixed time-like line $\mathbb{V}^\perp$. Define $$f=\frac{\hat{F}Y+\hat Y}{\sqrt{2\hat{F}}},\qquad g=\frac{\hat{F}Y-\hat Y}{\sqrt{2\hat{F}}},$$ which satisfy $\langle f,f\rangle=1,\langle g, g\rangle=-1$. So $g \in \mathbb{V}^\perp$ is a constant time-like vector, and $f: M^3 \rightarrow \mathbb{S}^5 \subset \mathbb{V}$ is a submanifold in the sphere. Assume $g=(1,\vec{0})$. This is without loss of generality since we can always apply a Lorentz transformation in $\mathbb{R}^{7}_{1}$ to $Y$ and its frame at the beginning if necessary, whose effect on $x(M)\subset\mathbb{S}^5$ is a Möbius transformation. Then from the geometric meaning of the mean curvature sphere $\xi_r$ explained in Remark \[rem-xi\], we know that $x: M^3\rightarrow \mathbb \mathbb{S}^{5}$ is a minimal Wintgen ideal submanifold (up to a conformal transformation). Note that this special minimal submanifold can now be identified with $f$ since we have $$f+g=\sqrt{2\hat{F}}Y=\sqrt{2\hat{F}}\rho (1,x).$$ Comparison shows $\sqrt{2\hat{F}}\rho=1$ and $x=f$. That $f:M^3 \rightarrow \mathbb{S}^5 \subset \mathbb{V}$ is minimal can be verified directly by $\sim$. [**Case 2:**]{} $\hat{F}<0$. In this case, $\mathbb{V}^6$ is a fixed Lorentz subspace orthogonal to a constant space-like line $\mathbb{V}^\perp$. Define $f=(\hat{F}Y+\hat Y)/\sqrt{-2\hat{F}}, g=(\hat{F}Y-\hat Y)/\sqrt{-2\hat{F}}.$ Then similar to Case 1 we know $x$ is Möbius equivalent to $f: M^3 \rightarrow \mathbb{H}^5 \subset \mathbb{V}\cong \mathbb{R}^6_1$ which is a minimal Wintgen ideal submanifold. [**Case 3:**]{} $\hat{F}\equiv 0$. Now $\hat \o_i \equiv0, ~i=1,2,3$. So $d\hat Y= -\o\hat{Y}$. That means $\hat Y$ determines a constant light-like direction. From $d\o=0$, we get $w=d\tau$ for some locally defined function $\tau$. Up to a Lorentz transformation one may take $e^{\tau}\hat Y=(-1, 1, \vec{0})$ which is still denoted by $g$. Since $\langle \xi_r,g\rangle=0$, from the geometric meaning of the mean curvature sphere $\xi_r$ explained in Remark \[rem-xi2\] we know that $x$ is a three dimensional minimal Wintgen ideal submanifold in $\mathbb{R}^5$ (up to a suitable conformal transformation).\ Finally we show (3) implies (1), i.e., for any minimal Wintgen ideal submanifold $x: M^3 \rightarrow \mathbb{Q}^5(c)$ whose distribution $\mathbb{D}=\mathrm{Span}\{E_1, E_2\}$ is not integrable, there is always $d\o=0$. By assumption, for this $x$ we can always take local orthonormal frame $\{e_1, e_2, e_3\}$ and $\{n_1, n_2\}$ for the tangent and normal bundles, such that the second fundamental form is given by $$h^1=\begin{pmatrix} 0& \nu&0\\ \nu&0&0\\ 0&0&0 \end{pmatrix}, \qquad h^2=\begin{pmatrix} \nu&0&0\\ 0&-\nu&0\\ 0&0&0 \end{pmatrix}.\label{h}$$ It follows that $\rho^2=6\nu^2$ by . From Remark 2.2, using we always have $$C^1_1=-C^2_2=-\frac{e_2(\nu)}{6\nu^2}, ~~~~~~~C_1^2=C^2_1=-\frac{e_1(\nu)}{6\nu^2}.$$ Consider the Möbius position vector $Y: M^3 \rightarrow \mathbb{R}^7_1$ of $x$ defined by . For the Möbius metric $$\mathrm{g}=\langle dY, dY\rangle=\rho^2 dx^2=6\nu^2 dx^2,$$ we can choose $\{E_1=\frac{e_1}{\sqrt{6}\nu}, \;E_2=\frac{e_2}{\sqrt{6}\nu}, \;E_3=\frac{e_3}{\sqrt{6}\nu}\}$ as a set of local orthonormal basis for $(M^3, \mathrm{g})$ with the dual basis $\{\o_1, \;\o_2, \;\o_3\}$. By , $$2\o_{12}+\theta_{12}=-U\o_1-V\o_2+L\o_3,$$ We have $$C^1_{1,1}=E_1(C^1_1)+C^1_2(\o_{21}+\theta_{21})(E_1) =\frac{1}{\sqrt{6}\nu}(-E_1(E_2(\nu))+E_1(\nu)\o_{21}(E_1)),$$ $$C^1_{2,2}=E_2(C^1_2)+C^1_1(\o_{12}+\theta_{12})(E_2) =\frac{1}{\sqrt{6}\nu}(-E_2(E_1(\nu))+E_2(\nu)\o_{12}(E_2)).$$ Using we get $$C^1_{1,1}-C^1_{2,2}=2\frac{E_3(\nu)}{\nu} \mu L.$$ So we get that $\frac{G}{L}=\frac{E_3(\nu)}{\nu}$. Since $$\o=-\sqrt{6}(C^1_2\o_1+C^1_1\o_2)+\frac{G}{L}\o_3 =\frac{E_1(\nu)}{\nu}\o_1+\frac{E_2(\nu)}{\nu}\o_2 +\frac{E_3(\nu)}{\nu}\o_3,\label{3.77}$$ it is obvious that $\o$ is an exact 1-form, and $d\o=0$. This finishes the proof. In the proof of (1)$\Rightarrow$(3), there is always a constant vector $g$ orthogonal to $\mathrm{Span}\{\xi_1,\xi_2,\eta_1,\eta_2\}$ in either of the three cases. Thus in the foliation described in (3) of Theorem \[thm-envelop\], each leave is now a geodesic in the corresponding space form. In other words, they are ruled submanifolds. This fact is already known in the study of austere submanifolds [@br],[@Dajczer3],[@Lu]. Two Möbius characterization results =================================== In this section we will give two characterization theorems (in terms of Möbius invariants) related with the following minimal Wintgen ideal submanifolds in $\mathbb{S}^5$. \[ex1\] Let $\gamma:N^2\to \mathbb{C}P^2$ be a holomorphic curve, and $\pi: \mathbb{S}^5 \rightarrow \mathbb{C}P^2$ be the Hopf fibration. Then the circle bundle $M^3 \subset \mathbb{S}^5$ over $N^2$ obtained by taking the Hopf fibers over $\gamma(N^2)$ is a three dimensional minimal Wintgen ideal submanifold in $\mathbb{S}^5$ as pointed out in [@Smet] (Example 6). Observe that $S^1$ acts by isometry on $\mathbb{S}^5$ whose orbits give the Hopf fibration. Thus for $M^3$ as above it has an induced $S^1$ symmetry. Consider the second fundamental forms given in ; the invariant $\nu$ must be a constant along every orbit of this $S^1$ action, which is exactly an integral curve of $E_3$ (see [@Smet] for details where they use $\xi$ to denote this $E_3$). So we have $G=E_3(\nu)=0$ in this special case. Since they are minimal, by Theorem \[thm-minimal\] we have $d\o=0$. These conditions characterize this class of submanifolds as below. \[thm-Hopf\] Let $x:M^3\rightarrow\mathbb{S}^5$ be a Wintgen ideal submanifold with non-integrable distribution $\mathbb{D}=\mathrm{Span}\{E_1,E_2\}$. If it satisfies $d\o=0, G=0$, then up to a Möbius transformation on $\mathbb{S}^5$, $x$ is the Hopf lift of a holomorphic curve given in Example \[ex1\]. When $G=0$, comparing the coefficients of $\o_1\wedge\o_3$ in and using yields $2\hat{F}=L^2$. From the proof to theorem \[thm-minimal\] we know that $x$ is Möbius equivalent to a minimal Wintgen ideal submanifold $$f=\frac{\hat{F}Y+\hat Y}{\sqrt{2\hat{F}}}: M^3\rightarrow \mathbb{S}^5 \subset \mathbb{R}^6,$$ where $\mathbb{R}^6=\mathrm{Span}_{\mathbb{R}} \{f,\eta_3,\eta_1,\eta_2,\xi_1,-\xi_2\}$. Using $\sim$ and $2\hat{F}=L^2, d\hat{F}=-2\o\hat{F}$, with respect to this frame we can write out the structure equations of $f$: $$\label{Theta1} d\begin{pmatrix} f\\ \eta_3\\ \eta_1\\ \eta_2\\ \xi_1\\ -\xi_2\end{pmatrix}=\begin{pmatrix} 0& L\o_3& L\o_1& L\o_2& 0& 0\\ -L\o_3& 0& -L\o_2& L\o_1& 0& 0\\ -L\o_1& L\o_2& 0& \Omega_{12}& \mu\o_2& -\mu\o_1\\ -L\o_2& -L\o_1& -\Omega_{12}& 0& \mu\o_1& \mu\o_2\\ 0& 0& -\mu\o_2& -\mu\o_1& 0& -\theta_{12}\\ 0& 0& \mu\o_1& -\mu\o_2& \theta_{12}& 0 \end{pmatrix}\begin{pmatrix} f\\ \eta_3\\ \eta_1\\ \eta_2\\ \xi_1\\ -\xi_2\end{pmatrix}.$$ Denote the frame as a matrix $T:M^3\to \mathrm{SO}(6)$ with respect to a fixed basis $\{{\bf e}_k\}_{k=1}^6$ of $\mathbb{R}^6$, we can rewrite as $$\label{Theta2} dT=\Theta T.$$ The algebraic form of $\Theta$ motivates us to introduce a complex structure ${\bf J}$ on $\mathbb{R}^6= \mathrm{Span}_\mathbb{R}\{f,\eta_3,\eta_1,\eta_2,\xi_1,\xi_2\}$ as below: $${\bf J}\begin{pmatrix}f\\ \eta_3\\ \eta_1\\ \eta_2\\ \xi_1\\ -\xi_2\end{pmatrix} =\begin{pmatrix} \begin{pmatrix}0& -1\\1 & 0\end{pmatrix} & & \\ & \begin{pmatrix}0& -1\\1 & 0\end{pmatrix} & \\ & &\begin{pmatrix}0& -1\\1 & 0\end{pmatrix} \end{pmatrix}\begin{pmatrix} f\\ \eta_3 \\ \eta_1 \\ \eta_2 \\ \xi_1 \\ -\xi_2 \end{pmatrix}.$$ Denote the diagonal matrix at the right hand side as $J_0$. Then the matrix representation of operator ${\bf J}$ under $\{{\bf e}_k\}_{k=1}^6$ is: $$J=T^{-1}J_0T.$$ Using $dT=\Theta T$ and the fact that $J_0$ commutes with $\Theta$, it is easy to verify $$dJ=-T^{-1}dT T^{-1}J_0T+T^{-1}J_0 dT =-T^{-1}\Theta J_0T+T^{-1}J_0 \Theta T=0.$$ So ${\bf J}$ is a well-defined complex structure on this $\mathbb{R}^6$. Another way to look at the structure equations is to consider the complex version: $$\begin{aligned} d(f+i\eta_3)&=-iL\o_3(f+i\eta_3)+L(\o_1-i\o_2)(\eta_1+i\eta_2), \label{5.1}\\ d(\eta_1+i\eta_2)&=-L(\o_1+i\o_2)(f+i\eta_3)-i\Omega_{12}(\eta_1+i\eta_2) +i\mu(\o_1-i\o_2)(\xi_1-i\xi_2), \notag\\ d(\xi_1-i\xi_2)&=i\mu(\o_1+i\o_2)(\eta_1+i\eta_2) +i\theta_{12}(\xi_1-i\xi_2).\label{5.3}\end{aligned}$$ Geometrically, this implies that $$\mathbb{C}^3=\mathrm{Span}_{\mathbb{C}}\{f+i\eta_3, \eta_1+i\eta_2,\xi_1-i\xi_2\},$$ is a fixed three dimensional complex vector space endowed with the complex structure $i$, which is identified with $(\mathbb{R}^6,{\bf J})$ via the following isomorphism between complex linear spaces: $$v\in \mathbb{C}^3~~\mapsto~~\mathrm{Re}(v)\in \mathbb{R}^6.$$ For example, $f+i\eta_3\mapsto f,if-\eta_3\mapsto -\eta_3$ and so on. The second geometrical conclusion is an interpretation of that $[f+i\eta_3]$ defines a holomorphic mapping from the quotient surface $\overline{M}=M^3/\Gamma$ to the projective plane $\mathbb{C}P^2$ (like the conclusion (1) in Theorem \[thm-envelop\]). Moreover, the unit circle in $$\mathrm{Span}_{\mathbb{R}}\{f,\eta_3\}=\mathrm{Span}_{\mathbb{R}}\{f,{\bf J}f\}=\mathrm{Span}_{\mathbb{C}}\{f+i\eta_3\}$$ is a fiber of the Hopf fibration of $\mathbb{S}^5\subset (\mathbb{R}^6,{\bf J})$. It corresponds to the subspace $\mathrm{Span}_{\mathbb{R}}\{Y,\hat{Y},\eta_3\}$, which is geometrically a leave of the foliation $(M^3,\Gamma)$ as described by conclusion (3) in Theorem \[thm-envelop\]. Thus the whole $M^3$ is the Hopf lift of $\overline{M}\to\mathbb{C}P^2$. In other words we have the following commutative diagram $$\begin{xy} (30,30)*+{M^3}="v1", (60,30)*+{\mathbb{S}^5}="v2", (90,30)*+{\mathbb{C}^3}="v3";% (30,0)*+{\overline{M}}="v4", (60,0)*+{\mathbb{C}P^2}="v5".% {\ar@{->}^{f} "v1"; "v2"}% {\ar@{->}^{\subset} "v2"; "v3"}% {\ar@{->}_{M^3/\Gamma} "v1"; "v4"}% {\ar@{->}_{[f+i\eta_3]} "v1"; "v5"}% {\ar@{->}^{} "v4"; "v5"}% {\ar@{->}^{\pi} "v2"; "v5"}% {\ar@{->}_{\pi} "v3"; "v5"}% \end{xy}$$ This finishes the proof. Among examples given above, there is a special one coming from the lift of the famous Veronese embedding $\gamma:\mathbb{C}P^1\to \mathbb{C}P^2$ which is homogeneous. Thus the lift $M^3$ is itself a homogeneous minimal Wintgen ideal submanifold in $\mathbb{S}^5$. This special example can also be described as below. \[ex2\] The orthogonal group $\mathrm{SO}(3)$ embedded in $\mathbb{S}^5$ homogeneously: $$\label{SO3} x: \quad \mathrm{SO}(3)~\rightarrow~ \mathbb{S}^5, \qquad (u, v, u\times v) \mapsto \frac{1}{\sqrt{2}}(u, v).$$ The orthonormal frames of the tangent and normal bundles can be chosen as $ e_1=(u\times v, 0), e_2=(0, u\times v), e_3=\frac{1}{\sqrt{2}}(-v, u);~~ n_1=\frac{-1}{\sqrt{2}}(v, u), n_2=\frac{-1}{\sqrt{2}}(u, -v). $ Direct computation verifies that it is a minimal Wintgen ideal submanifold. Consider the canonical lift $Y=\sqrt{6}(1,x): \mathrm{SO}(3)\rightarrow \mathbb{R}^7_1.$ The Möbius metric is given by $\mathrm{g}=6dx\cdot dx.$ It follows from that the Möbius form vanishes, i.e., $C^{r}_{j}=0$. Next, $\{E_j=e_j/\sqrt{6}\}$ form an orthonormal frame for $(M^3, \mathrm{g})$, with the dual 1-form $\{\o_j\}$. So the frame used in Section 3 is given by $$\eta_j=Y_j=(0,e_j), ~~\hat{Y}=N=\frac{1}{2\sqrt{6}}(-1,x), ~~\xi_r=(0,n_r).$$ The structure equations are $$d\begin{pmatrix} Y\\ \hat{Y} \\ \eta_1 \\ \eta_2 \\ \eta_3 \\ \xi_1 \\ \xi_2 \end{pmatrix}=\begin{pmatrix} 0&0&\o_1&\o_2&\o_3&0&0\\ 0&0&\frac{\o_1}{12}&\frac{\o_2}{12}&\frac{\o_3}{12}&0&0\\ \frac{-\o_1}{12}&-\o_1&0&0&\frac{\o_2}{\sqrt{6}}& \frac{\o_2}{\sqrt{6}}&\frac{\o_1}{\sqrt{6}}\\ \frac{-\o_2}{12}&-\o_2&0&0&\frac{-\o_1}{\sqrt{6}}& \frac{\o_1}{\sqrt{6}}&\frac{-\o_2}{\sqrt{6}}\\ \frac{-\o_3}{12}&-\o_3&\frac{-\o_2}{\sqrt{6}}& \frac{\o_1}{\sqrt{6}}&0&0&0\\ 0&0&\frac{-\o_2}{\sqrt{6}}&\frac{-\o_1}{\sqrt{6}}&0&0& \frac{\o_3}{\sqrt{6}}\\ 0&0&\frac{-\o_1}{\sqrt{6}}&\frac{\o_2}{\sqrt{6}}&0& \frac{-\o_3}{\sqrt{6}}&0 \end{pmatrix}\begin{pmatrix} Y\\ \hat{Y} \\ \eta_1 \\ \eta_2 \\ \eta_3 \\ \xi_1 \\ \xi_2 \end{pmatrix}.\label{4.15}$$ In [@Smet] they gave a characterization of this example as the unique Wintgen ideal submanifold $M^m\to\mathbb{Q}^{m+2}(c)$ with constant non-zero normal curvature. Here we provide another characterization of it in Möbius geometry. In the statement below, a connected submanifold $M$ in $\mathbb{S}^5$ is said to be *locally Möbius homogenous* if for any two points $p,q\in M$, there are two neighborhoods $U_p,U_q\subset M$ of them respectively and a Möbius transformation $T$ such that $T(p)=q, T(U_p)=U_q$. An essential property of a locally (Möbius) homogenous submanifold is that any well-defined (Möbius) invariant function on it must be a constant. \[thm-homog\] Let $x: M^3\rightarrow \mathbb{S}^5$ be a Wintgen ideal submanifold of dimension 3. If it is locally Möbius homogenous and the distribution $\mathbb{D}=\mathrm{Span}\{E_1, E_2\}$ is not integrable, then up to a Möbius transformation this is part of $x: \mathrm{SO}(3) \rightarrow \mathbb{S}^5$ given in Example \[ex2\]. According to Remark \[rem-transform\] and Remark \[rem-invariant\], the coefficients $\{\hat{F}, G, L\}$ appearing in the structure equations are geometric invariants. Thus under our assumption these functions must be constants; in particular $L$ is a non-zero constant. By we know $G=E_3(L)=0$. From $\sim$, we have $$\qquad 2\hat{F}=L^2,~~~ \hat\o_3=LU\o_1+LV\o_2+\hat{F}\o_3.\label{5.7}$$ By one can write out explicitly that $$\begin{aligned} \Omega_{12}&=\alpha\o_1+\beta\o_2+\gamma\o_3.\label{5.8}\\ \theta_{12}&=(U-2\alpha)\o_1+(V-2\beta)\o_2+(L-2\gamma)\o_3. \label{5.9}\end{aligned}$$ Note that in general $\alpha, \beta, \gamma$ are not geometric invariants, because they are components of the connection 1-form $\Omega_{12}$, and when the frame $\{E_1,E_2\}$ rotate by angle $t$ in , $\Omega_{12}$ will differ by a closed 1-form $dt$. On the other hand, by , $U^2+V^2$ is the square of the norm of the Möbius form $\Phi$ (up to a non-zero constant), hence a geometric invariant. It must also be a constant on $M^3$.\ **Claim**: The 1-form $\o=0$ identically; i.e., $U=V=0$ on $M^3$ everywhere and under any frame $\{E_1,E_2\}$. (As a consequence of this fact and the conclusion of Theorem \[thm-Hopf\], any of such examples is the Hopf lift of a complex curve in $\mathbb{C}P^2$.)\ We prove this claim by contradiction. Suppose $U^2+V^2$ is a non-zero constant. We can choose a canonical frame according to Remark \[rem-invariant\]. With respect to such a canonical frame, all coefficients $\alpha,\beta,\gamma$ in are now well-defined functions, hence be constants. From $\sim$ we have $$\begin{aligned} d\Omega_{12}=&(\alpha^2-\alpha U+\beta^2-\beta V+2L\gamma)\o_1\wedge\o_2 \notag \\ &+[\alpha(L-\gamma)+\gamma U]\o_2\wedge\o_3 -[\beta(L-\gamma)+\gamma V]\o_1\wedge\o_3,\label{dOmega12}\\ d\theta_{12}=&-[(U-2\alpha)(U-\alpha) +(V-2\beta)(V-\beta)-2(L-2\gamma)L]\o_1\wedge\o_2 \notag\\ &+[(U-2\alpha)(L-\gamma)+U(L-2\gamma)]\o_2\wedge\o_3 \notag\\ &-[(V-2\beta)(L-\gamma)+V(L-2\gamma)]\o_2\wedge\o_3. \label{dtheta12}\end{aligned}$$ Comparing the coefficients with and separately, we obtain $$\begin{gathered} \alpha^2-\alpha U+\beta^2-\beta V+2L\gamma=2\mu^2-2L^2, \label{5.10}\\ \alpha(L-\gamma)+\gamma U=0,\label{5.11}\\ \beta(L-\gamma)+\gamma V=0, \label{5.12}\\ 2\mu^2=-(U-2\alpha)(U-\alpha)-(V-2\beta)(V-\beta) +2(L-2\gamma)L,\label{5.13}\\ (U-2\alpha)(L-\gamma)+U(L-2\gamma)=0,\label{5.14}\\ (V-2\beta)(L-\gamma)+V(L-2\gamma)=0. \label{5.15}\end{gathered}$$ If $\gamma=L\ne 0$, then from and we have $U=V=0$. This does not only contradict with the assumption $U^2+V^2\ne 0$, but also implies from that $\mu^2=-\alpha^2-\beta^2-L^2$, a contradiction with $L\ne 0,\mu=1/\sqrt{6}\ne 0$. If $\gamma\ne L$, then and tell us $\alpha=-\frac{\gamma }{L-\gamma}U,\beta=-\frac{\gamma }{L-\gamma}V.$ Combined with and , we get $\gamma=2L, \alpha=2U, \beta=2V.$ Insert them into , we have $2\mu^2=-3U^2-3V^2-6L^2.$ So there must be $U=V=L=0$, which also contradicts to our assumption. Thus the claim is proved.\ Now that $U=V=\frac{G}{L}=0$, we have $\o=0$ and $2\Omega_{12}+\theta_{12}=L\o_3$. Differentiate the last equation. We get $$d(2\Omega_{12}+\theta_{12})=Ld\o_3=2L^2\o_1\wedge\o_2.$$ On the other hand, still by there is $$d(2\Omega_{12}+\theta_{12})=(6\mu^2-4L^2)\o_1\wedge\o_2.$$ Comparison shows $$L=\mu=1/\sqrt{6}$$ (assume $L>0$ without loss of generality). Then by and $2\hat{F}=L^2$ in , we obtain $$d\Omega_{12}=d\o_{12}=2(\mu^2-L^2)\o_1\wedge\o_2=0.$$ Thus $\Omega_{12}$ is a closed 1-form, which is locally an exact 1-form. Then we can use to find another frame such that $\Omega_{12}=\o_{12}=0$ and such frame is canonically chosen once it is fixed at an arbitrary point. With respect to this frame on a simply connected domain of $M^3$ we know $\alpha=\beta=\gamma=0$ in , and $\theta_{12}=\frac{1}{\sqrt{6}}\o_3$ in . The proof is finished by checking that the structure equations are the same as for $x(\mathrm{SO}(3))$. For Möbius homogeneous Wintgen ideal submanifolds $M^3\to \mathbb{S}^5$ with $L=0$ (integrable $\mathbb{D}$), the classification will not be difficult. By the conclusion of Theorem A in the introduction, such an example comes from super-minimal surface $\overline{M}$ in four dimensional space forms. This super-conformal $\overline{M}$ must also be homogeneous by itself. According to our classification of Willmore surfaces with constant Möbius curvature [@MaWang], this $\overline{M}$ should be the Veronese surface $\mathbb{R}P^2\to \mathbb{S}^4$, and the original $M^3$ is a cone in $\mathbb{R}^5$ over this surface. [99]{} R. Bryant, [*A duality theorem for Willmore surfaces,*]{} J. Diff. Geom. 20(1984), 20-53. R. Bryant, [*Some remarks on the geometry of austere manifolds,*]{} Bol. Soc. Bras. Mat., 21(1991), 122-157. F. Burstall, D. Ferus, K. Leschke, F. Pedit, U. Pinkall, [*Conformal geometry of surfaces in the 4-sphere and Quaternions,*]{} Lect. Notes Math., vol 1772, Springer-Verlag, 2002. M. Dajczer, R. Tojeiro, [*All superconformal surfaces in $R^4$ in terms of minimal surfaces,*]{} Math. Z. 261 (2009), no. 4, 869-890. M. Dajczer, R. Tojeiro, [*Submanifolds of codimension two attaining equality in an extrinsic inequality,*]{} Math. Proc. Cambridge Philos. Soc. 146 (2009), no. 2, 461-474. M. Dajczer, R. Tojeiro, [*A class of austere submanifolds,*]{} Illinois J. Math., 45(2001), no.3, 735-755. P. J. De Smet, F. Dillen, L. Verstraelen, L. Vrancken, [*A pointwise inequality in submanifold theory,*]{} Archivum Mathematicum (Brno), 35(1999), 115-128. J. Ge, Z. Tang, [*A proof of the DDVV conjecture and its equality case,*]{} Pacific J. Math. 237 (2008), 87-95. I. Guadalupe, L. Rodríguez, [*Normal curvature of surfaces in space forms,*]{} Pacific J. Math., 106(1983), 95-103. U. Hertrich-Jeromin, [*Introduction to Möbius Differential Geometry,*]{} Cambridge University Press, 2003. T. Li, X. Ma, C. Wang, [*Deformation of Hypersurfaces Preserving the Möbius Metric and a Reduction Theorem,*]{} arXiv:1204.1408, accepted by Adv. in Math. T. Li, X. Ma, C. Wang, [*Wintgen ideal submanifolds with a low-dimensional integrable distribution (I),*]{} arxiv:1301.4742. T. Li, X. Ma, C. Wang, Z. Xie, [*Wintgen ideal submanifolds with a low-dimensional integrable distribution (II),*]{} in preparation. T. Li, X. Ma, C. Wang, Z. Xie, [*Wintgen ideal submanifolds of codimension two, complex curves, and Möbius geometry,*]{} preprint. H. Liu, C. Wang, G. Zhao, [*Möbius isotropic submanifolds in $S^n$*]{}, Tohoku Math. J., 53(2001), 553-569. T. Choi and Z. Lu, [*On the DDVV conjecture and the comass in calibrated geometry (I),*]{} Math. Z., 260(2008), 409-429. Z. Lu, [*On the DDVV conjecture and the comass in calibrated geometry (II),*]{} arXiv:Math.DG/0708.2921. Z. Lu, [*Recent developments of the DDVV conjecture,* ]{} Bull. Transil. Univ. Brasov, ser.B, 14(2008), no. 49, 133-144. Z. Lu, [*Normal scalar curvature conjecture and its applications,*]{} Journal of Functional Analysis 261 (2011), 1284-1308. H. Ma, Y. Ohnita, [*Hamiltonian stability of the Gauss images of homogeneous isoparametric hypersurfaces,*]{} arxiv:1207.0338 X. Ma, C. Wang, [*Willmore Surfaces of Constant M¡§obius Curvature,*]{} Ann. Glob. Anal. Geom. 32 (2007), no.3, 297-310. C. Wang, [*Möbius geometry of submanifolds in $S^n$,*]{} Manuscripta Math., 96(1998), 517-534. P. Wintgen, [*Sur l’inégalité de Chen-Willmore,*]{} C. R. Acad. Sci. Paris, 288(1979),993-995. Zhenxiao Xie, [*School of Mathematical Sciences, Peking University, Beijing 100871, People’s Republic of China. e-mail: [[email protected]]{}*]{} Tongzhu Li, [*Department of Mathematics, Beijing Institute of Technology, Beijing 100081, People’s Republic of China. e-mail:[[email protected]]{}*]{} Xiang Ma, [*LMAM, School of Mathematical Sciences, Peking University, Beijing 100871, People’s Republic of China. e-mail: [[email protected]]{}*]{} Changping Wang [*School of Mathematics and Computer Science, Fujian Normal University, Fuzhou 350108, People’s Republic of China. e-mail: [[email protected]]{}*]{}

ArXiv

--- abstract: 'A computational tool for coarse-graining nonlinear systems of ordinary differential equations in time is discussed. Three illustrative model examples are worked out that demonstrate the range of capability of the method. This includes the averaging of Hamiltonian as well as dissipative microscopic dynamics whose ‘slow’ variables, defined in a precise sense, can often display mixed slow-fast response as in relaxation oscillations, and dependence on initial conditions of the fast variables. Also covered is the case where the quasi-static assumption in solid mechanics is violated. The computational tool is demonstrated to capture all of these behaviors in an accurate and robust manner, with significant savings in time. A practically useful strategy for accurately initializing short bursts of microscopic runs for the evolution of slow variables is integral to our scheme, without the requirement that the slow variables determine a unique invariant measure of the microscopic dynamics.' author: - 'Sabyasachi Chatterjee[^1]' - 'Amit Acharya[^2]' - 'Zvi Artstein[^3]' bibliography: - 'testbibliog.bib' title: 'Computing Singularly Perturbed Differential Equations[^4]' --- Introduction ============ This paper is concerned with a computational tool for understanding the behavior of systems of evolution, governed by (nonlinear) ordinary differential equations, on a time scale that is much slower than the time scales of the intrinsic dynamics. A paradigmatic example is a molecular dynamic assembly under loads, where the characteristic time of the applied loading is very much larger than the period of atomic vibrations. We examine appropriate theory for such applications and devise a computational algorithm. The singular perturbation problems we address contain a small parameter $\epsilon$ that reflects the ratio between the slow and the fast time scales. In many cases, the solutions of the problem obtained by setting the small parameter to zero matches solutions to the full problem with small $\epsilon$, except in a small region - a boundary/initial layer. But, there are situations, where the limit of solutions of the original problem as $\epsilon$ tends to zero does not match the solution of the problem obtained by setting the small parameter to zero. Our paper covers this aspect as well. In the next section we present the framework of the present study, and its sources. Before displaying our algorithm in Section \[sec:theory\_alg\], we display previous approaches to the computational challenge. It allows us to pinpoint our contribution. Our algorithm is demonstrated through computational examples on three model problems that have been specially designed to contain the complexities in temporal dynamics expected in more realistic systems. The implementation is shown to perform robustly in all cases. These cases include the averaging of fast oscillations as well as of exponential decay, including problems where the evolution of slow variables can display fast, almost-discontinuous, behavior in time. The problem set is designed to violate any ergodic assumption, and the computational technique deals seamlessly with situations that may or may not have a unique invariant measure for averaging fast response for fixed slow variables. Thus, it is shown that initial conditions for the fast dynamics matter critically in many instances, and our methodology allows for the modeling of such phenomena. The method also deals successfully with conservative or dissipative systems. In fact, one example on which we demonstrate the efficacy of our computational tool is a linear, spring-mass, *damped* system that can display permanent oscillations depending upon delicate conditions on masses and spring stiffnesses and initial conditions; we show that our methodology does not require *a-priori* knowledge of such subtleties in producing the correct response. The framework ============= A particular case of the differential equations we deal with is of the form $$\begin{aligned} \label{eq:1.1} \frac{dx}{dt} &= \frac{1}{\epsilon} F(x)+ G(x),\end{aligned}$$ with $ \epsilon > 0 $ a small real parameter, and $x \in \mathbb{R}^{n}$. For reasons that will become clear in the sequel we refer to the component $G(x)$ as the drift component. Notice that the dynamics in (\[eq:1.1\]) does not exhibit a prescribed split into a fast and a slow dynamics. We are interested in the case where such a split is either not tractable or does not exist. Another particular case where a split into a fast and slow dynamics can be identified, is also of interest to us, as follows. $$\begin{aligned} \label{eq:1.2} \frac{dx}{dt} &= \frac{1}{\epsilon} F(x,l)\\ \frac{dl}{dt} &= L(x,l),\nonumber\end{aligned}$$ with $x \in \mathbb{R}^n$ and $l \in \mathbb{R}^m$. We think of the variable $l$ as a load. Notice that the dynamics of the load is determined by an external “slow" equation, that, in turn, may be affected by the “fast” variable $x$. The general case we study is a combination of the previous two cases, namely, $$\begin{aligned} \label{eq:1.3} \frac{dx}{dt} &= \frac{1}{\epsilon} F(x,l)+ G(x,l)\\ \frac{dl}{dt} &= L(x,l),\nonumber\end{aligned}$$ which accommodates both a drift and a load. In the theoretical discussion we address the general case. We display the two particular cases, since there are many interesting examples of the type (\[eq:1.1\]) or (\[eq:1.2\]). An even more general setting would be the case where the right hand side of (\[eq:1.3\]) is of the form $H(x, l, \epsilon)$, namely, there is no a priori split of the right hand side of the equation into fast component and a drift or a slow component. A challenge then would be to identify, either analytically or numerically, such a split. We do not address this case here, but our study reveals what could be promising directions of such a general study. We recall that the parameter $\epsilon$ in the previous equation represents the ratio between the slow (or ordinary) part in the equation and the fast one. In Appendix B we examine one of our examples, and demonstrate how to derive the dimensionless equation with the small parameter, from the raw mechanical equation. In real world situations, $\epsilon$ is small yet it is not infinitesimal. Experience teaches us, however, that the limit behavior, as $\epsilon$ tends to 0, of the solutions is quite helpful in understanding of the physical phenomenon and in the computations. This is, indeed, demonstrated in the examples that follow. References that carry out a study of equations of the form (\[eq:1.1\]) are, for instance, Tao, Owhadi and Marsden [@Tao_2010_Noninstructive], Artstein, Kevrekidis, Slemrod and Titi [@artstein2007slow], Ariel, Engquist and Tsai [@ariel2009multiscale; @ariel2009numerical], Artstein, Gear, Kevrekidis, Slemrod and Titi [@Artstein_Gear_2011_KdV_burgers], Slemrod and Acharya [@Slemord_Acharya_2012_Time_average]; conceptually similar questions implicitly arise in the work of Kevrekidis et al. [@kevrekidis2003equation]. The form (\[eq:1.2\]) coincides with the Tikhonov model, see, e.g., O’Malley [@O_Malley_2014_Singular_perturb], Tikhonov, Vasileva and Sveshnikov [@Tikhonov_1985_DE], Verhulst [@Verhulst_2005_App_of_sing_perturb], or Wasow [@Wasow_1965_Asym]. The literature concerning this case followed, mainly, the so called Tikhonov approach, namely, the assumption that the solutions of the $x$-equation in (\[eq:1.2\]), for $l$ fixed, converge to a point $x(l)$ that solves an algebraic equation, namely, the second equation in (\[eq:1.3\]) where the left hand side is equal to 0. The limit dynamics then is a trajectory $(x(t), l(t))$, evolving on the manifold of stationary points $x(l)$. We are interested, however, in the case where the limit dynamics may not be determined by such a manifold, and may exhibit infinitely rapid oscillations. A theory and applications alluding to such a case are available, e.g., in Artstein and Vigodner [@Artstein_Vigonder_1996_dynamic_limit], Artstein [@Artstein_2002_perturbed], Acharya [@Acharya2007coarse_variable; @Acharya_2010_Coarsegraining_autonomous], Artstein, Linshiz and Titi [@Artstein_Linshiz_2007_Young_measure], Artstein and Slemrod [@Artstein_Slemord_2001_sigular_perturb_limit]. The goal ======== A goal of our study is to suggest efficient computational tools that help revealing the limit behavior of the system as $\epsilon$ gets very small, this on a prescribed, possibly long, interval. The challenge in such computations stems from the fact that, for small $\epsilon$, computing the ordinary differential equation takes a lot of computing time, to the extent that it becomes not practical. Typically, we are interested in a numerical description of the full solution, namely, the progress of the coupled slow/fast dynamics. At times, we may be satisfied with partial information, say in the description of the progress of a slow variable, reflecting a measurement of the underlying dynamics. To that end we first identify the mathematical structure of the limit dynamics on the given interval. The computational algorithm will reveal an approximation of this limit dynamics, that, in turn, is an approximation of the full solution for arbitrarily small $\epsilon$. If only a slow variable is of interest, it can be derived from the established approximation. The limit dynamics ================== In order to achieve the aforementioned goal, we display the limit structure, as $\epsilon \rightarrow 0,$ of the dynamics of (\[eq:1.3\]). To this end we identify the $\textit{fast time equation}$ $$\begin{aligned} \label{eq:3.1} \frac{dx}{d\sigma} = F(x, l),\end{aligned}$$ when $l$ is held fixed (recall that $l$ may not show up at all, as in (\[eq:1.1\])). The equation (\[eq:3.1\]) is the $\textit{fast part}$ of (\[eq:1.3\]) (as mentioned, $G(x)$ is the $\textit{drift}$ and the solution $l(t)$ of the load equation is the $\textit{load}$). Notice that when moving from (\[eq:1.3\]) to (\[eq:3.1\]), we have changed the time scale, with $t = \epsilon \sigma.$ We refer to $\sigma$ as the fast time scale. In order to describe the limit dynamics of (\[eq:1.3\]) we need the notions of: Probability measures and convergence of probability measures, Young measures and convergence in the Young measures sense, invariant measures and limit occupational measures. In particular, we shall make frequent use of the fact that when occupational measures of solutions of (\[eq:3.1\]), on long time intervals, converge, the limit is an invariant measure of (\[eq:3.1\]). A concise explanation of these notions can be found, e.g., in [@Artstein_Vigonder_1996_dynamic_limit; @artstein2007slow]. It was proved in [@Artstein_Vigonder_1996_dynamic_limit] for (\[eq:1.2\]) and in [@artstein2007slow] for (\[eq:1.1\]), that under quite general conditions, the dynamics converge, as $\epsilon\rightarrow0,$ to a Young measure, namely, a probability measure-valued map, whose values are invariant measures of (\[eq:3.1\]). These measures are drifted in the case of (\[eq:1.1\]) by the drift component of the equation, and in the case (\[eq:1.2\]) by the load. We display the result in the general case after stating the assumptions under which the result holds. **Assumption 4.1.** The functions $F(.,.), G(.,.)$ and $L(.,.)$ are continuous. The solutions, say $x(.)$, of the fast equation (\[eq:3.1\]), are determined uniquely by the initial data, say $x(\sigma_{0}) = x_{0}$, and stay bounded for $\sigma \geq \sigma_{0}$, uniformly for $x_{0}$ and for $l$ in bounded sets. Here is the aforementioned result concerning the structure of the limit dynamics. **Theorem 4.2.** For every sequence $\epsilon_{i}\rightarrow0$ and solutions $(x_{\epsilon_{i}}(t), l_{\epsilon_{i}}(t))$ of the perturbed equation (\[eq:1.3\]) defined on $[0, T]$, with $(x_{\epsilon_{i}}(0), l_{\epsilon_{i}}(0))$ in a bounded set, there exists a subsequence $\epsilon_{j}$ such that $(x_{\epsilon_{j}}(.), l_{\epsilon_{j}}(.))$ converges as $j\rightarrow\infty$, where the convergence in the $x$-coordinates is in the sense of Young measures, to a Young measure, say $\mu(.)$, whose values are invariant measures of the fast equation (\[eq:3.1\]), and the convergence in the $l$-coordinates is uniform on the interval, with a limit, say $l_{0}(.)$, that solves the differential equation $$\begin{aligned} \label{eq:3.2} \frac{dl}{dt}=\int_{\mathbb{R}^{n}}L(x,l)\mu(t)dx.\end{aligned}$$ The previous general result has not been displayed in the literature, but the arguments in [@artstein2007slow] in regard to (\[eq:1.1\]) or the proof given in [@Artstein_Vigonder_1996_dynamic_limit] for the case (\[eq:1.2\]), apply to the present setting as well. Measurements and slow observables ================================= A prime role in our approach is played by $\textit{slow observables}$, whose dynamics can be followed. The intuition behind the notion is that the observations which the observable reveals, is a physical quantity on the macroscopic level, that can be detected. Here we identify some candidates for such variables. The role they play in the computations is described in the next section. In most generality, an $\textit{observable}$ is a mapping that assigns to a probability measure $\mu(t)$ arising as a value of the Young measure in the limit dynamics of (\[eq:1.3\]), a real number, or a vector, say in $\mathbb{R}^{k}$. Since the values of the Young measure are obtained as limits of occupational measures (that in fact we use in the computations), we also demand that the observable be defined on these occupational measures, and be continuous when passing from the occupational measures to the value of the Young measure. An observable $v(.)$ is a $\textit{slow observable}$ if when applied to the Young measure $\mu(.)$ that determines the limit dynamics in Theorems 4.2, the resulting vector valued map $v(t) = v(\mu(t), l(t))$ is continuous at points where the measure $\mu(.)$ is continuous. An $\textit{extrapolation rule}$ for a slow observable $v(.)$ determines an approximation of the value $v(t + h)$, based on the value $v(t)$ and, possibly, information about the value of the Young measure $\mu(t)$ and the load $l(t)$, at the time $t$. A typical extrapolation rule would be generated by the derivative, if available, of the slow observable. Then $v(t + h) = v(t) + h\frac{dv}{dt}(t).$ A trivial example of a slow observable of (\[eq:1.3\]) with an extrapolation rule is the variable $l(t)$ itself. It is clearly slow, and the right hand side of the differential equation $\eqref{eq:3.2}$ determines the extrapolation rule, namely : $$\begin{aligned} \label{eq:4.1} l(t + h) = l(t) + h\frac{dl}{dt}(t).\end{aligned}$$ An example of a slow observable possessing an extrapolation rule in the case of (\[eq:1.1\]), is an $\textit{orthogonal observable}$, introduced in [@artstein2007slow]. It is based on a mapping $m(x, l):\mathbb{R}^{n}\rightarrow \mathbb{R}$ which is a first integral of the fast equation (\[eq:3.1\]) (with $l$ fixed), namely, it is constant along solutions of ($\ref{eq:3.1}$). Then we define the observable $v(\mu) = m(x,l)$ with $x$ any point in the support of $\mu$. But in fact, it will be enough to assume that the mapping $m(x,l)$ is constant on the supports of the invariant measures arising as values of a Young measure. The definition of $v(\mu) = m(x, l)$ with $x$ any point in the support of $\mu$ stays the same, that is, $m(x,l)$ may not stay constant on solutions away from the support of the limit invariant measure. It was shown in [@artstein2007slow] for the case (\[eq:1.1\]), that if $m(.)$ is continuously differentiable, then $v(t)$ satisfies, almost everywhere, the differential equation $$\begin{aligned} \label{eq:4.2} \frac{dv}{dt}= \int_{\mathbb{R}^{n}}\nabla m(x)G(x)\mu(t)dx.\end{aligned}$$ It is possible to verify that the result holds also when the observable satisfies the weaker condition just described, namely, it is a first integral only on the invariant measures that arise as values of the limit Young measure. The differential equation ($\ref{eq:4.2}$) is not in a closed form, in particular, it is not an ordinary differential equation. Yet, if one knows $\mu(t)$ and $v(t)$ at time $t$, the differentiability expressed in ($\ref{eq:4.2}$) can be employed to get an extrapolation of the form $v(t + h) = v(t) + h \frac{dv}{dt}(t)$ at points of continuity of the Young measure, based on the right hand side of (\[eq:4.2\]). A drawback of an orthogonal observable for practical purposes is that finding first integrals of the fast motion is, in general, a non-trivial matter. A natural generalization of the orthogonal observable would be to consider a moment or a generalized moment, of the measure $\mu(t)$. Namely, to drop the orthogonality from the definition, allowing a general $m$ : $\mathbb{R}^{n}\rightarrow \mathbb{R}$ be a measurement (that may depend, continuously though, on $l$ when $l$ is present), and define $$\begin{aligned} \label{eq:4.3} v(\mu) = \int_{\mathbb{R}^n} m(x) \mu(dx).\end{aligned}$$ Thus, the observable is an average, with respect to the probability measure, of the bounded continuous measurement $m(.)$ of the state. If one can verify, for a specific problem, that $\mu(t)$ is piecewise continuous, then the observable defined in (\[eq:4.3\]) is indeed slow. The drawback of such an observable is the lack of an apparent extrapolation rule. If, however, in a given application, an extrapolation rule for the moment can be identified, it will become a useful tool in the analysis of the equation. A generalization of (\[eq:4.3\]) was suggested in [@acharya2006computational; @ariel2009multiscale] in the form of running time-averages as slow variables, and was made rigorous in the context of delay equations in [@Slemord_Acharya_2012_Time_average]. Rather than considering the average of the bounded and continuous function $m(x)$ with respect $\mu(t)$, we suggest considering the average with respect to the values of the Young measure over an interval \[$t-\Delta,t$\], i.e, $$\begin{aligned} \label{eq:4.4} v(t)=\frac{1}{\Delta}\int_{t-\Delta}^{t} \int_{\mathbb{R}^n} m(x)\mu(s)(dx)ds.\end{aligned}$$ Again, the measurement $m$ may depend on the load. Now the observable (\[eq:4.4\]) depends not only on the value of the measure at $t$, but on the “history" of the Young measure, namely its values on \[$t-\Delta,t$\]. The upside of the definition is that $v(t)$ is a Lipschitz function of $t$ (the Lipschitz constant may be large when $\Delta$ is small) and, in particular, is almost everywhere differentiable. The almost everywhere derivative of the slow variable is expressed at the points $t$ where $\mu(.)$ is continuous at $t$ and at $t-\Delta$, by $$\begin{aligned} \label{eq:4.5} \frac{dv}{dt}=\frac{1}{\Delta} \left( \int_{\mathbb{R}^n} m(x)\mu(t)(dx)-\int_{\mathbb{R}^n} m(x)\mu(t-\Delta)(dx)\right).\end{aligned}$$ This derivative induces an extrapolation rule. For further reference we call an observable that depends on the values of the Young measure over an interval prior to $t$, an $\textit{H-observable}$ (where the $H$ stands for history). An $H$-observable need not be an integral of generalized moments, i.e., of integrals. For instance, for a given measure $\mu$ let $$\begin{aligned} \label{eq:4.6} r(\mu) = max\{x \cdot e_{1}: x \in supp(\mu)\},\end{aligned}$$ where $e_{1}$ is a prescribed unit vector and supp($\mu$) is the support of $\mu$. Then, when supp($\mu$) is continuous in $\mu$, (and recall Assumption 4.1) the expression $$\begin{aligned} \label{eq:4.7} v(t) = \frac{1}{\Delta}\int_{t-\Delta}^{t} r(\mu(\tau))d\tau,\end{aligned}$$ is a slow observable, and $$\begin{aligned} \label{eq:4.8} \frac{dv}{dt} = \frac{1}{\Delta} \left(r(\mu(t))- r(\mu(t-\Delta))\right)\end{aligned}$$ determines its extrapolation rule. The strategy we display in the next section applies whenever slow observables with valid extrapolation rules are available. The advantage of the $H$-observables as slow variables is that any smooth function $m(.)$ generates a slow observable and an extrapolation rule. Plenty of slow variables arise also in the case of generalized moments of the measure, but then it may be difficult to identify extrapolation rules. The reverse situation occurs with orthogonal observables. It may be difficult to identify first integrals of (\[eq:3.1\]), but once such an integral is available, its extrapolation rule is at hand. Also note that in all the preceding examples the extrapolation rules are based on derivatives. We do not exclude, however, cases where the extrapolation is based on a different argument. For instance, on information of the progress of some given external parameter, for instance, a control variable. All the examples computed in the present paper will use $H$-observables. The algorithm {#sec:theory_alg} ============= Our strategy is a modification of a method that has been suggested in the literature and applied in some specific cases. We first describe these, as it will allow us to pinpoint our contribution. A computational approach to the system (\[eq:1.2\]) has been suggested in Vanden-Eijnden [@Vanden_2003_Numtech] and applied in Fatkullin and Vanden-Eijnden [@Fatkullin_2004_Lorenz]. It applies in the special case where the fast process is stochastic, or chaotic, with a unique underlying measure that may depend on the load. The underlying measure is then the invariant measure arising in the limit dynamics in our Theorem 4.2. It can be computed by solving the fast equation, initializing it at an arbitrary point. The method carried out in [@Vanden_2003_Numtech] and [@Fatkullin_2004_Lorenz] is, roughly, as follows. Suppose the value of the slow variable (the load in our terminology) at time $t$ is known. The fast equation is run then until the invariant measure is revealed. The measure is then employed in the averaging that determines the right hand side of the slow equation at $t$, allowing to get a good approximation of the load variable at $t + h$. Repeating this scheme results in a good approximation of the limit dynamics of the system. The method relies on the property that the value of the load determines the invariant measure. The latter assumption has been lifted in [@Artstein_Linshiz_2007_Young_measure], analyzing an example where the dynamics is not ergodic and the invariant measure for a given load is not unique, yet the invariant measure appearing in the limit dynamics can be detected by a good choice of an initial condition for the fast dynamics. The method is, again, to alternate between the computation of the invariant measure at a given time, say $t$, and using it then in the averaging needed to determine the slow equation. Then determine the value of the load at $t+h$. The structure of the equation allows to determine a good initial point for computing the invariant measure at $t + h$, and so on and so forth, until the full dynamics is approximated. The weakness of the method described in the previous paragraph is that it does not apply when the split to fast dynamics and slow dynamics, i.e. the load, is not available, and even when a load is there, it may not be possible to determine the invariant measure using the value of the load. Orthogonal observables were employed in [@artstein2007slow] in order to analyze the system (\[eq:1.1\]), and a computational scheme utilizing these observables was suggested. The scheme was applied in [@Artstein_Gear_2011_KdV_burgers]. The suggestion in these two papers was to identify orthogonal observables whose values determine the invariant measure, or at least a good approximation of it. Once such observables are given, the algorithm is as follows. Given an invariant measure at $t$, the values of the observables at $t + h$ can be determined based on their values at $t$ and the extrapolation rule based on (\[eq:4.2\]). Then a point in $\mathbb{R}^n$ should be found, which is compatible with the measurements that define the observables. This point would be detected as a solution of an algebraic equation. Initiating the fast equation at that point and solving it on a long fast time interval, reveals the invariant measure. Repeating the process would result in a good approximation of the full dynamics. The drawback of the previous scheme is the need to find orthogonal observables, namely measurements that are constant along trajectories within the invariant measures in the limit of the fast flow, and verify that their values determine the value of the Young measure, or at least a good approximation of it. Also note that the possibility to determine the initialization point with the use of measurements, rather than the observables, relies on the orthogonality. Without orthogonality, applying the measurements to the point of initialization of the fast dynamics, yields no indication. In this connection, it is important to mention the work of Ariel, Engquist, and Tsai [@ariel2009multiscale; @ariel2009numerical] that demonstrates theory, and computations utilizing the Hetergeneous Multiscale Modeling (HMM) scheme of E and Engquist [@weinan2003heterognous], for defining a complete set of slow variables that determine the unique invariant measure for microscopic systems equipped with such. The scheme that we suggest employs general observables with extrapolation rules. As mentioned, there are plenty of these observables, in particular $H$-observables, as (\[eq:4.4\]) indicates. The scheme shows how to use them in order to compute the full dynamics, or a good approximation of it. It should be emphasized that none of our examples satisfy the ergodicity assumption placed in the aforementioned literature. Also, in none of the examples it is apparent, if possible at all, to find orthogonal observables. In addition, our third example exhibits extremely rapid variation in the slow variable (resembling a jump), a phenomenon not treated in the relevant literature so far. We provide two versions of the scheme. One for observables determined by the values $\mu(t)$ of the Young measure, and the second for $H$-observables, namely, observables depending on the values of the Young measure over an interval \[$t-\Delta,t$\]. The modifications needed in the latter case are given in parentheses. **The scheme**. Consider the full system (\[eq:1.3\]), with initial conditions $x(t_{0}) = x_{0}$ and $l(t_{0}) = l_{0}$. Our goal is to produce a good approximation of the limit solution, namely the limit Young measure $\mu(.)$, on a prescribed interval \[$t_{0}, T_{0}$\]. **A general assumption**. The function of time defined by the closest-point projection, in some appropriate metric, of any fixed point in state-space on the support of $\mu(t)$ for each $t$ in any interval in which $\mu(\cdot)$ is continuous, is smooth. A number of slow observables, say $v_1,\ldots, v_k$ can be identified, each of them equipped with an extrapolation rule, valid at all continuity points of the Young measure. **Initialization of the algorithm**. Solve the fast equation (\[eq:3.1\]) with initial condition $x_{0}$ and a fixed initial load $l_{0}$, long enough to obtain a good approximation of the value of the Young measure at $t_{0}$. (In the case of an $H$-observable solve the full equation with $\epsilon$ small, on an interval \[$t_{0}, t_{0} + \Delta$\], with $\epsilon$ small enough to get a good approximation of the Young measure on the interval). In particular the initialization produces good approximations of $\mu(t)$ for $t = t_{0}$ (for $t = t_{0} +\Delta$ in case of an $H$-observable). Compute the values $v_{1}(\mu(t)), ..., v_{k}(\mu(t))$ of the observables at this time $t$. **The recursive part.** **Step 1:** We assume that at time $t$ the values of the observables $v_{1}(\mu(t)), ..., v_{k}(\mu(t))$, applied to the value $\mu(t)$ of the Young measure, are known. We also assume that we have enough information to invoke the extrapolation rule to these observables (for instance, we can compute $\frac{dv}{dt}$ which determines the extrapolation when done via a derivative). If the model has a load variable, it should be one of the observables. **Step 2:** Apply the extrapolation rule to the observables and get an approximation of the values $v_{1}(\mu(t + h)),..., v_{k}(\mu(t + h))$, of the observables at time $t + h$ (time $t + h -\Delta$ in the case of an $H$-observable). Denote the resulting approximation by ($v_{1}$, ..., $v_{k}$). **Step 3:** Make an intelligent guess of a point $x(t + h)$ (or $x(t + h -\Delta)$ in the case of $H$-observable), that is in the basin of attraction of $\mu(t + h)$ (or $\mu(t + h -\Delta)$ in the case of $H$-observable). See a remark below concerning the intelligent guess. **Step 4:** Initiate the fast equation at $(x(t + h), l(t + h))$ and run the equation until a good approximation of an invariant measure arises (initiate the full equation, with small $\epsilon$ at $(x(t + h -\Delta), l(t + h -\Delta))$, and run it on \[$t + h -\Delta, t + h$\], with $\epsilon$ small enough such that a good approximation for a Young measure on the interval is achieved). **Step 5:** Check if the invariant measure $\mu(t + h)$ revealed in the previous step is far from $\mu(t)$ (this step, and consequently step 6.1, should be skipped if the Young measure is guaranteed to be continuous). **Step 6.1:** If the answer to the previous step is positive, it indicates that a point of discontinuity of the Young measure may exist between $t$ and $t + h$. Go back to $t$ and compute the full equation (initiating it with any point on the support of $\mu(t)$) on \[$t, t+h$\], or until the discontinuity is revealed. Then start the process again at Step 1, at the time $t + h$ (or at a time after the discontinuity has been revealed). **Step 6.2:** If the answer to the previous step is negative, compute the values of the observables $v_{1}(\mu(t + h)), ..., v_{k}(\mu(t + h))$ at the invariant measure that arises by running the equation. If there is a match, or almost a match, with $(v_{1}, ..., v_{k})$, accept the value $\mu(t + h)$ (accept the computed values on \[$t + h -\Delta,t + h$\] in the case of an $H$-observable) as the value of the desired Young measure, and start again at Step 1, now at time $t + h$. If the match is not satisfactory, go back to step 3 and make an improved guess. [ **Conclusion of the algorithm.** Continue with steps 1 to 6 until an approximation of the Young measure is computed on the entire time interval \[$t_{0}, T_{0}$\]. ]{} **Making the intelligent guess in Step 3.** The goal in this step is to identify a point in the basin of attraction of $\mu(t+h)$. Once a candidate for such a point is suggested, the decision whether to accept it or not is based on comparing the observables computed on the invariant measure generated by running the equation with this initial point, to the values as predicted by the extrapolation rule. If there is no match, we may improve the initial suggestion for the point we seek. In order to get an initial point, we need to guess the direction in which the invariant measure is drifted. We may assume that the deviation of $\mu(t+h)$ from $\mu(t)$ is similar to the deviation of $\mu(t)$ from $\mu(t-h)$. Then the first intelligent guess would be, say, a point $x_{t+h}$ such that $x_{t+h}-x_{t}= x_{t}-x_{t-h}$ where $x_{t}$ is a point in the support of $\mu(t)$ and $x_{t-h}$ is the point in the support of $\mu(t-h)$ closest, or near closest, to $x_{t}$. At this point, in fact, we may try several such candidates, and check them in parallel. If none fits the criterion in Step 6.2, the process that starts again at Step 3, could use the results in the previous round, say by perturbing the point that had the best fit in a direction of, hopefully, a better fit. A sequence of better and better approximations can be carried out until the desired result is achieved. The expected savings in computer time ===================================== The motivation behind our scheme of computations is that the straightforward approach, namely, running the entire equation (\[eq:1.3\]) on the full interval, is not feasible if $\epsilon$ is very small. To run (\[eq:3.1\]) in order to compute the invariant measure at a single point $t$, or to run (\[eq:1.3\]) on a short interval \[$t -\Delta,t$\], with $\Delta$ small, does not consume a lot of computing time. Thus, our scheme replaces the massive computations with computing the values of the Young measure at a discrete number of points, or short intervals, and the computation of the extrapolation rules to get an estimate of the progress of the observables. The latter step does not depend on $\epsilon$, and should not consume much computing time. Thus, if $h$ is large (and large relative to $\Delta$ in the case of $H$-observables), we achieve a considerable saving. These arguments are also behind the saving in the cited references, i.e., [@Vanden_2003_Numtech; @Fatkullin_2004_Lorenz; @Artstein_Linshiz_2007_Young_measure; @Artstein_Gear_2011_KdV_burgers]. In our algorithm there is an extra cost of computing time, namely, the need to detect points in the basin of attraction of the respective invariant measures, i.e., Step 3 in our algorithm. The extra steps amount to, possibly, an addition of a discrete number of computations that reveal the invariant measures. An additional computing time may be accrued when facing a discontinuity in the Young measure. The cost is, again, a computation of the full Young measure around the discontinuity. The possibility of discontinuity has not been address in the cited references. Error estimates =============== The considerations displayed and commented on previously were based on the heuristics behind the computations. Under some strong, yet common, conditions on the smoothness of the processes, one can come up with error estimate for the algorithm. We now produce such an estimate, which is quite standard (see e.g., [@Num1978]). Of interest here is the nature of assumptions on our process, needed to guarantee the estimate. Suppose the computations are performed on a time interval \[$0, T$\], with time steps of length $h$. The estimate we seek is the distance between the computed values, say $P(kh)$, $k = 0, 1, ...,N,$ of invariant measures, and the true limit dynamics, that is the values $\mu(kh)$ of the limit Young measure, at the same mesh points (here $hN$ is close to $T$). Recall that the Young measure we compute is actually the limit as $\epsilon\rightarrow0$ of solutions of . Still, we think of $\mu(t)$ as reflecting a limiting dynamics, and consider the value $\mu(0)$ as its initial state, in the space of invariant measures (that may indeed be partially supported on points, for instance, in case the load is affected by the dynamics, we may wish to compute it as well). We shall denote by $\mu(h,\nu)$ the value of the Young measure obtained at the time $h$, had $\nu$ been the initial state. Likewise, we denote by $P(h,\nu)$ the result of the computations at time $h$, had $\nu$ been the invariant measure to which the algorithm is applied. We place the assumptions on $\mu(h,\nu)$ and $P(h,\nu)$. After stating the assumption we comment on the reflection of them on the limit dynamics and the algorithm. Denote by $\rho(.,.)$ a distance, say the Prohorov metric, between probability measures. **Assumption 8.1.** There exists a function $\delta(h)$, continuous at $0$ with $\delta(0) = 0$, and a constant $\eta$ such that $$\begin{aligned} \label{eq:8.1} \rho(P(h,\nu), \mu(h,\nu))\leq\delta(h)h,\end{aligned}$$ and $$\begin{aligned} \label{eq:8.2} \rho(\mu(h,\nu_{1}),\mu(h,\nu_{2}))\leq (1+\eta h)\rho(\nu_{1},\nu_{2}).\end{aligned}$$ **Remark.** Both assumptions relate to some regularity of the limiting dynamics. The inequality (\[eq:8.2\]) reflects a Lipschitz property. It holds, for instance, in case $\mu(t)$ is a solution of an ordinary differential equations with Lipschitz right hand side. Recall that without a Lipschitz type condition it is not possible to get reasonable error estimates even in the ode framework. Condition (\[eq:8.1\]) reflects the accuracy of our algorithm on a time step of length $h$. For instance, If a finite set of observables determines the invariant measure, if the extrapolation through derivatives are uniform, if the mapping that maps the measures to the values of the observables is bi-Lipschitz, and if the dynamics $\mu(t)$ is Lipschitz, then (\[eq:8.1\]) holds. In concrete examples it may be possible to check these properties directly (we comment on that in the examples below). [**Theorem 8.2.**]{} Suppose the inequalities in Assumption 8.1 hold. Then $$\begin{aligned} \label{eq:8.3} \rho(P(kh),\mu(kh)) \leq \delta(h) \eta^{-1}(e^{\eta{T}}-1).\end{aligned}$$ [**Proof.**]{} Denote $E_{k} = \rho((P(kh),\mu(kh))$, namely, $E_{k}$ is the error accrued from $0$ to the time $kh$. Then $E_{0}= 0$. Since $P(kh) = P(h,P((k-1)h))$ and $\mu(kh) = \mu(h,\mu((k-1)h))$ it follows that $$\begin{aligned} \label{eq:8.4} E_{k} \leq \rho(P(h,P((k-1)h)),\mu(h,P((k-1)h))) +\nonumber\\ \rho(\mu(h,P((k-1)h)),\mu(h,\mu((k-1)))).\end{aligned}$$ From inequalities (\[eq:8.1\]) and (\[eq:8.2\]) we get $$\begin{aligned} \label{eq:8.5} E_{k} \leq \delta(h)h + (1 + \eta h)E_{k-1}.\end{aligned}$$ Spelling out the recursion we get $$\begin{aligned} \label{eq:8.6} E_{k} & \leq \delta(h)h(1+ (1+\eta h) + (1+\eta h)^{2} + ... + (1 + \eta h)^{k-1})\nonumber\\ &= \delta(h) \eta^{-1} ((1+ \eta h)^{k}-1).\end{aligned}$$ Since $hk \leq T$ the claimed inequality (\[eq:8.3\]) follows. **Remark.** The previous estimate imply that the error tends to zero as the step size $h$ tends to zero. Note, however, that the bigger the step size, the bigger the saving of computational time is. This interplay is common in computing in general. **Remark.** Needless to say, the previous estimate refers to the possible errors on the theoretical dynamics of measures. A different type of errors that occur relates to the numerics of representing and approximating the measures. The same issue arises in any numerical analysis estimates, but they are more prominent in our framework due to the nature of the dynamics. Computational Implementation {#impl_algo} ============================ We are given the initial conditions of the fine and the slow variables, $x(-\Delta) = x_0 $ and $l(-\Delta) = l_0 $. We aim to produce a good approximation, of the limit solution in the period $[0, T_0]$. Due to the lack of algorithmic specification for determining orthogonal observables, we concentrate on $H$-observables in this paper. In the implementation that follows, we get to the specifics of how the calculations are carried out. Recall that the $H$-observables are averages of the form $$\begin{aligned} \label{coarse_obs_impl} v(t)=\frac{1}{\Delta}\int_{t-\Delta}^{t} \int_{\mathbb{R}^n} m(x)\mu(s)(dx)ds.\end{aligned}$$ Namely, the $H$-observables are slow variables with a time-lag, i.e. the fine/microscopic system needs to have run on an interval of length $\Delta$ before the observable at time $t$ can be defined. *From here onwards, with some abuse of notation, whenever we refer to only a measure at some instant of time we mean the value of the Young measure of the fast dynamics at that time. When we want to refer to the Young measure, we mention it explicitly.* We will also refer to any component of the list of variables of the original dynamics as *fine* variables. We think of the calculations marching forward in the slow time-scale in discrete steps of size $h$ with $T_0 = nh$. Thus the variable $t$ below in the description of our algorithm takes values of $0h,1h,2h,\ldots,nh$. **Step 1: Calculate the rate of change of slow variable**\ We calculate the rate of change of the slow variable at time $t$ using the following: $$\begin{aligned} \label{eq:comp_impl_1} \frac{dv}{dt}(t)=\frac{1}{\Delta} \left( \int_{\mathbb{R}^n} m(x)\mu(t)(dx)-\int_{\mathbb{R}^n} m(x)\mu(t-\Delta)(dx)\right).\end{aligned}$$ Let us denote the term $\int_{\mathbb{R}^n} m(x)\mu(t)(dx)$ as $R_t^m$ and the term $\int_{\mathbb{R}^n} m(x)\mu(t-\Delta)(dx)$ as $R_{t-\Delta}^m$. The term $R_t^m$ is computed as $$\begin{aligned} \label{eq:comp_impl_R1} R_t^m = \frac {1}{N_t} \sum_{i=1}^{N_t} m(x_\epsilon(\sigma_i),l_\epsilon(\sigma_i)).\end{aligned}$$ The successive values ($x_\epsilon(\sigma_i)$, $l_\epsilon(\sigma_i)$) are obtained by running the fine system $$\label{eq:alg_fast_2} \begin{split} \frac{dx_\epsilon}{d\sigma} &= F(x_\epsilon,l_\epsilon) + \epsilon G(x_\epsilon,l_\epsilon)\\ \frac{dl_\epsilon}{d\sigma}& = \epsilon L(x_\epsilon,l_\epsilon), \end{split}$$ with initial condition $x_{guess}(\sigma = \frac{t}{\epsilon})$ and $l(\sigma=\frac{t}{\epsilon} )$. We discuss in Step 5 how we obtain $x_{guess}(\sigma)$. Here, $N_t$ is the number of increments taken for the value of $R_t^m$ to converge upto a specified value of tolerance. Also, $N_t$ is large enough such that the effect of the initial transient does not affect the value of $R_t^m$. Similarly, $R_{t-\Delta}^m$ is computed as: $$\begin{aligned} \label{eq:comp_impl_R2} R_{t-\Delta}^m = \frac {1}{N_{t-\Delta}} \sum_{i=1}^{N_{t-\Delta}} m\left(x_\epsilon(\sigma_i), l_\epsilon(\sigma_i)\right),\end{aligned}$$ where successive values $x_\epsilon(\sigma_i)$ are obtained by running the fine system (\[eq:alg\_fast\_2\]) with initial condition $x_{guess}(\sigma- \frac {\Delta} {\epsilon})$ and $l(\sigma- \frac {\Delta}{\epsilon})$. We discuss in Step 3 how we obtain $x_{guess}(\sigma- \frac {\Delta} {\epsilon})$. Here, $N_{t-\Delta}$ is the number of increments taken for the value of $R_{t-\Delta}^m$ to converge upto a specified value of tolerance. **Step 2: Find the value of slow variable**\ We use the extrapolation rule to obtain the predicted value of the slow variable at the time $t+h$: $$\begin{aligned} \label{eq:comp_impl_expl} v(t+h) = v(t) + \frac{dv}{dt}(t) \, h, \end{aligned}$$ where $\frac{dv}{dt} (t)$ is obtained from (\[eq:comp\_impl\_1\]). **Step 3: Determine the closest point projection**\ We assume that the closest-point projection of any fixed point in the fine state space, on the Young measure of the fine evolution, evolves slowly in any interval where the Young measure is continuous. We use this idea to define a guess $x_{guess}(t+h-\Delta)$, that is in the basin of attraction of $\mu(t+h-\Delta)$. The fixed point, denoted as $x^{arb}_{t-\Delta}$, is assumed to belong to the set of points, $x_\epsilon(\sigma_i)$ for which the value of $R_{t-\Delta}^m$ in (\[eq:comp\_impl\_R2\]) converged. Specifically, we make the choice of $x^{arb}_{t-\Delta}$ as $x_\epsilon(\sigma_{N_{t-\Delta}})$ where $x_\epsilon(\sigma_i)$ is defined in (\[eq:comp\_impl\_R2\]) and $N_{t-\Delta}$ is defined in the discussion following it. Next, we compute the closest point projection of this point (in the Euclidean norm) on the support of the measure at $t-h-\Delta$. This is done as follows. Define $x^{conv}_{t-h-\Delta}$ as the point $x_\epsilon(\sigma_{N_{t-h-\Delta}})$, where $x_\epsilon(\sigma_i)$ is defined in the discussion surrounding (\[eq:comp\_impl\_R2\]) with $\sigma$ replaced by $\sigma - \frac{h}{\epsilon}$, and $N_{t-h-\Delta}$ is the number of increments taken for the value of $R_{t-h-\Delta}^m$ to converge (the value of $x^{conv}_{t-h-\Delta}$ is typically stored in memory during calculations for the time $t-h-\Delta$). The fine system (\[eq:alg\_fast\_2\]) with $\sigma$ replaced by $\sigma - \frac{h}{\epsilon}$ is initiated from $x^{conv}_{t-h-\Delta}$, and we calculate the distance of successive points on this trajectory with respect to $x^{arb}_{t-\Delta}$ until a maximum number of increments have been executed. We set the maximum number of increments as $2\,N_{t-h-\Delta}$. The point(s) on this finite time trajectory that records the smallest distance from $x^{arb}_{t-\Delta}$ is defined as the closest point projection, $x^{cp}_{t-h-\Delta}$. Finally, the guess, $x_{guess}(t+h-\Delta)$ is given by $$x_{guess}(t+h-\Delta)=2\,x^{arb}_{t-\Delta} - x^{cp}_{t-h-\Delta},$$ for $t>0$ (see Remark associated with Step 3 in Sec. \[sec:theory\_alg\]). For $t = 0$, we set $$x_{guess}(h-\Delta)= x^{arb}_{0} + \frac { ( x^{arb}_{0} - x^{cp}_{-\Delta} ) } {\Delta} (h-\Delta),$$ where $x^{arb}_{.}$ and $x^{cp}_{.}$ are defined in the first and the second paragraph respectively in Step 3 above with the time given by the subscripts. This is because the computations start at $t=-\Delta$ and we do not have a measure at $t=-h-\Delta$ and hence cannot compute $x^{cp}_{-h-\Delta}$ to be able to use the above formula to obtain $x_{guess}(t+h-\Delta)$. Thus, the implicit assumption is that $x_{guess}(t+h-\Delta)$ is the closest-point projection of $x^{arb}_{t-\Delta}$ on the support of the measure $\mu(t+h-\Delta)$, and that there exists a function $x^{cp}(s)$ for $t-h-\Delta \leq s\leq t+h-\Delta$ that executes slow dynamics in the time interval if the measure does not jump within it. **Step 4: Accept the measure**\ We initiate the fine equation (\[eq:alg\_fast\_2\]) at $(x_{guess}(t+h-\Delta), l(t+h-\Delta))$ and run the equation from $\sigma+\frac{h}{\epsilon}-\frac{\Delta}{\epsilon}$ to $\sigma+\frac{h}{\epsilon}$ (recall $\sigma = \frac{t}{\epsilon}$). We say that there is a match in the value of a slow variable *if* the following equality holds (approximately): $$\label{eq:v(t+h)} v(t+h)=\frac{1}{N'}\sum_{i=1}^{N'} m\left(x_\epsilon(\sigma_i),l_\epsilon(\sigma_i)\right),$$ where $v(t + h)$ refers to the predicted value of the slow variable obtained from the extrapolation rule in Step 2 above. The successive values $(x_\epsilon(\sigma_i), l_\epsilon(\sigma_i))$ are obtained from system (\[eq:alg\_fast\_2\]). Here, $N'=\frac {\Delta}{\epsilon \, \Delta \sigma}$ where $\Delta \sigma$ is the fine time step. If there is a match in the value of the slow variable, we accept the measure which is generated, in principle, by running the fine equation with the guess $x_{guess}(t+h-\Delta)$ and move on to the next coarse increment. If not, we check if there is a jump in the measure. We say that there is a jump in the measure if the value of $R_{t+h}^m$ is significantly different from the value of $R_t^m$. This can be stated as: $$\begin{aligned} \left|\frac {R_{t+h}^m - R_t^m} {R_t^m}\right| \gg \frac{1}{N}\sum_n \left|\frac {R_{t-(n-1)h}^m - R_{t-nh}^m} {R_{t-nh}^m}\right|,\end{aligned}$$ where $n$ is such that there is no jump in the measure between $t-nh$ and $t-(n-1)h$ and $N$ is the maximal number of consecutive (integer) values of such $n$. If there is no jump in the measure, we try different values of $x_{guess}(t+h-\Delta)$ based on different values of $x^{arb}_{t-\Delta}$ and repeat Steps 3 and 4. If there is a jump in the measure, we declare $v(t+h)$ to be the the right-hand-side of (\[eq:v(t+h)\]). The rationale behind this decision is the assumption $x_{guess}(t+h-\Delta)$ lies in the basin of attraction of the measure at $t+h-\Delta$. [**Step 5: Obtain fine initial conditions for rate calculation**]{}\ Step 1 required the definition of $x_{guess}(t)$. We obtain it as $$x_{guess}(t)=x^{arb}_{t-\Delta} + \frac {\left( x^{arb}_{t-\Delta} - x^{cp}_{t-h} \right)}{\left(h-\Delta \right)} \, \Delta,$$ for $t>0$, and $x^{arb}_{.}$ and $x^{cp}_{.}$ are defined in the same way as in Step 3, but at different times given by the subscripts. For $t=0$, we obtain $x_{guess}(0)$, which is required to compute $R_{0}^m$, by running the fine equation (\[eq:alg\_fast\_2\]) from  $\sigma=-\frac{\Delta}{\epsilon}$ to $\sigma=0$. This is because the computations start at $t=-\Delta$ and we do not have a measure at $t=-h$ and hence cannot compute $x^{cp}_{-h}$ to be able to use the above formula to obtain $x_{guess}(t)$. Another possible way to obtain $x_{guess}(t)$ is using the same extrapolation rule used in Step 3 to obtain $x_{guess}(t+h-\Delta)$ which means we can also obtain $x_{guess}(t)$ as $x_{guess}(t)=x^{arb}_{t-h-\Delta}+\frac{(x^{arb}_{t-h-\Delta}-x^{cp}_{t-2h-\Delta})}{h}(h+\Delta)$. But we have not used it in the computational results that follow. We continue in this manner until an approximation of slow observables is computed on the entire time interval \[$0, T_{0}$\]. [**Discussion.**]{} The use of the guess for fine initial conditions to initiate the fine system to compute $R^m_{t+h}$ and $R^m_{t+h-\Delta}$ is an integral part of this implementation. This allows us to systematically use the coarse evolution equation (\[eq:comp\_impl\_1\]). This feature is a principal improvement over previous work [@tan2013coarse; @tan2014md]. We will refer to this scheme, which is a mixture of rigorous and heuristic arguments, as Practical Time Averaging (PTA) and we will refer to results from the scheme by the same name. Results obtained solely by running the complete system will be referred to as [*fine*]{} results, indicated by the superscript or subscript [*f*]{} when in a formula. Thus, if $v$ is a scalar slow variable, then we denote the slow variable value obtained using PTA scheme as $v^{PTA}$ while the slow variable value obtained by running the fine system alone is called $v^f$. The speedup, $S(\epsilon)$, in compute time between the $fine$ and $PTA$ calculations is presented in the results that follow in subsequent sections. This is defined to be the ratio of the time taken by the $fine$ calculations to that by the $PTA$ calculations for an entire simulation, say consisting of $n$ steps of size $h$ on the slow time-scale. Let $T^{cpu}_f (\epsilon)$ and $T^{cpu}_{PTA} (\epsilon)$ be the compute times to obtain the [*fine*]{} and $PTA$ results per jump on the slow time scale, respectively, for the specific value of $\epsilon$. The compute time to obtain the $PTA$ results for $n$ jumps in the slow time scale is $n T^{cpu}_{PTA}(\epsilon)$ which can be written as $$n T^{cpu}_{PTA}(\epsilon)=n T^{cpu}_{PTA,1}(\epsilon) + T^{cpu}_{PTA,2}(\epsilon),$$ where $T^{cpu}_{PTA,1}(\epsilon)$ is the compute time to perform the computations mentioned in Step 1 to Step 5 for every jump in the slow time scale. Since we cannot use the formula for $x_{guess}(t)$ mentioned in Step 5 to obtain $x_{guess}(0)$ and we have to run the fine equation from $\sigma=-\frac{\Delta}{\epsilon}$ to $\sigma=0$, an additional overhead is incurred in the compute time for the $PTA$ computations which we denote as $T^{cpu}_{PTA,2}(\epsilon)$. Thus $$S(\epsilon) = \frac{n T^{cpu}_f (\epsilon)}{n T^{cpu}_{PTA}(\epsilon)} \approx \frac{T^{cpu}_{f}(\epsilon)}{T^{cpu}_{PTA,1}(\epsilon)}.$$ for large $n$. Error in the PTA result is defined as: $$\begin{aligned} Error(\%)=\frac{v^{PTA}-v^f}{v^f} \times 100.\end{aligned}$$ We obtain $v^f$ as follows: [**Step 1:**]{} We run the fine system (\[eq:alg\_fast\_2\]) from $\sigma=-\frac{\Delta}{\epsilon}$ to $\sigma=\frac{T_0}{\epsilon}$ using initial conditions ($x_0$, $l_0$) to obtain ($x_\epsilon(\sigma_i),l_\epsilon(\sigma_i)$) where $\sigma_i = i \, \Delta \sigma$ and $i \in \mathbb Z_+$ and $i \leq \frac{T_0+\Delta}{\epsilon \, \Delta \sigma}$. [**Step 2:**]{} We calculate $v^f(t)$ using: $$\begin{aligned} \label{eq:v_fine} v^f(t)=\frac{1}{N'}\sum_{i=N^0(t)}^{N^0(t)+ N'} m\left(x_\epsilon(\sigma_i),l_\epsilon(\sigma_i)\right),\end{aligned}$$ where $N'=\frac {\Delta}{\epsilon \, \Delta \sigma}$ and $N^0(t)=\frac{t+\Delta}{\epsilon \, \Delta \sigma}$ where $\Delta \sigma$ is the fine time step. [**Remark.**]{} If we are aiming to understand the evolution of the slow variables in the slow time scale, we need to calculate them, which we do in Step 2. However, the time taken in computing the average of the state variables in Step 2 is much smaller compared to the time taken to run the fine system in Step 1. We will show this in the results sections that follow. [**Remark.**]{} All the examples computed in this paper employ $H$-observables. When, however, orthogonal observables are used, the time taken to compute their values using the PTA scheme ($T^{cpu}_{PTA}$) will not depend on the value of $\epsilon$. Example I: Rotating planes ========================== Consider the following four-dimensional system, where we denote by $x$ the vector $x = (x_{1},x_{2},x_{3},x_{4})$. $$\begin{aligned} \label{eq:9.1} \frac{dx}{dt} = \frac{F(x)}{\epsilon} + G(x),\end{aligned}$$ where: $$\begin{aligned} \label{eq:9.2} F(x) = ((1-\lvert x\rvert)x+\gamma(x))\end{aligned}$$ with $$\begin{aligned} \label{eq:9.3} \gamma(x) = (x_{3},x_{4},-x_{1},-x_{2}).\end{aligned}$$ The drift may be determined by an arbitrary function $G(x)$. For instance, if we let $$\begin{aligned} \label{eq:9.4} G(x) = (-x_{2}, x_{1}, 0, 0),\end{aligned}$$ then we should expect nicely rotating two-dimensional planes. A more complex drift may result in a more complex dynamics of the invariant measures, namely the two dimensional limit cycles. Discussion ---------- The right hand side of the fast equation has two components. The first drives each point $x$ which is not the origin[,]{} toward the sphere of radius 1. The second, $\gamma(x)$, is perpendicular to $x$. It is easy to see that the sphere of radius 1 is invariant under the fast equation. For any initial condition $x_0$ on the sphere of radius 1, the fast time equation is $$\begin{aligned} \dot{ {\bf x} } = \begin{pmatrix} 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ -1 & 0 & 0 & 0 \\ 0 &-1 & 0 & 0 \end{pmatrix} {\bf x} ~. \end{aligned} \label{ex1_rate}$$ It is possible to see that the solutions are periodic, each contained in a two dimensional subspace. An explicit solution (which we did not used in the computations) is $${\bf x} = cos \left( t \right) \begin{pmatrix} x_{0,1} \\ x_{0,2} \\ x_{0,3} \\ x_{0,4} \end{pmatrix} + sin \left( t \right) \begin{pmatrix} x_{0,3} \\ x_{0,4} \\ - x_{0,1} \\ -x_{0,2} \end{pmatrix}. \label{ex1_sol}$$ Thus, the solution at any point of time is a linear combination of ${\bf x}_0$ and $ { \gamma} ( {\bf x}_0 )$ and stays in the plane defined by them. Together with the previous observation we conclude that the limit occupational measure of the fast dynamics should exhibit oscillations in a two-dimensional subspace of the four-dimensional space. The two dimensional subspace itself is drifted by the drift $G(x)$. The role of the computations is then to follow the evolution of the oscillatory two dimensional limit dynamics. We should, of course, take advantage of the structure of the dynamics that was revealed in the previous paragraph. In particular, it follows that three observables of the form $r(\mu)$ given in (\[eq:4.6\]), with $e_{1}$, $e_{2}$ and $e_{3}$ being unit vectors in $\mathbb{R}^{4}$, determine the invariant measure. They are not orthogonal (it may be very difficult to find orthogonal observables in this example), hence we may use, for instance, the $H$-observables introduced in (\[eq:4.7\]). It is also clear that the circles that determine the invariant measures move smoothly with the planes. Hence employing observables that depend smoothly on the planes would imply that conditions (\[eq:8.1\]) and (\[eq:8.2\]) hold, validating the estimates of Theorem 8.2. Results ------- We chose the slow observables to be the averages over the limit cycles of the four rapidly oscillating variables and their squares since we want to know how they progress. We define the variables $w_i=x_i^2$ for $i=1,2,3$ and $4$. The slow variables are $x_1^f$, $x_2^f$, $x_3^f$, $x_4^f$ and $w_1^f $ , $w_2^f $ , $w_3^f$, $w_4^f $. The slow variable $x_1^f$ is given by (\[coarse\_obs\_impl\]) with $m(x)=x_1$. The slow variables $x_2^f$, $x_3^f$ and $x_4^f$ are defined similarly. The slow variable $w_1^f$ is given by (\[coarse\_obs\_impl\]) with $m(x)=w_1$. The slow variables $w_2^f$, $w_3^f$ and $w_4^f$ are defined similarly (we use the superscript $f$, that indicates the [*fine*]{} solution, since in order to compute these observables we need to solve the entire equation, though on a small interval). We refer to the $PTA$ variables as $x_1^{PTA}$, $x_2^{PTA}$, $x_3^{PTA}$, $x_4^{PTA}$ and $w_1^{PTA} $ , $w_2^{PTA} $ , $w_3^{PTA}$, $w_4^{PTA} $. A close look at the solution (\[ex1\_sol\]) reveals that the averages, on the limit cycles, of the fine variables, are all equal to zero, and we expect the numerical outcome to reflect that. The average of the squares of the fine variables evolve slowly in time. In our algorithm, non-trivial evolution of the slow variable does not play a role in tracking the evolution of the measure of the complete dynamics. Instead they are used only to accept the slow variable (and therefore, the measure) at any given discrete time as valid according to Step 4 of Section \[impl\_algo\]. It is the device of choosing the initial guess in Step 3 and Step 5 of Section \[impl\_algo\] that allows us to evolve the measure discretely in time. [0.4]{} {height="3.1in"}   [0.6]{} {height="3.1in"} Fig. \[fig:ex1\_phase\] shows the phase space diagram of $x_1$, $x_2$ and $x_3$. ![ *The rapidly oscillating solution of the full equation of $w_3$ is given by the plot marked $w_3$. The drift in the [*fine*]{} and $PTA$ values cannot be seen on the given scale. But the drift is visible in Fig. \[fig:ex1\_x3\_sq\]*. The $PTA$ and the [*fine*]{} results overlap.[]{data-label="fig:ex1_x3sq_fine"}](figures/ex1_x1.pdf){width="\linewidth"} ![ *The rapidly oscillating solution of the full equation of $w_3$ is given by the plot marked $w_3$. The drift in the [*fine*]{} and $PTA$ values cannot be seen on the given scale. But the drift is visible in Fig. \[fig:ex1\_x3\_sq\]*. The $PTA$ and the [*fine*]{} results overlap.[]{data-label="fig:ex1_x3sq_fine"}](figures/ex1_fine_x3sq.pdf){width="\linewidth"} Fig. \[fig:ex1\_x1\_fine\] shows the rapid oscillations of the rapidly oscillating variable $x_1$ and the evolution of the slow variable $x_1^f$. Fig. \[fig:ex1\_x3sq\_fine\] shows the rapid oscillations of $w_3$ and the evolution of the slow variable $w_3^f$. We find that $x_3$ and $x_4$ evolve exactly in a similar way as $x_1$ and $x_2$ respectively. We find from the results that $x_3^f$, $x_4^f$, $w_3^f$ and $w_4^f$ evolve exactly similarly as $x_1^f$, $x_2^f$, $w_1^f$ and $w_2^f$ respectively. ![*Evolution of $w_4^f$.* []{data-label="fig:ex1_x4_sq"}](figures/ex1_x3sq.pdf){width="\linewidth"} ![*Evolution of $w_4^f$.* []{data-label="fig:ex1_x4_sq"}](figures/ex1_x4sq.pdf){width="\linewidth"} The comparison between the [*fine*]{} and the PTA results of the slow variables $w_3^f$ and $w_4^f$ are shown in Fig. \[fig:ex1\_x3\_sq\] and Fig. \[fig:ex1\_x4\_sq\] (we have not shown the evolution of $w_1^f$ and $w_2^f$ since they evolve exactly similarly to $w_3^f$ and $w_4^f$ respectively). The error in the PTA results are shown in Fig. \[fig:ex1\_rel\_err\]. Since the values of $x_1^f$, $x_2^f$, $x_3^f$ and $x_4^f$ are very close to $0$, we have not provided the error in PTA results for these slow variables. [**Savings in computer time**]{} In Fig. \[fig:ex1\_cputime\], we see that as $\epsilon$ decreases, the compute time for the fine run increases very quickly while the compute time for the PTA run increases relatively slowly. The compute times correspond to simulations spanning $t=0.01$ to $t=0.02$ with $\Delta=0.001$. The speedup in compute time, $S$, obtained as a function of $\epsilon$, is given by the following polynomial: $$\begin{aligned} S(\epsilon)= 73.57 - 3.70 \times 10^{9} \, \epsilon + 6.76 \times 10^{16} \, \epsilon^2 - 3.74\times 10^{23} \, \epsilon^3.\end{aligned}$$ The function $S(\epsilon)$ is an interpolation of the computationally obtained data to a cubic polynomial. A more efficient calculation yielding higher speedup is to calculate the slow variable $v$ using Simpson’s rule instead of using , by employing the procedures outlined in section \[prb2:res\_num\_case1\], section \[prb2:res\_num\_case2\] and the associated *Remark* in Appendix \[prb2:res\_case2\]. In this problem, we took the datapoint of $\epsilon=10^{-8}$ and obtained $S(10^{-8}) = 43$. This speedup corresponds to an accuracy of $0.7\%$ error. However, as $\epsilon$ decreases and approaches zero, the asymptotic value of $S$ becomes 74. Example II: Vibrating springs {#vibrate} ============================= Consider the mass-spring system in Fig. \[fig:prb2\]. The governing system of equations for the system in dimensional time is given in Appendix \[prb2:equations\]. The system of equations posed in the slow time scale is $$\begin{aligned} \label{eq:10.3} \epsilon \frac{dx_{1}}{d t} &= T_f \, y_{1}\nonumber\\ \epsilon \frac{dy_{1}}{d t}&= - T_f \left(\frac{k_{1}}{m_{1}}(x_{1} - w_{1})-\frac{\eta}{m_{1}}(y_{2}-y_{1})\right)\nonumber\\ \epsilon \frac{dx_{2}}{d t}&= T_f \, y_{2} \nonumber\\ \epsilon \frac{dy_{2}}{d t}&= - T_f \left(\frac{k_{2}}{m_{2}}(x_{2}-w_{2}) + \frac{\eta}{m_{2}}(y_{2}-y_{1})\right)\nonumber\\ \frac{dw_{1}}{d t}&= T_s \, L_{1}(w_{1})\nonumber\\ \frac{dw_{2}}{d t}&= T_s \, L_{2}(w_{2})~.\end{aligned}$$ The derivation of (\[eq:10.3\]) from the system in dimensional time is given in Appendix \[prb2:equations\]. The small parameter $\epsilon$ arises from the ratio of the fast oscillation of the springs to the slow application of the load. A closed-form solution to (\[eq:10.3\]) can be computed, but in the present study the solution will be used just for verifying the computations, and will not be used in computations themselves. The closed-form solutions are presented in Appendices \[prb2:res\_case1\] and \[prb2:res\_case2\]. Discussion {#ex2:discussion} ---------- - For a fixed value of the slow dynamics, that is, for fixed positions $w_{1}$ and $w_{2}$ of the walls, the dynamics of the fine equation is as follow. If $\frac{k_{1}}{m_{1}} \neq {\frac{k_{2}}{m_{2}}}$, then all the energy is dissipated, and the trajectory converges to the origin (the reason behind this behavior is explained in the Remark of Appendix \[prb2:res\_case1\]). If the equality holds, only part of the energy possessed by the initial conditions is dissipated, and the trajectory converges to a periodic one (in rare cases it will be the origin), whose energy is determined by the initial condition (the reason behind this behavior is explained in Case 2.1 and Case 2.2 of Section \[prb2:res\_num\_case2\] and in the Remark of Appendix \[prb2:res\_case2\]). The computational challenge is when fast oscillations persist. Then the limiting periodic solution determines an invariant measure for the fast flow. When the walls move, slowly, the limit invariant measure moves as well. The computations should detect this movement. However, if the walls move very slowly, there is a possibility that in the limit the energy does not change at all as the walls move. Notice that the invariant measure is not determined by the position of the walls, and additional slow observables should be incorporated. A possible candidate is the total energy stored in the invariant measure. Since the total energy is constant on the limit cycle, it forms an orthogonal observable as described in section 4. Its extrapolation rule is given by (\[eq:4.1\]). In order to apply (\[eq:4.1\]) one has to derive the effect of the movement of the walls on the observable, namely, on the total energy. Two other observables could be the average kinetic energy and the average potential energy on the invariant measure. In both cases, the form of $H$-observables should be employed, as it is not clear how to come up with an extrapolation rules for these observables. It is clear that with the three observables just mentioned, if the two forcing elements $L_{1}$ and $L_{2}$ in are Lipschitz, then conditions (\[eq:8.1\]) and (\[eq:8.2\]) are satisfied, and, consequently, the conclusion of Theorem 8.2 holds. - We define *kinetic energy* ($K$), *potential energy* ($U$) and *reaction force* on the right wall ($R_2$) as: $$\begin{aligned} \label{ex2_statefn} K(\sigma)&=\frac{1}{2} \left(m_1 \, {y_{1,\epsilon}(\sigma)}^2 + m_2 \, {y_{2,\epsilon}(\sigma)}^2\right) \nonumber\\ U(\sigma)&=\frac{1}{2} k_1 {(x_{1,\epsilon}(\sigma)-w_{1,\epsilon}(\sigma))}^2 + \frac{1}{2} k_2 {(x_{2,\epsilon}(\sigma)-w_{2,\epsilon}(\sigma))}^2 \\ R_2(\sigma) &= -k_2\left(x_{2,\epsilon}(\sigma) - w_{2,\epsilon}(\sigma)\right)\nonumber.\end{aligned}$$ - The H-observables that we obtained in this example are the *average kinetic energy* ($K^f$), *average potential energy* ($U^f$) and *average reaction force* on the right wall ($R_2^f$) which are calculated as: $$\begin{aligned} \label{ex2_hobs} {K^f}(t)&=\frac{1}{N'} \sum_{i=1}^{N'} K(\sigma_i) \nonumber\\ {U^f}(t)&=\frac{1}{N'} \sum_{i=1}^{N'} U(\sigma_i) \\ {R_2^f}(t)&=\frac{1}{N'} \sum_{i=1}^{N'} R_2(\sigma_i), \nonumber \end{aligned}$$ where $N'$ is defined in the discussion following (\[eq:v(t+h)\]) and successive values $x_{1,\epsilon}(\sigma_i)$, $x_{2,\epsilon}(\sigma_i)$, $y_{1,\epsilon}(\sigma_i)$ and $y_{2,\epsilon}(\sigma_i)$ are obtained by solving the fine system associated with (see in Appendix \[prb2:equations\]) with appropriate initial conditions which is discussed in detail in Step 3 of Section \[impl\_algo\]. The computations are done when $L_1(w_1) = 0$ and $L_2(w_2) = c_2$. - To integrate the fine system (\[eq:10.4\]), we use a modification of the velocity Verlet integration scheme to account for damping (given in [@Sandvik_2016_py502]). This is done so that the energy of the system does not diverge in time due to energy errors of the numerical method. - As we will show in Section \[prb2:res\_num\_case1\] (where we show results for the case corresponding to the condition $\frac{k_1}{m_1} \neq \frac{k_2}{m_2}$ which we call Case 1) and Section \[prb2:res\_num\_case2\] (where we show results for the case corresponding to the condition $\frac{k_1}{m_1} = \frac{k_2}{m_2}$ which we call Case 2) respectively, in Case 1, the fine evolution converges to a singleton (in the case without forcing) while in Case 2, the fine evolution generically converges to a limit set that is not a singleton (which will be shown in Case 2.2 in Section \[prb2:res\_num\_case2\]), which shows the distinction between the two cases. This has significant impact on the results of average kinetic and potential energy. Our computational scheme requires no *a-priori* knowledge of these important distinctions and predicts the correct approximations of the limit solution in all of the cases considered. - For the sake of comparison with our computational approximations, in Appendix C we provide solutions to our system corresponding to the Tikhonov framework [@Tikhonov_1985_DE] and the quasi-static assumption commonly made in solid mechanics for mechanical systems forced at small loading rates. We show that the quasi-static assumption does not apply for this problem. The Tikhonov framework applies in some situations and our computation results are consistent with these conclusions. As a cautionary note involving limit solutions (even when valid), we note that evaluating nonlinear functions like potential and kinetic energy on the weak limit solutions as a reflection of the limit of potential and kinetic energy along sequences of solutions of as $\epsilon \to 0$ (or equivalently $T_s \to \infty$) does not make sense in general, especially when oscillations persist in the limit. Indeed, we observe this for all results in Case 2. - All results shown in this section are obtained from numerical calculations, with no reference to the closed-form solutions. The closed-form solution for Case 1 is derived in Appendix \[prb2:res\_case1\] while the closed-form solution for Case 2 is derived in Appendix \[prb2:res\_case2\]. Results - Case 1 : $\left(\frac{k_1}{m_1} \neq \frac{k_2}{m_2}\right)$ {#prb2:res_num_case1} ---------------------------------------------------------------------- We do not use the explicit limit dynamics displayed in Appendix \[prb2:res\_case1\]. Rather, we proceed with the computations employing the kinetic and potential energies as our $H$-observables. All simulation parameters are grouped in Table \[tab:simulation\_details\_uneql\]. The total physical time over which the simulation runs is $T_0 T_s$, where $T_0$ is defined in Section \[sec:theory\_alg\] (in all computed problems here, we have chosen $T_0=1$). The PTA computations done in this section are with the load fixed while calculating $R^m_t$ and $R^m_{t-\Delta}$ using and respectively (by setting $\frac{dl_\epsilon}{d\sigma}=0$ in ). The slow variable value ($v(t+h)$ in ) is calculated using Simpson’s rule as described in *Remark* in Appendix \[prb2:res\_case2\]. The $PTA$ results and the [*closed-form*]{} results (denoted by “$cf$" in the superscript) match for all values of $t$ ($K^{PTA}= 10^{-10} \approx 0 = K^{cf} $, $U^{PTA} = 10^{-10} \approx 0 = U^{cf}$ and $R_2^{PTA} = 10^{-5} \approx 0 = R_2^{cf}$ - note that following the discussion around in Appendix \[prb2:res\_case2\], all the results presented here are non-dimensionalized). In this case, the results from the Tikhonov framework match with our computational approximations. This is because after the initial transient dies out, the whole system displays slow behavior in this particular case. However, the solution under the quasi-static approximation does not match our computational results (even though the loading rate is small), and we indicate the reason for its failure in Appendix \[prb2:res\_case1\]. Name Physical definition Values ------------ ------------------------------------ ----------------------------------------------- $k_1$ Stiffness of left spring $ 10^7 \,\mathit{N/m}$ $k_2$ Stiffness of right spring $ 10^7 \,\mathit{N/m}$ $m_1$ Left mass $1\,\mathit{kg}$ $m_2$ Right mass $2\,\mathit{kg}$ $\eta$ Damping coefficient of dashpot $ 5 \times 10^3 \,\mathit{N \, s/m}$ $c_2$ Velocity of right wall $10^{-6}$ = $\frac{0.01}{10^4}\,\mathit{m/s}$ $ h $ Jump size in slow time scale $ 0.25 $ $ \Delta $ Parameter used in rate calculation $ 0.05 $ : Simulation parameters.[]{data-label="tab:simulation_details_uneql"} ### Power Balance {#ex2_power_balance} It can be show from that $$\begin{aligned} \label{power_balance} {1 \over T_s} \frac{d}{dt} \left( \frac{1}{ 2} m_1 y_1^2 + \frac{1}{ 2} m_2 y_2^2 + \frac{1}{ 2} m_{w_1} {v_{w_1}}^2 + \frac{1}{ 2} m_{w_2} {v_{w_2}}^2 \right)& \nonumber \\ + {1 \over T_s} \frac{d}{dt} \left( \frac{1}{ 2} k_1 {(x_1 - w_1)}^2 + \frac{1}{ 2} k_2 {(x_2 - w_2)}^2 \right) &= R_1 v_{w_1} + R_2 v_{w_2} - \eta {(y_2 - y_1)}^2,\end{aligned}$$ where $v_{w_1}= {1 \over T_s} \frac{d w_1}{dt}$ and $v_{w_2}= {1 \over T_s} \frac{d w_2}{dt}$. Equation is simply the statement that at any instant of time the rate of change of kinetic energy and potential energy is the external power supplied through the motion of the walls less the power dissipated as viscous dissipation. This means that the sum of the kinetic and potential energy of the system, is equal to the sum of the initial kinetic and potential energy, plus the integral of the external power supplied to the system, minus the viscous dissipation. The [*fine*]{} solution indicates that, for $c_2$ small, the dashpot kills all the initial potential and kinetic energy supplied to the system. The two springs get stretched based on the value of $c_2$. The stretches remain fixed for large times on the fast time scale and the mass $m_2$ and the right wall move with the same velocity with mass $m_1$ remaining fixed. Thus the right spring moves like a rigid body. Based on this argument and from , the viscous, dissipated power in the system at large fast times is $\eta {c_2}^2$, which is equal to the external power provided to the system (noting that even though mass $m_2$ moves for large times, it does so with uniform velocity in this problem resulting in no contribution to the rate of change of kinetic energy of the system). [**Savings in Computer time**]{} Fig.\[fig:ex2\_cputime\_uneql\] shows the comparison between the time taken by the fine and the PTA runs for simulations spanning $t=0.25$ to $t=0.5$ with $\Delta=0.05$. The speedup in compute time is given by the following polynomial: $$\begin{aligned} S(\epsilon) = 2.28 \times 10^3 - 1.29 \times 10^{10} \, \epsilon + 6.46\times 10^{15} \, \epsilon^2 + 2.94 \times 10^{20} \, \epsilon^3. \end{aligned}$$ The function $S(\epsilon)$ is an interpolation of the computationally obtained data to a cubic polynomial. We used the datapoint of $\epsilon=1.98 \times 10^{-7}$ and obtained $S(1.98 \times 10^{-7})=231$. Results obtained in this case have accuracy of $0.0000\%$ error. As $\epsilon$ is decreased and approaches zero, the asymptotic value of $S$ is $2.28\times10^3$. Results - Case 2: $\left(\frac{k_1}{m_1} = \frac{k_2}{m_2} \right)$ {#prb2:res_num_case2} -------------------------------------------------------------------- As already noted, in this case the quasi-static approach is not valid. The closed-form solution to this case is displayed in Appendix \[prb2:res\_case2\] (but it is not used in the computations). Recall that the computations are carried out when $L_1(w_1) = 0$ and $L_2(w_2) = c_2$. The PTA computations done in this section are with the load fixed while calculating $R^m_t$ and $R^m_{t-\Delta}$ using and respectively (by setting $\frac{dl_\epsilon}{d\sigma}=0$ in ). The slow variable value ($v(t+h)$ in ) is calculated using Simpson’s rule as described in *Remark* in Appendix \[prb2:res\_case2\]. All results in this section are non-dimensionalized following the discussion around in Appendix \[prb2:res\_case2\]. The following cases arise: - [**Case 2.1.**]{} When $c_2=0 $ and the initial condition does not have a component on the modes describing the dashpot being undeformed ($x_1=x_2$ and $y_1=y_2$), then the solution will go to rest. For example, the initial conditions ${x_1}^0=1.0$, ${x_2}^0=-0.5$ and ${y_1}^0 = {y_2}^0 = 0.0$ makes the solution (\[sol\]) of Appendix \[prb2:res\_case2\] go to rest ($\kappa_3=\kappa_4=0$ in (\[eq:coeffs\]) of Appendix \[prb2:res\_case2\]). The simulation results agree with the *closed-form* results and go to zero. - [**Case 2.2.**]{} When $c_2=0 $ and the initial condition has a component on the modes describing the dashpot being undeformed, then in the fast time limit the solution shows periodic oscillations whose energy is determined by the initial conditions. This happens, of course, for almost all initial conditions. One such initial condition is ${x_1}^0 = 0.5 $, ${x_2}^0 = -0.1 $ and ${y_1}^0={y_2}^0 = 0 $ ($\kappa_3=0$ but $\kappa_4=-0.4472$ in (\[eq:coeffs\]) of Appendix \[prb2:res\_case2\]). The simulation results agree with the *closed-form* results. This is in contrast with Case 1 where it is impossible to find initial conditions for which the solution shows periodic oscillations. ![ *Case 2.2 - Error in $K^{PTA}$ and $U^{PTA}$.*[]{data-label="fig:ex2_err_c2eql0"}](figures/ex2_energy_case2_2.pdf){width="\linewidth"} ![ *Case 2.2 - Error in $K^{PTA}$ and $U^{PTA}$.*[]{data-label="fig:ex2_err_c2eql0"}](figures/ex2_err_case2_2.pdf){width="\linewidth"} In Fig. \[fig:ex2\_pta\_c2eql0\], we see that the PTA results are very close to the [*closed-form*]{} results. The error in PTA results are presented in Fig. \[fig:ex2\_err\_c2eql0\]. Oscillations persist in the limit and the potential and kinetic energies computed based on the Tikhonov framework as well as the quasi-static solution (\[ex2\_uneql\_quasistatic\]) derived in Appendix \[prb2:res\_case1\] are not expected to, and do not, yield correct answers. - [**Case 2.3.**]{} When $c_2 \neq 0$ and the initial condition does not have a component on the modes describing the dashpot being undeformed, then the solution on the fast time scale for large values of $\sigma$ does not depend on the initial condition. One such initial condition is ${x_1}^0=1.0$, ${x_2}^0=-0.5$ and ${y_1}^0 = 0.0$ and $ {y_2}^0= 10^{-4} $. The *closed-form average kinetic energy* ($K^{cf}$) and *closed-form average potential energy* ($U^{cf}$) do not depend on the magnitude of the initial conditions in this case. - [**Case 2.4.**]{} The initial condition has a component on the modes describing the dashpot being undeformed. But when $c_2 \neq 0 $, the dashpot gets deformed due to the translation of the mass $m_2$. The *closed-form average kinetic energy* ($K^{cf}$) and *closed-form average potential energy* ($U^{cf}$) depend on the initial conditions. ![ *Case 2.4 - Error in $K^{PTA}$ and $U^{PTA}$*. []{data-label="fig:ex2_err"}](figures/ex2_energy_case2_4.pdf){width="\linewidth"} ![ *Case 2.4 - Error in $K^{PTA}$ and $U^{PTA}$*. []{data-label="fig:ex2_err"}](figures/ex2_err_case2_4.pdf){width="\linewidth"} In Fig. \[fig:ex2\_pta\], we see that the PTA results are very close to the [*closed-form*]{} results. The errors in PTA results are presented in Fig. \[fig:ex2\_err\]. - The comparison between $R_2^{PTA}$ and $R_2^{cf}$ for Case 2.1 to 2.4 is shown in Fig.\[fig:ex2\_rf\_pta\]. The *closed-form* result goes to zero and the PTA result becomes very small. The displacement of the right wall ($w_2$) can be expressed as a function of time for Case 2.3 and Case 2.4 (given by in Appendix \[prb2:res\_case2\]). - Again, oscillations persist in the limit, and the Tikhonov framework and the quasi-static approximation (see Appendix \[prb2:res\_case1\]) do not work in this case. - The results do not change if we decrease the value of $\epsilon$. However, the speedup changes as will be shown in *Savings in Computer time* later in this section. We used the same simulation parameters as in Case 1 ( shown in Table \[tab:simulation\_details\_uneql\] ) but with $k_2=2 \times 10^7 N/m$ so that $ \frac{k_1}{m_1} = \frac{k_2}{m_2} $. Let us assume that the strain rate is $10^{-4} s^{-1} $. Then the slow time period, $T_s = {1 \over{\dot{\bar{\epsilon} } } }= 10000 s$. The fast time period is obtained as the period of fast oscillations of the spring, given by, $T_f = 2 \pi \sqrt{ \frac { m_1} { k_1} }= 0.002 s$. Thus, we find $\epsilon = \frac { T_f } {T_s} = 1.98 \times 10^{-7} $. While running the PTA code with $\epsilon = 0.002$, we have seen that the PTA scheme is not able to give accurate results and it breaks down. [**Power Balance**]{}\ The input power supplied to the system is ${1 \over T_s}R_2 \frac{d w_2}{dt} $. Since $R_2 = k_2(w_2 - x_2)$ and $\frac{d w_2}{dt}=T_s \, c_2$, the average value of input power at time $t$ is ${1 \over \Delta} \int_t^{t+\Delta} k_2 (w_2 - x_2) c_2 \, dt'$. From the results in of Appendix \[prb2:res\_case2\] and noting that the average of oscillatory terms over time $\Delta$ is approximately 0, we see that ${1 \over \Delta} \int_t^{t+\Delta}{(w_2-x_2)}\,dt'={1 \over \Delta} \int_t^{t+\Delta} \{ c_2 T_s t' - (c_2 T_s t' - \frac {\eta c_2}{k_2}) \} \, dt' = \frac {\eta c_2}{k_2}$. Hence average input power supplied is $\eta c_2^2$. The average dissipation at time $t$ is $ {1 \over \Delta} \int_t^{t+\Delta} \frac {\eta}{T_s^2} {(\frac{d x_2}{dt'}-\frac{d x_1}{dt'}) }^2 \, dt' $. Using the result from of Appendix \[prb2:res\_case2\], and using the same argument  that the average of oscillatory terms over time $\Delta$ is approximately 0, we can say that the dissipation is $\eta c_2^2$. Thus, the average input power supplied to the system is equal to the dissipation in the damper. A part of the input power also goes into the translation of the mass $m_2$. But its value is very small compared to the total kinetic energy of the system. [**Savings in Computer time**]{}\ It takes the PTA run around 62 seconds to compute the calculations that start at slow time which is a multiple of $h$ (steps 1 through 5 in section \[impl\_algo\] ). It takes the fine theory run around 8314 seconds to evolve the fine equation starting at slow time $nh$ to slow time $(n+1)h$, where $n$ is a positive integer. Thus, we could achieve a speedup of 134. We expect that the speedup will increase if we decrease the value of $\epsilon$. Fig.\[fig:ex2\_cputime\_eql\] shows the comparison between the time taken by the fine and the PTA runs for simulations spanning $t=0.25$ to $t=0.5$ with $\Delta=0.05$. The speedup in compute time is given by the following polynomial: $$\begin{aligned} S(\epsilon) = 164.34 - 1.62\times 10^{8} \, \epsilon + 5.23 \times 10^{13} \, \epsilon^2 -2.25 \times 10^{18} \epsilon^3.\end{aligned}$$ The function $S(\epsilon)$ is an interpolation of the computationally obtained data to a cubic polynomial. We used the datapoint of $\epsilon=1.98 \times 10^{-7}$ and obtained $S(1.98 \times 10^{-7})=134$. This speedup corresponds to an accuracy of $0.026\%$ error. As $\epsilon$ is decreased and approaches zero, the asymptotic value of $S$ is 164. Example III: Relaxation oscillations of oscillators =================================================== This is a variation of the classical relaxation oscillation example(see, e.g., [@Artstein_2002_perturbed]). Consider the four-dimensional system $$\begin{aligned} \label{eq:11.1} \frac{dx}{dt}&= z\nonumber\\ \frac{dy}{dt}&= \frac{1}{\epsilon}(-x+y-y^{3})\\ \frac{dz}{dt}&= \frac{1}{\epsilon}(w+(z - y)(\frac{1}{8}-w^{2}-(z-y)^{2}))\nonumber\\ \frac{dw}{dt}&= \frac{1}{\epsilon}(-(z-y)+w(\frac{1}{8}-w^{2}-(z-y)^{2}))\nonumber.\end{aligned}$$ Notice that the $(z,w)$ coordinates oscillate around the point $(y,0)$ (in the $(z,w)$-space), with oscillations that converge to a circular limit cycle of radius ${1 \over \sqrt{8} }$. The coordinates $(x,y)$ follow the classical relaxation oscillations pattern (for the fun of it, we replaced $y$ in the slow equation by $z$, whose average in the limit is $y$). In particular, the limit dynamics of the $y$-coordinate moves slowly along the stable branches of the curve $0 = -x + y - y^{3}$, with discontinuities at $x = -\frac{2}{3 \sqrt{3}}$ and $x = \frac{2}{3 \sqrt{3}}$. In turn, these discontinuities carry with them discontinuities of the oscillations in the $(z,w)$ coordinates. The goal of the computation is to follow the limit behavior, including the discontinuities of the oscillations. Discussion {#discussion-1} ---------- The slow dynamics, or the load, in the example is the $x$-variable. Its value does not determine the limit invariant measure in the fast dynamics, which comprises a point $y$ and a limit circle in the $(z,w)$-coordinates. A slow observable that will determine the limit invariant measure is the $y$-coordinate. In particular, conditions (\[eq:8.1\]) an (\[eq:8.2\]) hold except at points of discontinuity, and so does the conclusion of Theorem 8.2. Notice, however, that this observable does go through periodic discontinuities. Results ------- We see in Fig. \[fig:ex3\_fast\] that the $y$-coordinate moves slowly along the stable branches of the curve $0 = -x + y - y^{3}$ which is evident from the high density of points in these branches of the curve as can be seen in Fig. \[fig:ex3\_fast\]. There are also two discontinuities at $x = -\frac{2}{3 \sqrt{3}}$ and $x = \frac{2}{3 \sqrt{3}}$. The pair $(z,w)$ oscillates around $(y,0)$ in circular limit cycle of radius ${ 1 \over \sqrt{8} }$. In Fig. \[fig:ex3\_pta\], we see that the average of $z$ and $y$ which are given by the $y$-coordinate in the plot, are the same which acts as a verification that our scheme works correctly. Also, average of $w$ is 0 as expected. Since there is a jump in the evolution of the measure at the discontinuities (of the Young measure), the observable value obtained using extrapolation rule is not able to follow this jump. However, the observable values obtained using the guess for fine initial conditions at the next jump could follow the discontinuity. This is the principal computational demonstration of this example. ![ [*PTA result. The portion with the arrows correspond to very rapid evolution on the slow time scale*]{}.[]{data-label="fig:ex3_pta"}](figures/ex3_fine.pdf){width="\linewidth"} ![ [*PTA result. The portion with the arrows correspond to very rapid evolution on the slow time scale*]{}.[]{data-label="fig:ex3_pta"}](figures/ex3_pta1.pdf){width="\linewidth"} [0.5]{} {height="3.1in"}   [0.5]{} {height="3.1in"} Fig. \[fig:pta\_working\] shows the working of the PTA scheme when there is a discontinuity in the Young measure. We obtain the initial guess at time $t+h-\Delta$ by extrapolating the closest point projection at time $t-h-\Delta$ of a point on the measure at time $t-\Delta$. The details of the procedure are mentioned in Step 3 of section \[impl\_algo\]. When there is a discontinuity in the Young measure, the results of the slow observables obtained using coarse evolution (Point 6) is unable to follow the discontinuity. But when the fine run is initiated at the initial guess at time $t+h-\Delta$ which is given by Point 3 in the figure, the PTA scheme is able to follow the jump in the measure and we obtain the correct slow observable values (Point 5) which is very close to the slow observable value obtained from the fine run (Point 4). **Savings in computer time.** Fig.\[fig:ex3\_cputime\] shows the comparison between the time taken by the fine and the PTA runs for simulations spanning $t=0.2$ to $t = 0.4$ with $\Delta = 0.01$. The speedup in compute time as a function of $\epsilon$ for $\Delta=0.01$, is given by the following polynomial: $$\begin{aligned} S (\epsilon) = 1.29 \times 10^3 - 1.19 \times 10^{12} \, \epsilon + 1.17 \times 10^{20} \, \epsilon ^2 - 1.05 \times 10^{27} \, \epsilon^3.\end{aligned}$$ The function $S (\epsilon)$ is an interpolation of the computationally obtained data to a cubic polynomial. A more efficient calculation yielding higher speedup is to calculate the slow variable $v$ using Simpson’s rule instead of using , by employing the procedures outlined in section \[prb2:res\_num\_case1\], section \[prb2:res\_num\_case2\] and the associated *Remark* in Appendix \[prb2:res\_case2\]. We used the datapoint of $\epsilon=10^{-10}$ for this problem and obtained $S(10^{-10})=1.17 \times 10^3$. This speedup corresponds to an accuracy of $4.33\%$ error. As we further decrease $\epsilon$ and it approaches zero, the asymptotic value of $S$ becomes $1.29 \times 10^3$. [**Remark.**]{} As a practical matter, it seems advantageous to set $\epsilon = 0$ in for the computations of $R^m_{t-\Delta}$ in and $R^m_t$ in . Related to this, when the value of $\epsilon$ is decreased, calculating the slow variable value ($v(t+h)$ in ) using Simpson’s rule as described in *Remark* in Appendix \[prb2:res\_case2\] reduces $T^{cpu}_{PTA}$ considerably and improves the speedup $S(\epsilon)$. Concluding remarks ================== The focus of this paper has been the precise definition and demonstration of a computational tool to probe slow time-scale behavior of rapidly evolving microscopic dynamics, whether oscillatory or exponentially decaying to a manifold of slow variables, or containing both behaviors. A prime novelty of our approach is in the introduction of a general family of observables ($H$-observables) that is universally available, and a practical computational scheme that covers cases where the invariant measures may not be uniquely determined by the slow variables in play, and one that allows the tracking of slow dynamics even at points of discontinuity of the Young measure. We have solved three model problems that nevertheless contain most of the complications of averaging complex multiscale temporal dynamics. It can be hoped that the developed tool is of substantial generality for attacking real-world practical problems related to understanding and engineering complex microscopic dynamics in relatively simpler terms. Verlet Integration {#app:verlet} =================== The implementation of Verlet scheme that we used in Example II to integrate the fine equation (\[eq:10.4\]) is as follows [@Sandvik_2016_py502]: $$\begin{aligned} \label{eq:verlet} x_1(\sigma + \Delta \sigma) &= x_1(\sigma) + \Delta \sigma \, y_1(\sigma) + \frac {1} {2} {\Delta \sigma}^2 \, \frac {d y_1}{d \sigma} \nonumber\\ x_2(\sigma + \Delta \sigma) &= x_2(\sigma) + \Delta \sigma \, y_2(\sigma) + \frac {1} {2} {\Delta \sigma}^2 \, \frac {d y_2}{d \sigma} \nonumber\\ \hat{y_1}(\sigma + \Delta \sigma) &= y_1(\sigma) + \frac {1}{2} \Delta \sigma \left( \frac {d y_1}{d \sigma} + a_1 \left\{x_1(\sigma + \Delta \sigma), y_1(\sigma) + \Delta \sigma \frac {d y_1}{d \sigma}, y_2(\sigma) + \Delta \sigma \frac {d y_2}{d \sigma}, w_1(\sigma) \right\} \right) \nonumber\\ \hat{y_2}(\sigma + \Delta \sigma) &= y_2(\sigma) + \frac {1}{2} \Delta \sigma \left( \frac {d y_2}{d \sigma} + a_2 \left\{x_2(\sigma + \Delta \sigma), y_1(\sigma) + \Delta \sigma \frac {d y_1}{d \sigma}, y_2(\sigma) + \Delta \sigma \frac {d y_2}{d \sigma}, w_2(\sigma) \right\} \right) \nonumber\\ y_1(\sigma + \Delta \sigma) &= y_1(\sigma) + \frac{1}{2} \Delta \sigma \left( \frac{d y_1}{d \sigma} + a_1\left\{x_1(\sigma+\Delta \sigma), \hat{y_1} (\sigma + \Delta \sigma), \hat{y_2}(\sigma + \Delta \sigma), w_1(\sigma) \right\} \right) \nonumber\\ y_2(\sigma + \Delta \sigma) &= y_2(\sigma) + \frac{1}{2} \Delta \sigma \left( \frac{d y_2}{d \sigma} + a_2 \left\{x_2(\sigma+\Delta \sigma), \hat{y_1} (\sigma + \Delta \sigma), \hat{y_2}(\sigma + \Delta \sigma), w_2(\sigma)\right\} \right). \end{aligned}$$ Here $a_1(x_1,y_1,y_2,w_1)=\frac{d y_1}{d\sigma}$ and $a_2(x_2,y_1,y_2,w_2)=\frac{d y_2}{d\sigma}$ where $\frac{d y_1}{d\sigma}$ and $\frac{d y_2}{d\sigma}$ are given by (\[eq:10.4\]). Example II: Derivation of system of equations {#prb2:equations} ============================================= Two massless springs and masses $m_{1}$ and $m_{2}$ are connected through a dashpot damper, and each is attached to two bars, or walls, that may move very slowly, compared to possible oscillations of the springs. The system is described in Fig. \[fig:prb2\] of Section \[vibrate\]. Let the springs constants be $k_{1}$ and $k_{2}$ respectively, and let $\eta$ be the dashpot constant. Denote by $x_{1}$ and $x_{2}$, and by $w_{1}$ and $w_{2}$, the displacements from equilibrium positions of the masses of the springs and the positions of the two walls. We agree here that all positive displacements are toward the right. We think of the movement of the two walls as being external to the system, determined by a “slow" differential equation. The movement of the springs, however, will be “fast", which we model as singularly perturbed. Let $m_{w_1}$ and $m_{w_2}$ be the masses of the left and the right walls respectively. The displacements $x_1$, $x_2$, $w_1$ and $w_2$ have physical dimensions of length. The spring constants $k_1$ and $k_2$ have physical dimensions of force per unit length while $m_1$ and $m_2$ have physical dimensions of mass. In view of the assumptions just made, a general form of the dynamics of the system is given by following set of equations: $$\begin{aligned} \label{eq:10.1} m_1 \frac{d^2 x_1}{d {t^*}^2} &= -k_1 (x_1 - w_1) + \eta \left( \frac{d x_2}{d t^*} - \frac{d x_1}{d t^*} \right) \nonumber\\ m_2 \frac{d^2 x_2}{d {t^*}^2} &= -k_2 (x_2 - w_2) - \eta \left( \frac{d x_2}{d t^*} - \frac{d x_1}{d t^*} \right) \\ m_{w_1} \frac{d^2 w_1}{d {t^*}^2} &= k_1 (x_1 - w_1) + R_1 \nonumber\\ m_{w_2} \frac{d^2 w_2}{d {t^*}^2} &= k_2 (x_2 - w_2) + R_2, \nonumber\end{aligned}$$ where $R_1$ and $R_2$ incorporate the reaction forces on the left and right walls respectively due to their prescribed motion. We agree that forces acting toward the right are being considered positive. We make the assumption that $m_{w_1}=m_{w_2}=0$. The time scale $t^*$ is a time scale with physical dimensions of time. In our calculations, however, we address a simplified version, of first order equations, that can be obtained from the previous set by appropriately specifying what the forces on the system are: $$\begin{aligned} \label{eq:10.2} \frac{dx_{1}}{dt^*} &= y_{1} \nonumber\\ \frac{dy_{1}}{dt^*}&= -\frac{k_{1}}{m_{1}}(x_{1}-w_{1}) + \frac{\eta}{m_{1}}(y_{2}-y_{1})\nonumber\\ \frac{dx_{2}}{dt^*}&= y_{2}\\ \frac{dy_{2}}{dt^*}&= -\frac{k_{2}}{m_{2}}(x_{2}-w_{2})-\frac{\eta}{m_{2}}(y_{2}-y_{1})\nonumber\\ \frac{dw_{1}}{dt^*}&=L_{1}(w_{1})\nonumber\\ \frac{dw_{2}}{dt^*}&=L_{2}(w_{2})\nonumber.\end{aligned}$$ The motion of the walls are determined by the functions $L_1(w_1)$ and $L_2(w_2)$. In the derivations that follow, we use the form $L_1(w_1) = c_1$ and $L_2(w_2) = c_2$, with $c_1=0$ and $c_2$ being a constant. The terms $c_1$ and $c_2$ have physical dimensions of velocity. We define a coarse time period, $T_s$, in terms of the applied loading rate as $T_s=\frac{const}{L_2}$. Hence, $T_s c_2$ is a constant which is independent of the value of $T_s$. The fine time period, $T_f$, is defined as the smaller of the periods of the two spring mass systems. We then define the *non-dimensional* slow and fast time scales as $t = \frac{t^*}{T_s}$ and $\sigma = \frac{t^*}{T_f}$, respectively. The parameter $\epsilon$ is given by $\epsilon = \frac{T_f}{T_s}$. Then the dynamics on the slow time-scale is given by (\[eq:10.3\]) in Section \[vibrate\]. The dynamics on the fast time-scale is written as: $$\begin{aligned} \label{eq:10.4} \frac{dx_{1}}{d\sigma} &= T_f \, y_{1}\nonumber\\ \frac{dy_{1}}{d\sigma}&= -T_f \left(\frac{k_{1}}{m_{1}}(x_{1}-w_{1}) - \frac{\eta}{m_{1}}(y_{2}-y_{1})\right)\nonumber\\ \frac{dx_{2}}{d\sigma}&= T_f \, y_{2} \nonumber\\ \frac{dy_{2}}{d\sigma}&= -T_f \left(\frac{k_{2}}{m_{2}}(x_{2}-w_{2}) + \frac{\eta}{m_{2}}(y_{2}-y_{1})\right)\nonumber\\ \frac{dw_{1}}{d\sigma}&= \epsilon \, T_s \, L_{1}(w_{1})\nonumber\\ \frac{dw_{2}}{d\sigma}&=\epsilon \, T_s \, L_{2}(w_{2}).\end{aligned}$$ [**Remark.**]{} A special case of (\[eq:10.3\]) is when $c_1=0$ and $c_2=0$. This represents the unforced system i.e. the walls remain fixed. Then (\[eq:10.3\]) is modified to: $$\begin{aligned} \label{eq:unforced system} \dot{ {\bf x} } = {\bf B} {\bf x}, \end{aligned}$$ where ${\bf x}={(x_1, y_1, x_2, y_2)}^T$. The overhead dot represent time derivatives w.r.t. $t$. The matrix [**B**]{} is given by $${\bf B} = T_s \begin{pmatrix} 0 & 1 & 0 & 0 \\ -\frac{k_1}{m_1} & -\frac{\eta}{m_1} & 0 & \frac{\eta}{m_1} \\ 0 & 0 & 0 & 1 \\ 0 & \frac{\eta}{m_2} & -\frac{k_2}{m_2} & -\frac{\eta}{m_2} \end{pmatrix} .$$ Example II: Case 1 - Validity of commonly used approximations {#prb2:res_case1} ============================================================= The mechanical system can be written in the form $$\begin{aligned} \label{appC_qs} {\left(\frac{T_i}{T_s} \right)}^2 {\bf A}_1 \frac{d^2 {\bf x}}{dt^2} &+ \left(\frac{T_\nu}{T_s} \right) {\bf A}_2 \frac{d {\bf x}}{dt} + {\bf A}_3 {\bf x} = \frac {\bf F}{k} \\ \frac{d {\bf w}}{dt} &={\bf L}({\bf w}) \nonumber,\end{aligned}$$ where $t = \frac{t^*}{T_s}$ where $t^*$ is dimensional time and $T_s$ is a time-scale of loading defined below, ${T_i}^2=\frac{m}{k}$, $T_\nu=\frac{D}{k}$ (the mass $m$, damping $D$ and stiffness $k$ have physical dimensions of $mass$,   $\frac{Force \times time}{length}$ and $\frac{Force}{Length}$ respectively), [**x**]{} and [**w**]{} are displacements with physical units of $length$, ${\bf L}$ is a function, *independent of $T_s$*, with physical units of $length$ , and ${\bf A}_1$, ${\bf A}_2$ and ${\bf A}_3$ are non-dimensional matrices. In this notation, $\frac{T_i}{T_s}=\epsilon$. In the examples considered, ${\bf L} = \tilde{\bf c} = T_s {\bf c}$, where ${\bf c}$ has physical dimensions of $velocity$ and is assumed given in the form ${\bf c} = \frac{\tilde{\bf c}}{T_s}$ thus serving to define $T_s$; $\tilde {\bf c}$ has dimensions of $length$. Necessary conditions for the application of the Tikhonov framework are that $\frac{T_i}{T_s} \to 0$, $\frac{T_\nu}{T_s} \to 0$ as $T_s \to \infty$. Those for the quasi-static assumption, commonly used in solid mechanics when loading rates are small, are that $\frac{T_i}{T_s} \to 0$ and $\frac{T_\nu}{T_s} \approx 1$ as $T_s \to \infty$. In our example, we have $m=1 kg$,$k=2 \times 10^7 N/m$, $D=5 \times 10^3 N s /m$ and $T_s=100 s$. Hence $\frac{T_i}{T_s}=2.24 \times 10^{-6}$ and $\frac{T_\nu}{T_s}=2.5 \times 10^{-6}$, and the damping is not envisaged as variable as $T_s \to \infty$, which shows that the quasi-static approximation is not applicable. Nevertheless, due to the common use of the quasi-assumption under slow loading in solid mechanics (which amounts to setting $\epsilon \frac{dy_1}{dt} =0$, $\epsilon \frac{dy_2}{dt}=0$ in ), we record the quasi-static solution as well. [**The Tikhonov framework.**]{} It is easy to see that under the conditions in Case 1, when the walls do not move, all solutions tend to an equilibrium (that may depend on the position of the walls). Indeed, the only way the energy will not be dissipated is when $y_1(t) = y_2(t)$ along time intervals, a not sustainable situation. Thus, we are in the classical Tikhonov framework, and, as we already noted toward the end of the introduction to the paper, the limit solution will be of the form of steady-state equilibrium of the springs, moving on the manifold of equilibria determined by the load, namely the walls. Computing the equilibria in (equivalently ) is straightforward. Indeed, for fixed ($w_1$,$w_2$) we get $$\begin{aligned} \label{appC_tikhonov} x_1 &= w_1, \quad x_2 = w_2 \nonumber\\ y_1 &= 0, \qquad y_2 = 0. \end{aligned}$$ Under the assumption that $L_1 = 0$ and $L_2 = c_2$ we get $$\begin{aligned} \label{appC_tikhonov2} w_1(t) &= 0 \nonumber\\ w_2(t) &= c_2 T_s t.\end{aligned}$$ Plugging the dynamics in yields the limit dynamics of the springs. The real world approximation for $\epsilon$ small would be a fast movement toward the equilibrium set (i.e., a boundary layer which has damped oscillations), then an approximation of -. Our computations in Section \[prb2:res\_num\_case1\] corroborate this claim. Solutions to seem to suggest that this is one example where the limit solution - is attained by a sequence of solutions of as $\epsilon \to 0$ or $T_s \to \infty$ in a ‘strong’ sense (i.e. not in the ‘weak’ sense of averages); e.g. for small $\epsilon > 0$, $y_2$ takes the value $c_2$ in the numerical calculations and this, when measured in units of slow time-scale $T_s$ (note that $y$ has physical dimensions of velocity), yields $T_s c_2$ which equals the value of the (non-dimensional) time rates of $w_2$ and $x_2$ corresponding to the limit solution -; of course, $c_2 \to 0$ as $T_s \to \infty$, by definition, and therefore $y_2 \to 0$ as well. Thus, the kinetic energy and potential energy evaluated from the limit solution -, i.e. 0 respectively, are a good approximation of the corresponding values from the actual solution for a specific value of small $\epsilon > 0$, as given in Sec. \[prb2:res\_num\_case1\]. [**The quasi-static assumption.**]{} We solve the system of equations: $$\begin{aligned} \label{eq:10.5} -k_1 (x_1 - w_1) + {\eta}( {y_2} - {y_1} ) &= 0 \nonumber\\ -k_2 (x_2 - w_2) - {\eta}( {y_2} - {y_1} ) &= 0 \nonumber\\ y_1 = {1 \over T_s} \frac{d x_1}{dt} \nonumber\\ y_2 = {1 \over T_s} \frac{d x_2}{dt}.\end{aligned}$$ We assume the left wall to be fixed and the right wall to be moving at a constant velocity of magnitude $c_2$, so that $w_1=0$ and $w_2=c_2 \, T_s \, t$. This results in $$\begin{aligned} \frac{d x_1}{dt} + \frac{k_1 T_s}{\eta (1+ \frac{k_1}{k_2} ) } x_1 = \frac {c_2 T_s} {1+ \frac{k_1}{k_2}}. \end{aligned}$$ Solving for $x_1$ and using , we get the following solution $$\begin{aligned} \label{ex2_uneql_quasistatic} x_1 &= \frac{c_2 \eta}{k_1} + \alpha \, e^{- \beta T_s t }, \quad y_1 = -\alpha \, \beta e^{-\beta T_s t}, \nonumber\\ x_2 &= c_2 T_s t - \frac{c_2 \eta} { k_2} - \frac{k_1}{k_2} \alpha e^{-\beta T_s t }, \quad y_2 = c_2 - \frac{k_1}{k_2} \alpha \beta e^{-\beta T_s t}, \end{aligned}$$ where $\alpha$ is a constant of integration and $\beta=\frac{k_1}{\eta (1 + \frac{k_1}{k_2})}$. [**Remark.**]{} The solution of the unforced system given by (\[eq:unforced system\]) in Appendix \[prb2:equations\] is of the form $$\begin{aligned} \label{eq:sol_complex} {\bf x} (t) = \sum_{i=1}^4 Q_i e^{\lambda_i t} {\bf V}_i, \end{aligned}$$ where ${\lambda_i}$ and ${\bf V}_i$ are the eigenvalues and eigenvectors of [**B**]{} respectively. Using the values provided in Table \[tab:simulation\_details\_uneql\] to construct [**B**]{}, we find that ${\lambda_i}$ and ${\bf V}_i$ are complex. The general real-valued solution to the system can be written as $\sum_{i=1}^4 \psi_i e^{\gamma_i t} {\bf M}_i(t)$ where \[tab:uneql\_eigval\] We see that all the real (time-dependent) modes ${\bf M}_i(t)$ are decaying. Moreover, none of the modes describe the dashpot as being undeformed i.e. ${\bf M}_{i,1}(t)={\bf M}_{i,3}(t) $ and ${\bf M}_{i,2}(t)={\bf M}_{i,4}(t)$ (where ${\bf M}_{i,j}(t)$ is the $j^{th}$ row of the mode ${\bf M}_i(t)$). Therefore, solution [**x**]{}(t) goes to rest when $t$ becomes large and the initial transient dies. Example II: Case 2 - Closed-form Solution {#prb2:res_case2} ========================================= We can convert to the following: $$\begin{pmatrix} \ddot{x_1} \\ \ddot{x_2} \end{pmatrix} + \alpha {T_s}^2 \begin{pmatrix} x_1 \\ x_2 \end{pmatrix} + {T_s} \begin{pmatrix} \frac {\eta} {m_1} & -\frac {\eta} {m_1} \\ -\frac {\eta} {m_2} & \frac {\eta} {m_2} \end{pmatrix} \begin{pmatrix} \dot{x_1} \\ \dot{x_2} \end{pmatrix} = \begin{pmatrix} \alpha \, w_1 \\ \alpha \, w_2 \end{pmatrix}, \label{sys2}$$ where $\alpha = \frac {k_1} {m_1} = \frac{ k_2} {m_2} $ and $w_1$ and $w_2$ are defined as : $$\begin{aligned} \label{ex2:load} w_1(t)=0 \nonumber\\ w_2(t)=c_2 \, T_s \, t .\end{aligned}$$ The overhead dots represent time derivatives w.r.t. $t$. The above equation is of the form: $$\ddot{ {\bf x} } + \alpha \, {T_s}^2 \, {\bf x} + T_s { \bf A} \dot{ {\bf x} } = {\bf g} (t),$$ with the general solution \[sol\] $$x_1= C_1 cos (\sqrt{\alpha} T_s t) + C_2 sin (\sqrt{\alpha} T_s t) - \frac {m_2} {m_1} C_3 e^{- p_1 T_s t} - \frac {m_2} {m_1} C_4 e^{- p_2 T_s t} + \frac { \eta c_2 } {k_1}, \label{sol1}$$ $$x_2= C_1 cos (\sqrt{\alpha} T_s t) + C_2 sin (\sqrt{\alpha} T_s t) + C_3 e^{- p_1 T_s t} + C_4 e^{- p_2 T_s t} + c_2~T_s~t - \frac { \eta c_2 } {k_2}, \label{sol2}$$ where $ p_1 = \frac { \eta(m_1 + m_2) + \sqrt{ \eta^2 (m_1 + m_2)^2 - 4 \alpha m_1^2 m_2^2 } } { 2 m_1 m_2} $ and $ p_2 = \frac { \eta(m_1 + m_2) - \sqrt{ \eta^2 (m_1 + m_2)^2 - 4 \alpha m_1^2 m_2^2 } } { 2 m_1 m_2} $. In the computational results in Section \[prb2:res\_num\_case1\] and \[prb2:res\_num\_case2\], we found that $p_1,p_2 > 0$. Imposing initial conditions ${x_1}^0$ and ${x_2}^0$ on displacement and ${v_1}^0$ and ${v_2}^0$ on velocity of the two masses $m_1$ and $m_2$ respectively, we obtain \[tab:ex2\_coeff\] The *closed-form average kinetic energy* ($K^{cf}$), *closed-form average potential energy* ($P^{cf}$)and *closed-form average reaction force* ($R_2^{cf}$) are $$\begin{aligned} \label{eq:analytical_sol} K^{cf}(t) &= {1 \over \Delta} \int_{t-\Delta}^{t} \left({1 \over 2} m_1 y_1(s)^2 + {1 \over 2} m_2 y_2(s)^2 \right)ds \nonumber \\ P^{cf}(t) &= {1 \over \Delta} \int_{t-\Delta}^{t} \left( {1 \over 2} k_1 x_1(s)^2 + {1 \over 2} k_2 \left(x_2(s)-w_2(s)\right)^2 \right) ds \\ R_2^{cf}(t) &= {1 \over \Delta} \int_{t-\Delta}^{t} \left( -k_2 \left((x_2(s)-w_2(s)\right) \right) ds \nonumber,\end{aligned}$$ where $x_1(s)$ and $x_2(s)$ can be substituted from and $y_1(s)={1\over T_s} \frac{dx_1}{ds}$ and $y_2(s)={1\over s} \frac{dx_2}{ds}$. [**Non-dimensionalization.**]{} Let us denote $$m_{max}=\max \limits_i ~m(x_{\epsilon}(\sigma_i),l_{\epsilon}(\sigma_i)) ,$$ where $m(x_{\epsilon}(\sigma_i),l_{\epsilon}(\sigma_i))$ is given in and . Please note that $i$ is chosen such that there is no effect of the initial transient. Using and computational results in Case 2.4 in Section \[prb2:res\_num\_case2\], we find that $$\begin{aligned} K_{max}={1 \over 2} C_1^2 (k_1 + k_2), \nonumber ~~ P_{max}={1 \over 2} C_1^2 (k_1 + k_2), \nonumber ~~ R_{2,max}= C_1 k_2. \nonumber \end{aligned}$$ We introduce the following non-dimensional variables: $$\begin{aligned} \label{non-dim} \tilde{K}^{cf}=\frac {K^{cf}}{K_{max}}, ~~ \tilde{P}^{cf}=\frac {P^{cf}}{P_{max}}, ~~ \tilde{R_2}^{cf}=\frac {R_2^{cf}}{R_{2,max}}.\end{aligned}$$ Henceforth, while referring to the dimensionless variables, we drop the overhead tilde for simplicity. ![ *$P^{cf}$ as a function of $t$*. []{data-label="fig:ex2_P_analytical"}](figures/ex2_K_anal.pdf){width="\linewidth"} ![ *$P^{cf}$ as a function of $t$*. []{data-label="fig:ex2_P_analytical"}](figures/ex2_P_anal.pdf){width="\linewidth"} We evaluated numerically (since the analytical expressions become lengthy) with $2 \times 10^6$ integration points over the interval $[t-\Delta,t]$ using Simpson’s rule and we used $2000$ discrete points for $t$. We substituted the values provided in Table \[tab:simulation\_details\_uneql\]. We repeated the calculations with $4 \times 10^6$ integration points over the interval $[t-\Delta,t]$ and $4000$ discrete points for $t$, and found the results to be the same. We see in Fig. \[fig:ex2\_K\_analytical\] and Fig.\[fig:ex2\_P\_analytical\] that $K^{cf}$ and $P^{cf}$ oscillate around 0.5 with very small amplitude for different values of $\epsilon$ (recall that $\epsilon = \frac{T_f}{T_s}$ and we think of $T_f$ being fixed with $T_s \rightarrow \infty$ to effect $\epsilon \rightarrow 0$). However, as we decrease $\epsilon$, the amplitude of oscillations of $R_2^{cf}$ decreases and it goes to zero for $\epsilon=1.98 \times 10^{-7}$ as we see in Fig. \[fig:ex2\_R\_analytical\]. We use these ‘closed-form’ results to compare with the PTA results in Section \[prb2:res\_num\_case1\] and Section \[prb2:res\_num\_case2\]. [**Remark.**]{} The criteria for convergence of $R^m_t$ (as mentioned in the discussion following in Section \[impl\_algo\]) is discussed as follows. Let us denote $$m_{I,t}={1 \over I} \sum_{i=1}^{I} m(x_\epsilon(\sigma_i),l_\epsilon(\sigma_i)),$$ where $x_\epsilon(\sigma_i),l_\epsilon(\sigma_i)$ is defined in , $I \in \mathbb{Z_+}$ and $m(x_\epsilon(\sigma_i),l_\epsilon(\sigma_i))$ is non-dimensionalized $\forall i \in [1,I]$. Then, we say that $m_{I,t}$ has converged if $$\begin{aligned} \label{eq:conv_crit} |\frac{m_{I-jk,t} ~~ - ~~ m_{I-pk,t}}{m_{I-pk,t}}| \leq tol_1^m, \end{aligned}$$ $\forall j \in [0,p]$ where $p,k \in \mathbb{Z_+}$ and $p,k<I$ and $tol_1^m$ is a specified value of tolerance (which is a small value generally around $10^{-2}$). We declare $N_t=I$ (where $N_t$ is defined in the discussion following ) and $R^m_t=m_{I,t}$. In situations where $m_{I-pk,t}$ becomes very small so that the convergence criteria in cannot be practically implemented, we say that $m_{I,t}$ has converged if $$\begin{aligned} |m_{I-jk,t}| \leq tol_2^m,\end{aligned}$$ $\forall j \in [0,p]$ and $tol_2^m$ is a specified value of tolerance (generally around $10^{-5}$). [**Remark.**]{} We used the Simpson’s rule of numerical integration to obtain the value of the slow variable instead of using to obtain the results in Section \[prb2:res\_num\_case1\] and \[prb2:res\_num\_case2\]. The Simpson’s rule of numerical integration for any function $f(t)$ over the interval $[a,b]$ where $a,b \in \mathbb{R}$ is $$\int_a^b f(t)dt \approx \frac{\Delta t}{n} [f(t_0)+4f(t_1)+2f(t_2)+...+2f(t_{n-2})+4f(t_{n-1})+f(t_n)],$$ where $\Delta t=\frac{b-a}{n}$, $n \in \mathbb{Z_+}$, $n$ is even and $t_i=t_0+i\Delta t$ for $i \in \mathbb{Z}$ and $i \leq n$. All the function evaluations at points with odd subscripts are multiplied by 4 and all the function evaluations at points with even subscripts (except for the first and last) are multiplied by 2. Since we are calculating the value of the slow variable $v$ given by , the function $f(t)$ is given by $R^m_t$ which is defined in , with $\epsilon = 0$ in . We chose $n=2$ and calculated the values of $R^m_{t-\Delta+\frac{i}{n}\Delta}$ using for $i = 0,1,2$. We obtained the fine initial conditions $x_{guess}(t-\Delta+\frac{i}{n}\Delta)$ as $$x_{guess}(t-\Delta+\frac{i}{n}\Delta)=x^{arb}_{t-\Delta} + \frac {\left( x^{arb}_{t-\Delta} - x^{cp}_{t-h} \right)}{\left(h-\Delta \right)} \, \frac{i\Delta}{n},$$ where $x^{arb}_{.}$ and $x^{cp}_{.}$ are defined in Step 3 in Section \[impl\_algo\]. Please note that we also used Simpson’s rule to evaluate with $n=2\times10^6$ and the details of the calculation are mentioned in *Non-dimensionalization* in this Appendix above. [**Remark.**]{} To see the effect of the initial condition on the solution, we solve (\[eq:unforced system\]) in a particular case using the values provided in Table \[tab:simulation\_details\_uneql\] but with $k_2=2 \times 10^7 N/m$ and $c_2=0$. The general real-valued solution to the system can be written as $\sum_{i=1}^4 \psi_i e^{\gamma_i t} {\bf M}_i(t)$ where \[tab:uneql\_eigval\] We see that while ${\bf M}_1(t)$ and ${\bf M}_2(t)$ are decaying real (time-dependent) modes, ${\bf M}_3(t)$ and ${\bf M}_4(t)$ are the non-decaying real (time-dependent) modes. Moreover, both ${\bf M}_3(t)$ and ${\bf M}_4(t)$ describe the dashpot as being undeformed i.e. ${\bf M}_{i,1}(t)={\bf M}_{i,3}(t) $ and ${\bf M}_{i,2}(t)={\bf M}_{i,4}(t)$ (where ${\bf M}_{i,j}(t)$ is the $j^{th}$ row of the mode ${\bf M}_i(t)$). The solution ${\bf x}(t)$ described in (\[eq:sol\_complex\]) of Appendix \[prb2:res\_case1\] can be written as $$\begin{aligned} \label{eq:coeffs} {\bf x}(t) = \sum_{i=1}^4 \kappa_i {\bf M}_i(t) ,\end{aligned}$$ where the coefficients $\kappa_i$ are obtained from the initial condition ${\bf x}_0 $ using $$\kappa_i = {\bf x}_0 \cdot {\bf M}_i^d (0)$$ where ${\bf M}_i^d (t)$ are the dual basis of ${\bf M}_i (t)$. Acknowledgment {#acknowledgment .unnumbered} ============== We thank the anonymous referees for their valuable comments that helped improve the presentation of our paper. S. Chatterjee acknowledges support from NSF grant NSF-CMMI-1435624. A. Acharya acknowledges the support of the Rosi and Max Varon Visiting Professorship at The Weizmann Institute of Science, Rehovot, Israel, and the kind hospitality of the Dept. of Mathematics there during a sabbatical stay in Dec-Jan 2015-16. He also acknowledges the support of the Center for Nonlinear Analysis at Carnegie Mellon and grants NSF-CMMI-1435624, NSF-DMS-1434734, and ARO W911NF-15-1-0239. Example I is a modification of an example that appeared in a draft of \[AKST07\], but did not find its way to the paper. [^1]: Dept. of Civil & Environmental Engineering, Carnegie Mellon University, Pittsburgh, PA 15213. [email protected]. [^2]: Dept. of Civil & Environmental Engineering, and Center for Nonlinear Analysis, Carnegie Mellon University, Pittsburgh, PA 15213. [email protected]. [^3]: Dept. of Mathematics, The Weizmann Institute of Science, Rehovot, Israel, 7610001. [email protected]. [^4]: [**To appear in Journal of Computational Physics.**]{}

ArXiv

--- abstract: 'We develop an interference alignment (IA) technique for a downlink cellular system. In the uplink, IA schemes need channel-state-information exchange *across base-stations of different cells*, but our downlink IA technique requires feedback *only within a cell*. As a result, the proposed scheme can be implemented with a few changes to an existing cellular system where the feedback mechanism (within a cell) is already being considered for supporting multi-user MIMO. Not only is our proposed scheme implementable with little effort, it can in fact provide substantial gain especially when interference from a dominant interferer is significantly stronger than the remaining interference: it is shown that in the two-isolated cell layout, our scheme provides four-fold gain in throughput performance over a standard multi-user MIMO technique. We show through simulations that our technique provides respectable gain under a more realistic scenario: it gives approximately 20% gain for a 19 hexagonal wrap-around-cell layout. Furthermore, we show that our scheme has the potential to provide substantial gain for *macro-pico* cellular networks where pico-users can be significantly interfered with by the nearby macro-BS.' bibliography: - 'IA\_BIB.bib' title: Downlink Interference Alignment --- Interference Alignment, Downlink, Multi-User MIMO, Macro-Pico Cellular Networks Introduction ============ One of the key performance metrics in the design of cellular systems is that of cell-edge spectral efficiency. As a result, fourth-generation (4G) cellular systems, such as WiMAX [@80216m] and 3GPP-LTE [@3GPPLTE], require at least a doubling in cell-edge throughput over previous 3G systems [@3GPPLTE]. Given the disparity between average and cell-edge spectral efficiencies (ratios of about 4:1) [@80216m], the desire to improve cell-edge throughput performance is likely to continue. Since the throughput of cell-edge users is greatly limited by the presence of co-channel interference from other cells, developing an intelligent interference management scheme is the key to improving cell-edge throughput. One interesting recent development, called *interference alignment* (IA) [@Mohammad; @Jafar:IC], manages interference by aligning multiple interference signals in a signal subspace with dimension smaller than the number of interferers. While most of the work on IA [@Jafar:IC; @Jafar:DB; @Heath:First] has focused on $K$ point-to-point interfering links, it has also been shown in [@SuhTse:IA; @Caire:IA; @Tresch:CellularIA] that IA can be used to improve the cell-edge user throughput in a cellular network. Especially, it was shown in [@SuhTse:IA] that *near interference-free throughput* performance can be achieved in the cellular network. While IA promises substantial theoretical gain in cellular networks, it comes with challenges in implementation. First, the uplink IA scheme in [@SuhTse:IA] requires extensive channel-state-information (CSI) to be exchanged over the backhaul *between base-stations (BSs) of different cells*. A second challenge comes from realistic cellular environments that involve multiple unaligned out-of-cell interferers. Lastly, the integration of IA with other system issues, such as scheduling, needs to be addressed. We propose a new IA technique for downlink cellular systems that addresses many of these practical concerns. Unlike the uplink IA, our downlink IA scheme requires feedback only within a cell. As a consequence, our technique can be implemented with small changes to existing 4G standards where the within-a-cell feedback mechanism is already being considered for supporting multi-user MIMO. Our proposed technique builds on the idea of the IA technique in [@SuhTse:IA] that aims for a two-isolated cell layout and can thus cancel interference only from one neighboring BS. We observe that the IA technique in [@SuhTse:IA] may give up the opportunity of providing matched-filtered gain (also called beam-forming gain in the case of multiple antennas) in the presence of a large number of interferers. Our new technique balances these two scenarios, inspired by the idea of the standard MMSE receiver that unifies a zero-forcing receiver (optimum in the high $\sf SNR$ regime) and a matched filter (optimum in the low $\sf SNR$ regime). Through simulations, we show that our scheme provides approximately 55% and 20% gain in cell-edge throughput performance for a linear cell layout and 19 hexagonal wrap-around-cell layout, respectively, as compared to a standard multi-user MIMO technique. We also find that our scheme has the potential to provide significant performance for heterogeneous networks [@Heterogeneous], e.g., macro-pico cellular networks where a dominant interference can be much stronger than the residual interference. For instance, pico-users can be significantly interfered with by the nearby macro-BS, as compared to the aggregated remaining BSs. We show that for these networks our scheme can give around 30% to 200% gain over the standard technique. Furthermore, our scheme is easily combined with a widely-employed opportunistic scheduler [@Tse:Book] for significant multi-user-diversity gain. Interference Alignment {#sec-IAtechnique} ====================== Review of Uplink IA ------------------- We begin by reviewing uplink IA in [@SuhTse:IA]. Fig. \[fig:uplinkIA\] illustrates an example for the case of two isolated cells $\alpha$ and $\beta$. Suppose that each cell has $K$ users and each user (e.g., user $k$ in cell $\alpha$) sends one symbol (or stream) along a transmitted vector $\mathbf{v}_{\alpha k} \in \mathbb{C}^{M}$. Each user can generate multiple dimensions by using subcarriers (in an OFDM system), antennas, or both: $$\begin{aligned} M = \textrm{(\# of subcarriers)} \times \textrm{(\# of antennas)}.\end{aligned}$$ In this paper, we assume that each BS has the same number of dimensions: the $M$-by-$M$ symmetric configuration. The asymmetric case will be discussed in Section \[sec-discussion\]. The idea of interference alignment is to design the transmitted vectors so that they are aligned onto a one-dimensional linear subspace at the other BS. Due to the randomness in wireless channels, the transmitted vectors are likely to be linearly independent at the desired BS. Note that for $M=K+1$, the desired signals span a $K$-dimensional linear space while the interference signals only occupy a one-dimensional subspace. Hence, each BS can recover $K$ desired symbols using $K+1$ dimensions. The performance in the interference-limited regime can be captured by a notion of degrees-of-freedom ($\sf dof$). Here, ${\sf dof \; per \; cell} =\frac{K}{K+1} $, so as $K$ gets large, we can asymptotically achieve interference-free $\sf dof = 1$. On the other hand, one implementation challenge comes from the overhead of exchanging CSI needed for enabling the IA technique. The IA scheme requires each user to know its *cross*-channel information to the other BS. While in a time-division-multiplexing system, channels can be estimated using reciprocity, in a frequency-division-multiplexing system, *backhaul cooperation* is required to convey such channel knowledge. Fig. \[fig:uplinkIA\] shows a route to obtain the CSI of $\mathbf{G}_{\beta 1}$: $BS\; \beta \rightarrow backhaul \rightarrow BS \; \alpha \rightarrow feedback \rightarrow user \;1 \;of\; cell \; \alpha$. Here $\mathbf{G}_{\beta 1} \in \mathbb{C}^{(K+1) \times (K+1)}$ indicates the cross-channel from user 1 of cell $\alpha$ to BS $\beta$. On the contrary, in the downlink, we show that IA can be applied without backhaul cooperation. Downlink Interference Alignment {#sec:downlinkIA} ------------------------------- Fig. \[fig:downlinkIA\] illustrates an example of downlink IA where there are two users in each cell. The uplink-downlink duality says that the $\sf dof$ of the uplink is the same as that of the downlink. Hence, ${\sf dof \; per \; cell} = \frac{K}{K+1}= \frac{2}{3}$. To achieve this, each BS needs to send two symbols (streams) over three dimensions. The idea is similar to that of the uplink IA in a sense that two dimensions are used for transmitting desired signals and the remaining one dimension is reserved for interference signals. However, the method of interference alignment is different. We first set a 3-by-2 precoder matrix $\mathbf{P}$ at BS $\alpha$ and BS $\beta$, respectively. This spreads two data streams over three-dimensional resources. Next, each user, such as user $1$ in cell $\alpha$, estimates the interference $\mathbf{G}_{\beta 1} \mathbf{P}$ using pilots or a preamble. User $1$ then generates a vector $\mathbf{u}_{\alpha 1}$ that lies in the null-space of the $\mathbf{G}_{\beta 1} \mathbf{P}$. Since the $\mathbf{G}_{\beta 1} \mathbf{P}$ is of dimension 3-by-2, such a vector $\mathbf{u}_{ \alpha 1}$ always exists, and when applied to the received signal, it will null out the out-of-cell interference. Note that the receive vector $\mathbf{u}_{\alpha 1}$ does not guarantee the cancellation of intra-cell interference from user 2 in the same cell $\alpha$. This is accomplished as follows. In cell $\alpha$, each user feeds back its equivalent channel $\mathbf{u}_{\alpha k}^* \mathbf{H}_{\alpha k}\mathbf{P}$ (obtained after applying the receive vector) to its own BS $\alpha$, where $\mathbf{H}_{\alpha k} \in \mathbb{C}^{3 \times 3}$ indicates the direct-channel from BS $\alpha$ to user $k$ in the cell. BS $\alpha$ then applies an additional zero-forcing precoder formed by the pseudo-inverse of the composite matrix $[\mathbf{u}_{\alpha 1}^* \mathbf{H}_{\alpha 1}\mathbf{P}; \mathbf{u}_{\alpha 2}^* \mathbf{H}_{\alpha 2}\mathbf{P}]$. This zero-forcing precoder guarantees user 2’s transmitted signal $\mathbf{H}_{\alpha 1} \mathbf{P} \mathbf{v}_{\alpha 2}$ to lie in the interference space $\mathbf{G}_{\beta 1} \mathbf{P}$. Note that $\mathbf{u}_{\alpha 1}^* (\mathbf{H}_{\alpha 1} \mathbf{P} \mathbf{v}_{\alpha 2}) = 0$. A series of operations enables interference alignment. Let us call this scheme *zero-forcing IA*. To see this, let us observe the interference plane of user 1 in cell $\alpha$. Note that there are three interference vectors: two *out-of-cell* interference vectors and one *intra-cell* interference vector. These three vectors are aligned onto a two-dimensional linear subspace. Interference alignment is achieved between out-of-cell and intra-cell interference signals to save one dimension. Similarly, user 2 in the cell can save one dimension. Hence, two dimensions can be saved in total by sacrificing only one dimension. If the number of users is $K$, each cell can save $K$ dimensions by sacrificing one dimension. The loss will become negligible with the increase of $K$, as was seen in the uplink IA. While the downlink $\sf dof$ is the same as that of the uplink, the way interference is aligned is quite different. Note in Fig. \[fig:uplinkIA\] that in uplink IA, interference alignment is achieved among out-of-cell interference vectors only. On the other hand, in downlink IA, interference alignment is achieved between out-of-cell and intra-cell interference vectors at multiple users *at the same time*. **Feedback Mechanism:** Note two key system aspects of the technique. First, the exchange of cross-channel information between BSs or between users in different cells is not needed. Each BS can fix precoder $\mathbf{P}$, *independent of channel gains*. Each user can then specify the space orthogonal to the out-of-cell interference signal space. This enables the user to design a zero-forcing receive vector *without knowing the actually transmitted vectors*. Each user then feeds back the equivalent channel $\mathbf{u}_{\alpha k} \mathbf{H}_{\alpha k}\mathbf{P}$ and the BS forms the zero-forcing transmit vectors *only with the feedback of the equivalent channels*. Hence, the scheme requires only within-a-cell feedback mechanism. This is in stark contrast to the uplink IA which requires backhaul cooperation between different BSs. Secondly, while feedback is required from the user to the BS, this feedback is the same as the feedback used for standard multi-user MIMO techniques. The only difference is that in downlink IA, two cascaded precoders are used and the receive vector of each user is chosen as a null vector of out-of-cell interference signal space. As a result, the scheme can be implemented with little change to an existing cellular system supporting multi-user MIMO. Performance and Limitations --------------------------- Fig. \[fig:Performance2cellonly\] shows the sum-rate performance for downlink zero-forcing IA in a two-isolated cell layout where $M=4$ (e.g., a 4-by-4 antenna configuration). As a baseline scheme, we use a *matched filter receiver*: one of the standard multi-user MIMO techniques [@Pan:IterativeMF; @Gesbert:MUMIMO]. The scheme uses the dominant left-singular vector of the direct-channel as a receive vector. We assume a zero-forcing vector at the transmitter to null out intra-cell interference. Nulling intra-cell interference is important as its power has the same order as the desired signal power. $$\begin{aligned} \label{eq-MFreceiveweight} & \mathbf{u}_{\alpha k}^{\sf MF} = \textrm{a maximum left-singular vector of } \mathbf{H}_{\alpha k}, \\ \label{eq-ZFtransmitweight} & \mathbf{v}_{\alpha k}^{\sf ZF}= k\textrm{th normalized column of } \mathbf{H} (\mathbf{H}\mathbf{H}^*)^{-1}, \; \mathbf{H}:= \left[ \begin{array}{c} \vdots \\ \mathbf{u}_{\alpha k}^{{\sf MF}*} \mathbf{H}_{\alpha k} \\ \vdots \\ \end{array} \right].\end{aligned}$$ Note that the matched filter receiver maximizes beam-forming gain but it ignores the interference signal space. Also notice that the receive-and-transmit vectors are interconnected, i.e., a receiver vector can be updated as a function of a transmit vector and vice versa. One way to compute the transmit-and-receive vectors is to employ an iterative algorithm [@Pan:IterativeMF; @Gesbert:MUMIMO]. We call this scheme *iterative matched filtering*. See Appendix \[appen:IterativeMF\] for further details. In Fig. \[fig:Performance2cellonly\], we assume no iteration for fair comparison of CSI overhead. An opportunistic scheduler [@Tse:Book] is employed to choose a set of 3 users out of 10 such that the sum rate is maximized. We also consider uncoordinated schedulers, i.e., scheduling information is not exchanged between different BSs. One can clearly see that the zero-forcing IA provides significant (asymptotically optimum for large $\sf SNR$) performance gain for the two-isolated-cell case, as there are no residual interferers. However, for realistic multi-cellular environments, the performance may not be very good due to the remaining interferers. In order to take multi-cellular environments into account, we introduce a parameter $\gamma$ that captures the relative strength of the interference power from a dominant interferer to the remaining interference power (summed from the other BSs): $$\begin{aligned} \label{eq-gamma} \gamma:= \frac{\sf INR_{rem}}{\sf INR_{dom}},\end{aligned}$$ where $\sf INR_{dom}$ and $\sf INR_{rem}$ denote the ratios of the dominant and aggregate interference power over the noise power, respectively. Note that by adapting $\gamma$, one can cover arbitrary mobile location and cellular layouts. While, at one extreme ($\gamma =0$), the zero-forcing IA provides significant performance, at the other extreme ($\gamma \gg 1$), the scheme may not be good as it completely loses receive beam-forming gain (the zero-forcing IA receiver is independent of the direct-channel since it depends only on the interference space). In this case, one can expect that matched filtering will perform much better than the IA scheme. This motivates the need for developing a new IA technique that can balance the degrees-of-freedom gain with the matched-filtered power gain depending on the value of $\gamma$. Proposed New IA Scheme {#sec-proposedscheme} ====================== The zero-forcing IA and matched filtering schemes remind us of a conventional zero-forcing receiver and a matched filter receiver in a point-to-point channel with colored noise. So it is natural to think of a unified technique like the standard MMSE receiver. However, in our cellular context, a straightforward design of an MMSE receiver usually requires the knowledge of transmitted vectors from the other cell. Moreover, a chicken-and-egg problem arises between different cells, due to the interconnection of the transmit-and-receive vector pairs. In order to *decouple* the vector design between cells, we consider uncoordinated systems, i.e., transmit vector information is not exchanged between different cells. Under this assumption, a goal is to mimic an MMSE receiver. The idea is to *color* an interference signal space by using two cascaded precoders, one of which is a *fixed* precoder $\mathbf{\bar{P}}$ located at the front-end. With the fixed precoder, we can color the interference space, to some extent, to be independent of actually transmitted vectors. To see this, consider the covariance matrix of interference-plus-noise: $$\begin{aligned} \mathbf{\Phi}_{ k} = (1 + {\sf INR_{rem} }) \mathbf{I} + \frac{ \sf SNR }{S} (\mathbf{G}_{\beta k} \mathbf{\bar{P}} \mathbf{B}_{\beta } \mathbf{B}_{\beta }^* \mathbf{\bar{P}}^* \mathbf{G}_{\beta k}^* ),\end{aligned}$$ where $S$ is the total number of streams assigned to the scheduled users in the cell ($S \leq M$) and $\mathbf{B}_{\beta}$ indicates the zero-forcing precoder of a dominant interferer (BS $\beta$): $\mathbf{B}_{\beta} = [\mathbf{v}_{\beta 1}, \cdots, \mathbf{v}_{\beta S}] \in \mathbb{C}^{M \times S}$. Assume that the aggregate interference except the dominant interference is white Gaussian[^1]. Without loss of generality, we assume that Gaussian noise power is normalized to 1. Assume the total transmission power is equally allocated to each stream. We control the coloredness of interference signals by differently weighting the last $(M-S)$ columns of $\mathbf{\bar{P}}$ with a parameter $\kappa$ ($0 \leq \kappa \leq 1$): $$\begin{aligned} \mathbf{\bar{P}} = \left[ \mathbf{f}_1, \cdots, \mathbf{f}_{S}, \kappa \mathbf{f}_{S+1}, \cdots, \; \kappa \mathbf{f}_M \right] \in \mathbb{C}^{M \times M},\end{aligned}$$ where $\left[ \mathbf{f}_1, \cdots, \mathbf{f}_{M} \right]$ is a unitary matrix. Since we consider uncoordinated systems, $\mathbf{B}_{\beta}$ is unknown. Hence, we use the expectation of the covariance matrix over $\mathbf{B}_{\beta}$: $$\begin{aligned} \mathbf{\bar{\Phi}}_{ k}:= \mathbb{E} [\mathbf{\Phi}_{ k}] = (1 + {\sf INR_{rem} }) \mathbf{I} + \frac{ \sf SNR }{S} (\mathbf{G}_{\alpha k} \mathbf{\bar{P}} \mathbf{\bar{P}}^* \mathbf{G}_{\alpha k}^* ),\end{aligned}$$ where we assume that each entry of $\mathbf{B}_{\beta}$ is i.i.d. $\mathcal{CN}(0, \frac{1}{S} )$. Two extreme cases give insights into designing $\kappa$. When the residual interference is negligible, i.e., $\gamma \ll 1$, the scheme should mimic the zero-forcing IA, so $\mathbf{\bar{P}}$ should be rank-deficient, i.e., $\kappa =0$. Note in this case that the null space of the interference signals can be specified, independent of $\mathbf{B}_{\beta}$. As a result, the expected covariance matrix acts as the actual covariance matrix to induce the solution of the zero-forcing IA. At the other extreme ($\gamma \gg 1$), the scheme should mimic matched filtering. This motivates us to choose a unitary matrix $\mathbf{\bar{P}}$. One way for smoothly sweeping between the two cases is to set: $$\begin{aligned} \label{eq-KAPPA} \kappa = \min \left( \sqrt{ \gamma } , 1 \right).\end{aligned}$$ Note that for $\gamma \ll 1$, $\kappa \approx 0$ and for $\gamma \gg 1$, $\kappa$ is saturated as 1. Considering system aspects, however, the $\kappa$ needs to be carefully chosen. In the above choice, the $\kappa$ varies with mobile location, since $\sf INR_{rem}$ is a function of mobile location. This can be undesirable because it requires frequent adaptation of BS precoder which supports users from the cell center to the cell edge. Therefore, we propose to fix $\kappa$. For example, we can fix $\kappa$ based on the case of $\sf SNR=20$ dB, a cell-edge mobile location, and a fixed network layout, e.g., $\kappa \approx 0.34$ for the linear cell layout and $\kappa \approx 0.64$ for the 19 hexagonal wrap-around cell layout (See Fig. \[fig:gamma\]). With $\mathbf{\bar{\Phi}}_k$, we then use the standard formula of an MMSE receiver. Similar to the iterative matched filtering technique, we also employ an iterative approach to compute transmit-and-receive vector pairs. **$<$Proposed New IA Scheme$>$** 1. (*Intialization*): Each user initializes its receive vector as follows: $\forall k \in \{1,\cdots, K \},$ $$\begin{aligned} \begin{split} \mathbf{u}_{\alpha k}^{(0)} = \textrm{normalization} \left\{ \mathbf{\bar{\Phi}}_{ k}^{-1} \mathbf{H}_{\alpha k} \mathbf{\bar{P}} \mathbf{v}_{\alpha k}^{(0)} \right \}, \end{split} \end{aligned}$$ where we set $\mathbf{v}_{\alpha k}^{(0)}$ as a maximum eigenvector of $\mathbf{\bar{P}}^* \mathbf{H}_{\alpha k}^* \mathbf{\bar{\Phi}}_{ k}^{-1} \mathbf{H}_{\alpha k} \mathbf{\bar{P}}$ to initially maximize beam-forming gain. Each user then feeds back the equivalent channel $\mathbf{u}_{\alpha k}^{(0)*} \mathbf{H}_{\alpha k} \mathbf{\bar{P}}$ to its own BS. With this feedback information, the BS computes zero-forcing transmit vectors: $\forall k$ $$\begin{aligned} \begin{split} \mathbf{v}_{\alpha k}^{(1)} = k\textrm{th normalized column of } \mathbf{H}^{(1)*} (\mathbf{H}^{(1)} \mathbf{H}^{(1)*})^{-1}, \end{split} \end{aligned}$$ where $$\begin{aligned} \begin{split} \mathbf{H}^{(1)}:= \left[ \begin{array}{c} \vdots \\ \mathbf{u}_{\alpha k}^{(0)*} \mathbf{H}_{\alpha k} \mathbf{\bar{P}} \\ \vdots \\ \end{array} \right] \end{split} \end{aligned}$$. 2. (*Opportunistic Scheduling*): The BS finds $A^*$ such that $$\begin{aligned} A^{*} = \arg \max_{ A \in \mathcal{K} } \sum_{k \in A } \log \left( 1 + \frac{ \frac{ {\sf SNR} }{S} || \mathbf{u}_{\alpha k}^{(0)*} \mathbf{H}_{\alpha k} \mathbf{v}_{\alpha k}^{(1)} ||^2 }{1 + {\sf INR_{rem} } } \right),\end{aligned}$$ where $\mathcal{K}$ is a collection of subsets $ \subset \{1,\cdots,K\}$ that has cardinality $|\mathcal{K}| = \binom{K}{S}$. 3. (*Iteration*): For $A^*$, we iterate the following. The BS informs each user of $\mathbf{v}_{\alpha k}^{(i)}$ via precoded pilots. Each user updates the receive vector as follows: $$\begin{aligned} \mathbf{u}_{\alpha k}^{(i)} = \textrm{normalization} \left\{ \mathbf{\bar{\Phi}}_{ k}^{-1} \mathbf{H}_{\alpha k} \mathbf{\bar{P}} \mathbf{v}_{\alpha k}^{(i)} \right \}. \end{aligned}$$ Each user then feeds back the updated equivalent channel to its own BS. With this feedback information, the BS computes zero-forcing transmit vectors $\mathbf{v}_{\alpha k}^{(i+1)}$. **Remarks:** Although users can see out-of-cell interference, the scheduler at BS cannot compute it. Hence, we assume that the scheduler makes a decision assuming no dominant interference. Note that the denominator inside the logarithmic term contains only noise and residual interference. To reduce CSI overhead, we assume that a scheduler decision is made before the *iteration* step. In practice, we may not prefer to iterate, since it requires more feedback information. Note that the feedback overhead is exactly the same as that of iterative matched-filtering (baseline). The only difference is that we use the fixed precoder $\mathbf{\bar{P}}$ and the MMSE-like receiver employing the $\mathbf{\bar{\Phi}}_{ k}$. *This requires very little change to an existing cellular system supporting multi-user MIMO*. Simulation Results {#sec-simulation} ================== Through simulations, we evaluate the performance of the proposed scheme for downlink cellular systems. We consider one of the possible antenna configurations in the 4G standards [@80216m; @3GPPLTE]: $4$ transmit and 4 receive antennas. To minimize the change to the existing 4G systems, we consider using only antennas for the multiple dimensions, i.e., $M=4$. We focus on three different cellular layouts, illustrated in Fig. \[fig:gamma\]. We consider a specific mobile location (the mid-point between two adjacent cells), as the cell-edge throughput performance is of our main interest. We use the standard ITU-Ped path-loss model, with i.i.d. Rayleigh fading components for each of the antenna. Fig. \[fig:19hexa\] shows the throughput performance for 19 hexagonal cellular systems where $\gamma \approx 0.4$. We consider $K=10$ and $S=3$. We find through simulations that using three streams provides the best performance for a practical number of users per cell (around 10). See Appendix \[appendix:NumberofStreams\] for further details. Note that the zero-forcing IA scheme is worse than the matched filtering (baseline). This implies that when $\gamma \approx 0.4$ (residual interference is not negligible), boosting power gain gives better performance than mitigating dominant out-of-cell interference. However, the proposed unified IA technique outperforms both of them for all regimes. It gives approximately 20% throughput gain when ${\sf SNR} = 20$ dB. We also investigate the convergence of the proposed scheme. Note in Fig. \[fig:19hexa\]$(b)$ that the proposed scheme converges to the limits very fast, i.e., even one iteration is enough to derive most of the asymptotic performance gain. This means that additional iterations provide marginal gain, while requiring a larger overhead of CSI feedback. Another observation is that the converged limits of the proposed technique is invariant to the initial values of transmit-and-receive vectors. Note that random initialization induces the same limits as that of our carefully chosen initial values, but it requires more iterations to achieve the limits. Therefore, initial values need to be carefully chosen to minimize the overhead of CSI feedback. Fig. \[fig:linearcell\] shows throughput performance for a linear cell layout. In this case, the residual interference is significantly reduced at $\gamma \approx 0.1$, so mitigating dominant out-of-cell interference improves the performance more significantly than beam-forming does. The gain of the proposed scheme over the matched filtering is significant, i.e., approximately 55% in the high $\sf SNR$ regime of interest. Notice that a crossover point between the zero-forcing IA and the matched filtering occurs at around ${\sf SNR}=0$ dB. The benefit of the zero-forcing IA is substantial. Macro-Pico Cellular Networks {#sec-Application} ============================ We have observed that our scheme shows promise especially when dominant interference is much stronger than the remaining interference, i.e., $\gamma \ll 1$. Such scenario occurs often in heterogeneous networks [@Heterogeneous] which use a mix of macro, pico, femto, and relay BSs to enable flexible and low-cost deployment. In this section, we focus on a scenario of the macro-pico cell deployment, illustrated in Fig. \[fig:macro-pico\]. As shown in the figure, suppose that pico-BS is deployed at a distance $d$ from the nearby macro-BS and a user is connected to the pico-BS. The pico-user can then see significant interference from the nearby macro-BS, and this interference can be much stronger than the aggregated interference from the remaining macro-BSs, especially when $d$ is small. The interference problem can be further aggravated due to range extension techniques[^2] [@Heterogeneous] and the disparity between the transmit power levels of the macro-BS and the pico-BS. This motivates the need for intelligent interference management techniques. We show that our IA scheme can resolve this problem to provide substantial gain. To show this, we evaluate the throughput performance of pico-users in the simple scenario shown in Fig. \[fig:macro-pico\]. We assume the 19 hexagonal wrap-around cellular layout, and on top of it we deploy one pico-BS. Based on [@Heterogeneous], we consider the power levels of $46$ dBm and $30$ dBm for the macro-BS and the pico-BS, respectively, so the difference is $16$ dB. Consistent with previous simulation setups, we consider a specific mobile location where the downlink received power from the pico-BS is the same as that from the nearby macro-BS. Due to the disparity of the power levels, the pico-users are closer to the pico-BS. We assume a $4$-by-$4$ antenna configuration where $M=4$. Fig. \[fig:macro-pico\_SNR\] shows the throughput performance of the pico-users as a function of $\sf SNR$. We assume that $K=10$, $S=3$ and no iterations. We employ the opportunistic scheduler to choose the best 3 users out of 10. Fig. \[fig:macro-pico\_SNR\] $(a)$ considers the case of $\frac{d}{R}=0.5$ where pico-users are significantly interfered with by the nearby macro-BS. In this case, as one can expect, our IA scheme provides significant gain of $150$% over the matched filtering, similar to the two-isolated cell case. In Fig. \[fig:macro-pico\_SNR\] $(b)$, we also consider the case of $\frac{d}{R}=1$ where the minimum gain of our scheme is expected. Even in this worst case, our proposed scheme gives approximately $28$% gain over the matched filtering. Recall that in this simulation we consider the specific mobile location where the downlink received power from the two BSs are the same. In fact, this is a conservative case. As mentioned earlier, the use of the range extension technique expands the footprint of pico-cells and therefore aggravates the interference problem. One can expect a larger gain of our IA scheme when range extension is employed. **Comparison to Resource Partitioning:** In this scenario, as an alternative to our IA scheme, one may consider resource partitioning to resolve the interference problem. This is because unlike the conventional macro cellular networks containing many neighboring cells, this macro-pico network scenario has a fewer number of dominant interferers, thus making resource coordination simpler. For example, we can use a frequency reuse of $\frac{1}{2}$ for the scenario in Fig. \[fig:macro-pico\]. However, resource partitioning requires *explicit* coordination of frequency resources which can increase the control channel overhead. On the contrary, our IA scheme does not require explicit coordination, as it adapts only the number of streams under frequency reuse of $1$. In addition to this implementation advantage, our scheme shows respectable gain over resource partitioning. Fig. \[fig:macro-pico\_comparison\] shows the throughput performance as a function of $\frac{d}{R}$ when ${\sf SNR}=20$ dB and $K=10$. We use $S=3$ for the IA schemes and the matched filtering, while for resource partitioning we optimize the number of streams to plot the best performance curve. Notice that our scheme gives approximately $20$% gain for $\frac{d}{R}=0.5$. The smaller ratio of $\frac{d}{R}$, the larger the gain, while for large $\frac{d}{R}$, the gain becomes marginal. Extension {#sec-discussion} ========= Asymmetric Antenna Configuration -------------------------------- As one natural extension, we consider asymmetric antenna configuration where the BSs are equipped with more antennas. A slight modification of our technique can cover this case. Consider $M$-by-$N$ antenna configuration where $M>N$. Compared to the symmetric case, the only difference is that the number of streams is limited by the number $N$ of receive antennas, i.e., $S \leq N$. Other operations remain the same. Specific operations are as follows. Each BS sets the precoder $\mathbf{\bar{P}}$ as follows: $$\begin{aligned} \mathbf{\bar{P}} = [\mathbf{f}_1, \cdots, \mathbf{f}_S, \kappa \mathbf{f}_{S+1}, \cdots, \kappa \mathbf{f}_M] \in \mathbb{C}^{M \times M},\end{aligned}$$ where $0 \leq \kappa \leq 1$. Notice that $S \leq N$. Each user computes the expected covariance matrix by averaging over the transmitted signals from the other cell and then applies the standard MMSE formula for a receive vector. The BS then computes the zero-forcing transmit vectors with the feedback information. These steps can then be iterated. While our technique can be extended to any antenna configuration, interpretation needs to be carefully made for some cases. For example, consider 4-by-2 antenna configuration in a two-cell layout. Our scheme allows each BS to send one stream out of two and therefore each user sees only one interference vector from the other cell. This induces no interference alignment. Even in this configuration, however, interference alignment can be achieved if multiple subcarriers are incorporated. This will be discussed in the following section. Using Subcarriers ----------------- Recall in our simulations that only antennas are employed to generate multiple dimensions. We can also increase $M$ by using multiple subcarriers, thereby improving performance as the dimension reserved for interference signals becomes negligible with the increase of $M$. For example, we can create 8-by-4 configuration by using two subcarriers in a 4-by-2 antenna configuration. Interestingly, unlike the 4-by-2 configuration, this 8-by-4 configuration enables interference alignment. To see this, consider a two-cell layout where each cell has three users. Our scheme allows each BS to transmit three streams out of four and thus each user sees five interfering vectors in total: three *out-of-cell* and two *intra-cell* interfering vectors. Notice the five interfering vectors are aligned onto a three dimensional linear subspace, thereby achieving interference alignment. Open-Loop Multi-User MIMO ------------------------- Since the feedback mechanism of our scheme is the same as that of standard multi-user MIMO techniques, any CSI feedback reduction scheme used for standard techniques can also be applied to our proposed scheme. For example, an open-loop multi-user MIMO technique can be easily applied to our scheme. Our scheme has only two differences: (1) each BS employs two cascaded precoders, including a fixed precoder $\mathbf{\bar{P}}$; (2) each user employs an MMSE-like receiver using $\mathbf{\bar{\Phi}}_k$. Multiple Interferers -------------------- Our IA technique removes the interference from a single dominant interferer. A slight modification can be made to cope with multiple dominant interferers. For example, consider a 19 hexagonal cell layout in Fig. \[fig:gamma\] and suppose that mobiles are located at the middle point of three neighboring BSs. In this case, mobiles see two dominant interferers. One simple way is to take multiple dominant interferers into account in the process of computing the expected covariance matrix. Specifically, we can use: $$\begin{aligned} \begin{split} \mathbf{\bar{\Phi}}_{k}&:= \mathbb{E} \left[ (1 + {\sf INR_{rem}}) \mathbf{I} + \frac{\sf SNR}{S} \mathbf{G}_{ \beta k} \mathbf{\bar{P}} \mathbf{B}_{\beta} \mathbf{B}_{\beta}^* \mathbf{\bar{P}}^* \mathbf{G}_{\beta k}^{*} + \frac{\sf SNR}{S} \mathbf{G}_{ \gamma k} \mathbf{\bar{P}} \mathbf{B}_{\gamma} \mathbf{B}_{\gamma}^* \mathbf{\bar{P}}^* \mathbf{G}_{\gamma k}^{*} \right] \\ & = (1 + {\sf INR_{rem}}) \mathbf{I} + \frac{\sf SNR}{S} \mathbf{G}_{ \beta k} \mathbf{\bar{P}} \mathbf{\bar{P}}^* \mathbf{G}_{\beta k}^{*} + \frac{\sf SNR}{S} \mathbf{G}_{ \gamma k} \mathbf{\bar{P}} \mathbf{\bar{P}}^* \mathbf{G}_{\gamma k}^{*}, \end{split}\end{aligned}$$ where $\mathbf{G}_{\beta k}$ denotes cross-link channel from BS $\beta$ to user $k$ in cell $\alpha$ and $\mathbf{B}_{\beta}$ indicates the zero-forcing precoder of BS $\beta$, and we use similar notation ($\mathbf{G}_{\gamma k}, \mathbf{B}_{\gamma}$) for cell $\gamma$. We further assume that each entry of $\mathbf{B}_{\beta}$ and $\mathbf{B}_{\gamma}$ is i.i.d. $\mathcal{CN}(0,\frac{1}{S})$. Optimization of $\kappa$ ------------------------ Our proposed scheme employs a parameter $\kappa$ in constructing the precoder $\mathbf{\bar{P}}$. We have considered one particular choice of (\[eq-KAPPA\]), and simulation results are based on this choice. However, the performance can be improved by optimizing $\kappa$. It could be future work to find the optimum $\kappa$ for different cellular layouts. Conclusion ========== We have observed that the zero-forcing IA scheme is analogous to the zero-forcing receiver, and the iterative matched-filtering technique corresponds to the conventional matched-filter receiver. Based on this observation, we proposed a unified IA technique similar to an MMSE receiver that outperforms both techniques for all values of $\gamma$, where the power of the dominant interferer may be much greater or smaller than the power of the remaining aggregate interference. Of practical importance is the fact that our proposed scheme can be implemented with small changes to an existing cellular system supporting multi-user MIMO, as it requires only a localized *within-a-cell* feedback mechanism. This technique can be extended to asymmetric antenna configurations, scenarios with more than one dominant interferer, and low CSI schemes such as open-loop MU-MIMO. Our technique also shows even greater performance gains for macro-pico cellular networks where the dominant interference is much stronger than the remaining interference. Iterative Matched Filtering (Baseline) {#appen:IterativeMF} ====================================== We compute the transmit-and-receive vector pairs using an iterative algorithm [@Pan:IterativeMF; @Gesbert:MUMIMO]. We describe the algorithm combined with opportunistic scheduler. 1. (*Intialization*): Each user initializes a receive vector so as to maximize beam-forming gain: $\forall k \in \{ 1,\cdots, K\}$, $$\begin{aligned} \begin{split} %& \mathbf{v}_{\alpha k}^{(0)}= \textrm{a maximum \emph{right}-singular vector of } \mathbf{H}_{ k}, \\ \mathbf{u}_{\alpha k}^{(0)} = \textrm{a maximum left-singular vector of } \mathbf{H}_{\alpha k}. \end{split} \end{aligned}$$ Each user then feeds back the equivalent channel $\mathbf{u}_{\alpha k}^{(0)*} \mathbf{H}_{\alpha k}$ to its own BS. With this feedback information, the BS computes zero-forcing transmit vectors: $\forall k$, $$\begin{aligned} \begin{split} \mathbf{v}_{\alpha k}^{(1)} = k\textrm{th normalized column of } \mathbf{H}^{(1)*} (\mathbf{H}^{(1)} \mathbf{H}^{(1)*})^{-1}, \end{split} \end{aligned}$$ where $$\begin{aligned} \begin{split} \mathbf{H}^{(1)}:= \left[ \begin{array}{c} \vdots \\ \mathbf{u}_{\alpha k}^{(0)*} \mathbf{H}_{\alpha k} \\ \vdots \\ \end{array} \right] \end{split} \end{aligned}$$ 2. (*Opportunistic Scheduling*): The BS finds $A^*$ such that $$\begin{aligned} A^{*} = \arg \max_{ A \in \mathcal{K} } \sum_{k \in A } \log \left( 1 + \frac{ \frac{ {\sf SNR} }{S} || \mathbf{u}_{\alpha k}^{(0)*} \mathbf{H}_{\alpha k} \mathbf{v}_{\alpha k}^{(1)} ||^2 }{1 + {\sf INR_{dom} } + {\sf INR_{rem} } } \right).\end{aligned}$$ where $\mathcal{K}$ is a collection of subsets $ \subset \{1,\cdots,K\}$ that has cardinality $|\mathcal{K}| = \binom{K}{S}$. 3. (*Iteration*): For $A^*$, we iterate the following. The BS informs each user of $\mathbf{v}_{\alpha k}^{(i)}$ via precoded pilots. Each user updates the receive vector: $$\begin{aligned} \begin{split} \mathbf{u}_{\alpha k}^{(i)} = \textrm{normalization} \left\{ \mathbf{H}_{\alpha k} \mathbf{v}_{\alpha k}^{(i)} \right\}, k \in A^*. \end{split} \end{aligned}$$ Each user then feeds back the updated equivalent channel $\mathbf{u}_{\alpha k}^{(i)*} \mathbf{H}_{\alpha k}$ to its own BS. With this feedback information, the BS computes zero-forcing transmit vectors: $$\begin{aligned} \begin{split} \mathbf{v}_{\alpha k}^{(i+1)} = k\textrm{th normalized column of } \mathbf{H}^{(i+1)*} (\mathbf{H}^{(i+1)} \mathbf{H}^{(i)*})^{-1}, \end{split} \end{aligned}$$ where $$\begin{aligned} \begin{split} \mathbf{H}^{(i+1)}:= \left[ \begin{array}{c} \vdots \\ \mathbf{u}_{\alpha k}^{(i)*} \mathbf{H}_{\alpha k} \\ \vdots \\ \end{array} \right]. \end{split} \end{aligned}$$ **Remarks:** Although users can see out-of-cell interference, the scheduler at BS cannot compute it. We assume the scheduler uses the average power of the dominant interference. Note that the denominator inside the logarithmic term contains noise, dominant interference and residual interference. To reduce CSI overhead, we assume a scheduler decision is made before an iteration. In practice, we may prefer not to iterate, since it requires more feedback information. Discussion on the Number of Streams {#appendix:NumberofStreams} =================================== The number of streams is related to the effect of scheduling. We investigate the relationship through simulations. Fig. \[fig:BLStreams\] shows the sum-rate performance for the matched filtering (baseline) as a function of $K$. Note that with an increase in $K$, using more streams gives better performance. This is because for a large value of $K$, an opportunistic scheduler provides good signal separation and power gain, thereby inducing the high [SINR]{} regime where multiplexing gain affects the performance more significantly than beam-forming gain does. Notice that for a practical range of $K$ (around 10), using 3 streams provides the best performance. Acknowledgment {#acknowledgment .unnumbered} ============== We gratefully acknowledge Alex Grokhov, Naga Bhushan and Wanshi Chen for discussions on the heterogeneous network scenario. [^1]: To be more accurate, we may consider two or three dominant interferers for an actual realization. See Section \[sec-discussion\] for details. [^2]: Range extension extends the footprint of pico-cells by allowing more users to connect even if users do not see the pico-BS as the strongest downlink received power. The purpose for this is to better utilize cell-splitting and maximize cell offloading gain.

ArXiv

--- abstract: 'We show that the associated production $pp\rightarrow W^{\prime }W$, and the rare decay $pp\rightarrow W''\rightarrow \bar\ell\ell W$ are useful tests of $W''$ couplings to fermions at future hadron colliders. For $M_{W''}\sim (1 - 3)$ TeV they would allow a clean determination on whether the $W''$ couples to $V-A$ or $V+A$ currents. As an illustration a model in which the $W^{\prime\pm}$ couples only to $V-A$ currents is contrasted to the left-right symmetric models which involve $V+A$ currents.' author: - | Mirjam Cvetič, Paul Langacker and Jiang Liu\ *Department of Physics\ *University of Pennsylvania\ *Philadelphia, PA 19104\ UPR-0579T*** title: August 1993 --- =6.5in =2ex =5ex plus.3em minus.5em plus.3em minus.5em plus.2em minus.4em plus.2em minus.4em =10000 PACS \# 12.15, 12.10, 11.15 Many types of new physics, including some grand unified and superstring theories, predict the existence of additional charged and neutral gauge bosons ($W', Z'$). While their masses are [*a priori*]{} arbitrary, it is at least possible that they may be in the experimentally accessible range of a few TeV [@rfi]. The present direct and indirect limits on additional gauge bosons are very model dependent. The bounds on the mass of a new $Z'$ are $160-400$ GeV [@rfv; @rfii; @rfiii; @delA], although the limits are stronger, e.g., $500-1000$ GeV, in some models in which the $Z'$ mass and the $Z-Z'$ mixing are related. In the version of left-right symmetric models [@rfip] with equal left- and right-handed gauge couplings and magnitudes of quark-mixing matrix elements one has the stringent limit $M_{W'}>1.4\hbox{~TeV} $ from the $K_L-K_S$ mass difference [@rfiip]. In general left-right models, however, one has the weaker limit [@rfivp] $g_L M_{W'}/g_R>300\hbox{~GeV}$. Stronger limits follow from CP violation unless there is fine tuning [@rfLoWy]. Heavy $Z'$ and $W'$ can be produced and detected by their leptonic decays at the Large Hadron Collider (LHC) and Superconducting Super Collider (SSC) for masses up to $\sim 5$ TeV [@rfi], [@rfx]-[@HRII]. To identify the origin of such bosons, more detailed diagnostic probes of their couplings will be needed. Recent detailed studies [@rfxiiip]-[@25a] have demonstrated that the rare decay process [@rfxiiip; @rfxiv] $Z'\to \bar f_1f_2 V\ (V=W,Z)$, where $f_{1,2}$ are ordinary fermions, the associated production [@asso] $pp\to Z'V $, and the rapidity distribution in $pp\to Z'\to \ell^+\ell^-$ [@25a] are useful diagnostics of the $Z'$ couplings to the ordinary fermions. Another clean probe for the gauge couplings of $Z'$ and $W'$ is the forward-backward asymmetry [@LRR]. For $pp\to Z'\to \ell^+\ell^-\ (\ell=e$ or $\mu$) and $pp\to W^{\prime\pm}\to \ell^{\pm}\nu_{\ell}$ the asymmetries can distinguish between different models for $M_{Z'(W')}$ up to a few TeV, and test some combinations of the couplings of $Z'$ and $W'$ to quarks and leptons. However, the forward-backward asymmetry for $W^{\prime\pm}$ does not distinguish $V+A$ couplings from $V-A$. Although the most likely extension of the standard model involving a $W'$ is the left-right symmetric model [@rfip] with $V+A$ couplings, it is possible to construct viable models with $V-A$ couplings as well [@GJS]. It is therefore important to be able to distinguish $V+A$ from $V-A$. Possibilities for distinguishing the handedness of $W'$ have been recently pointed out [@rfxiv; @rfix; @asso]. The basic idea is that the ordinary $W^{\pm}$ has only $V-A$ couplings, which acts as a filter for testing the handedness of $W^{\prime\pm}$. For example, if $W'$ has only $V+A$ couplings, the decay $W^{\prime\pm}\to W^{\pm}\ell^+\ell^-$ will not occur at the lowest order except for small corrections from lepton masses. For the same reason the process $pp\to W^{\prime\pm}W^{\mp}$ would be strongly suppressed if $W'$ has the opposite handedness as $W$: in the left-right-symmetric model the suppression is proportional to the square of the $W'-W$ mixing angle or to the ratio $m_f^2/M_W^2$, where $m_f$ is a small fermion mass. On the other hand, if the $W'$ couples to $V-A$ currents these processes would not be suppressed by the mismatch of the handedness. In this paper we will examine this possibility in more detail. We will show that the number of events in the $V-A$ case can be sufficiently large to allow a clean determination of the handedness of a $W'$ with mass of the order of $(1-3)$ TeV. As an illustration we consider a theory [@GJS] in which $W'$ couples to $V-A$ currents. This is an ‘un-unified’ theory of weak interaction with a gauge structure $SU(2)_{q}\times SU(2)_{\ell}\times U(1)_Y$, in which the left-handed quarks and leptons transform as doublets of their own $SU(2)$. One set of linear combinations of the gauge bosons of $SU(2)_q$ and $SU(2)_{\ell}$ give the standard $W$ and $Z$, and the other become $W'$ and $Z'$. In this model both $W'$ and $Z'$ couple to $V-A$ currents. While this model was originally proposed as an alternative to the standard electroweak model with relatively light $W'$ and $Z'$, for our purpose we only consider situations in which the extra gauge bosons are heavy, i.e., $M_{W'}, M_{Z'}\geq1$ TeV. Then to leading order of $M_W^2/M_{W'}^2$ one finds $M_{W'}= M_{Z'}$. Neglecting fermion mixings the charged current interaction is given by $$\begin{aligned} {\cal{L}}_{\rm CC}=-{g\over 2\sqrt 2} &\Bigl\{& \Bigl[W^-_{\mu} + \cot \phi W^{\prime -}_{\mu}\Bigr]\bar u\gamma^{\mu}(1-\gamma_5) d\nonumber\\ &+& \Bigl[W^-_{\mu}-\tan\phi W^{\prime -}_{\mu}\Bigr]\bar\ell\gamma^{\mu}(1-\gamma_5) \nu_{\ell}\Bigr\},\label{CC}\end{aligned}$$ where $\tan\phi=g_{\ell}/g_{q}$ and $g_{\ell(q)}$ is the gauge coupling constant of $SU(2)_{\ell(q)}$. To have a meaningful perturbation calculation in what follows we consider that $\phi$ is not close to $0$ or $\pi/2$. The neutral current interaction of $Z$ has the standard form, whereas that of $Z'$ is $$\begin{aligned} {\cal{L}}_{NC}(Z')=-{g\over 4}Z'_{\mu}&\Bigl\{& \cot\phi\Bigl[\bar u\gamma^{\mu}(1-\gamma_5)u - \bar d\gamma^{\mu} (1-\gamma_5)d\Bigr]\nonumber\\ &-& \tan\phi\Bigl[\bar\nu_{\ell}\gamma^{\mu}(1-\gamma_5)\nu_{\ell} -\bar\ell\gamma^{\mu}(1-\gamma_5)\ell\Bigr]\Bigr\}.\label{NC}\end{aligned}$$ The coupling constant for the trilinear $W'Z'W$ vertex is $g$, those involving a $W'ZW$ or $W'\gamma W$ vertex are further suppressed by the ratio $M_W^2/M_{W'}^2$. Gauge invariance relates the couplings in (\[NC\]) and the trilinear gauge interactions, resulting in a destructive interference for the physical processes discussed below. The heavy charged gauge boson $W'$, assuming its existence, can be produced at future hadron colliders (SSC and LHC) and can be detected via the resultant leptonic decays $pp\to W'\to \bar{\ell}\nu_{\ell} (\bar\nu_{\ell}\ell)$. For given $W'$ couplings the total cross section $\sigma(pp\to W')$ can be computed quite accurately. The cross sections are given in [@LRR; @Bar; @DuLa; @HRII]. For definiteness, we assume that the neutrinos to which the $W'$ couples are massless or light. This is the case for the un-unified model [@GJS], for which $\nu_{\ell}$ is the ordinary neutrino, and in some versions of the left-right symmetric model. The same ideas would apply to models involving heavy (e.g. Majorona) neutrinos. We first address the associated production. In the un-unified model there are two tree-level graphs (Fig. 1). Contributions from these two graphs are equally important. In fact, gauge invariance requires that they interfere destructively to enforce unitarity. The squared amplitude for the quark process $q\bar q\to W'W$ averaged (summed) over initial (final) polarizations is $$\begin{aligned} {d\sigma_{W'W}\over dt} = {g^4\cot^2\phi\over 16\pi s^2}{\cal{M}}, \label{sigma}\end{aligned}$$ where $$\begin{aligned} {\cal{M}} = &-& {1\over 4t^2}\Bigl[ 3t^2 + t(s + M_{W'}^2 + M_W^2) + M_{W'}^2M_W^2\Bigr]\nonumber \\ &+& {1\over (s-M_{W'}^2)}\Bigl[ -{M_{W'}^2\over 2} + {M_W^2\over 8} + {M_W^4\over 8 M_{W'}^2} -{M_{W'}^4\over 2t} -{M_{W'}^2M_W^2\over t} - {tM_W^2\over 16M_{W'}^2} -t\Bigr]\label{M}\\ &+& {1\over (s-M_{W'}^2)^2} \Bigl[-{M_{W'}^2\over 2} - {7M_{W'}^2M_W^2\over 8} + {M_W^4\over 16} + {t M_W^2 \over 2} + { t M_W^4 \over 16 M_{W'}^2} -{t^2\over 2} - { t^2M_W^2 \over 16 M_{W'}^2}\Bigr],\nonumber\end{aligned}$$ and $s, t$ are the Mandelstam variables. The total cross section for $\sigma_{W'W}$ is obtained in a straightforward manner using the quark distribution functions of Ref. [@EHLQ]. We define the cross section for $pp\to W'W$ as the sum over $W^{\prime +}W^-$ and $W^{\prime -}W^+$. For a one year ($10^7\ s$) run at the LHC (SSC) with the projected luminosity of $10^{34}(10^{33})\ cm^{-2}\ s^{-1}$, the number of events along with typical statistical errors for the process $pp\to W'W$, with $W'$ subsequently decaying into $\bar\ell\nu_{\ell}$ and $\bar\nu_{\ell}\ell$ ($\ell= e$ and $\mu$) is presented in Table 1. In obtaining these results we have assumed for simplicity that the $W'$ only decays to ordinary fermions, with the leading term of its total rate given by $$\begin{aligned} \Gamma(W'\to\bar ff')={g^2M_{W'}\over 16\pi}\Bigl[ \tan^2\phi + 3\cot^2\phi\Bigr].\label{Gammatotal}\end{aligned}$$ These numbers are presented only for illustration. They should be contrasted with number zero, which would be the result if the $W'$ coupled to $V+A$ currents. On average the numbers for the $W'W$ associated production are about two orders of magnitude larger than those from the $Z'W$ associated production [@rfxiv]. One major reason is that in this model $W'$ has a larger coupling. Thus, the signal is still significant even for $M_{W'}= 3$ TeV. The production of $W'W$, with $W'$ subsequently decaying into leptons and $W$ into hadrons are clean events without major background. The standard-model background from $pp\to WW$ with one $W$ decaying into leptons can be cleanly eliminated at a loss of only a few percent of the signal by requiring the transverse invariant mass of the lepton system to be larger than $90$ GeV. We now turn to the rare decay process $W'\to \bar\ell\ell W$. There are two tree-level graphs displayed in Fig. 2, and their contributions are equally important. Averaging (summing) over the initial (final) polarizations of the squared amplitude we find $$\begin{aligned} {d\Gamma(W'\to \bar\ell\ell W)\over d\omega} = {g^4\tan^2\phi\over 24 (2\pi)^5 M_{W'}^3 t^2(s-M_{W'}^2)^2} \delta^4(P_{W'}-P_{W}-P_{\ell}-P_{\bar\ell}){\cal{M'}}, \label{rate}\end{aligned}$$ where $d\omega=(d^3\vec{P}_{W}/2P_{W}^0) (d^3\vec{P}_{\bar\ell}/2P_{\bar\ell}^0) (d^3\vec{P}_{\ell}/2P_{\ell}^0)$, $t=(P_{W'}-P_{\ell})^2, s=(P_{\bar\ell} + P_{\ell})^2$, with $P_{W'}, P_{W}, P_{\bar\ell}$ and $P_{\ell}$ referring to the momenta of the corresponding particles, and $$\begin{aligned} {\cal{M}}'= &\ & 16 t^2M_{W'}^2\Bigl[M_{W'}^2(s - M_W^2) - t(s - M_{W'}^2)\Bigr]\nonumber \\ &+& 4 t M_{W'}^2\Bigl[3 M_{W'}^4M_W^2 - 3t ( s + M_{W'}^4) - s(M_{W'}^4 +s M_W^2 + s^2)\Bigr]\nonumber\\ &+& 4 M_{W'}^2\Bigl[M_{W'}^2(s^2 + M_{W'}^4) (t-M_W^2) - 2sM_{W'}^2M_W^2 (t-M_{W'}^2 ) - 2t^4\Bigr]\label{M'}\\ &+& t^2 M_W^2\Bigl[2s(M_{W'}^2 + M_W^2) - t(t+s-9M_{W'}^2) + M_W^2(t-M_{W'}^2)\Bigr].\nonumber\end{aligned}$$ Due to the destructive interference of the two graphs in Fig. 2, there are no terms in (\[M’\]) proportional to $1/M_W^2$. This preserves unitarity. A simple analytic expression for $\Gamma(W'\to \bar\ell\ell W)$ can be obtained in the large $M_{W'}$ limit. The result is $$\begin{aligned} \Gamma(W'\to \bar\ell\ell W) = {2g^2\Gamma(W'\to \bar ff') \over 192\pi^2(1 + 3 \cot^4\phi)} \Bigl[\Bigl(\ln{M_{W'}^2\over M_W^2}\Bigr)^2 - 5 \ln{M_{W'}^2\over M_W^2} - {\pi^2\over 3} +{37\over 3} +{\cal{O}}\Bigl({M_W^2\over M_{W'}^2}\Bigr)\Bigr].\label{GammaW'}\end{aligned}$$ The double log term in (\[GammaW’\]) arises from the interference of the two graphs in Fig. 2 in the kinematic region in which $W$ is soft. For $M_{W'}\sim 1$ TeV numerical evaluation of $\Gamma(W'\to \bar\ell\ell W)$ using (\[rate\]) and the analytic formula (\[GammaW’\]) are in excellent agreement with less than a few percent difference. Although the rare decay $W'\to \bar\ell\ell W$ is suppressed by a factor of $\alpha/2\pi$ compared to $W'\to \bar\ell\nu_{\ell}$, the double log factor provides an enhancement. The observation of this logarithmic enhancement has led to a series of diagnostic studies [@rfxiv; @PACOF; @HR] on the properties of $Z'$. The origin of these log factors is related to the infrared and collinear singularities of $S$-matrix elements, and is well known in QED and QCD. To compare $W'\to \bar\ell\ell W$ with the basic process $W'\to \bar\ell\nu_{\ell}$ we define a ratio $$\begin{aligned} R_{\rm lep}= {B(W'\to \bar\ell\ell W)\over B(W'\to \bar\ell\nu_{\ell})}.\label{ratio}\end{aligned}$$ We plot the distribution of $R_{\rm lept}$ with respect to the invariant mass of the charged lepton pair for $M_{W'}=1$ TeV in Fig. 3. In accordance with the aforementioned logarithmic enhancement, the distribution is clearly dominated by configurations in which the dilepton invariant mass is large, implying that the $W$ is soft and/or collinear. The number of events for the process $pp\to W^{\prime\pm}\to \bar\ell\ell W^{\pm}$ in the narrow width approximation of $W'$ is given by ${\cal{L}}\sigma(pp\to W^{\prime\pm})B(W^{\prime\pm}\to\bar\ell\ell W^{\pm})$. The results along with their typical statistical errors are summarized in Table 2. Again, they should be contrasted with number zero for a right-handed $W'$. Due to the large $W'$ gauge coupling the numbers for the $W'$ rare decay are about one order of magnitude larger than those of the corresponding $Z'$ decays [@rfxiv; @PACOF; @HR]. The signal of the production of $W'$ followed by the rare decay $W'\to \bar\ell\ell W$ is very clean. The major background comes from the process $pp\to W'\to WZ$, with $Z$ decaying into a charged lepton pair. Although in the present model the coupling for the interaction $W'\to WZ$ is suppressed by $M_W^2/M_{W'}^2$, this background can be significant because of the enhancement of $W'$ decaying into longitudinally polarized $W$ and $Z$. However, the background events can cleanly be eliminated by requiring the invariant mass of the charged lepton system to be bigger than $100$ GeV. This cut has been built into the numerical calculation. The loss of the signal associated with this cut is insignificant (a few percent), as expected from the kinematic distribution of $R_{\rm lep}$ (Fig. 3). Another source of background comes from the standard model process $pp\to WZ\to W\ell^+\ell^-$. However, requiring the $WZ$ invariant mass to be equal to $M_{W'}\pm 10$ GeV already puts the total cross section $\sigma(pp\to WZ)$ below $\sigma(pp\to W')$ for $M_{W'}\sim (1- 3 )$ TeV. The background from the $WZ$ production is thus eliminated by employing the dilepton invariant mass cut. In this paper we have shown that the processes $pp\to W'W$ and $W'\to \bar\ell\ell W$ can be useful tests of whether a $W'$ has a $V-A$ or $V+A$ coupling at future hadron colliders. To illustrate the idea we have considered a specific example in which the $W'$ couples to $V-A$ currents. For $M_{W'}\sim (1 - 3)$ TeV, it is shown that the LHC and SSC can produce sufficient numbers of events from these processes for such a left-handed $W'$. On the other hand, the absence of such events would be a clean signal that the $W'$ is right-handed. In addition, the rates for the above processes allow for a determination of the relative strength of the $V-A$ gauge coupling of $W'$ to quarks and lepton. #### Acknowledgments : {#acknowledgments .unnumbered} This work was supported in part by the U.S. Department of Energy Grand No. DE-AC02-76-ERO-3071, and an SSC Fellowship (J.L.) from Texas National Research Laboratory Commission. [\[00\]]{} M. Cvetič and P. Langacker, Phys. Rev. D[**42**]{}, 1797 (1990). P. Langacker and M. Luo, Phys. Rev. [**D45**]{}, 278 (1992); U. Amaldi, [*et al.*]{}, Phys. Rev. [**D36**]{}, 1385 (1987); L. S. Durkin and P. Langacker, Phys. Lett. [**166B**]{}, 436 (1986). The sensitivity of future precision experiments is discussed in P. Langacker, M. Luo and A. K. Mann, Rev. Mod. Phys. [**64**]{}, 87 (1992). G. Altarelli, [*et al.*]{}, Phys. Lett. [**B263**]{}, 459 (1991), [*ibid.*]{} [**B261**]{}, 146 (1991); J. Layssac, F. M. Renard, and C. Verzegnassi, Z. Phys. [**C53**]{}, 97 (1992); F. M. Renard and C. Verzegnassi, Phys. Lett. [**B260**]{}, 225 (1991). K. T. Mahanthappa and P. K. Mohapatra, Phys. Rev. D [**43**]{}, 3093 (1991); P. Langacker, Phys. Lett. [**B256**]{}, 277 (1991). F. del Aguila, W. Hollik J. M. Moreno, and M. Quiros, Nucl. Phys. [B372]{}, 3 (1992); F. del Aguila, J. M. Moreno, and M. Quiros, Nucl. Phys. [**B361**]{}, 45 (1991); Phys. Lett. [**B254**]{}, 479 (1991); M. C. Gonzalez-Garcia and J. W. F. Valle, Phys. Lett. [**B259**]{}, 365 (1991). J. C. Pati and A. Salam, Phys. Rev. Lett. [**31**]{}, 661 (1973); Phys. Rev. [**D10**]{}, 275 (1974); R. Mohapatra and J. C. Pati, Phys. Rev. [**D11**]{}, 566 (1975). G. Beall, M. Bander, and A. Soni, Phys. Rev. Lett. [**48**]{}, 848 (1982). P. Langacker and U. Sankar, Phys. Rev. [**D40**]{}, 1569 (1989). D. London and D. Wyler, Phys. Lett. [**B232**]{}, 503 (1989). F. del Aguila, M. Quiros, and F. Zwirner, Nucl. Phys. B [**284**]{}, 530 (1987); P. Kalyniak and M. Sundaresan, Phys. Rev. D [**35**]{}, 75 (1987). N. Deshpande, J. Gunion, and F. Zwirner, in the [*Experiments Detectors and Experimental Areas for the Supercollider*]{}. Proceedings of the Workshop, Berkeley, California, 1987, edited by R. Donaldson and M. G. D. Gilchriese (World Scientific, Singapore, 1988); N. Deshpande, J. Grifols, and A. Méndez, Phys. Lett. [**B208**]{}, 141, (1988). F. del Aguila, L. Ammetller, R. Field, and L. Garrido, Phys. Lett. [**B201**]{}, 375 (1988); Phys. Lett. [ **B 221**]{}, 408 (1989). G. Kane and C. P. Yuan, Phys. Rev. [**D40**]{}, 2231 (1989). M. J. Duncan and P. Langacker, Nucl. Phys. [**B277**]{}, 285 (1986). V. Barger, [*et al.*]{}, Phys. Rev. [**D35**]{}, 2893 (1987). L. S. Durkin and P. Langacker, Phys. Lett. [**B166**]{}, 436 (1986); F. del Aguila, M. Quiros, and F. Zwirner, Nucl. Phys. [**B287**]{}, 419 (1987); [**284**]{}, 530 (1987); J. Hewett and T. Rizzo in [*Proceedings of the 1988 Snowmass Summer Study on High Energy Physics in the 1990’s*]{}, Snowmass, CO 1988; P. Chiappetta [*et al.*]{}, in the [*Proceedings of the Large Hadron Collider Workshop*]{}, Aachen, Germany, 1990, edited by G. Jarlskog and D. Rein (CERN Report No. 90-10, Geneva, Switzerland, 1990). J.  Hewett and T.  Rizzo, Phys. Rev. [**D45**]{}, 161 (1992). T. Rizzo, Phys. Lett. [**B192**]{}, 125 (1987). M. Cvetič and P. Langacker, Phys. Rev. D [**46**]{}, R 14 (1992). F. del Aguila, B. Alles, L. Ametller and A. Grau, Phys. Rev. D [**48**]{}, 425 (1993). J. Hewett and T. Rizzo, Phys. Rev. D [**47**]{}, 4981 (1993). M. Cvetič, B. Kayser, and P. Langacker, Phys. Rev. Lett. [**68**]{}, 2871 (1992). M. Cvetič and P. Langacker, Phys. Rev. D [**46**]{}, 4943 (1992). F. del Aguila, M. Cvetič, and P. Langacker, UPR-0547T (1993), (to appear in Phys. Rev. D (Rapid Communication)). P. Langacker, R. Robinett, and J. Rosner, Phys. Rev. [**D30**]{}, 1470 (1984). H. Georgi, E. E. Jenkins and E. H. Simmons, Phys. Rev. Lett. [**62**]{}, 2789 (1989); [*ibid*]{} [**63**]{}, 1540 (1989) (E); Nucl. Phys. [**B331**]{}, 541 (1990). E. Eichten, I. Hinchliffe, K. Lane, and C. Quigg, Rev. Mod. Phys. [**56**]{}, 579 (1984). **Figure Captions** Feynman diagrams for the associated production process $pp\to W'W$. \[f:production\] Feynman diagrams for the rare decay process $W'\to \bar\ell\ell W$. \[f:raredecay\] Distribution of $R_{\rm lep}$ defined in (\[ratio\]) with respect to the invariant mass of the charged lepton pair. \[f:distribution\] ---------- ----------- -------------- ---------------- --------------- Collider $W'$ mass $(TeV)$ $\phi=\pi/8$ $\phi=\pi/4$ $\phi=3\pi/8$ $1$ $1514\pm 39$ $6681\pm 82$ $4213\pm 65$ SSC $2$ $131\pm 11$ $579\pm 24$ $365 \pm 19$ $3$ $29\pm 5$ $128\pm 11$ $81\pm 9$ $1$ $3200\pm 57$ $14120\pm 119$ $8907\pm 94$ LHC $2$ $146\pm 12$ $644\pm 25$ $406\pm 20$ $3$ $15\pm 4$ $67\pm 8 $ $42\pm 6$ ---------- ----------- -------------- ---------------- --------------- [**Table 1.**]{} Number of events of the process $pp\to W'W$, with $W'$ subsequently decaying into leptons ($e$ and $\mu$), at the SSC and LHC. The errors are statistical. ---------- ----------- -------------- -------------- --------------- -------------- -------------- --------------- Collider $W'$ mass $(TeV)$ $\phi=\pi/8$ $\phi=\pi/4$ $\phi=3\pi/8$ $\phi=\pi/8$ $\phi=\pi/4$ $\phi=3\pi/8$ $1$ $384\pm 20$ $1694\pm 41$ $1068\pm 33$ $219\pm 15$ $967\pm 31$ $611\pm 25$ SSC $2$ $ 79\pm 9 $ $348\pm 19$ $220\pm 15$ $ 40\pm 6 $ $175\pm 13$ $111\pm 11$ $3$ $ 24\pm 5 $ $107\pm 10$ $ 68\pm 8 $ $ 11\pm 3 $ $ 49\pm 7 $ $ 31\pm 6 $ $1$ $1065\pm 33$ $4703\pm 69$ $2965\pm 54$ $ 513\pm 23$ $2263\pm 48$ $1427\pm 38$ LHC $2$ $ 128\pm 11$ $566\pm 24$ $ 357\pm 19$ $ 51\pm 7$ $225\pm 15$ $ 142\pm 12$ $3$ $ 21\pm 5$ $ 95\pm 10$ $ 60\pm 8$ $ 7\pm 3$ $ 33\pm 6$ $ 21\pm 5$ ---------- ----------- -------------- -------------- --------------- -------------- -------------- --------------- [**Table 2.**]{} Number of events of the process $pp\to W^{\prime\pm}\to \bar\ell\ell W^{\pm}$ at the SSC and LHC, where $\ell=e$ and $\mu$. The errors are statistical.

ArXiv

--- abstract: 'Resonant tunneling between fractional quantum Hall edge states is studied in the Luttinger liquid picture. For the Laughlin parent states, the resonance line shape is a universal function whose width scales to zero at zero temperature. Extensive quantum Monte Carlo simulations are presented for $\nu = 1/3$ which confirm this picture and provide a parameter-free prediction for the line shape.' address: - ' [\[1\]]{} Department of Physics, Indiana University, Bloomington, Indiana 47405 ' - ' [\[2\]]{} Department of Physics, University of Pennsylvania, Philadelphia, PA 19104 ' - ' [\[3\]]{} IBM Research, T.J. Watson Research Cent., PO Box 218, Yorktown Heights, NY 10598 ' author: - 'K. Moon\[1\], H. Yi\[2\], C.L. Kane\[2\], S.M. Girvin\[1\], Matthew P.A. Fisher\[3\]' title: Resonant Tunneling Between Quantum Hall Edge States --- Many body correlations often play an important role in tunneling and resonant tunneling in mesoscopic structures, such as quantum dots[@averin; @meirwingreenlee]. In addition to the Coulomb correlations in the vicinity of a tunneling structure, it has recently been emphasized that the electron interactions in the “leads” feeding the structure can be crucial. Specifically, when the leads are one-dimensional, interactions are believed to de-stablize the Fermi liquid, forming a Luttinger liquid. In this case, the tunneling conductance through a point contact is predicted to vanish as a power law of temperature, due to the absence of long-lived single-particle excitations in the incident electron gas[@kanefisherprl]. Unfortunately, in real one-dimensional wires extraneous impurities away from the point contact will complicate matters, tending to backscatter and localize the electrons. In this article, we consider the situation where the “leads” feeding a point contact are 2d fractional quantum Hall fluids. We argue that the correlations built into Laughlin states play a crucial role in determining the nature of resonant tunneling at low temperatures. In particular, we show that as the temperature is lowered, resonances become sharper and have a universal shape. We present the results of extensive quantum Monte Carlo simulations which give the universal line shape expected for the $\nu =1/3$ state. Provided the temperature is below the gap in the fractional quantum Hall fluid, the incoming current will be carried by edge states. In a beautiful series of articles, Wen[@WEN] has demonstrated that the gapless edge excitations of a quantum Hall system are “chiral” Luttinger liquids. Since the electrons in a chiral Luttinger liquid move only in one direction, backscattering is only possible when opposite edges of the sample are close together, namely at the point contact. Thus localization in the edge state “leads” is completely unimportant. Recently, two of us have studied the problem of resonant tunneling in a Luttinger liquid[@kaneandfisher]. A renormalization group analysis reveals that the resonance lineshape at low temperatures is universal, depending only on the two terminal conductance, $G = g e^2/h$ of the Luttinger liquid. The parameter $g$ depends in a complicated way on the strength of the interactions between left- and right-moving electrons. In a quantum Hall edge state the electrons move only in one direction. Most strikingly, in this case $g$ is a topological invariant controlled by the quantum Hall state in the bulk[@WEN]. To see this, consider raising the chemical potential of the right-movers relative to that of the left by an amount $\delta \mu$. This corresponds to applying a Hall voltage $V_{\rm H} = \delta\mu/e$ and the resulting current is given by the quantized Hall coefficient as $I = \nu (e^2/h) V_{\rm H}$. This immediately establishes the universal result within a given Hall plateau: $g = \nu$. [*This remarkable fact makes the resonance line shape completely universal, model-independent and fully determined (up to an overall temperature scale)*]{}. The fractional quantum Hall regime ($g<1$) is thus a far more promising place to observe pure Luttinger liquid behavior than in a single-channel quantum wire, where it is difficult to eliminate disorder and where the value of $g$ is unknown. Chamon and Wen[@chamonandwen] have recently considered a theory of resonant tunneling between quantum Hall edge states which is valid in the limit in which the peak conductance of the resonances are much less than $e^2/h$, or in the tails of a stronger resonance. A similar theory for resonant tunneling in a Luttinger liquid has been developed by Furusaki and Nagaosa[@furusakinagaosa]. In contrast, the scaling theory presented below is valid over the entire width of the resonance for resonances whose peak conductance approaches the ‘perfect’ value $\nu e^2/h$. The analog of a weak impurity that causes back-scattering is a narrow constriction which brings the left- and right-movers close enough together to communicate via tunneling of Laughlin quasiparticles through a ‘weak link’ as illustrated in Fig.(\[fig1\]). This can be achieved by means of a gate which electrostatically defines a narrow region in the Hall bar. The analog of the two-impurity resonance geometry considered by Kane and Fisher[@kaneandfisher] would be two nearby, symmetric tunneling paths for quasiparticles[@geometry]. For some value of magnetic field or gate voltage one will (randomly or intentionally) achieve the condition of destructive interference[@interference] which shuts off the inter-edge quasiparticle tunneling. This is the resonance (no backscattering) condition which will be manifested experimentally by the appearance of a two-terminal source-drain conductance[@fourterm] which peaks at a value which at low temperatures approaches the quantized value, $G = \nu e^2/h$. Away from resonance the quasi-particle tunneling causes current to leak from one edge to the other, thereby reducing the source to drain conductance (see Fig.(\[fig1\]). In fact, in the fractional quantum Hall effect the quasiparticle tunneling is expected to increase upon cooling, driving the conductance all the way to zero in the zero temperature limit. Following the analysis of ref. , the weak quasiparticle tunneling regime can be analyzed perturbatively. As the temperature is lowered and the effective tunneling amplitude grows, perturbative analysis fails. However at low enough temperatures the conductance from source to drain becomes tiny and a perturbative analysis in terms of electron transmission becomes possible. The peak and tail regimes of the resonance are tied together by a universal scaling function. Below we establish a model and calculate the scaling function for $g=1/3$ using quantum Monte Carlo. We begin our analysis by briefly reviewing the logic behind Wen’s edge state theory. For simplicity we focus here on the primary Hall states with inverse filling factor $\nu^{-1}$ equal to an odd integer. In this case the edge state has only one branch. Conservation of electron three-current $j_\mu$ permits us to introduce a fictitious gauge field $a_\mu$ via $$j_\mu = {1 \over {2 \pi}} \epsilon_{\mu \nu \lambda} \partial_\nu a_\lambda . \label{2}$$ The bulk 2D electron gas is in an incompressible quantum Hall state with an excitation gap, which means that the low-energy, long-length-scale physics must be described by a massive theory. In 2+1-D the only massive gauge theory is the Chern-Simons theory which has (Euclidean) action[@zee] (ignoring irrelevant terms): $$S_{bulk} = {i \over {4 \pi \nu }} \int a_\mu \partial_\nu a_\lambda \epsilon_{\mu \nu \lambda} d^2 x d \tau , \label{3}$$ The coefficient $\nu^{-1}$ is uniquely fixed by the quantized Hall conductivity and specifies the number of zeros bound to the electrons in the Laughlin wave function[@smgandmacd; @kivelsonzhang; @leeandfisher] Wen has shown that in the presence of a boundary, say at y=0, an effective action for the edge state can be obtained as follows[@WEN]: First integrate out $a_\tau$ in the bulk, which gives an incompressibility constraint on the electron density, $\epsilon_{ij} \partial_i a_j =0$. Then solve the constraint in terms of a scalar field, $a_j = \partial_j \phi$. After an integration by parts the final Euclidean action for the edge state takes the form, $$S_{\rm edge} = {1 \over {4 \pi g}} \int dx d \tau (\partial_x \phi) ( i \partial_\tau \phi + v \partial_x \phi ) . \label{4}$$ Here $v$ is the velocity of the edge excitation, which is non-universal, and will depend on the details of the edge confining potential and the Coulomb interaction at the edge. The dimensionless parameter $g$ on the other hand is [*universal*]{} and depends only on the quantum Hall state, $|g| = \nu$. As emphasized by Wen the requirement that the Hamiltonian associated with eq.(\[3\]) be bounded below requires that $v/g$ be positive. This analysis neglects the effects of a long range Coulomb interaction which are expected to become relevant at very low temperatures (roughly $T<10mK$ for a Hall bar of length $100 \mu{\rm m}$ and width $1\mu{\rm m}$) [@WEN; @Glazman]. It follows from eq.(\[2\]) that the one-dimensional electron density along the edge, $\rho (x)$, is given by $\rho = \partial_x \phi /2 \pi$. An expression for the electron creation operator at the edge can be obtained by combining this with the fact that the momentum operator conjugate to $\phi$ is $\Pi_\phi = \rho /g$. Since adding an extra electron to the edge is equivalent to creating an ‘instanton’ in $\phi$, in which $\phi$ changes by $2 \pi$ in the region near $x$, the electron creation operator at the edge is simply $$\psi (x) \sim \exp\left[2\pi i \int^x \Pi_\phi (x') dx'\right] = e^{i \phi(x)/g} . \label{5}$$ A ‘vortex’ or Laughlin quasiparticle at the edge is created simply by $e^{i \phi(x)}$, which carries fractional charge $ge$. We now suppose that the right-moving and left-moving edges are coupled via a tunneling term, say at $x=0$. The total action will then have the form, $S_L[\phi_L] + S_R[\phi_R] + V(\phi_L,\phi_R)$. Lacking a specific model we can not say whether quasiparticles or electrons will tunnel more easily. Presumably the fractionally-charged quasiparticles see a lower barrier, but the matrix elements may compensate. Instead, we write down the most general form allowed by symmetry. Taking the weak link region to be at $x=0$, we have $$V = \sum_{m=1}^\infty v_m \exp\left[im(\phi_L(x=0)-\phi_R(x=0))\right] + {\rm c.c.}, \label{7}$$ where the $v_m$ are (complex) tunneling amplitudes. The term $m=1$ represents the combined amplitudes for a quasi-electron to tunnel from one edge to the other [*or*]{} a quasi-hole to tunnel in the opposite direction. These physically distinct processes lead to the same final state and hence add coherently to produce $v_1$. The term $m=1/g$ corresponds to electron tunneling. We have no [*a priori*]{} knowledge of the $v_m$. Fortunately, for $g=1/3$, all terms except $v_1$ are irrelevant, having a negative renormalization group eigenvalue, $1-gm^2$, [@kaneandfisher]. Thus at low enough temperatures ($T\ll\Delta$, where $\Delta \sim 1 {\rm K}$ is the bulk excitation gap) and small enough $v_1$, the irrelevant variables $v_m, m> 1$, will flow to zero before $v_{\rm eff} \sim v_1/T^{1-g}$ has grown large. Thus the RG flow will follow a [*universal*]{} trajectory away from the resonance fixed point ($v_1=0$) into the insulating fixed point. At finite temperature the renormalization group flows will be cut off, and the system will end up somewhere along that universal trajectory. From this it follows that in the limit of low temperature, the conductance as a function of the resonance tuning parameter $\delta\sim |v_1|$ and the temperature will obey the scaling form, $$G(T,\delta) = \tilde G_g(c\delta/T^{1-g}), \label{8}$$ The scaling function $\tilde G_g(X)$ is [*universal*]{} in the sense that it does not depend on microscopic details, but is a property of the universal trajectory connecting the two fixed points. Since $g=\nu$ in the quantum Hall effect, $\tilde G_g$ is completely determined by the theory. The parameter $c$ is a non-universal dimensionful factor which sets the temperature scale. By demanding that the scaling form (6) matches onto the off-resonance conductance, which vanishes as $G(T) \sim T^{2/g -2}$, implies that the tails of the scaling function should decay like $X^{-2/g}$, or $X^{-6}$ for $g=1/3$. Though the general properties of the scaling function, such as the temperature dependence of the width and the exponent in the tails are known, the detailed shape of the scaling function has been calculated analytically[@kaneandfisher] only for $g=1/2$. This problem is idealy suited to Monte Carlo simulation, and we have explicitly computed $\tilde G(X)$, verified the predicted scaling behavior, and determined the entire scaling function. Following ref. we note that the action is gaussian in $\phi(x)$ for $x\ne 0$ and so we integrate out all degrees of freedom except $\phi(\tau)\equiv \phi_L(x=0,\tau)-\phi_R(x=0,\tau)$. This gives the action, $$S = {1\over {4\pi |g|}} \sum_{i\omega_n} |\omega_n| |\phi(\omega_n)|^2 + v_1 \int_0^\beta d\tau \cos \phi(\tau), \label{9}$$ where we have retained only the single relevant operator. We have computed the finite frequency ‘two-terminal’ conductance from the Kubo formula, $$G(\omega_n) = {e^2\over 2\pi h} |\omega_n| < |\phi(\omega_n)|^2 >. \label{10}$$ A hard cutoff $\Lambda$ is introduced by keeping only a finite number of Matsubara frequencies L (typically $L <100$). We also simulated a dual version of the model in which the tunneling events are represented by a plasma of logarithmically interacting ‘charges’[@kaneandfisher]. Essentially identically results were obtained in the two approaches. In order to extract information about the temperature dependence of the D.C. conductance, analytic continuation to zero frequency is necessary. Though difficult to do exactly, this may be done with sufficient accuracy by fitting the finite frequency data to a rational function \[2/3\] Padé form in order to extrapolate to $\omega=0$. In order to check the code and to test our analytic continuation procedure, we compare our results to an exact solution which is available for $g=1/2$ in Fig.(\[fig2\]) The solid line is the exact solution derived in ref. , and the data points correspond to Monte Carlo simulations of $G$ as a function of $cv_1/T^{1-g}$ for different system sizes corresponding to $T=\Lambda/(31\pi)$ and $T=\Lambda/(41\pi)$. $c$ is the single non-universal parameter which is adjusted to obtain the fit. In Fig.(\[fig3\]) we display the results of our Monte Carlo simulation for $g=1/3$. In this case the tails of the resonance are predicted to decay much faster, like $X^{-6}$. The data clearly scales. If the coefficient $c$ in eq.(\[9\]) is chosen such that the scaling function varies as $\tilde G(X) = g(1-X^2)$ for small $X$, this simulation allows us to determine that for large $X$ $\tilde G(X) = K X^{-6}$ with $K = 2.6\pm 0.2$. It should be emphasized that this scaling behavior is to be expected for the fractional quantum Hall effect $\nu = 1/3$, but not for the integer effect $\nu=1$ or higher-order fractions. In the integer case the edge state is equivalent to a non-interacting Fermi-liquid, and at low temperatures the resonances should be temperature-independent and Lorentzian. For all higher-order fractions in Laughlin’s sequence, $1/\nu$ an odd integer, multi-quasiparticle backscattering processes, $v_m$ in (5) for $m>1$, are also relevant and grow at low temperatures. Thus in these cases the conductance at the peak of the “resonance” ($v_1=0$) will decrease upon cooling, eventually killing completely the resonance in the zero temperature limit. The higher order hierarchical quantum Hall fluids, such as $\nu=2/3,2/5,2/7,...$, have more than one branch of edge states[@WEN; @macdonald], which complicates the analysis. Resonant tunneling between hierarchical edges will be considered in a subsequent paper. For a review, see D.V. Averin and K.K. Likharev, in Mesoscopic Phenomena in Solids, edited by B. L. Altshuler, P.A. Lee and R.A. Webb (Elsevier, Amsterdam, 1990). Y. Meir, N.S. Wingreen and P.A. Lee, Phys. Rev. Lett. [**66**]{}, 3048 (1991). C.L. Kane and M.P.A. Fisher, Phys. Rev. Lett. [**68**]{}, 1220 (1992). X.G. Wen, Phys. Rev. B [**43**]{}, 11025 (1991); Phys. Rev. Lett. [**64**]{}, 2206 (1990). X.G. Wen, Phys. Rev. B [**44**]{} 5708 (1991). C. de C. Chamon and X. G. Wen (preprint, 1993). A. Furusaki and N. Nagaosa, Phys. Rev. B [**47**]{}, 3827 (1993). C. L. Kane and M. P. A. Fisher, Phys. Rev. B[**46**]{}, 7268 (1992); [*op. cit.*]{}, 15233 (1992). These tunneling paths (weak links) may be due to random impurity potentials or an intentionally created quantum dot. If the tunneling occurs through a dot with significant charging energy U, or the tunneling matrix element between the dot from the leads is very small, a more careful analysis is required in order to see how the resonance fixed point is reached. This has been studied in detail in ref. . Since the Laughlin quasiparticles are vortices, the relative phase for the tunneling on the two weak links is given by the number of enclosed electrons in the area between the two links. This is adjustable by means of the B field and possibly electrostatically. Or equivalently the four-terminal conductance $G^{13}$ with the voltage probes on opposite sides of the sample and straddling the weak-link region as shown in Fig.(\[fig1\]). J. Frölich and A. Zee, Nucl. Phys. B[**364**]{}, 517 (1991). S. M. Girvin and A. H. MacDonald, Phys. Rev. Lett. [**58**]{}, 1252 (1987). S-C. Zhang, T. H. Hansson and S. Kivelson, Phys. Rev. Lett. [**62**]{}, 82 (1989). D. H. Lee and M. P. A. Fisher, Phys. Rev. Lett. [**63**]{}, 903 (1989). L.I. Glazman, I.M. Ruzin and B.I. Shklovskii, Phys. Rev. B [**45**]{} 8454 (1992). A. H. MacDonald, Phys. Rev. Lett. [**64**]{}, 222 (1990); M. D. Johnson and A. H. MacDonald, Phys. Rev. Lett. [**67**]{}, 2060 (1991).

ArXiv

[**Interdisciplinary Monte Carlo Simulations**]{} D. Stauffer Institute for Theoretical Physics, Cologne University, D-50923 Euroland [Biological, linguistic, sociological and economical applications of statistical physics are reviewed here. They have been made on a variety of computers over a dozen years, not only at the NIC computers. A longer description can be found in [@newbook], an emphasis on teaching in [@europe].]{} Introduction ============ The Monte Carlo methods invented for physics problems half a century ago were later also applied to fields outside of physics, like economy [@stigler], biology [@kauffman], or sociology [@schelling]. Instead of atoms one simulates animals, including people. These physics methods are often called “independent agents” when applied outside physics, to distinguish them from “representative agent” approximations and other mean field theories. “Emergence” in these fields is what physicists call self-organization, that means systems of many simple particles showing complex behaviour (like freezing or evaporating) which is not evident from the single-particle properties. The three people cited in Refs.3-5 were not physicists; two got the economics Nobel prize. But also physicists have entered these fields intensively in the last years (and much earlier for biology; see Erwin Schrödinger’s question: What is life?). The German Physical Society has since several years a working group on socio-economic problems, started by Frank Schweitzer. And our university just got approved a new Special Research Grant (SFB) where geneticists and theoretical physicists are supposed to work together. The NIC Research Group in Jülich is an earlier physics-biology example. An important difference between physics and applications outside physics is the thermodynamic limit. A glass of Cologne beer has about $10^{25}$ water molecules, which is close enough to infinity for physicists. Economists, in contrast, are less interested in stock markets with $10^{25}$ traders. Thus finite-size effects, which often are a nuisance in Statistical Physics simulations, may be just what we need outside of physics. Of this large area of computer simulations by physicists for fields outside physics I now select: population genetics, language competition, opinion dynamics, and market fluctuations, mostly following [@newbook; @europe]. Population Genetics =================== Darwinian Evolution is similar to thermal physics in that two effects compete: Mother Nature wants to select the fittest and to minimize energy; but more or less random accidents (mutations in biology, thermal noise or entropy in statistical physics) lead to deviations from ideality, like biological ageing or minimization of the free energy. The following example is ongoing work together with Cebrat, Pȩkalski, Moss de Oliveira and de Oliveira and can be regarded as an improved Eigen quasispecies model. Each individual in the population has a genome, which consists of two bit-strings inherited from the mother and the father, respectively. Each bit-string has $L$ bits with $L$ = 8,16,32,64, as is convenient for Fortran words (byte to integer\*8). A bit set to one means a bad mutation in the DNA, while a zero bit is a healthy gene. All mutations are assumed to be recessive, that means they diminish the survival probability by a factor $x < 1$ if and only if both the paternal and the maternal bit-string have their corresponding bits mutated. At reproduction, the bit-strings in both the father and the mother are mutated with probability $M$ at a randomly selected position; then with probability $R$ they undergo a crossover (recombination) at some randomly selected position (like in genetic algorithms); then the bits neighbouring the crossover point are mutated with probability $M_R$; and finally one bit-string of the mother and one of the father give one child genome, with $B$ such births per iteration and per female. (The mother selects the father at random.) Mutation attempts for an already mutated bit leave this bit unchanged. At each iteration the genetic survival probability is $x^n$ where $n$ is the number of active mutations (bit-pairs set to 1) and $x$ an input parameter. To account for limitations in space and food, as well as for infections from other individuals, additional Verhulst death probabilities proportional to the current number of individuals are applied to both the newborns and at each iteration to the adults. For very small $x$, only mutation-free individuals survive: $n=0$. With growing $x$ the survival chances grow, but so does the mutation load $<n>$ which in turn reduces the survival chances. As a result, for $L = 64$ three different phase transitions can be found in Fig.1: For $0 < x < 0.45$ the population dies out; for $0.45 < x < 0.96$ it survives; for $0.96 < x < 0.98$ it dies out again, and for $0.98 < x < 1$ it survives again. The transitions at 0.45 and 0.96 seem to be first-order (jump in population and load) while the one at 0.98 is second-order (continuous). For $x > 0.98$ all bits of both bit-strings are mutated to one, which allows a simple scaling prediction of the population for general $L$ in agreement with the simulations: Results depend on $x^L$ as seen in Fig.2. For example, the critical point at birth rate $B$ is at $x = (1 + B/2)^{-1/L}$. Real animals get old with increasing age, and that can be simulated with similar techniques. The more complicated Penna bit-string model [@penna] simulates the ageing of individuals and agrees well with the empirical Gompertz law of 1825, that the mortality of adult humans increases exponentially with age [@newbook]. Language Competition ==================== Every ten days on average one human language dies out. Simulations of the bit-string Schulze model are very similar to the above population genetics, with random mutations, transfer of language bits from one language to another, and flight from small to large languages [@schulzestauffer]. The alternative Viviane model [@viviane] simplifies mutation and flight from small to large languages into one process, and ignores transfer. It gives in Fig.3 a wide range of language sizes, i.e. of the number of people speaking one language, from dying languages with only one speaker, to Chinese with $10^9$ speakers. The Schulze model gives a more realistic nearly log-normal shape for this distributions, but not the wide range of language sizes. Both the proper shape and the large size range of reality (bottom part of Fig.3) might come from non-equilibrium statistics. In the last version of the Schulze model, each language (better interpretation: its grammar) is characterized by $F$ features each of which can adopt one of $Q$ different integer values $1,2,...Q$. Each site of a large square lattice is occupied by a person speaking one language. At each iteration, each feature of each person is mutated with probability $p$. This mutation is random with probability $1-q$ while with probability $q$ the corresponding feature from one of the four lattice neighbours is adopted. Also, at each iteration, each person independently, with a probability proportional to $1-x^2$ abandons the whole language and adopts the language of one randomly selected person in the population. In the last version of the Viviane model, each lattice site is either empty of carries a population with a size randomly fixed between 1 and, say, like 127. Initially one lattice site is occupied and all others are empty. Then at each time step one empty neighbour of an occupied site is occupied with a probability proportional to the number of people which can live there. Then this new site adopts the language of one of its four lattice neighbours, with a probability proportional to the size of the language spoken at that neighbour site. However, this adopted language is mutated to a new language with probability inversely proportional to the new size of the adopted language. (This denominator is not allowed to exceed a maximum, set randomly between 1 and, say, 2048.) The whole process ends once the last lattice site has become occupied. Opinion Dynamics ================ Can a single person make a difference in public life? In chaos theory we ask whether a single butterfly in Brazil can influence a hurrican in the Caribbean. Kauffman [@kauffman] asked the analogous question whether a single biological mutation has a minor effect or disturbs the whole genetic network [@kauffman]. Physicists call this damage spreading and ask, for example, how the evolution of an Ising model is changed if one single spin is flipped and otherwise the system, including the random numbers to simulate it, remains unperturbed. This question was discussed [@fortunato; @newbook] for three models: The opportunists of Krause and Hegselmann [@krause], the negotiators of Deffuant et al [@deffuant], and the missionaries of Sznajd [@sznajd]. The opportunists take as their new opinion the average opinion of the large population to which they belong, except that they ignore those who differ too much from their own opinion. Also the negotiators ignore opinions which differ too much from their own; otherwise a randomly selected pair gets closer in their two opinions without necessarily agreeing fully. A randomly selected pair of missionaries, neighbouring on a lattice or network, convinces its neighbours if and only if the two people in the pair have the same opinion. Simulations show that the opinion change of a single person may influence the whole population for suitable parameters [@fortunato; @newbook]. For the missionaries on a scale-free network, simulations agreed nicely with election results in Brazil, apart from fitted scale factors, Fig.4. Market Fluctuations =================== How can we get rich fast by speculating on the stock market? This writer earned about one Heugel (a local currency unit of about $10^4$ Euro) by believing some theory for the Tokyo stock market [@sornette]. Details, of course, are given out only for more JUMP time. Instead this section summarizes the Cont-Bouchaud model of stock market fluctuations [@cont], because it is closest to the pre-existing physics model of percolation. Each site of a large square lattice is either occupied by an investor (with probability $p$), or empty with probability $1-p$. Sets of occupied neighbours are called clusters and are identified with groups of investors which act (buy or sell) together. At each iteration a cluster either buys (with probability $a$), sells (also with probability $a$) or sleeps (with probability $1-2a$). The traded amount is proportional to the number of investors in the trading cluster. The difference between supply and demand drives the market values up and down. This basic model gives on average: i) as many ups as downs on the market; ii) a power-law decay (“fat tail”) for the probability to have a large price change, and with modifications also: iii) volatility clustering (markets have turbulent and calm times), iv) effective multi-fractality, v) sharp peaks and flat valleys for the prices, but no prediction on how the market will move tomorrow. Apart from these nice basic properties also practical applications were made [@ehrenstein]: Does a small “Tobin” tax of a few tenths of a percent on all transactions reduce fluctuations and earn tax revenue without killing the whole market? It does, but apart from more government control over individuals there is another danger which can be simulated: If the tax revenue increases with increasing tax rate, then governments will be tempted to increase this tax again and again (as Germans just saw in fall 2005 and German student may observe in future tuition hikes.) Much better is a maximum of tax revenue at some moderate tax rate; then the government should settle on this moderate tax rate, provided it regards the simulations as reliable. Fig.5 shows that in this model such a desirable maximum exists for some parameters but not for all. Another application is the confirmation that halting the trade when excessive price changes are observed indeed helps to calm the market. Discussion ========== Interdisciplinary applications of physics methods are no longer as exotic as they were years ago; biologists and economists have started to publish papers together with computational physicists on these non-physics fields. Thanks for S. Cebrat, P.M.C. de Oliveira and S. Moss de Oliveira for comments on the manuscript. [10]{} D. Stauffer, S. Moss de Oliveira, P.M.C. de Oliveira and J.S. Sá Martins, [*Biology, Sociology, Geology by Computational Physicist*]{}. Elsevier, Amsterdam 2006 in press. ISBN-13: 978-0-444-52146-0 and ISBN-10: 0-444-52146-1. S. Moss de Oliveira, P.M.C. de Oliveira and D. Stauffer, [*Evolution, Money, War and Computers*]{}, Teubner, Leipzig and Stuttgart 1999. D. Stauffer, Eur. J. Phys. 26, S 79 (2005) and AIP Conf. Proc. 779, 49, 56, 69 and 75 (2005). G.J. Stigler, Journal of Business 37, 117 (1964). S.A. Kauffman, J. Theoretical Biology 22, 437 (1969). T.C. Schelling, J. Mathematical Sociology 1 143 (1971). T.J.P. Penna, J. Stat. Phys. 78, 1629 (1995). C. Schulze and D. Stauffer, Int. J. Mod. Phys. C 16, issue 5 (2005) and Physics of Life Reviews 2, 89 (2005); T. Teşileanu and H. Meyer-Ortmanns, 2006, Int. J. Mod. Phys. C17, issue 3 = physics/0508229. For other models of language competition see D.M.Abrams and S.H. Strogatz, Nature 424, 900 (2003); M. Patriarca and T. Leppänen, Physica A 338, 296. (2004); J. Mira and A. Paredes, Europhys. Lett. 69, 1031 (2005); K. Kosmidis, J.M. Halley and P. Argyrakis, Physica A 353, 595 (2005) and in press; V. Schwämmle, Int. J. Mod. Phys. C 16, 1519 (2005) and 17, issue 3; W.S.Y. Wang, J.W. Minett, Trans. Philological Soc. 103, 121 (2005), A. Baronchelli et al., physics/059075, 0511201 and 0512045. V.M. de Oliveira, M.A.F. Gomes and I.R. Tsang, Physica A, in press = physics/0505197 and 0510249 (2006). W.J. Sutherland, Nature 423, 276 (2003). S. Fortunato and D. Stauffer, page 231 in: [ *Extreme Events in Nature and Society*]{}, edited by S. Albeverio, V. Jentsch and H. Kantz. Springer, Berlin - Heidelberg 2005. R. Hegselmann and U. Krause, Journal of Artificial Societies and Social Simulation 5, issue 3, paper 2 (jasss.soc.surrey.ac.uk) (2002). G. Deffuant, F. Amblard, G. Weisbuch and T. Faure, Journal of Artificial Societies and Social Simulation 5, issue 4, paper 1 (jasss.soc.surrey.ac.uk) (2002). K. Sznajd-Weron and J. Sznajd, Int. J. Mod. Phys. C 11, 1157 (2000). A. T. Bernardes, D. Stauffer and J. Kertész: Eur. Phys. J. B 25, 123 (2002). A. Johansen and D. Sornette, Int. J. Mod. Phys. C 10, 563 (1999). R. Cont and J.P. Bouchaud, eprint cond-mat/9712318 = Macroeconomic Dynamics 4, 170 (2000); D. Stauffer, Adv. Complex Syst. 4, 19 (2001). G. Ehrenstein, F. Westerhoff and D. Stauffer, Quantitative Finance 5, 213 (2005); G. Ehrenstein and F. Westerhoff, Int. J. Mod. Phys. C 17, issue 5 (2006).

ArXiv

--- abstract: | We present a unifying approach to the efficient evaluation of propositional answer-set programs. Our approach is based on backdoors which are small sets of atoms that represent “clever reasoning shortcuts” through the search space. The concept of backdoors is widely used in the areas of propositional satisfiability and constraint satisfaction. We show how this concept can be adapted to the nonmonotonic setting and how it allows to augment various known tractable subproblems, such as the evaluation of Horn and acyclic programs. In order to use backdoors we need to find them first. We utilize recent advances in fixed-parameter algorithmics to detect small backdoors. This implies fixed-parameter tractability of the evaluation of propositional answer-set programs, parameterized by the size of backdoors. Hence backdoor size provides a structural parameter similar to the treewidth parameter previously considered. We show that backdoor size and treewidth are incomparable, hence there are instances that are hard for one and easy for the other parameter. We complement our theoretical results with first empirical results. author: - | Johannes Klaus Fichte and Stefan Szeider\ Vienna University of Technology, Austria\ [email protected], [email protected] title: 'Backdoors to Tractable Answer-Set Programming[^1] [^2]' --- Introduction ============ *Answer-Set Programming* (ASP) is an increasingly popular framework for declarative programming [@MarekTruszczynski99; @Niemela99]. ASP allows to describe a problem by means of rules and constraints that form a disjunctive logic program. Solutions to the program are so-called stable models or answer sets. Many important problems of AI and reasoning can be represented and successfully solved within the ASP framework. However, the main computational problems for ASP (such as deciding whether a program has a solution, or if a certain atom is contained in at least one or in all solutions) are of high worst-case complexity and are located at the second level of the Polynomial Hierarchy [@EiterGottlob95]. The known complexity results do not rule out the possibility for exact algorithms that work efficiently for real-world instances by exploiting the presence of a “hidden structure.” In this paper we follow a new approach of making the vague notion of a hidden structure precise. Our approach is based on the concept of *backdoors* which is widely used in the areas of propositional satisfiability and constraint satisfaction (see, e.g., [@WilliamsGomesSelman03; @GottlobSzeider08; @SamerSzeider08c]), and also for quantified Boolean formulas and argumentation [@SamerSzeider09a; @OrdyniakSzeider11]. A backdoor is a small set of key atoms that represent a “clever reasoning shortcut” through the search space. By deciding the status of the atoms in the backdoor, we can reduce a given program to several tractable programs belonging to a *target class* of programs. Consequently the evaluation of the given program is *fixed-parameter tractable* in the size of the backdoor, i.e., polynomial for fixed backdoor size $k$ where the order of the polynomial is independent of $k$ [@DowneyFellows99]. By allowing backdoors of increasing size $k=1,2,3,\dots$ we can gradually augment a known tractable class of programs. Our results are as follows: - We show that the most important computational problems of propositional answer-set programming, including credulous/skeptical reasoning (and even counting all answer sets) are fixed-parameter tractable in the size of the backdoor. - We show that the detection of backdoors is fixed-parameter tractable for various target classes, including the class of all Horn programs and classes based on various notions of acyclicity. This way we make recent results of fixed-parameter algorithmics accessible to the field of answer-set programming. - We show that the concept of backdoors entails fixed-parameter tractability results for answer-set programming [@Ben-Eliyahu96] and so provides a unifying framework. - We compare backdoor size with respect to various base classes with each other and with the recently studied parameter incidence treewidth [@JaklPichlerWoltran09]. - We present first empirical results where we consider the backdoor size of structured programs and random programs of varied density. Formal Background ================= We consider a universe $U$ of propositional *atoms*. A *literal* is an atom $a\in U$ or its negation $\neg a$. A *disjunctive logic program* (or simply a *program*) $P$ is a set of *rules* of the form $ x_1{\vee}\dots {\vee}x_l \leftarrow y_1,\dots,y_n,{\neg}z_1,\dots,{\neg}z_m $ where $x_1,\dots,x_l, y_1,\dots,y_n, z_1,\dots, z_m$ are atoms and $l,n,m$ are non-negative integers. We write $\{x_1,\dots,x_l\}=H(r)$ (the *head* of $r$) and $\{y_1,\dots,y_n,z_1,\dots,z_m\}=B(r)$ (the *body* of $r$), $B^+(r)= \{y_1,\dots,y_n\}$ and $B^-(r)= \{z_1,\dots,z_n\}$. We denote the sets of atoms occurring in a rule $r$ or in a program $P$ by ${\text{\normalfont at}}(r)=H(r) \cup B(r)$ and ${\text{\normalfont at}}(P)=\bigcup_{r\in P} {\text{\normalfont at}}(r)$, respectively. A rule $r$ is *negation-free* if $B^-(r)=\emptyset$, $r$ is *normal* if ${|H(r)|}=1$, $r$ is a *constraint* if ${|H(r)|}=0$, $r$ is *disjunction-free* if ${|H(r)|}\leq 1$, $r$ is *Horn* if it is negation-free and disjunction-free, and $r$ is *tautological* if $B^+(r) \cap (H(r) \cup B^-(r))\neq \emptyset$. We say that a program has a certain property if all its rules have the property. We denote by ${\text{{\textnormal{\textbf{Horn}}}}}$ the classes of all Horn programs. A set $M$ of atoms *satisfies* a rule $r$ if $(H(r)\cup B^-(r)) \cap M \neq \emptyset$ or $B^+(r) \setminus M \neq \emptyset$. $M$ is a *model* of $P$ if it satisfies all rules of $P$. The *GL reduct* of a program $P$ under a set $M$ of atoms is the program $P^M$ obtained from $P$ by first removing all rules $r$ with $B^-(r)\cap M\neq \emptyset$ and second removing all $\neg z$ where $z \in B^-(r)$ from all remaining rules $r$ [@GelfondLifschitz91]. $M$ is an *answer set* (or *stable set*) of a program $P$ if $M$ it is a minimal model of $P^M$. We denote by ${\text{\normalfont AS}}(P)$ the set of all answer sets of $P$. We also need some notions from *propositional satisfiability*. A *clause* is a finite set of literals, a CNF formula is a finite set of clauses. A *truth assignment* is a mapping $\tau:X\rightarrow \{0,1\}$ defined for a set $X\subseteq U$ of atoms. For $x\in X$ we put $\tau({\neg}x)=1 - \tau(x)$. By ${\text{\normalfont ta($X$)}}$ we denote the set of all truth assignments $\tau:X\rightarrow \{0,1\}$. The *truth assignment reduct* of a CNF formula $F$ with respect to $\tau \in {\text{\normalfont ta($X$)}}$ is the CNF formula $F_\tau$ obtained from $F$ by first removing all clauses $c$ that contain a literal set to $1$ by $\tau$, and second removing from the remaining clauses all literals set to $0$ by $\tau$. $\tau$ *satisfies* $F$ if $F_\tau=\emptyset$, and $F$ is *satisfiable* (in symbols $\text{sat}(F)$) if it is satisfied by some $\tau$. #### ASP Problems The main computational problems for ASP are as follows. <span style="font-variant:small-caps;">Consistency</span>: given a program $P$, does $P$ have an answer-set? <span style="font-variant:small-caps;">Credulous/Skeptical Reasoning</span>: given a program $P$ and an atom $a\in {\text{\normalfont at}}(P)$, is $a$ contained in some/all answer-set(s) of $P$? <span style="font-variant:small-caps;">AS Counting</span>: how many answer sets does $P$ have? <span style="font-variant:small-caps;">AS Enumeration</span>: list all answer sets of $P$. <span style="font-variant:small-caps;">Consistency</span> and <span style="font-variant:small-caps;">Credulous Reasoning</span> are $\Sigma^P_2$-complete, <span style="font-variant:small-caps;">Skeptical Reasoning</span> is $\Pi^P_2$-complete [@EiterGottlob95]. The problems remain ${\text{\normalfont NP}}$ (or ${\text{\normalfont co-NP}}$) hard for normal programs [@MarekTruszczynski91], but are polynomial-time solvable for Horn programs [@GelfondLifschitz88]. <span style="font-variant:small-caps;">AS Counting</span> is easily seen to be $\#P$-hard even for normal programs in view of the $\#P$-completeness of $\#$SAT. #### Fixed-Parameter Tractability We give some basic background on parameterized complexity. For more detailed information we refer to other sources  [@DowneyFellows99; @GottlobSzeider08]. A *parameterized problem* $L$ is a subset of $\Sigma^* \times {\mathbb{N}}$ for some finite alphabet $\Sigma$. For an instance $(I,k) \in \Sigma^* \times {\mathbb{N}}$ we call $I$ the *main part* and $k$ the *parameter*. $L$ is *fixed-parameter tractable* if there exist a computable function $f$ and a constant $c$ such that we can decide whether $(I,k)\in L$ in time $O(f(k){\|I\|}^c)$ where ${\|I\|}$ denotes the size of $I$. ${\text{\normalfont FPT}}$ is the class of all fixed-parameter tractable decision problems. The *Weft Hierarchy* consists of parameterized complexity classes ${\text{\normalfont W[1]}}\subseteq {\text{\normalfont W[2]}}\subseteq \cdots$ which are defined as the closure of certain parameterized problems under parameterized reductions. There is strong theoretical evidence that parameterized problems that are hard for classes ${\text{\normalfont W[xxxx]}}[$i$]$ are not fixed-parameter tractable. Backdoors ========= Before we introduce the notion of backdoors to the ASP domain we review it in the domain where it originates from. #### Satisfiability Backdoors Let $F$ be a CNF formula and $X$ a set of atoms. The following is obvious from the definitions: - $F$ is satisfiable if and only if $F_\tau$ is satisfiable for at least one truth assignment $\tau\in {\text{\normalfont ta($X$)}}$. This observation leads to the definition of a strong backdoor relative to a class ${\mathcal{C}}$ of polynomially solvable CNF formulas: a set $X$ of atoms is a *strong ${\mathcal{C}}$-backdoor* of a CNF formula $F$ if $F_\tau\in {\mathcal{C}}$ for all truth assignments $\tau\in {\text{\normalfont ta($X$)}}$. Assume that the satisfiability of formulas $F\in {\mathcal{C}}$ of size ${\|F\|}=n$ can be decided in time $O(n^c)$. Then we can decide the satisfiability of an arbitrary formula $F$ for which we know a strong ${\mathcal{C}}$-backdoor of size $k$ in time $O(2^k n^c)$ which is efficient as long as $k$ remains small. Before we can use the strong backdoor we need to find it first. For most reasonable target classes ${\mathcal{C}}$ the detection of a strong backdoor of size at most $k$ is ${\text{\normalfont NP}}$-hard if $k$ is part of the input. However, as we are interested in finding small backdoors, it makes sense to parameterized the backdoor search by $k$ and consider the parameterized complexity of backdoor detection. Indeed, with respect to the classes of Horn CNF formulas and 2-CNF formulas, the detection of strong backdoors of size $\leq k$ is fixed-parameter tractable [@NishimuraRagdeSzeider04-informal; @SamerSzeider08c]. For other target classes (clustering formulas and renamable Horn formulas) the detection of deletion backdoors (a subclass of strong backdoors) of size at most $k$ is fixed-parameter tractable [@NishimuraRagdeSzeider07; @RazgonOSullivan08]. #### ASP Backdoors In order to translate the notion of backdoors to the domain of ASP, we first need to come up with a suitable concept of a reduction with respect to a truth assignment. The following is a natural definition which generalizes a concept of [@GottlobScarcelloSideri02 ]{}. Let $P$ be a program, $X$ a set of atoms, and $\tau\in {\text{\normalfont ta($X$)}}$. The *truth assignment reduct* of $P$ under $\tau$ is the logic program $P_\tau$ obtained from $P$ by 1. removing all rules $r$ with $H(r)\cap \tau^{-1}(1)\neq \emptyset$ or $H(r)\subseteq X$; 2. removing all rules $r$ with $B^+(r) \cap \tau^{-1}(0)\neq \emptyset$; 3. removing all rules $r$ with $B^-(r) \cap \tau^{-1}(1)\neq \emptyset$; 4. removing from the heads and bodies of the remaining rules all literals $v,{\neg}v$ with $v\in X$. Let ${\mathcal{C}}$ be a class of programs. A set $X$ of atoms is a *strong ${\mathcal{C}}$backdoor* of a program $P$ if $P_{\tau}\in {\mathcal{C}}$ for all truth assignments $\tau\in {\text{\normalfont ta($X$)}}$. \[ex1\] Consider the program $P=\{ s \leftarrow w;\, u\leftarrow s,q;\, r \leftarrow w,s;\, t \leftarrow {\neg}r;\, q \leftarrow {\neg}s, u;\, w \leftarrow {\neg}r,u \}$.\ The set $X=\{r,s\}$ is a strong ${\text{{\textnormal{\textbf{Horn}}}}}$backdoor since all four truth assignment reducts $P_{r=0,s=0}=P_{00}=\\\{ t \leftarrow;\, q \leftarrow u;\, w \leftarrow u\}$, $P_{01}=\{u \leftarrow q;\, t \leftarrow;\, w \leftarrow u \}$, $P_{10}=\{ q \leftarrow u \}$, and $P_{11}=\{u \leftarrow q\}$ are Horn programs. A direct equivalence similar to ($\ast$) does not hold for ASP, even if we consider the most basic problem <span style="font-variant:small-caps;">Consistency</span>. Take for example the program $P={\{\,}\newcommand{\SM}{\;{:}\;}\newcommand{\SE}{\,\}}x \leftarrow {\neg}x;\, y \leftarrow \SE $ and the set $X=\{x\}$. Both reducts $P_{x=0}={\{\,}\newcommand{\SM}{\;{:}\;}\newcommand{\SE}{\,\}}y \SE$ and $P_{x=1}={\{\,}\newcommand{\SM}{\;{:}\;}\newcommand{\SE}{\,\}}y \SE$ have answer sets, but $P$ has no answer set. However, we can show a somewhat weaker asymmetric variant of ($\ast$), where we can map each answer set of $P$ to an answer set of $P_\tau$ for some $\tau\in {\text{\normalfont ta($X$)}}$. This is made precise by the following definition and lemma. Let $P$ be a program and $X$ a set of atoms. We define $${\text{\normalfont AS}}(P,X) = {\{\,}\newcommand{\SM}{\;{:}\;}\newcommand{\SE}{\,\}}M\cup \tau^{-1}(1) \SM \tau\in {\text{\normalfont ta($X\cap\, {\text{\normalfont at}}(P)$)}}, M \in {\text{\normalfont AS}}(P_\tau)\SE.$$ \[lem:subset\] ${\text{\normalfont AS}}(P) \subseteq {\text{\normalfont AS}}(P,X)$ holds for every program $P$ and every set $X$ of atoms. Let $M\in {\text{\normalfont AS}}(P)$ be chosen arbitrarily. We put $X_0=(X\setminus M)\cap {\text{\normalfont at}}(P)$ and $X_1=X \cap M$ and define a truth assignment $\tau\in {\text{\normalfont ta($X\cap {\text{\normalfont at}}(P)$)}}$ by setting $\tau^{-1}(i)=X_i$ for $i\in \{0,1\}$. Let $M'=M\setminus X_1$. Observe that $M'\in {\text{\normalfont AS}}(P_\tau)$ implies $M\in {\text{\normalfont AS}}(P,X)$ since $M= M'\cup \tau^{-1}(1)$ by definition. Hence, to establish the lemma, it suffices to show that $M'\in {\text{\normalfont AS}}(P_\tau)$. We have to show that $M'$ is a model of $P_\tau^{M'}$, and that no proper subset of $M'$ is a model of $P_\tau^{M'}$. In order to show that $M'$ is a model of $P_\tau^{M'}$, choose $r'\in P_\tau^{M'}$ arbitrarily. By construction of $P_\tau^{M'}$ there is a corresponding rule $r\in P$ with $H(r')=H(r)\setminus X_0$ and $B^+(r')=B^+(r)\setminus X_1$ which gives rise to a rule $r''\in P_\tau$, and in turn, $r''$ gives rise to $r'\in {\text{\normalfont P}}_\tau^{M'}$. Since $B^-(r)\cap X_1=\emptyset$ (otherwise $r$ would have been deleted forming $P_\tau$) and $B^-(r)\cap M'=\emptyset$ (otherwise $r''$ would have been deleted forming $P_\tau^{M'}$), it follows that $B^-(r)\cap M=\emptyset$. Thus $r$ gives rise to a rule $r^*\in P^M$ with $H(r)=H(r^*)$ and $B^+(r)=B^+(r^*)$. Since $M\in {\text{\normalfont AS}}(P)$, $M$ satisfies $r^*$, i.e., $H(r)\cap M\neq \emptyset$ or $B^+(r)\setminus M\neq \emptyset$. However, $H(r)\cap M=H(r')\cap M'$ and $B^+(r)\setminus M = B^+(r')\setminus M'$, thus $M'$ satisfies $r'$. Since $r'\in P_\tau^{M'}$ was chosen arbitrarily, we conclude that $M'$ is a model of $P_\tau^{M'}$. In order to show that no proper subset of $M'$ is a model of $P_\tau^{M'}$ choose arbitrarily a proper subset $N'\subsetneq M'$. Let $N=N'\cup X_1$. Since $M'=M\setminus X_1$ and $X_1\subseteq M$ it follows that $N \subsetneq M$. Since $M$ is a minimal model of $P^M$, $N$ cannot be a model of $P^M$. Consequently, there must be a rule $r\in P$ such that $B^-(r)\cap M= \emptyset$ (i.e., $r$ is not deleted by forming $P^M$), $B^+(r)\subseteq N$ and $H(r)\cap N=\emptyset$. However, since $M$ satisfies $P^M$, and since $B^+(r)\subseteq N\subseteq M$, $H(r)\cap M\neq \emptyset$. Thus $r$ is not a constraint. Moreover, since $H(r)\cap M\neq \emptyset$ and $M\cap X_0=\emptyset$, it follows that $H(r)\setminus X_0\neq \emptyset$. Thus, since $H(r)\cap X_1 =\emptyset$, $H(r) \setminus X \neq \emptyset$. We conclude that $r$ is not deleted when forming $P_\tau$ and giving rise to a rule $r'\in P_\tau$, which in turn is not deleted when forming $P_\tau^{M'}$, giving rise to a rule $r''$, with $H(r'')=H(r)\setminus X_0$, $B^+(r'')=B^+(r)\setminus X_1$, and $B^-(r'')=\emptyset$. Since $B^+(r'')\subseteq N'$ and $H(r'')\cap N=\emptyset$, $N'$ is not a model of $P_\tau^{M'}$. Thus we have established that $M'$ is a stable model of $P_\tau$, and so the lemma follows. In view of Lemmas \[lem:subset\] and \[lem:mincheck\], we can compute ${\text{\normalfont AS}}(P)$ by (i) computing ${\text{\normalfont AS}}(P_\tau)$ for all $\tau\in {\text{\normalfont ta($X$)}}$ (this produces the set ${\text{\normalfont AS}}(P,X)$ of candidates for ${\text{\normalfont AS}}(P )$, and (ii) checking for each $M \in {\text{\normalfont AS}}(P,X)$ whether it is an answer-set of $P$. The check (ii) entails (iia) checking whether $M\in {\text{\normalfont AS}}(P,X)$ is a model of P and (iib) whether $M\in {\text{\normalfont AS}}(P,X)$ is a minimal model of $P^M$. We would like to note that in particular any constraint contained in $P$ is removed in the truth assignment reduct $P_\tau$ but considered in check (iia). Clearly check (iia) can be carried out in polynomial time for each $M$. Check (iib), however is ${\text{\normalfont co-NP}}$hard in general [@EiterGottlob95], but polynomial for normal programs [@CadoliLenzerini94]. Fortunately, for our considerations it suffices to perform check (iib) for programs that are “close to Horn”, and so the check is fixed-parameter tractable in the size of the given backdoor, as we shall show in the following lemma[^3]. \[lem:mincheck\] Let ${\mathcal{C}}$ be a class of normal programs. Given a program $P$, a strong ${\mathcal{C}}$backdoor set ${X}$ of $P$, and ${M}\in {\text{\normalfont AS}}(P,X)$. Then deciding whether ${M}$ is an answer set of $P$ is fixed-parameter tractable for parameter ${|{X}|}$. In particular, this decision can be made in time $O(2^k n)$ where $n$ denotes the input size of $P$ and $k={|{X}|}$. Let ${\mathcal{C}}$ be a class of normal programs, $P$ a program, and $X$ a strong ${\mathcal{C}}$backdoor set ${X}$ of $P$ with ${|X|}=k$. We can check in polynomial time whether ${M}$ is a model of $P^{M}$. If it is not, we can reject ${M}$, and we are done. Hence assume that ${M}$ is a model of $P^{M}$. In order to check whether $M \in {\text{\normalfont AS}}(P)$ we still need to decide whether ${M}$ is a minimal model of $P^{M}$. We may assume, w.l.o.g., that $P$ contains no tautological rules, as it is clear that the test for minimality does not depend on tautological rules. Let ${X}_1\subseteq {M}\cap {X}$. We construct from $P^{M}$ a program $P^{M}_{{X}_1\subseteq {X}}$ by (i) removing all rules $r$ for which $H(r)\cap {X}_1\neq \emptyset$, and (ii) replacing for all remaining rules $r$ the head $H(r)$ with $H(r)\setminus {X}$, and the positive body $B^+(r)$ with $B^+(r)\setminus {X}_1$. *Claim: $P^{M}_{{X}_1\subseteq {X}}$ is Horn.* To show the claim, consider some rule $r'\in P^{M}_{{X}_1\subseteq {X}}$. By construction, there must be a rule $r\in P$ that gives raise to a rule in $P^M$, which in turn gives raise to $r'$. Let $\tau \in {\text{\normalfont ta(${X}$)}}$ be the assignment that sets all atoms in ${X}\cap H(r)$ to 0, and all atoms in ${X}\setminus H(r)$ to 1. Since $r$ is not tautological, it follows that $r$ is not deleted when we obtain $P_\tau$, and it gives rise to a rule $r^*\in P_\tau$, where $H(r^*)=H(r)\setminus {X}$. However, since ${\mathcal{C}}$ is a class of normal programs, $r^*$ is normal. Hence $1\geq {|H(r^*)|}={|H(r)\setminus {X}|}=H(r')$, and the claim follows. To test whether ${M}$ is a minimal model of $P^M$, we run the following procedure for every set ${X}_1\subseteq {M}\cap {X}$. If $P^{M}_{{X}_1\subseteq {X}}$ has no model, then stop and return TRUE. Otherwise, compute the unique minimal model $L$ of the Horn program $P^{M}_{{X}_1\subseteq {X}}$. If $L \subseteq {M}\setminus {X}$, $L \cup {X}_1 \subsetneq {M}$, and $L \cup {X}_1$ is a model of $P^{M}$, then return FALSE. Otherwise return TRUE. For each set ${X}_1\subseteq {M}\cap {X}$ the above procedure runs in linear time. This follows directly from the fact that we can compute the unique minimal model of a Horn program in linear time [@DowlingGallier84]. As there are $O(2^k)$ sets ${X}_1$ to consider, we have a total running time of $O(2^k n)$ where $n$ denotes the input size of $P$ and $k={|X|}$. It remains to establish the correctness of the algorithm in terms of the following claim. *Claim: ${M}$ is a minimal model of $P^{M}$ if and only if the algorithm returns TRUE for each ${X}_1\subseteq {M}\cap {X}$.* ($\Rightarrow$). Assume that ${M}$ is a minimal model of $P^{M}$, and suppose to the contrary that there is some ${X}_1\subseteq {M}\cap {X}$ for which the algorithm returns FALSE. Consequently, $P^{M}_{{X}_1\subseteq {X}}$ has a unique minimal model $L$ with $L \subseteq {M}\setminus {X}$, $L \cup {X}_1 \subsetneq {M}$, and where $L \cup {X}_1$ is a model of $P^{M}$. This contradicts the assumption that ${M}$ is a minimal model of $P^{M}$. Hence the only-if direction of the lemma is shown. ($\Leftarrow$). Assume that the algorithm returns TRUE for each ${X}_1 \subseteq {M}\cap {X}$. We show that ${M}$ is a minimal model of $P^{M}$. Suppose to the contrary that $P^{M}$ has a model ${M}'\subsetneq {M}$. We run the algorithm for ${X}_1:={M}' \cap {X}$. By assumption, the algorithm returns TRUE. There are two possibilities: (i) $P^{M}_{{X}_1\subseteq {X}}$ has no model, or (ii) $P^{M}_{{X}_1 \subseteq {X}}$ has a model, and for its unique minimal model $L$ the following holds: $L$ is not a subset of ${M}\setminus {X}$, or $L \cup {X}_1$ is not a proper subset of ${M}$, or $L \cup {X}_1$ is not a model of $P^{M}$. We show that case (i) is not possible, by showing that ${M}' \setminus {X}$ is a model of $P^{M}_{{X}_1\subseteq {X}}$. To see this, consider a rule $r'\in P^{M}_{{X}_1\subseteq {X}}$, and let $r\in P^{M}$ such that $r'$ is obtained from $r$ by removing ${X}$ from $H(r)$ and by removing ${X}_1$ from $B^+(r)$. Since ${M}'$ is a model of $P^{M}$, we have (a) $B^+(r) \setminus {M}' \neq\emptyset$ or (b) $H(r)\cap {M}'\neq \emptyset$. Moreover, since $B^+(r')=B^+(r)\setminus {X}_1$ and ${X}_1={M}'\cap {X}$, (i) implies $\emptyset \neq B^+(r) \setminus {M}' = B^+(r) \setminus {X}_1 \setminus {M}' = B^+(r')\setminus {M}' \subseteq B^+(r') \setminus ({M}' \setminus {X})$, and since $H(r)\cap{X}_1=\emptyset$, (ii) implies $\emptyset \neq H(r)\cap{M}'= H(r)\cap ({M}'\setminus {X}_1)= H(r)\cap ({M}'\setminus {X})= (H(r)\setminus {X}) \cap ({M}'\setminus {X})= H(r') \cap ({M}'\setminus {X})$. Hence ${M}' \setminus {X}$ satisfies $r'$. Since $r'\in P^{M}_{{X}_1\subseteq {X}}$ was chosen arbitrarily, we conclude that ${M}' \setminus {X}$ is a model of $P^{M}_{{X}_1\subseteq {X}}$. Case (ii) is not possible either, as we can see as follows. Assume $P^{M}_{{X}_1 \subseteq {X}}$ has a model, and let $L$ be its unique minimal model. Since ${M}'\setminus {X}$ is a model of $P^{M}_{{X}_1 \subseteq {X}}$, as shown above, we have $L\subseteq {M}'\setminus {X}$. We have $L \subseteq {M}\setminus {X}$ since $L\subseteq {M}' \setminus {X}$ and ${M}' \setminus {X}\subseteq {M}\setminus {X}$. Further we have $L \cup {X}_1 \subsetneq {M}$ since $L \cup {X}_1 \subseteq ({M}' \setminus {X}) \cup {X}_1 = ({M}' \setminus {X}) \cup ({M}'\cap {X}) ={M}' \subsetneq {M}$. And finally $L \cup {X}_1$ is a model of $P^{M}$, as can be seen as follows. Consider a rule $r\in P^M$. If ${X}_1 \cap H(r)\neq \emptyset$, then $L\cup {X}_1$ satisfies $r$; thus it remains to consider the case ${X}_1 \cap H(r) = \emptyset$. In this case there is a rule $r'\in P^{M}_{{X}_1 \subseteq {X}}$ with $H(r') = H(r)\setminus {X}$ and $B^+(r')=B^+(r)\setminus {X}_1$. Since $L$ is a model of $P^{M}_{{X}_1 \subseteq {X}}$, $L$ satisfies $r'$. Hence (a) $B^+(r')\setminus L \neq\emptyset $ or (b) $H(r')\cap L \neq \emptyset$. Since $B^+(r')=B^+(r)\setminus {X}_1$, (a) implies that $B^+(r)\setminus ( L\cup {X}_1)\neq \emptyset$; and since $H(r') \subseteq H(r)$, (b) implies that $H(r)\cap (L \cup {X}_1) \neq \emptyset$. Thus $L \cup {X}_1$ satisfies $r$. Since $r\in P^M$ was chosen arbitrarily, we conclude that $L \cup {X}_1$ is a model of $P^M$. Since neither case (i) nor case (ii) is possible, we have a contradiction, and we conclude that ${M}$ is a minimal model of $P^{M}$. Hence the second direction of the claim is established, and so the lemma follows. Thus, in view of Lemmas \[lem:subset\] and \[lem:mincheck\], the computation of ${\text{\normalfont AS}}(P)$ is fixed-parameter tractable for parameter $k$ if we know a strong ${\mathcal{C}}$backdoor set $X$ of size at most $k$ for $P$, and each program in ${\mathcal{C}}$ is normal and its stable sets can be computed in polynomial time. This consideration leads to the following definition and result. A class ${\mathcal{C}}$ of programs is *enumerable* if for each $P\in {\mathcal{C}}$ we can compute ${\text{\normalfont AS}}(P)$ in polynomial time. For any class ${\mathcal{C}}$ of programs we denote by ${\mathcal{C}}^*$ the class containing all programs that belong to ${\mathcal{C}}$ after removal of tautological rules and constraints. It is easy to see that whenever ${\mathcal{C}}$ is enumerable, then so is ${\mathcal{C}}^*$. Note that all classes considered in this paper are enumerable. \[the:evaluation\] Let ${\mathcal{C}}$ be an enumerable class of normal programs. Problems <span style="font-variant:small-caps;">Consistency</span>, <span style="font-variant:small-caps;">Credulous</span> and <span style="font-variant:small-caps;">Skeptical Reasoning</span>, <span style="font-variant:small-caps;">AS Counting</span> and <span style="font-variant:small-caps;">AS Enumeration</span> are all fixed-parameter tractable when parameterized by the size of a strong ${\mathcal{C}}$backdoor, assuming that the backdoor is given as an input. Let $X$ be the given backdoor, $k={|X|}$ and $n$ the input size of $P$. Since $P_\tau \in {\mathcal{C}}$ and ${\mathcal{C}}$ is enumerable, we can compute ${\text{\normalfont AS}}(P_\tau)$ in polynomial time for each $\tau \in {\text{\normalfont ta($X$)}}$, say in time $O(n^c)$. Observe that therefore ${|{\text{\normalfont AS}}(P_\tau)|}\leq O(n^c)$ for each $\tau \in {\text{\normalfont ta($X$)}}$. Thus we obtain ${\text{\normalfont AS}}(P,X)$ in time $O(2^k n^c)$, and ${|{\text{\normalfont AS}}(P,X)|}\leq O(2^k n^c)$. By Lemma \[lem:subset\], ${\text{\normalfont AS}}(P)\subseteq {\text{\normalfont AS}}(P,X)$. By means of Lemma \[lem:mincheck\] we can decide whether $M \in AS(P)$ in time $O(2^k n)$ for each $M\in {\text{\normalfont AS}}(P,X)$. Thus we determine from ${\text{\normalfont AS}}(P,X)$ the set of all answer sets of $P$ in time $O(2^k\cdot n^c \cdot 2^k \cdot n + 2^k \cdot n^c)=O(2^{2k} n^{c+1})$. Once we know ${\text{\normalfont AS}}(P)$, then we can also answer any of the listed problems within polynomial time. If we know that each program in ${\mathcal{C}}$ has at most one answer set, and $P$ has a strong ${\mathcal{C}}$backdoor of size $k$, then we can conclude that $P$ has at most $2^k$ answer sets. Thus, we obtain an upper bound on the number of answer sets of $P$ by computing a small strong ${\mathcal{C}}$backdoor of $P$. We consider program $P$ of Example \[ex1\]. The answer sets of $P_\tau$ are ${\text{\normalfont AS}}(P_{00})=\{\{t \}\}$, ${\text{\normalfont AS}}(P_{01})=\{\{t\}\}$, ${\text{\normalfont AS}}(P_{10})=\{\emptyset\}$, and ${\text{\normalfont AS}}(P_{11})=\{\emptyset\}$ for $\tau \in {{\text{\normalfont ta($\{r,s\}$)}}}$. ${\text{\normalfont AS}}(P,X)=\{ \{t\},\{t,s\},\{r\},\{r,s\}\}$, and since only $\{t\} \in {\text{\normalfont AS}}(P,X)$ is an answer set of $P$, we obtain ${\text{\normalfont AS}}(P)=\{\{t\}\}$. #### Deletion Backdoors We will see below that the following variant of strong backdoors is often useful. For a program $P$ and a set $X$ of atoms we define $P-X$ as the program obtained from $P$ by deleting all atoms contained in $X$ from all the rules (heads and bodies) of $P$. Let ${\mathcal{C}}$ be a class of programs. A set $X$ of atoms is a *deletion ${\mathcal{C}}$backdoor* of a program $P$ if $P-X\in {\mathcal{C}}$. In general, not every strong ${\mathcal{C}}$backdoor is a deletion ${\mathcal{C}}$backdoor, and not every deletion ${\mathcal{C}}$backdoor is a strong ${\mathcal{C}}$backdoor. We call ${\mathcal{C}}$ to be *rule induced* if for each $P\in {\mathcal{C}}$, $P'\subseteq P$ implies $P'\in {\mathcal{C}}$. Note that many natural classes of programs (and all classes considered in this paper) are rule induced. \[lem:rule-induced\] If ${\mathcal{C}}$ is rule induced, then every deletion ${\mathcal{C}}$backdoor is a strong ${\mathcal{C}}$backdoor. The statement follows from the fact that $P_\tau \subseteq P-X$ for every $\tau\in {\text{\normalfont ta($X$)}}$ and every program $P$. #### Backdoor Detection In order to use Theorem \[the:evaluation\] we need to find the backdoor first. Each class ${\mathcal{C}}$ of programs gives rise to the following parameterized problem: <span style="font-variant:small-caps;">Strong ${\mathcal{C}}$-Backdoor Detection</span>: given a program $P$ and an integer $k$, find a strong ${\mathcal{C}}$backdoor $X$ of $P$ of size at most $k$, or report that such $X$ does not exist. We also consider the problem <span style="font-variant:small-caps;">Deletion ${\mathcal{C}}$-Backdoor Detection</span>, defined similarly (which is in some cases easier to solve). Target Class ${\text{{\textnormal{\textbf{Horn}}}}}$ ==================================================== We first consider the important case ${\text{{\textnormal{\textbf{Horn}}}}}^*$ as the target class for backdoors. It is well known that normal Horn programs have a unique answer set and this set can be found in linear time [@DowlingGallier84], hence ${\text{{\textnormal{\textbf{Horn}}}}}$ and ${\text{{\textnormal{\textbf{Horn}}}}}^*$ are enumerable. The following lemma shows that ${\text{{\textnormal{\textbf{Horn}}}}}^*$ is particularly well suited as a target class. \[lem:Horn-strong-deletion\] A set $X$ is a strong ${\text{{\textnormal{\textbf{Horn}}}}}^*$-backdoor of a program $P$ if and only it is a deletion ${\text{{\textnormal{\textbf{Horn}}}}}^*$-backdoor of $P$. It suffices to show the lemma for ${\text{{\textnormal{\textbf{Horn}}}}}$, and in view of Proposition \[lem:rule-induced\] it suffices to show the only-if direction. Assume $X$ is a strong ${\text{{\textnormal{\textbf{Horn}}}}}$-backdoor of $P$. Consider a rule $r'\in P-X$ which is neither tautological nor a constraint. Let $r\in P$ be a rule from which $r'$ was obtained in forming $P-X$. We define $\tau\in {\text{\normalfont ta($X$)}}$ by setting all atoms in $H(r)\cup B^-(r)$ to 0, all atoms in $B^+(r)$ to 1, and all remaining atoms in $X\setminus {\text{\normalfont at}}(r)$ arbitrarily to 0 or 1. Since $r$ is not tautological, this definition of $\tau$ is sound. It remains to observe that $r'\in P_\tau$. Since $X$ is a strong ${\text{{\textnormal{\textbf{Horn}}}}}$-backdoor of $P$, the rule $r'$ is Horn. \[the:horn\] <span style="font-variant:small-caps;">Strong ${\text{{\textnormal{\textbf{Horn}}}}}^*$Backdoor Detection</span> is fixed-parameter tractable in $k$. In fact, given a program $P$ with $n$ atoms we can find a strong ${\text{{\textnormal{\textbf{Horn}}}}}^*$backdoor of size $\leq k$ in time $O(1.2738^k + kn)$ or decide that no such backdoor exists. (Sketch) Let $G$ be the undirected graph defined on the set of atoms of the given program $P$, where two atoms $x,y$ are joined by an edge if and only if $P$ contains a non-tautological rule $r$ with $x,y\in H(r)$ or $x\in H(r)$ and $y\in B^-(r)$. Now it is easy to see that a set $X\subseteq {\text{\normalfont at}}(P)$ is a vertex cover of $G$ if and only if $X$ is a deletion ${\text{{\textnormal{\textbf{Horn}}}}}^*$-backdoor of $P$. A vertex cover of size $\leq k$, if it exists, can be found in time $O(1.2738^k + kn)$ [@ChenKanjXia06]. The theorem follows by Lemma \[lem:Horn-strong-deletion\]. For instance, the undirected graph $G$ of the program $P$ of Example \[ex1\] consists of the two paths $(w,r,t)$ and $(s,q)$. Then $\{r,s\}$ is a vertex cover of $G$. We observe easily that there exists no vertex cover of size $1$. Thus $\{r,s\}$ is a smallest strong ${\text{{\textnormal{\textbf{Horn}}}}}^*$backdoor of $P$. [@Ben-Eliyahu96 ]{} showed that evaluation of normal logic programs is fixed-parameter tractable in (i) the number of atoms that appear in negative rule bodies, and (ii) the total number of non-Horn rules. It is not difficult to see that both numbers are greater or equal to the size of a smallest strong ${\text{{\textnormal{\textbf{Horn}}}}}^*$backdoor, and so entailed by our approach. Acyclicity-Based Target Classes =============================== There are two causes for a program to have a large number of answer sets: (i) disjunctions in the heads of rules, and (ii) certain cyclic dependencies between rules. Disallowing both causes yields so-called *stratified* programs [@GelfondLifschitz88]. In the following we will study backdoor detection for various classes of stratified programs. We define the classes by requiring normality and acyclicity (the absence of certain types cycles). In order to define acyclicity we associate with each program $P$ its *directed dependency graph* $D_P$ and its *undirected dependency graph* $U_P$ where $D_P$ is an extended version of the dependency graph in [@AptBlairWalker88] and $U_P$ of the undirected dependency graph in [@GottlobScarcelloSideri02]. $D_P$ has as vertices the atoms of $P$, a directed edge $(x,y)$ between any two atoms $x$, $y$ for which there is a rule $r\in P$ with $x\in H(r)$ and $y \in B(r)$ or a rule $r\in P$ with $x,y\in H(r)$; if there is a rule $r\in P$ with $x\in H(r)$ and $y\in B^-(r)$ or there is a rule $r \in P$ with $x,y\in H(r)$, then the edge $(x,y)$ is called a *negative edge*. $U_P$ is obtained from $D_p$ by replacing each negative edge $e=(x,y)$ with two undirected edges $\{x,v_e\},\{v_e,y\}$ where $v_e$ is a new *negative vertex*, and by replacing each remaining directed edge $(u,v)$ with an undirected edge $\{u,v\}$. By a *directed cycle of $P$* we mean a directed cycle in $D_P$, by an *undirected cycle of $P$* we mean an undirected cycle in $U_P$. Figure \[fig:DpUp\] visualizes $D_P$ and $U_p$ of the program $P$ of Example \[ex1\]. A directed (undirected) cycle is *bad* if it contains a negative edge (a negative vertex), otherwise it is *good*. Various classes of programs arise by requiring the programs to have no directed bad cycles ([[**DBC-Acyc**]{.nodecor}]{}), no undirected bad cycles ([[**BC-Acyc**]{.nodecor}]{}), no directed cycles ([[**DC-Acyc**]{.nodecor}]{}), and no undirected cycles ([[**C-Acyc**]{.nodecor}]{}). [[**DBC-Acyc**]{.nodecor}]{}, the largest class among the considered classes, is exactly the class ${{\textnormal{\textbf{Strat}}}}$ of stratified programs [@AptBlairWalker88]. For instance in the program $P$ of Example \[ex1\], $(r,$$v_{(w,r)},$$w,$$r)$ is an undirected cycle, $(u,q,u)$ is a directed cycle, $(s,$$v_{q,s},$$q,u,s)$ is an undirected bad cycle, and $(w,r,w)$ is a directed bad cycle (see Figure \[fig:DpUp\]). In order to compare the size of backdoors for the various classes, we need to compare the classes themselves (evidently, if ${\mathcal{C}}\subsetneq {\mathcal{C}}'$, then every strong (deletion) ${\mathcal{C}}'$-backdoor is also a strong (deletion) ${\mathcal{C}}'$-backdoor, but not necessarily the other way around). By definition we have ${\textnormal{\textbf{DC-Acyc}}}\subsetneq {\textnormal{\textbf{DBC-Acyc}}}$ and ${\textnormal{\textbf{C-Acyc}}}\subsetneq {\textnormal{\textbf{BC-Acyc}}} \subsetneq {\textnormal{\textbf{DBC-Acyc}}}$; it is easy to see that the inclusions are proper. However, contrary to what one expects, ${\textnormal{\textbf{C-Acyc}}}\not\subseteq {\textnormal{\textbf{DC-Acyc}}}$, which can be seen by considering the program $P_0=\{x\leftarrow y,\ y\leftarrow x\}$, hence ${\textnormal{\textbf{C-Acyc}}}$ and ${\textnormal{\textbf{DC-Acyc}}}$ are incomparable. Requiring that a program has no directed cycles but may have directed good cycles of length 2 (as in $P_0$) gives rise to the class ${\textnormal{\textbf{DC2-Acyc}}}$, which generalizes both classes ${\textnormal{\textbf{C-Acyc}}}$ and ${\textnormal{\textbf{DC-Acyc}}}$. The diagram in Figure \[fig:diagram\] shows the relationship between the various program classes. **** (0.3,0) node (A) [ ]{} (-1,-0.5) node (B) (-1,0.5) node (C) (-2,-0.5) node (D) (-2,0.5) node (E) (-2,-1.5) node (F) ; (D) edge (B) (D) edge (C) (E) edge (C) (B) edge (A) (C) edge (A) (F) edge (B) ; \[the:w2\] For each class ${\mathcal{C}}\in \{$[[**C-Acyc, BC-Acyc, DC-Acyc, DC2-Acyc**]{.nodecor}]{}, ${{\textnormal{\textbf{Strat}}}}\}$ the problem <span style="font-variant:small-caps;">Strong ${\mathcal{C}}$Backdoor Detection</span> is ${\text{\normalfont W[2]}}$-hard and therefore unlikely to be fixed-parameter tractable. (Sketch) We give a reduction from the ${\text{\normalfont W[2]}}$-complete problem <span style="font-variant:small-caps;">Hitting Set</span> [@DowneyFellows99]. An instance of this problem is a pair $({\mathsf{S}},k)$ where ${\mathsf{S}}=\{S_1,\dots,S_m\}$ is a family of sets and $k$ is an integer (the parameter). The question is whether there exists a set $H$ of size at most $k$ which intersects with all the $S_i$; such $H$ is a hitting set. We construct a program $P$ as follows. As atoms we take the elements of $X=\bigcup_{i=1}^m S_i$ and new atoms $a_i^j$ and $b_i^j$ for $1\leq i \leq m$, $1\leq j \leq k+1$. For each $1\leq i \leq m$ and $1\leq j \leq k+1$ we take two rules $r_i^j$, $s_i^j$ where $H(r_i^j)=\{a_i^j\}$, $B^-(r_i^j)=S_i$, $B^+(r_i^j)=S_i\cup \{b_i^j\}$; $H(s_i^j)=\{b_i^j\}$, $B^-(s_i^j)=\{a_i^j\}$, $B^+(s_i^j)=\emptyset$. The result now follows by showing that ${\mathsf{S}}$ has a hitting set of size $\leq k$ if and only if $P$ has a strong ${\mathcal{C}}$backdoor of size $\leq k$ where ${\mathcal{C}}$ is any of the classes mentioned. For [[**DC-Acyc**]{.nodecor}]{}, [[**DC2-Acyc**]{.nodecor}]{}, and [[[**Strat**]{.nodecor}]{}]{} we can avoid the use of tautological rules in the reduction and so strengthen Theorem \[the:w2\] as follows (it would be interesting to know if this is also possible for the remaining two classes mentioned in Theorem \[the:w2\]). \[the:w2\*\] For each class ${\mathcal{C}}\in \{$[[**DC-Acyc, DC2-Acyc**]{.nodecor}]{}, ${{\textnormal{\textbf{Strat}}}}\}$ the problem <span style="font-variant:small-caps;">Strong ${\mathcal{C}}^*$Backdoor Detection</span> is ${\text{\normalfont W[2]}}$-hard and therefore unlikely to be fixed-parameter tractable. (Sketch) We modify the above reduction from <span style="font-variant:small-caps;">Hitting Set</span> by redefining the rules $r_i^j$, $s_i^j$. We put $H(r_i^j)=\{a_i^j\}$, $B^-(r_i^j)=S_i$, $B^+(r_i^j)=\{b_i^j\}$; $H(s_i^j)=\{b_i^j\}$, $B^-(s_i^j)=\{a_i^j\}$, $B^+(s_i^j)=X$. The ${\text{\normalfont W[2]}}$-hardness results suggests to relax the considered problems and look for *deletion* backdoors: Which of the classes mentioned in Theorem \[the:w2\] admit fixed-parameter tractable detection of deletion backdoors? Using very recent results from fixed-parameter algorithmics we can answer this question positively for all considered classes except for ${{\textnormal{\textbf{Strat}}}}$ whose complexity remains open. The ${\text{\normalfont W[2]}}$-hardness results suggest to relax the considered problems and to look for *deletion* backdoors: Which of the classes mentioned in Theorem \[the:w2\] admit fixed-parameter tractable detection of deletion backdoors? Using very recent results from fixed-parameter algorithmics we can answer this question positively for all considered classes except for ${{\textnormal{\textbf{Strat}}}}$ whose complexity remains open. \[the:fvs\] For each class ${\mathcal{C}}\in \{$[[**C-Acyc, BC-Acyc, DC-Acyc, DC2-Acyc**]{.nodecor}]{}$\}$ the problem <span style="font-variant:small-caps;">Deletion ${\mathcal{C}}^*$Backdoor Detection</span> is fixed-parameter tractable. (1) <span style="font-variant:small-caps;">Deletion [[**C-Acyc**]{.nodecor}]{}Backdoor Detection</span> can be solved by solving the <span style="font-variant:small-caps;">Feedback Vertex Set</span> (FVS) problem for $U_p$, which is well-known to be FPT [@DowneyFellows99], this was already observed by [@GottlobScarcelloSideri02 ]{}. (2) <span style="font-variant:small-caps;">Deletion [[**DC-Acyc**]{.nodecor}]{}Backdoor Detection</span> is equivalent to the <span style="font-variant:small-caps;">Directed FVS</span> problem on $D_p$. The parameterized complexity of <span style="font-variant:small-caps;">Directed FVS</span> remained open for many years and was recently shown FPT by [@ChenLiuLuOsullivanRazgon08 ]{} with a break-through result. (3) <span style="font-variant:small-caps;">Deletion [[**BC-Acyc**]{.nodecor}]{}Backdoor Detection</span> can be solved by solving the <span style="font-variant:small-caps;">Edge Subset FVS</span> problem, where only cycles need to be covered that contain an edge from a given set $S$. This problem was recently shown FPT by [@CyganPilipczukPilipczukWojtaszczyk10 ]{} and [@KawarabayashiKobayashi10 ]{}. (4) Finally, the problem <span style="font-variant:small-caps;">Deletion [[**DC2-Acyc**]{.nodecor}]{}Backdoor Detection</span> can be solved by finding a feedback vertex set in a *mixed graph* (a graph containing directed and undirected edges), as we can replace a good cycle on two edges by one undirected edge. <span style="font-variant:small-caps;">FVS for Mixed Graphs</span> was recently shown FPT by [@BonsmaLokshtanov10 ]{}. The classes mentioned in Theorem \[the:w2\] are rule-induced. Hence we can use this theorem to strengthen the fixed-parameter tractability result of Theorem \[the:evaluation\] by dropping the assumption that the backdoor is given. The considered classes can be generalized by taking the parity of the number of negative edges on bad cycles into account. In recent research [@Fichte11 ]{} generalized the tractability results of [@LinZhao04 ]{} by considering backdoors with respect to such parity classes. Theoretical Comparison of Parameters ==================================== In this section we compare ASP parameters in terms of their generality. [@JaklPichlerWoltran09 ]{} applied the graph parameter treewidth to ASP and showed that the main reasoning problems for ASP are fixed-parameter tractable by the treewidth of the *incidence graph* of the program. The incidence graph of a program $P$ is the bipartite graph on the rules and atoms of $P$, where a rule and an atom are joined by an edge if and only if the atom occurs in the rule. It turns out that incidence treewidth is incomparable with backdoor size for various target classes considered. Let ${\mathcal{C}}\in\{$, [[**C-Acyc**]{.nodecor}]{}, [[**DC-Acyc**]{.nodecor}]{}, [[**DC2-Acyc**]{.nodecor}]{}, [[**BC-Acyc**]{.nodecor}]{}, [[**Strat**]{.nodecor}]{}$\}$. There are programs whose incidence graphs have constant treewidth but require arbitrarily large strong and deletion ${\mathcal{C}}$backdoors, and there are programs where the converse prevails. (Sketch) Clearly each of the considered classes contains programs of arbitrary high treewidth of the incidence graph. Conversely, we consider a program $P\notin {\mathcal{C}}$. We denote by $nP$ the program consisting of the union of $n$ atom-disjoint copies of $P$. By basic properties of treewidth it follows that the treewidth of the incidence graph of $P$ equals to that of $nP$, however, smallest ${\mathcal{C}}$backdoors are of size $\geq n$. The treewidth approach is based on dynamic programming which is of high space complexity and therefore only practical for instances of treewidth below 10 [@JaklPichlerWoltran09]. The backdoor approach is more space efficient since for each partial truth assignment $\tau\in {\text{\normalfont ta($X$)}}$ of a backdoor $X$, the computations of ${\text{\normalfont AS}}(P_\tau)$ and the corresponding elements of ${\text{\normalfont AS}}(P_\tau)$ can be carried out independently. One might ask whether it makes sense to consider restrictions on the treewidth of the undirected dependency graph, defined above. However, this restriction does not yield tractability, as the reduction of [@EiterGottlob95] produces programs with undirected dependency graphs of treewidth 2. Empirical Comparison of Parameters ================================== We have determined the size of smallest backdoors for various programs, including *structured* real-world instances and *random* instances. The results are summarized in Table \[tab:horn\]. It is known that so-called tight programs are closely related to SAT [@LinZhao03]. The QueensEqTest instances and two of the Daimler-Chrysler instances are tight, all other instances considered are not tight. instance set vars bd (%) stdev ----------------------- ----------- --------- -------- `Daimler-Chrysler-MT` $1785.14$ $21.46$ $1.56$ `Daimler-Chrysler-NC` $1793.0$ $22.94$ $2.92$ `Daimler-Chrysler-RZ` $1562.5$ $11.53$ $2.21$ `Daimler-Chrysler-SZ` $1567.53$ $13.97$ $3.47$ `Daimler-Chrysler-UC` $1781.74$ $21.4$ $2.01$ `Daimler-Chrysler-UT` $1781.23$ $23.53$ $3.95$ `Mutex` $6449.0$ $49.94$ $0.09$ `RLP-3` $150.0$ $58.28$ $1.37$ `RLP-4` $150.0$ $64.53$ $0.92$ `RLP-5` $150.0$ $68.4$ $0.97$ `RLP-6` $150.0$ $70.9$ $0.82$ `RLP-7` $150.0$ $73.68$ $0.87$ `RLP-8` $150.0$ $75.54$ $0.74$ `RG-40` $40.0$ $93.5$ $1.24$ `RG-50` $50.0$ $94.05$ $0.96$ `RG-60` $60.0$ $94.38$ $0.82$ : Size of smallest strong ${\text{{\textnormal{\textbf{Horn}}}}}$backdoors (bd) for various benchmark sets, given as % of the total number of variables (vars) by the mean over the instances. `Daimler-Chrysler`-$\langle$test$\rangle$: 554 Real-world instances encoding logistics problems from car configurations. The disjunctive programs have been compiled from SAT instances provided by Sinz et al. \[2003\] grouped by the kind of consistency test. The instances are produced using the simple encoding where a clause $\{a,b,\neg c, \neg d\}$ becomes the rule $a, b\leftarrow c,d$. `Mutex`: Disjunctive programs that encode the equivalence test of partial implementations of circuits, provided by Maratea et al. \[2008\] based on QBF instances of Ayari and Basin \[2000\] `RLP`-$\langle \rho \rangle$: Randomly generated normal programs provided by Zhao and Lin \[2003\] of various density $\rho$ (number of rules divided by the number of variables) with 10 instances per step. `RG`-$\langle n \rangle$: Randomly generated instances provided by Gebser \[Asp, 2009\] with $n=40$, 50, and 60 variables, respectively with 40 instances per step.[]{data-label="tab:horn"} For pragmatic reasons we have used ${\text{{\textnormal{\textbf{Horn}}}}}^*$ as the target class as smallest backdoors are easy to compute, even for large inputs. Our experimental results indicate that structured instances have smaller backdoors than random instances. It also seems that random instances with higher density have larger backdoors. We have conducted a second series of experiments on random instances where we have analyzed how much we gain by considering the more general acyclicity-based target classes instead of ${\text{{\textnormal{\textbf{Horn}}}}}$. It appears that smallest deletion ${\text{{\textnormal{\textbf{Horn}}}}}^*$backdoors are indeed significantly larger than deletion backdoors for acyclicity-based target classes. The distinction between directed and undirected cycles seems to have a significant effect on the backdoor size, whereas the distinction between good and bad cycles seems to be less significant. However these results are not fully conclusive as the considered programs were rather small. Conclusion ========== We have introduced the backdoor approach to the domain of propositional answer-set programming. The backdoor approach allows to augment a known tractable class and makes the efficient solving methods for the tractable class generally applicable. Our approach makes recent results in fixed-parameter algorithmics applicable to nonmonotonic reasoning. The comparison results show that the parameters based on backdoor size are incomparable with treewidth and therefore provide fixed-parameter tractability for programs that are hard for the treewidth approach. The results and concepts of this paper give rise to several interesting research questions. For instance, it would be interesting to consider backdoors for target classes that contain programs with an exponential number of answer-sets, but where the set of all answer-sets can be succinctly represented. A simple example is the class of programs that consist of independent components of bounded size. Other questions are concerned with alternative ways of using backdoors. For instance, by means of “backdoor trees” [@SamerSzeider08b] one can avoid the consideration of all $2^k$ partial assignments of the backdoor and thus make the backdoor approach feasible for programs with larger backdoors. A further use of backdoors that seems worth exploring is the control of heuristics of ASP solvers. K. R. Apt, H. A. Blair, and A. Walker. , Morgan Kaufmann, 1988. Asparagus. <http://asparagus.cs.uni-potsdam.de>, 2009. A. Ayari and D. Basin. Bounded model construction for monadic second-order logics. . R. Ben-Eliyahu. A hierarchy of tractable subsets for computing stable models. , 5, 1996. P. Bonsma and D. Lokshtanov. Feedback vertex set in mixed graphs. Algorithms and Data Structures, 122–133, Springer, 2011. M. Cadoli and M. Lenzerini. The complexity of propositional closed world reasoning and circumscription. Journal of Computer and System Sciences, 48(2), 255–310, Elsevier, 1994. J. Chen, I. A. Kanj, and G. Xia. Improved parameterized upper bounds for vertex cover. . J. Chen, Y. Liu, S. Lu, B. O’Sullivan, and I. Razgon. A fixed-parameter algorithm for the directed feedback vertex set problem., 55(5), 2008. M. Cygan, M. Pilipczuk, M. Pilipczuk, and J. O. Wojtaszczyk. Subset feedback vertex set is fixed parameter tractable. arXiv 1004.2972, 2010. W. F. Dowling and J. H. Gallier. Linear-time algorithms for testing the satisfiability of propositional [H]{}orn formulae. , 1(3), 1984. R. G. Downey and M. R. Fellows. . Springer, 1999. T. Eiter and G. Gottlob. On the computational cost of disjunctive logic programming: propositional case. , 15(3-4), 1995. Fichte, J.K., Szeider, S.: Backdoors to tractable answer-set programming.\ . J. K. Fichte. The good, the bad, the odd – cycles in answer set programming. [*ESSLLI 2011 Student Session*]{}, 2011. M. Gelfond and V. Lifschitz. The stable model semantics for logic programming. . M. Gelfond and V. Lifschitz. Classical negation in logic programs and disjunctive databases. , 9(3/4), 1991. G. Gottlob and S. Szeider. Fixed-parameter algorithms for artificial intelligence, constraint satisfaction, and database problems. , 51(3), 2006. G. Gottlob, F. Scarcello, and M. Sideri. Fixed-parameter complexity in [AI]{} and nonmonotonic reasoning. , 138(1-2), 2002. M. Jakl, R. Pichler, and S. Woltran. Answer-set programming with bounded treewidth. . T. Janhunen and E. Oikarinen. Testing the equivalence of logic programs under stable model semantics. [*Logics in AI*]{}, 2002. K. Kawarabayashi and Y. Kobayashi. Fixed-parameter tractability for the subset feedback set problem and the s-cycle packing problem. Technical report, University of Tokyo, Japan, 2010. F. Lin and J. Zhao. On tight logic programs and yet another translation from normal logic programs to propositional logic. . F. Lin and X. Zhao. On odd and even cycles in normal logic programs. . M. Maratea, F. Ricca, W. Faber, and N. Leone. Look-back techniques and heuristics in [D]{}[L]{}[V]{}: Implementation, evaluation, and comparison to [Q]{}[B]{}[F]{} solvers. , 63, 2008. V. W. Marek and M. Truszczy[ń]{}ski. Autoepistemic logic. , 38(3), 1991. V. W. Marek and M. Truszczy[ń]{}ski. Stable models and an alternative logic programming paradigm. , 1999. I. Niemel[ä]{}. Logic programs with stable model semantics as a constraint programming paradigm. , 25(3-4), 1999. N. Nishimura, P. Ragde, and S. Szeider. Detecting backdoor sets with respect to [Horn]{} and binary clauses. . N. Nishimura, P. Ragde, and S. Szeider. Solving \#[S]{}[A]{}[T]{} using vertex covers. , 44(7-8), 2007. S. Ordyniak and S. Szeider Augmenting Tractable Fragments of Abstract Argumentation. . I. Razgon and B. O’Sullivan. Almost 2-sat is fixed-parameter tractable (extended abstract). . M. Samer and S. Szeider. Backdoor trees. . M. Samer and S. Szeider. Fixed-parameter tractability. , ch. 13, 2009. M. Samer and S. Szeider. Backdoor sets of quantified [B]{}oolean formulas. , 42(1):77–97, 2009. C. Sinz, A. Kaiser, and W. K[ü]{}chlin. . . R. Williams, C. Gomes, and B. Selman. Backdoors to typical case complexity. . Y. Zhao and F. Lin. . . [^1]: Research supported by ERC (COMPLEX REASON 239962). [^2]: This is an extended and updated version of a paper that appeared in the Proceedings of IJCAI’11. [^3]: In [@FichteSzeider11] we overlooked that the minimality check requires some additional attention.

ArXiv

--- abstract: 'We carried out a feasibility study on the measurement of the branching ratio of $H \rightarrow c\bar{c}$ at a future $e^{+}e^{-}$ linear collider. We used topological vertex reconstruction algorithm for accumulating secondary vertex information and neural network for optimizing $c$ quark selection. With an assumption of a Higgs mass of 120 GeV/$c^{2}$ we estimated statistical uncertainty of Br($H \rightarrow c\bar{c}$) to be 20.1% or 25.7% , depending on the number of vertex detector layers, at the center-of-mass energy of 250 GeV and the integrated luminosity of 500 ${\rm fb^{-1}}$.' address: - 'Department of Physics, Korea University, Seoul, Korea, 136-701' - 'KEK, 1-1 Oho, Tsukuba-Shi, Ibaraki-Ken, Japan, 305-0801' - 'Department of Physics, Kyungpook National University, Taegu, Korea, 702-701' author: - 'Yu, Geum Bong[^1]' - 'Kang, JooSang' - 'Miyamoto, Akiya' - 'Park, Hwanbae[^2]' title: | A Study of the Branching Ratio of $H \rightarrow c\bar{c}$\ at a Future $e^{+}e^{-}$ Linear Collider --- \[sec:level1\]Introduction ========================== In the theory of elementary particles, Higgs boson is introduced to elucidate the origin of particle masses. The Standard Model (SM) proposed the single Higgs doublet that gives rise to a scalar particle, while extended supersymmetric models hypothesized that two Higgs doublets give separate vacuum expectation values to the up-type and down-type quarks. Among the extended models beyond SM, the Minimal Supersymmetric Standard Model (MSSM) sets the upper bound of the lightest neutral Higgs mass around 130 GeV/$c^{2}$[@susy]. Since the mass of SM Higgs boson is expected to be between 114.4 GeV/$c^{2}$ and 211 GeV/$c^{2}$ by LEP experiment[@lep], the Higgs boson should be identified if it is found. Therefore a precise measurement of Higgs couplings to fermions at a future $e^{+}e^{-}$ linear collider is indispensable to unveil physics of the Higgs sector. In particular, a measurement of the branching ratio of $H \rightarrow c\bar{c}$ would provide a unique opportunity to study Higgs to up-type quark coupling. However, the $H \rightarrow b\bar{b}$ is the dominant process for Higgs mass below 140 GeV/$c^{2}$ . A high efficient charm quark identification is necessary for the precise branching ratio study and it is possible only with a high performance vertex detector. Development of this vertex detector has been a major issue for the linear collider [@lcrnd]. To this end, understanding detector performance is extremely important to physics. In this paper, we introduce our method of measuring the branching ratio of $H \rightarrow c\bar{c}$ against SM $e^{+}e^{-}$ processes and other decay branches from Higgs. In addition, we would like to see the significance of this precision study with four layers of Charge Coupled Device (CCD) and an additional inner layer of vertex detector. \[sec:level1\]HIGGS SIMULATION ============================== A mass of SM Higgs boson is assumed to be 120 GeV/$c^2$ and the center-of-mass energy of 250 GeV is selected by the highest $s$-channel production cross-section at this energy, 226 fb. It is concerned with an integrated luminosity of 500 $\rm fb^{-1}$ for the first few years running of the future $e^{+}e^{-}$ linear collider experiment. Events were generated using PYTHIA 5.7 [@pythia] and only the Higgsstrahlung process ($e^{+}e^{-} \rightarrow Z^{\ast} \rightarrow Z^{\circ}H$) was taken into account. The branching ratios of the SM Higgs boson was estimated by the HDecay program [@hdecay]. Depending on the decay modes of $Z^{\circ}$, the events are categorized into 4-jet mode ($Z^{\circ} \rightarrow q\bar q$ and $H \rightarrow q\bar q$), 2-jet mode ($Z^{\circ} \rightarrow \nu\bar\nu$ and $H \rightarrow q\bar q$), and charged lepton pair mode ($Z^{\circ} \rightarrow \ell^+\ell^-$ and $H \rightarrow q\bar q$). Since the branching ratio of Higgs to $c\bar{c}$ is estimated to be very small ($\sim$3%), we concentrated on the 2-jet and 4-jet modes only. As background processes from $e^{+}e^{-}$ collision, $W^{+}W^{-},~Z^{\circ}Z^{\circ}$, and $q\bar{q}$ events were considered with corresponding cross-sections of 15460 fb, 1250 fb, and 47300 fb, respectively. The detector simulation was performed using a fast parameterized simulator [@jlc], which is implemented in the Joint Linear Collider (currently called as Global Linear Collider) detector [@report]. In this simulator five parameters of the helical track and their error matrices including non-diagonal elements were generated; thus the quality of the vertices is similar to that of a full simulation. The JLC vertex detector is equipped with four layers of CCD at the radius from 2.4 cm to 6 cm of which intrinsic spatial resolutions are 4 ${\rm \mu m}$ in ${\rm r \phi}$ and z directions. The solenoidal magnet field of the detector is 3 tesla. With a vertex detector constraint, the impact parameter resolution in xy plane (${\rm \sigma_{r \phi}}$) is ${\rm \sqrt{(25/p sin^{2/3} \theta)^{2} +4^2} ~\mu m}$, and the momentum resolution for charged track (${\rm \Delta p_t/p_t }$) is ${\rm \sqrt{(1\times 10^{-4} p_t)^{2}+ 10^{-3} } }$. This study was done with four/five CCD layers of vertex detector parameter, repectively, to see the influence of the vertex detector options on the physics result. The vertex detector parameter sets are shown in the Table \[tab:table1\]. number of CCD layers Beam pipe 1st 2nd 3rd 4th 5th ---------------------- ----------- ----- ----- ----- ----- ----- -- -- -- 4 2.0 2.4 3.6 4.8 6.0 5 1.0 1.2 2.4 3.6 4.8 6.0 : \[tab:table1\] Two vertex detector parameter set. the number of CCD layers and their positions(in cm) from the center of the beam pipe, the radius of beam pipe are listed for each set. In the event reconstruction, jets were reconstructed using JADE clustering algorithm [@jade] and the vertices were reconstructed using the topological vertex reconstruction algorithm (ZVTOP program) [@zvtop]. Based on kinematic variables and reconstructed vertex information, an artificial neural network (NN) [@nn] is formed to identify $H \rightarrow c\bar{c}$ events more efficiently. We considered the non-$c\bar{c}$ Higgs decays ($H \rightarrow b\bar{b},~gg,~ WW^{\ast}$; Higgs background) and other processes from $e^{+}e^{-}$ collision ($e^{+}e^{-} \rightarrow Z^{\circ}Z^{\circ},~W^{+}W^{-},~q\bar{q}$; non-Higgs background) as backgrounds of $H \rightarrow c\bar{c}$ measurement. \[sec:level1\]Selection of $H \rightarrow c\bar{c}$ =================================================== \[sec:level2\]2-jet mode ------------------------ In 2-jet mode all clustered particles are forced to make 2 jets by adjusting the maximum value of ycut [^3], Ymax. The following selections are applied to reduce non-Higgs background in the sample: (1)visible energy between 110 GeV and 143 GeV, (2)missing transverse momentum between 25 GeV/$c$ and 70 GeV/$c$, (3)Higgs mass between 105 GeV/$c^{2}$ and 125 GeV/$c^{2}$, (4)recoiled $Z^{\circ}$ mass between 82 GeV/$c^{2}$ and 120 GeV/$c^{2}$, (5)thrust between 0.75 and 0.99, (6)Ymax between 0.72 and 0.84, and (7)mass of each jet between 2 GeV/$c^{2}$ and 40 GeV/$c^{2}$. Here the Higgs mass is an invariant mass of all the observed particles and the recoiled $Z^{\circ}$ mass is calculated disregarding the initial state radiation. We also require a successful secondary vertex reconstruction and at least one ${\rm P_{t}}$ corrected invariant mass (MSPTM, ${\rm \sqrt{ {M^2}_{VTX}+{P_t}^{2} }+|P_{t}|}$) of the secondary vertex to be between 0.1 GeV/$c^{2}$ and 7.0 GeV/$c^{2}$, where ${\rm P_t}$ is the total transverse momentum of the secondary tracks with respect to the flight direction of the vertex. The separation between signal and non-Higgs background is clearly seen in Fig. \[fig:2jsel\], while Higgs background is not distinguishable in this aspect. After the sample selection, NN is trained to maximize the selection quality. Two sets of training are used; one against non-Higgs backgrounds (Background NN-training) and the other against Higgs backgrounds ( Higgs NN-training). For each NN-training $H \rightarrow WW^{\ast}$ and $e^{+}e^{-} \rightarrow q\bar{q}$ processes are not used due to the featureless patterns and small statistics, respectively.\ The normalized input patterns for NN-trainings were given from the ZVTOP program; the number of vertices, MSPTM, transverse momentum of the secondary vertex, decay length, invariant mass of the secondary vertex, number of tracks in the secondary vertex, and momentum of the secondary vertex divided by total momentum (corrected secondary momentum). Distinctive input patterns for Higgs NN-training and Background NN-training are shown in Fig. \[fig:2h\] and Fig. \[fig:2b\], respectively. In the Background NN-training Ymax and jet decay angle were included in the input patterns especially for this 2-jet mode. Using both results of Higgs NN-training and Background NN-training, the signal region is evaluated in two dimensional plane for best significance. The signal selection area is shown in the top right corner of the Fig. \[fig:2d2j\]. \[sec:level2\]4-jet mode ------------------------ Here we used forced 4-jet clustering for Higgs jet pair and $Z^{\circ}$ jet pair. The jets are identified by minimum $\chi^{2}$, which is defined as the squared sum of the differences between reconstructed invariant mass of jet pair and the expected mass divided by each mass resolution: $$\begin{aligned} \chi^2 = (\frac{M_{Z^{\circ}}-91.2}{\rm width})^2 + (\frac{M_{{\rm recoil}Z^{\circ}}-91.2}{\rm width})^2 \\ + (\frac{M_H-120}{\rm width})^2 + (\frac{M_{{\rm recoil} H}-120}{\rm width})^2. \end{aligned}$$ The selection cuts were made: (1) visible energy greater than 210 GeV, (2) $\cos \theta$ of thrust axis(thrust angle) between $-0.85$ and 0.85, (3) Higgs mass between 103 GeV/$c^{2}$ and 130 GeV/$c^{2}$, (4) $Z^{\circ}$ mass between 78 GeV/$c^{2}$ and 102 GeV/$c^{2}$, (5) mass recoiled to the Higgs jet pair greater than 80 GeV/$c^{2}$, (6) number of particles in each jet of Higgs pair greater than 5, and (7) Ymax value greater than 0.01. Only 2 jets which are assigned to Higgs are used for the vertex tagging. We also required at least one successful secondary vertex reconstruction with MSPTM between 0.1 GeV/$c^{2}$ and 7.0 GeV/$c^{2}$ out of 2 jets from Higgs. The distributions of Higgs mass and thrust angle are shown in the Fig. \[fig:4j\]. With the same procedure as the 2-jet study, we trained the NN against non-Higgs background and Higgs background separately using normalized vertex information from the ZVTOP program. The MSPTM and the corrected secondary momentum distributions of the $c\bar{c}$ and the non-$c \bar{c}$ Higgs decays are seen in the Fig. \[fig:4h\]. We did not use $H \rightarrow WW^{\ast}$ events in Higgs NN-training due to their featureless patterns. In the Background NN-training, thrust angle is added in the input pattern. As seen in the Fig. \[fig:4b\], the thrust angle shows a clear separation between the signal and non-Higgs backgrounds. We use both results of Higgs NN-training and Background NN-training same as 2-jet mode. The 2D NN-results are shown in the Fig. \[fig:2d4j\] with signal box at the top right. Number of CCD layers ----------------------------------------------- -------------- ------------- ------------- ------------- -- -- Decay mode $\backslash$ Events 2-jet 4-jet 2-jet 4-jet ${H \rightarrow c\bar{c}}$ 122.7(17.6) 306.3(12.6) 112.7(16.2) 316.8(13.0) ${H \rightarrow b\bar{b}}$ 641.8(4.2) 2686.7(5.0) 231.2(1.5) 1807.2(3.4) ${H \rightarrow gg}$ 28.1(1.8) 217.6(3.9) 8.9(0.6) 143.4(2.6) ${H \rightarrow WW^{\ast}}$ 23.6(0.8) 226.7(2.1) 10.1(0.3) 178.0(1.7) ${e^{+}e^{-} \rightarrow W^{+}W^{-}}$ 640(0.008) 9790(0.130) 330(0.004) 8530(0.110) ${e^{+}e^{-} \rightarrow Z^{\circ}Z^{\circ}}$ 100(0.016) 1710(0.274) 30(0.004) 1305(0.209) ${e^{+}e^{-} \rightarrow q\bar{q}}$ 20($<$0.001) 1730(0.007) 5($<$0.001) 1345(0.006) S/N 0.0843 0.0187 0.1832 0.0238 ${S / \sqrt{S+N}}$ 3.09 2.37 4.17 2.71 Statistical uncertainty \[sec:level1\]Discussion ======================== The results of $c\bar{c}$ event selection are summarized in Table  \[tab:table2\]. As seen in the table, roughly 120 and 310 $H \rightarrow c\bar{c}$ events are selected in 2-jet and 4-jet modes, respectively, with a reasonable significance despite its small branching ratio of $H \rightarrow c\bar{c}$ and large backgrounds from non-Higgs processes and non-$c \bar{c}$ Higgs decays. In the $H \rightarrow c\bar{c}$ measurement in 2-jet mode, major backgrounds are those from $H \rightarrow b\bar{b}$ where $b$ is misidentified as $c$ and $W^{+}W^{-}$ events which are reconstructed as 2-jet events due to imperfect detector acceptance. $e^{+}e^{-} \rightarrow e \nu W$ is not considered here due to relatively small cross-section than other background processes. In the case of 4-jet mode, $W^{+}W^{-}$ and $Z^{\circ}Z^{\circ}$ events are increased after event selection because two $c$ jets can be produced in the final state of these processes. In addition to the c-jet contamination from backgrounds, there is additional ambiguity in selecting two Higgs jets out of four jets. Combining these effects, the significance in the 4-jet mode is less than 3$\sigma$ and worse than the 2-jet mode. The statistical uncertainty combining 2-jet and 4-jet analyses is estimated as 25.7% in the case of four CCD layers of the vertex detector. Similar analysis with different detector configuration was report[@JB] for assumption of the same Higgs mass and luminosity at $\sqrt s$=350 GeV and 500 GeV that the relative braching ratio errors be 19% and 39%, respectively. When an additional CCD layer of the vertex detector is included near the interaction point, the impact parameter resolution for low momentum track is improved since the lever arm for track extrapolation to the interaction point is reduced. Thus we can reconstruct secondary vertices much closer to the interaction point. The better separation of $b$- and $c$- jets reduces backgrounds in the event selection, especially in 2-jet mode. Improvements in 4-jet selection is not decisive, suffering ambiguities in Higgs jet selections. Combining 2-jet and 4-jet, relatively 20% improvement in background reduction is achieved compared to the study with four CCD layers of the vertex detector. \[sec:level1\]Summary ===================== In this study we focused on the $H \rightarrow c\bar{c}$ measurements in 2-jet and 4-jet modes in the case of 120 GeV/$c^{2}$ Higgs mass at the center-of-mass energy of 250 GeV in the future $e^{+}e^{-}$ linear collider. In the study, the topological vertex finding algorithm was used for tagging the $c$-jet and the neural network was used to optimize the $H \rightarrow c\bar{c}$ selection. With consideration of 500 ${\rm fb^{-1}}$ data, we obtained the statistical uncertainty of 25.7% for the measurement of $H \rightarrow c\bar{c}$ with four CCD layers vertex detector. The statistical uncertainty would be improved to 20.1% with an additional CCD layer of vertex detector near the interaction point. \[sec:ack\]Acknowledgment ========================= We would like to thank the members of ACFA Joint Linear Collider Physics and Detector Working group [@acfa] for valuable discussions during the course of the analysis. We also thank Dr. David Jackson for allowing us to use his ZVTOP program. We acknowledge the support by Korea Research Foundation Grant funded by Korea Government (MOEHRD) (KRF-2005-070-C000321) and Korea Science and Engineering Foundation. Akiya Miyamoto is partially supported by Japan-Europe Research Cooperative Program. [1]{} J. Kamoshita, Y. Okada, and M. Tanaka, Phys. Lett. [**328B**]{}, 67 (1994) ; T. Moroi and Y. Okada, Phys. Lett. [**295B**]{}, 73 (1992). The LEP Collaborations, the LEP Electroweak Working Group, and the SLAC heavy flavour group, LEPEWWG/2003-01 (2003). “Report on the international detector R & D”, See “http://blueox.uoregon.edu/$\sim$lc/randd.html”. T. Sjostrand, Comput. Phys. Commun. [**82**]{}, 74 (1994). A. Djouadi, J. Kalinowski, and M. Spira, Comput. Phys. Commun. [**108**]{}, 56 (1998). JLC Study Framework, See “http://www-jlc.kek.jp/subg/offl/jsf”. K. Abe [*et al.*]{}, KEK Report 2001-11 (2001). JADE Collaboration, Phys. Lett. [**213B**]{}, 235 (1988) D. J. Jackson, Nucl. Instrum. Meth. [**A 388**]{}, 247 (1997). C. Peterson, T. Rognvaldsson, and L. Lonnblad, Comput. Phys. Commun. [**81**]{}, 185 (1994). See “http://acfahep.kek.jp”. C.T. Potter, J.E. Brau and N.B. Sinev, Nucl. Instrum. Meth. [**A 511**]{}, 225 (2003). [^1]: Current address: Department of Physics and Astronomy, University of Rochester, Rochester, NY 14627, U.S.A. [^2]: Corresponding author: [email protected] [^3]: ${\rm y_{ij}=\frac{2E_{i}E_{j} (1-cos\theta_{ij})}{E_{visible}^2}}$, where ${\rm E_i~and~E_j}$ are energy of i-th and j-th cluster, respectively, and $\theta_{ij}$ is angle between two clusters.

ArXiv

--- abstract: 'Nonsequential multiple ionization of atoms in intense laser fields is initiated by a recollision between an electron, freed by tunneling, and its parent ion. Following recollision, the initial electron shares its energy with several bound electrons. We use a classical model based on rapid electron thermalization to interpret recent experiments. For neon, good agreement with the available data is obtained with an upper bound of 460 attoseconds for the thermalization time.' author: - 'X. Liu' - 'C. Figueira de Morisson Faria' - 'W. Becker' - 'P. B. Corkum' title: 'Attosecond electron thermalization by laser-driven electron recollision in atoms' --- Atoms exposed to intense laser fields ionize. The freed electron and its ionic partner are accelerated by the laser field away from each other. When the field changes sign, it may drive the electron into a recollision with the ion. This simple mechanism, which for high-intensity low-frequency fields is largely classical, governs many laser-atom processes such as high-order harmonic generation (HHG), high-order above-threshold ionization (HATI), and nonsequential double ionization (NSDI) and explains the gross features of the spectra observed [@corkum]. The period of the commonly applied titanium-sapphire (Ti:Sa) laser is about 2.7 fs. The recollision physics unfold on the time scale of a small fraction of the laser period. Therefore, the analysis of laser-induced recollision phenomena provides access to the inner-atomic dynamics on the attosecond time scale and, indeed, the focus of recent investigations has moved into this temporal domain. For example, it has brought molecular imaging with subangstrom spatial and subfemtosecond temporal resolution within reach [@Niikura; @molimag]. The advent of phase-stabilized infrared few-cycle pulses and of uv pulses of attosecond duration allows even more control [@Baltuska], but neither are necessary for a study of the attosecond dynamics. In this Letter, we analyze recent experiments on nonsequential multiple ionization (NSMI) of neon \[5,6\] in which the momentum distributions of Ne$^{N+}$ $(N=3,4)$ were measured at two intensities. We use the fact that the time-dependent laser field provides a clock — the field accelerates the ion to a final velocity that depends on the times at which the $N$ electrons ionized. This “streak camera” [@streak] therefore measures the range of times of ionization. Comparing the measured momentum distributions to those predicted by a classical model, we infer that the recolliding electron thermalizes with the $N-1$ bound electrons in less than 500 attoseconds. To our knowledge, no other method allows such a low upper bound for thermalization times within atoms to be measured. Nonsequential double and multiple ionization is defined by the fact that it is not sequential, that is, it is not the product of a sequence of uncorrelated single-ionization events. NSDI and NSMI require electron-electron correlation as a necessary precondition [@Ffm2000; @FrMBI2000]. Even for the very simplest such process – NSDI of helium – a fully quantum-mechanical description from first principles, i.e. by solution of the time-dependent Schrödinger equation in six spatial dimensions, has not been accomplished yet [@taylor], and for the heavier atoms it is clearly out of the question. This leaves approximate quantum-mechanical approaches, such as density-functional methods [@bauer], $S$-matrix methods that try to identify the most relevant terms of an appropriate perturbative expansion [@faisal; @KBRS], or classical-trajectory methods [@china; @eberly]. For *multiple* (triple and higher) ionization, which occurs under the same conditions as NSDI if the laser intensity is high enough, any description from first principles appears to be utterly out of reach. It seems equally hopeless, for three electrons and more, to identify the relevant diagrams in the microscopic $S$-matrix approach. The model we here propose is in the spirit of the one of Ref. [@KBRS], but essentially classical. We investigate the scenario wherein NSMI is effected by one single recollision. We assume that the pertinent electron tunnels into the continuum with zero velocity at the ionization time $t'$ according to the time-dependent rate $R(t')$, for which we adopt the standard quasi-static rate [@LL]. Thereafter, we turn to an entirely classical description: The laser field may drive the electron back to its parent ion at a later time $t$, which is a function $t(t')$ of the ionization time and can be easily evaluated. We assume that the energy $E_\mathrm{ret}(t)$ of the returning electron be completely thermalized among the ensemble of participating electrons, that is, the returning electron and the $N-1$ electrons to be freed. These $N$ electrons then form an excited complex with the total energy (with respect to the continuum threshold) $E_\mathrm{ret}(t) - E_0^{(N)}$, where $E_0^{(N)}>0$ denotes the total ionization potential of the $N-1$ (up to the recollision time $t$ inactive) electrons. The distribution of energy and momentum over the $N$ electrons is assumed to be completely statistical and only governed by the available phase space. At the time $t+\Delta t$, the $N$ electrons become free to move in the laser field, which is described by the vector potential ${\mathbf{A}}(t)$ such that ${\mathbf{A}}(t)=\mathbf{0}$ outside the pulse. The corresponding distribution of the final electron momenta ${\mathbf{p}}_n\ (n=1,\dots,N)$ is proportional to $$\begin{aligned} F({\mathbf{p}}_1,{\mathbf{p}}_2,\dots,{\mathbf{p}}_N) = \int dt' R(t') \delta \left(E_0^{(N)} -E_{\rm ret}(t)\right. \nonumber\\ \left.+{\frac{1}{2}}\sum_{n=1}^N [{\mathbf{p}}_n+{\mathbf{A}}(t+\Delta t)]^2 \right), \label{F}\end{aligned}$$ where the integral extends over the ionization time $t'$. The $\delta$ function expresses the fact that the total kinetic energy of the $N$ participating electrons at the time $t+\Delta t$ is fixed by the first-ionized electron at its recollision time $t$. This constitutes the one and only condition on the final momenta ${\mathbf{p}}_n$. The only free parameter of this model is the time delay $\Delta t$ between the recollision time and the time when the electrons become free. It is the sum of a thermalization time ${\Delta t_\mathrm{th}}$ – the time it takes to establish the statistical ensemble – and a possible additional “dwell time”, until the electrons become free. By comparing the predictions of the model with the data, we will be able to infer a value of $\Delta t$, which in turn provides an upper limit for the thermalization time ${\Delta t_\mathrm{th}}$. This model is an extension to NSMI of a classical model introduced for NSDI for $\Delta t=0$ [@FFetal04R; @FFetal04]. Sufficiently high above threshold, it produced momentum distributions that were virtually indistinguishable from their quantum-mechanical counterparts. Statistical models similar to the one above have been used in many areas of physics. For example, the statistical Rice-Ramsperger-Kassel-Marcus (RRKM) theory [@RRKM] describes thermalization of molecular vibrational degrees of freedom, and for high-energy collisions of elementary particles and heavy ions statistical models have been utilized to predict the momentum spectra of the reaction products [@hagedorn]. An excited complex as the doorway to NSMI was also considered in Ref. [@sachaeckh]. A convenient feature of the ansatz (\[F\]) is that integration over unobserved momentum components is easily carried out. To this end, we exponentialize the $\delta$ function in Eq. (\[F\]) with the help of its Fourier representation $$\delta (x) = \int_{-\infty}^\infty \frac{d\lambda}{2\pi} \exp (-i\lambda x).\label{delta}$$ Infinite integrations over the momenta ${\mathbf{p}}_n$ can then be done by Gaussian quadrature. The remaining integration over the variable $\lambda$ is taken care of by the formula [@GR] $$\int^\infty_{-\infty}\frac{d\lambda}{(i\lambda+\epsilon)^{\nu}} e^{ip\lambda}=\frac{2\pi}{\Gamma (\nu)}p^{\nu -1}_+,$$ where $x^\nu_+ = x^\nu \theta(x)$, with $\theta(x)$ the unit step function and $\epsilon \rightarrow +0$. For comparison with the experiments [@FrMBI2000; @multi], we calculate the distribution of the momentum ${\mathbf{P}}$ of the ion. Provided the momentum of the absorbed laser photons can be neglected, momentum conservation implies ${\mathbf{P}}=-\sum_{n=1}^N {\mathbf{p}}_n$, so that $$\begin{aligned} F_\mathrm{ion}({\mathbf{P}}) \equiv \int \prod_{n=1}^N d^3{\mathbf{p}}_n \delta \left({\mathbf{P}}+\sum_{n=1}^N {\mathbf{p}}_n \right) F({\mathbf{p}}_1,{\mathbf{p}}_2,\dots, {\mathbf{p}}_N)\nonumber\\ =\frac{(2\pi)^{\frac{3N}{2}-\frac{3}{2}}}{N^{3/2}\Gamma(\frac32(N-1))}\int dt' R(t') \left(\Delta E_{N,\mathrm{ion}}\right)^{\frac{3N}{2}-\frac{5}{2}}_+\label{ion}\end{aligned}$$ with $\Delta E_{N,\mathrm{ion}} \equiv E_\mathrm{ret}(t) -{E_0^{(N)}}-\frac{1}{2N}[{\mathbf{P}}-N{\mathbf{A}}(t+\Delta t)]^2$. In the experiments thus far, the ion momentum transverse to the direction of the laser polarization is entirely or partly integrated over. In the first case, the remaining distribution of the longitudinal ion momentum $P_\parallel$ is $$\begin{aligned} F_\mathrm{ion}(P_\parallel) \equiv \int d^2{\mathbf{P}}_\perp F_\mathrm{ion}({\mathbf{P}})\nonumber\\ =\frac{(2\pi)^{\frac{3N}{2}-\frac{1}{2}}}{\sqrt{N} \Gamma((3N-1)/2)}\int dt' R(t') \left(\Delta E_{\mathrm{ion}\parallel}\right)^{\frac{3N}{2} -\frac{3}{2}}_+,\label{ionpar}\end{aligned}$$ where now $\Delta E_{\mathrm{ion}\parallel} \equiv E_\mathrm{ret}(t) -{E_0^{(N)}}-\frac{1}{2N}[P_\parallel -NA(t+\Delta t)]^2$. If just one transverse-momentum component ($P_{\perp,2}$, say) is integrated while the other one ($P_{\perp,1} \equiv P_\perp$) is observed, the corresponding distribution is $$\begin{aligned} F_\mathrm{ion}(P_\parallel,P_\perp) \equiv \int dP_{\perp,2} F_\mathrm{ion}({\mathbf{P}})\nonumber\\ =\frac{(2\pi)^{\frac{3N}{2} -1}}{N\Gamma(3N/2-1)} \int dt' R(t') \left(\Delta E_{N,\mathrm{ion}\parallel\perp} \right)^{\frac{3N}{2}-2}_+\label{ionparperp}\end{aligned}$$ with $\Delta E_{N,\mathrm{ion}\parallel\perp} \equiv \Delta E_{\mathrm{ion}\parallel}-\frac{1}{2N}P_\perp^2$. ![(Color online) Distribution of the longitudinal ion momentum for triple (upper panel) and quadruple (lower panel) nonsequential ionization of neon at 2 PWcm$^{-2}$ calculated from Eq. (\[ionpar\]) for various delays $\Delta t$ as indicated in the lower panel. Note that in the upper panel the curves for $\Delta t=0$ and $\Delta t=0.1T$ almost completely overlap.[]{data-label="fig1"}](ThFig1.EPS){width="7cm"} ![The kinetic energy $E_\mathrm{ret}(t)$ (lower panel) and recollision time $t$ (upper panel) of electrons liberated at the tunneling time $t'$. The classically allowed intervals ${\Delta t_\mathrm{rec}}$ of the recollision time, for which the corresponding electron return energy $E_\mathrm{ret}(t)$ is greater than $E_0^{(3)}$ (and $E_0^{(4)}$), are shown in the upper panel. The calculation is for neon at 2.0PWcm$^{-2}$.[]{data-label="therm-time"}](thermtime2.EPS){width="7.5cm"} In Fig. \[fig1\] we present calculations of the ion-momentum distributions for triple (upper panel) and quadruple (lower panel) NSMI of neon according to Eq. (\[ionpar\]), for various values of $\Delta t$ between 0 and $0.2T$. The parameters are for the experimental data of neon presented in Fig. 2 of Ref. [@multi]. The figure displays the double-hump structure of the ion momentum in NSMI of neon. When the time delay $\Delta t$ increases from 0, initially, the effect on the momentum distribution is small. Later, however, the center positions of the humps start moving towards zero momentum and the widths of the humps increase until the two humps begin to merge. This behavior can easily be understood from the recollision kinematics, which are illustrated in Fig. \[therm-time\]. The final electrons undergo maximal acceleration if they are released near a zero crossing of the electric field, which in the figure occurs at $t=T$. For the example of triple ionization of neon, the earliest recollision with $E_\mathrm{ret}> E_0^{(3)}$ takes place at $t=0.74T$. Already with a delay of $\Delta t > 0.26T$, all electrons will be released after the zero crossing, which results in significantly lower ion momenta. ![(Color online) Distribution of the longitudinal momentum of nonsequential triple ionization of neon at 1.5 PWcm$^{-2}$ (upper panel), 2.0 PWcm$^{-2}$ (middle panel), and of nonsequential quadruple ionization of Ne at 2.0 PWcm$^{-2}$ (lower panel). The rugged (black) curves represent the data of Fig. 2 of Ref. [@multi]. The outermost smooth (red) curve and the innermost smooth (green) curve are calculated from Eq. (\[ionpar\]) for $\Delta t=0$ and $\Delta t=0.17T$. []{data-label="fig3"}](ThFig3.EPS){width="6.5cm"} ![Ion-momentum distribution of nonsequential triple ionization of neon at 1.0 PWcm$^{-2}$ for $\Delta t=0$ (upper panel) and $\Delta t=0.17T$ (lower panel), calculated from Eq. (\[ionparperp\]), to be compared with the data of Ref. [@FrMBI2000].[]{data-label="neon00"}](ThFig4.EPS){width="6.5cm"} In Fig. \[fig3\] we compare the results of the statistical model with the data of Ref. [@multi]. The intensities are those given for the experiment. We display momentum distributions calculated from Eq. (\[ionpar\]) for zero delay and for $\Delta t= 0.17T$. The latter value was chosen to yield optimal agreement for the entire set of data. We notice, in particular, that with this nonzero delay the model reproduces the maxima of the experimental ion-momentum distribution. This removes a longstanding discrepancy between models of the type discussed in Refs. [@FFetal04R; @FFetal04; @KBRS] and the data. For triple ionization, the calculated momentum distributions are wider than those of the data, in particular for the higher intensity, even though the data may include a contribution of the partly sequential channel Ne $\to$ Ne$^+ \to$ Ne$^{3+}$, or the contribution of a recollision-excitation channel, both of which are, of course, not part of the model. Figure \[neon00\] exhibits the results of our thermalization model for $N=3$ for the conditions of Ref. [@FrMBI2000]. The distribution of two components of the ion momentum is presented, the component parallel to the laser field and one transverse component, which therefore provides a more stringent test of the model. The third component is integrated over in the data, which corresponds to Eq. (\[ionparperp\]). The intensity given in the experiment is 1.5PWcm$^{-2}$; we obtain good agreement with the data for the lower intensity of 1.0PWcm$^{-2}$ and $\Delta t=0.17T$ as before [@footnote]. The nonzero delay $\Delta t$ has little effect on the transverse width of the distribution, but it causes an elongation in the longitudinal direction and moves the centers to lower momenta, markedly improving the agreement with the data. Encouraged by the good agreement between the model and the data, we interpret the delay for which we observed optimal agreement as an upper bound of the thermalization time ${\Delta t_\mathrm{th}}$, as argued above. This yields ${\Delta t_\mathrm{th}}< \Delta t_\mathrm{opt} = 0.17T \simeq 460$ as. A lower limit of the thermalization time should be given by the inverse of the plasma frequency for an electron density $\rho_\mathrm{e} = N$ in atomic units. This produces ${\Delta t_\mathrm{th}}< \sqrt{\pi/N}$, which is of the order of the atomic unit of time. All of the above discussion has been for neon. NSDI of argon appears to be governed by a recollision-excitation scenario [@ArvsNe]. We note in passing that for (quadruple) NSMI of argon we get good agreement of our statistical thermalization model with the data [@multi] for the experimental intensity and $\Delta t=0.35T$, twice as long as for neon. Details will be given elsewhere. To test the model further, and possibly to set a tighter upper limit on the thermalization time, it is necessary to restrict the time range $\Delta t_\mathrm{CM}$ of recollision (Fig. \[therm-time\]). If thermalization is more rapid than 460 as, then the width of the ion-momentum distribution will decrease as we restrict ${\Delta t_\mathrm{rec}}$. There are a number of ways to minimize ${\Delta t_\mathrm{rec}}$. Within limits, all that is needed is to lower the light intensity or to increase the laser frequency. However, the best way to minimize ${\Delta t_\mathrm{rec}}$ and to control the time of recollision is to use a second harmonic field, polarized perpendicular to the fundamental. The combined requirement that the electron and ion recollide in both directions allows the time of recollision to be precisely determined and controlled via the relative phase of the two beams. The limit on the thermalization time that we determine (as well as the much tighter bounds that seem feasible in future experiments) should apply to stationary electron-atom scattering in general. From a collision-physics perspective, as a result of streaking, the time-dependent laser field reveals information that is hard to obtain by other means. The streaking principle should also be applicable for studying nuclear dynamics. As in NSMI, nuclear processes also can be initiated by laser-controlled re-collision [@nuclear] (of course, at much higher intensity). Any nuclear decay process that results in a mass or charge change of the fragments will be streaked by the laser field just as electrons and ions are streaked in our case. To summarize, by comparison of experimental data for triple and quadruple nonsequential ionization of neon with a simple statistical recollision model where the returning electron thermalizes with a subset of the bound electrons, we have been able to conclude that (i) for neon such a model appears to contain the most relevant physics, and (ii) the time for this thermalization to occur is extremely fast, well below one femtosecond. We gratefully acknowledge stimulating discussions with G.G. Paulus, H. Rottke, and W. Sandner. [99]{} P.B. Corkum, Phys. Rev. Lett. **71**, 1994 (1993). H. Niikura, F. Legare, R. Hasbani, A.D. Bandrauk, M.Yu. Ivanov, D.M. Villeneuve, and P.B. Corkum, Nature (London) **417**, 917 (2002). S.N. Yurchenko, S. Patchkovskii, I.V. Litvinyuk, P.B. Corkum, and G.L. Yudin, Phys. Rev. Lett. **93**, 223003 (2004). A. Baltuška, Th. Udem, M. Uiberacker, M. Hentschel, E. Goulielmakis, Ch. Gohle, R. Holzwarth, V.S. Yakovlev, A. Scrinzi, T.W. Hänsch, and F. Krausz, Nature (London) **421**, 611 (2003). R. Moshammer, B. Feuerstein, W. Schmitt, A. Dorn, C.D. Schröter, J. Ullrich, H. Rottke, C. Trump, M. Wittmann, G. Korn, K. Hoffmann, and W. Sandner, Phys. Rev. Lett. **84**, 447 (2000). A. Rudenko, K. Zrost, B. Feuerstein, V.L.B. de Jesus, C.D. Schröter, R.Moshammer, and J. Ullrich, Phys. Rev. Lett. **93**, 253001 (2004). E. Constant, V.D. Taranukhin, A. Stolow, and P.B. Corkum, Phys. Rev. A **56**, 3870 (1997). Th. Weber, M. Weckenbrock, A. Staudte, L. Spielberger, O. Jagutzki, V. Mergel, F. Afaneh, G. Urbasch, M. Vollmer, H. Giessen, and R. Dörner, Phys. Rev. Lett. **84**, 444 (2000). K.T. Taylor, J.S. Parker, K.J. Meharg, and D. Dundas, Eur. Phys. J. D **26**, 67 (2003). D. Bauer and F. Ceccherini, Opt. Express **8**, 377 (2001). A. Becker and F.H.M. Faisal, Phys. Rev. Lett. **84**, 3546 (2000); *ibid.* **89**, 193003 (2002). R. Kopold, W. Becker, H. Rottke, and W. Sandner, Phys. Rev. Lett. **85**, 3781 (2000). L.B. Fu, J. Liu, S.G. Chen, Phys. Rev. A **65**, 021406(R)(2002); J. Chen and C.H. Nam, Phys. Rev. A **66**, 053415 (2002). P.J. Ho, R. Panfili, S.L. Haan, and J.H. Eberly, Phys. Rev. Lett. **94**, 093002 (2005). L.D. Landau and E.M. Lifshitz, *Quantum Mechanics (Nonrelativistic Theory)* (Pergamon Press, Oxford, 1977). C. Figueira de Morisson Faria, X. Liu, W. Becker, and H. Schomerus, Phys. Rev. A **69**, 012402(R)(2004). C. Figueira de Morisson Faria, H. Schomerus, X. Liu, and W. Becker, Phys. Rev. A **69**, 043405 (2004). See, e.g., W. Forst, *Theory of Unimolecular Reactions* (Academic, New York, 1973); P.J. Robinson and K.A. Holbrook, *Unimolecular Reactions* (Wiley-Interscience, New York, 1972). R. Hagedorn, Nuovo Cimento Suppl. **3**, 147 (1965). K. Sacha and B. Eckhardt, Phys. Rev. A **64**, 053401 (2001); J. Phys. B **36**, 3923 (2003). I.S. Gradsteyn and I.M. Ryzhik, *Table of Integrals, Series, and Products* (Academic Press, New York, 1980). A delay of the same order has also been observed in classical-trajectory simulations of NSDI of helium [@eberly]. V.L.B. de Jesus, B. Feuerstein, K. Zrost, D. Fischer, A. Rudenko, F. Afaneh, C.D. Schröter, R. Moshammer, and J. Ullrich, J. Phys. B **37**, L161 (2004). N. Milosevic, P.B. Corkum, and Th. Brabec, Phys. Rev. Lett. **92**, 013002 (2004); S. Chelkowski, A.D. Bandrauk, and P.B. Corkum, Phys. Rev. Lett. **93**, 083602 (2004).

ArXiv

--- abstract: 'Recent work on the engines of active galactic nuclei jets suggests their power depends strongly and perhaps counter-intuitively on black hole spin. We explore the consequences of this on the radio-loud population of active galactic nuclei and find that the time evolution of the most powerful radio galaxies and radio-loud quasars fits into a picture in which black hole spin varies from retrograde to prograde with respect to the accreting material. Unlike the current view, according to which jet powers decrease in tandem with a global downsizing effect, we argue for a drop in jet power resulting directly from the paucity of retrograde accretion systems at lower redshift $z$ caused by a continuous history of accretion dating back to higher $z$. In addition, the model provides simple interpretations for the basic spectral features differentiating radio-loud and radio-quiet objects, such as the presence or absence of disk reflection, broadened iron lines and signatures of disk winds. We also briefly describe our models’ interpretation of microquasar state transitions. We highlight our result that the most radio-loud and most radio-quiet objects both harbor highly spinning black holes but in retrograde and prograde configurations, respectively.' date: Released 2010 Xxxxx XX title: The Evolution of Radio Loud Active Galactic Nuclei as a Function of Black Hole Spin --- \[firstpage\]  Galaxy evolution - black hole spin  Introduction ============ Over the past decade and a half, our understanding of the dynamics of active galaxies (AGN) has made significant strides. Once of marginal interest astrophysically, black holes have taken center stage, becoming an integral part of galactic dynamics and evolution. The picture that has emerged involves the presence of supermassive black holes at the center of most if not all galaxies, with active galaxies interacting with these black holes via accretion, producing winds and jets. There is overwhelming evidence that galaxies are interconnected with black holes to the extent that the cosmic evolution of both is coupled (Kormendy & Richstone 1995; Magorrian et al 1998; Gebhardt et al 2000; Ferrarese & Merritt 2000; Tremaine et al 2002; Marconi & Hunt 2003). Despite this, there are uncertainties in our understanding of how black hole engines produce jets, how such structures are collimated over large scales, and how galaxies that harbor these powerful engines evolve over cosmic time. The theoretical framework in which we have attempted to address such questions over the past two decades involves the “spin paradigm” (Blandford 1990; Wilson & Colbert, 1995; Moderski, Sikora & Lasota, 1998), whereby high black hole spin can lead to powerful, radio-loud, jetted AGN, while low black hole spin to radio-quiet, weak or non-jetted AGN where jet power is related to black hole spin via the Blandford-Znajek mechanism (Blandford & Znajek 1977; henceforth BZ). The literature on the BZ effect and black hole spin-related processes for jet formation in relation to radio-loud AGN is extensive. In its most recent guise, the spin paradigm enters the literature in various places and with differing results. Using the BZ mechanism, Mangalam, Gopal-Krishna & Wiita (2009) propose a scenario leading to low black hole spins in advection dominated accretion flows under the assumption that the magnetic flux threading the black hole is insensitive to spin. Similar ideas grounded in general relativistic magnetohydrodynamic simulations suggest that high spin for higher order spin power dependence in thick accretion geometries vs thin accretion flows, produce the conditions for the $3$ order of magnitude difference in power between radio-loud and radio-quiet AGN (Tchekhovskoy et al 2010). In work that closely follows the results of numerical simulations that supports the work of Nemmen et al (2007) on high black hole spin in geometrically thick flows, Benson & Babul (2009), also using advection dominated thick accretion flow models, arrive at a maximum spin of $0.92$ for black holes, suggesting that a spin-equilibrium scenario is compatible with the linear accretion rate dependence of jet power found in recent observations of AGN jets (Allen et al 2006). In addition to the question of the radio-loud/radio-quiet division in active galaxies, a division within the group of jetted AGN is observed and a classification is produced according to differences in power and jet collimation (Fanaroff & Riley 1974). Attempts at understanding this Fanaroff-Riley (FR) division have also been framed in the context of the spin-paradigm in an attempt to explore the possibility of differences in spin between FRII and FRI objects. Meier (1999) suggested a magnetic switch operating at high spin triggers the production of relativistic jets. Using the BZ effect and hybrid Meier model, Daly (2009b) calculated the cosmological evolution of black hole spins, showing that they are largest (about unity) for the FRII sources at high redshift and decrease (to about 0.7) for radio-loud sources at lower redshift. The idea that the engine of jets determines their FR morphology spans the works from Rees (1982) to Hardcastle et al (2007). A theoretical approach to determining black hole spin that is model-independent, but assumes that spin changes only by extraction of the reducible black hole mass, applied to a small subset of powerful radio galaxies, finds that they harbor low spinning black holes (Daly 2009a). This work suggests also that the FRI/FRII division is due to environment. Such environmental effects on jet morphology, unrelated at least directly to the properties of the central engine of the jet, have been addressed starting with De Young (1993) and more recently with Gopal-Krishna and Wiita (2000). Black hole mass dependence (Ghisellini & Celotti 2001) and accretion dependence of the FRI/FRII division have also been explored (Marchesini et al 2004). Under the assumption that jets are produced by the combined effort of the BZ and Blandford-Payne (Blandford & Payne 1982; henceforth BP) effects, recent numerical studies of general relativistic magnetohydrodynamics of black hole accretion flows suggest a tight link between jet power and black hole spin (Garofalo, 2009a, 2009b). Whereas BZ involves jets produced via spin-energy extraction from the black hole, BP is a jet mechanism that originates in the accretion disk via mass-loading of disk gas onto large scale magnetic fields. The black hole spin dependencies of these processes have direct implications for the jets produced in AGN. According to these studies, largest BZ power occurs for highly retrograde accretion systems with respect to the black hole and less so for prograde ones, transitioning to zero power at zero spin while the BP power is also maximized for highly retrograde accretion but monotonically decreases toward high prograde ones. In the context of this framework, we attempt in this paper to construct a physical foundation for the morphology and evolution of jet-producing AGN. At the heart of our picture lies the notion that accretion onto spinning supermassive black holes tends, in time, to produce increasingly prograde accretion systems, in which lower jet output is hosted by more stable accretion configurations. In terms of the radio-loud AGN population, “high excitation radio galaxies”, evolve over cosmic timescales toward “low-excitation radio galaxies”, as black holes conspire with their host galaxy in evolving toward prograde accretion states from an earlier phase of retrograde accretion flow. In addition, we suggest a retrograde vs prograde division for the engines of radio-loud objects whose interaction with the external environment produces the distribution of the Fanaroff & Riley classification, according to which the powerful, highly collimated jets of the FRII class, are retrograde spin systems while the mostly less powerful, less collimated FRI sources embedded in gas-rich environments are the late-state evolution toward prograde accretion systems. Our interpretation of the radio-loud/radio-quiet division will be that the most radio-loud and the most radio-quiet AGN both harbor rapidly rotating black holes with their accretion angular momentum vectors determining the difference, while the FRI/FRII division in our model is based on a combination of nuclear properties and environment. In addition to this, our model unifies radio-loud, radio-quiet, and FRI, FRII objects, to the extent, as we will illustrate, that radio-loud objects of the FRII class evolve either into radio-loud FRI objects or into radio-quiet AGN. We point out, furthermore, that the theoretical frameworks grounded in the spin-paradigm discussed above, all suffer from a “spin paradox” (David L. Meier) to be discussed in a follow-up paper in which we illustrate its resolution within our framework (Meier & Garofalo, in preparation). In section \[observation\] we highlight the relevant observed properties of radio-loud AGN which constitute the pieces of the theoretical puzzle we construct in Section \[Theory\]. Section \[Conclusions\] briefly addresses the radio-loud/radio-quiet dichotomy in this model and its extension to microquasars and then concludes. Radio-loud AGN: Observational picture ===================================== In this section we emphasize the main observational features concerning the radio-loud population and give a broad-brush description of past attempts and difficulties in combining such features into a theoretical framework. \[observation\] The Excitation Dichotomy in Radio-Loud AGN ------------------------------------------ Extensive optical and X-ray surveys show clear evidence for a fundamental dichotomy in the properties of radio-loud AGN, which is directly related to the mode of accretion onto the central supermassive black hole. “High-excitation radio galaxies” (HERGs), those with prominent emission lines in their optical spectra, have standard, geometrically thin accretion disks and accrete at a significant fraction of their Eddington limits. These sources are heavily obscured in the X-ray by columns in excess of $10^{23}$ cm$^{-2}$, consistent with AGN unification (Donato et al. 2004; Evans et al. 2006). HERGs show weaker, if any, neutral, inner-disk, broadened Compton reflection continuua compared to radio-quiet AGN (Reeves & Turner 2000; Grandi et al. 2002). Furthermore, most of these sources have narrow, typically unresolved neutral Fe K$\alpha$ lines (Evans et al. 2004, 2006), which indicates that the primary X-ray emission is being reprocessed far from the inner disk regions. HERGs tend to inhabit isolated environments, at least at low redshift, and often show evidence for recent mergers, consistent with the idea that they derive their power from the accretion of cold gas (Hardcastle, Evans, & Croston 2007). [*Spitzer*]{} IRS spectroscopy (Ogle et al. 2006) shows that HERGs are luminous MIR emitters, and the detection of strong 9.4$\mu$m silicate absorption implies that they possess molecular tori. On the other hand, “low-excitation radio galaxies” (LERGs) – those with few or no observed optical emission lines – lack any of the features required by standard AGN unification models. Their X-ray emission is dominated by a parsec scale jet, they have radiatively inefficient accretion flows ($L/L_{Edd}\sim10^{-(5-7)}$ - where $L$ is luminosity and $L_{Edd}$ is the Eddington luminosity), and they show no evidence at all for an obscuring torus (Hardcastle, Evans, & Croston 2006; Ogle et al 2006). These sources tend to inhabit hot, gas-rich environments, such as groups and clusters. It has been suggested by Hardcastle, Evans, & Croston (2007) that the jet outbursts in LERGs derive their power from the Bondi accretion of this IGM/ICM gas. This result holds for the majority of LERGs, but cannot be applied to the most powerful jets in clusters of galaxies, such as MS0735.6+7421 (McNamara et al. 2009), whose kinetic power exceeds that available from Bondi accretion. Finally, while LERGs show no evidence of ionized outflows of gas or winds from their central regions, evidence is emerging for such outflows in HERGs, at least in X-rays (Reeves et al 2009; Tombesi et al submitted), although less so than for their radio-quiet counterparts. Low-excitation (LERG) High-excitation (HERG) ------------------------------- -------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------- Definition No narrow optical line emission. Prominent optical emission lines, either narrow (NLRG) or broad (BLRG), or quasar. Fanaroff-Riley classification Almost all FRIs are LERGs at $z$=0. Significant population of FRIIs at $z \sim$0.5. Most FRIIs are HERGs, as are a handful of FRIs (e.g., Cen A). X-ray spectra Jet-related unabsorbed power law only. Upper limits only to ’hidden’ accretion-related emission. Jet-related unabsorbed power law + significant accretion contribution (heavily absorbed in NLRGs). Accretion-flow type Highly sub-Eddington. Likely radiatively inefficient. Reasonable fraction of Eddington. Likely standard accretion disk. Optical constraints Strong radio/optical/soft X-ray correlations. Optical emission is jet-related. Strong radio/optical/soft X-ray correlations. Optical emission is jet-related. : Overview of the properties of low- and high-excitation radio galaxies \[summary\] The Fanaroff-Riley Dichotomy in LERGs and HERGs and its Redshift Dependence --------------------------------------------------------------------------- HERGs and LERGs can be further classified according to whether they belong to the Fanaroff & Riley classification FRI or FRII (Fanaroff & Riley 1974), whose differences are related to the power of the jet and the physics of mass-loading (entrainment). FRI-type radio galaxies display “edge-darkened” radio morphology, with generally weak jets that are poorly collimated on kpc-scales. FRIIs are typically more powerful and more collimated. Whereas most FRIIs are HERGs and most FRIs are LERGs, [*Chandra*]{} observations indicate the existence of mixed FRII LERG and FRI HERG states. FRI HERGs are very rare. One example is the nearest AGN, Centaurus A (Evans et al. 2004). On the other hand, there is a significant population of FRII LERGs, which occur at what is arguably an intermediate redshift ($z=0.5-1$) for these objects, and essentially zero cases at low redshift ($z < 0.1$). FRII LERGs tend to lie in gas-rich groups or cluster-scale environments. The appearance of radio-loudness in the forms discussed above and the connection to radio-quiet AGN forms the subject of the following sections. Combinations of radiatively inefficient/efficient accretion flow with high/low spinning black holes (the spin paradigm) are only partially successful in modeling the observations. The observed weakness of disk winds and broadened reflection features in FRII HERGs compared to radio-quiet objects remains unexplained. Recent work points to a possible resolution of such problems by suggesting that radiative efficiency in very high accretion rate systems may be accompanied by fully ionized inner regions, thereby explaining the weakness of the reflection component (Sambruna et al 2009 and references therein). Explaining LERG jet systems in gas-rich clusters and the energetics of the FRII class is also problematic as mentioned above. For LERG jets, the observed linear relation between jet power and accretion power (Allen et al 2006) is expected assuming the BZ effect is operating but energetically can only be explained if the black holes are maximally spinning (Nemmen et al 2007; Benson & Babul 2009), which, in turn, implies there is fine-tuning in the spin parameter. The attempt to extend this linear relation between BZ jet power and accretion rate fails for the FRIIs because the observed accretion power is too weak. And finally, models suggesting high spin leads to powerful jets are statistically incompatible with mounting evidence for high-spinning black holes in Seyfert galaxies as it becomes problematic to assume such objects are all currently living within their radio-quiet phase (Wilms et al 2001; Brenneman & Reynolds 2006; Fabian et al 2009; Zoghbi et al 2010). In the next section we attempt to produce a black hole engine-based theoretical framework that addresses these observed properties. We will argue that misaligned, or retrograde systems, in which disk gas rotates opposite to that of the black hole, create the conditions for powerful jets albeit in dynamically unstable configurations, whose time evolution is toward energetically more stable conditions in which the black hole and disk coexist in rotationally aligned states. We will suggest that in addition to radiative efficiency or lack thereof and black hole spin, alignment vs misalignment between the angular momentum of the black hole and the disk, is crucial. Radio loud AGN: Theoretical framework ===================================== In this section we describe the theoretical framework within which we fit the observational elements of the previous section. The fundamental distinguishing feature of the theory (Garofalo, 2009a) involves the change in size of the gap region that exists between the inner edge of accretion disks located near the innermost stable circular orbit (ISCO) and the black hole horizon. The size of the gap region changes as a function of black hole spin because the ISCO depends on spin. We quantitatively determine the effect that the change in size of the gap region has on the BZ mechanism, on the BP mechanism, and on the overall effect of the gap region on jet power and collimation under the assumption that BZ and BP are the dominant and mutually dependent mechanisms involved in jet production. Finally, we include the effect of the gap region on accretion efficiency. We should emphasize that the fundamental feature of our model involves the inverse relationship between accretion efficiency and jet efficiency, and that section \[gap\_paradigm\] is focused on illustrating the dependence of such efficiencies on the size of the gap region. \[Theory\] BZ and BP: quantitative approach -------------------------------- We assume the spacetime is that of a rotating black hole and proceed in Boyer-Lindquist coordinates for which the Kerr metric takes the form, $$\begin{aligned} dS^{2}&=&-\left(1-\frac{2Mr}{\rho^{2}}\right) dt^{2}-\frac{4Mar\sin^{2}\theta}{\rho^{2}}dt\,d\phi\\ &+&\frac{\Sigma}{\rho^{2}}\sin^{2}\theta\, \nonumber d\phi^{2}+\frac{\rho^{2}}{\Delta}dr^{2}+\rho^{2}d\theta^{2},\end{aligned}$$ where $M$ is black hole mass, $a$ is the dimensionless spin parameter, $$\rho^{2} = r^{2} + a^{2}\cos^{2}\theta,$$ $$\Delta = r^{2}-2Mr+a^{2},$$ and $$\Sigma = (r^{2}+a^{2})^{2}-a^{2}\Delta\sin^{2}\theta.$$ The basic equation describing the evolution of the large-scale magnetic field within a Novikov & Thorne (1973) acccretion disk is obtained by following the relativistic analogue (Garofalo 2009b) of the non-relativistic treatment of Reynolds et al (2006), by combining Maxwell’s equation $$\triangledown_{b}F^{ab} = \mu J^{a},$$ with a simplified Ohm’s law $$\label{Ohm} J^{a} = \sigma F^{ab}u_{b},$$ where $F^{ab}$ is the standard Faraday tensor, $\mu$ is the permeability of the plasma, $J^a$ the 4-current, $u^a$ the 4-velocity of the accretion disk flow, and $\sigma$ the effective conductivity of the turbulent plasma. This gives $$\triangledown_{b}F^{ab} = \frac {1}{\eta} F^{ab}u_{b}, \label{disk}$$ where $\eta = 1/\mu\sigma$ is the effective magnetic diffusivity. The equations are cast in terms of the vector potential, which is related to the Faraday tensor via $$F_{ab}=A_{b,a}-A_{a,b}, \label{vec_pot}$$ and, in particular, in terms of the component $A_{\phi}$ in the coordinate basis of the Boyer-Lindquist coordinates. Ultimately, to examine BZ and BP powers, we need to derive the magnetic flux and angle of the flux contours threading a hoop placed at a given radius $r$. The magnetic flux function is related to the vector potential via Stokes’ Theorem applied to the Faraday tensor $$\psi \equiv \int_{S}F=\int_{S}dA=\int_{\partial{S}}A=2 \pi A_{\phi},$$ where $S$ is a space-like surface with boundary $\partial S$ consisting of a ring defined by $r={\rm constant}$, $\theta={\rm constant}$, and $t={\rm constant}$. Since $A_{b}$ is specified up to the gradient of a scalar function $\Gamma$, $$A_{b}^{'} = A_{b} + \triangledown _{b} \Gamma,$$ the assumption of time-independence and axisymmetry gives us $$A^{'}_{t} = A_{t}$$ and $$A_{\phi}^{'} = A_{\phi}.$$ Thus, we need not specify the gauge uniquely beyond the statement of t and $\phi$ independence. The region outside the black hole and accretion disk is modeled as force-free, satisfying $$F^{ab}J_{b} = 0 \label{force-free}$$ and $$\triangledown_{b}F^{ab} = \mu J^{a}. \label{magnetosphere1}$$ Outside the accretion disk we also impose the ideal MHD condition $$F^{ab}u_{b} = 0, \label{ideal}$$ where $u^b$ is the 4-velocity of the (tenuous) plasma in the magnetosphere and is determined by the condition that field lines rigidly rotate. The numerical details can be found in Garofalo 2009b. We now consider the BZ and BP powers that result from the numerical solution of the above equations under the assumption that the gap region is not threaded by magnetic flux (i.e. the Reynolds conjecture of a zero-flux boundary condition - Reynolds et al 2006; Garofalo 2009b), where this ability of the gap region to enhance magnetic flux on the black hole is supported also by general relativistic MHD simulations (McKinney & Gammie, 2004). The foundation of the Reynolds conjecture stems from noting that within the gap region, circular orbits are no longer stable and the accretion flow plunges into the black hole. In Reynolds et al (2006), it is argued that the inertial forces within the gap region prevent magnetic flux that is threading the black hole from expanding back into the disk. Accretion of magnetic field can result in a strong flux-bundle threading the black hole, confined in the disk plane by the gap region. We start by evaluating the horizon-threading magnetic field as measured by ZAMO observers from the flux values we obtain, $$\label{zamo_b} B_{H}=\sqrt{g_{11}}B^{r}$$ with $$B^{r} = \ast F^{rb}u_{b},$$ where $\ast F^{ab}$ is the dual Faraday tensor and $u^{b}$ is the four-velocity of the ZAMO observers evaluated in the equatorial plane on the stretched horizon or membrane in the sense of the Membrane Paradigm (Thorne et al 1986), and from this magnetic field value we determine BZ power as (Thorne et al 1986), $$L_{BZ} = 2 \times 10^{47}ergs \; s^{-1}(\frac{B_{H}}{10^{5}G})^{2}m_{9}^{2}j^{2}$$ where $B_{H}$ is the poloidal magnetic field threading the black hole, $j$ is the normalized angular momentum of the black hole or $a/M$, and $m_{9}$ is the black hole mass in units of $10^{9}$ solar masses. BP power is similarly constructed but depends on magnetic field strength threading the accretion disk, the bend angle with which the magnetic field presents itself to the disk surface, or more precisely, the radial extent over which the bend angle is sufficient for mass-loading, and the Keplerian rotation rate which is also spin-dependent. Following Cao (2003) we have $$L_{BP} = \int B_{d}^{2} r^{2} \Omega dr$$ where $B_{d}$ is the field strength threading the disk and $\Omega$ is the rotation of the disk magnetic field. The radial extent over which the bend angle is large enough for mass-loading, increases with increase of spin in the retrograde regime and comes from the numerical solution (Garofalo 2009b). We show the power in BZ in Figure \[L\_BZ\], that in BP in Figure \[L\_BP\] and the overall power in Figure \[Ledlow\] assuming they combine as follows, where the functions $\alpha$ and $\beta$ capture the effects on BP and BZ powers from the numerical solution in the context of the Reynolds conjecture (Garofalo 2009b). $$L_{jet} = 2 \times 10^{47}ergs \; s^{-1} \alpha\beta^{2}(\frac{B_{d}}{10^{5}G})^{2}m_{9}^{2}j^{2} \label{jet_power}$$ where $$\alpha = \delta(\frac{3}{2}-j)$$ and $$\beta = -\frac{3}{2}j^{3} + 12j^{2} - 10j + 7 - \frac{0.002}{(j-0.65)^{2}} +\frac{0.1}{(j+0.95)} +\frac{0.002}{(j-0.055)^{2}}.$$ While $j$ spans negative values for retrograde spin and positive values for prograde spin, a conservative value for $\delta$ of about $2.5$ is adopted but our ignorance of how jets couple the BZ and BP components restrict our ability to specify it and suggest that it might well be larger by an order of magnitude or more. While $\alpha$ can be thought of as the parameter that determines the effectiveness of the BP jet as a function of spin, $\beta$ captures the enhancement on the black hole of the disk-threading field, both within the context of the Reynolds conjecture. The parameter $\delta$ determines the effective contribution of BP to overall jet power, with larger values shifting jet power efficiency more toward the retrograde regime. Because, as we point out in the next section, our focus is on the collimating properties of BP jets, we have been conservative in choosing a small $\delta$ of order unity. It is important to point out that the no-flux boundary condition’s effect is to increase the BZ and BP efficiency toward the retrograde regime. The consequence of this for prograde spin is to produce a flattening of the overall jet power so that low prograde black hole spins have powers that increase. The behavior of the power in Figure \[Ledlow\] follows from the assumption of a force-free magnetosphere but changes if this assumption is relaxed. A non-negligible inertia for fast rotating magnetospheres and the added centrifugal barrier that would ensue, would make it more difficult for plasma to be advected inward through the diffusionless gap region, a situation that arises more for the high prograde regime as the rotation is greater. As a result, a more realistic magnetosphere should further shift the overall jet power efficiency toward the retrograde regime, thereby further flattening the spin dependence in the prograde one. Jets vs. accretion: the gap paradigm ------------------------------------ We combine these results and produce the following illustration of the combined effect of BZ and BP on jet power and collimation. 1. In the left column in Figure \[BZ\] we show the difference between accretion systems around spinning black holes as a function of the spin parameter and type of accretion (i.e. prograde vs. retrograde) and its effect on the BZ power. We use negative spin values to indicate retrograde accretion while positive spin values indicate prograde accretion. Note how the gap region is larger as the spin increases from large prograde to large retrograde. The shrinking of the gap region in the prograde direction, produces an asymmetry in BZ power that favors retrograde systems. As a result, retrograde BZ power is larger as indicated by the length of the arrows. The basic physics behind this power dependence on spin is that larger gap region allows greater magnetic flux accumulation on the black hole (Garofalo, 2009b). 2. In the center column of Figure \[BZ\] we present a schematic of the monotonic dependence of BP on spin by using arrows that originate in the disk (since BP jets are disk-launched). The length again indicates the magnitude of the jet and we highlight how the power increases from high prograde to high retrograde. Here, as well, the size of the gap region is instrumental in that larger gap regions produce greater magnetic flux on the black hole, which leads to greater bending of inner-disk-threading magnetic field lines, which, in turn, leads to greater mass outflow and more effective BP jets (Garofalo 2009a). 3. By considering both the left and central columns of Figure \[BZ\] we combine the effects and produce an overall picture for the jets in retrograde vs prograde black hole accretion systems. Assuming that BP jets originating in the disk/corona are essential for jet collimation (McKinney & Narayan 2007; Bogovalov & Tsinganos 2005; Meier et al 2001), and that such collimation is directly related to jet acceleration (Tchekhovskoy, McKinney & Narayan, 2009), the strongest and most collimated jets occur for high retrograde systems while high prograde systems produce low-power, weakly collimated jets due to the weakness of BP, despite the presence of strong BZ. Our heuristic, qualitative scenario proposes that the jet itself comes from the BP effect but is sparked, in some sense, by the BZ mechanism. It is worth pointing out that we are not illustrating a precise mechanism for producing jets, but simply highlighting the compatibility that arises between theory and observation if BP and BZ conspire as described above. In addition, surviving MHD disruptions such as the kink mode instability (Nakamura & Meier 2004), appears to occur in combination with the presence of a disk jet and relativistic bulk motion (Hardee & Hughes, 2003; McKinney & Blandford, 2009) The diagram, therefore, illustrates these ideas with uncollimated jets for the high prograde case resulting from small BP, no jet in the zero spin case, and a powerful, collimated jet for the high retrograde case resulting from strong BP. We highlight two aspects of the BZ/BP scenario. The first is that whereas jet power displays a roughly flat spin dependence for a range of prograde spin values, this does not occur for retrograde spin. This indicates that unlike for retrograde spins, there is a range of prograde spin values for which jet power is weakly dependent on spin. We also re-emphasize the fact that BP jets do not simply add to the overall jet power according to our picture; they serve the fundamental purpose of collimation (a feature motivated in the diagrams of Figure \[BZ\]). In addition to the effect that the location of the disk inner edge has on jet production, the size of the gap region is also connected to the energetics of the disk itself. Recent work extending standard, radiatively efficient accretion disk theory to self-consistently include the effects of magnetized coronae and disk winds, shows that larger gap regions inhibit or limit the presence of disk winds (Kuncic & Bicknell 2004, 2007). The further out in radial position for the disk inner edge, the less gravitational power is available to be reprocessed in the disk to produce mass outflows further out in the disk. This is illustrated in the right column of Figure \[BZ\]. With only these considerations on the spin-dependent size of the gap region, we introduce the predictions of our model for the cosmological evolution of AGN (S.\[cosmo\_evolution\]), and its implications for the Fanaroff-Riley dichotomy (S.\[Fanaroff-Riley\]) and the observed X-ray characteristics of radio-loud AGN (S.\[Fe\_alpha\]). \[gap\_paradigm\] Cosmological Evolution of Radio-Loud AGN ---------------------------------------- Our story focuses on the fraction of objects that involve retrograde accretion onto rapidly spinning black holes, most likely formed in mergers of equal mass black holes or in mergers of black holes where the larger one spins rapidly and the merger with the smaller one is in the prograde direction (Hughes & Blandford 2003). As long as the angular momenta of the disk, $J_{d}$, and that of the black hole, $J_{h}$, satisfy $$cos\theta < -\frac {J_{d}}{2J_{h}},$$ where $\theta$ is the angle between the two angular momentum vectors (King et al. 2005), counteralignment between disk and black hole occurs (i.e. $\theta=\pi$) and a retrograde accretion state is formed. Because the merger produces a gas-rich environment mainly in the form of cool, molecular gas, that feeds the black hole at relatively high accretion rates (Barnes & Hernquist 1991), we associate this initial phase with HERGs. In other words, the relatively high accretion rates are such that the accretion flow is close to a standard, radiatively efficient Novikov & Thorne disk. The powerful, highly collimated jet, on the other hand, is a direct consequence of the fact that the system is in a highly retrograde accretion state (i.e. the black hole engine is operating at maximum efficiency due to the presence of strongest BZ and BP effects). In other words, these FRII HERGs are powered by nuclear engines in the lower panel of Figure \[BZ\]. We now consider the evolutionary paths of two initial FRII HERGs whose jet powers differ, due perhaps to different ratios of black hole mass to the amount of post-merger cold gas. Recent work indicates that high redshift, $z\sim2$, radio-loud quasars, can deliver $\sim10\%$ of the jet energy to the ISM, sufficient to expel the cold gas via outflows up to 1000 $km/s$ in $\sim10^{7}yrs.$ (Nesvadba et al. 2008). If the loss of cold gas due to this radio-mode AGN feedback leads to a drop in the pressure and density of the ISM, it is expected that accretion will switch to hot Bondi-fed advection-dominated accretion flow (ADAF - Narayan & Yi 1995) from cold, thin-disk accretion at higher redshift as galaxies expand to $\sim3$ times their size (Mangalam, Gopal-Krishna & Wiita 2009 and possibly as envisioned in Antonuccio-Delogu & Silk 2010). If the initial FRII HERG is of lower jet power (Fig. \[BP\]), the expulsion of cold gas is less effective, and the transition to hot Bondi-fed ADAF accretion is slower. In this case, the initial FRII HERG transitions to an FRI HERG. The FRI nature of the system originates in the fact of prograde accretion as mentioned in the Introduction, while the HERG label corresponds to the relatively high accretion rate and thus to radiatively efficient accretion flow. In other words, the spin-up toward the prograde direction occurs faster than the transition from radiatively efficient accretion to ADAF accretion. If, on the other hand, the initial FRII HERG jet is more powerful and more effective in expelling the cold gas, the initial FRII HERG transitions to an FRII LERG, due to the fact that the system remains retrograde but the accreting gas is now hot and enters the ADAF phase (i.e. the black hole has not accreted enough to have been spun down and then up again in the prograde regime-Fig. \[BZ+BP\]). There are two basic processes working together here. The first involves accretion feeding a black hole in a retrograde state. [*Continued accretion initially spins the black hole down toward zero spin and then up again in a prograde state where the angular momentum vector of the black hole is parallel to that of the accretion flow.*]{} The other process involves the accretion state itself. If the accreting gas is cold, the accretion flow radiates in an efficient manner. However, as described, the FRII HERG jets expel cold gas and the timescales on which they accomplish this depends on jet power. Therefore, more or less powerful FRII HERG, cold gas-expelling jets, create the conditions that produce a mixing of FRIIs and FRIs with the HERG and LERG states. Nevertheless, continued cosmological evolution of both FRII LERGs and FRI HERGs inevitably leads toward FRI LERG states, so the mixed states appear at intermediate redshifts as is observed (Hardcastle, Evans & Croston, 2006). In other words, the appearance of these transitional states between FRIIs and FRIs is observationally in agreement with our model in that mixed states are sandwiched between FRII HERGs and FRI LERGs. From a general perspective, cosmological evolution in this model produces FRI LERGs from FRII HERGs. As a direct consequence of this, LERGs have higher black hole masses than HERGs, as is observed (Smolcic et al. 2009; Smolcic 2009). The boundary between FRIIs and FRIs produces other interesting morphologies. As a consequence of the fact that BZ power drops to zero at zero black hole spin, there must be a transition period in which the jet engine turns off. The timescale of this transition depends on the accretion rate so our emphasis remains qualitative, but we could imagine the following scenario. Because the transition involves going from an FRII to an FRI state, a jet morphology might appear in which further away from the black hole engine is the presence of the relic, well-collimated FRII double-sided jet, whereas closer to the black hole appears the younger double-sided FRI jet, such as in 3C288 (Bridle et al. 1989; Lal et al. in preparation). Alternatively, the likelihood is that of a spin-flip without transition through zero spin. The spin value of that flip would produce an even larger break in the boundary between FRIIs and FRIs. This transition or gap in the efficiency of the engine near zero spin or due to a spin-flip is important in that it provides the possibility of a sharp break between states that have high and low BP, suggesting a natural location for a break between collimated and uncollimated or less collimated jets. We also emphasize the fact that our model is founded on the prescription of prolonged accretion as opposed to chaotic accretion scenarios (Volonteri, Sikora & Lasota 2007; Berti & Volonteri 2008; King, Pringle & Hoffman 2008) so that most of the mass is provided by an accretion disk with fixed orientation (Miller 2002), and thus, as pointed out in more detail in the discussion section, we expect black holes in all galaxies to evolve toward highly spinning prograde values as a result. The chaotic accretion scenario is usually invoked to produce low spins in radio-quiet spiral galaxies under the assumption that the spin paradigm operates; but, given that radio-quiet AGN are not low-spinning black hole accretion systems in the gap paradigm, no such mechanism need be invoked (as pointed out by Berti & Volonteri (2008), the spin of $0.99$ claimed by Brenneman & Reynolds for MCG-06-30-15, is very unlikely in the chaotic accretion scenario.) In other words, [*accretion in the gap paradigm simply spins all black holes up in the prograde direction in all galaxies, ellipticals as well as spirals*]{}. Accordingly, the observed distribution of powerful radio-loud galaxies reflects the ability of mergers to produce retrograde accretion flows in our model, which implies that mergers are constrained to produce a peak of such occurrences at $z\approx2$, a distribution that dies off at higher and lower $z$. \[cosmo\_evolution\] FRI/II Dichotomy ---------------- The FRI/FRII dichotomy between low radio-luminosity, poorly collimated radio jets and high radio-luminosity, highly collimated radio jets, has been explained according to two different physical scenarios. The first assumes that external environmental factors influence the jet structure (De Young 1993; Laing 1994; Bicknell 1995; Kaiser & Alexander 1997; Gopal-Krishna & Wiita 2000), while the second attributes the differences in morphology to parameters associated with the jet production mechanism itself (Rees 1982; Baum et al 1995; Reynolds et al 1996; Meier 1999, 2001; Ghisellini & Celotti 2001; Marchesini et al. 2004). Our model weighs in on this issue by naturally reproducing an FRI/FRII, Ledlow-Owen-like diagram as we illustrate below. Figure \[Ledlow\] shows the power in the jet as a function of spin state. According to our scenario, FRIIs are retrograde accretion systems and so are located on the left side of the break centered at zero spin. FRIs, instead, are modeled as prograde spin states so are located on the right-hand-side of the break. The time flow in our model runs from retrograde accretion toward prograde accretion. Therefore, the x-axis can be replaced by time. Such a jet power vs time diagram serves the purpose of illustrating the cosmological evolution of a single initial FRII HERG. In attempting to recover a Ledlow-Owen-like diagram from our model, we must consider a combination of the time evolution of multiple initial FRII HERGs. The simplest way to do this is to imagine a series of FRII HERGs forming at different times, producing a separation between otherwise similar paths. In other words, we have combined, side-by-side, paths like those of Fig. \[Ledlow\] and simply shifted them along the x-axis, producing multiple plots of translated but equal configurations. Since supermassive black hole mass increases in time due to accretion, from a qualitative perspective, similar information appears on both a time and black hole mass plot despite a relative stretching between such diagrams due to short lifetimes and smaller mass increase during the retrograde phase compared to the prograde one. In other words, one can produce a qualitatively valid diagram by exchanging time with black hole mass. Finally, via the black hole mass-galaxy luminosity relation, one can exchange black hole mass with optical galaxy luminosity on the x-axis. These steps constitute a qualitative, zeroth-order attempt, at illustrating the consequences of our model on the evolution of a $\it{family}$ of powerful radio-loud objects. The result is Figure \[Owen-Ledlow\]. From a qualitative perspective, then, the Ledlow-Owen diagram marks a division between objects that are young and powerful and others that are old and weak. Although that is the general trend, near the transition region of Figure \[Ledlow\], we can appreciate the existence of FRIIs that have weaker jets compared to FRIs further up beyond the break in the prograde regime. The presence of a large population of FRIIs (blue) in the lower part of Figure \[Owen-Ledlow\] as opposed to the actual Owen-Ledlow diagram is simplistic in two ways. First, the diagram fails to capture the fact that the lower power FRII engines are less collimated than their more powerful counterparts suggesting a less superficial classification in which “FRII-like” or “blue-like” objects populate the lower regions of Figure \[Owen-Ledlow\]. Second, we must also consider the effects of the external environment. As galaxy luminosity increases to the right on the x-axis, the galactic environment is more prone to turning these otherwise low power FRIIs into FRIs. Applying the same argument to the FRIIs (blue) that are mixed in with the FRIs (red) further up but to the right in the diagram of Figure \[Owen-Ledlow\] (i.e. where the objects are living in more dense stellar environments compared to their counterparts at equal jet power further back along the x-axis) suggests that environment might be responsible for turning some of them into FRIs. In other words, the further to the right we move on the diagram, the further up in the vertical direction we expect to replace FRIIs with FRIs. Because, to re-emphasize, our theoretical paths of Figure \[Owen-Ledlow\] are based solely on the properties of the jet engine, it is not surprising that it is precisely that which we are not modeling (i.e. environment) that is required to make theory compatible with observation. In other words, the Fanaroff-Riley dichotomy remains a function of jet power (governed by black hole spin) and environmental factors. We recognize that producing Figure \[Owen-Ledlow\] does not yield additional predictive power. However, we suggest that the simple compatibility at even a superficial level between the gap paradigm and the actual Owen-Ledlow diagram is statistically non-trivial. \[Fanaroff-Riley\] Fe K$\alpha$ Lines and Disk Winds --------------------------------- One interpretation of the narrowness of iron lines in HERGs compared to radio-quiet AGN is that the innermost stable circular orbit is further away from the black hole for higher retrograde accretion systems (we will suggest that true radio-quiet AGN are maximally spinning prograde systems). As far as reflection features in the inner disk regions go, even mild relativistic motion in coronal material above an accretion disk can reduce the reprocessed radiation from the disk (Beloborodov 1999). At velocities of just under $0.3c$, fluorescent line emission is effectively washed out (Reynolds & Fabian 1997). Therefore, the weakness of reflection features in HERGs may simply be the natural consequence of highly relativistic motion away from the disk in the inner regions where both BZ and BP conspire to produce relativistic jets while their narrowness indicates disk inner edges at larger radii. The largeness of the gap region that produces the powerful, collimated jets of the FRII HERGs is also associated with an absence of inner-disk regions that are close to the black hole where gravitational potential energy can be reprocessed in the disk. As mentioned above, the absence of this reservoir of gravitational energy is compatible with a decrease or absence of disk winds further out in the disk. FRI LERGs, on the other hand, also fail to produce disk winds but due to the fact that the flow is advective dominated and not thin-disk-like. FRI HERGs, according to our model, should possess disk winds because the flow is thin-disk-like and the gap region is smaller so the reservoir of reprocessable gravitational energy near the black hole for use further out in the disk is comparatively larger. We will argue, however, that as the spin increases in the prograde direction, the increase in disk winds resulting from smaller ISCO radii, inhibit jets, so that FRI HERGs fails to be appropriate nomenclature once jets are absent. As a result, FRI HERGs applies only to lower prograde HERGs. The fact that FRI LERGs are located on the right-hand side of Figure \[Ledlow\] implies that such objects should display a weak dependence on black hole spin since the power is roughly a flat function of spin for a range of intermediate spins. As pointed out (Garofalo 2009b), this alleviates the fine-tuning issue of Nemmen et al (2007) and Benson & Babul (2009). In addition, the retrograde or FRII region involves a much steeper function of black hole spin so the dependence on accretion rate should not be the same as it is for the FRIs. In other words, our model removes the need for even higher accretion rates in order to model the most powerful FRIIs, i.e. radio-loud quasars. It might be worth emphasizing that although jet power in the BZ effect does not come from the accretion power per se, the magnetic field which allows extraction of black hole rotational energy depends on the accretion rate in such a way that for fixed black hole spin, BZ power depends linearly on accretion rate in ADAFs. We also point out that retrograde accretion states are short-lived. From purely accretion considerations, a highly-spinning retrograde black hole will be spun down to zero spin when the black hole has accreted about 0.2 of its original mass (Moderski & Sikora 1996). The prograde evolution regime, on the other hand, is slower not only because the black hole must reach a mass of $\sim 2.5M$ where $M$ is the original black hole mass (Volonteri, Sikora & Lasota 2007), but also because the FRI regime is characterized by low angular momentum black hole-feeding in radiatively inefficient accretion. While FRIs have lifetimes reaching $\sim 10^{8} years$, FRII lifetimes are about a factor of ten smaller at a few times ($10^{6}-10^{7}$) years (O’Dea et al. 2009). \[Fe\_alpha\] Discussion and Conclusion ========================= We presented a framework for the cosmological evolution of radio-loud galaxies in which the gap region between accretion disks and black holes is key. In this gap paradigm, jet production is most effective when black hole magnetospheres conspire with their large gap regions to produce strong black hole-threading magnetic fields that are geometrically favorable to both the BZ and BP effects. As time evolution through accretion spins the black hole up in the prograde direction and the gap region size decreases, the interaction of rotating black holes with their magnetospheres becomes inefficient and jets weaken, become less collimated and if accretion spins the black hole up to high prograde values, likely fails entirely to produce jets. Maximally spinning prograde black hole magnetospheres that are the result of cosmological evolution via accretion, thus, may be systems that have now become inactive. The radio-quiet quasars formed in ellipticals may, instead, involve the population of post-merger systems characterized by high accretion onto spinning black holes in a prograde configuration. According to our scenario, the origin of the radio-loud/radio-quiet dichotomy may lie in the existence of two types of engine efficiencies. On the one hand, retrograde black holes conspiring with their magnetized flows are highly efficient in producing non-thermal, jet outflows, but shorter-lived configurations. These engines produce the observed radio-loud population. On the other hand, prograde, high accretion systems with small gap regions, while incapable of producing powerful jets, are capable of tapping into the gravitational potential energy of their inner accretion flow, making them highly efficient in producing strong thermal emission. The reprocessed gravitational potential energy of the inner regions near the rapidly spinning black hole is fundamental in producing the disk winds from larger disk radii. It has recently been shown that GRS 1915+105, a galactic microquasar, exhibits a soft state in which the radiation field drives a hot wind off the accretion disk carrying enough mass to halt the flow of matter into the radio jet (Neilsen & Lee 2009), a picture that is underscored by theoretical work (Meier 1996) and for which observation is mounting in other systems as well (Blum et al. 2010). This scenario suggests that most FRI HERGs are systems in which the disk winds eventually inhibit jet production making such objects part of the radio-quiet population, thereby explaining the rarity of FRI HERGs. The handful of observed FRI HERGs are then systems where the black hole spin is small, producing relatively larger gap regions, which decreases the production of winds further out from the disk and allows the presence of an FRI jet. In other words, in the cosmological evolution of radio-loud objects, FRI HERGs are objects that are transitioning away from strong jet production and weak winds, to a scenario involving weaker jets and stronger winds but that have not reached the high prograde spin states that would produce strongest winds and weakest or no jets. We highlight the compatibility of this scenario with the existence of FRI type radio quasars (Heywood et al 2007) as pointed out to us by David L. Meier. The fact that accretion states transition from HERGs to LERGs, however, allows such systems to remain radio-loud as they increase their prograde spins as a result of the failure of ADAFs to produce the strong jet-inhibiting disk winds. Thus, spinning black holes produce both FRII HERGs and radio-quiet quasars depending on retrograde vs prograde accretion. If mergers are responsible for both the high accretion rates and the occurrence of retrograde flows, it does not surprise that spiral galaxies are not competitive in their radio-loudness, and less accretion powered than the post-merger ellipticals. If retrograde accretion is formed in mergers, it also does not suprise that spirals are predominantly radio-quiet. In fact, only the low prograde spin configurations would lead to radio-loudness. Given that higher prograde accretion systems surrounded by radiatively efficient thin disks have smallest gap regions and the strongest jet-inhibiting winds, higher prograde spin systems in spirals should be radio-quiet; and, therefore, maximally spinning prograde configurations should be the most radio-quiet sources. In short, prograde vs retrograde, radiatively efficient accretion flows, produce the radio-quiet/radio-loud dichotomy. And finally, we conclude our discussion with a glance at the implications for microquasars. Because such objects are prograde accretion systems, their accretion state changes produce the microquasar counterpart to AGN HERGs and LERGs, or more appropriately, a transition between high prograde spin, high excitation accretion states (or radiatively efficient accretion states) to high prograde spin, low excitation accretion states (or radiatively inefficient accretion states). If the parallel with the FRII feedback phase in AGN is valid, these high prograde spin, high excitation accretion states in microquasars produce strong disk winds that expel the gas (Neilsen & Lee 2009), thereby producing an ADAF phase during which disk winds drop and jets are no longer hindered while the system enters the radio dominated hard state. Eventually, the accretion flow from the companion allows the system to re-enter the soft or radiatively efficient accretion state. The extension of our model to microquasars suggests that the approach to the high excitation accretion (soft) state from the low excitation accretion (hard) state, involves the brief combined presence of a BZ component and a BP-like component produced by the onset of inner disk winds. In other words, there is a short time during which the microquasar mimicks the conditions in FRII HERGs, thereby producing a powerful, collimated jet. Note that no spin change occurs. This suggests that microquasars undergo state transitions that take them from high prograde spinning high excitation accretion states (soft states) to high prograde spinning low excitation accretion states (hard states) and back, and that during the low excitation to high excitation transition, the physical conditions of an FRII HERG are briefly produced. We summarize our main conclusions as follows: 1. FRIIs have retrograde spin while FRIs have prograde spin 2. FRII HERGs evolve toward FRI LERGs 3. FRIIs have larger jet efficiencies while FRIs have lower jet efficiencies 4. High prograde spin systems in HERG states have large disk efficiencies (and thus strong disk winds) while LERG states and retrograde accretion states have lower disk efficiencies (and thus lower disk winds) 5. Highly efficient jet engines are highly inefficient disk engines and viceversa 6. FRI HERGs have low prograde spin 7. Radio quiet AGN are high spinning prograde, radiatively efficient systems 8. Maximally spinning prograde, radiatively efficient systems, are the most radio quiet AGN 9. Spiral galaxies involve prograde accretion 10. Spiral galaxies should be more radio loud when the spin is lower and more radio quiet, quasar-like, when the spin is high 11. A radiatively inefficient to radiatively efficient transition, in high prograde systems, produces a short-lived, powerful, collimated jet as observed in microquasars. We conclude by highlighting the fundamental role or impact of general relativity in galaxy evolution in this model. The validity of our paradigm suggests that aspects of the spacetime metric that are assumed to govern physics in regions provincially relegated to the near black hole, are in fact dominant on much larger scales, to the extent that galaxy morphology, energetics, and evolution, are tightly linked to the details of strong-field general relativistic effects. Whereas a Newtonian treatment of spacetime is successful just a few gravitational radii from the black hole in the sense that relativistic corrections are negligible there, the scale of influence of black hole spin that our paradigm forces us to grapple with is daunting, with up to more than eight orders of magnitude beyond its local sphere of influence. It is the details of tiny regions of highly curved spacetime that have the greatest effect on the large-scale properties of galaxies. \[Conclusions\] acknowledgments =============== D.G. thanks David L. Meier for discussion on the FRI/FRII division, the physics of jet launching, propagation and collimation, and for commenting on the draft as a whole, Peter Polko for stimulating ideas on the effects of BP jets on accretion disk coronae, Christopher S. Reynolds for issues concerning broad iron lines and jet energetics and Cole Miller and Marta Volonteri for discussion on retrograde accretion. We acknowledge the role of and thank the three referees and three editors that were involved in this project. The research described in this paper was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. D.G. is supported by the NASA Postdoctoral Program at NASA JPL administered by Oak Ridge Associated Universities through contract with NASA. References {#references .unnumbered} ========== Allen S.W. et al, MNRAS, 2006, 372, 21 Antonuccio-Delogu V. & Silk J., 2010, MNRAS, submitted Arshakian, T., Beck, R., Krause, M., Sokoloff, D., & Stepanov, R., Proceedings to Panoramic Radio Astronomy, Groningen, 2009 Barnes J.E.& Hernquist, L. 1991, ApJ, 370, L65 Baum S.S., Zirbel E.L. & O’Dea C.P., 1995, ApJ, 451, 88 Beloborodov, A.M., 1999, ApJ, 510, L123 Benson A.J. & Babul A., 2009, MNRAS, 397, 1302 Bicknell G.V., 1995, ApJS, 101, 29 Blandford, R. D., & Payne, D. G. 1982, MNRAS, 199, 883 Blandford, R. D., & Znajek, R. L. 1977, MNRAS, 179, 433 Blandford, R.D., in Active Galactic Nuclei, ed. T.J.-L. Courvoisier & M. Mayor, 161-275 Blum J.L. et al, 2010, ApJ, in press Bogovalov, S. & Tsinganos, K. 2005, MNRAS, 357, 918 Brenneman, L., & Reynolds, C.S., 2006, ApJ, 652, 1028 Bridle A.H. et al, 1989, AJ, 97, 674 Cao X., 2003, ApJ, 599, 147 Daly, R.A., ApJ, 2009, 691, L72 Daly, R.A., ApJ, 2009, 696, L32 De Young D.S., 1993, ApJ, 405, L13 Donato D., Sambruna, R.M., Gliozzi, M, 2004, ApJ, 617, 915 Evans, D., 2004, ApJ, 612, 786 Evans, D., et al, 2006, ApJ, 642, 96 Fabian, A.C. et al, Nature, 2009, 459, 540 Fanaroff, B.L.& Riley, J.M., 1974, MNRAS, 167, 31P Ferrarese, L., & Merritt, D., 2000, ApJ, 539, L9 Garofalo, D., 2009, ApJ, 699, L52 (a) Garofalo, D., 2009, ApJ, 699, 400 (b) Gebhardt, K., et al. 2000, ApJ, 539, L13 Ghisellini G., Celotti A., 2001, A&A, 379, L1 Gopal-Krishna & Wiita P.J, 2000, A&A, 363, 507 Grandi P., Urry C.M., Maraschi L., 2002, NewAR, 46, 221 Kormendy, J., & Richstone, D., 1995, ARA&A, 33, 581 Hardee P.E & Hughes P.A., 2003, ApJ, 583, 116 Hardcastle M.J., Evans D.A. & Croston J.H., 2006, MNRAS, 370, 1893 Hardcastle M.J., Evans D.A. & Croston J.H., 2007, MNRAS, 376, 1849 Heywood I., Blundell K.M.& Rawlings S., 2007, MNRAS, 381, 1093 Hughes, S.A. & Blandford, R.D., 2003, ApJ, 585, L101 Kaiser C.R. & Alexander P., MNRAS, 1997, 286, 215 King A.R., Lubow S.H., Ogilvie G.I. & Pringle J.E., 2005, MNRAS, 363, 49 King, A.R., Pringle,J.E. & Hofmann, J.A., 2008, MNRAS, 385, 1621 Kormendy, J., & Richstone, D., 1995, ARA&A, 33, 581 Kuncic, Z. & Bicknell, G.V., 2004, ApJ, 616, 669 Kuncic, Z. & Bicknell, G.V., 2007, Ap&SS,311, 127 Laing R.A., 1994, ASPC, 54, 227 Lal et al, in preparation Livio, M., Ogilvie, G.I. & Pringle, J.E., 1999, ApJ, 512, 100 Magorrian, J., et al. 1998, AJ, 115, 2285 Mangalam A., Gopal-Krishna & Wiita P.J., 2009, MNRAS, 397, 2216 Marchesini D., Celotti A. & Ferrarese L., 2004, MNRAS, 351, 733 Marconi, A., & Hunt, L.K., 2003, ApJ, 589, L21 McKinney J.C. & Narayan R. 2007, MNRAS, 375, 513 McKinney J.C. & Blandford R.D., 2009, MNRAS, 394, L126 McKinney J.C. & Gammie C.F., 2004, ApJ, 611, 977 McNamara B.R., et al,2009, ApJ, 698, 594 Menon, G. & Dermer, C.D., 2005, ApJ, 635, 1197 Meier D. L., 1999, ApJ, 522, 753 Meier D.L. et al., Science 291 (2001), 84 Meier D.L., 1996, ApJ, 459, 185 Miller, M.C., 2002, ApJ, 581, 438 Moderski R.& Sikora M., 1996, A&AS, 120, C591 Moderski R., Sikora M.& Lasota J.-P., MNRAS, 1998, 301, 142 Nakamura, M. & Meier, D.L., 2004, ApJ, 617, 123 McClintock, J.E., Shafee, R., Narayan, R., Remillard, R.A., Davis, S.W.& Xin-Li, L., ApJ, 652, 518 Narayan, R. & Yi I., 1995, ApJ, 452, 710 Neilsen J. & Lee J., Nature, 458, 481 Neemen R.S. et al, 2007, MNRAS, 377, 1652 Nesvadba N.P.H, Lehnert M.D., De Breuck C., Gilbert A. & van Breugel W., 2008, A&A, 491, 407 Novikov I. & Thorne K., 1973, Black Holes (Les astres occlus), p. 343 O’Dea et al, 2009, A&A, 494, 471 Ogle P., Whysong D. & Antonucci R., 2006, ApJ, 647, 161 Rees M.J., 1982, IAUS, 97, 211 Reeves J.N.& Turner M.J.L., 2000, MNRAS, 316, 234 Reeves J.N., Sambruna R.M., Braito V. & Eracleous M., 2009, ApJ, 702, L187 Reynolds C.S. et al, 1996, MNRAS, 283, L111 Reynolds C.S. & Fabian A.C. 1997, MNRAS, 290, L1 Reynolds C.S., Garofalo D.& Begelman M.C., 2006, ApJ, 651, 1023 Sambruna et al, 2009, ApJ, 700, 1473 Smolcic V., 2009, ApJ, 699, L43 Smolcic V. et al, 2009, ApJ, 696, 24 Tchekhovskoy A., McKinney J.C. & Narayan R. 2009, ApJ, 699, 1789 Tchekhovskoy A., Narayan R & McKinney J.C. 2010, ApJ, in press Thorne K.S., Price R.H. & Macdonald D.A., 1986, Black Holes: The Membrane Paradigm (New Haven: Yale Univ. Press) Tremaine S., et al 2002, ApJ, 574, 740 Volonteri M., Sikora M., & Lasota J-P, 2007, ApJ, 667, 704 Wilms J. et al, 2001, MNRAS, 328, L27 Wilson A.S. & Colbert E.J.M, 1995, ApJ, 438, 62 Zoghbi A. et al, 2010, MNRAS, 401, 2419

ArXiv

--- abstract: 'The slot filling task aims at extracting answers for queries about entities from text, such as “Who founded Apple”. In this paper, we focus on the relation classification component of a slot filling system. We propose type-aware convolutional neural networks to benefit from the mutual dependencies between entity and relation classification. In particular, we explore different ways of integrating the named entity types of the relation arguments into a neural network for relation classification, including a joint training and a structured prediction approach. To the best of our knowledge, this is the first study on type-aware neural networks for slot filling. The type-aware models lead to the best results of our slot filling pipeline. Joint training performs comparable to structured prediction. To understand the impact of the different components of the slot filling pipeline, we perform a recall analysis, a manual error analysis and several ablation studies. Such analyses are of particular importance to other slot filling researchers since the official slot filling evaluations only assess pipeline outputs. The analyses show that especially coreference resolution and our convolutional neural networks have a large positive impact on the final performance of the slot filling pipeline. The presented models, the source code of our system as well as our coreference resource is publicy available.' author: - | Heike Adel [email protected]\ Hinrich Schütze [email protected]\ Center for Information and Language Processing (CIS)\ LMU Munich, Germany bibliography: - 'refs.bib' title: 'Type-aware Convolutional Neural Networks for Slot Filling' --- Introduction {#sec:Introduction} ============ Knowledge bases provide structured information about entities and concepts of the world. They are important resources for artificial intelligence (AI) and natural language processing (NLP) tasks, such as entity disambiguation, question answering or information retrieval . Given a knowledge base, answering a question like “Who founded Apple?” would require only a simple lookup. Similarly, an automatic assistant or dialogue system could satisfy the needs of users more easily with the access to a background knowledge base. If a user is, for example, looking for popular sights nearby, an automatic assistant could look up points of interest and information about them in a knowledge base. Popular large-scale knowledge bases, such as Freebase or Wikipedia (@wikiurl-URL) are often created in a large collaborative effort. Despite a lot of (manual) effort spent on their creation and maintenance, they are usually incomplete. Missing facts, however, limit their applicability in down-stream tasks. At the same time, there is a lot of unstructured text data available – e.g., on the internet – that mentions information missing in knowledge bases. Therefore, automatic methods for extracting structured information from text data to populate knowledge bases are important. One specific incarnation of knowledge base population (KBP) is slot filling [@overviewSF2013; @overviewSF2014], a shared task (@sfurl-URL) which is yearly organized by the Text Analysis Conference (TAC). Given a large document collection and a query like “X founded Apple”, the task is to extract “fillers” for the slot “X” from the document collection. The extraction of answers to the queries from large amounts of natural language text involves a variety of challenges, such as document retrieval, entity identification, coreference resolution or cross-document inferences. To cope with those challenges, most slot filling systems are pipelines of different NLP components. One of the most important components validates whether a candidate (e.g., “Steve Jobs”) is a correct filler of the slot (e.g., “X” in “X founded Apple”). We take a *relation classification* approach to candidate validation in this paper. For example, the relation classifier validates whether the relation between the noun phrases “Steve Jobs” and “Apple” in the sentence “Steve Jobs started Apple” is the relation “X founded Y”. Traditional methods to slot filling relation classification rely on (hand-crafted) patterns or linear classifiers with manually designed features. Given the variability of language, it is desirable to learn relation-specific characteristics automatically from data instead. Therefore, we design convolutional neural network architectures for the special characteristics of the slot filling task (e.g., long sentences, many inverse relations) which learn to recognize relation-specific n-gram patterns. Contributions {#sec:contributions} ------------- We now describe our contributions in this paper which are centered around *convolutional neural network* (CNN) architectures for *relation classification* in the context of *slot filling*. ### Architectures and Extensive Experimentation Using Convolutional Neural Networks for Slot Filling We were one of the first groups to use CNNs for relation classification and demonstrate their effectiveness for slot filling [@adelNaacl2016; @vuNaacl2016]. The system based on this work (which is described in detail in Section \[sec:overview\]) is state of the art for distantly supervised slot filling [@cis2015], see Section \[sec:soaSF\]. In contrast to scenarios where carefully labeled gold training sets are available, relation classifiers in slot filling are trained on data that is noisy – due to error propagation through the pipeline and due to distant supervision [@cis2015; @adelNaacl2016]. We show that CNNs are robust enough to be successfully applied in this noisy environment if the generic CNN architecture is adapted for relation classification and if hyperparameters are carefully tuned on a per-relation basis (see Section \[sec:neuralNets\]). We also show that multi-class CNNs perform better than per-relation binary CNNs in the slot filling pipeline (Section \[sec:end-to-end\]) probably because imposing a 1-out-of-k constraint models the data better – even though there are rare cases where more than one relation holds true. ### Type-Aware Relation Classification We define relation classification as the problem of assigning one of several relations to a 5-tuple $(c_{-1},e_1,c_0,e_2,c_{+1})$ consisting of two entites $(e_1$, $e_2)$ and preceding, central and following contexts $(c_{-1}$, $c_0$, $c_{+1})$. If entities are represented as they occur in the raw text, the classifier is likely to overfit to the idiosyncrasies of the entities mentioned in the training data. On the other hand, removing all entity information from the input is also harmful since entities provide valuable information for disambiguating relations; consider “Apple launches iPhone” (to start selling) vs. “SpaceX launches Falcon 9” (to send into orbit). A major focus of this work is *type-aware relation classification*, a middle ground between complete entity information and no entity information: only the predicted types of the entities are made available. Type information is arguably the key information needed for disambiguation in relation classification – e.g., this is the case for our “launch” example); and it prevents overfitting to entity idiosyncrasies. We design three type-aware architectures: a simple pipeline of first type classification and then relation classification (see Section \[sec:pipeline\]); a joint model (Section \[sec:jointTraining\]); and a structured prediction model that more directly takes into account the dependencies between entity and relation classes (Section \[sec:structuredPrediction\]). In our experiments in Section \[sec:end-to-end\], we show that the structured prediction model outperforms the other models in terms of macro $F_1$, the best measure of performance for difficult cases because it gives equal weight to rare and frequent relations. ### Analysis The TAC KBP organizers only evaluate the final results of the entire slot filling pipeline. We perform an extensive and detailed analysis (Sections \[sec:recallAnalysis\] and \[sec:errorAnalysis\]) and several ablation studies (Section \[sec:ablation\]) on individual modules of the pipeline. Inter alia, we quantify the impact of entity linking, coreference resolution and type-aware CNNs on the overall pipeline performance. We hope that this will be of great benefit to the community because this analysis – in contrast to the official TAC KBP evaluation – allows researchers to assess the impact of individual components and which components are worth investing more research effort in. ### Resources We make our complete slot filling system, including the source code, publicly available at <http://cistern.cis.lmu.de/CIS_SlotFilling>. Since slot filling poses many NLP challenges, building such a system is a substantial software development and research effort. Through publication of the system, we share our experience with the community and lower the barriers to entry for researchers wishing to work on slot filling. The component of our pipeline that has the longest runtime – a runtime of several months on the entire TAC source corpus – is coreference resolution. Therefore, we publish the output of the coreference resolver for the two million documents of the TAC source corpus at <http://cistern.cis.lmu.de/corefresources>. It consists of 198 million mentions linked in 54 million coreference chains. This will make it easier for other researchers to take advantage of coreference resolution in their systems. By making these resources available to the NLP community, we aim to promote research in knowledge base population in general and slot filling in particular. Relation to Our Prior Publications ---------------------------------- In this subsection, we delineate our contributions in this paper from our prior publications. In earlier work [@cis2015; @adelNaacl2016], we compared binary convolutional neural networks to traditional models for slot filling (patterns and support vector machines). Binary models facilitate extensions of the slot list with a few more slots since new models can be trained for the new slots but the existing models do not need to be retrained. However, the more slots there are, the more models need to be optimized, maintained and evaluated. Therefore, we explore multi-class convolutional neural networks for slot filling in this paper. In our experiments, we compare the novel multi-class models to the previously trained binary models. The contribution of type-aware models which lead to our best results on the official slot filling evaluation data, is entirely novel to this paper. Although entity types are well-studied features for traditional slot filling models , this is the first work to explore end-to-end type-aware neural networks for slot filling and show their positive impact in the pipeline setting. Our three approaches for type-aware neural networks build on models we proposed earlier [@EACLnoiseMitigation; @adelEmnlp2017]. In contrast to these two prior studies, we adapt the architectures of the models to the requirements of the slot filling task, e.g., making them more robust against unknown rare test entities and against the existence of inverse relations. This is also the first work to evaluate the CNNs with structured prediction in a noisy scenario which is arguably conceptually different to both clean data with manual annotations and distantly supervised data used without pipelines. For structured prediction, we formalize the task of joint entity and relation classification as a triple of predictions (similar to a knowledge base triple) which enables the model to learn which entity and relation classes often co-occur together. This is a novelty to slot filling which has been approached mainly with pattern matching or classification so far. Structure --------- The remainder of the paper is organized as follows. First, the slot filling task and its challenges are described (Section \[sec:task\]). Section \[sec:overview\] presents our slot filling system which we used in the official shared task competition in 2015. In Section \[sec:neuralNets\], we describe our convolutional neural network for slot filling relation classification and introduce multi-class models as well as models for the joint task of entity and relation classification. Afterwards, we present our experiments and discuss our results in Section \[sec:results\]. Section \[sec:analysis\] provides the results of a recall analysis, a manual categorization of the errors of our system and several ablation studies. Section \[sec:relWork\] presents related work. Finally, Section \[sec:conclusion\] concludes the paper. Slot Filling ============ In this section, we describe the slot filling task and its challenges and present the most important aspects of our slot filling system. Task and Challenges {#sec:task} ------------------- The TAC KBP slot filling task addresses the challenge of gathering information about entities (persons, organizations or geo-political entities) from a large amount of unstructured text data [@overviewSF2013; @overviewSF2014]. The input is a query similar to the following one (fictional query with a random query id and document id): <query id="CSSF15_ENG_012abc3456"> <name>Apple</name> <docid>NYT_ENG_20131203.4567</docid> <beg>222</beg> <end>226</end> <enttype>org</enttype> <slot0>org:founded_by</slot0> <slot1>per:date_of_birth</slot1> </query> This sample query asks for the founders of the company Apple as well as for their dates of birth. It consists of a unique identifier (`query id`), the name of an entity (`name`), which we will call query entity in the following, and the type of this entity (`enttype`) which can be either person, organization or geo-political entity. Furthermore, it contains the slots to be filled (`slot0, slot1`), i.e., the questions that should be answered for the query entity, as well as a starting point in the document collection (`docid` with begin (`beg`) and end offset (`end`)) which points to a mention of the query entity. The starting point usually does not provide the answer to the query but it can be used to disambiguate different entities with the same name. ![Overview of the slot filling task with a multiple-hop query.[]{data-label="fig:task"}](overview_task.pdf){width=".8\textwidth"} Figure \[fig:task\] illustrates the slot filling task which is described in the following. The system has access to a large document collection which needs to be processed in order to answer the query – in the example, to find the founders of Apple and their dates of birth. The query can be split into two parts: First, the question posed by `slot0` (“Who are the founders of Apple?”) needs to be answered. Based on the results of the system on this slot, `slot1` is processed (“What is their date of birth?”). Since this corresponds to taking a hop in the corresponding knowledge graph (from Apple to its founders to their date of birth), a query with two slots is also called multiple-hop query, and the question posed by `slot0` is called hop 0 and the question posed by `slot1` is called hop 1. The shared task organizers provide a script to generate new sub-queries for hop 1 given the system results for hop 0 and the original set of queries. The output of the system should contain the answer for each given slot (e.g., “Steve Jobs” and “Steve Wozniak” for `org:founded_by`), a supporting sentence from the document collection (e.g., “Founders Steve Jobs and Steve Wozniak created Apple Computers on April 1, 1976”) as well as a confidence score. The slots can be single-valued (for instance, `per:date_of_birth`: a person has only one date of birth) or list-valued (for instance, `org:founded_by`: a company might have more than one founder). In total, there are 65 slots, out of which 18 are single-valued and 47 are list-valued. The answer of the slot filling system is assessed as correct if both the slot filler and the supporting sentence are correct. In the official evaluations, human annotators assess system outputs manually. Based on these assessments, individual results for both hops are reported as well as an overall result. Previous work on slot filling showed that this task includes a variety of NLP challenges [@analysisRecall; @analysis2012; @overviewSF2014], such as alternate names for the same entity, ambiguous names (i.e., the same name for different entities), misspellings, coreference resolution, location inference, cross-document inference and relation extraction / classification. Our slot filling system addresses most of these challenges (except for cross-document inference which we only consider in the context of location inference). The CIS Slot Filling System {#sec:overview} --------------------------- For filling slots for persons, organizations and geo-political entities, i.e., for answering the questions posed by the input queries, a variety of natural language processing steps need to be performed. Our system addresses the slot filling task in a modular way. This has several advantages, including extensibility, component-wise analyzability (see Section \[sec:errorAnalysis\]) and modular development. In this section, an overview of the different components of our system is given. They are also depicted in Figure \[fig:systemoverview\]. More details can be found in our shared task system description paper [@cis2015]. ![System overview: Basic components of the CIS slot filling system.[]{data-label="fig:systemoverview"}](systemoverview_new.pdf){width="70.00000%"} ### Alias component {#component:alias} The alias component expands the query with possible aliases for the entity name. For this purpose, we employ a preprocessed list of possible aliases based on Wikipedia redirects which we extracted using <span style="font-variant:small-caps;">Jwpl</span> [@JWPL] on a Wikipedia dump from July 2014. If the query entity is an organization, we also add various company-specific suffixes to the list of aliases, such as “Corp”, “Co”, “Inc”. If the query entity is a person, we include nicknames taken from the web (@nicknameurl-URL) into the list of aliases. ### Information retrieval component {#component:IR} Based on the entity name (and aliases), documents mentioning this name are retrieved to reduce the large search space. For this, we apply the open-source system <span style="font-variant:small-caps;">Terrier</span> with the following set of queries: - AND combination of the tokens of the entity name as given in the input query - AND combination of the tokens of an alias - OR combination of the tokens of the entity name as given in the input query For geo-political entities, we only use the two AND queries. In prior experiments, we also investigated phrase queries but found that they did not work well with spelling variations, resulting in a considerably lower overall recall of the system. Instead, we filter the resulting list of relevant documents by fuzzy string matching with the name and aliases to skip documents mentioning both the first and the last name of a person but not in a phrase. For each entity, the results of the subqueries are ordered according to the relevance score assigned by <span style="font-variant:small-caps;">Terrier</span>, concatenated and limited to the top 300 documents. ### Entity linking component {#component:entityLinking} For disambiguating entities with the same name, we apply the entity linking system <span style="font-variant:small-caps;">Wat</span> [@WAT]. It takes a sentence as input and determines the Wikipedia id of every entity in that sentence. In order to get the Wikipedia id of the query entity, we apply it to the sentence specified by the starting point of the query. Afterwards, we check for each document returned by the information retrieval component whether the mention in the document which matches the name of the query entity refers to the same Wikipedia entity as the query, i.e., whether <span style="font-variant:small-caps;">Wat</span> assigns it the same Wikipedia id as the query entity. In the case of a mismatch, the document is ignored in the following steps. After entity linking, we limit the set of documents to the 100 documents with the highest relevance score according to <span style="font-variant:small-caps;">Terrier</span>. This number has been determined empirically in prior experiments: On data from previous slot filling evaluations (2013 and 2014), we observed that 100 documents are a good trade-off between recall and processing time. ### Candidate extraction component {#component:candidateExtraction} From the remaining documents, possible slot fillers (filler candidates) are extracted. #### Genre-Specific Document Processing. First, the documents are split into sentences (with Stanford <span style="font-variant:small-caps;">CoreNlp</span>) and cleaned, i.a., by removing html tags. Since the TAC 2015 evaluation corpus consists of two genres (news and discussion forums), our document processing and cleaning steps are genre-dependent. Prior analysis showed that this is crucial to reduce the noise in the input to the following pipeline components. #### Coreference Resolution. Second, fuzzy string matching (based on Levenshtein distance) and automatic coreference resolution (with Stanford <span style="font-variant:small-caps;">CoreNlp</span>) is performed in order to retrieve sentences mentioning the query entity. Different studies show the importance of coreference resolution for slot filling [@analysis2012; @analysisRecall; @overviewSF2014]. While most systems apply coreference resolution only for matching the query entity, we also use it for the filler candidates if the named entity type of the filler is `PERSON`. This improves the recall of the system considerably (e.g., consider the slot `org:students` and the sentence “He went to University of Munich”). In Section \[sec:ablationCoref\], we show the positive impact of coreference resolution on the slot filling pipeline results. Given only raw text data, the runtime of the slot filling system is mainly determined by coreference resolution. Therefore, we preprocessed the TAC source corpus (which consists of over two million documents) and stored the coreference information. In total, we have extracted about 54M coreference chains with a total number of about 198M mentions. The processing of all documents of the source corpus takes a considerable amount of time and may be infeasible in case of restricted computational resources. Therefore, we make this resource publicly available to the community (see Section \[sec:contributions\]). #### Filler Candidate Extraction. Given sentences with mentions of the query entity, the system extracts possible filler candidates based on named entity tags. For named entity tagging, we apply <span style="font-variant:small-caps;">CoreNlp</span>. For example, for the slot `per:date_of_birth`, the system would only consider dates as filler candidates while for a slot like `org:members`, the fillers can be organizations, locations or persons. For slots with string fillers, such as `per:title` or `per:charges`, we have automatically assembled lists of possible filler values from Freebase [@Freebase]. The lists have been cleaned manually in order to improve their precision. ### Slot Filler Classification Component {#component:classificationComponent} The classification component identifies valid fillers for the given slot based on the textual context of the extracted filler candidates. This is mainly a relation classification task with the additional challenges that no designated training data is available and that the classifier inputs are the results from previous pipeline steps and can, thus, be noisy (e.g., due to wrong coreference resolution, wrong named entity typing or erroneous sentence splitting). Given our previous results [@adelNaacl2016], we combine traditional models with convolutional neural networks (CNNs) in this component. The weights for combining the model scores are tuned on previous slot filling evaluation data. Section \[sec:neuralNets\] describes the CNNs in more detail. ### Postprocessing component {#component:postprocessing} As a last step, the filler candidates are postprocessed and output along with the confidence scores from the classification component and the contexts they appear in. #### Output Thresholds and Ranking. Based on the classification scores, filler candidates are selected for output or discarded. This decision is done based on slot-specific thresholds which were tuned on previous evaluation data. For the second slot (hop 1) of multiple-hop queries, we increase the thresholds by 0.1 in order to reduce false positive answers. (An answer to a hop 1 sub-query is only scored as correct if both the hop 0 answer and the hop 1 answer are correct. Thus, errors are propagated from hop 0 to hop 1.) The selected filler candidates are ranked according to their classification score. For single-valued slots, only the top filler candidate is output. For list-valued slots, the top $N$ filler candidates are output. ($N$ is slot-dependent and has been determined heuristically on previous evaluation data in order to increase the precision of the system.) #### Location Disambiguation and Inference. As we will describe below in Section \[sec:neuralNets-remarks\], we do not distinguish between cities, states-or-provinces, and countries in the classification component. Before outputting the results, however, the extracted locations need to be disambiguated. The system uses city-, state- and country lists from Freebase, Wikipedia (@statelisturl-URL) and an online list of countries (@countrylisturl-URL) to decide to which category a location belongs. If the system extracted a city or state while the slot given in the query is a state or country, the system automatically infers the answer for the desired slot based on city-to-state, city-to-country and state-to-country mappings extracted from Freebase. Convolutional Neural Networks for Slot Filling Relation Classification {#sec:neuralNets} ====================================================================== Convolutional neural networks (CNNs) have been applied successfully to natural language processing in general (; ) and relation classification in particular (; ). We integrate them into a slot filling pipeline. This poses two additional challenges to them which prior work usually does not consider in combination: noisy labels at training time due to distantly supervised training data; and noisy or wrong inputs at test time due to error propagations in the slot filling pipeline. Examples for the latter are wrong sentence boundaries (resulting in incomplete or very long inputs), wrong coreference resolution or wrong named entity tags (resulting in incorrect candidate entites for relation classification). Our results show that CNNs are still able to classify the relations and improve the final performance of the system. There are three reasons why CNNs are promising models for slot filling relation classification: (i) Convolutional filters of length $n$ automatically create features for every possible $n$-gram in the sentence. Since relation-indicative phrases are often $n$-grams (examples: “was born in” or “subsidiary of”), each convolutional filter can learn to recognize a particular $n$-gram and assign a high score to it. (ii) Max pooling, i.e., only considering the highest activations from each filter application result, helps extracting the most relevant $n$-grams independent of their position in the sentence. Thus, the following network layers can focus on those most relevant parts of the sentence only. (iii) The representation of input words with word embeddings and the internal computation of phrase and sentence representations based on them enables the network to recognize words or phrases which are similar to the ones seen during training. Thus, if a convolutional filter has learned to assign high scores to the $n$-gram “was founded by” during training, it can also recognize a phrase like “was established by” during testing even if this phrase did not occur in the training data (assuming that the embeddings of “founded” and “established” are similar). This is an advantage of neural models compared to pattern matching or bag-of-word approaches for which this generalization is more difficult. General Remarks {#sec:neuralNets-remarks} --------------- For training and evaluating the convolutional neural networks, we replace the query entity with the tag `<name>` and the candidate filler with the tag `<filler>`. This prevents the model from remembering entities from the training data and helps it to focus on the context words instead. Furthermore, we only train one model for each slot and its inverse, for example `per:parents` and `per:children`. For this, we transform all training examples of `per:parents` into training examples for `per:children` by reversing the `<name>` and `<filler>` tags in the sentences. This avoids redundant training and helps the model to discriminate between the two inverse slots. To extract the probability for an inverse slot during test time (e.g., for `per:parents`), we again reverse the `<name>` and `<filler>` tags of the sentence and extract the probability for the corresponding slot (e.g., `per:children`). We also merge the “city”, “country” and “state-or-province” slots to one “location” slot since we expect their fillers to appear in very similar contexts. Basic Architecture ------------------ This section describes the basic architecture of our convolutional neural network for slot filling relation classification (without entity types). ### Model Input {#sec:neuralNets-input} We only use features which are directly available from the input context, i.e., words and combinations of words, but no hand-crafted features, such as part-of-speech tags or dependency paths. One reason for that is to avoid potential noise in the inputs due to wrong tags or wrong dependency paths. As described before, the slot filling pipeline introduces different kinds of noise anyway. Therefore, we aim for limiting additional noise as much as possible. Another reason is previous results: In earlier work [@adelNaacl2016], we showed that models without dependency paths as features are able to outperform models using dependency paths (<span style="font-variant:small-caps;">Mintz++</span> and <span style="font-variant:small-caps;">MimlRe</span>) on slot filling relation classification. Similarly, achieved the best results in the slot filling evaluations with only a minimalistic feature set. We split the input sentence into three parts: left of the candidate relation arguments (left context), between the arguments (middle context) and right of them (right context). These three contexts form the input to the CNN, together with a flag indicating whether the query entity or the candidate filler appears first in the sentence. This flag is important for disambiguating inverse relations (see Section \[sec:neuralNets-remarks\]). ### Convolutional Layer The words of the input sentence are represented with word embeddings pre-trained with word2vec on English Wikipedia. As described in Section \[sec:neuralNets-input\], the input sentence is split into three parts: left, middle and right context. The network applies the following Equations \[eq:cnn1\]-\[eq:cnn6\] for convolution and 3-max pooling [@kalchbrenner] to each of these three parts individually but with convolutional filters $H$ and bias terms $b$ shared across the three contexts. $$\begin{aligned} C_{\text{left}} & = \tanh(I_{\text{left}} \ast H + b) \label{eq:cnn1}\\ C_{\text{middle}} & = \tanh(I_{\text{middle}} \ast H + b) \label{eq:cnn2}\\ C_{\text{right}} & = \tanh(I_{\text{right}} \ast H + b) \label{eq:cnn3}\end{aligned}$$ The symbol $\ast$ denotes convolution, $I$ is an input context, $H$ is the filter matrix and $b$ the bias term. We apply Equations \[eq:cnn1\]-\[eq:cnn3\] with multiple filter matrices $H$. The number of filter matrices $m \in \left\{100, 300, 1000\right\}$ is tuned on the development set. After convolution, 3-max pooling is applied which extracts the three maximum values of each $C$ (in the same order as they appeared in the input sequence), yielding the pooling results $P_{\text{left}}, P_{\text{middle}}, P_{\text{right}}$: $$\begin{aligned} P_{\text{left}}[i] & = [C_{\text{left}}[i,t] | R_t(C_{\text{left}}[i,t]) \le 3] \label{eq:cnn4}\\ P_{\text{middle}}[i] & = [C_{\text{middle}}[i,t] | R_t(C_{\text{middle}}[i,t]) \le 3] \label{eq:cnn5}\\ P_{\text{right}}[i] & = [C_{\text{right}}[i,t] | R_t(C_{\text{right}}[i,t]) \le 3] \label{eq:cnn6}\end{aligned}$$ with $[i]$ denoting the $i$-th row and $[i,t]$ the cell in the $i$-th row and $t$-th column of a matrix. The sequence $[p | P(p)]$ contains all elements $p$ that satisfy predicate $P$, in this case all elements whose rank $R_t$ along the time axis $t$ is 1, 2 or 3. For more details on k-max pooling, see . Because of convolution and pooling, the network is able to recognize relevant n-grams independent of their position in the input sentence. Afterwards, the results are concatenated to one large vector and extended with a flag $v$ indicating whether the entity or the filler candidate appears first in the sentence. The final vector is passed to a multi-layer perceptron with one hidden layer. $$s = \tanh(W_1^\top P_{\text{left}} + W_2^\top P_{\text{middle}} + W_3^\top P_{\text{right}} + W_4^\top v + d) \label{eq:hidden}$$ The matrices $W_1$, $W_2$, $W_3$ and $W_4$ are the weights of the hidden layer, $d$ is its bias term. ### Output Layer Finally, a softmax layer is applied to the sentence representation $s$. In earlier work [@cis2015; @adelNaacl2016], we only trained binary models which output 1 if $s$ expresses the given slot or 0 if it does not. In this paper, we explore multi-class models. In the case of multi-class models, the output vector contains one output entry for each slot (except for inverse slots, see Section \[sec:neuralNets-remarks\]). Figure \[cnnFig\] depicts the structure of the CNN. ![Convolutional neural network for slot filling.[]{data-label="cnnFig"}](CNN3.pdf){width="60.00000%"} Type-aware Convolutional Neural Networks {#sec:types} ---------------------------------------- The relation arguments of the input for the binary models all have named entity types corresponding to the expected types of the slots. The model for the relation `per:date_of_birth`, for example, only gets sentences with one relation argument being a `PERSON` and the other relation argument being a `DATE`. In our training data, this is ensured by the design of the extraction process of positive and negative examples (see Section \[sec:trainingdata\]). In the slot filling pipeline, the candidate extraction component extracts filler candidates based on their named entity types. In contrast to the input of the binary models, the input of the multi-class model can contain relation arguments of all available types. This complicates the relation classification task of the model. A context for the relation `per:date_of_birth`, for example, might be similar to a context for the relation `per:location_of_birth`. To simplify the classification, we propose to provide the models with the named entity types of the relation arguments. In particular, we investigate three different settings for augmenting the input of the multi-class model with named entity types. They are described in the following paragraphs. For all settings, we use the same coarse-grained types as we process in the slot filling pipeline, namely `PERSON, ORGANIZATION, LOCATION, DATE, NUMBER, O`. ### Pipeline Approach {#sec:pipeline} The first approach we explore is a pipeline approach based on the slot to evaluate. We use two binary (one-hot) type vectors $t_1$ and $t_2$ of the size of the type vocabulary as additional input to the network from Figure \[cnnFig\]. For the slot `per:employee_or_member_of`, for example, the type vector $t_1$ for the first relation argument would consist of only one 1 at the position of `PERSON` and 0 otherwise. The type vector $t_2$ for the second relation argument would consist of a 1 at the position of `ORGANIZATION` and a 1 at the position of `LOCATION` since a person can be employed by an organization or by a geo-political entity. Since the type vectors are based on the slots and the filler extraction with named entity recognition (see Section \[component:candidateExtraction\]), we call this approach “pipeline”. The type vectors are then fed into a hidden layer which creates type embeddings $E_1$ and $E_2$: $$E_i = \tanh(V^\top t_i + c)$$ with $V$ being the weight matrix and $c$ the bias term of the hidden layer. Then, we concatenate the type embeddings $E_1$ and $E_2$ with the pooling results of the CNN for relation classification to calculate a type-aware sentence representation $s$. Thus, Equation \[eq:hidden\] becomes: $$s = \tanh(W_1^\top P_{\text{left}} + W_2^\top P_{\text{middle}} + W_3^\top P_{\text{right}} + W_4^\top v + W_5^\top E_1 + W_6^\top E_2 + d) \label{eq:hidden2}$$ This is depicted in Figure \[fig:combination\]. ![Integration of entity type information into multi-class CNN.[]{data-label="fig:combination"}](CNN_combination_new.pdf){width=".6\textwidth"} ### Joint Training {#sec:jointTraining} ![Convolutional neural network for entity type classification; $t_1$ and $t_2$ are the predictions of the types of the first and second relation argument, respectively.[]{data-label="fig:typePrediction"}](CNN_types.pdf){width=".6\textwidth"} Instead of using prior knowledge about the slots and the expected types of their arguments, we propose to jointly learn entity and relation classification. Following , we use a convolutional neural network to predict scores for the different types (see Figure \[fig:typePrediction\]). We treat this as a multi-label classification task and use the sigmoid function to ensure that the scores for each class are between 0 and 1. We then use the scores as type vectors $t_1$ and $t_2$ in Figure \[fig:combination\]. This is similar to the architecture <span style="font-variant:small-caps;">Predicted-Hidden</span> from with the following differences: We do not use entity embeddings for modeling the relation arguments (see Section \[sec:neuralNets-remarks\] for our motivation) and also integrate the flag for the order of the relation arguments. This flag is highly relevant for slot filling since there is an inverse slot for almost all slots. For jointly training the CNN for entity classification and the CNN for relation classification, we use the following loss function: $$L = (1-\alpha) \cdot L_{\text{rel}} + \frac{\alpha}{2} \cdot L_{\text{type1}} + \frac{\alpha}{2} \cdot L_{\text{type2}}$$ The weight $\alpha$ controls the ratio between the relation classification loss $L_{\text{rel}}$ and the losses of entity type classification $L_{\text{type1}}$ and $L_{\text{type2}}$. It is tuned on the development set. ### Structured Prediction {#sec:structuredPrediction} ![CNN with structured prediction for type-aware slot filling relation classification.[]{data-label="fig:global"}](global_small.pdf){width=".9\textwidth"} The third approach for integrating entity information into a convolutional neural network for relation classification is based on structured prediction, as we originally presented for a table filling evaluation of entity and relation recognition [@adelEmnlp2017]. While we applied it only to a manually labeled dataset in that previous work, we now adopt it to the slot filling pipeline setting with distantly supervised training data for the first time. Figure \[fig:global\] shows the architecture of the model. Again, the context is split into three parts: left, middle and right context. For each context, a convolutional and 3-max pooling layer computes a representation with weights shared across contexts. For predicting the class of the first entity, we use the representation of the left and middle context; for predicting the class of the second entity, we use the representation of the middle and right context. To calculate scores for the possible relations, all three contexts are used. In contrast to the model we proposed earlier [@adelEmnlp2017], we again use the flag for the order of the relation arguments for classifying the relation. Also, we do not compute representations for entities for slot filling, as motivated in Section \[sec:neuralNets-remarks\]. The scores of the two entity classes and the relation class are then fed into a conditional random field (CRF) output layer which optimizes the following sequence of predictions: $$[\text{class of } e_1, \text{relation } r_{12} \text{ between } e_1 \text{ and } e_2, \text{class of } e_2]$$ In particular, we apply a linear-chain conditional random field. Thus, the model learns scores for transitions $T$ between the class of the first entity and the relation and between the relation and the class of the second entity. As a result, it approximates the joint probability of entity classes $C_{e_1}$, $C_{e_2}$ and relations $R_{e_1e_2}$ as follows: $$P(C_{e_1}, R_{e_1e_2}, C_{e_2}) \approx P(C_{e_1}) \cdot P(R_{e_1e_2}|C_{e_1}) \cdot P(C_{e_2}|R_{e_1e_2})$$ Our intuition behind this is that the dependency between relations and entities is stronger than the dependency between the two entities. Given neural network activations $v$ for the different entity and relation classes, the input sequence $d$ to the CRF layer is $$d = [v(e_1), v(r_{12}), v(e_2)] \label{eq:inputSeq}$$ This sequence is padded with a begin and end tag and used to compute the score for a particular sequence $s$ in the following way: $$\text{score}(s) = \sum_{i = 0}^{n}T_{s_is_{i+1}} + \sum_{i=1}^nd_{is_i} \label{eq:score}$$ with $T$ being the transition scores (randomly initialized and learned during training) and $d$ storing the neural network activations (see Equation \[eq:inputSeq\]). Following , we assume that all variables live in log space and, therefore, use the sum in Equation \[eq:score\]. For training, we normalize the score of the gold sequence over the scores of all possible sequences. We compute all possible sequences with the forward algorithm. To compute the best path during testing and get probabilities for the different relation classes, we apply the viterbi and forward-backward algorithm, respectively [@hmmPaper]. Experiments and Results {#sec:results} ======================= In this section, we describe our datasets (Section \[sec:data\]), and present and discuss our results. We conduct two sets of experiments: First, we evaluate our models in a pure relation classification setup (Section \[sec:resultsBenchmark\]). Second, we show the performance of the slot filling pipeline when using our models in the slot filler classification component (Section \[sec:end-to-end\]). Section \[sec:discussion\] discusses the difference between the results presented in Section \[sec:resultsBenchmark\] and Section \[sec:end-to-end\]. Finally, Section \[sec:soaSF\] sets our results in the context of other state-of-the-art slot filling pipelines. Data {#sec:data} ---- This subsection describes the different datasets we created for our experiments: Section \[sec:trainingdata\] presents the training data, Section \[sec:benchmark\] reviews the slot filling relation classification benchmarks we use to optimize our models (development set) and to evaluate their performance outside of the slot filling pipeline (test set). Section \[sec:dataMulti\] describes how we transform the training data which has been created for binary models into a dataset for training multi-class models. ### Training Data {#sec:trainingdata} The slot filling shared task does not offer a training dataset for relation classification models. Therefore, it is necessary to create one. Since manual labeling is expensive and does not scale to large amounts of data, we choose a distantly supervised labeling approach . In particular, we create a large set of training examples using distant supervision over Freebase relation instances [@Freebase] and the following corpora: - TAC source corpus (LDC2013E45) - NYT corpus (LDC2008T19) - subset of ClueWeb (@clueweburl-URL) - Wikipedia - Freebase description fields Negative examples for each relation are created by extracting sentences with entity pairs with the correct named entity tags for the given slot but without the given relation according to Freebase. However, as mentioned before, Freebase is incomplete. Thus, if a relation between two entities is not stored in Freebase, it does not mean that it does not exist in reality. Therefore, we clean the negative examples with trigger words and patterns: If a trigger/pattern of the given relation appears in the sentence, we do not include it in the set of negative examples. The list of triggers has been compiled manually based on the official slot descriptions and examples provided by TAC (@slotsurl-URL). It consists of a few high-precision patterns for each slot. By manually investigating random subsets of the filtered examples, we find the pattern set to be very effective in reducing the number of false negative labels. To reduce the number of false positive labels (and further improve the negative labels), we perform an automatic training data selection process. First, the extracted training samples are divided into $k$ batches. Then, we train one SVM per slot on the annotated slot filling dataset released by . Thus, the classifiers are trained on data with presumably correct labels and should, therefore, be able to help in the process of selecting additional data. For each batch of training samples, we use the classifiers to predict labels for the samples and select those samples for which the distantly supervised label corresponds to the predicted label with a high confidence of the classifier. Those samples are, then, added to the training data and the SVMs are retrained to predict the labels for the next batch. This process is depicted in Figure \[dataSelection\]. ![Training data selection process.[]{data-label="dataSelection"}](trainingdataSelection.pdf){width=".5\textwidth"} ### Development Data: Slot Filling Benchmark {#sec:benchmark} To optimize the parameters and to test the slot filling relation classifiers outside of the slot filling pipeline, we build a slot filling relation classification dataset, leveraging the existing manually labeled system outputs from the previous slot filling evaluations [@adelNaacl2016]: We extract the supporting sentences from the system outputs and automatically determine the position of the entity and the filler. Then, we label each sentence as correct or wrong according to the manual assessment provided by the shared task organizers. Due to differences in the offset calculation of some systems, we cannot use all available data: We extract 39,386 relation classification instances out of the 59,755 system output instances which have been annotated as either completely correct or completely incorrect by the shared task organizers. Thus, the resulting dataset has a reasonable number of examples with presumably clean labels. For our experiments, we split the data into a development part (data from slot filling evaluations 2012-2013) and a test part (data from slot filling evaluations 2014). We tune the parameters of our models on the development part and use the test part as a first indicator of their performance on unseen data. For more details on the data and a script to reproduce the data, see . ### Data for Multi-class Models {#sec:dataMulti} Since our training dataset has been created for binary models, the negative examples for each slot need to be processed for the multi-class setting: For example, a negative instance for the slot `per:date_of_birth` is not automatically a negative instance for the slot `per:date_of_death`. Therefore, we filter the negative instances in the training data with pattern lists: A negative instance that includes a trigger for any of our positive slots is deleted from the set. The remaining negative instances are labeled with an artificial class `N`. Finally, we found it beneficial on the slot filling relation classification benchmark to subsample the number of negative instances. Thus, we use the same number of negative instances as non-negative instances. Note that we only modify the training set and still use the original development and test sets from the slot filling relation classification benchmark for our experiments in order to compare the multi-class models with the binary models. Results on Slot Filling Benchmark {#sec:resultsBenchmark} --------------------------------- In this subsection, we present the performance of our models on the slot filling benchmark dataset (which is described in Section \[sec:benchmark\]). Section \[sec:baselines\] summarizes the baseline models to which we compare our type-aware convolutional neural networks, and Section \[sec:evaluationBenchmark\] describes the evaluation measure. Finally, Section \[sec:benchmarkResults\] provides and discusses the performance of our different models. ### Baselines and Support Vector Machines {#sec:baselines} In earlier work [@adelNaacl2016], we compared various models on the slot filling relation classification benchmark dataset described in Section \[sec:benchmark\]. In this paper, we use the same baseline models, namely <span style="font-variant:small-caps;">Mintz++</span> [@distant] and <span style="font-variant:small-caps;">MimleRe</span> . For another comparison, we also train multi-class variants of the SVM models from . For training the multi-class SVM, we apply LinearSVC from scikit learn (@svmurl-URL) with automatically adjusted class weights and compare the one-vs-rest training strategy (which actually trains binary classifiers by contrasting the examples from one class against the examples from all other classes) with the multi-class training strategy by (which changes the objective function to optimizing multiple classes jointly). The SVM is implemented using liblinear . ### Evaluation {#sec:evaluationBenchmark} The different models are evaluated using $F_1$, the harmonic mean of precision $P$ and recall $R$ of the classifiers: $$F_1 = \frac{2 \cdot P \cdot R}{P + R}$$ We calculate both slot-wise $F_1$ scores and a macro $F_1$ score which is the average of the scores over all slots. ------------------------- ----------------------------------------------------- ---------------------------------------------------- --------- --------- --------- --------- --------- --------- --------- --------- <span style="font-variant:small-caps;">Mintz</span> <span style="font-variant:small-caps;">Miml</span> binary o-v-r c-s binary multi +p +j +s per:age .71 .73 .74 .70 .70 **.76** .68 .72 .66 .67 per:alternate\_names .03 .03 .02 .00 .00 **.04** .00 .00 .00 .00 per:children .43 .48 **.68** .39 .45 .61 .48 .44 .44 .36 per:cause\_of\_death .42 .36 .32 .00 .00 **.52** .11 .29 .00 .06 per:date\_of\_birth .60 .60 .67 .33 .57 .77 **.80** **.80** .73 .73 per:date\_of\_death .45 .45 .54 .60 **.62** .48 .51 .46 .39 .59 per:empl\_memb\_of .36 **.37** .36 .21 .15 **.37** .28 .29 .25 .28 per:location\_of\_birth .22 .22 .27 .31 .33 .23 **.36** .20 .34 .35 per:loc\_of\_death .41 **.43** .34 .32 .35 .28 .28 .19 .25 .21 per:loc\_of\_residence .11 .18 **.33** .11 .08 .23 .15 .06 .22 .25 per:origin .48 .46 **.64** .04 .02 .39 .11 .30 .13 .17 per:schools\_att **.78** .75 .71 .58 .62 .55 .45 .47 .56 .68 per:siblings .59 .59 .68 .71 .68 .70 **.73** .54 .63 .68 per:spouse .23 .27 .32 .45 **.49** .30 .39 **.49** .36 .30 per:title .39 .40 .48 **.51** .45 .46 .42 .43 .44 .48 org:alternate\_names .46 .48 .62 .59 .52 **.66** .58 .55 .50 .58 org:date\_founded .71 .73 .70 .60 .60 .71 .63 .65 **.74** .69 org:founded\_by .62 .65 **.74** .70 .68 .68 .71 .43 **.74** .73 org:loc\_of\_headqu .19 .20 .42 .14 .11 **.45** .24 .42 .21 .34 org:members .06 .16 .13 .15 **.31** .04 .17 .07 .17 .11 org:parents .14 .17 **.20** .10 .11 .16 .14 .09 .12 .10 org:top\_memb\_empl .44 .46 .55 .54 .51 .53 .55 .49 **.58** **.58** macro $F_1$ .40 .42 **.48** .37 .38 .45 .40 .38 .38 .41 ------------------------- ----------------------------------------------------- ---------------------------------------------------- --------- --------- --------- --------- --------- --------- --------- --------- : $F_1$ results on slot filling benchmark test data (from 2014). The columns show our binary and multi-class SVMs and CNNs as well as our type-aware CNN models in comparison to two standard baseline models for relation classification: <span style="font-variant:small-caps;">Mintz</span> and <span style="font-variant:small-caps;">Miml</span> (see Section \[sec:baselines\] for more information). o-v-r: one-vs-rest; c-s: Crammer and Singer; p: pipeline (Section \[sec:pipeline\]), j: joint training (Section \[sec:jointTraining\]), s: structured prediction (Section \[sec:structuredPrediction\]).[]{data-label="tab:yearwiseResults"} ### Results {#sec:benchmarkResults} Table \[tab:yearwiseResults\] provides slot-wise results for the baseline models, the SVMs as well as for the different CNN setups: binary CNNs, a multi-class CNN without entity type information, a multi-class CNN with slot-based entity types (+p), a multi-class CNN with entity type probabilities jointly trained with the relation classification CNN (+j), and a multi-class CNN with a structured prediction (CRF) output layer which optimizes a sequence of entity and relation classes (+s). Note that Table \[tab:yearwiseResults\] shows results for 22 slot types instead of 65 slot types mentioned in Section \[sec:task\]. The reason is that we merged all location slots and all inverse slot pairs into one slot type (see Section \[sec:neuralNets-remarks\]). For example, our slot type `per:children` actually covers both original slot types `per:children` and `per:parents`, and our slot type `org:founded_by` covers even four slot types, namely `org:founded_by`, `per:organizations_founded`, `org:organizations_founded` and `gpe:organizations_founded`. As a result, our merged slot types in Table \[tab:yearwiseResults\] actually cover 54 out of all 65 slot types. We provide an overview which original slot types we cover with our models in Table \[tab:coveredslots\] in the appendix. For the remaining slot types, we were not able to extract enough training data with distant supervision to train machine learning models. For those slots, our slot filling pipeline falls back to pattern matching. In general, the binary models perform better than the multi-class models, even when adding entity type information to the latter. For example, the binary CNN achieves better results than the multi-class CNN without entity type information for 13 out of 22 slot types and better results than any multi-class CNN (with or without entity type information) for nine out of 22 slot types. We assume that one reason might be that the convolutional filters of the binary models can concentrate on learning features for just one relation type while the filters of the multi-class models need to learn features which can be used to discriminate all relation types. All binary models outperform the baseline models <span style="font-variant:small-caps;">Mintz</span> and <span style="font-variant:small-caps;">Miml</span>. The binary SVM performs slightly better than the binary CNN. In contrast to these results, the multi-class CNN outperforms the multi-class SVM. However, both models seem to struggle with the large number of output classes. While the model with jointly trained entity classification achieves slightly better results on the development set than the multi-class model without named entity information (0.53 vs. 0.52), this improvement is not transferred to the test set. The structured prediction model generalizes better to an unseen test set than the other type-aware models and comes closer to the results of the binary CNN. Slots for which entity type information seems to help the most are `per:date_of_birth` and `per:location_of_birth`, two slots with similar contexts. Similarly, type-aware CNNs achieve the best results for `org:date_founded` and `org:founded_by`. Despite their lower results on average, the multi-class models generalize better to unseen test data than the binary models for some slots, such as `per:date_of_death`, `per:schools_attended` or `per:spouse`. In the next section, we provide results for using the different models in the slot filling pipeline. Slot Filling Pipeline Results {#sec:end-to-end} ----------------------------- In this subsection, we present the performance of our slot filling pipeline (described in detail in Section \[sec:overview\]) on the official slot filling evaluation data from 2015. It consists of 1951 queries with manual assessments of the outputs of the systems that were submitted to the shared task evaluations. 930 of those queries are multi-hop queries, i.e., they require the system to first fill one slot and then fill a second slot based on the answers of the first slot (see Section \[sec:task\]). Section \[sec:systems\] presents the different system configurations we evaluate. Section \[sec:evaluationPipeline\] describes the evaluation measures and Section \[sec:resultsPipeline\] provides and discusses the results. ### System Configurations {#sec:systems} In this experiment, we run the whole slot filling pipeline as described in Section \[sec:overview\]. The different configurations we evaluate differ from one another in terms of the relation classification models that are used in the slot filler classification component. Note that all configurations use the pattern matching module in addition to the machine learning models. The numbers we provide in the following to distinguish the different configurations correspond to the numbers in Table \[tab:results-ldcMax2\]. System I only uses binary SVMs in combination with patterns. Thus, it can be used to assess the impact of CNNs in general. System II combines binary SVMs and binary CNNs. To assess the impact of joint training with entity types also for the binary models, we add another configuration, system III, that uses binary SVMs and binary CNN models which are jointly trained on entity typing and relation classification. Thus, in contrast to system II, it has the ability of correcting wrong entity types from the candidate extraction component of the pipeline. System IV uses multi-class SVMs instead of binary SVMs as in system I. For the multi-class SVMs, we use the one-vs-rest training strategy since this led to slightly better results on the development part of the slot filling relation classification benchmark (0.51 vs. 0.50). System V combines multi-class SVMs with multi-class CNNs. Thus, comparing system II to system V allows us to assess the performance difference of binary and multi-class models for slot filling. Finally, systems VI, VII and VIII combine multi-class SVMs with type-aware multi-class CNNs, our main contribution in this paper. System VI integrates pipeline-based CNNs, as described in Section \[sec:pipeline\], into the slot filling pipeline. In system VII, we apply CNNs that have been jointly trained on entity and relation classification, as presented in Section \[sec:jointTraining\]. The last configuration, system VIII, uses CNN models trained with structured prediction of both entity and relation classes. Note that we only compare pure binary to pure multi-class classification modules in this paper. We also experimented with combining binary SVMs with multi-class CNNs but did not obtain additional performance gains with this setup. ### Evaluation {#sec:evaluationPipeline} The slot filling pipeline is evaluated using the official measures from the shared task. In particular, we report two measures, one based on the micro $F_1$ score (the $F_1$ score over all examples, giving slots with more examples higher weight), and another one based on the macro $F_1$ score (the average of slot-wise $F_1$ scores, weighting all slots equally): “CSLDC max micro” is the micro precision, recall and $F_1$ score over all queries (thus, “micro”). If the hop 0 sub-query occurred several times in the query set (with different hop 1 sub-queries), that answer to the hop 0 sub-query is scored which leads to the maximum results over both hops (thus, “max”). “CSLDC max macro” is the average $F_1$ score over all slots (thus, “macro”). All scores in this section are calculated using the official shared task scoring scripts. Their readme file provides more details on the scoring procedure (@scoringurl-URL). ### Results {#sec:resultsPipeline} Table \[tab:results-ldcMax2\] provides the scores of our slot filling system with our newly introduced classification models in comparison to SVMs and CNNs (in both binary and multi-class variants) without entity type information. -- -------- -------- ------------ ----------- ----------- ----------- ----------- macro SVM CNN $P$ $R$ $F_1$ $F_1$ (I) binary - 29.12 26.18 27.57 32.50 (II) binary binary 31.79 28.23 29.91 34.20 (III) binary binary + j 30.11 28.23 29.14 **34.81** (IV) multi - 25.25 12.07 16.33 18.44 (V) multi multi **34.42** 26.66 30.04 32.82 (VI) multi multi + p 23.58 **28.55** 25.83 30.91 (VII) multi multi + j 32.42 27.84 29.95 33.14 (VIII) multi multi + s 33.33 27.68 **30.25** 33.98 (I) binary - 7.36 4.78 5.80 6.13 (II) binary binary 9.80 **7.00** **8.17** 8.28 (III) binary binary + j 8.90 6.56 7.55 8.73 (IV) multi - 7.11 3.67 4.84 4.34 (V) multi multi 12.59 3.89 5.94 7.78 (VI) multi multi + p 6.62 3.00 4.13 4.66 (VII) multi multi + j **13.47** 5.00 7.29 8.15 (VIII) multi multi + s 12.24 5.22 7.32 **9.24** (I) binary - 21.75 17.30 19.27 23.06 (II) binary binary 23.80 **19.42** 21.39 24.92 (III) binary binary + j 22.52 19.23 20.75 **25.47** (IV) multi - 17.38 8.58 11.49 13.39 (V) multi multi **29.60** 17.20 21.76 23.86 (VI) multi multi + p 20.02 17.94 18.92 21.51 (VII) multi multi + j 27.97 18.36 **22.17** 24.20 (VIII) multi multi + s 27.70 18.36 22.08 25.12 -- -------- -------- ------------ ----------- ----------- ----------- ----------- : Slot filling pipeline results for our different models and model combinations. Evaluation measure: CSLDC max micro/macro. p: pipeline, j: joint training, s: structured prediction as in Table \[tab:yearwiseResults\].[]{data-label="tab:results-ldcMax2"} In contrast to the slot filling relation classification results (see Table \[tab:yearwiseResults\]), most of the multi-class models (systems IV–VIII in Table \[tab:results-ldcMax2\]) perform comparable or even better than the binary models (systems I–III). The best micro $F_1$ results for hop 0 and all (both hops) are achieved by using multi-class classification models (systems V,VII). The multi-class models have higher precision than the binary models across all hops (hop 0, hop 1 and all). The multi-class models with joint training of entity types and relation classification (system VII) achieves the highest overall micro $F_1$ score. In terms of macro $F_1$, the binary model with jointly learned types (system III) and the multi-class model with structured prediction (system VIII) perform best. This suggests that the opportunity to correct wrong entity types from the candidate extraction component is crucial and that joint modeling with entity types especially improves the performance on rare relations. Discussion of Results: Benchmark vs. Pipeline {#sec:discussion} --------------------------------------------- The performance ranking of the models on the benchmark dataset for slot filling relation classification (Table \[tab:yearwiseResults\]) is different to the performance ranking in the slot filling pipeline evaluation (Table \[tab:results-ldcMax2\]). This seems to be in contradiction to the positive correlation of results we reported earlier [@adelNaacl2016]. We assume that the reason is domain mismatch. While the positive correlation was calculated for running the pipeline on 2013/2014 slot filling evaluation data [@adelNaacl2016], the pipeline is now run on 2015 evaluation data. In 2015, the evaluation corpus for slot filling was changed, introducing many more discussion forum documents and significantly changing the ratio of domains (see Table \[tab:domains\]). This leads to a severe domain mismatch challenge for the components of the slot filling pipeline and reduces the correlation with the benchmark dataset which has been built based on 2012-2014 data. Nevertheless, participants of the slot filling shared task only have previous evaluation data available for developing and tuning their models. Therefore, we argue that it is still important to also evaluate models on the slot filling relation classification benchmark. ------------------ ----------- -------- -------- -------- number ratio number ratio news documents 1,000,257 47.65% 8,938 18.19% web documents 999,999 47.63% 0 0% discussion forum 99,063 4.19% 40,186 81.81% ------------------ ----------- -------- -------- -------- : Domains in slot filling corpora.[]{data-label="tab:domains"} Comparison with State of the Art {#sec:soaSF} -------------------------------- Finally, we set our results in the context of state of the art on the 2015 evaluation dataset for slot filling. rank team micro $F_1$ distant supervision? -------- ------------------------------- ------------- ---------------------- 1 Stanford 31.06 no 2 UGhent 22.38 yes CIS (with type-aware CNN) 22.17 yes 3 CIS (our official submission) 21.21 yes 4 UMass 17.20 yes 5 UWashington 16.44 yes median - 15.32 - : State-of-the-art results for slot filling. Stanford: [@stanford2015], UGhent: [@ughent2015], CIS: [@cis2015], UMass: [@umass2015], UWashington: [@uw2015]. The third line (“CIS with type-aware CNN”) is our best result from Table \[tab:results-ldcMax2\]. Our official submission (forth line) did not include type-aware models.[]{data-label="tab:soa"} Table \[tab:soa\] shows that our official submission to the shared task in 2015, which did not include type-aware models, was ranked at third position. With our type-aware convolutional neural networks, we can improve our result. Our pipeline performs comparable to the second rank now. It is considerably better than the system on rank 4 and performs clearly above median. Only the results of the top-performing system are still superior. As also described in Section \[sec:relWork\], there are two differences relevant to this: First, the top-performing system does not use information retrieval, like our system and most other systems, but stores preprocessed versions of the corpus in a database, including an index for all occurring entities. This requires extensive corpus preprocessing (such as the identification of all entities along with their positions in the documents) and data storage (the source corpus contains millions of documents) but makes it possible to access the query entities directly at test time. However, this approach is only possible with prior access to the whole corpus and cannot be applied to changing environments. In contrast, our retrieval-based pipeline is more flexible since it processes only those documents relevant to the input query. The second, and arguably more important, difference in terms of final performance is that the training dataset of the top-performing system has been labeled manually via crowdsourcing. In contrast, the datasets of the other systems are created with distant supervision. When training our models, we observed large performance differences depending on the quality of the training data. Therefore, we suspect that the main reason for the superior performance of the system by is their training data. Unfortunately, obtaining manual labels is time-consuming and challenging, even in the context of crowdsourcing. An example is the extension of the knowledge base schema to new relations or a more fine-grained distinction of existing relations which would always require manual relabeling. Therefore, automatic methods like distant supervision are still of high relevance. Among the systems using automatically created training data, our pipeline is state of the art. Analysis {#sec:analysis} ======== In this section, we analyze the behavior of the slot filling system in more detail in order to see which pipeline components need to be improved in the future. Section \[sec:recallAnalysis\] analyzes the recall of the different components of the slot filling pipeline, showing which components are responsible for which recall loss. Section \[sec:errorAnalysis\] presents a manual analysis of wrong system outputs, categorizing the errors with respect to which pipeline component is responsible for them. Finally, Section \[sec:ablation\] provides several ablation studies, indicating the impact of entity linking, coreference resolution and type-aware neural networks. Recall Analysis {#sec:recallAnalysis} --------------- Our first analysis investigates the recall of the different components and is similar to the analysis by . In particular, we evaluate the components of our system before the slot filler classification module. Thus, we measure which recall our system could achieve with a perfect slot filler classification module that does not lose any recall. ![Analysis of recall after the application of the different pipeline components.[]{data-label="fig:recallanalysis"}](recall_analysis.pdf){width=".8\textwidth"} Figure \[fig:recallanalysis\] shows the results on the slot filling assessment data from 2015 for hop 0: Information retrieval with <span style="font-variant:small-caps;">Terrier</span> IR and fuzzy string match is able to achieve a recall of 78.82%. The entity linking component hurts recall a bit. However, it also increases precision which leads to better overall results (cf., Section \[sec:ablationEL\]). Evaluating only the top 100 documents instead of all extracted documents from <span style="font-variant:small-caps;">Terrier</span> (maximum 300), leads to a recall loss of almost 5%. Thus, allowing the slot filling system a longer run time for processing all extracted documents could lead to a higher final recall (but potentially also to more false positive extractions and, thus, a lower precision). As mentioned before, choosing only the 100 most relevant documents has led to the best time-performance trade-off on data from previous evaluations (2013 and 2014). The sentence extraction component extracts the relevant sentences quite successfully with an additional recall loss of only 6.95%. Evaluating this component in more detail shows the importance of coreference resolution and aliases: The recall loss without coreference resolution is almost 10%, the recall loss without aliases is more than 13%. Finally, the candidate extraction component is able to extract most of the relevant candidates, yielding an overall recall of 59.64% before slot filling relation classification. Without coreference resolution for sentence extraction, the overall recall is 56.23%, without alias information for sentence extraction, the overall recall is 53.69%. Assuming a perfect slot filler classification component with $P=100\%$ and $R=100\%$, the maximum $F_1$ score of the whole slot filling system would be 74.72%. This number is about twice as high as the performance of the best slot filling system 2015 [@stanford2015] (see Table \[tab:soa\]) but still low compared to other NLP tasks. This illustrates the difficulties of the slot filling task and the importance of all individual components of the pipeline since especially recall losses cannot be recovered by subsequent components. Error Analysis {#sec:errorAnalysis} -------------- In our second analysis, we manually analyze 120 errors of our system from the official 2015 evaluations, i.e., its wrong (false positive) predictions. While Section \[sec:recallAnalysis\] shows the recall loss of the different pipeline components, this analysis categorizes which component is responsible for which false positive prediction, and as a result, for a precision loss of the pipeline. Table \[tab:errorAnalysis\] shows which pipeline component is responsible for how many errors. The numbers do not sum to 1 since for 7% of the cases, we could not unambiguously identify a single component as the error source. Error category ratio -------------------------------- ------- Alias component 9% Entity linking component 2% Candidate extraction component 21% Classification component 61% : Error analysis of the pipeline.[]{data-label="tab:errorAnalysis"} The alias component especially struggles with acronyms which can refer to several entities. An example is NL which is an acronym for “National League” in the document collection but got wrongly recognized as an acronym for the query entity “Nest Labs”. In the candidate extraction component, most errors (16% of 21%) occur in the named entity recognition part. For example, “Bloomberg” is wrongly tagged as organization although it is a person in the given context (“... people like Bloomberg ...”). Similarly, “Heinz” gets tagged as person although it is an organization in the given context (“MacDonald’s dropping Heinz after CEO change”). Nested named entities are also a challenge: In “Tom Clancy games”, for example, “Tom Clancy” gets tagged as a person although the whole phrase actually forms a single entity. For some instances (4% of 21%), the document has been incorrectly split into sentences and in the remaining cases (1% of 21%), coreference resolution failed. The classification component faces a very challenging task since most extracted filler candidates are false positives. Thus, it has to establish precision while keeping as much recall as possible. Based on a manual inspection of errors, the most important challenge for the classification component is long contexts which mention several relations between several entity pairs. An example is “Mikhail Kalashnikov, designer of the famed Russian AK-47 assault rifle, died on Monday in his home city of Izhevsk, an industrial town 1,300 km east of Moscow, local media reported” from which the relation `per:location_of_death` between Mikhail Kalashnikov and Moscow is extracted. Thus, our classification component correctly recognizes the relation trigger “died \[...\] in” but assigned it to the wrong relation arguments. This finding is in line with the study by who also identified long context as one of the main challenges in slot filling relation classification. Ablation Studies {#sec:ablation} ---------------- In our last analysis, we present ablation studies showing the impact of entity linking, coreference resolution and type-aware neural networks. We focus on entity linking and coreference resolution since the design choices of whether or not to integrate them into a slot filling system are among those with the highest disagreement among slot filling researchers (see our description of related work in Section \[sec:relWork\]). Furthermore, the type-aware neural networks for slot filling are the main contribution of this paper. Table \[tab:ablationEL\] compares the performance of our slot filling system with and without entity linking, coreference resolution, and type-aware neural networks, respectively. $P$ $R$ $F_1$ $\Delta$$F_1$ -- ------------------------ ------- ------- ------- --------------- joint 32.42 27.84 29.95 without entity linking 31.30 27.92 29.51 -0.44 without coreference 32.13 25.87 28.66 -1.29 without CNN 25.25 12.07 16.33 -13.62 joint 13.47 5.00 7.29 without entity linking 11.81 4.78 6.80 -0.49 without coreference 11.80 4.22 6.22 -1.07 without CNN 7.11 3.67 4.84 -2.45 joint 27.97 18.36 22.17 without entity linking 26.56 18.31 21.68 -0.49 without coreference 27.25 16.88 20.85 -1.32 without CNN 17.38 8.58 11.49 -10.68 : Impact of entity linking and coreference resolution and type-aware convolutional neural networks on the slot filling pipeline.[]{data-label="tab:ablationEL"} ### Impact of Entity Linking {#sec:ablationEL} The system performance is slightly reduced by omitting entity linking. However, the difference of the $F_1$ scores is rather small. This shows that the main challenges of the system lie in other components and ambiguous names play a rather small role for the final results of the system. ### Impact of Coreference {#sec:ablationCoref} The $F_1$ score drops by 1.3 points when omitting coreference resolution. As expected, the impact on recall is higher than the impact on precision. However, also precision is reduced. This is because the number of true positives is reduced considerably (from 398 to 366) when the system does not use coreference information. The number of false positives is also lower, but the final results show that the impact of the number of true positives is larger. ### Impact of Neural Networks Type-aware CNNs have the largest impact on performance. They improve both precision and recall considerably. In contrast to SVMs, the usage of word embeddings allows the CNNs to detect synonyms or phrases which are similar but not the same as the ones learned during training. Training them jointly with entity classification allows them to benefit from the mutual dependencies between entity and relation classes. Discussion: Lessons Learned --------------------------- When developing the slot filling system and training the relation classification models, we had to solve several challenges. In this section, we report on lessons we have learned when developing the slot filling pipeline. Those are mainly qualitative statements which we have found by manually inspecting intermediate system outputs and results in the development process. Thus, many of them are based on prior experiments and not on concrete numbers reported in the paper. For the slot filling pipeline, it is especially useful to keep the extracted filler candidates and contexts as clean as possible, e.g., by applying genre-specific document processing steps and manually cleaning filler lists for string slots. Moreover, it is important for the recall of the system to extend the integration of coreference resolution, for instance, by applying it to both relation arguments. Our recall analysis and ablation study also show the positive impact of coreference resolution on the overall results. Based on our error analysis, the candidate extraction module and the classification module are responsible for most of the errors of the overall system. Thus, for future work, it is essential to focus on those two components. The performance of the relation classification models is mainly influenced by their training data and the input they get from the pipeline. If the input is very long, the models might extract wrong relations or relations between different entities than the query entity and the filler candidate. This is a particular challenge of slot filling since the quality of the inputs to the relation classification models directly depends on the previous system components. The significant difference of the top-ranked slot filling system to all the other systems in 2015 emphasizes the importance of a high-quality training dataset for slot filling relation classification. In an analysis of random subsamples of our training dataset before and after reducing wrong labels with patterns and self-training, we saw the importance of cleaning the noisy labels from distant supervision. Despite our automatic cleaning steps, our dataset still includes noise and a promising future research direction might be the exploration of techniques for further enhancing the data or collecting new data without distant supervision. Furthermore, our results suggest that multi-class models with entity type information are a promising direction for future research on slot filling. Related Work {#sec:relWork} ============ The slot filling shared task has been held since 2009. There are about 20 teams participating each year. Most systems apply a modular pipeline structure and combine multiple approaches, such as distant supervision and patterns [@overviewSF2014]. In 2015, we were one of the first teams using neural networks [@cis2015; @stanford2015; @umass2015]. In this section, different approaches for implementing the slot filling pipeline are described, followed by a more detailed description of two systems that are most relevant to our work: the top-ranked system from 2013 [@roth2013] since we use their distantly supervised patterns and similar features in our support vector machines; and the winning system from 2015 [@stanford2015] since we evaluate our system on the assessment data from 2015. Finally, we summarize more recent developments in slot filling research. Most slot filling systems consist of an information retrieval-based pipeline of different modules. Exceptions are, for example, the systems by and , which rely on relational databases consisting of one table storing all sentences from the corpus and another table storing all entity mentions. Most groups expand the query with aliases . Our system follows this line of work and uses information retrieval and query expansion to extract relevant documents and cover alternate names and spelling variations. For sentence extraction, only a subset of systems use coreference information . In all our experiments and analysis, coreference information improves the final results though. mention the long computing time of coreference resolution systems as a major drawback. This is why we make our coreference resource publicly available. Together with our positive results, it can help convince other researchers to integrate this component which we consider very important, especially for the recall of a slot filling system. Even fewer systems apply entity linking or another form of disambiguating different entities with the same name . Our results with entity linking are mixed: Although it slightly improves the final pipeline results, it leads to recall losses due to wrong links. Especially in 2012 and 2013, many systems relied only on pattern matching for identifying slot fillers . Now, more and more teams use machine learning models for slot filling relation classification, such as naive Bayes [@ucd2014], logistic regression [@nyu2014; @stanford2015; @cmu2016; @stanford2016], conditional random fields or support vector machines . <span style="font-variant:small-caps;">Mintz</span> and <span style="font-variant:small-caps;">Miml</span>, our baseline systems in Section \[sec:resultsBenchmark\], are used by, i.a., and . More recently, participants also train neural networks, such as bidirectional gated recurrent units [@pris2016], bidirectional long short-term memory (LSTM) networks [@pris2015; @stanford2015; @cmu2016; @rpi2016; @stanford2016; @umass2016] or convolutional neural networks . In 2015, we were one of the first to show the success of neural networks, especially convolutional neural networks, on this task which has led to their increasing popularity. With the exception of, e.g., or who use human labels or manual cleaning of noisy labels, e.g., in connection with active learning, the machine-learning models are trained with distant supervision. In this work, we approach the problem of noisy labels by an automatic self-training procedure. A subset of the data we use for self-training includes crowdsourced annotations but the actual cleaning process is fully automatic and uses machine learning methods. Thus, it scales better to larger datasets than manual cleaning or manual labeling. While some of the participants use binary models, i.a., or , others train multi-class models, i.a., and . In this study, we provide a direct comparison of the performance of binary and multi-class models. formulate entity type constraints and use integer linear programming (ILP) to combine them with relation classification. In contrast, our type-aware models are trained end-to-end and do not rely on hand-crafted hard constraints. Instead, they are able to learn correlations between entity and relation classes from data. In the relation extraction community, the joint modeling of entity types and relations is known to improve results . However, only very few approaches use neural models for joint modeling as we do in this research. Examples are or but both of them apply their models to clean datasets which have been manually labeled with entity types and relations. In contrast, we conduct our experiments on distantly supervised slot filling which provides neither clean labels for entity nor for relation types. We show that it is possible to use joint training of entity and relation classification in order to reduce the problem of error propagation in the slot filling pipeline. The positive results of our type-aware multi-class CNNs may motivate other researchers in slot filling or general relation extraction to extend their neural models with entity type information or other features which are known to be useful for relation classification with traditional models. The top-ranked system in 2013 [@roth2013] follows the main trends in slot filling and applies a modular system based on distant supervision which is called <span style="font-variant:small-caps;">RelationFactory</span>. Its pipeline is similar to ours except that it uses neither entity linking nor coreference resolution. Following , we use their distantly supervised patterns and add skip n-gram features to the feature set of our support vector machines. An important difference of our models, however, is that we also integrate neural networks and train not only binary models but also multi-class models with and without entity type information. The top-performing system in 2015 [@stanford2015] uses manually labeled training data [@active] as well as a bootstrapped self-training strategy in order to avoid distant supervision. In contrast to most other slot filling systems, they do not apply a pipeline system based on information retrieval but store preprocessed versions of all sentences and entity mentions from the source corpus in a relational database which they access during evaluation. As relation extractors, they apply a combination of patterns, an open information extraction system, logistic regression, a bidirectional long short-term memory network and special extractors for website and alternate names slots. In contrast to their system, we apply a traditional slot filling pipeline based on information retrieval and train convolutional neural networks. In earlier work [@cis2015], we also combined convolutional and recurrent neural networks and found that adding recurrent neural networks increased the performance only slightly. Thus, we assume that the main reason for their better performance is the less noise in the labels of their training data. Last, we summarize more recent developments in slot filling research: present a method based on trigger extraction from dependency trees which does not require (distantly) supervised labels and can work for any language as long as named entity recognition, part-of-speech tagging, dependency parsing and trigger gazetteers are available. follow our work in extracting training and development data and in using convolutional neural networks for slot filling relation classification. They input dependency paths into the network and apply attention in order to account for the larger middle contexts in slot filling relation classification. In contrast to their work, we extend the convolutional neural network in this paper to not only doing relation classification but jointly learning to classify entities and relations. Recently, propose position-aware attention which calculates attention weights based on the current hidden state of their LSTM, the output state of the LSTM and the position embeddings which encode the distance of the current word to the two relation arguments. Moreover, they publish a supervised relation extraction dataset, obtained by crowdsourcing, for training slot filling relation classification models. address the issue of evaluating new slot filling systems outside of the official shared task evaluations. They build an evaluation method based on importance-sampling and crowdsourcing which they make publicly available. Conclusion {#sec:conclusion} ========== In this paper, we proposed different type-aware convolutional neural network architectures for slot filling. After describing our slot filling pipeline, we focused on its relation classification component. Slot filling relation classification is a task with a particularly noisy setup due to distantly supervised data at training time and error propagation through the pipeline at test time. We were one of the first groups to show that convolutional neural networks are successful classification models in this scenario. We proposed three techniques of incorporating entity type information: a pipeline-based, a joint training and a structured prediction approach. In our experiments, we compared binary and multi-class models and showed that the multi-class models improved the final performance of the slot filling pipeline. The model trained jointly on entity and relation classification achieved the best micro $F_1$ scores while the model with structured prediction performed best in terms of macro $F_1$. Finally, we presented several analyses to assess the impact and errors of the different components of the pipeline, a very important aspect which is not evaluated in the official slot filling shared task. Our recall analysis showed the importance of aliases and coreference resolution. Our manual error analysis revealed that the candidate extraction component (especially the named-entity-recognition module) and the classification component are responsible for most of the errors of the system. Finally, ablation studies confirmed the large positive impact of our type-aware convolutional neural network on the performance of the whole slot filling pipeline. We publish our complete slot filling system, including the source code and the presented models (<http://cistern.cis.lmu.de/CIS_SlotFilling>), as well as our coreference resource (<http://cistern.cis.lmu.de/corefresources>) along with this paper. Overview of Slots ================= As described in Section \[sec:neuralNets-remarks\], we merge different location slots and inverse slots into one slot type to use our training data most effectively and avoid redundant training. In Table \[tab:coveredslots\], we provide an overview of the official slot types we cover with our models. [.9]{}[l|&gt;X]{} our label & covered original slot types\ per:age & per:age\ per:alternate\_names & per:alternate\_names\ per:cause\_of\_death & per:cause\_of\_death\ per:children & per:children, per:parents\ per:date\_of\_birth & per:date\_of\_birth\ per:date\_of\_death & per:date\_of\_death\ per:employee\_or\_member\_of & per:employee\_or\_member\_of, org:employees\_or\_members, gpe:employees\_or\_members\ per:location\_of\_birth & per:city\_of\_birth, per:country\_of\_birth, per:stateorprovince\_of\_birth, gpe:births\_in\_city, gpe:births\_in\_country, gpe:deaths\_in\_stateorprovince\ per:loc\_of\_death & per:city\_of\_death, per:country\_of\_death, per:stateorprovince\_of\_death, gpe:deaths\_in\_city, gpe:deaths\_in\_country, gpe:deaths\_in\_stateorprovince\ per:loc\_of\_residence & per:cities\_of\_residence, per:countries\_of\_residence, per:statesorprovinces\_of\_residence, gpe:residents\_of\_city, gpe:residents\_of\_country, gpe:residents\_of\_stateorprovince\ per:origin & per:origin\ per:schools\_attended & per:schools\_attended, org:students\ per:siblings & per:siblings\ per:spouse & per:spouse\ per:title & per:title\ org:alternate\_names & org:alternate\_names\ org:loc\_of\_headquarters & org:city\_of\_headquarters, org:country\_of\_headquarters, org:stateorprovince\_of\_headquarters, gpe:headquarters\_in\_city, gpe:headquarters\_in\_country, gpe:headquarters\_in\_stateorprovince\ org:date\_founded & org:date\_founded\ org:founded\_by & org:founded\_by, per:organizations\_founded, org:organizations\_founded, gpe:organizations\_founded\ org:members & org:members, org:member\_of, gpe:member\_of\ org:parents & org:parents, org:subsidiaries, gpe:subsidiaries\ org:top\_members\_employees & org:top\_members\_employees, per:top\_member\_employee\_of\ For the following slots, we could not extract enough training data to train machine learning models: per:charges, per:other\_family, per:religion, org:date\_dissolved, org:number\_of\_employees\_members, org:political\_religious\_affiliation, org:shareholders, org:website, {per, org, gpe}:holds\_shares\_in. For those slots, our slot filling pipeline uses only the pattern matcher. Hyperparameters =============== For tuning the CNN models, we performed grid-search over the following ranges of hyperparameters: filter width $\in \left\{3,5\right\}$, \# conv filters $\in \left\{100,300,1000,3000\right\}$, hidden units for relation extraction $\in \left\{100, 300, 1000\right\}$, hidden units for entity classification $\in \left\{25,100\right\}$. Other hyperparameters are the same for all models. For example, we use stochastic gradient descent with minibatches of size 10 and a learning rate of 0.1. For regularization, we add a L2 penalty with a weight of 1e-5. For the SVM models, we only tune $C$, the penalty of the error term. All the other parameters are set to the default parameters of the implementation. For example, the SVMs are trained with squared hinge loss and L2 regularization, the tolerance value for the stopping criteria is 1e-4. -- ---------------------------- -------------- ----------------- ---------- -------- ------- SVM filter width \# conv filters relation entity per:age 3 300 300 - 0.03 per:alternate\_names 5 300 300 - 0.03 per:cause\_of\_death 5 300 1000 - 1.00 per:children 3 300 100 - 0.01 per:date\_of\_birth 5 300 1000 - 0.01 per:date\_of\_death 5 300 100 - 0.01 per:empl\_memb\_of 3 300 100 - 0.01 per:location\_of\_birth 3 300 300 - 0.01 per:loc\_of\_death 3 300 300 - 0.01 per:loc\_of\_residence 3 300 100 - 0.01 per:origin 5 300 100 - 0.10 per:schools\_att 3 300 300 - 0.10 per:siblings 3 300 1000 - 0.01 per:spouse 3 300 300 - 0.10 per:title 3 300 100 - 10.00 org:alternate\_names 5 300 1000 - 10.00 org:date\_founded 5 300 100 - 0.01 org:founded\_by 5 300 300 - 0.01 org:loc\_of\_headqu 3 300 100 - 0.01 org:members 5 300 100 - 3.00 org:parents 5 300 300 - 0.01 org:top\_memb\_empl 3 300 100 - 0.30 without entity information 3 3000 100 - 0.30 + p 3 300 100 100 - + j 3 300 100 25 - + s 3 3000 100 100 - -- ---------------------------- -------------- ----------------- ---------- -------- ------- : Hyperparameters of CNNs and SVMs.[]{data-label="tab:hyperparamsCNN"} Table \[tab:hyperparamsCNN\] shows the hyperparameters tuned on the development part of the slot filling relation classification benchmark dataset. The configuration files which we use to specify the hyperparameters for the different models are included in the code which we publish along with this paper. 0.2in

ArXiv

--- abstract: 'We discuss the results of simulations of an intruder pulled through a two-dimensional granular system by a spring, using a model designed to lend insight into the experimental findings described by Kozlowski et al. \[Phys. Rev. E [**100**]{}, 032905 (2019)\]. In that previous study the presence of basal friction between the grains and the base was observed to change the intruder dynamics from clogging to stick–slip. Here we first show that our simulation results are in excellent agreement with the experimental data for a variety of experimentally accessible friction coefficients governing interactions of particles with each other and with boundaries. Then, we use simulations to explore a broader range of parameter space, focusing on the friction between the particles and the base. We consider a range of both static and dynamic basal friction coefficients, which are difficult to vary smoothly in experiments. The simulations show that dynamic friction strongly affects the stick–slip behaviour when the coefficient is decreased below 0.1, while static friction plays only a marginal role in the intruder dynamics.' author: - 'C. Manuel Carlevaro' - Ryan Kozlowski - 'Luis A. Pugnaloni' - Hu Zheng - 'Joshua E. S. Socolar' - Lou Kondic bibliography: - 'granulates.bib' title: 'Intruder in a two-dimensional granular system: Effects of dynamic and static basal friction on stick-slip and clogging dynamics' --- Introduction ============ Granular media respond in a variety of ways to applied loads, such as boundary shear [@granularfaultdanielshayman; @localglobalavalanchesbares; @SimGranularSeismicPicaCiamarra2011], intruding rods [@stickslipalbert; @stickslippulldiskoutMetayer], or surface sliders [@stickslipnasuno; @stickslipcracklingagheal; @SimGranularSeismicPicaCiamarra2011], exhibiting behaviors that include fluid-like flow, solid-like rigidity, and (sometimes periodic) cycles of stability and failure [@jaeger96b]. The response of a granular medium to a point-load, or single-grain perturbation, is a particularly sensitive probe of the connection between grain-scale dynamics and large-scale material stability and failure. Driving with a single particle avoids averaging over interactions of many grains with the extended objects (system boundaries, sliders, etc.) that are typically used to apply stress. Recent granular point-load studies have focused on the dynamical response of a single-grain intruder, mostly in quasi-two-dimensional (2D) beds of disks [@probeintruderreichhardt; @intrudervibrationgdauchot; @dragforcecavityformationintruderkolb; @slowdraggeng; @rheologyintruderexperimentseguin; @sticksliptordesillas]. In all cases, the dynamics is affected by the packing fraction and the strength of the driving mechanism, whether the intruder is driven at a constant velocity, by a constant force, or by a continually loading spring. An experiment on intruder dynamics in a 2D Couette geometry showed that friction between the particles and the supporting substrate (basal friction) also has a strong effect [@Kozlowski2019], as had previously been observed only in the context of quasistatic shear-jamming [@shearjamnobfhu]. In the present work, we develop numerical simulations and validate them against the experimental findings in  [@Kozlowski2019] with the aim of elucidating the influence of basal friction on the dynamics of an intruder moving through a granular medium, which is important for understanding the role of basal friction in supporting force-bearing clusters of grains subjected to a point load. We use numerical simulations to extend our study to parameter regimes that are difficult to access in the laboratory. In particular, we smoothly vary the basal dynamic and static friction coefficients from zero to the experimentally relevant value. We find, on the one hand, that dynamic friction ${\mu_{\rm s}}$ controls the overall dynamics with a clear stick–slip behavior for ${\mu_{\rm s}}> 0.1$ and an intermittent clogging-like flow for ${\mu_{\rm s}}< 0.1$. Static friction, on the other hand, plays a marginal role in determining the intruder’s behavior. The paper is organized as follows. In Sec. \[sec:exp\], we review the experimental setup, and in Sec. \[sec:simulations\], we describe the simulation procedures. Section \[sec:results\] contains our key results. After describing the basic dynamics of the intruder, we compare the experimental and simulation results for different packing fractions for both frictional and frictionless substrates. We then present simulation results for a range of static and dynamic basal friction coefficients. Section \[sec:conclusions\] is devoted to our conclusions regarding the effect of basal friction on the intruder dynamics. The Experiment {#sec:exp} ============== The experimental system consists of a layer of bidisperse plastic disks confined to an annular region by solid walls. An intruder the size of one disk is pushed in the azimuthal direction around the annulus by means of an externally controlled arm attached to a torsion spring whose other end is driven at constant angular speed. The setup, shown in Fig. \[fig:experimentapp\], and observations have been described in detail in Ref. [@Kozlowski2019]; here we provide a brief summary. ![A top-down schematic of the experiment with a sample image of grains. []{data-label="fig:experimentapp"}](ExperimentSchematic.pdf){width="0.5\columnwidth"} A central axle supports a pusher arm rigidly attached to an intruder rod suspended in a bidisperse granular bed (grain diameters $d_{\rm s} = 1.28\,$cm and $d_{\rm l} = 1.60\,$cm). The grains are confined in an annular Couette geometry having a channel width of $17.8$ cm, or roughly 14 small particle diameters, and boundaries lined with ribbed rubber to minimize boundary slipping. The central axis is composed of a bottom rotating shaft (indicated as [**ii**]{} in Fig. \[fig:experimentapp\]) driven by a stepper motor at constant angular velocity $\omega = 0.119 \pm 0.006$ rad/s, and a top rotating shaft (indicated as [**i**]{} in Fig. \[fig:experimentapp\]) that is coupled to the bottom shaft by a torque spring ($\kappa = 0.431 \pm 0.001$ Nm/rad). The intruder (diameter $d_I = 1.59 \pm 0.01\,$cm, the size of a large grain) is held at a fixed radius from the annulus center ($R = 19.7 \pm 0.1$ cm). Load cells in the pusher arm measure the force between the granular system and the intruder. The torque spring loads the system until the granular medium yields and the intruder slips. Cameras above the annular bed allow for intruder and particle tracking. The force on the load cells is recorded every $0.01\,$s, and the cameras capture images every $0.02\,$s. For experiments with basal friction, particles are dry and lightly coated in powder to reduce friction. The static friction coefficient between the particles and the glass table was measured to be $\sim 0.36$, and we assume that the dynamic friction coefficient is slightly below this value. To perform experiments without basal friction, the annular region is filled with water, and particles float on water. In this case, the particle diameters are $d_{\rm s} = 1.30\,$cm and $d_{\rm l} = 1.62\,$cm, slightly larger than the dry case due to swelling of the plastic. In both experiments, the number ratio of large to small particles is approximately fixed at $1\,:\,2.75$. Simulations {#sec:simulations} =========== We have carried out discrete element method (DEM) simulations of two-dimensional (2D) systems of particles. The simulations were implemented by means of the Box2D library [@box2d], which uses a constraint solver to handle rigid bodies. Before each time step, a series of iterations (typically 100) is used to resolve constraints on overlaps and on static friction between bodies through a Lagrange multiplier scheme [@catto]. After resolving overlaps, the inelastic collision at each contact is solved and new linear and angular velocities are assigned to each body. The equations of motion are integrated through a symplectic Euler algorithm. Solid friction between grains is also handled by means of a Lagrange multiplier scheme that implements the Coulomb criterion with the dynamic and static friction coefficients set to be equal. The approach yields realistic dynamics for granular bodies [@pytlos2015modelling] with complex shapes, including sharp corners, and has been successfully used to study grains under tapping protocols [@carlevaro_jsm11; @irastorza2013exact] and under vigorous vibration [@sanchez2014effect]. ![Snapshot of a sample simulation. The light blue particles are bidisperse (see text for details), and the static inner and outer boundaries are formed by equilateral triangles. The blue disk is the intruder (which is moving counter-clockwise around the annulus). The red dot indicates the position of the end of the torque spring that is driven at constant velocity; the other end of the spring is attached to the blue particle. A movie of the dynamics is available as Supplementary Material. []{data-label="fig:setup"}](setup_rotated.jpg){width="0.5\columnwidth"} Cell ---- Figure \[fig:setup\] shows a snapshot of a typical simulation. The cell consists of two immobile concentric boundary “rings” forming an annular two-dimensional (2D) Couette cell. The boundaries are formed by small equilateral triangles facing inward (toward the annular channel), which prevent slippage at the boundary, serving the role of the ribbed rubber in the experiments. The inner ring consists of 72 triangles of side length $0.682 d$, where $d$ is the diameter of the small particles used for the granular pack (see below); the outer ring consists of 180 triangles of side length $0.683 d$. The inner and outer rings are $8.810 d$ and $22.800 d$ in radius, Gravity (with acceleration $g$) acts in the direction perpendicular to the Couette plane. The base on which the circular particles rest is modeled by implementing an effective solid friction as follows. If a particle is moving at a speed above a small threshold $v'$ (i.e., $|\bm{v}|>v'$), then a dynamic friction force with the base $F^{\rm pb}_{\rm d} = -{\mu_{\rm d}^{\rm pb}}mg \bm{v}/|\bm{v}|$ is applied to the center of mass of the particle. Whenever $|\bm{v}|<v'$, the particle is immobilized by setting $\bm{v}=0$. If at rest, the particle will only resume translational motion if the total external force exerted by other particles exceeds the static friction force with the base set to ${F_{\rm s}^{\rm pb}}={\mu_{\rm s}^{\rm pb}}mg$, with ${\mu_{\rm s}^{\rm pb}}=v'/(g \:dt)$, where $dt$ is the simulation time step. This ensures that a particle will resume motion only if its initial velocity due to the collisions in the previous time step exceeds the velocity threshold $v'$. Therefore, the static friction is controlled via $v'$. We do not implement rotational friction forces between the particles and the base in this model. Disks ----- The annulus is filled with a large to small $1:2.75$ binary mixture of circular particles. The small particles have diameter $d$ and mass $m$. The larger particles have diameter $d_{\rm l}=1.25d$ and $m_{\rm l}=(d_{\rm l}/d)^2 m =1.565m$. The packing fraction is set by inserting a given number of disks ($N=N_{\rm s}+N_{\rm l}$, where $N_{\rm s}$ and $N_{\rm l}$ are the number of small and large particles, respectively), conserving the $1:2.75$ number ratio. Table \[tab:packing-fractions\] lists all packing fractions explored. For the calculation of the packing fraction we exclude both the area of the triangles and the space between them. $\phi$ $N_{\rm l}$ $N_{\rm s}$ -------- ------------- ------------- 0.6480 245 676 0.6589 250 685 0.6691 253 697 0.6797 257 708 0.6899 260 720 0.6998 264 730 0.7104 268 741 0.7203 272 751 0.7308 276 762 0.7414 280 773 0.7513 284 783 0.7619 288 794 0.7724 292 805 0.7824 296 815 : List of packing fractions explored in the simulations and the corresponding numbers of large and small particles.[]{data-label="tab:packing-fractions"} Each disk interacts with other disks, the boundaries of the cell, and the intruder disk (see below) as a perfectly rigid impenetrable object. The result of a collision is controlled by a restitution coefficient $\epsilon$ and the static ${\mu_{\rm s}}$ and dynamic ${\mu_{\rm d}}$ friction coefficients. Unless explicitly noted, we take ${\mu_{\rm s}}={\mu_{\rm d}}$. Note that the particle–base interaction has ${\mu_{\rm s}^{\rm pb}}\neq{\mu_{\rm d}^{\rm pb}}$. Table \[tab:interactions\] lists the parameters for the various pair-wise interactions in the system. Acronym ${\mu_{\rm s}}$ ${\mu_{\rm d}}$ $\epsilon$ ------------------- --------- ----------------- ----------------- ------------ particle–particle pp 1.20 1.20 0.05 particle–annulus pa 0.77 0.77 0.05 particle–intruder pi 0.41 0.41 0.05 particle–base pb \[0.36;1.00\] \[0;0.36\] – intruder–base ib 0.00 0.00 – : List of friction and restitution coefficients for the various pair-wise interactions in the system. The numerical values for the parameters are motivated by the ones measured in the experiments.[]{data-label="tab:interactions"} Intruder -------- A circular particle of diameter $1.25 d$ (the size of a large particle) is used as an intruder. The intruder interacts only with the other disks, not with the base (see Table \[tab:interactions\]) as in the experiments, where the intruder is suspended above the base at all times. The intruder is constrained to move along a circle of radius $15.8 d$ centered at the annulus center. This is done by binding the intruder to a very stiff radial spring. A “soft” torque spring ($K=3591.98 mgd/{\rm rad}$) is connected to the intruder and rotated counter-clockwise at a low constant angular speed, $\omega = 0.00432\sqrt{g/d}$. This drives the intruder through the pack of disks. The attached spring can only pull the intruder; if the spring becomes shorter than its equilibrium length, no force is applied. The mass of the intruder is set to $380m$, a mass that yields the same moment of inertia with respect to the center of the annulus as the relevant moment of inertia in the experiment, which includes both the pushing arm and the intruder. The simulation time step is $0.001\sqrt{d/g}$ and the instantaneous intruder position and velocity (and spring force) are recorded every 100 time steps. The time step is sufficiently small to avoid numerical instabilities, and the results are consistent for smaller time steps. Results {#sec:results} ======= For both experiments and simulations, the intruder is driven completely around the annulus at least twice and at most ten times in a given run. For all statistical analyses, we ignore the first revolution, in which transient effects are observed as the intruder moves through an initially random configuration of grains. The friction coefficients and packing fractions of the experiments, as well as the post-processing performed on intruder velocity and force data, are matched in simulation. We first validate simulations by comparing the statistics of intruder velocity, spring force, creep velocity during sticking periods and waiting times between them with those of the experiment, and then continue to explore the influence of basal friction. Stick-slip dynamics \[sec:stickslip\] ------------------------------------- Figure \[fig:comp-timeseries\] shows examples of time series for the intruder velocity and force, where time is measured in terms of the cumulative drive angle $\theta = \omega t$. The force measured in experiments is the force of the grains acting on the intruder (compression of the load cells that hold the intruder) while in simulations the force presented is that exerted by the torque spring on the intruder. In a static configuration, these forces are the same, but while the intruder is moving, as during slips, the experimental force fluctuates more, as discussed further below. The packing fractions used for Fig. \[fig:comp-timeseries\] are the highest packing fractions $\phi_{\rm max}$ explored in both simulation and experiment (which are slightly different in the simulations for different values of the basal friction coefficient). Higher packing fractions could not be studied with the present experimental apparatus due to buckling of the particles out of plane. The thick, black overlay plots on intruder velocity are detected sticking periods, which are defined as series of consecutive data points in the intruder velocity that fall below threshold 0.04 rad/s for a duration of at least 0.4 s. Figure \[fig:comp-timeseries\] shows that simulations (a, c) and experiments (b, d) at comparable packing fractions produce qualitatively similar results for two different values of the basal friction coefficient. For a frictional base (${\mu_{\rm d}^{\rm pb}}=0.36$), the intruder displays a clear stick–slip dynamics, characterized by extended sticking periods followed by rapid slip events. During sticking periods the intruder velocity is nominally zero and the force of the grains (and torque spring) acting on the intruder increases approximately linearly with time. The granular medium eventually yields under the increasing point load, and a slip event occurs. During a slip, the intruder’s velocity fluctuates irregularly as it collides with many grains until the medium forms a stable structure again. In the experiments, the measured force fluctuates rapidly as the load cells register numerous collisions with grains. In the simulations, the measured force is of the torque spring acting on the intruder, with fluctuations that are relatively small compared to the force itself. For a frictionless base, both experiments and simulations show long periods of time during which the intruder moves at the drive speed with superimposed fluctuations. Occasionally, the intruder does get stuck for a short period and then slips. This behavior is reminiscent of clogging of grains that flow through a restricted aperture; we refer to it as intermittent flow dynamics [@Kozlowski2019]. ![Sample time series at maximum packing fraction for cases with basal friction (a,b) and without (c,d) in both simulation (a,c) and experiment (b,d). Thick black overlay plots indicate detected sticking periods. []{data-label="fig:comp-timeseries"}](TimeSeries.pdf){width="1\columnwidth"} Comparison between simulations and experiments {#sec:comparison} ---------------------------------------------- Figures \[fig:pdf-velocity\] and \[fig:pdf-force\] show the probability distribution functions (PDF) of the intruder velocity and the spring force in simulations and experiments. We have tested a range of packing fractions, $\phi$, for the frictional and the frictionless base. The range explored for the frictionless base is narrower because in this case even at relatively high $\phi$ the intruder simply moves at the drive velocity, very rarely getting stuck. Figure \[fig:pdf-velocity\] shows that the agreement between experiments and simulations is remarkably good. Note that the range of velocities observed are consistent as well as the qualitative forms of the distributions. In the frictional case, for low $\phi$, the velocity distribution has its maximum at the drive speed. However, at high $\phi$, this peak disappears and the maximum occurs at zero velocity, due to the fact that the intruder is stuck most of the time. These two distinct regimes are separated by a smooth transition region. Following Ref. [@Kozlowski2019], we refer to these two regimes as a stick–slip regime for high $\phi$ and as an intermittent flow (clogging-like) regime at low $\phi$. In the case of the frictionless base, we observe only the intermittent flow, independent of the value of $\phi$. ![Intruder velocity distribution for frictional (a-c) and frictionless (d-f) bases. A range of packing fractions have been studied as indicated by the color scale. Experimental (a,d) and simulation (b, e) results show the same trends and range of velocities. Panels (c) and (f) show a direct comparison of the experimental and simulation PDF for two specific values of $\phi$. The vertical dotted line in (a,b,d,e) marks the drive velocity.[]{data-label="fig:pdf-velocity"}](VelocityDistributions.pdf){width="1\columnwidth"} Figure \[fig:pdf-force\] shows that the spring force PDFs are very similar in experiments and simulations. The range of forces and the positions of the maxima in the experiments are very well captured by the simulations. For the frictional base, the PDFs are rather broad with a maximum at non-zero force for high $\phi$. However, for the low $\phi$ (intermittent flow) the PDF has its maximum very near zero force. Negative forces are not present in the simulations due to the modeling of the torque spring as incapable of sustaining tension. In experiments, negative forces are only measured if the spring completely decompresses and the intruder rebounds due to a collision within the central axis (see Ref. [@Kozlowski2019] for more details). For the frictionless base, all $\phi$s lead to PDFs with a maximum at or close to zero force, consistent with the relatively free flow of the intruder. Only at very high $\phi$ do occasional sticking periods lead to some larger forces and therefore longer distribution tails. ![Intruder force distribution for frictional (a-c) and frictionless (d-f) bases. A range of packing fractions have been studied as indicated by the color scale. Experimental (a,d) and simulation (b, e) results show the same trends and range of forces. Panels (c) and (f) show a direct comparison of the experimental and simulation PDF for two specific values of $\phi$.[]{data-label="fig:pdf-force"}](ForceDistributions.pdf){width="1\columnwidth"} Figure \[fig:summary-phi\](a) shows the average force as a function of $\phi$ for both the frictional and the frictionless base. Overall, experimental and simulation values are very similar, with the exception that at high $\phi$ the average forces are higher in simulations, corresponding to the fact that the upper cutoffs in the distributions of Fig. \[fig:pdf-force\] are somewhat larger in simulations than in experiments. As we show in Sec. \[sec:basalfriction\], this quantitative difference is a consequence of running these simulations with a higher static basal friction coefficient than that measured in experiments. ![(a) The average force of the force time series as a function of $\phi$, for the frictional (green) and the frictionless (blue) base. Open triangles correspond to simulation data and filled circles to experiments. Each shaded region ranges from the lower 10% cutoff to the upper 90% cutoff of the distribution. Error bars indicate the standard error of the average. Inset: Zoom in to highlight the case with no basal friction. (b) The average creep velocity of the intruder during detected sticking periods as a function of $\phi$. (c) The average waiting time as a function of $\phi$. Inset: Zoom in to highlight the case with basal friction.[]{data-label="fig:summary-phi"}](AverageForce_AverageWaitingTime.pdf){width="1\columnwidth"} We have measured two additional features of the intruder dynamics: (i) the creep velocity during sticking periods and (ii) the waiting time between consecutive sticking periods. For these statistical quantities, unless noted otherwise, we only present data from runs for which at least 20 sticking periods are detected. During a sticking period, the intruder does not remain fully static but creeps forward as the spring force increases. Figure \[fig:summary-phi\](b) shows that the creep velocity is independent of $\phi$ when basal friction is present and is substantially larger for the frictionless base, both in simulations and experiments. Figure \[fig:summary-phi\](c) shows that the average waiting time is about ten times longer for the frictionless base than for the frictional one at the high packing fractions where sticking events do occur in the frictionless case; sticking events occur less frequently for ${\mu_{\rm d}^{\rm pb}}=0$. With the frictional base, for which sticking events occur at lower $\phi$, we find a mild increase of the waiting time as $\phi$ is decreased below $0.67$. Here too, simulations and experiments are consistent. Given the close agreement between experiment and simulation, it appears that our simulations are a reliable tool for describing, studying, and explaining the dynamics observed in experiments. We next use simulations to explore in more detail the effect of basal friction on the intruder dynamics, which is a difficult task to carry out experimentally. Effect of basal friction {#sec:basalfriction} ------------------------ As we have discussed, there is a dramatic change in the intruder dynamics when friction with the base is removed. This naturally raises the issue of whether the dynamics can be tuned continuously by changing the basal friction, or whether there is a sharp transition at ${\mu_{\rm d}^{\rm pb}}=0$. Controlling basal friction in the experiments is prohibitively difficult, and this is where the simulations provide novel insights. We have run simulations with dynamic friction coefficient ${\mu_{\rm d}^{\rm pb}}$ in the range $[0.0,\, 0.36]$ at a high packing fraction ($\phi=0.7724$) while keeping the static friction coefficient constant at ${\mu_{\rm s}^{\rm pb}}=1.0$. Figure \[fig:pdf-dynamic-friction\] shows the intruder velocity and spring force distributions for a range of ${\mu_{\rm d}^{\rm pb}}$. The spring force PDFs shown in Fig. \[fig:pdf-dynamic-friction\](a) reveal that the probabilities for large forces within the stick-slip regime decrease as ${\mu_{\rm d}^{\rm pb}}$ is decreased from 0.36 down to 0.1, indicating that the lowering of ${\mu_{\rm d}^{\rm pb}}$ (in the stick-slip regime) induces shorter sticking periods. Within the intermittent flow regime (${\mu_{\rm d}^{\rm pb}}<0.1$), lowering ${\mu_{\rm d}^{\rm pb}}$ leads to the appearance of a sharp peak at zero force, caused mainly by the longer periods of continuous flow between sticks. The velocity distribution shown in Fig. \[fig:pdf-dynamic-friction\](b) seems to be nearly independent of ${\mu_{\rm d}^{\rm pb}}$ for ${\mu_{\rm d}^{\rm pb}}> 0.1$, showing a maximum at zero velocity, which is consistent with stick–slip behavior. However, for ${\mu_{\rm d}^{\rm pb}}<0.1$ the maximum of the velocity PDF abruptly shifts to the spring drive speed, a characteristic of the intermittent flow regime. ![PDF of the spring force (a) and the intruder velocity (b) for a range of basal dynamic friction coefficient ($0 \leq {\mu_{\rm d}^{\rm pb}}\leq 0.36$) and ${\mu_{\rm s}^{\rm pb}}=1.0$. Insets: Zoom into the peak values of the PDFs.[]{data-label="fig:pdf-dynamic-friction"}](Distributions_ChangingBF_update.pdf){width="0.75\columnwidth"} Figure \[fig:summary-dynamic-friction\] shows the average force, average creep velocity, and average waiting time as a function of the dynamic friction coefficient ${\mu_{\rm d}^{\rm pb}}$. Interestingly, the creep velocity and the average waiting time between the end of a sticking period and the beginning of the next one grow very little as the dynamic friction is decreased from $0.36$ to $0.1$. Thus, above ${\mu_{\rm d}^{\rm pb}}\approx 0.1$, a reduction in ${\mu_{\rm d}^{\rm pb}}$ does not significantly change the number of sticking periods (longer waiting times) and stiffness of the packing (larger creep velocities). However, a dramatic increase in both creep velocity and waiting time happens for ${\mu_{\rm d}^{\rm pb}}<0.1$, indicating a rather sharp transition between the stick-slip and intermittent flow regimes. ![Average force (a), average creep velocity during sticking periods (b), and average waiting time between sticking periods (c) as a function of ${\mu_{\rm d}^{\rm pb}}$ for ${\mu_{\rm s}^{\rm pb}}=1.0$. Error bars indicate the standard error of the average and the shaded area indicates the 10%–90% percentile in the force distribution. Note that the ${\mu_{\rm d}^{\rm pb}}=0$ data point in (b) and (c) is computed from only 11 events.[]{data-label="fig:summary-dynamic-friction"}](AverageForce_AverageWaitingTime_ChangingBF_update.pdf){width="1\columnwidth"} Figure \[fig:summary-static-friction\] shows the average force, average creep velocity, and average waiting time for different values of basal [*static*]{} friction coefficient ${\mu_{\rm s}^{\rm pb}}$, both for ${\mu_{\rm d}^{\rm pb}}=0$ and ${\mu_{\rm d}^{\rm pb}}=0.36$. In all cases we ensure ${\mu_{\rm d}^{\rm pb}}\leq {\mu_{\rm s}^{\rm pb}}$. Static friction seems to play a marginal role in the dynamics. The most salient feature is a marked drop in the spread of the forces (indicated by the width of the shaded region in Fig. \[fig:summary-static-friction\](a)) for ${\mu_{\rm d}^{\rm pb}}= 0.36$ when ${\mu_{\rm s}^{\rm pb}}$ is changed from 1.0 to 0.36 . This indicates that ${\mu_{\rm s}^{\rm pb}}> {\mu_{\rm d}^{\rm pb}}$ induces the occurrence of some longer lasting sticking periods, without affecting the average force significantly. This finding is confirmed in Fig. \[fig:pdf-static-friction\], which shows the PDFs for forces and intruder velocities. We note that the PDF in Fig. \[fig:pdf-static-friction\](a) for ${\mu_{\rm s}^{\rm pb}}=0.36$ suggests a cutoff at large forces, in agreement with the experimental observation shown in Fig. \[fig:pdf-force\]. We recall here that in Fig. \[fig:pdf-force\] we used ${\mu_{\rm s}^{\rm pb}}=1.0$, which is somewhat above the experimental static friction coefficient between the discs and the dry base. ![Average force (a), Average creep velocity during sticking periods (b), and average waiting time between sticking periods (c) as a function of the static particle-base friction ${\mu_{\rm s}^{\rm pb}}$ for two values of the dynamic friction: ${\mu_{\rm d}^{\rm pb}}=0$ (open diamonds) and ${\mu_{\rm d}^{\rm pb}}=0.36$ (filled diamonds). Error bars indicate the standard error of the average and the shaded area indicates the 10%–90% percentile in the force distribution. Note that all of the ${\mu_{\rm d}^{\rm pb}}=0$ data points are computed from fewer than 20 events. For the filled diamonds, error bars are smaller than the data points.[]{data-label="fig:summary-static-friction"}](AverageForce_AverageWaitingTime_ChangingStaticBF.pdf){width="1\columnwidth"} ![PDF of the spring force (a) and the intruder velocity (b) for a range of basal static friction coefficient ($0.36 \leq {\mu_{\rm s}^{\rm pb}}\leq 1.0$) for two values of the dynamic friction coefficient: ${\mu_{\rm d}^{\rm pb}}=0$ (thick lines) and ${\mu_{\rm d}^{\rm pb}}=0.36$ (thin lines). []{data-label="fig:pdf-static-friction"}](Distributions_ChangingStaticBF.pdf){width="0.9\columnwidth"} Conclusions {#sec:conclusions} =========== Inspired by a recent experimental finding on the dynamics of an intruder the size of a single grain dragged though a two-dimensional granular system [@Kozlowski2019], we have developed numerical simulations that allow us to deepen our understanding of the effect of basal friction. Experimentally, it has been observed that the intruder can flow rather smoothly with occasional short sticks (intermittent flow regime) or show a fully developed stick–slip dynamics. The former is observed if the base on which the particles sit is frictionless, the latter when the base is frictional. Our simulations yield results consistent with these dynamics and are also in fair quantitative agreement with the experiment. Having the numerical model validated against the experiments, we use the simulations to investigate how the transition from intermittent flow to stick–slip is controlled by both the dynamic and the static friction coefficient of the particle–base interaction. We have shown that the transition between the two dynamic regimes (intermittent flow and stick-slip) is clearly controlled by the dynamic friction coefficient with the base and not by the static friction coefficient. One may intuitively expect that static friction would play a major role by contributing to the stability of jammed configurations. However, one has to bear in mind that jammed configurations are reached through a dynamic process. The development of rigidity is a result of the interactions at play during motion of the particles, making ${\mu_{\rm d}^{\rm pb}}$ the more relevant factor in determining the waiting times between sticking events. Our interpretation is that, for ${\mu_{\rm d}^{\rm pb}}>0.1$, particle motion is damped strongly enough to allow rapid formation of stable, static force network structures. Further increase in ${\mu_{\rm d}^{\rm pb}}$ can only marginally reduce these waiting times. Decreasing ${\mu_{\rm d}^{\rm pb}}$ below $0.1$, however, reduces the damping enough so that particles do not quickly come to rest after a slip event is initiated. Particle–particle dissipative interactions (inelasticity and friction) then play the dominant role in slowing particles down, and this leads only to occasional clogging rather than strong sticking periods. Static friction is expected to affect the stability of the jammed states once they occur. One of the affected features is the decay of the intruder force PDF for large forces, which is associated with the stability of the jammed states. The larger ${\mu_{\rm s}^{\rm pb}}$ is, the higher the probability of finding strong intruder forces. This is indeed confirmed in Fig. \[fig:pdf-static-friction\], both for ${\mu_{\rm d}^{\rm pb}}=0$ and ${\mu_{\rm d}^{\rm pb}}=0.36$. Increasing ${\mu_{\rm s}^{\rm pb}}$ does not, however, induce a change of dynamic regime (intermittent flow or stick–slip); the main peaks in the force and velocity distributions remain unaltered. In short, static friction can only help the stick periods last longer when they do occur. The initiation of these sticking events, however, is determined solely by the basal dynamic friction. Our results show that the additional background dissipation (especially dynamic friction) provided by the supporting substrate is essential for observing a clear stick–slip dynamics in the case of single-grain perturbations. However, stick–slip is observed without the presence of a frictional substrate if a driving force is applied over a length scale that is much larger than the characteristic particle size [@SimGranularSeismicPicaCiamarra2011; @stickslipalbert; @stickslippulldiskoutMetayer; @stickslipnasuno; @stickslipcracklingagheal; @SimGranularSeismicPicaCiamarra2011]. One open question is whether an additional background dissipation in the case of larger perturbations can alter the dynamics. Existing studies on granular systems immersed in viscous fluids (which provide an additional dissipation) indeed do show a strong effect on the stick–slip dynamics for large scale perturbations [@Higashi2009]. However, the effects observed seem to be more connected to lubrication than to additional “background” dissipation. A natural question is then how parameters of an immersed system, such as particle size and fluid viscosity, can possibly strengthen or weaken stick-slip behavior as lubrication and drag compete, and whether the effects of fluid drag might be similar to those due to frictional dissipation. Finally, we note that simulations provide a much higher resolution for velocities and forces than those achieved in the experiments. In order to compare results we have used in the simulation data analysis the same velocity threshold to detect sticking periods as was used in analyzing the experimental data. This overlooks a number of detailed features observed in the simulations during the sticking periods that may also be present in the experiments but were not resolvable in the experimental system. In particular, the creep observed in simulations has a very rich structure of *micro-slips*. With the simulation model now validated, it will be interesting to explore the details of the dynamics that are difficult to measure in experiments, including the interparticle forces and particle displacements during micro-slip events. We thank Karen Daniels for valuable discussions. This work was supported by the US Army Research Office through grant W911NF1810184 and by the Keck Foundation. L. A. P. and C. M. C. acknowledge support by Universidad Tecnológica Nacional through grants PID-MAUTNLP0004415 and PID-MAIFIBA0004434TC and CONICET through grant RES-1225-17. C. M. C. also thanks the Norma Hoermann Foundation for partial funding for his visit to NJ. L. K. was supported in part by NSF Grant No. 1521717.

ArXiv

--- abstract: 'It is shown that the method of infinitesimal generators (“Racah’s method”) can be broadly and systematically formulated as a method applicable to the calculation of reduced coupling coefficients for a generic subalgebra chain $G\supset H$, provided the reduced matrix elements of the generators of $G$ and the recoupling coefficients of $H$ are known. The calculation of $\grpso{5}\supset\grpso{4}$ reduced coupling coefficients is considered as an example, and a procedure for transformation of reduced coupling coefficients between canonical and physical subalegebra chains is presented. The problem of calculating coupling coefficients for generic irreps of $\grpso{5}$, reduced with respect to any of its subalgebra chains, is completely resolved by this approach.' author: - 'M. A. Caprio' - 'K. D. Sviratcheva' - 'A. E. McCoy' title: | Racah’s method for general subalgebra chains:\ Coupling coefficients of $\grpso{5}$ in canonical and physical bases --- Introduction {#sec-intro} ============ Continuous symmetries and their associated Lie algebras facilitate the description of many-body systems both directly and indirectly. When a symmetry occurs as a dynamical symmetry of the system, the corresponding algebra immediately gives the spectroscopic properties of the system. However, even when a symmetry is strongly broken, the algebraic structure nonetheless provides a calculational tool, classifying the basis states used in a full computational treatment of the many-body problem and greatly simplifying the underlying calculational machinery. Lie algebras have a long history of application, in both these capacities, to nuclear spectroscopy and related problems.[@racah1965-group-spectro; @hecht1973:nuclear-symmetries; @iachello2006:liealg] The fundamental quantities underlying calculations within a Lie algebraic framework are the coupling coefficients of the algebra, also known as generalized Clebsch-Gordan coefficients or Wigner coefficients. These are needed in order to couple states (or operators) of good symmetry to yield new states (or operators) of good symmetry, and they are required for the calculation of matrix elements through the generalized Wigner-Eckart theorem of the algebra. The Lie algebra $\grpso{5}$, isomorphic to $\grpsp{4}$, has several distinct applications in nuclear theory, involving different physical realizations of the operators, and in which different subalgebra chains are relevant to the symmetry properties. The natural construction of $\grpso{5}$ in terms of generators of rotation in five-dimensional space gives rise to a canonical $\grpso{4}\sim\grpso{3}\otimes\grpso{3}$ subalgebra.[@fn-apologia] However, application as the proton-neutron pairing quasispin algebra[@helmers1961:shell-sp; @flowers1964:quasispin; @ichimura1965:seniority-isospin; @ginocchio1965:so5-quasispin; @hecht1967:so5-shell-wigner] requires reduction with respect to the $\grpu{1}\otimes\grpso{3}$ algebra of isospin and occupation number operators. For the dynamics of spin-$2$ bosons (as in the interacting boson model[@arima1976:ibm-u5; @iachello1987:ibm]) and for the Bohr collective model,[@bohr1998:v2; @eisenberg1987:v1; @rowe2004:spherical-harmonics; @debaerdemacker2008:collective-cartan; @rowe2009:acm] the appropriate reduction is instead with respect to a physical angular momentum $\grpso{3}$ subalgebra. In this article, it is shown that the method of infinitesimal generators (“Racah’s method”) can be broadly and systematically formulated as a method applicable to the calculation of reduced coupling coefficients for a generic subalgebra chain $G\supset H$, provided the reduced matrix elements of the generators of $G$ and the recoupling coefficients of $H$ are known (Sec. \[sec-method\]). More specifically, the problem of calculating coupling coefficients for generic irreps of $\grpso{5}$, reduced with respect to any of the subalgebra chains, is completely resolved by this approach. The calculation of reduced coupling coefficients for the $\grpso{5}\supset\grpso{4}$ canonical chain is considered in detail (Sec. \[sec-so5\]). Coupling coefficients reduced with respect to the noncanonical subalgebra chains of $\grpso{5}$ may be obtained by a similar application of Racah’s method, or they can be deduced from the canonical chain coupling coefficients by unitary transformation. The general formulation in the presence of outer multiplicities for $H$, numerical examples for $\grpso{5}$, and a detailed account of the transformation procedure between subalgebra chains are given in the appendices. Method {#sec-method} ====== Background and definitions {#sec-background} -------------------------- Consider a Lie algebra $G$ and subalgebra $H$. States which reduce this subalgebra chain may be identified by the irrep labels $\Gamma$ of $G$, the irrep labels $\Lambda$ of $H$, and a label $\lambda$ (typically the Cartan weights) to distinguish basis states within $\Lambda$, as $\ket{\Gamma\Lambda\lambda}$. The coupling coefficients, or generalized Clebsch-Gordan coefficients, for $G$ relate the uncoupled product states of two irreps of $G$ to the coupled states, as $$\label{eqn-coupling-simple} \ket[4]{\Gamma_1~~\Gamma_2\\\Gamma\\\Lambda\\\lambda} = \sum_{\substack{\Lambda_1\Lambda_2\\\lambda_1\lambda_2}} \ccg{\Gamma_1}{\Lambda_1}{\lambda_1}{\Gamma_2}{\Lambda_2}{\lambda_2}{\Gamma}{\Lambda}{\lambda} \ket[3]{\Gamma_1&\Gamma_2\\\Lambda_1&\Lambda_2\\\lambda_1&\lambda_2}$$ In general, additional labels will be required to resolve multiplicities. There may be “outer” multiplicities in the Clebsch-Gordan series for the outer product of $G$ (that is, $\Gamma_1\otimes\Gamma_2$ may contain the irrep $\Gamma$ more than once), and there may be “branching” multiplicites under the restriction of $G$ to $H$ (that is, the given irrep $\Gamma$ of $G$ may contain an irrep $\Lambda$ of $H$ more than once). The coupling relation (\[eqn-coupling-simple\]) generalizes, with multiplicities, to $$\label{eqn-coupling-mult} \ket[4]{\Gamma_1~~\Gamma_2\\\rho\Gamma\\a\Lambda\\\lambda} = \sum_{\substack{a_1\Lambda_1a_2\Lambda_2\\\lambda_1\lambda_2}} \ccg{\Gamma_1}{a_1\Lambda_1}{\lambda_1}{\Gamma_2}{a_2\Lambda_2}{\lambda_2}{\rho\Gamma}{a\Lambda}{\lambda} \ket[3]{\Gamma_1&\Gamma_2\\a_1\Lambda_1&a_2\Lambda_2\\\lambda_1&\lambda_2},$$ where $\rho$ is the outer multiplicity index for $G\otimes G\rightarrow G$, and the $a$ indices resolve the branching multiplicities for $G\rightarrow H$. Furthermore, $H$ may be subject to outer multiplicites ($H\otimes H \rightarrow H$). In the following discussion, we shall for simplicity take the subalgebra $H$ to be multiplicity free. Such is the case for the commonly encountered situation in which the physically relevant subalgebra $H$ is $\grpso{3}$, as well as for the subalgebra $\grpso{4}$ considered in Sec. \[sec-so5\]. However, the necessary generalizations in the presence of outer multiplicities on $H$ are given in Appendix \[app-mult\], as would be needed for consideration of, [*e.g.*]{}, chains involving $\grpsu{3}$ as a subalgebra. Racah’s factorization lemma[@racah1949:complex-spectra-part4-f-shell] allows the coupling coefficient appearing in (\[eqn-coupling-mult\]) to be decomposed as the product $$\label{eqn-factorization} \ccg{\Gamma_1}{a_1\Lambda_1}{\lambda_1}{\Gamma_2}{a_2\Lambda_2}{\lambda_2}{\rho\Gamma}{a\Lambda}{\lambda} = \cg{\Lambda_1}{\lambda_1}{\Lambda_2}{\lambda_2}{\Lambda}{\lambda} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda}$$ of a coupling coefficient of $H$, embodying all the dependence upon weights $\lambda$, with a [*reduced coupling coefficient*]{} (or [*isoscalar factor*]{}) for $G\supset H$. The reduced coupling coefficient is nonzero only if $\Gamma$ is contained in the outer product of $\Gamma_1$ and $\Gamma_2$ ([*i.e.*]{}, $\Gamma_1\otimes \Gamma_2\rightarrow \Gamma$), each irrep of $H$ is contained in the corresponding irrep of $G$ ([*i.e.*]{}, $\Gamma_1\rightarrow\Lambda_1$, $\Gamma_2\rightarrow\Lambda_2$, and $\Gamma\rightarrow\Lambda$), and $\Lambda$ is contained in the outer product of $\Lambda_1$ and $\Lambda_2$ ([*i.e.*]{}, $\Lambda_1\otimes \Lambda_2\rightarrow \Lambda$). The reduced coupling coefficients satisfy the orthonormality conditions[@wybourne1974:groups] $$\begin{aligned} \label{eqn-ortho-bra-sum} \sum_{a_1\Lambda_1a_2\Lambda_2} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho'\Gamma'}{a'\Lambda} &=\delta_{(\rho\Gamma)(\rho'\Gamma')}\delta_{aa'} \intertext{and} \label{eqn-ortho-ket-sum} \sum_{\rho\Gamma a} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda} \cg{\Gamma_1}{a_1'\Lambda_1'}{\Gamma_2}{a_2'\Lambda_2'}{\rho\Gamma}{a\Lambda} &=\delta_{(a_1\Lambda_1)(a_1'\Lambda_1')}\delta_{(a_2\Lambda_2)(a_2'\Lambda_2')},\end{aligned}$$ for any irrep $\Lambda$ such that $\Gamma\rightarrow\Lambda$. If $T^{\Lambda_T}$ is an irreducible tensor operator with respect to $H$, the Wigner-Eckart theorem for $H$ permits the expression of a general matrix element of $T^{\Lambda_T}_{\lambda_T}$ as[@wybourne1974:groups] $$\label{eqn-we} \me[3]{\Gamma'\\a'\Lambda'\\\lambda'}{T^{\Lambda_T}_{\lambda_T}}{\Gamma\\a\Lambda\\\lambda} = \cg{\Lambda}{\lambda}{\Lambda_T}{\lambda_T}{\Lambda'}{\lambda'} \rme[2]{\Gamma'\\a'\Lambda'}{T^{\Lambda_T}}{\Gamma\\a\Lambda},$$ in terms of a coupling coefficient for $H$ and a [*reduced matrix element*]{} with respect to $H$. A Wigner-Eckart theorem of this form \[or its generalization (\[eqn-we-mult\])\] may be obtained whenever $H$ is a compact, semi-simple Lie algebra. Several methods may be considered, in general, for constructing the reduced coupling coefficients of Lie algebras: 1. Recurrence relations among coupling coefficients may be obtained by considering the action of an infinitesimal generator $G_i=G_i^{(1)}+G_i^{(2)}$ on uncoupled and coupled states. This approach, used in the present construction, is broadly termed “Racah’s method” (see Ref. ) and generalizes the classic recurrence method for evaluating $\grpsu{2}\sim\grpso{3}$ Clebsch-Gordan coefficients.[@edmonds1960:am] 2. Recurrence relations and seed values may be obtained by considering the action of a “shift tensor”, lying outside the algebra, which connects different irreps of the algebra.[@rowe1997:su3-cg-algebraic] 3. Consistency relations among coupling and recoupling coefficients serve as the basis for a “building up” process,[@hecht1967:so5-shell-wigner; @hecht1969:su4-wigner] in which unknown coupling coefficients can be deduced from a few known coefficients. 4. Explicit realizations of an algebra can be obtained in terms of bosonic or fermionic creation and annihilation operators. Relations among coupling coefficients follow from considering the matrix elements of tensor operators acting on bosonic or fermionic states ([*e.g.*]{}, Ref. ). This approach is generally restricted to symmetric irreps, antisymmetric irreps, or irreps which can be obtained as simple combinations thereof. Indeed, all of these approaches have been applied or suggested, in various forms, for the calculation of specific classes of $\grpso{5}$ coupling coefficients.[@racah1965-group-spectro; @hecht1965:so5-wigner; @hecht1967:so5-shell-wigner; @hemenger1970:so5-quasispin; @wybourne1974:groups; @iachello1981:ibfm-spin6; @vanisacker1987:spin6-spin5-isf; @hecht1989:so5-vcs-cg; @hecht1993:so5-u2-vcs; @han1993:so5-coupling; @caprio2007:geomsuper2] For the [*symmetric*]{} irreps of $\grpso{5}$, one may also work with an explicit realization in terms of five-dimensional spherical harmonics as functions on the four-sphere. Their triple overlap integrals are then proportional to $\grpso{5}$ coupling coefficients.[@rowe2004:spherical-harmonics; @caprio2009:gammaharmonic] Racah relations among reduced coupling coefficients {#sec-relations} --------------------------------------------------- Let us now consider how the first approach, [*i.e.*]{}, Racah’s method[@racah1949:complex-spectra-part4-f-shell; @racah1965-group-spectro] based on the action of infinitesimal generators, can be generally and systematically formulated as a method applicable to the calculation of reduced coupling coefficients involving generic irreps of an arbitrary subalgebra chain. Consider the action of an infinitesimal generator $G_i$ of $G$ on the coupled product state of (\[eqn-coupling-mult\]). The generator on the product space is of the form $G_i=G_i^{(1)}+G_i^{(2)}$, where $G_i^{(1)}$ acts only on the space carrying the irrep $\Gamma_1$ and $G_i^{(2)}$ acts only on the space carrying the irrep $\Gamma_2$. The equivalence of the action of $G_i$ on the two sides of (\[eqn-coupling-mult\]) imposes conditions on the coupling coefficients connecting the different basis states used on the two sides. For effective application of Racah’s method, it is most convenient to recast these relations among coupling coefficients so that they involve only (1) [*reduced*]{} coupling coefficients of $G$ with respect to $H$, (2) [*reduced*]{} matrix elements of the generators of $G$, and (3) recoupling coefficients of $H$, as obtained in this section. Racah’s approach requires that the action of the generators on the basis states of an irrep be known explicitly. In general, if coupling coefficients are to be determined for states which reduce $G\supset H$, it is necessary to consider the action of the generators which are in $G$ but not in $H$, since only these generators can connect different irreps $\Lambda$ of $H$. Note that Racah’s method is essentially an extension of the classic scheme[@edmonds1960:am] for calculating the ordinary $\grpso{3}$ Clebsch-Gordan coefficients, via recurrence relations obtained by considering the known matrix elements of $J_\pm=J_\pm^{(1)}+J_\pm^{(2)}$, between the uncoupled product states $\smallket[2]{J_1&J_2\\M_1&M_2}$ and coupled states $\smallket[2]{J\\M}$ (see also Sec. \[sec-solution\]). If $T^{\Lambda_T}_{\lambda_T}$ is a generator of $G$, expressed as an irreducible tensor operator with respect to $H$, we begin by considering the matrix element of $T^{\Lambda_T}_{\lambda_T}=T^{\Lambda_T\,(1)}_{\lambda_T}+T^{\Lambda_T\,(2)}_{\lambda_T}$, between uncoupled and coupled product states, $$\label{eqn-racah-raw} \methreefour{\Gamma_1&\Gamma_2\\a_1\Lambda_1&a_2\Lambda_2\\\lambda_1&\lambda_2}{T^{\Lambda_T}_{\lambda_T}}{\Gamma_1~\Gamma_2\\\rho\Gamma\\a\Lambda\\\lambda} = \methreefour{\Gamma_1&\Gamma_2\\a_1\Lambda_1&a_2\Lambda_2\\\lambda_1&\lambda_2}{T^{\Lambda_T\,(1)}_{\lambda_T}}{\Gamma_1~\Gamma_2\\\rho\Gamma\\a\Lambda\\\lambda} + \methreefour{\Gamma_1&\Gamma_2\\a_1\Lambda_1&a_2\Lambda_2\\\lambda_1&\lambda_2}{T^{\Lambda_T\,(2)}_{\lambda_T}}{\Gamma_1~\Gamma_2\\\rho\Gamma\\a\Lambda\\\lambda}.$$ The coupling relation (\[eqn-coupling-mult\]) may be used to express each ket on the right hand side of (\[eqn-racah-raw\]) entirely in terms of uncoupled states, and its inverse, obtained by orthonormality of coupling coefficients, may be used to express the bra on the left hand side entirely in terms of coupled states. Since $T^{\Lambda_T}$, as a generator of $G$, does not connect different irreps of $G$, and since the matrix elements of $T^{\Lambda_T}$ between states within an irrep of $G$ depends only upon the irrep labels, the result simplifies to $$\begin{gathered} \label{eqn-racah-coupled-1-2} \sum_{\substack{a'\Lambda'\\(\lambda')}} \me[3]{\Gamma\\a'\Lambda'\\\lambda'}{T^{\Lambda_T}_{\lambda_T}}{\Gamma\\a\Lambda\\\lambda} \ccg{\Gamma_1}{a_1\Lambda_1}{\lambda_1}{\Gamma_2}{a_2\Lambda_2}{\lambda_2}{\rho\Gamma}{a'\Lambda'}{\lambda'} = \sum_{\substack{a_1'\Lambda_1'\\(\lambda_1')}} \me[3]{\Gamma_1\\a_1\Lambda_1\\\lambda_1}{T^{\Lambda_T}_{\lambda_T}}{\Gamma_1\\a_1'\Lambda_1'\\\lambda_1'} \ccg{\Gamma_1}{a_1'\Lambda_1'}{\lambda_1'}{\Gamma_2}{a_2\Lambda_2}{\lambda_2}{\rho\Gamma}{a\Lambda}{\lambda} \\ + \sum_{\substack{a_2'\Lambda_2'\\(\lambda_2')}} \me[3]{\Gamma_2\\a_2\Lambda_2\\\lambda_2}{T^{\Lambda_T}_{\lambda_T}}{\Gamma_2\\a_2'\Lambda_2'\\\lambda_2'} \ccg{\Gamma_1}{a_1\Lambda_1}{\lambda_1}{\Gamma_2}{a_2'\Lambda_2'}{\lambda_2'}{\rho\Gamma}{a\Lambda}{\lambda}.\end{gathered}$$ To introduce reduced coupling coefficients and reduced matrix elements, we apply Racah’s factorization lemma (\[eqn-factorization\]) and the Wigner-Eckart theorem (\[eqn-we\]), yielding $$\begin{gathered} \label{eqn-racah-semireduced} \sum_{\substack{a'\Lambda'\\(\lambda')}} \cg{\Lambda}{\lambda}{\Lambda_T}{\lambda_T}{\Lambda'}{\lambda'} \rme[2]{\Gamma\\a'\Lambda'}{T^{\Lambda_T}}{\Gamma\\a\Lambda} \cg{\Lambda_1}{\lambda_1}{\Lambda_2}{\lambda_2}{\Lambda'}{\lambda'} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a'\Lambda'} \\= \sum_{\substack{a_1'\Lambda_1'\\(\lambda_1')}} \cg{\Lambda_1'}{\lambda_1'}{\Lambda_T}{\lambda_T}{\Lambda_1}{\lambda_1} \rme[2]{\Gamma_1\\a_1\Lambda_1}{T^{\Lambda_T}}{\Gamma_1\\a_1'\Lambda_1'} \cg{\Lambda_1'}{\lambda_1'}{\Lambda_2}{\lambda_2}{\Lambda}{\lambda} \cg{\Gamma_1}{a_1'\Lambda_1'}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda} \\+ \sum_{\substack{a_2'\Lambda_2'\\(\lambda_2')}} \cg{\Lambda_2'}{\lambda_2'}{\Lambda_T}{\lambda_T}{\Lambda_2}{\lambda_2} \rme[2]{\Gamma_2\\a_2\Lambda_2}{T^{\Lambda_T}}{\Gamma_2\\a_2'\Lambda_2'} \cg{\Lambda_1}{\lambda_1}{\Lambda_2'}{\lambda_2'}{\Lambda}{\lambda} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2'\Lambda_2'}{\rho\Gamma}{a\Lambda}.\end{gathered}$$ It now remains to eliminate the coupling coefficients of $H$, and thus all reference to weights. The orthogonality relations allow these coefficients to be moved to the right hand side, resulting in sums of quadruple products of coupling coefficients. These sums are recognized as recoupling coefficients of $H$, specifically, the “unitary $6$-$\Lambda$ symbols”, or transformation brackets between basis states in the coupling schemes $[(\Lambda_1\Lambda_2)^{\Lambda_{12}}\Lambda_3]^{\Lambda}$ and $[\Lambda_1(\Lambda_2\Lambda_3)^{\Lambda_{23}}]^{\Lambda}$, which are given by $$\label{eqn-usixj} \usixj{\Lambda_1}{\Lambda_2}{\Lambda_{12}}{\Lambda_3}{\Lambda}{\Lambda_{23}} = \sum_{\substack{\lambda_1\lambda_2\lambda_3\\\lambda_{12}\lambda_{13}}} \cg{\Lambda_1}{\lambda_1}{\Lambda_2}{\lambda_2}{\Lambda_{12}}{\lambda_{12}} \cg{\Lambda_{12}}{\lambda_{12}}{\Lambda_3}{\lambda_3}{\Lambda}{\lambda} \cg{\Lambda_2}{\lambda_2}{\Lambda_3}{\lambda_3}{\Lambda_{23}}{\lambda_{23}} \cg{\Lambda_1}{\lambda_1}{\Lambda_{23}}{\lambda_{23}}{\Lambda}{\lambda}.$$ The $6$-$\Lambda$ symbol is nonvanishing only if the Clebsch-Gordan series relations $\Lambda_1\otimes\Lambda_2\rightarrow\Lambda_{12}$, $\Lambda_{12}\otimes\Lambda_3\rightarrow\Lambda$, $\Lambda_2\otimes\Lambda_3\rightarrow\Lambda_{23}$, and $\Lambda_1\otimes\Lambda_{23}\rightarrow\Lambda$ are satisfied. Let $\Phi(\Lambda_1\Lambda_2;\Lambda)$ denote the phase factor incurred by interchange of the first and second irreps in a coupling coefficient of $H$, [*i.e.*]{}, $\smallcg{\Lambda_2}{\lambda_2}{\Lambda_1}{\lambda_1}{\Lambda}{\lambda}= \Phi(\Lambda_2\Lambda_1;\Lambda)\smallcg{\Lambda_1}{\lambda_1}{\Lambda_2}{\lambda_2}{\Lambda}{\lambda}$. Then the condition (\[eqn-racah-semireduced\]) becomes, with labels renamed for simplicity, $$\begin{gathered} \label{eqn-racah-reduced} \sum_{a} \rme[2]{\Gamma\\a\Lambda}{T^{\Lambda_T}}{\Gamma\\a'\Lambda'} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda} \\= \sum_{a_1'\Lambda_1'}\Phi(\Lambda_1\Lambda_2;\Lambda)\Phi(\Lambda_1'\Lambda_2;\Lambda') \usixj{\Lambda_2}{\Lambda_1'}{\Lambda'}{\Lambda_T}{\Lambda}{\Lambda_1} \rme[2]{\Gamma_1\\a_1\Lambda_1}{T^{\Lambda_T}}{\Gamma_1\\a_1'\Lambda_1'} \cg{\Gamma_1}{a_1'\Lambda_1'}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a'\Lambda'} \\+ \sum_{a_2'\Lambda_2'} \usixj{\Lambda_1}{\Lambda_2'}{\Lambda'}{\Lambda_T}{\Lambda}{\Lambda_2} \rme[2]{\Gamma_2\\a_2\Lambda_2}{T^{\Lambda_T}}{\Gamma_2\\a_2'\Lambda_2'} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2'\Lambda_2'}{\rho\Gamma}{a'\Lambda'},\end{gathered}$$ expressed entirely in terms of the reduced coupling coefficients to be calculated, reduced matrix elements, and recoupling coefficients of the lower algebra $H$. For the important special case in which $H$ is the angular momentum algebra $\grpso{3}\sim\grpsu{2}$, the relation (\[eqn-racah-reduced\]) becomes $$\begin{gathered} \label{eqn-racah-reduced-so3} \sum_{a} \rme[2]{\Gamma\\aJ}{T^{(J_T)}}{\Gamma\\a'J'}_{\grpso{3}} \cg{\Gamma_1}{a_1J_1}{\Gamma_2}{a_2J_2}{\rho\Gamma}{aJ} \\= \sum_{a_1'J_1'} (-)^{J_2+J_T+J'}\Jhat \Jhat' \sixj{J_2}{J_1'}{J'}{J_T}{J}{J_1} \rme[2]{\Gamma_1\\a_1J_1}{T^{(J_T)}}{\Gamma_1\\a_1'J_1'}_{\grpso{3}} \cg{\Gamma_1}{a_1'J_1'}{\Gamma_2}{a_2J_2}{\rho\Gamma}{a'J'} \\+ \sum_{a_2'J_2'} (-)^{J_1+J_2'+J} \Jhat \Jhat' \sixj{J_1}{J_2'}{J'}{J_T}{J}{J_2} \rme[2]{\Gamma_2\\a_2J_2}{T^{(J_T)}}{\Gamma_2\\a_2'J_2'}_{\grpso{3}} \cg{\Gamma_1}{a_1J_1}{\Gamma_2}{a_2'J_2'}{\rho\Gamma}{a'J'},\end{gathered}$$ where $\Jhat\equiv(2J+1)^{1/2}$. Note that the customary form[@edmonds1960:am] of the Wigner-Eckhart theorem for $\grpso{3}$ is defined in terms of a $3$-$J$ symbol rather than a Clebsch-Gordan coefficient. This results in a reduced matrix element which differs in normalization and phase from the definition implied by the generic statement of the Wigner-Eckhart theorem in (\[eqn-we\]). The reduced matrix elements under these two conventions are related by $\rme{J_3}{T^{(J_2)}}{J_1}_{\grpso{3}}=(-)^{2J_2}\Jhat_3 \rme{J_3}{T^{(J_2)}}{J_1}$. Solution of the homogeneous system {#sec-solution} ---------------------------------- The condition (\[eqn-racah-reduced\]) yields a different relation among specific reduced coupling coefficients for each choice of values for the four irrep labels $a_1\Lambda_1$, $a_2\Lambda_2$, $\Lambda$, and $a'\Lambda'$ (the multiplicity index $a$ is summed over). If there are $N$ coupling coefficients for the coupling $\Gamma_1\otimes\Gamma_2\rightarrow\rho\Gamma$, then the relations obtained from (\[eqn-racah-reduced\]) constitute a linear, homogeneous system of equations in $N$ unknowns for these coupling coefficients. Note that the most familiar and traditional approach to extracting coupling coefficients, after obtaining some set of relations among them, is to proceed by recurrence (*e.g.*, in the familiar case of $\grpso{3}$[@edmonds1960:am] and in the “building-up process”,[@wybourne1974:groups] as well as in prior applications of Racah’s method to higher algebras[@hecht1965:so5-wigner; @han1993:so5-coupling]). That is, a seed value is given for one coupling coefficient, and further coefficients are deduced inductively (one by one) from those already obtained. ![The classic problem of constructing the $\grpso{3}$ Clebsch-Gordan coefficients $\smallcg{J_1}{M_1}{J_2}{M_2}{J}{M_1+M_2}$ by use of the relations (\[eqn-jpm-relations\]). Dots indicate allowed non-zero coefficients. Coefficients at the vertices of a dashed triangle are connected by (\[eqn-jpm-relations\]). (a) The conventional recurrence approach, in which coefficients are calculated inductively from a seed coefficient, making use of relations which in some cases invoke known-zero Clebsch-Gordan coefficients. (b) A full set of relations among allowed Clebsch-Gordan coefficients, yielding a linear, homogeneous system of equations in the Clebsch-Gordan coefficients. []{data-label="fig-so3-recurrence"}](racah_fig01.eps){width="0.9\hsize"} A recurrence approach is indeed natural in the case of $\grpso{3}$. The relations obtained by considering the actions of $J_\pm$ are $$\label{eqn-jpm-relations} K_{\pm}(JM)\smallcg{J_1}{M_1}{J_2}{M_2}{J}{M\pm1} = K_{\pm}(J_1M_1\mp1)\smallcg{J_1}{M_1\mp1}{J_2}{M_2}{J}{M} + K_{\pm}(J_2M_2\mp1)\smallcg{J_1}{M_1}{J_2}{M_2\mp1}{J}{M},$$ where $M=M_1+M_2\mp1$, and $K_{\pm}(JM)\equiv\me{JM\pm1}{J_\pm}{JM}=[(J\mp M)(J\pm M+1)]^{1/2}$ is the generator matrix element. These relations connect at most three coupling coefficients, and a natural order for traversing the coefficients can easily be chosen, such that only one unknown arises at each step, as illustrated in Fig. \[fig-so3-recurrence\](a). \[This is accomplished by involving certain known-zero or “forbidden” Clebsch-Gordan coefficients, represented by the triangle vertices without dots in Fig. \[fig-so3-recurrence\](a), in the relations.\] Since classic treatments (*e.g.*, Ref. ) apply orthonormality relations interspersed with the recurrence relations at intermediate stages of the calculation, we stress that recourse to orthonormality conditions is not actually necessary. It may be seen from the figure that all coefficients are accessible by the relations (\[eqn-jpm-relations\]). In anticipation of the treatment of higher algebras, we also observe that all allowed coefficients may be connected by the relations directly, without involving any forbidden coefficients, as in Fig. \[fig-so3-recurrence\](b). This yields a system of equations which fully determines the coefficients, to within an overall phase and normalization, although in this case the system is not amenable to solution by recursive calculation of successive coefficients from a single seed coefficient. For higher algebras, many irreps of $H$ may be connected by the generator $T^{(\Lambda_T)}$, and therefore each relation obtained from (\[eqn-racah-reduced\]) may involve many unknown coupling coefficients. A simple recurrence pattern, as in Fig. \[fig-so3-recurrence\](a), may be impractical to devise. A more generally applicable and straightforward approach is to directly solve the linear, homogeneous system of equations for the unknown coupling coefficients, by standard linear algebraic methods, [*e.g.*]{}, Euler row reduction.[@fn-nullmethod] Let us therefore summarize the system of equations which must be constructed and solved. The $N$ unknown coupling coefficients for $\Gamma_1\otimes\Gamma_2\rightarrow\rho\Gamma$ may be labeled with a single counting index as $C_i\equiv\smallcg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda}$ with $i=1,\ldots,N$. Each value of the index $i$ therefore designates a specific combination $(a_1\Lambda_1a_2\Lambda_2a\Lambda)$. The numerical coefficients of the unknown quantities $C_i$ in the relations (\[eqn-racah-reduced\]) do [*not*]{} depend upon the outer multiplicity index $\rho$, to be discussed further below. For each generator $T^{(\Lambda_T)}$ in $G$ but not in $H$, and for each quadruplet of irrep labels $(a_1\Lambda_1a_2\Lambda_2\Lambda a'\Lambda')$, the relation (\[eqn-racah-reduced\]) yields an equation (which we label by a counting index $k$) of the form $\sum_{i=1}^N a_{ki}C_i=0$, *i.e.*, linear and homogeneous in the $C_i$. The equation is nonnull (*i.e.*, some of the coefficients $a_{ki}$ are nonvanishing) only if the Clebsch-Gordan series conditions $\Lambda_1\otimes\Lambda_2\rightarrow\Lambda$ and $\Lambda'\otimes\Lambda_T\rightarrow\Lambda$ are met.[@fn-terms] The resulting equations must be aggregated to yield the full system, which may be expressed in matrix form as $$\label{eqn-system-matrix} \begin{matrix} & { \text{Coupling coefficient}\atop \xrightarrow{(a_1\Lambda_1a_2\Lambda_2a\Lambda)} } \\ \rotatebox[origin=c]{90}{ \begin{math} { \text{Racah relation} \atop \xleftarrow{(a_1\Lambda_1a_2\Lambda_2\Lambda a' \Lambda')} } \end{math} } & \underbrace{ \begin{bmatrix} &\vdots &\\ \cdots & a_{ki}&\cdots\\ &\vdots &\\ \end{bmatrix} }_{\equiv A} & \begin{bmatrix} C_1\\\vdots\\C_N \end{bmatrix} = \begin{bmatrix} 0\\ \vdots\\ 0 \end{bmatrix} \end{matrix} .$$ Normally, it suffices to consider the conditions obtained with $(\Lambda_1\Lambda_2\Lambda \Lambda')$ such that $\Gamma_1\rightarrow\Lambda_1$, $\Gamma_2\rightarrow\Lambda_2$, $\Gamma\rightarrow\Lambda$, and $\Gamma\rightarrow\Lambda'$, that is, relations involving only “allowed” coupling coefficients, as in Fig. \[fig-so3-recurrence\](b). However, in certain exceptional cases,[@fn-identitycoupling] additional conditions involving known-zero coupling coefficients may be necessary, analogous to Fig. \[fig-so3-recurrence\](a). These may be obtained, *e.g.*, by considering some $\Lambda'$ with $\Gamma\nrightarrow\Lambda'$. The problem of solving this linear homogeneous system of equations (\[eqn-system-matrix\]) is equivalent to finding the null vector (or vectors) of the matrix $A$ appearing on the left hand side of (\[eqn-system-matrix\]). In general, there may be many more rows (equations) than columns (unknown coupling coefficients). However, these rows are not linearly independent. In the case where $\Gamma_1\otimes\Gamma_2\rightarrow\Gamma$ is free of outer multiplicity, the matrix $A$ can be expected to be of rank $N-1$. The null vector is then uniquely determined, to within normalization and phase, and its entries are the coupling coefficients $[C_1\,C_2\,\cdots\,C_N]$. The proper normalization, yielding coefficients satisfying the condition (\[eqn-ortho-bra-sum\]), is obtained by evaluating $$\label{eqn-ortho-N} \scrN^2\equiv\sum_{\substack{i\\(\text{same~}a\Lambda)}} C_i^2.$$ That is, the summation runs over the subset of entries $C_i$ sharing the same value for $a\Lambda$. An identical result for $\scrN$ must be obtained, regardless of the choice of $a\Lambda$, provided $\Gamma\rightarrow a\Lambda$. (In fact, the requirement of equality may be used as an internal consistency check on the calculation.) The normalized coupling coefficients are then obtained by dividing the null vector by $\scrN$. An overall sign remains to be chosen for the entire set of coupling coefficients for $\Gamma_1\otimes\Gamma_2\rightarrow\Gamma$. For instance, Refs.  suggest a“generalized Condon-Shortley phase convention”, such that $\smallcg{\Gamma_1}{\Lambda_{1m}}{\Gamma_2}{a_2\tilde{\Lambda}_{2}}{\Gamma}{\Lambda_{m}}>0$, that is, a positive value is adopted for the coupling coefficient involving the highest weight irreps of $H$ contained in $\Gamma_1$ and $\Gamma$ and the highest weight irrep $\tilde{\Lambda}_{2}$ consistent with these. More generally, when the coupling $\Gamma_1\otimes\Gamma_2\rightarrow\Gamma$ has outer multiplicity $D$ ($\rho=1,\ldots,D$), the matrix $A$ may be expected to be of rank $N-D$. That is, the system of equations given by (\[eqn-system-matrix\]) yields $D$ linearly independent null vectors (or its null space has dimension $D$). The null vectors obtained by Euler row reduction must be orthonormalized,[@fn-inner] *e.g.*, by the Gram-Schmidt procedure, to yield a set of coupling coefficients satisfying the orthonormality relation (\[eqn-ortho-bra-sum\]). Note that the appropriate inner product for this orthonormalization is [*not*]{} the standard vector dot product on $\bbR^N$. Rather, if we label the entries of each null vector ${\boldsymbol{\mathbf{\mathrm{C}}}}_{\rho}$ as $[C_{\rho1}\,C_{\rho2}\,\cdots\,C_{\rho N}]$, then the inner product to be used for orthonormalization is $$\label{eqn-ortho-M} \scrM_{\rho'\rho}\equiv\sum_{\substack{i\\(\text{same~}a\Lambda)}} C_{\rho' i}C_{\rho i}.$$ The same value of $\scrM_{\rho'\rho}$ is obtained regardless of the choice of $a\Lambda$ used in evaluating the sum. (Again, requirement of this equality provides an internal consistency check on the calculation.) The orthonormal coupling coefficients are then simply the entries of the orthonormalized null vectors. When an outer multiplicity is present, it should be noted that the coupling coefficients are defined only to within a unitary transformation, arising from the arbitrariness inherent in defining the resolution of the outer multiplicity, *i.e.*, in choosing the basis states $\ket{\rho\Gamma\cdots}$ ($\rho=1,\ldots,D$) spanning the $D$-dimensional space of irreps of type $\Gamma$. In the present calculational procedure, the freedom in resolution of the multiplicity is manifested in the freedom to choose different sets of orthogonal basis vectors for the null space of $A$. Coupling coefficients for $\grpso{5}$ in the canonical basis {#sec-so5} ============================================================ Overview {#example-overview} -------- For a concrete example of the application of Racah’s method in terms of reduced coupling coefficients, as developed in Sec. \[sec-method\], we consider the calculation of coupling coefficients for $\grpso{5}$, reduced with respect to the canonical subalgebra $\grpso{4}$. That is, we have $\grpso{5}\supset\grpso{4}$ as the algebras $G\supset H$. Both essential criteria for application of the method are met: (1) the coupling and recoupling coefficients (Wigner calculus) for $\grpso{4}$ are known,[@biedenharn1961:so4-wigner] and (2) the reduced matrix elements of the $\grpso{5}$ generators, considered as tensor operators with respect to $\grpso{4}$, are also known.[@hecht1965:so5-wigner; @kemmer1968:so5-irreps-1] The algebra $\grpso{5}$ contains several subalgebra chains, involving distinct $\grpso{3}\sim\grpsu{2}$ subalgebras, $$\label{eqn-so5-chains} \begin{aligned} \subqn{\grpso{5}}{[l_1l_2]} &\supset\subqn{\grpso{4}}{[pq]}\supset\subqn{\grpso[J]{3}}{J}\supset\subqn{\grpso[J]{2}}{M_J} &\quad&(\chain{I}) \\ &\supset\subqn{\grpso{4}}{}\sim\subqn{\grpso[X]{3}}{X}\otimes\subqn{\grpso[Y]{3}}{Y}\supset\subqn{\grpso[X]{2}}{M_X}\otimes\subqn{\grpso[Y]{2}}{M_Y} &&(\chain{I'}) \\ &\subqnsupset{\kappa}\subqn{\grpu[N]{1}}{M_S}\otimes\subqn{\grpso[T]{3}}{T}\supset\subqn{\grpso[T]{2}}{M_T} &&(\chain{II}) \\ &\subqnsupset{\alpha}\subqn{\grpso[L]{3}}{L}\supset\subqn{\grpso[L]{2}}{M_L} &&(\chain{III}), \end{aligned}$$ where the irrep label has been noted beneath each subalgebra. Branching multiplicity labels are indicated by $\kappa$ and $\alpha$ in the last two chains. Chain () is the standard canonical chain, while in () the canonical $\grpso{4}$ subalgebra is reexpressed using the isomorphism $\grpso{4}\sim\grpso{3}\otimes\grpso{3}$. \[As far as definition of reduced coupling coefficients is concerned, the two chains () and () are equivalent, but branching rules, coupling coefficients, *etc.*, are simpler when expressed with respect to the latter chain ().\] The prerequisite definitions and algebraic results are summarized in Sec. \[sec-alg\], and the calculation of reduced coupling coefficients for the canonical chain is discussed in Sec. \[sec-canonical\]. Physical applications require the coupling coefficients of $\grpso{5}$ reduced with respect to the noncanonical subalgebras of chains () and (). The isospin algebra $\grpso[T]{3}$ of chain () is the relevant subalgebra for the description of proton-neutron pairing.[@helmers1961:shell-sp; @flowers1964:quasispin; @ichimura1965:seniority-isospin; @ginocchio1965:so5-quasispin; @hecht1967:so5-shell-wigner; @engel1996:so5-isovector-pairing; @sviratcheva2005:isospin-breaking-sp4; @sviratcheva2006:realistic-symmetries] In this context, the $\grpso{5}$ generators arise as quasispin operators for pairing of protons and neutrons occupying the same $j$-shell. On the other hand, the “physical” or “geometric” angular momentum subalgebra $\grpso[L]{3}$ of chain () is the relevant subalgebra for application to systems of spin-$2$ bosons[@arima1976:ibm-u5; @iachello1987:ibm] or the nuclear collective model.[@bohr1998:v2; @eisenberg1987:v1; @rowe2004:spherical-harmonics; @rowe2009:acm] Explicit constructions of these subalgebras and further algebraic properties for the noncanonical chains are detailed in Appendix \[app-transform\], where the transformation between canonical and noncanonical bases is considered. Definitions and algebraic properties {#sec-alg} ------------------------------------ Let us begin with a concise but comprehensive summary of the construction of $\grpso{5}$ and the algebraic properties needed for the application of Racah’s method. Such a review is particularly necessary since notations and conventions for nearly all aspects of the treatment of $\grpso{5}$ vary widely (*e.g.*, Refs. ), and phases and normalizations play an essential role in the calculation of coupling coefficients. The basic construction proceeds from the generators of rotation, $$\label{eqn-L-defn} L_{rs}\equiv -i(x_r\partial_s-x_s\partial_r).$$ These operators are Hermitian ($L_{rs}^\dagger=L_{rs}$), are antisymmetric in the indices, and have commutators $$\label{eqn-L-comm} [L_{pq},L_{rs}]= -i(\delta_{qr}L_{ps}+\delta_{ps}L_{qr}+\delta_{sq}L_{rp}+\delta_{rp}L_{sq}).$$ First, for $\grpso{4}$, let $J_r\equiv\half\varepsilon_{rst}L_{st}$ and $N_r\equiv L_{r4}$ ($1\leq r,s,t\leq3$), *i.e.*, $$\label{eqn-so4-gen} \begin{aligned} J_1&=L_{23}&J_2&=L_{31}&J_3&=L_{12}\\ N_1&=L_{14}&N_2&=L_{24}&N_3&=L_{34}. \end{aligned}$$ Then the $J_r$ span the usual three-dimensional angular momentum algebra, which we denote by $\grpso[J]{3}$. The $J_r$ and $N_r$ together span $\grpso{4}$, with commutators $[J_r,J_s]=i\varepsilon_{rst}J_t$, $[N_r,N_s]=i\varepsilon_{rst}J_t$, and $[J_r,N_s]=i\varepsilon_{rst}N_t$. The standard Cartan weight operators for $\grpso{4}$ are $J_3$ and $N_3$. An $\grpso{4}$ irrep is labeled by the highest weight defined by these operators, which is of the form $[pq]$, with $p\geq\abs{q}$, both integer or both odd half integer. The isomorphism $\grpso{4}\sim\grpso[X]{3}\otimes\grpso[Y]{3}$ is realized by taking $$\label{eqn-so4-iso} X_k\equiv\half(J_k+N_k) \quad Y_k\equiv\half(J_k-N_k),$$ so $[X_r,X_s]=i\varepsilon_{rst}X_t$, $[Y_r,Y_s]=i\varepsilon_{rst}Y_t$, and $[X_r,Y_s]=0$. The ladder operators for each $\grpso{3}$ algebra are thus $X_\pm\equiv X_1\pm i X_2$ and $Y_\pm\equiv Y_1\pm i Y_2$. The natural Cartan weight operators in this scheme are then the $\grpso{3}$ angular momentum projections $X_0\equiv X_3$ and $Y_0\equiv Y_3$, defining weight labels $M_X$ and $M_Y$. An $\grpso{4}$ irrep is then labeled by the highest weight $(XY)$, *i.e.*, the angular momenta associated with the $\grpso[X]{3}$ and $\grpso[Y]{3}$ subalgebras. The $\grpso[X]{3}\otimes\grpso[Y]{3}$ irrep labels are related to the standard $\grpso{4}$ labels by $X=\half(p+q)$ and $Y=\half(p-q)$ or, conversely, $[p,q]=[X+Y,X-Y]$. Note that the canonical $\grpso[J]{3}$ is obtained as the sum angular momentum algebra of $\grpso[X]{3}$ and $\grpso[Y]{3}$, since $J_k=X_k+Y_k$, and thus the basis states reducing chains () and () are related to each other by ordinary angular momentum coupling. The Clebsch-Gordan series and coupling and recoupling coefficients for $\grpso{4}$ follow immediately from the $\grpso{3}\otimes\grpso{3}$ structure,[@biedenharn1961:so4-wigner] most transparently with the $(XY)$ labeling scheme for the irreps. The weights contained within $(XY)$ are $M_X=-X,\ldots,X-1,X$ and $M_Y=-Y,\ldots,Y-1,Y$. The Clebsch-Gordan series is given by application of the triangle inequality separately to each of the $\grpso{3}$ algebras, that is, for $(X_1Y_1)\otimes(X_2Y_2)\rightarrow(XY)$, $X=\abs{X_1-X_2},\abs{X_1-X_2}+1,\ldots,X_1+X_2$ and $Y=\abs{Y_1-Y_2},\abs{Y_1-Y_2}+1,\ldots,Y_1+Y_2$. Hence, no inner or outer multiplicities are obtained for $\grpso{4}$. Coupling coefficients factorize into products of ordinary $\grpso{3}$ Clebsch-Gordan coefficients, as $$\label{eqn-so4-coupling} \cg{(X_1Y_1)}{M_{X1}M_{Y1}}{(X_2Y_2)}{M_{X2}M_{Y2}}{(XY)}{M_{X}M_{Y}} = \cg{X_1}{M_{X1}}{X_2}{M_{X2}}{X}{M_{X}} \cg{Y_1}{M_{Y1}}{Y_2}{M_{Y2}}{Y}{M_{Y}}.$$ By inspection of (\[eqn-usixj\]), it is immediately apparent that the recoupling coefficients factorize as well, as $$\begin{gathered} \label{eqn-so4-recoupling} \usixj{(X_1Y_1)}{(X_2Y_2)}{(X_{12}Y_{12})}{(X_3Y_3)}{(XY)}{(X_{23}Y_{23})} = \usixj{X_1}{X_2}{X_{12}}{X_3}{X}{X_{23}} \usixj{Y_1}{Y_2}{Y_{12}}{Y_3}{Y}{Y_{23}}\\ = (-)^{X_1+X_2+X_3+X} (-)^{Y_1+Y_2+Y_3+Y} \hat{X}_{12}\hat{X}_{23} \hat{Y}_{12}\hat{Y}_{23} \sixj{X_1}{X_2}{X_{12}}{X_3}{X}{X_{23}} \sixj{Y_1}{Y_2}{Y_{12}}{Y_3}{Y}{Y_{23}}.\end{gathered}$$ An equivalent result is given with standard $\grpso{4}$ labels in Ref. . However, note that the result is considerably more cumbersome to derive if one uses the standard canonical chain ().[@ding2003:so4-recoupling-tetratomic] The algebra $\grpso{5}$ includes the additional four generators $L_{r5}$ ($r=1,\ldots,4$). A tensor operator with respect to $\grpso{4}$ is simply a simultaneous spherical tensor with respect to both the $\grpso[X]{3}$ and $\grpso[Y]{3}$ algebras, *i.e.*, a spherical “bitensor”. For the $\grpso{5}$ generators, we have bitensor expressions[@fn-so5gen] $$\label{eqn-so5-gen} \begin{aligned} X^{(10)}_{\pm10}&\equiv X_{\pm1}=\mp\tfrac1{2\sqrt{2}}[(L_{23}+L_{14})\pm i(L_{31}+L_{24})] & X^{(10)}_{00}&\equiv X_{0}=\half(L_{12}+L_{34}) \\ Y^{(01)}_{0\pm1}&\equiv Y_{\pm1}=\mp\tfrac1{2\sqrt{2}}[(L_{23}-L_{14})\pm i(L_{31}-L_{24})] & Y^{(01)}_{00}&\equiv Y_{0}=\half(L_{12}-L_{34})\\ \Tfull{+}{+}&\equiv\Tpp=-\half(L_{15}+iL_{25}) & \Tfull{+}{-}&\equiv\Tpm=\half(L_{35}+iL_{45})\\ \Tfull{-}{+}&\equiv\Tmp=\half(L_{35}-iL_{45}) & \Tfull{-}{-}&\equiv\Tmm=\half(L_{15}-iL_{25}). \end{aligned}$$ The phases are chosen so that these operators obey $A^{(XY)\,\dagger}_{M_XM_Y}=(-)^{M_X+M_Y} A^{(XY)}_{-M_X-M_Y}$, a generalization of the usual condition for a self-adjoint spherical tensor.[@edmonds1960:am] All commutators involving $X_\mu$ or $Y_\mu$ have the values implied by the spherical bitensor notation of (\[eqn-so5-gen\]), *e.g.*, $[X_{\pm1},A^{(\lambda\lambda')}_{\mu\mu'}]=\mp[\half(\lambda\mp\mu)(\lambda\pm\mu+1)]^{1/2}A^{(\lambda\lambda')}_{(\mu\pm1)\mu'}$. The commutators between components of $\Tss$ are given explicitly in Table \[tab-so5-comm\]. ![ Root vector diagram for $\grpso{5}$ and its subalgebras. (a) The generators of $\grpso{5}$, labeled by their Cartan weights $M_X$ and $M_Y$. (b) The canonical subalgebra $\grpso{4}\sim\grpso[X]{3}\otimes\grpso[Y]{3}$, which begins chains () and (). (c) The $\grpu[N]{1}\otimes\grpso[T]{3}$ subalgebra of chain (). (d) The physical angular momentum $\grpso[L]{3}$ subalgebra of chain (). These generators are obtained as linear combinations of the canonical generators, as indicated by the dashed lines, such that all generators have good $\grpso[L]{3}\supset\grpso[L]{2}$ tensorial character. []{data-label="fig-root-chains"}](racah_fig02.eps){width="0.8\hsize"} The root vector diagram of $\grpso{5}$ is shown for reference in Fig. \[fig-root-chains\](a), with the generators (\[eqn-so5-gen\]) placed according to their $\grpso[X]{3}\otimes\grpso[Y]{3}$ weights $(M_XM_Y)$. The canonical subalgebra is highlighted in Fig. \[fig-root-chains\](b), and the construction of the physical subalgebras, of chains () and (), is indicated in Fig. \[fig-root-chains\](c,d) (see Appendix \[app-transform\]). The standard Cartan highest weight labels for an $\grpso{5}$ irrep are defined with respect to weight operators $J_3$ and $N_3$ and have the form $[l_1l_2]$, with $l_1\geq l_2$, both integer or both odd half integer. It is more convenient in the present context to label $\grpso{5}$ irreps by the highest weight defined by the $\grpso[X]{3}\otimes\grpso[Y]{3}$ weight operators $X_0$ and $Y_0$ (following Hecht[@hecht1965:so5-wigner]). The resulting label has the form $(RS)$, with $R\geq S$, each independently either integer or odd half integer. This label may also be considered as representing the angular momenta $(X_mY_m)$ of the highest weight $\grpso[X]{3}\otimes\grpso[Y]{3}$ irrep contained in the $\grpso{5}$ irrep. The relation to the standard labels is $R=\half(l_1+l_2)$ and $S=\half(l_1-l_2)$. A plethora of labeling schemes for $\grpso{5}$ irreps are in use in the physics literature, interrelated as summarized in Table \[tab-so5-label\] (even more schemes arise if we consider the translation to physical labels, such as reduced isospin[@flowers1952:jj-coupling-part1]). The branching rule for $\grpso{5}$ to $\grpso{4}$, *i.e.*, $(RS)\rightarrow(XY)$, is given by[@hecht1965:so5-wigner; @kemmer1968:so5-irreps-1] $$\label{eqn-so5-so4-branch} \begin{aligned} X&=R-\half n-\half m\\ Y&=S+\half n-\half m, \end{aligned}$$ with $0\leq n\leq2(R-S)$ and $0\leq m\leq2S$, $m$ and $n$ integers. Graphically, the $\grpso{4}$ irreps form a lattice bounded by a tilted rectangle, as illustrated by the dotted line in Fig. \[fig-weights\](c). The rectangle’s “right” corner is at the highest weight $(RS)=(X_mY_m)$, its “bottom” corner lies on the $M_X$ axis, and the remaining two corners are specified by symmetry about the line $M_X=M_Y$. The branching rule is shown for example irreps of $\grpso{5}$ in Fig. \[fig-weights\]: symmetric \[Fig. \[fig-weights\](a)\], antisymmetric \[Fig. \[fig-weights\](b)\], and generic \[Fig. \[fig-weights\](c)\]. ![ Branching diagrams for $\grpso{5}\supset\grpso{4}$, shown for (a) the symmetric irrep $(22)$, (b) the antisymmetric irrep $(20)$, and (c) a representative generic irrep $(\tfrac32\half)$. The open circle indicates the highest weight for the $\grpso{5}$ irrep, and the solid dots indicate $\grpso{4}$ highest weights, according to branching rule (\[eqn-so5-so4-branch\]). The shaded rectangles are the boundaries of the weight sets for these $\grpso{4}$ irreps. The dashed rectangle in panel (c) is the boundary of the $\grpso{4}$ highest weight set, as discussed in the text.[]{data-label="fig-weights"}](racah_fig03.eps){width="\hsize"} The Clebsch-Gordan series for $\grpso{5}$ may be obtained by a relatively efficient elementary approach based on the method of weights, but “reduced” with respect to $\grpso{4}$. Since the $\grpso{5}\rightarrow\grpso{4}$ branching rules (\[eqn-so5-so4-branch\]) are known, as is the $\grpso{4}$ Clebsch-Gordan series, the tabulation of weights can be replaced by tabulation of $\grpso{4}$ irrep labels, which then imply all the weights contained within these irreps. To decompose the $\grpso{5}$ outer product $(R_1S_1)\otimes(R_2S_2)$, first the $\grpso{4}$ irreps in the branchings $(R_1S_1)\rightarrow(X_1Y_1)$ and $(R_2S_2)\rightarrow(X_2Y_2)$ are enumerated. Then, for each pair of $\grpso{4}$ irreps $(X_1Y_1)$ and $(X_2Y_2)$, the product irreps $(X_1Y_1)\otimes(X_2Y_2)\rightarrow(XY)$ are enumerated. The aggregate set of these product irreps represents the $\grpso{4}$ content of $(R_1S_1)\otimes(R_2S_2)$. The $\grpso{5}$ content can now be extracted. Namely, the highest weight $\grpso{4}$ label in the set gives the highest weight $\grpso{5}$ irrep contained in $(R_1S_1)\otimes(R_2S_2)$ \[which will, incidentally, simply be the sum of $(R_1S_1)$ and $(R_2S_2)$ as weights\]. The $\grpso{4}$ content of this $\grpso{5}$ irrep can now be deleted from the set, after which the next highest remaining $\grpso{4}$ label gives the next highest weight $\grpso{5}$ irrep in $(R_1S_1)\otimes(R_2S_2)$, [*etc.*]{} The process is repeated until the set of $\grpso{4}$ irreps has been exhausted. The Clebsch-Gordan series for $\grpso{5}$ may also be obtained by group character methods (see Refs. ). The remaining ingredients needed for application of Racah’s method are the $\grpso{4}$-reduced matrix elements of the “additional” generators of $\grpso{5}$ not contained in $\grpso{4}$, *i.e.*, $\Tss$. These were obtained in closed form, by solving certain recurrence relations obtained from the commutators of the algebra, by Hecht[@hecht1965:so5-wigner] and by Kemmer, Pursey, and Williams,[@kemmer1968:so5-irreps-1] as $$\label{eqn-T-rme} \begin{aligned} \rme[2]{(RS)\\(X+\half\,Y+\half)}{\Thh}{(RS)\\(XY)}&= \frac{ \begin{multlined} \bigl[ (R+S-X-Y)(R+S+X+Y+3) \\[-1ex] \times(-R+S+X+Y+1)(R-S+X+Y+2) \bigr]^{1/2} \end{multlined} } {2\widehat{X+\half}\widehat{Y+\half}} \\ \rme[2]{(RS)\\(X+\half\,Y-\half)}{\Thh}{(RS)\\(XY)}&= \frac{ \begin{multlined} \bigl[ (R+S-X+Y+1)(R+S+X-Y+2) \\[-1ex] \times(R-S-X+Y)(R-S+X-Y+1) \bigr]^{1/2} \end{multlined} } {2\widehat{X+\half}\widehat{Y-\half}}. \end{aligned}$$ These expressions are appropriate to the Wigner-Eckart theorem normalization defined in (\[eqn-we\]) and the normalization of $\Tss$ defined by (\[eqn-so5-gen\]). The remaining matrix elements, connecting $(XY)$ with $(X-\half\,Y+\half)$ or $(X-\half\,Y-\half)$, follow from these by the self-adjoint property of the generators,[@kemmer1968:so5-irreps-1] as $$\rme[2]{(RS)\\(XY)}{\Thh}{(RS)\\(X'Y')} = \frac{\hat{X}'\hat{Y}'}{\hat{X}\hat{Y}} (-)^{X-X'+Y-Y'} \rme[2]{(RS)\\(X'Y')}{\Thh}{(RS)\\(XY)}.$$ Calculation of $\grpso{5}\supset\grpso{4}$ coupling coefficients {#sec-canonical} ---------------------------------------------------------------- Consider calculation of the set of $\grpso{5}\supset\grpso{4}$ reduced coupling coefficients for $(R_1S_1)\otimes(R_2S_2)\rightarrow(RS)$. It is now straightforward to construct the terms appearing in the Racah condition (\[eqn-racah-reduced\]), by use of the chain () branching rules, $\grpso{4}$ Clebsch-Gordan series, $\grpso{4}$ Wigner calculus, and $\grpso{4}$-reduced matrix elements of $\Tss$, compiled in the preceding section. For $\grpso{5}\supset\grpso{4}$, the relation (\[eqn-racah-reduced\]) may be written $$\begin{gathered} \label{eqn-racah-reduced-so5} \rme[2]{(RS)\\(XY)}{\Thh}{(RS)\\(X'Y')} \cg{(R_1S_1)}{(X_1Y_1)}{(R_2S_2)}{(X_2Y_2)}{(RS)}{(XY)} \\ \begin{aligned} = &\sum_{(X_1'Y_1')}\Phi[(X_1Y_1)(X_2Y_2);(XY)]\Phi[(X_1'Y_1')(X_2Y_2);(X'Y')] \\ &\quad \times \usixj{(X_2Y_2)}{(X_1'Y_1')}{(X'Y')}{(\half\half)}{(XY)}{(X_1Y_1)} \rme[2]{(R_1S_1)\\(X_1Y_1)}{\Thh}{(R_1S_1)\\(X_1'Y_1')} \cg{(R_1S_1)}{(X_1'Y_1')}{(R_2S_2)}{(X_2Y_2)}{(RS)}{(X'Y')} \\ +& \sum_{(X_2'Y_2')} \usixj{(X_1Y_1)}{(X_2'Y_2')}{(X'Y')}{(\half\half)}{(XY)}{(X_2Y_2)} \rme[2]{(R_2S_2)\\(X_2Y_2)}{\Thh}{(R_2S_2)\\(X_2'Y_2')} \cg{(R_1S_1)}{(X_1Y_1)}{(R_2S_2)}{(X_2'Y_2')}{(RS)}{(X'Y')}. \end{aligned}\end{gathered}$$ The system of equations for the coupling coefficients is assembled following the approach of Sec. \[sec-solution\]. A different condition is obtained from (\[eqn-racah-reduced-so5\]) for each quadruplet of $\grpso{4}$ irreps $(X_1Y_1)$, $(X_2Y_2)$, $(XY)$, and $(X'Y')$, chosen from the branchings $(R_1S_1)\rightarrow(X_1Y_1)$, $(R_2S_2)\rightarrow(X_2Y_2)$, $(RS)\rightarrow(XY)$, and $(RS)\rightarrow(X'Y')$ \[for couplings involving the identity irrep $(00)$, see endnote \]. A nonnull condition on the coupling coefficients is obtained only if $(X_1Y_1)\otimes(X_2Y_2)\rightarrow(XY)$ and $(X'Y')\otimes(\half\half)\rightarrow(XY)$. Since all quantities involved in the conditions (\[eqn-racah-reduced-so5\]) are known exactly, in the form of square roots of rational numbers, and since Euler row reduction can be carried out in exact (*i.e.*, symbolic) arithmetic, all coupling coefficients can be obtained exactly through the present process, again as (signed) square roots of rational numbers. Two concrete numerical examples are provided as illustrations of the method in Appendix \[app-canonical\]. A simple low-dimensional example is provided by the coupling $(\half\half)\otimes(\half0)\rightarrow(\half0)$, which involves only a $4\times4$ coefficient matrix, and an example involving an outer multiplicity is provided by the coupling $(10)\otimes(1\half)\rightarrow(1\half)$, for which the coefficient matrix has dimensions $36\times18$. (In the canonical labeling scheme, these examples are $[10]\otimes[\half\half]\rightarrow[\half\half]$ and $[11]\otimes[\tfrac32\half]\rightarrow[\tfrac32\half]$, respectively.) Noncanonical chains {#sec-noncanonical} ------------------- In the previous section, it was seen how the coupling coefficients for $\grpso{5}$ reduced with respect to the [*canonical*]{} chain may be evaluated by Racah’s method, as formulated in Sec. \[sec-method\]. The necessary ingredients take on a particularly simple form for the canonical chain, in that the reduced matrix elements of the generators are given by closed form expressions (\[eqn-T-rme\]). However, the matrix elements of the generators reduced with respect to the noncanonical isospin subalgebra \[chain ()\] and physical angular momentum algebra \[chain ()\] are also known. Recurrence relations for the reduced matrix elements have been obtained from vector coherent state realizations,[@hecht1982:sp4-vcs; @turner2006:so5-so3-vcs] or an elementary construction has also been demonstrated for chain ().[@han1993:so5-coupling] Hence, Racah’s method may be applied directly to the calculation of chain () and chain () reduced coupling coefficients. Alternatively, once coupling coefficients for the canonical chain have been obtained, the coupling coefficients for the noncanonical chains can readily be deduced by a unitary transformation. For instance, the coupling coefficients of $\grpsu{3}$ reduced with respect to its physical angular momentum subalgebra are conventionally obtained from the canonical $\grpsu{3}\supset\grpu{1}\otimes\grpsu{2}$ coupling coefficients[@draayer1973:su3-cg; @rowe2000:su3-cg] through such a process. The transformation brackets between basis states reducing the canonical and noncanonical chains are only known in closed form for a few special cases.[@hecht1967:so5-shell-wigner] However, they can be obtained in a straightforward fashion, either (1) by diagonalizing the appropriate Casimir operator, [*i.e.*]{}, ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$ or ${\boldsymbol{\mathbf{\mathrm{L}}}}^2$, in the canonical basis, as in Ref. , or (2) by a combination of laddering and orthogonalization operations. The procedure for transformation of coupling coefficients to either of the noncanonical chains is discussed in Appendix \[app-transform\]. Conclusion {#sec-concl} ========== It has been shown that Racah’s method of infinitesimal generators can be systematically generalized to the calculation of reduced coupling coefficients for an arbitrary subalgebra chain, provided the matrix elements of the generators (reduced with respect to the lower algebra) and the recoupling coefficients of the lower algebra are known. For the algebra $\grpso{5}$, the problem of calculating coupling coefficients for generic irreps, reduced with respect to the canonical or noncanonical chains, is thereby completely resolved. The specific example of $\grpso{5}$ coupling coefficients may be considered as a prototype for the systematic calculation of coupling coefficients for other higher algebras. For instance, the computational machinery for $\grpsu{3}$ is well established[@draayer1973:su3-cg; @rowe2000:su3-cg] and may therefore be used as the starting point for calculation of $\grpsp{6}\supset\grpu{3}$ reduced coupling coefficients, for the fermion dynamical symmetry model,[@wu1986:fdsm] or $\grpsp{6,\bbR}\supset\grpu{3}$ reduced coupling coefficients, for the symplectic shell model.[@rosensteel1980:sp6r-shell] The requisite generator matrix elements for $\grpsp{6}$ and $\grpsp{6,\bbR}$ may be calculated from vector coherent state realizations.[@rowe1984:sp6r-vcs; @hecht1990:sp6-u3-vcs] The $\grpsp{6,\bbR}\supset\grpu{3}$ coupling coefficients are required, for instance, if large-scale calculations are to be carried out in the [*ab initio*]{} symplectic scheme of Dytrych [*et al.*]{}[@dytrych2007:sp-ncsm-evidence; @dytrych2008:sp-ncsm] Discussions with C. Bahri, T. Dytrych, J. P. Draayer, S. De Baerdemacker, D. J. Rowe, and F. Iachello are gratefully acknowledged. This work was supported by the US DOE (under grant DE-FG02-95ER-40934), the US NSF (under grants NSF-PHY-0500291, NSF-OCI-0904874, and NSF-PHY05-52843), and the Southeastern Universities Reasearch Association (SURA). General relations in the presence of outer multiplicities for $H$ {#app-mult} ================================================================= The derivation of Racah’s method in terms of reduced quantities, as given in Sec. \[sec-method\], can readily be generalized to the case in which the subalgebra $H$ has outer multiplicities, [*i.e.*]{}, its Kronecker product is not simply reducible. In this appendix, the necessary generalizations of the algebraic relations ([*e.g.*]{}, Ref. ) entering into the derivation of Sec. \[sec-method\] are summarized, and the fundamental relation (\[eqn-racah-reduced\]) for Racah’s method is extended to incorporate outer multiplicities of $H$. The general form of Racah’s factorization lemma for $G\supset H$ is $$\label{eqn-factorization-mult} \ccg{\Gamma_1}{a_1\Lambda_1}{\lambda_1}{\Gamma_2}{a_2\Lambda_2}{\lambda_2}{\rho\Gamma}{a\Lambda}{\lambda} = \sum_\sigma \cg{\Lambda_1}{\lambda_1}{\Lambda_2}{\lambda_2}{\sigma\Lambda}{\lambda} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda}_\sigma,$$ where $\sigma$ is the multiplicity index for the coupling $\Lambda_1\otimes\Lambda_2\rightarrow\Lambda$. The reduced coupling coefficients satisfy orthonormality relations $$\begin{aligned} \label{eqn-ortho-bra-sum-mult} \sum_{\substack{a_1\Lambda_1a_2\Lambda_2\\\sigma}} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda}_\sigma \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho'\Gamma'}{a'\Lambda}_\sigma &=\delta_{(\rho\Gamma)(\rho'\Gamma')}\delta_{aa'} \intertext{and} \label{eqn-ortho-ket-sum-mult} \sum_{\rho\Gamma a} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda}_{\sigma} \cg{\Gamma_1}{a_1'\Lambda_1'}{\Gamma_2}{a_2'\Lambda_2'}{\rho\Gamma}{a\Lambda}_{\sigma'} &=\delta_{(a_1\Lambda_1)(a_1'\Lambda_1')}\delta_{(a_2\Lambda_2)(a_2'\Lambda_2')} \delta_{\sigma\sigma'},\end{aligned}$$ for any irrep $\Lambda$ such that $\Gamma\rightarrow\Lambda$ and, in the second relation, any values of the multiplicity indices $\sigma$ and $\sigma'$, which resolve $\Lambda_1\otimes\Lambda_2\rightarrow\Lambda$ and $\Lambda_1'\otimes\Lambda_2'\rightarrow\Lambda$, respectively. For the subalgebra $H$, the Wigner-Eckart theorem is of the form $$\label{eqn-we-mult} \me[3]{\Gamma'\\a'\Lambda'\\\lambda'}{T^{\Lambda_T}_{\lambda_T}}{\Gamma\\a\Lambda\\\lambda} = \sum_\sigma \cg{\Lambda}{\lambda}{\Lambda_T}{\lambda_T}{\sigma\Lambda'}{\lambda'} \rme[2]{\Gamma'\\a'\Lambda'}{T^{\Lambda_T}}{\Gamma\\a\Lambda}_\sigma.$$ The recoupling coefficients (unitary $6$-$\Lambda$ symbols) of $H$ are $$\label{eqn-usixj-mult} \usixj{\Lambda_1}{\Lambda_2}{\sigma_{12}\Lambda_{12}}{\Lambda_3}{\Lambda}{\sigma_{23}\Lambda_{23}}_{\sigma\sigma'} = \sum_{\substack{\lambda_1\lambda_2\lambda_3\\\lambda_{12}\lambda_{13}}} \cg{\Lambda_1}{\lambda_1}{\Lambda_2}{\lambda_2}{\sigma_{12}\Lambda_{12}}{\lambda_{12}} \cg{\Lambda_{12}}{\lambda_{12}}{\Lambda_3}{\lambda_3}{\sigma\Lambda}{\lambda} \cg{\Lambda_2}{\lambda_2}{\Lambda_3}{\lambda_3}{\sigma_{23}\Lambda_{23}}{\lambda_{23}} \cg{\Lambda_1}{\lambda_1}{\Lambda_{23}}{\lambda_{23}}{\sigma'\Lambda}{\lambda}.$$ This represents the transformation bracket between basis states in the coupling schemes $[(\Lambda_1\Lambda_2)^{\sigma_{12}\Lambda_{12}}\Lambda_3]^{\sigma\Lambda}$ and $[\Lambda_1(\Lambda_2\Lambda_3)^{\sigma_{23}\Lambda_{23}}]^{\sigma'\Lambda}$. Under interchange of the first and second irreps, the coupling coefficients may be expected to satisfy a symmetry relation of the form \[[*e.g.*]{}, for $\grpsu{3}$, see Refs. \] $$\label{eqn-cg-symm-mult} \cg{\Lambda_2}{\lambda_2}{\Lambda_1}{\lambda_1}{\sigma\Lambda}{\lambda} =\sum_{\sigma'}\Phi_{\sigma\sigma'}(\Lambda_2\Lambda_1;\Lambda) \cg{\Lambda_1}{\lambda_1}{\Lambda_2}{\lambda_2}{\sigma'\Lambda}{\lambda}.$$ Following the same arguments as in Sec. \[sec-method\], the reduced coupling coefficients for $G\supset H$ are found to satisfy a homogeneous system of equations of the form $$\begin{gathered} \label{eqn-racah-reduced-mult} \sum_{a} \rme[2]{\Gamma\\a\Lambda}{T^{\Lambda_T}}{\Gamma\\a'\Lambda'}_{\tau} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a\Lambda}_{\sigma} \\= \begin{lgathered}[t] \sum_{\substack{a_1'\Lambda_1'\\\tau_1\sigma'\delta\delta'}} \Phi_{\sigma\delta}(\Lambda_1\Lambda_2;\Lambda)\Phi_{\sigma'\delta'}(\Lambda_1'\Lambda_2;\Lambda') \\\quad\times \usixj{\Lambda_2}{\Lambda_1'}{\delta'\Lambda'}{\Lambda_T}{\Lambda}{\tau_1\Lambda_1}_{\tau\delta} \rme[2]{\Gamma_1\\a_1\Lambda_1}{T^{\Lambda_T}}{\Gamma_1\\a_1'\Lambda_1'}_{\tau_1} \cg{\Gamma_1}{a_1'\Lambda_1'}{\Gamma_2}{a_2\Lambda_2}{\rho\Gamma}{a'\Lambda'}_{\sigma'} \\+ \sum_{\substack{a_2'\Lambda_2'\\\tau_2\sigma'}} \usixj{\Lambda_1}{\Lambda_2'}{\sigma'\Lambda'}{\Lambda_T}{\Lambda}{\tau_2\Lambda_2}_{\tau\sigma} \rme[2]{\Gamma_2\\a_2\Lambda_2}{T^{\Lambda_T}}{\Gamma_2\\a_2'\Lambda_2'}_{\tau_2} \cg{\Gamma_1}{a_1\Lambda_1}{\Gamma_2}{a_2'\Lambda_2'}{\rho\Gamma}{a'\Lambda'}_{\sigma'}. \end{lgathered}\end{gathered}$$ A different equation is obtained for each choice of $(a_1\Lambda_1a_2\Lambda_2\Lambda a'\Lambda')$ and indices $\sigma$ and $\tau$, where $\sigma$ resolves $\Lambda_1\otimes\Lambda_2\rightarrow\Lambda$ and $\tau$ resolves $\Lambda'\otimes\Lambda_T\rightarrow\Lambda$. Numerical examples for the calculation of $\grpso{5}\supset\grpso{4}$ reduced coupling coefficients {#app-canonical} =================================================================================================== For a simple numerical example of the calculation of $\grpso{5}\supset\grpso{4}$ reduced coupling coefficients, according to the methods of Sec. \[sec-canonical\], consider the coupling $(\half\half)\otimes(\half0)\rightarrow(\half0)$. The coupling coefficients to be obtained are $\smallcg{(\shalf\,\shalf)}{(0\,0)}{(\shalf\,0)}{(0\,\shalf)}{(\shalf\,0)}{(0\,\shalf)}$, $\smallcg{(\shalf\,\shalf)}{(\shalf\,\shalf)}{(\shalf\,0)}{(\shalf\,0)}{(\shalf\,0)}{(0\,\shalf)}$, $\smallcg{(\shalf\,\shalf)}{(0\,0)}{(\shalf\,0)}{(\shalf\,0)}{(\shalf\,0)}{(\shalf\,0)}$, and $\smallcg{(\shalf\,\shalf)}{(\shalf\,\shalf)}{(\shalf\,0)}{(0\,\shalf)}{(\shalf\,0)}{(\shalf\,0)}$. The system of four equations in four unknowns obtained from (\[eqn-racah-reduced-so5\]) has the coefficient matrix shown in Fig. \[fig-system-small\]. This matrix is of rank $3$, admitting the null vector $$\begin{bmatrix} -\half & -1 & \vline& \half & 1 \end{bmatrix},$$ shown as a row vector for easier comparison with the coupling coefficient labels across the top of Fig. \[fig-system-small\]. It remains to obtain the proper normalization (and phase, if a phase convention is to be enforced). The vertical bar demarcates coupling coefficients involving the same product irrep $(XY)$, $(0\half)$ for the first two coefficients and $(\half0)$ for the remaining two coefficients, that is, the coefficients which appear together in the normalization sum (\[eqn-ortho-N\]). The squared norm, calculated using either pair of coefficients, is $\scrN^2=\tfrac54$. Hence, the normalized coupling coefficients are given by $$\begin{bmatrix} -\sqrt{\tfrac15} & -\sqrt{\tfrac45} &\vline& \sqrt{\tfrac15} & \sqrt{\tfrac45} \end{bmatrix},$$ which incidentally also conforms to the generalized Condon-Shortly phase convention. One of the simplest cases in which an outer multiplicity occurs is in the coupling $(10)\otimes(1\half)\rightarrow(1\half)$, which has multiplicity $D=2$. There are $18$ coupling coefficients for these irreps, related by a system of $36$ equations. The first few rows of the coefficient matrix are shown in Fig. \[fig-system-large\](a). The matrix is of rank $16$, admitting two null vectors, given in Fig. \[fig-system-large\](b). Again, these are shown as row vectors, and groups of coefficients sharing the same $(XY)$ are delimited by vertical bars. The inner product matrix (\[eqn-ortho-M\]), which may be obtained using any of these four groups of coefficients, is $$\scrM=\begin{bmatrix} \tfrac{15}{8} & \sqrt{\tfrac{45}{32}}\\ \sqrt{\tfrac{45}{32}} & \tfrac{31}{4} \end{bmatrix}.$$ The orthonormal null vectors obtained by the Gram-Schmidt procedure with respect to $\scrM$ are then given in Fig. \[fig-system-large\](c). The entries of the upper and lower rows may be taken as the $\grpso{5}\supset\grpso{4}$ reduced coupling coefficients for $\rho=1$ and $2$, respectively. Transformation of coupling coefficients {#app-transform} ======================================= In this appendix, a general algorithm is outlined for determination of the transformation brackets between canonical and $\grpu[N]{1}\otimes\grpso[T]{3}$ \[chain ()\] bases of an $\grpso{5}$ irrep. These are the necessary ingredients for deducing the chain () reduced coupling coefficients from the canonical reduced coupling coefficients. The analogous procedure for transformation to the $\grpso[L]{3}$ basis \[chain ()\] is also briefly considered. First, let us review the relevant properties of the proton-neutron quasispin realization of $\grpso{5}$ and the chain () basis and branching rules. For protons and neutrons occupying a single level of angular momentum $j$ (degeneracy $2j+1$), consider the proton pair quasispin operators ($X_+$, $X_0$, and $X_-$), neutron pair quasispin operators ($Y_+$, $Y_0$, and $Y_-$), proton-neutron pair quasispin operators ($S_+$, $S_0$, and $S_-$), and isospin operators ($T_+$, $T_0$, and $T_-$). Each set spans an angular momentum algebra, which we denote by $\grpso[X]{3}$, $\grpso[Y]{3}$, $\grpso[S]{3}$, or $\grpso[T]{3}$, respectively. Explicitly, $$\begin{aligned} \label{eqn-chainII-gen} X_+&=\tfrac12(\ad_p\cdot\ad_p) & X_0&=-\tfrac14(\ad_p\cdot\at_p+\at_p\cdot\ad_p) & X_-&=-\tfrac12(\at_p\cdot\at_p)\\ Y_+&=\tfrac12(\ad_n\cdot\ad_n) & Y_0&=-\tfrac14(\ad_n\cdot\at_n+\at_n\cdot\ad_n) & Y_-&=-\tfrac12(\at_n\cdot\at_n)\\ S_+&=\tfrac12(\ad_p\cdot\ad_n+\ad_n\cdot\ad_p) & S_0&=-\tfrac12(\ad_p\cdot\at_p+\at_n\cdot\ad_n) & S_-&=-\tfrac12(\at_n\cdot\at_p+\at_p\cdot\at_n)\\ T_+&=-(\ad_p\cdot\at_n) & T_0&=-\tfrac12(\ad_p\cdot\at_p-\ad_n\cdot\at_n) & T_-&=-(\ad_n\cdot\at_p), \end{aligned}$$ where $\at^{(j)}_m\equiv (-)^{j-m}a^{(j)}_{-m}$, and we define $A^{(j)}\cdot B^{(j)} \equiv \jhat (A\times B)^{(0)}_0=\sum_m A_m \tilde{B}_m$ for half-integer $j$.[@fn-sph-dot] Thus, for instance, $X_+=\half\sum_m(-)^{j-m}\ad_{p,m}\ad_{p,-m}$, $X_0=\tfrac14\sum_m(\ad_{p,m}a_{p,m}-a_{p,m}\ad_{p,m})$, and $X_-=\half\sum_m(-)^{j-m}a_{p,-m}a_{p,m}$. There are only ten independent operators, since $S_0=X_0+Y_0$ and $T_0=X_0-Y_0$. The operators defined in (\[eqn-chainII-gen\]) obey the commutation relations of the $\grpso{5}$ generators of Sec. \[sec-alg\], with the identifications $$\label{eqn-sot3-reln} S_+=-2\Tpp \quad S_-=2\Tmm \quad T_+=2\Tpm \quad T_-=2\Tmp,$$ and thus span an $\grpso{5}$ algebra. The generators $T_+$, $T_0$, $T_-$, and $S_0$ span the $\grpu[N]{1}\otimes\grpso[T]{3}$ subalgebra of $\grpso{5}$, as shown in Fig. \[fig-root-chains\](c). Here $\grpu[N]{1}$ is the one-dimensional algebra of $S_0$, so denoted in recognition of the relation of this operator to the total proton-neutron number operator $N=N_p+N_n$. The natural weights for chain (), $M_S$ and $M_T$, are related to the chain () weights by $M_S=M_X+M_Y$ and $M_T=M_X-M_Y$.[@fn-weights-I-II] For the proton-neutron quasispin realization of $\grpso{5}$ defined in (\[eqn-chainII-gen\]), the weights are simply related to proton and neutron occupation numbers by $M_X=\half(N_p-\Omega)$ and $M_Y=\half(N_n-\Omega)$, or $M_S=\half(N-2\Omega)$ and $M_T=\half(N_p-N_n)$, where $\Omega=\half(2j+1)$ is the half-degeneracy. ![Weight diagrams for $\grpso{5}$ irreps, illustrating the decomposition into $\grpu[N]{1}\otimes\grpso[T]{3}$ irreps, [*i.e.*]{}, the chain () branching. The area of each dot indicates the multiplicity of the weight point, and the rectangles indicate grouping of weight points into isospin multiplets. (a) The general characteristics, in the presence of branching multiplicites, are illustrated with the irrep $(\tfrac72\tfrac32)$. The branching is obtained by decomposing the weights at a given value of $M_S$ into multiplets, [*e.g.*]{}, for $M_S=+2$, in this example, $T=1^2$, $2^2$, $3^2$, $4$, and $5$, where the exponents indicate branching multiplicities. The mathematical labeling schemes for the $\grpso{5}$ irrep are based on the highest weight point (circle), but the $(v,t)$ labeling scheme for pairing applications is based on the quantum numbers of the “pair vacuum” $\grpu[N]{1}\otimes\grpso[T]{3}$ irrep (dashed box). The arrows indicate the action of the isospin laddering operators $T_\pm$ considered in the algorithms presented in the text. (b) Decomposition of the low-dimensional symmetric irrep $(11)$. (c) Decomposition for the low-dimensional antisymmetric irrep $(10)$, which is the adjoint (or generator) irrep. (d) Decomposition of the low-dimensional generic irrep $(1\half)$, illustrating isospin multiplets of half-integer isospin. []{data-label="fig-weights-iso"}](racah_fig06.eps){width="0.75\hsize"} The $\grpso{5}$ irrep $(RS)$ may be decomposed into $\grpu[N]{1}\otimes\grpso[T]{3}$ irreps labeled by $M_S$ and $T$, as indicated in (\[eqn-so5-chains\]). By inspection of the weight diagram \[Fig. \[fig-weights-iso\](a)\] and considering the relation between canonical weights (horizontal and vertical axes) and chain () weights (diagonal axes), it can be seen that the labels are taken from the possible values $M_S=-(R+S)$, $\ldots$, $R+S-1$, $R+S$ and $T=0$ or $\half$, $\ldots$, $R+S-1$, $R+S$. For a given value of $M_S$, the highest weight $M_T$ and thus highest isospin is given by $T_\text{max}(RS;M_S)=R+S$ for $\abs{M_S}\leq R-S$, and $T_\text{max}(RS;M_S)=2R-\abs{M_S}$ for $\abs{M_S}>R-S$. Note that one of the $\grpu[N]{1}\otimes\grpso[T]{3}$ irreps, at the bottom left of the weight diagram, is a “pair vacuum”, annihilated by $X_-$, $Y_-$, and $S_-$ \[dashed box in Fig. \[fig-weights-iso\](a)\]. In applications to proton-neutron pairing, $\grpso{5}$ irreps are conventionally labeled not by the usual mathematical labels (Table \[tab-so5-label\]), but rather by the seniority $v$ and reduced isospin $t$.[@flowers1952:jj-coupling-part1] These are the occupation number and isospin of the pair vacuum, [*i.e.*]{}, $M_{S\,\text{vac}}\equiv\half(v-2\Omega)=-(R+S)$ and $T_\text{vac}\equiv t=R-S$. The isospin content may be seen by decomposing the weights $M_T$ for each given value of $M_S$, [*i.e.*]{}, occuring on single diagonal of the weight diagram, into isospin multiplets \[solid boxes in Fig. \[fig-weights-iso\](a)\]. A given pair of $\grpu[N]{1}\otimes\grpso[T]{3}$ labels may occur more than once within an $\grpso{5}$ irrep and is therefore labeled by a multiplicity index $\kappa=1$, $2$, $\ldots$, $\operatorname{mult}(RS;M_ST)$. The algorithm for constructing the branching $\grpso{5}\rightarrow\grpu[N]{1}\otimes\grpso[T]{3}$ is derived in, [*e.g.*]{}, Ref. . Simple counting arguments then give a closed form multiplicity formula $$\label{eqn-chain-II-mult} \operatorname{mult}(RS;M_ST)=\begin{cases} f(2S,T;M_S)&T\leq R-S\\ f(R+S-T,R-S;M_S)&T> R-S, \end{cases}$$ where $f(u,v;w)=\lfloor\min[u,\half(u+v+w)]\rfloor -\lceil\max[0,\half(u-v+w)]\rceil+1$.[@fn-f-count] This relation fully defines the branching rule for chain (). The branching into $\grpu[N]{1}\otimes\grpso[T]{3}$ irreps is depicted in Fig. \[fig-weights-iso\], for example irreps of $\grpso{5}$: symmetric \[Fig. \[fig-weights-iso\](b)\], antisymmetric \[Fig. \[fig-weights-iso\](c)\], and generic \[Fig. \[fig-weights-iso\](d)\]. Basis states for chain () involve only a linear combination of canonical chain () basis states at the [*same*]{} point in weight space. This point may be labeled interchangably either by $M_X$ and $M_Y$ or by $M_S$ and $M_T$. Thus, $$\label{eqn-chainII-bracket} \ket[3]{(RS)\\M_{S} \kappa T\\M_{T}}= \sum_{(XY)} \overlap[3]{(RS)\\(XY)\\M_{X}M_{Y}}{(RS)\\M_{S} \kappa T\\M_{T}} \, \ket[3]{(RS)\\(XY)\\M_{X}M_{Y}},$$ where it is to be understood that $M_X=\tfrac12(M_S+M_T)$ and $M_Y=\tfrac12(M_S-M_T)$. The $\grpso{4}$ irreps $(XY)$ of the canonical basis states contributing to a given chain () basis state are constrained by the branching condition $(RS)\rightarrow(XY)$ and by the usual angular momentum projection rules ($\abs{M_X}\leq X$ and $\abs{M_Y}\leq Y$). The transformation brackets in (\[eqn-chainII-bracket\]) are only known in closed form for a restricted set of cases involving low branching multiplicites.[@hecht1967:so5-shell-wigner] However, the transformation brackets can be systematically calculated for any $\grpso{5}$ irrep, regardless of multiplicity, through two possible procedures, outlined here, involving isospin laddering operations and either matrix diagonalization or orthogonalization. The first method relies upon the construction of chain () basis states as eigenstates of ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$. Suppose ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$ is written as its matrix realization in the () basis. The transformation brackets appearing in (\[eqn-chainII-bracket\]) are the coefficients for decomposition of the eigenstate with respect to the chain () basis. Therefore, they must constitute the entries of an eigenvector of the ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$ matrix, of eigenvalue $T(T+1)$, and may be computed by diagonalization of this matrix. Since each chain () basis vector involves only a single weight point, the diagonalization can be carried out separately on subspaces corresponding to different weight points, in which case the eigenvalue $T(T+1)$ occurs with degeneracy given by $\operatorname{mult}(RS;M_ST)$ from (\[eqn-chain-II-mult\]). Note that the matrix realization of ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$ at weight point $(M_XM_Y)$ is readily obtained from the known matrix elements (\[eqn-T-rme\]) of the $\grpso{5}$ generators. From (\[eqn-sot3-reln\]), ${\boldsymbol{\mathbf{\mathrm{T}}}}^2=(X_0-Y_0)^2+2(\Tpm\Tmp+\Tmp\Tpm)$. This operator can be reexpressed in terms of spherical bitensor coupled products of the generator $\Tss$ as ${\boldsymbol{\mathbf{\mathrm{T}}}}^2=(X_0-Y_0)^2+2[(T\times T)^{(11)}_{00}-(T\times T)^{(00)}_{00}]$. The requisite matrix elements are therefore $$\begin{gathered} \label{eqn-Tsqr-rme} \me[3]{(RS)\\(X'Y')\\M_{X}M_{Y}}{{\boldsymbol{\mathbf{\mathrm{T}}}}^2}{(RS)\\(XY)\\M_{X}M_{Y}} = (M_X-M_Y)^2 +2\cg{(XY)}{M_XM_Y}{(11)}{00}{(X'Y')}{M_XM_Y} \\\times \rme[2]{(RS)\\(X'Y')}{(T\times T)^{(11)}}{(RS)\\(XY)} -2 \rme[2]{(RS)\\(X'Y')}{(T\times T)^{(00)}}{(RS)\\(XY)}.\end{gathered}$$ These may be evaluated using Racah’s reduction formula,[@edmonds1960:am] as naturally extended to spherical bitensors, [*i.e.*]{}, $\grpso{4}$ tensors, giving $$\begin{gathered} \label{eqn-T11-rme} \rme[2]{(RS)\\(X'Y')}{(T\times T)^{(11)}}{(RS)\\(XY)} = \sum_{(X''Y'')} \usixj{(\half\half)}{(\half\half)}{(11)}{(XY)}{(X'Y')}{(X''Y'')} \\\times \rme[2]{(RS)\\(X'Y')}{T}{(RS)\\(X''Y'')} \rme[2]{(RS)\\(X''Y'')}{T}{(RS)\\(XY)}\end{gathered}$$ and $$\label{eqn-T00-rme} \rme[2]{(RS)\\(X'Y')}{(T\times T)^{(00)}}{(RS)\\(XY)} = -\frac12 \delta_{(XY)(X'Y')} \sum_{(X''Y'')} \rme[2]{(RS)\\(XY)}{T}{(RS)\\(X''Y'')}^2.$$ These expressions involve only $\grpso{4}$ coupling coefficients (\[eqn-so4-coupling\]), $\grpso{4}$ recoupling coefficients (\[eqn-so4-recoupling\]), and $\grpso{5}\supset\grpso{4}$ reduced matrix elements (\[eqn-T-rme\]), all of which are readily calculated. However, it is important to note that the transformation brackets must be consistent among weight points within an isospin multiplet, since these are connected by the isospin laddering operators $T_\pm$ \[Fig. \[fig-weights-iso\](a)\]. The chain () basis states of a given $M_S$, $\kappa$, and $T$ but different $M_T$ must be related under laddering by $T_\pm$ with the correct (positive) phase. Furthermore, in the presence of branching multiplicities, the choice of basis vectors (resolution of the multiplicity) must be consistent between weight points, [*i.e.*]{}, laddering should not connect different $\kappa$ values at adjacent weight points. The relation among transformation brackets at adjacent weight points along a diagonal of given $M_S$, obtained by comparing the action of $T_\pm$ on the chain () and chain () states, is $$\begin{gathered} \label{eqn-iso-ladder} [(T\mp M_T)(T\pm M_T+1)]^{1/2}\overlap[3]{(RS)\\(X'Y')\\M_X'M_Y'}{(RS)\\M_S\kappa T\\M_T'} \\ =2\sum_{(XY)} \me[3]{(RS)\\(X'Y')\\M_X'M_Y'}{T_{\pm\mp}}{(RS)\\(XY)\\M_XM_Y} \, \overlap[3]{(RS)\\(XY)\\M_XM_Y}{(RS)\\M_S\kappa T\\M_T},\end{gathered}$$ where $M_X'=M_X\pm\half$, $M_Y'=M_Y\mp\half$, and $M_T'=M_T\pm1$, and where it is understood that $M_X=\tfrac12(M_S+M_T)$, $M_Y=\tfrac12(M_S-M_T)$, $M_X'=\tfrac12(M_S+M_T')$, and $M_Y'=\tfrac12(M_S-M_T')$. At most four $\grpso{4}$ irreps $(XY)$ contribute to the sum, under the triangularity condition $(XY)\otimes(\half\half)\rightarrow(X'Y')$. The matrix element of $T_{\pm\mp}$ can be calculated from (\[eqn-we\]) and (\[eqn-T-rme\]). If ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$ were diagonalized independently at each weight point, this would not guarantee consistency between weight points under laddering. As usual in diagonalization problems: (1) Eigenvectors, even in the absence of degenerate eigenvalues, are only defined to within an arbitrary phase (sign). (2) In the presence of degenerate eigenvalues (here, multiple occurence of the same isospin at given $M_S$), the choice of orthogonal basis vectors for the degenerate eigenspace is arbitrary. Diagonalization of ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$ must therefore be augmented by further conditions. A consistent prescription for the transformation brackets is provided by diagonalizing ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$ only [*once*]{}, for each $M_S$, at the most central weight point on the diagonal \[Fig. \[fig-weights-iso\](a)\], either $M_T=0$ for integer isospin \[Fig. \[fig-weights-iso\](b,c)\] or $M_T=+\half$ for half-integer isospin \[Fig. \[fig-weights-iso\](d)\]. The remaining transformation brackets are obtained by laddering outward, to more peripheral weight points of larger $\abs{M_T}$ \[the “$+$” arrow in Fig. \[fig-weights-iso\](a)\]. If the transformation brackets (\[eqn-chainII-bracket\]) are considered as entries of numerical eigenvectors, then the laddering operation (\[eqn-iso-ladder\]) consists of multiplication by the (generally nonsquare and sparse) matrix realization of $T_{\pm\mp}$ between weight points. For a vector of isospin $T$, laddering past the weight point $M_T=T$ gives a null result. The process terminates at $M_T= T_\text{max}(RS;M_S)$, or $M_T= -T_\text{max}(RS;M_S)$ for laddering towards negative $M_T$. Although the laddering procedure provides a [*consistent*]{} set of coupling coefficients, the resolution of multiplicities arising from the diagonalization (at $M_T=0$ or $+\half$) remains arbitrary and thus still does not provide a [*unique*]{} and reproducible prescription. Uniqueness (to within phase) can be obtained by further diagonalizing a “second” operator, as detailed in Ref. .[@fn-ms-conj] The second, alternative approach to constructing the transformation brackets does not involve explicitly diagonalizing ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$. Rather, it is based on laddering inward along a diagonal of constant $M_S$, from the most peripheral weight point \[the “$-$” arrow in Fig. \[fig-weights-iso\](a)\] in conjunction with orthogonalization at each weight point. The weight point $M_T= T_\text{max}(RS;M_S)$ contains only a single basis state, which therefore is also the $T=T_\text{max}$ basis vector for chain (), to within sign. We are free to choose this sign [*e.g.*]{}, as always positive. Laddering inward to $M_T=T_\text{max}-1$ yields the $T=T_\text{max}$ basis vector at this weight point. If the number of basis states ([*i.e.*]{}, the dimension of the subspace) at this point is larger than one, then the remaining orthogonal vector (or vectors) needed to span the space must be the chain () basis vector (or vectors) of isospin $T=T_\text{max}-1$. Ths degeneracy will be $\operatorname{mult}(RS;M_S,T_\text{max}-1)$. The orthogonal basis of good isospin may therefore be found by Gram-Schmidt orthogonalization of a complete but nonorthogonal basis, starting from the [*known*]{} $T=T_\text{max}$ basis vector, which is supplemented by $\operatorname{mult}(RS;M_S,T_\text{max}-1)$ further linearly independent vectors as needed to span the subspace. A [*unique*]{} set of transformation brackets, both in terms of phases and resolution of multiplicities, is obtained if a well-defined prescription is used for specifying these further independent vectors. For instance, one might use basis states taken from chain (), in order of increasing weight for the $\grpso{4}$ label $(XY)$, starting with the lowest-weight $\grpso{4}$ label available at the weight point. (Alternatively, uniqueness can be enforced by diagonalizing a “second” operator within the $T=T_\text{max}-1$ space, supplemented by a phase convention.) The process of laddering followed by orthonormalization must then be repeated for $M_T=T_\text{max}-2$, $T_\text{max}-3$, [*etc.*]{}, until $M_T=0$ or $+\half$ is reached. For large-scale computations, the choice between the two methods, (1) [*diagonalization*]{} of ${\boldsymbol{\mathbf{\mathrm{T}}}}^2$ followed by [*outward laddering*]{} or (2) [*inward laddering*]{} alternating with [*orthogonalization*]{}, will be dictated by considerations of numerical efficiency and accuracy. These will depend upon the numerical linear algebra algorithms being used. The latter method provides the simplest route to a unique, reproducible set of phases and resolution of the branching multiplicity. Once the transformation brackets between bases reducing chains () and () have been obtained, the transformation of reduced coupling coefficients follows immediately. The full (unreduced) coupling coefficient may be interpreted as the inner product of a coupled state with an uncoupled product of two states, each described by (\[eqn-chainII-bracket\]). Then, for the $\grpso{5}\supset\grpu[N]{1}\otimes\grpso[T]{3}$ reduced coupling coefficient it follows from the factorization lemma and orthonormality that $$\begin{gathered} \label{eqn-chainII-transform} \underbrace{ \cg{(R_1S_1)}{M_{S 1} \kappa_1 T_1}{(R_2S_2)}{M_{S 2} \kappa_2 T_2}{\rho(RS)}{M_{S} \kappa T} }_{\grpso{5}\supset\grpu[N]{1}\otimes\grpso[T]{3}} \\ = \sum_{\substack{(X_1Y_1)(X_2Y_2)(XY)\\M_{T 1}(M_{T 2})}} \underbrace{ \cg{T_1}{M_{T 1}}{T_2}{M_{T 2}}{T}{M_{T}} }_{\grpso{3}} \underbrace{ \cg{(X_1Y_1)}{M_{X 1}M_{Y 1}}{(X_2Y_2)}{M_{X 2}M_{Y 2}}{(XY)}{M_{X}M_{Y}} }_{\grpso{4}} \\\times \underbrace{ \overlap[3]{(R_1S_1)\\M_{S 1} \kappa_1 T_1\\M_{T 1}}{(R_1S_1)\\(X_1Y_1)\\M_{X 1}M_{Y 1}} \overlap[3]{(R_2S_2)\\M_{S 2} \kappa_2 T_2\\M_{T 2}}{(R_2S_2)\\(X_2Y_2)\\M_{X 2}M_{Y 2}} \overlap[3]{(RS)\\M_{S} \kappa T\\M_{T}}{(RS)\\(XY)\\M_{X}M_{Y}} }_{\grpso{5}\supset[\grpu[N]{1}\otimes\grpso[T]{3}\leftrightarrow\grpso{4}]} \\\times \underbrace{ \cg{(R_1S_1)}{(X_1Y_1)}{(R_2S_2)}{(X_2Y_2)}{\rho(RS)}{(XY)} }_{\grpso{5}\supset\grpso{4}} ,\end{gathered}$$ for any value of $M_T$ allowed given isospin $T$, where again it is to be understood that $M_X=\tfrac12(M_S+M_T)$, $M_Y=\tfrac12(M_S-M_T)$, and similarly for $M_{X1}$, $M_{Y1}$, $M_{X2}$, and $M_{Y2}$. The sum over $\grpso{4}$ irrep labels is subject to the usual branching and triangularity constraints on the canonical reduced coupling coefficients. For the resulting chain () reduced coupling coefficient to be nonvanishing, it must obey the $\grpso{5}\rightarrow\grpu[N]{1}\otimes\grpso[T]{3}$ branching rules (\[eqn-chain-II-mult\]), the $\grpu[N]{1}$ additivity condition $M_{S1}+M_{S2}=M_{S}$, and the $\grpso[T]{3}$ triangularity condition $T_1\otimes T_2\rightarrow T$. In Appendix \[app-canonical\], the canonical reduced coupling coefficients for the $\grpso{5}$ coupling $(10)\otimes(1\half)\rightarrow(1\half)$ were calculated \[Fig. \[fig-system-large\](c)\], as a relatively simple numerical example involving an outer multiplicity ($D=2$). For a concrete illustration of the transformation procedure just described, let us consider the transformation of these coefficients into chain () reduced coupling coefficients. The decomposition of the irreps $(10)$ and $(1\half)$ into $\grpu[N]{1}\otimes\grpso[T]{3}$ irreps is shown in Fig. \[fig-weights-iso\](c,d). For the transformation brackets, we follow the second method above. For instance, consider the $M_S=+\half$ diagonal of the weight diagram for the $(1\half)$ irrep of $\grpso{5}$ \[Fig. \[fig-weights-iso\](d)\]. The $T=\tfrac32$ seed state at $M_T=+\tfrac32$ is $\ket{{+\tfrac12}\tfrac32{+\tfrac32}}=+\ket{(1\tfrac12){+1}{-\tfrac12}}$, where we abbreviate chain () basis states as $\ket{(XY)M_XM_Y}$ and chain () basis states as $\ket{M_STM_T}$. Laddering inward to $M_T=+\tfrac12$ yields $T=\tfrac32$ state $$\ket{{+\tfrac12}\tfrac32{+\tfrac12}}= +\sqrt{\tfrac56}\ket{(\tfrac12 0){+\tfrac12}0} +\sqrt{\tfrac16}\ket{(\tfrac12 1){+\tfrac12}0}.$$ Orthogonalization, using $\ket{(\tfrac12 0){+\tfrac12}0}$ as the independent Gram-Schmidt basis vector, gives $T=\tfrac12$ state $$\ket{{+\tfrac12}\tfrac12{+\tfrac12}}= +\sqrt{\tfrac16}\ket{(\tfrac12 0){+\tfrac12}0} -\sqrt{\tfrac56}\ket{(\tfrac12 1){+\tfrac12}0}.$$ Continued laddering to negative $M_T$ gives the remaining transformation brackets along the diagonal. Straightforward application of (\[eqn-chainII-transform\]), using the full set of transformation brackets derived in this fashion, then yields the chain () reduced coupling coefficients in Table \[tab-chainII-coeffs\]. Note that the coefficients are either symmetric or antisymmetric under the $\grpu[N]{1}$ particle-hole conjugation operation $M_S\rightarrow -M_S$ (see Ref. ). The $\grpso[L]{3}$ subalgebra of chain () is the maximal $\grpso{3}$ subalgebra, [*i.e.*]{}, one which is not contained within any larger proper subalgebra of $\grpso{5}$. This subalgebra is obtained by letting[@corrigan1976:oscillator; @rowe1994:so5-so4-vcs; @debaerdemacker2007:so5-cartan] $$\label{eqn-chainIII-gen} \begin{gathered} L^{(1)}_{+1}= -\sqrt2X_+-\sqrt6\Tmp \quad L^{(1)}_{0}= X_0+3Y_0 \quad L^{(1)}_{-1}= \sqrt2X_-+\sqrt6\Tpm \\ O^{(3)}_{+3}= -\sqrt5Y_+ \quad O^{(3)}_{+2}= \sqrt{10}\Tpp \quad O^{(3)}_{+1}= -\sqrt3X_++2\Tmp \quad O^{(3)}_{0}= 3X_0-Y_0 \\ O^{(3)}_{-1}= \sqrt3X_--2\Tpm \quad O^{(3)}_{-2}= -\sqrt{10}\Tmm \quad O^{(3)}_{-3}= \sqrt5Y_-. \end{gathered}$$ Then $\grpso[L]{3}$ has generators $L^{(1)}_{M_L}$, where, as usual, $L^{(1)}_{\pm1}=\mp\tfrac1{\sqrt2}L_\pm$. The remaining generators $O^{(3)}_{M_L}$ constitute an octupole tensor with respect to $\grpso[L]{3}$. The commutation relations of the generators are given (most compactly in spherical tensor coupled form[@french1966:multipole]) by $$\label{eqn-chainIII-comm} \begin{aligned} [L,L]^{(1)}&=-\sqrt2L \quad & [L,O]^{(3)}&=-2\sqrt3O\\ [O,O]^{(1)}&=-2\sqrt7L & [O,O]^{(3)}&=\sqrt6O. \end{aligned}$$ The $\grpso[L]{3}$ weight $M_L$ is related to the canonical weights by $M_{L}=M_{X}+3M_{Y}$, according to (\[eqn-chainIII-gen\]), and thus defines an oblique axis in the weight space. The $\grpso{5}$ generators are shown classified according to this weight in Fig. \[fig-root-chains\](d). The dashed lines connect the canonical generators of Sec. \[sec-alg\], which have good $M_L$ but not $L$, to the linear combinations $L^{(1)}_{M_L}$ and $O^{(3)}_{M_L}$, which have both good $M_L$ and good $L$. The labels for basis states reducing the $\grpso[L]{3}$ subalgebra are indicated in (\[eqn-so5-chains\]). Basis states for chain () involve a linear combination of chain () basis states of the same $M_L$. The transformation between bases therefore involves a sum not only over $(XY)$, as in (\[eqn-chainII-bracket\]), but also over distinct weight points $(M_XM_Y)$, $$\label{eqn-chainIII-bracket} \ket[3]{(RS)\\ \alpha L\\M_{L}}= \sum_{\substack{(XY)\\M_X(M_Y)}} \overlap[3]{(RS)\\(XY)\\M_{X}M_{Y}}{(RS)\\ \alpha L\\M_{L}} \, \ket[3]{(RS)\\(XY)\\M_{X}M_{Y}},$$ subject to the constraint $M_{L}=M_{X}+3M_{Y}$. The transformation brackets may be systematically evaluated using an adaptation of either of the methods proposed above for Chain (): (1) diagonalization of ${\boldsymbol{\mathbf{\mathrm{L}}}}^2$ for the $M_L=0$ or $+\half$ subspace of the $\grpso{5}$ irrep, followed by laddering outward to larger-$M_L$ spaces, or (2) laddering inward from the maximal-$M_L$ space, in conjunction with orthogonalization within each successive lower-$M_L$ space. For a unique resolution of the $\grpso{5}\supset\grpso[L]{3}$ branching multiplicity, it has been suggested that chain () basis states be chosen in which the octupole tensor is diagonal, equivalent to diagonalizing the Hermitian “second” operator $(L\times L \times O \times L)^{(0)}+ (L\times O \times L \times L)^{(0)}$.[@rowe1995:k-canonical] Once transformation brackets have been obtained, the transformation of reduced coupling coefficients is given by $$\begin{gathered} \label{eqn-chainIII-transform} \underbrace{ \cg{(R_1S_1)}{\alpha_1 L_1}{(R_2S_2)}{\alpha_2 L_2}{\rho(RS)}{\alpha L} }_{\grpso{5}\supset\grpso[L]{3}} \\ = \sum_{\substack{(X_1Y_1)(X_2Y_2)(XY)\\M_{X 1}(M_{Y 1})M_{X 2}(M_{Y 2})M_{X}(M_{Y})\\M_{L 1}(M_{L 2})}} \underbrace{ \cg{L_1}{M_{L 1}}{L_2}{M_{L 2}}{L}{M_{L}} }_{\grpso{3}} \underbrace{ \cg{(X_1Y_1)}{M_{X 1}M_{Y 1}}{(X_2Y_2)}{M_{X 2}M_{Y 2}}{(XY)}{M_{X}M_{Y}} }_{\grpso{4}} \\\times \underbrace{ \overlap[3]{(R_1S_1)\\ \alpha_1 L_1\\M_{L 1}}{(R_1S_1)\\(X_1Y_1)\\M_{X 1}M_{Y 1}} \overlap[3]{(R_2S_2)\\ \alpha_2 L_2\\M_{L 2}}{(R_2S_2)\\(X_2Y_2)\\M_{X 2}M_{Y 2}} \overlap[3]{(RS)\\ \alpha L\\M_{L}}{(RS)\\(XY)\\M_{X}M_{Y}} }_{\grpso{5}\supset[\grpso[L]{3}\leftrightarrow\grpso{4}]} \\\times \underbrace{ \cg{(R_1S_1)}{(X_1Y_1)}{(R_2S_2)}{(X_2Y_2)}{\rho(RS)}{(XY)} }_{\grpso{5}\supset\grpso{4}} ,\end{gathered}$$ for any value of $M_L$ allowed given angular momentum $L$, where the summations over $M_{X1}$, $M_{Y1}$, $M_{X2}$, $M_{Y2}$, $M_X$, and $M_Y$ are subject to the constraints $M_{L 1}=M_{X 1}+3M_{Y 1}$, $M_{L 2}=M_{X 2}+3M_{Y 2}$, and $M_{L}=M_{X}+3M_{Y}$.

ArXiv

--- abstract: | The superfluid weight, free carrier density, and specific heat of the three-dimensional $tJ$ model are calculated by renormalization-group theory. We find that optimal hole doping for superfluidity occurs in the electron density range of $\langle n_i \rangle \approx$ 0.63–0.68, where the superfluid weight $n_s/m^\ast$ reaches a local maximum. This density range is within the novel $\tau$ phase, where the electron hopping strength renormalizes to infinity, the system remains partially filled at all length scales, and the electron-hopping expectation value remains distinctively non-zero at all length scales. The calculated superfluid weight drops off sharply in the overdoped region. Under hole doping, the calculated density of free carriers increases until optimal doping and remains approximately constant in the overdoped region, as seen experimentally in high-$T_c$ materials. Furthermore, from calculation of the specific heat coefficient $\gamma$, we see clear evidence of a gap in the excitation spectrum for the $\tau$ phase. PACS numbers: 74.72.-h, 71.10.Fd, 05.30.Fk, 74.25.Dw author: - 'Michael Hinczewski and A. Nihat Berker' title: | Superfluid Weight, Free Carrier Density, and Specific Heat of the $d=3$ $tJ$ Model at Finite Temperatures --- Introduction ============ The variation of the superfluid number density $n_s$ with temperature and carrier doping is of fundamental importance in describing the unique properties of the superconducting state in high-$T_c$ cuprates. Experimentally, muon-spin-rotation techniques are used to determine the closely related quantity $n_s/m^\ast$ (also known as the superfluid weight), where $m^\ast$ is the effective mass of the carriers in the superfluid. In the underdoped region of high-$T_c$ materials, $n_s/m^\ast$ increases with doping, and the low-temperature superfluid weight is correlated with $T_c$.[@Uemura; @Bernhard1] As the materials are doped past the optimal value (where $T_c$ is the highest), $n_s/m^\ast$ peaks and rapidly decreases.[@Niedermayer; @Locquet; @Bernhard2] The decrease in $n_s/m^\ast$ is surprising since the total density of free carriers saturates at optimal doping and remains almost constant in the overdoped region.[@Puchkov] By contrast, in a conventional superconductor, described by BCS theory, these two quantities have the same doping dependence. The $tJ$ model is a promising starting point in understanding these properties of cuprate superconductors. Mean-field $U(1)$ and $SU(2)$ slave-boson theories of the $tJ$ Hamiltonian have reproduced some aspects of the doping and temperature dependences of $n_s/m^\ast$.[@LeeSalk1; @LeeSalk2] More direct, unbiased numerical techniques applied to a $4\times 4$ $tJ$ cluster have observed a large peak in $n_s/m^\ast$ in the same region where pairing correlations indicate a superconducting ground state.[@DagottoRiera] A general limitation of these types of studies is that no finite-cluster approach can unambiguously identify phase transitions in the system, or exhibit the non-analytic behavior of thermodynamic quantities at these transitions. Alternatively, the physics of the bulk model can be studied through the position-space renormalization-group method, which has been used to determine the phase structure and thermodynamic properties of the $tJ$ and Hubbard models at finite temperatures.[@FalicovBerkerT; @FalicovBerker; @Hubshort; @Hubnew] In particular, Falicov and Berker’s calculation for the $tJ$ model in $d=3$ with the realistic coupling $J/t = 0.444$ produced a rich, multicritical phase diagram [@FalicovBerkerT; @FalicovBerker], with a novel low-temperature phase (called “$\tau$”) for $30-40\%$ hole doping where the electron hopping strength in the Hamiltonian renormalizes to infinity under repeated scale changes, while the system remains partially filled. This is the possible signature of a superconducting phase, and it is notable that a similar phase was also observed in the $d=3$ Hubbard model.[@Hubshort; @Hubnew] Our present study further develops this renormalization-group method, to yield the superfluid weight of the $tJ$ model as a function of temperature and hole doping. Our approach reproduces phenomenological features of high-$T_c$ materials. In particular we find that optimal doping is located in the vicinity of the $\tau$ phase, where $n_s/m^\ast$ peaks and then is sharply reduced with overdoping. Moreover, we also find that the density of free carriers increases until optimal doping, and saturates in overdoped region. These results suggest that the $\tau$ phase might indeed correspond to the superconducting phase in cuprates. Further supporting this idea, we present specific heat calculations that show clear evidence of a gap in the quasiparticle spectrum for the $\tau$ phase. The $tJ$ Hamiltonian ==================== We consider a $d$-dimensional hypercubic lattice where the $tJ$ model for electron conduction is defined by the Hamiltonian $$\begin{gathered} \label{eq:1} H = P \left[ \tilde{t} \sum_{\langle ij \rangle,\sigma} \left(c^\dagger_{i\sigma}c_{j\sigma} + c^\dagger_{j\sigma}c_{i\sigma}\right)\right.\\ \left. + \tilde{J} \sum_{\langle ij \rangle} \mathbf{S}_i\cdot\mathbf{S}_j - \tilde{V} \sum_{\langle ij \rangle} n_i n_j - \tilde{\mu} \sum_i n_i \right] P\,,\end{gathered}$$ where $c^\dagger_{i\sigma}$ and $c_{i\sigma}$ are creation and annihilation operators, obeying anticommutation rules, for an electron with spin $\sigma =\: \uparrow$ or $\downarrow$ at lattice site $i$, $n_{i\sigma} = c^\dagger_{i\sigma}c_{i\sigma}$, $n_i = n_{i\uparrow} + n_{i\downarrow}$ are the number operators, and $\mathbf{S}_i = \sum_{\sigma\sigma^\prime} c^\dagger_{i\sigma} \mathbf{s}_{\sigma\sigma^\prime} c_{i\sigma^\prime}$ is the single-site spin operator, with $\mathbf{s}$ the vector of Pauli spin matrices. The entire Hamiltonian is sandwiched between projection operators $P = \prod_{i} (1-n_{i\downarrow}n_{i\uparrow})$, which project out states with doubly-occupied sites. The interaction constants $\tilde{t}$, $\tilde{J}$, $\tilde{V}$ describe the following physical features: electron hopping ($\tilde{t}$), a nearest-neighbor antiferromagnetic coupling ($\tilde{J} >0$), and a nearest-neighbor interaction ($\tilde{V}$). The standard $tJ$ Hamiltonian is a special case of Eq.  with $\tilde{V}/\tilde{J} = 1/4$. For convenience, we introduce dimensionless interaction constants $t, J, V, \mu$, and rearrange the $\tilde{\mu}$ chemical potential term to group the Hamiltonian into a single lattice summation: $$\label{eq:2} \begin{split} -\beta H = & \sum_{\langle ij \rangle} P \biggl[ -t \sum_{\sigma} \left(c^\dagger_{i\sigma}c_{j\sigma} + c^\dagger_{j\sigma}c_{i\sigma}\right)\\ & - J \mathbf{S}_i\cdot\mathbf{S}_j + V n_i n_j + \mu ( n_i +n_j)\biggr] P\\ \equiv &\sum_{\langle ij \rangle} \{-\beta H (i,j)\}\,. \end{split}$$ Here $\beta = 1/k_B T$, so that the interaction constants are related by $t = \beta \tilde{t}$, $J = \beta \tilde{J}$, $V= \beta \tilde{V}$, $\mu = \beta\tilde{\mu}/2d$. The total Hamiltonian is now written as a sum of pair Hamiltonians $-\beta H(i,j)$. The sum over nearest-neighbor sites $(i,j)$ is taken so that the position of site $j$ is $\mathbf{r}_j = \mathbf{r}_i + \mathbf{a}_k$, where $\mathbf{a}_k$ is one of the $d$ lattice vectors. Since changing the sign of $t$ is equivalent to redefining the phase at every other site in the system, we shall choose $t>0$ with no loss of generality. The effective temperature variable will be $1/t = k_B T $, where we have taken $\tilde t = 1$ as the unit of energy. In order to study the superfluid weight, we introduce periodic boundary conditions, by considering the system as a ring in each axis direction threaded by a magnetic flux. We choose the vector potential $\mathbf{A}$ associated with the flux to have components $A/\sqrt{d}$ along each axis, so that the pair Hamiltonian becomes $$\begin{gathered} \label{eq:3} -\beta H(i,j) = P \biggl[ -t \sum_{\sigma} \left(e^{i\phi} c^\dagger_{i\sigma}c_{j\sigma} + e^{-i\phi} c^\dagger_{j\sigma}c_{i\sigma}\right)\\ - J \mathbf{S}_i\cdot\mathbf{S}_j + V n_i n_j + \mu ( n_i +n_j)\biggr] P\,,\end{gathered}$$ where $\phi= a A/\sqrt{d}$ and $a$ is the lattice spacing. For simplicity, we have adopted units so that $\hbar = c = e = 1$. In the presence of the applied phase twist $\phi$, the superfluid weight $n_s/m^\ast$ is related to the curvature of the total free energy $F$ near $\phi=0$,[@Fisher; @Scalapino] $$\label{eq:4} \frac{n_s}{m^\ast} = \frac{1}{Na^2} \lim_{A \to 0} \frac{\partial^2 F}{\partial A^2} = \frac{1}{Nd} \lim_{\phi \to 0} \frac{\partial^2 F}{\partial \phi^2}\,,$$ where $N \to \infty$ is the total number of lattice sites. In Sec.IIIE we shall show how this quantity can be calculated from the renormalization-group transformation developed below. Renormalization-Group Transformation ==================================== Recursion Relations ------------------- The position-space renormalization-group method used here starts with an approximate decimation in $d=1$, which is then generalized to higher dimensions by the Migdal-Kadanoff procedure [@FalicovBerkerT; @FalicovBerker]. In $d=1$, the Hamiltonian of Eq.  takes the form: $$\label{eq:14} -\beta H = \sum_i \left\{-\beta H(i,i+1) \right\}\,,$$ where $i=1,2,3,\ldots$. The decimation consists of finding a thermodynamically equivalent system, described by the Hamiltonian $-\beta^\prime H^\prime$, which depends only on the states of the odd-numbered sites. Since the quantum operators in the Hamiltonian do not commute, an exact decimation even in one dimension is not possible. We can carry out an approximate decimation as follows [@SuzTak; @TakSuz]: $$\label{eq:15} \begin{split} \text{Tr}_{\text{even}} e^{-\beta H} =&\text{Tr}_{\text{ even}}e^{\sum_{i}\left\{ -\beta H(i,i+1)\right\} }\\ =&\text{Tr}_{\text{even}} e^{\sum_{i}^{\text{ even}}\left\{ -\beta H(i-1,i)-\beta H(i,i+1) \right\} }\\ \simeq& \prod_{i}^{\text{even}}\text{Tr}_{i}e^{\left\{ -\beta H(i-1,i)-\beta H(i,i+1)\right\} }\\ =&\prod_{i}^{\text{ even}}e^{-\beta ^{\prime }H^{\prime }(i-1,i+1)}\\ \simeq& e^{\sum_{i}^{\text{even}}\left\{ -\beta ^{\prime }H^{\prime }(i-1,i+1)\right\} } =e^{-\beta ^{\prime }H^{\prime }}. \end{split}$$ Here $-\beta^\prime H^\prime$ is the Hamiltonian for the renormalized system, and $\text{Tr}_{\text{even}}$ is a trace over the degrees of freedom at all even-numbered sites. In the two approximate steps, marked by $\simeq$ in Eq. , we ignore the non-commutation of operators separated beyond three consecutive sites of the unrenormalized system (conversely, this means that anticommutation rules are taken into account within three consecutive sites at all successive length scales, as the renormalization-group procedure is repeated). These two steps involve the same approximation but in opposite directions, which gives some mutual compensation. Earlier studies of quantum spin systems have shown the success of this approximation at predicting finite-temperature behavior.[@SuzTak; @TakSuz] The renormalization-group mapping can be extracted from the third and fourth lines of Eq.(\[eq:15\]): $$e^{-\beta ^{\prime }H^{\prime }(i,k)}=\mbox{Tr}_{j}e^{-\beta H(i,j)-\beta H(j,k)}, \label{eq:16}$$ where $i,j,k$ are three consecutive sites of the unrenormalized system. The operators $-\beta ^{\prime }H^{\prime }(i,k)$ and $-\beta H(i,j)-\beta H(j,k)$ act on the space of two-site and three-site states respectively, so that, in terms of matrix elements, $$\begin{gathered} \langle u_{i}v_{k}|e^{-\beta ^{\prime }H^{\prime }(i,k)}|\bar{u}_{i}^{{}}% \bar{v}_{k}^{{}}\rangle = \label{eq:17}\\ \sum_{w_{j}}\langle u_{i}\,w_{j}\,v_{k}|e^{-\beta H(i,j)-\beta H(j,k)}|\bar{u}_{i}\,w_{j}\,\bar{v}_{k}^{{}}\rangle \:,\end{gathered}$$ where $u_{i},w_{j},v_{k},\bar{u}_{i},\bar{v}_{k}^{{}}$ are single-site state variables. Eq.(\[eq:17\]) is the contraction of a $27\times 27$ matrix on the right into a $9\times 9$ matrix on the left. We block-diagonalize the left and right sides of Eq.(\[eq:17\]) by choosing basis states which are the eigenstates of total particle number, total spin magnitude, total spin $z$-component, and parity. We denote the set of 9 two-site eigenstates by $\{|\phi _{p}\rangle \}$ and the set of 27 three-site eigenstates by $\{|\psi _{q}\rangle \}$, and list them in Tables I and II. Eq.(\[eq:17\]) is rewritten as $$\begin{gathered} \langle \phi _{p}|e^{-\beta ^{\prime }H^{\prime }(i,k)}|\phi _{\bar{p}% }\rangle = \label{eq:18}\\ \sum_{\substack{u,v,\bar{u},\\ \bar{v},w}} \sum_{\substack{q,\bar{q}}} \langle\phi _p|u_iv_k\rangle \langle u_iw_jv_k|\psi_q\rangle \langle \psi _q|e^{-\beta H(i,j)-\beta H(j,k)}|\psi _{\bar{q}}\rangle\cdot \\ \langle \psi_{\bar{q}}|\bar{u}_iw_j\bar{v}_k\rangle \langle \bar{u}_i\bar{v}_k|\phi _{\bar{p}}\rangle\:.\end{gathered}$$ Eq.  yields six independent elements for the matrix $\langle \phi _{p}|e^{-\beta ^{\prime }H^{\prime }(i,k)}|\phi_{\bar{p}}\rangle$, which we label $\gamma_p$ as follows: $$\label{eq:19} \begin{split} \gamma_p &\equiv \langle \phi _{p}|e^{-\beta ^{\prime }H^{\prime }(i,k)}|\phi_{p}\rangle \quad \text{for}\: p = 1,2,4,6,7,\\ \gamma_0 &\equiv \langle \phi _{2}|e^{-\beta ^{\prime }H^{\prime }(i,k)}|\phi_{4}\rangle\,. \end{split}$$ To calculate the $\gamma_p$, we determine the matrix elements of $-\beta H(i,j) -\beta H(j,k)$ in the three-site basis $\{\psi_q\}$, as listed in Table III, and exponentiate the matrix blocks to find the elements $\langle \psi _q|e^{-\beta H(i,j)-\beta H(j,k)}|\psi _{\bar{q}}\rangle$ which enter on the right-hand side of Eq. . In this way the $\gamma_p$ are functions of the interaction constants in the unrenormalized Hamiltonian, $\gamma_p = \gamma_p (t,\phi,J,V,\mu)$. ---------------------------------------------------------------------------------------------------------- $n$ $p$ $s$ $m_s$ Two-site basis states ----- ----- ------- ------- ------------------------------------------------------------------------------ $0$ $+$ $0$ $0$ $|\phi_{1}\rangle=|\circ\circ\rangle$ $1$ $+$ $1/2$ $1/2$ $|\phi_{2}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow \circ\rangle+|\circ\uparrow\rangle\}$ $1$ $-$ $1/2$ $1/2$ $|\phi_{4}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow \circ\rangle-|\circ\uparrow\rangle\}$ $2$ $-$ $0$ $0$ $|\phi_{6}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow\downarrow\rangle -|\downarrow\uparrow\rangle\}$ $2$ $+$ $1$ $1$ $|\phi_{7}\rangle=|\uparrow\uparrow\rangle$ $2$ $+$ $1$ $0$ $|\phi_{9}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow\downarrow\rangle+|\downarrow \uparrow\rangle\}$ ---------------------------------------------------------------------------------------------------------- : The two-site basis states, with the corresponding particle number ($n$), parity ($p$), total spin ($s$), and total spin $z$-component ($m_s$) quantum numbers. The states $|\phi_{3}\rangle$, $|\phi_{5}\rangle$, and $|\phi_{8}\rangle$, are obtained by spin reversal from $|\phi_{2}\rangle$, $|\phi_{4}\rangle$, and $|\phi_{7}\rangle$, respectively. ----------------------------------------------------------------------------------------------------------------------------- $n$ $p$ $s$ $m_s$ Three-site basis states ----- ----- ------- ------- ------------------------------------------------------------------------------------------------- $0$ $+$ $0$ $0$ $|\psi_{1}\rangle=|\circ\circ\,\circ\rangle$ $1$ $+$ $1/2$ $1/2$ $|\psi_{2}\rangle=|\circ \uparrow \circ\rangle,\: |\psi_{3}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow \circ\,\circ\rangle+|\circ\,\circ\uparrow\rangle\}$ $1$ $-$ $1/2$ $1/2$ $|\psi_{6}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow \circ\,\circ\rangle-|\circ\,\circ\uparrow\rangle\}$ $2$ $+$ $0$ $0$ $|\psi_{8}\rangle=\frac{1}{2}\{|\uparrow\downarrow\circ\rangle- |\downarrow\uparrow\circ\rangle-|\circ\uparrow\downarrow\rangle+ |\circ\downarrow\uparrow\rangle\}$ $2$ $-$ $0$ $0$ $|\psi_{9}\rangle=\frac{1}{2}\{|\uparrow\downarrow\circ\rangle- |\downarrow\uparrow\circ\rangle+|\circ\uparrow\downarrow\rangle- |\circ\downarrow\uparrow\rangle\},$ $|\psi_{10}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow\circ\downarrow\rangle-|\downarrow\circ\uparrow \rangle\}$ $2$ $+$ $1$ $1$ $|\psi_{11}\rangle=|\uparrow\circ\uparrow\rangle,\: |\psi_{12}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow\uparrow\circ\rangle+|\circ\uparrow\uparrow \rangle\}$ $2$ $+$ $1$ $0$ $|\psi_{13}\rangle=\frac{1}{2}\{|\uparrow\downarrow\circ\rangle+ |\downarrow\uparrow\circ\rangle+|\circ\uparrow\downarrow\rangle+ |\circ\downarrow\uparrow\rangle\},$ $|\psi_{14}\rangle=\frac{1}{\sqrt{2}} \{|\uparrow\circ\downarrow\rangle+|\downarrow\circ\uparrow \rangle\}$ $2$ $-$ $1$ $1$ $|\psi_{17}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow\uparrow\circ\rangle-|\circ\uparrow\uparrow \rangle\}$ $2$ $-$ $1$ $0$ $|\psi_{18}\rangle=\frac{1}{2}\{|\uparrow\downarrow\circ\rangle+ |\downarrow\uparrow\circ\rangle-|\circ\uparrow\downarrow\rangle- |\circ\downarrow\uparrow\rangle\}$ $3$ $+$ $1/2$ $1/2$ $|\psi_{20}\rangle=\frac{1}{\sqrt{6}}\{2|\uparrow\downarrow\uparrow\rangle-|\uparrow\uparrow \downarrow\rangle-|\downarrow\uparrow\uparrow\rangle\}$ $3$ $-$ $1/2$ $1/2$ $|\psi_{22}\rangle=\frac{1}{\sqrt{2}}\{|\uparrow\uparrow \downarrow\rangle-|\downarrow\uparrow\uparrow\rangle\}$ $3$ $+$ $3/2$ $3/2$ $|\psi_{24}\rangle=|\uparrow\uparrow\uparrow\rangle$ $3$ $+$ $3/2$ $1/2$ $|\psi_{25}\rangle=\frac{1}{\sqrt{3}}\{|\uparrow\downarrow\uparrow\rangle+|\uparrow\uparrow \downarrow\rangle+|\downarrow\uparrow\uparrow\rangle\}$ ----------------------------------------------------------------------------------------------------------------------------- : The three-site basis states, with the corresponding particle number ($n$), parity ($p$), total spin ($s$), and total spin $z$-component ($m_s$) quantum numbers. The states $|\phi_{4-5}\rangle$, $|\phi_{7}\rangle$, $|\phi_{15-16}\rangle$, $|\phi_{19}\rangle$, $|\phi_{21}\rangle$, $|\phi_{23}\rangle$, $|\phi_{26-27}\rangle$, are obtained by spin reversal from $|\phi_{2-3}\rangle$, $|\phi_{6}\rangle$, $|\phi_{11-12}\rangle$, $|\phi_{17}\rangle$, $|\phi_{20}\rangle$, $|\phi_{22}\rangle$, $|\phi_{24-25}\rangle$, respectively. $$\begin{gathered} \begin{array}{|c||c|}\hline & \psi_{1}\\ \hhline{|=#=|} \psi_{1} & 0\\ \hline \end{array}\\ \begin{array}{|c||c|c|c|}\hline & \psi_{2} & \psi_{3} & \psi_{6}\\ \hhline{|=#=|=|=|} \psi_{2} & 2\mu & -\sqrt{2}t \cos\phi & i \sqrt{2} t \sin\phi\\ \hline \psi_{3} & -\sqrt{2}t \cos\phi & \mu & i \Delta_1\\ \hline \psi_{6} & -i \sqrt{2} t \sin\phi & -i \Delta_1 & \mu\\ \hline \end{array}\\ \begin{array}{|c||c|c|c|}\hline & \psi_{8} & \psi_{9} & \psi_{10}\\ \hhline{|=#=|=|=|} \psi_{8} & -\frac{3}{4}J + V + 3\mu & -i \Delta_3 & -i \sqrt{2} t\sin \phi\\ \hline \psi_{9} & i \Delta_3 & -\frac{3}{4}J + V + 3\mu & -\sqrt{2} t\cos \phi\\ \hline \psi_{10} & i \sqrt{2} t\sin \phi & -\sqrt{2} t\cos \phi & 2\mu \\ \hline \end{array}\\ \begin{array}{|c||c|c|c|}\hline & \psi_{11} & \psi_{12} & \psi_{17}\\ \hhline{|=#=|=|=|} \psi_{11} & 2\mu & -\sqrt{2} t \cos\phi & i \sqrt{2} t\sin \phi\\ \hline \psi_{12} & -\sqrt{2}t\cos\phi & \frac{1}{4}J + V + 3\mu & i \Delta_2\\ \hline \psi_{17} & -i \sqrt{2} t\sin \phi & -i\Delta_2 & \frac{1}{4}J + V + 3\mu \\ \hline \end{array}\\ \begin{array}{|c||c|c|c|}\hline & \psi_{13} & \psi_{14} & \psi_{18}\\ \hhline{|=#=|=|=|} \psi_{13} & \frac{1}{4}J + V + 3\mu & -\sqrt{2} t \cos\phi & i \Delta_4\\ \hline \psi_{14} & -\sqrt{2}t\cos\phi & 2\mu & i \sqrt{2}t\sin\phi\\ \hline \psi_{18} & -i \Delta_4 & -i\sqrt{2} t\sin \phi & \frac{1}{4}J + V + 3\mu \\ \hline \end{array}\\ \begin{array}{|c||c|}\hline & \psi_{20}\\ \hhline{|=#=|} \psi_{20} & -J + 2V + 4\mu \\ \hline \end{array}\quad \begin{array}{|c||c|}\hline & \psi_{22}\\ \hhline{|=#=|} \psi_{22} & 2V + 4\mu \\ \hline \end{array}\\ \begin{array}{|c||c|}\hline & \psi_{24}\\ \hhline{|=#=|} \psi_{24} & \frac{1}{2}J +2V +4\mu \\ \hline \end{array}\quad \begin{array}{|c||c|}\hline & \psi_{25}\\ \hhline{|=#=|} \psi_{25} & \frac{1}{2}J +2V +4\mu \\ \hline \end{array}\end{gathered}$$ Since the matrix $\langle \phi _{p}|e^{-\beta ^{\prime }H^{\prime }(i,k)}|\phi_{\bar{p}}\rangle$ is determined by six independent elements $\gamma_p$, the renormalized pair Hamiltonian $-\beta ^{\prime }H^{\prime }(i,k)$ involves six interaction constants, namely those of the original types of interactions and an additive constant: $$\begin{gathered} \label{eq:20} -\beta^\prime H^\prime(i,k) = P \biggl[ -t^\prime \sum_{\sigma} \left(e^{i\phi^\prime} c^\dagger_{i\sigma}c_{j\sigma} + e^{-i\phi^\prime} c^\dagger_{j\sigma}c_{i\sigma}\right)\\ - J^\prime \mathbf{S}_i\cdot\mathbf{S}_j + V^\prime n_i n_j + \mu^\prime ( n_i +n_j) + G^\prime\biggr] P\,,\end{gathered}$$ The matrix elements of $-\beta ^{\prime }H^{\prime }(i,k)$ in the $\{\phi_p\}$ basis are shown in Table IV. Exponentiating this matrix, we can solve for the renormalized interaction constants $(t^\prime,\phi^\prime,J^\prime,V^\prime,\mu^\prime,G^\prime)$ in terms of the $\gamma_p$: $$\begin{gathered} t^\prime = \text{sign}\,(\gamma_4-\gamma_2)\cosh^{-1} \left(\frac{\gamma_2+\gamma_4}{2e^{v}}\right),\nonumber\\[5pt] \phi^\prime = \tan^{-1} \left(\frac{2\, \text{Im}\,\gamma_0}{\gamma_4-\gamma_2}\right),\qquad J^\prime = \ln\frac{\gamma_7}{\gamma_6},\nonumber\\[5pt] V^\prime = \frac{1}{4}\left\{\ln(\gamma_1^4 \gamma_6 \gamma_7^3)-8v\right\},\qquad \mu^\prime = v - \ln \gamma_1,\nonumber\\[5pt] G^\prime=\ln\gamma_1,\label{eq:21}\end{gathered}$$ where $$\begin{gathered} v = \frac{1}{2}\ln\left(\gamma_2\gamma_{4}-\gamma_0^\ast \gamma_0\right)\,.\end{gathered}$$ The approximate $d=1$ decimation contained in Eqs. - can be expressed as a mapping of a Hamiltonian with interaction constants $\mathbf{K} = \{ G,t,J,V,\mu,\phi\}$ onto another Hamiltonian with interactions constants $$\label{eq:22} \mathbf{K}^\prime = \mathbf{R}(\mathbf{K})\,.$$ The Migdal-Kadanoff procedure [@Migdal; @Kadanoff] is used to construct the renormalization-group transformation for $d>1$. We ignore a subset of the nearest-neighbor interactions in the $d$-dimensional hypercubic lattice, leaving behind a new $d$-dimensional hypercubic lattice where each point is connected to its neighbor by two consecutive nearest-neighbor segments of the original lattice. We apply the decimation described above to the middle site between the two consecutive segments, giving the renormalized nearest-neighbor couplings for the points in the new lattice. We compensate for the interactions that are ignored in the original lattice by multiplying by a factor of $b^{d-1}$ the interactions after the decimation, $b=2$ being the length rescaling factor. Thus, the renormalization-group transformation of Eq.  generalizes, for $d>1$, to $$\label{eq:13} \mathbf{K^\prime}=b^{d-1} \mathbf{R}(\mathbf{K}).$$ Renormalization-Group Transformation in the Presence of Magnetic Flux --------------------------------------------------------------------- In order to correctly model the response of the system to an applied magnetic flux, the renormalization-group approximation described in the last two sections needs to be extended. To see this, we first review the formalism for calculating thermodynamic densities from the renormalization-group flows.[@BerkerOstlundPutnam] Conjugate to each interaction $K_\alpha$ of $\mathbf{K} = \{ K_\alpha\}$, there is a density $M_\alpha$ (e.g., kinetic energy, electron density), $$\label{eq:23a} M_\alpha = \frac{1}{Nd} \frac{\partial \ln Z}{\partial K_\alpha}\,,$$ where $Z(\mathbf{K})$ is the partition function. We can relate the densities at the two consecutive points along a renormalization-group trajectory by $$\label{eq:23b} M_\alpha = b^{-d} M^\prime_\beta T_{\beta \alpha}\,, \quad \text{where} \qquad T_{\beta\alpha} \equiv \frac{\partial K^\prime_\beta}{\partial K_\alpha}\,,$$ with summation over repeated indices implied. At a fixed point of the renormalization-group transformation, corresponding to a phase transition or a phase sink, the densities $M_\alpha = M_\alpha^\prime \equiv M_\alpha^\ast$ are the left eigenvector with eigenvalue $b^d$ of the recursion matrix $\mathbf{T}$ evaluated at the fixed point. The densities at the starting point of the trajectory (the actual physical system) are computed by iterating Eq.  until a fixed point is effectively reached. If $\mathbf{T}^{(k)}$ is the recursion matrix of the $(k)$th renormalization-group iteration, then for large $k$, we can express the densities of the actual system $\mathbf{M}$ as $$\label{eq:25} \mathbf{M} \simeq b^{-kd} \mathbf{M}^\ast\ \cdot [\mathbf{T}^{(k)}] \cdot [\mathbf{T}^{(k-1)}] \cdot \cdots \cdot [\mathbf{T}^{(1)}].$$ The renormalization-group transformation incorporated in Eqs. - gives $$\begin{gathered} \frac{\partial t^\prime}{\partial \phi} = \frac{\partial J^\prime}{\partial \phi} = \frac{\partial V^\prime}{\partial \phi} =\frac{\partial G^\prime}{\partial \phi}= 0\,,\nonumber\\ \frac{\partial \phi^\prime}{\partial t} = \frac{\partial \phi^\prime}{\partial J} = \frac{\partial \phi^\prime}{\partial V} =\frac{\partial \phi^\prime}{\partial \mu}= 0\,, \quad \frac{\partial \phi^\prime}{\partial \phi} = 2\,,\label{eq:28}\end{gathered}$$ for all $\phi$. The $6\times 6$ recursion matrix $\mathbf{T}$ will then have the form $$\label{eq:29} \mathbf{T} = \left(\begin{array}{cccc|c} b^d & \frac{\partial G^\prime}{\partial t} & \cdots & \frac{\partial G^\prime}{\partial \mu} & \\ 0 & \frac{\partial t^\prime}{\partial t} & \cdots & \frac{\partial t^\prime}{\partial \mu} & 0 \\ \vdots & \vdots & \ddots & \vdots & \\ 0 & \frac{\partial \mu^\prime}{\partial t} & \cdots & \frac{\partial \mu^\prime}{\partial \mu} & \\ \hline & 0 & & & 2\\ \end{array}\right)$$ at every step in the flow. This leads to $$\label{eq:30} M^\ast_6 = 0\ \quad \text{and} \quad \frac{\partial}{\partial \phi} \ln Z = M_6 =0\,,$$ for all points of the phase diagram. This superfluid weight of zero for all temperatures and electronic densities is clearly due to the oversimplification in our initial approximation. The source of the problem is the three-site cluster approximation used in deriving the recursion relations. In modifying the original approximation scheme, we seek to incorporate the effect of the non-commutations extending beyond the three-site cluster. Turning to the matrix elements of $-\beta H(i,j) -\beta H(j,k)$ listed in Table III, we note the terms $\Delta_i$, $i=1,\ldots,4$. Using the original Hamiltonian of Eq.  restricted to the three-cluster, the matrix elements involving these terms are all zero. However, non-commutativity extending beyond the three-cluster makes, as we see below, these matrix elements non-zero. We can estimate the magnitude of the matrix elements $\Delta_i$ by considering a five-site cluster, described by Hamiltonian $-\beta H(h,i) -\beta H(i,j) -\beta H(j,k) -\beta H(k,l)$, where $(h,i,j,k,l)$ are consecutive sites. In the spirit of Eq. , we generate effective couplings for the three-cluster by tracing over the degrees of freedom at the outside sites in the five-cluster, $$\begin{gathered} \langle u_{i}\,v_{j}\, w_{k}|e^{-\tilde{\beta} \tilde{H}(i,j,k)}|\bar{u}_{i}\, \bar{v}_{j}\,\bar{w}_{k} \rangle = \label{eq:32}\\ \sum_{t_{h},x_{l}}\langle t_{h}\, u_{i}\,v_{j}\,w_{k}\,x_{l}|e^{-\beta H(h,i) -\beta H(i,j)-\beta H(j,k)-\beta H(k,l)}\\ \cdot|t_h\,\bar{u}_{i}\,\bar{v}_{j}\,\bar{w}_{k}\,x_l\rangle \:,\end{gathered}$$ where the subscripted variables refer to single-site states. From the above equation, we can extract the matrix elements of an effective three-cluster Hamiltonian $-\tilde{\beta} \tilde{H}(i,j,k)$. Eq.  is the contraction of a $243 \times 243$ matrix on the right-hand side into a $27 \times 27$ matrix on the left. We simplify our task by using the $\{\psi_p\}$ basis on the left, and choosing an appropriate five-site basis to block-diagonalize the $243 \times 243$ right-hand matrix. Since $-\tilde{\beta} \tilde{H}(i,j,k)$ is derived from the decimation of a five-cluster, it will have a more general form than $-\beta H(i,j) -\beta H(j,k)$, and approximately reflect the effect of the three-cluster non-commutations with the external sites. However our approximation scheme must also satisfy an important constraint: the $\phi \to 0$ limit should yield the same renormalization-group transformation used in earlier studies of the $tJ$ model [@FalicovBerkerT; @FalicovBerker]. To achieve this, we modify only a subset of the matrix elements of $-\beta H(i,j) -\beta H(j,k)$, namely those which are zero in the original scheme when $\phi \ne 0$, but whose corresponding elements in $-\tilde{\beta} \tilde{H}(i,j,k)$ are non-zero: $$\label{eq:33} \begin{split} \Delta_1 &= \text{sign}(\phi t)|\langle \psi_3 | \tilde{\beta}\tilde{H}(i,j,k) |\psi_6\rangle|,\\ \Delta_2 &= \text{sign}(\phi t)|\langle \psi_{12} | \tilde{\beta}\tilde{H}(i,j,k) |\psi_{17}\rangle|,\\ \Delta_3 &= \text{sign}(\phi t)|\langle \psi_{8} | \tilde{\beta}\tilde{H}(i,j,k) |\psi_{9}\rangle|,\\ \Delta_4 &= \text{sign}(\phi t)|\langle \psi_{13} | \tilde{\beta}\tilde{H}(i,j,k) |\psi_{18}\rangle|. \end{split}$$ The $\text{sign}(\phi t)$ prefactors guarantee that couplings between the same types of three-cluster states have the same sign. For example, $|\psi_2\rangle$ and $|\psi_3\rangle$ share the same $n$, $p$, $s$, and $m_s$ quantum numbers, as can be seen from Table II. A nonzero $\phi$ couples $|\psi_2\rangle$ to $|\psi_6\rangle$, a state with the same $n$, $s$, and $m_s$, but opposite parity. From the second block in Table III, the associated matrix element is $\langle \psi_2 |\cdots|\psi_6\rangle = i\sqrt{2}t\sin(\phi)$. The $\Delta_1$ elements in that block have an analogous role, coupling $|\psi_3\rangle$ to $|\psi_6\rangle$. The prefactor in the $\Delta_1$ expression of Eq.  sets the sign of the element $\langle \psi_3 |\cdots|\psi_6\rangle = i\Delta_1$ to equal that of $\langle \psi_2 |\cdots|\psi_6\rangle$. Since our calculations are all done for small $\phi$, $\text{sign}(\sin \phi) = \text{sign}(\phi)$. Similar reasoning applies to the prefactors of the other $\Delta_i$ elements. Through Eq. , the $\Delta_i$ are functions of the interactions strengths in the unrenormalized Hamiltonian, $\Delta_i = \Delta_i(t,J,V,\mu,\phi)$. They scale like $\phi$ for small $\phi$, and duly vanish in the limit $\phi \to 0$. As will be explained in Sec.IIIE, finding the superfluid weight involves calculating a thermodynamic density in the $\phi \to 0$ limit, so we shall be working in the regime where the $\Delta_i$ are vanishingly small. The result of the extended calculation, taking into account the quantum mechanical backflow into the three-cluster, is that Eqs.  no longer hold, $\partial \ln Z /\partial \phi \ne 0$ in general, and we obtain interesting nontrivial results for $n_s/m^\ast$. Calculation of the Superfluid Weight ------------------------------------ The superfluid weight of Eq.  is expressed as a derivative of the total free energy $F = F(n,T,\phi)$, where $n = \langle n_i \rangle$ is the electron density. In terms of the conjugate current $$\label{eq:34} j(n,T,\phi) = \frac{1}{Nd}\left.\frac{\partial F}{\partial \phi}\right|_{n,T}\,,$$ Eq.  becomes $$\label{eq:35} \frac{n_s}{m^\ast}(n,T) = \lim_{\phi \to 0} \left.\frac{\partial j}{\partial \phi}\right|_{n,T}\,.$$ In terms of the grand potential $\Omega(\mu,T,\phi) = -(1/\beta)\ln Z$, $$\label{eq:36} j(\mu,T,\phi) = \frac{1}{Nd} \left.\frac{\partial \Omega}{\partial \phi}\right|_{\mu,T}\,,$$ and $$\label{eq:40} n(\mu,T,\phi) = -\frac{\beta}{2Nd} \left.\frac{\partial \Omega}{\partial \mu}\right|_{T,\phi}\,.$$ Relating the partial derivatives of $j$ with respect to $\phi$ through $$\label{eq:37} \left.\frac{\partial j}{\partial \phi}\right|_{\mu,T} = \left.\frac{\partial j}{\partial n}\right|_{\phi,T} \left.\frac{\partial n}{\partial \phi}\right|_{\mu,T} + \left.\frac{\partial j}{\partial \phi}\right|_{n,T}\,,$$ and using the Maxwell relation $\left.\frac{\partial n}{\partial \phi}\right|_{\mu,T} = -\frac{\beta}{2}\left.\frac{\partial j}{\partial \mu}\right|_{\phi,T}$, $$\label{eq:38} \left.\frac{\partial j}{\partial \phi}\right|_{\mu,T} = -\frac{\beta}{2}\left.\frac{\partial j}{\partial n}\right|_{\phi,T} \left.\frac{\partial j}{\partial \mu}\right|_{\phi,T} + \left.\frac{\partial j}{\partial \phi}\right|_{n,T}\,.$$ The current $j$ is zero when $\phi=0$, so that the first term on the right-hand side above is also zero in the limit $\phi \to 0$, and we find that $\lim_{\phi\to 0} \left.\frac{\partial j}{\partial \phi}\right|_{\mu,T} = \lim_{\phi\to 0} \left.\frac{\partial j}{\partial \phi}\right|_{n,T}$. Thus Eq.  can be equivalently written as $$\begin{gathered} \label{eq:39} \frac{n_s}{m^\ast}(\mu,T) = \lim_{\phi \to 0} \left.\frac{\partial j}{\partial \phi}\right|_{\mu,T} = \frac{1}{Nd} \lim_{\phi \to 0} \left.\frac{\partial^2 \Omega}{\partial \phi^2}\right|_{\mu,T}\\ =-\frac{1}{\beta Nd} \lim_{\phi \to 0} \left.\frac{\partial^2 \ln Z}{\partial \phi^2}\right|_{\mu,T} \,.\end{gathered}$$ This is the form we shall use when calculating the superfluid weights. Results ======= Global Phase Diagram for $d=3$ ------------------------------ Each sink, or completely stable fixed point of the renormalization-group flows, corresponds to a thermodynamic phase, and we find the global phase diagram by determining the basin of attraction for every sink [@BerkerWortis]. Flows that start at the boundaries between phases have their own fixed points, distinguished from phase sinks by having at least one unstable direction. Analysis of these fixed points determines whether the phase transition is first- or second-order. As explained in Sec.IIIC, the thermodynamic densities, which are the expectation values of operators occurring in the Hamiltonian, can also be calculated from the renormalization-group flows. In particular, we determine the single-site electron density $\langle n_i \rangle$. For the coupling $J/t = 0.444$ and $\phi =0$, the phase diagram in terms of $\langle n_i \rangle$ and temperature $1/t$ is shown in Fig. 1 [@FalicovBerkerT; @FalicovBerker]. ![Phase diagram for the $d=3$ $tJ$ model with $J/t = 0.444$, $\phi = 0$, in temperature versus electron density.[@FalicovBerker] The antiferromagnetic (A), dense disordered (D), dilute disordered (d), and $\tau$ phases are seen. The second-order phase boundaries are drawn with full curves. The coexistence boundaries of first-order transitions are drawn with dotted curves, with the unmarked areas inside corresponding to coexistence regions of the two phases at either side. The dashed lines are not phase transitions, but disorder lines between the dilute disordered and dense disordered phases.](tjcondfig1.eps) ------------ ------------------------------------------------------- ----------------------- ----------------------------- --------------------------- Phase sink $\sum_\sigma \langle c^\dagger_{i\sigma}c_{j\sigma} + $\langle n_i \rangle$ $\langle \mathbf{S}_i \cdot $\langle n_i n_j \rangle$ c^\dagger_{j\sigma}c_{i\sigma}\rangle$ \mathbf{S}_j \rangle$ d 0 0 0 0 D 0 1 0 1 A 0 1 $\frac{1}{4}$ 1 $\tau$ $-\frac{2}{3}$ $\frac{2}{3}$ $-\frac{1}{4}$ $\frac{1}{3}$ ------------ ------------------------------------------------------- ----------------------- ----------------------------- --------------------------- : Expectation values at the phase-sink fixed points. The nature of the various phases is epitomized by the thermodynamic densities $\mathbf{M}^\ast$ calculated at each phase sink (Table V), which underpin the calculation of densities throughout their respective phases (Eq. ). We summarize the phase properties below (for a more detailed discussion, see [@FalicovBerkerT; @FalicovBerker]): [**Dilute disordered phase (d):**]{} The electron density $\langle n_i \rangle = 0$ at the sink and, as a result, the $\langle n_i \rangle$ calculated inside this phase are low. [**Dense disordered phase (D):**]{} The electron density $\langle n_i \rangle = 1$ at the sink and, as a result, the $\langle n_i \rangle$ calculated inside this phase are close to 1. [**Antiferromagnetic phase (A):**]{} The electron density $\langle n_i \rangle = 1$ at the sink, so that this phase is also densely filled. The nearest-neighbor spin-spin correlation $\langle \mathbf{S}_i \cdot \mathbf{S}_j\rangle = 1/4$ at the sink. Two spins that are nearest neighbors at the sink are distant members of the same sublattice in the original cubic lattice. The non-zero value of the correlation function at the sink leads to $\langle \mathbf{S}_i \cdot \mathbf{S}_j\rangle < 0$ for nearest-neighbor sites of the original, unrenormalized system. [**$\tau$ phase:**]{} This is a novel phase, characterized by partial-filling at the phase sink, $\langle n_i\rangle = 2/3$. It is the only phase where the electron hopping strength $t$ does not renormalize to zero after repeated rescalings; instead, $t \to \infty$ at the sink. As a result, the expectation value of the electron hopping operator at the sink is non-zero, $\sum_\sigma \langle c^\dagger_{i\sigma}c_{j\sigma} + c^\dagger_{j\sigma}c_{i\sigma}\rangle = -2/3$. This property makes it a possible $tJ$ model analogue to the superconducting phase in high-$T_c$ materials. The superfluid weight and thermodynamic results discussed below certainly support this idea. In the limit $\langle n_i \rangle = 1$, the system exhibits antiferromagnetic order at low temperatures, as expected from the spin-spin coupling in the Hamiltonian. Upon hole doping, there is a competition between the A and D phases, which respectively minimize antiferromagnetic potential energy and hole kinetic energy. Note the extent of the A phase near $\langle n_i \rangle = 1$, which persists only up to a small amount of hole doping $\delta = 1- \langle n_i \rangle \lesssim 0.05$. This feature is directly reminiscent of the antiferromagnetic phase in certain high-$T_c$ materials, for example La$_{2-x}$Sr$_x$CuO$_4$ [@Imada]. At intermediate dopings $\delta \approx 0.32 - 0.37$, we have a low-temperature $\tau$ phase, surrounded by islands of antiferromagnetism. (When the hopping strength $t$ increases under rescaling, this also lowers the free energy of antiferromagnetically long-range ordered states, which leads to these islands of A in the vicinity of the $\tau$ phase.[@FalicovBerker]) At hole dopings $\delta \gtrsim 0.37$, there is a transition to a dilute disordered phase, with a narrow region of first-order coexistence at lower temperatures. Superfluid Weight and Kinetic Energy ------------------------------------ Using the method of calculating thermodynamic densities described in Sec.IIIE, we determine $(1/Nd) \partial \ln Z/\partial \phi$ at small $\phi$. Taking the numerical derivative of this quantity at $\phi=0$, we find $n_s/m^\ast$ through Eq. . The superfluid weight is plotted as a function of electron density in Fig. 2, along four different constant temperature cross-sections of the phase diagram. For comparison, we also show in the same figure the calculated average kinetic energy per bond $\langle K \rangle$, where $K = - \sum_\sigma \left( c^\dagger_{i\sigma} c_{j\sigma}+ c^\dagger_{j\sigma} c_{i\sigma}\right)$. $\langle K \rangle$ and the total weight of $\sigma_1(\omega,T)$, the real part of the optical conductivity, are related by the sum rule [@Tan], $$\label{eq:8} \int_0^\infty d\omega\, \sigma_1(\omega,T) = \frac{\pi e^2}{2} \langle K \rangle\,.$$ In comparing the properties of the $tJ$ model to those of high-$T_c$ materials, we keep in mind that the $tJ$ Hamiltonian describes a one-band model, so cannot account for interband transitions. For real materials, the full conductivity sum rule has the form $$\label{eq:10} \int_0^\infty d\omega\, \sigma_1(\omega,T) = \frac{\pi e^2 n}{2m}\,,$$ where $n$ is the total density of electrons and $m$ is the free electron mass. The right-hand side of Eq.  is independent of electron-electron interactions, in contrast to the right-hand side of Eq. , where $\langle K\rangle$ will vary with the interaction strengths in the Hamiltonian. The optical conductivity of actual materials incorporates both transitions within the conduction band and those to higher bands, while the $tJ$ model contains only the conduction band. We can look at Eq.  as a partial sum rule [@Baeriswyl; @Tan], which reflects the spectral weight of the free carriers in the conduction band. The experimental quantity we are interested in modeling is the effective density of free carriers, which in actual materials is calculated from the low-frequency spectral weight [@Orenstein], $$\label{eq:10b} n_{\text{free}}(T) = \frac{2m}{\pi e^2} \int_0^{\omega_0} d\omega\,\sigma_1(\omega,T)\,.$$ For high $T_c$ materials, the cutoff frequency is typically chosen around $\hbar \omega_0 \approx 1$ eV so as to include only intraband transitions. For comparison with the $tJ$ model, we identify the right-hand side of Eq.  with $$\label{eq:10b} \begin{split} \frac{\pi e^2}{2} \langle K \rangle &= \frac{\pi e^2 n_{\text{free}}(T)}{2m}\,. \end{split}$$ The superfluid weight satisfies the inequality [@Paramekanti] $$\label{eq:41} \frac{n_s}{m^\ast} \le \langle K \rangle = \frac{n_{\text{free}}}{m}\,,$$ which is obeyed in our results in Fig.2.   The superfluid weight graphs at the sampled temperatures show a clear bipartite structure, with a peak at low $\langle n_i \rangle$, and another peak at high $\langle n_i \rangle$ (which develops into two closely spaced peaks at lower temperatures). In between these is a region of low superfluid weight, with a minimum near $\langle n_i\rangle \simeq 0.385$, approximately independent of temperature. Looking at the nearest-neighbor density-density correlation $\langle n_i n_j \rangle$ as shown in Fig. 3, we see that $\langle n_i \rangle \simeq 0.385$ is also the electron density separating two different regimes of the system: an extremely dilute regime, where $\langle n_i n_j \rangle \simeq 0$, and a partially-to-densely filled regime, where $\langle n_i n_j \rangle > 0$. It is therefore useful to discuss the superfluid weight and kinetic energy results in terms of these two regimes. ### Extremely dilute regime, $\langle n_i\rangle \lesssim 0.385$ The system in this regime is a dilute gas of electrons. For low $\langle n_i \rangle$, the kinetic energy per bond $\langle K \rangle \simeq 2 \langle n_i \rangle$, which follows if the density of free carriers is just the density of electrons, $n_\text{free} = \langle n_i \rangle$, and the mass of the carriers $m = 1/2$. The interaction terms in the $tJ$ Hamiltonian create an attractive potential of strength $-\tilde{J}$ between electrons in singlet-states on neighboring sites. For a coupling $J/t = 0.444$, this attraction is too weak to form two-body bound states, but since we are in three dimensions, even a weak attractive potential is sufficient for the formation of an electron superfluid at low temperatures [@Emery; @Randeria]. In fact, we see a peak in $n_s/m^\ast$ develop around $\langle n_i \rangle \approx$ 0.3–0.35, and this peak grows as the temperature is lowered from $1/t = 0.315$ to $0.1$. For low $\langle n_i\rangle$, the superfluid weight increases with electron density and $\langle K \rangle$. The location of the peak in $n_s/m^\ast$ is just before $\langle K \rangle$ comes to its maximum and levels off. As the density of free carriers saturates near $\langle n_i \rangle \simeq 0.385$, there is a sharp drop in $n_s/m^\ast$, and $\langle n_i n_j \rangle$ begins to increase from zero. At this density the physical characteristics of the system abruptly change, without however inducing a phase transition. ### Partially-to-densely filled regime, $\langle n_i \rangle\gtrsim 0.385$ For intermediate densities $\langle n_i \rangle \approx$ 0.385–0.63, the kinetic energy $\langle K \rangle$ remains approximately constant. Near $\langle n_i \rangle \simeq 0.63$, there is a phase transition to a densely filled phase (either $D$ or $A$). We go from a physical picture where the carriers are electrons in a mostly empty background to one where the carriers are holes moving in a mostly filled background. These holes condense into a superfluid at lower temperatures, and the peak in $n_s/m^\ast$ occurs in the vicinity of the dilute-dense narrow first-order phase transition. For $1/t \lesssim 0.16$, the maximum superfluid weight is reached inside the $\tau$ phase. In the densely filled regime, $\langle n_i \rangle \gtrsim 0.63$, the kinetic energy goes linearly as $\langle K \rangle \simeq 2 (1-\langle n_i \rangle) = 2\delta$, as expected if the free carriers are holes. For hole-doped high-$T_c$ materials, the density of free carriers increases with $\delta$ until the doping level optimal for superconductivity is reached, and remains approximately constant in the overdoped regime.[@Puchkov] The superfluid weight, in contrast, peaks near optimal doping and sharply decreases with overdoping. These trends are reproduced in our numerical results, identifying, from our calculated $n_s/m^\ast$ maxima, the optimal doping for the $tJ$ model as $\delta \approx$ 0.32–0.37, the range of densities where the $\tau$ phase occurs. Note that optimal doping for high-$T_c$ materials is lower than this, typically around $\delta = 0.15$, and the closely spaced double-peak structure of $n_s/m^\ast$ at low temperatures near optimal doping is not observed. On the other hand, our approximation for the $d=3$ $tJ$ model is closer to experiment in this respect than earlier numerical studies of the $tJ$ model, which focused mostly on finite-cluster techniques applied to the $d=2$ system [@DagottoRev]. In these earlier studies optimal doping is identified near $\langle n_i \rangle = 0.5$ on the basis of d-wave pairing correlations and the peak in the superfluid weight [@DagottoRiera]. Also in these earlier studies, the kinetic energy has a maximum at $\langle n_i \rangle = 0.5$, but, unlike experiments, does not saturate with overdoping [@Dagotto2]. To complete the description of the superfluid weight in this regime, in Fig. 4 we show $n_s/m^\ast$ as a function of temperature $1/t$ at various electron densities $\langle n_i \rangle$. For systems with small to optimal hole dopings, shown in Fig. 4(a), there is a clear onset temperature near $1/t \simeq 0.2$ below which the superfluid weight rises rapidly, until it levels off near zero temperature. This behavior is in good comparison with experimental results with YBa$_2$Cu$_3$O$_{6+x}$ [@Lemberger]. As we move past optimal doping to the overdoped systems of Fig. 4(b), we see a marked change in behavior, with the low temperature $n_s/m^\ast$ suppressed.  Specific Heat -------------   Since the superfluid weight peaks inside the $\tau$ phase at low temperatures, it is interesting to check whether the $\tau$ region has any other general characteristics of a superconducting phase. We have added a magnetic field spin coupling term to the $tJ$ Hamiltonian and have shown that the $\tau$ phase continues to exist when $H\ne 0$, up to a critical field $H_c(T)$, which decreases with increasing temperature and goes to zero at the temperature of the $\tau$ phase boundary. In our present study, we look at the spectrum of excitations of the system through the specific heat per bond $$\label{eq:42} C(n,T) = \left. \frac{\partial \langle H(i,j) \rangle}{\partial T} \right|_n\,,$$ calculated for $\phi =0$. If the $\tau$ phase corresponds to the superconducting phase in real materials, we should see evidence of a gap in the excitation spectrum. The results for $C$ as a function of temperature $1/t$ are plotted in Fig. 5 for a series of different electronic densities $\langle n_i \rangle$. Starting at $\langle n_i \rangle = 0.9995$, the smallest hole doping shown in Fig. 5(a), we observe a broad peak around $1/t \simeq 0.33$, corresponding to $k_B T \simeq 0.75 \tilde{J}$. We can identify this peak with the thermal excitation of the spin degrees of freedom. As we dope the system with holes, the weight under the curve at lower temperatures increases due to excitation of charge degrees of freedom. As we approach optimal doping, a second peak develops around $1/t \simeq 0.2$. Note that this approximately coincides with the onset temperature below which we see a dramatic increase in $n_s/m^\ast$ in Fig. 4. The spin-excitation peak is also enhanced for $\langle n_i \rangle\approx$ 0.65–0.75, which is related to the appearance of an antiferromagnetic island around $1/t \simeq 0.3$ in that density range. The peak at $1/t \simeq 0.2$ grows rapidly near optimal doping, reminiscent of the specific heat anomaly of high-$T_c$ materials [@Loram1; @Loram2]. For $\langle n_i \rangle = 0.655$ we see the appearance of two subsidiary peaks below the main one at $1/t \simeq 0.2$. These smaller peaks may be related to the complicated lamellar structure of A and D regions above the $\tau$ phase boundary. For temperatures $1/t \lesssim 0.16$, inside the $\tau$ phase, the specific heat is strongly suppressed, reflecting the opening up of a gap in the excitation spectrum. We can see this gap more directly by looking at the low-temperature limit of the specific heat. Quasiparticle excitations contribute a linear term to the specific heat $C \simeq \gamma T$ for small $T$. In Fig. 6, we plot $\gamma = C/T$ as a function of electron density at a low temperature, $1/t = 0.015$. The specific heat coefficient $\gamma \simeq 0$ in the A phase near half-filling, but then grows with increasing hole doping. At the onset of the $\tau$ phase a gap opens in the quasiparticle spectrum, $\gamma$ falls sharply, and stays small until it rises again near the phase boundary. Qualitatively, this doping-dependence of the low-temperature specific heat coefficient agrees well with the experimental results for high-$T_c$ superconducting materials [@Loram1]. Conclusions =========== We have developed a position-space renormalization-group approximation to study the superfluid weight of the three-dimensional $tJ$ model. Our results indicate that optimal hole doping for this system occurs in the density range of the $\tau$ phase, $\langle n_i \rangle \approx $ 0.63–0.68, where $n_s/m^\ast$ reaches a local maximum. While the superfluid weight drops off sharply in the overdoped region, the density of free carriers, proportional to the kinetic energy, remains approximately constant, as seen experimentally in high-$T_c$ materials. From calculations of the specific heat coefficient $\gamma$, we see clear evidence of a gap in the excitation spectrum for the $\tau$ phase. Earlier renormalization group studies [@FalicovBerkerT; @FalicovBerker] had suspected that the $\tau$ phase corresponds to the superconducting phase of high-$T_c$ materials, and this idea was reinforced when an analogous phase was discovered in the Hubbard model [@Hubshort; @Hubnew]. Our present results justify this suspicion. This research was supported by the U.S. Department of Energy under Grant No. DE-FG02-92ER-45473, by the Scientific and Technical Research Council of Turkey (TÜBITAK) and by the Academy of Sciences of Turkey. MH gratefully acknowledges the hospitality of the Feza Gürsey Research Institute and of the Physics Department of Istanbul Technical University. Y.J. Uemura, G.M. Luke, B.J. Sternlieb, J.H. Brewer, J.F. Carolan, W.N. Hardy, R. Kadono, J.R. Kempton, R.F. Kiefl, S.R. Kreitzman, P. Mulhern, T.M. Riseman, D.L. Williams, B.X. Yang, S. Uchida, H. Takagi, J. Gopalakrishnan, A.W. Sleight, M.A. Subramanian, C.L. Chien, M.Z. Cieplak, G. Xiao, V.Y. Lee, B.W. Statt, C.E. Stronach, W. J. Kossler, and X.H. Yu, Phys. Rev. Lett. [**62**]{}, 2317 (1989). C. Bernhard, C. Niedermayer, U. Binninger, A. Hofer, C. Wenger, J.L. Tallon, G.V.M. Williams, E.J. Ansaldo, J.I. Budnick, C.E. Stronach, D.R. Noakes, and M.A. Blankson-Mills, Phys. Rev. B [**52**]{}, 10488 (1995). C. Niedermayer, C. Bernhard, U. Binninger, H. Gl[ü]{}ckler, J.L. Tallon, E.J. Ansaldo, and J.I. Budnick, Phys. Rev. Lett. [**71**]{}, 1764 (1993). J.–P. Locquet, Y. Jaccard, A. Cretton, E.J. Williams, F. Arrouy, E. M[ä]{}chler, T. Schneider, Ø. Fischer, and P. Martinoli, Phys. Rev. B [**54**]{}, 7481 (1996). C. Bernhard, J.L. Tallon, T. Blasius, A. Golnik, and C. Niedermayer, Phys. Rev. Lett. [**86**]{}, 1614 (2001). A.V. Puchkov, P. Fournier, T. Timusk, and N.N. Kolesnikov, Phys. Rev. Lett. [**77**]{}, 1853 (1996). S.-S. Lee and S.-H.S. Salk, Phys. Rev. B [**64**]{}, 052501 (2001). S.-S. Lee and S.-H.S. Salk, cond-mat/0212582. E. Dagotto and J. Riera, Phys. Rev. Lett. [**70**]{}, 682 (1993). A. Falicov and A.N. Berker, Phys. Rev. B [**51**]{}, 12458 (1995). A. Falicov and A.N. Berker, Turk. J. Phys. [**19**]{}, 127 (1995). G. Migliorini and A.N. Berker, Eur. Phys. J. B [**17**]{}, 3 (2000). M. Hinczewski and A.N. Berker, cond-mat/0503226. M.E. Fisher, M.N. Barber, and D. Jasnow, Phys. Rev. A [**8**]{}, 1111 (1973). D.J. Scalapino, S.R. White, and S. Zhang, Phys. Rev. B [**47**]{}, 7995 (1993). M. Suzuki and H. Takano, Phys. Lett. A [**69**]{}, 426 (1979). H. Takano and M. Suzuki, J. Stat. Phys. [**26**]{}, 635 (1981). A.A. Migdal, Zh. Eksp. Teor. Fiz. [**69**]{}, 1457 (1975) \[Sov. Phys. JETP [**42**]{}, 743 (1976)\]. L.P. Kadanoff, Ann. Phys. (N.Y.) [**100**]{}, 359 (1976). A.N. Berker, S. Ostlund, and F.A. Putnam, Phys. Rev. B [**17**]{}, 3650 (1978). A.N. Berker and M. Wortis, Phys. Rev. B [**14**]{}, 4946 (1976). M. Imada, A. Fujimori, and Y. Tokura, Rev. Mod. Phys. [**70**]{}, 1039 (1998). L. Tan and J. Callaway, Phys. Rev. B [**46**]{}, 5499 (1992). D. Baeriswyl, C. Gros, and T.M. Rice, Phys. Rev. B [**35**]{}, 8391 (1987). J. Orenstein, G.A. Thomas, A.J. Millis, S.L. Cooper, D.H. Rapkine, T. Timusk, L.F. Schneemeyer, and J.V. Waszczak, Phys. Rev. B [**42**]{}, 6342 (1990). A. Paramekanti, N. Trivedi, and M. Randeria, Phys. Rev. B [**57**]{}, 11639 (1998). V.J. Emery, S.A. Kivelson, and H.Q. Lin, Phys. Rev. Lett. [**64**]{}, 475 (1990). M. Randeria, J.-M. Duan, and L.-Y. Shieh, Phys. Rev. Lett. [**62**]{}, 981 (1989). E. Dagotto, Rev. Mod. Phys. [**66**]{}, 763 (1994). E. Dagotto, A. Moreo, F. Ortolani, D. Poilblanc, and J. Riera, Phys. Rev. B [**45**]{}, 10741 (1992). Y. Zuev, M.S. Kim, and T.R. Lemberger, cond-mat/0410135. J.W. Loram, K.A. Mirza, J.R. Cooper, and W.Y. Liang, Phys. Rev. Lett. [**71**]{}, 1740 (1993). J.W. Loram, K.A. Mirza, J.M. Wade, J.R. Cooper, and W.Y. Liang, Physica C [**235-240**]{}, 134 (1994).

ArXiv

--- abstract: 'Generative adversarial networks have led to significant advances in cross-modal/domain translation. However, typically these networks are designed for a specific task (e.g., dialogue generation or image synthesis, but not both). We present a unified model, M$^{3}$D-GAN, that can translate across a wide range of modalities (e.g., text, image, and speech) and domains (e.g., attributes in images or emotions in speech). Our model consists of modality subnets that convert data from different modalities into unified representations, and a unified computing body where data from different modalities share the same network architecture. We introduce a universal attention module that is jointly trained with the whole network and learns to encode a large range of domain information into a highly structured latent space. We use this to control synthesis in novel ways, such as producing diverse realistic pictures from a sketch or varying the emotion of synthesized speech. We evaluate our approach on extensive benchmark tasks, including image-to-image, text-to-image, image captioning, text-to-speech, speech recognition, and machine translation. Our results show state-of-the-art performance on some of the tasks.' author: - | Shuang Ma\ SUNY Buffalo\ Buffalo, NY\ [[email protected]]{} - | Daniel McDuff\ Microsoft Research\ Redmond, WA\ [[email protected]]{} - | Yale Song\ Microsoft Cognition\ Redmond, WA\ [[email protected]]{} bibliography: - 'egbib.bib' title: 'M$^{3}$D-GAN: Multi-Modal Multi-Domain Translation with Universal Attention' --- ![We present a unified model that can translate across multiple modalities (synthesize text, images or audio from text, images or audio) and multiple domains (synthesize a diverse set of examples from a single source input with different attributes) in a controllable fashion.[]{data-label="fig:summary"}](Summary.PNG){width="1\linewidth"} Introduction ============ Generative adversarial networks [@GANs] learn a mapping from source to target distributions, and have shown great performances in data translation tasks involving different modalities such as image, text, audio. These problems are typically multi-modal (e.g., text-to-speech or text-to-image), and the mappings are inherently one-to-many (e.g., the same sentence said with different emotions should sound different). Thus, it would be ideal if a model is able to learn a mapping across multiple *modalities* that also allows for the *domain* to be explicitly controlled. We define each *modality* as a set of data of the same type, and each *domain* as a set of data with the same attribute value. For example, in Fig \[fig:summary\], text would be one modality, images another and audio yet another. Images captured at night represent one domain and images captured in the daytime represent another domain. A unified architecture would make the modeling more efficient and allow representations to be shared, even if they come from different domains or modalities. In this work, we introduce a multi-modal multi-domain generative adversarial network (M$^{3}$D-GAN) – a unified model that can translate across a wide range of modalities (e.g. image, text and speech) and domains (e.g., styles and attributes). We specifically focus on the ability to explicitly control the domain aspect, rather than randomly generating results as in the previous work [@Bicycle-GAN]. Creating such an architecture is non-trivial for several reasons: (1) Different models are designed for different data types; it is non-trivial to extend them across different modalities (e.g., a network designed for text-to-image translation would not work for speech-to-text translation). (2) The mode collapse issue widely observed in GANs makes it difficult to produce diverse results. (3) The alternating training process in GANs makes it difficult to explicitly control the domain aspect in the synthesized output. Our M$^{3}$D-GAN consists of modality-specific sub-networks and a unified computing body. The former converts input samples from different modalities into unified representations. The majority of computation is done in the unified computing body, where the representations from all modalities share the same network (but not the weights). We achieve the *cross-modality* property by letting different tasks from the same modality share their modality subnets, which avoids creating networks for every tasks. For example, all image translation tasks share the same image-modality-net, no matter which image domains they come from (e.g., shoes or scenes). This encourages generalization across tasks and makes it easy to add additional tasks. To produce diverse results, we introduce a *universal attention module* that encodes multi-domain distributions in a highly structured latent space by means of information bottleneck [@tishby2000information]. At inference time, we synthesize diverse samples by providing a condition, or randomly sampling from the latent space induced by this module. In summary, our contributions are: (1) We propose multi-modal multi-domain GAN (M$^3$D-GAN) that can translate data across multiple modalities and domains. (2) We introduce a universal attention module that encodes multi-domain variations in a latent space. (3) We evaluate our approach on a broad range of tasks and achieve superior results compared to baseline models. ![Results of our model on image-to-image translation given an input image (left). On the right we show reference images (top), reference-conditioned synthesis results (middle), and unconditional synthesis results. []{data-label="fig:paris"}](paris.jpg){width="1\linewidth"} {width="1.0\linewidth"} Related Work ============ **Cross-Domain Synthesis:** In the image modality, conditioning GANs with labels has been used to generate higher resolution images [@odena2017conditional]. An important aspect of synthesis is the ability to frame the task as the description of attributes or features [@attribute]. Modeling images in layers (i.e, foreground and background) allows for explicit control over the image attributes [@attribute]. Pix2pix [@pix2pix] made the first attempt at translating across different image domains by training on paired data (e.g., sketches to photos). Since then, more image translation work has used unsupervised training [@cycleGAN; @cross-domain-image-generation; @DA-GAN; @distance-GAN]. However, most GAN architectures require retraining for each specific tasks. Bridging the gaps between domains, Choi et al., proposed StarGAN [@choi2017stargan], which is a unified model for image-to-image translation that produced compelling results for controllable image generation, enabling cross-domain image synthesis. **Cross-Modality Synthesis**: GANs have had a significant impact on data synthesis cross modalities, from image-to-text to speech synthesis. Text-to-image generation is an example of cross-domain synthesis [@stackgan; @text2image_reed; @DA-GAN]. Image captioning can be considered as an inverse process of generating text from images . Analogously, speech recognition aims to turn a segment of audio into a textual string and similarly deep networks have been successfully employed for this [@zhang2017very; @xiong2018microsoft]. Text-to-speech synthesis presents the challenge of modeling the style in which the speech is synthesized [@ma-iclr19]. To this end, the recently proposed Tacotron-based approaches [@Tacotron; @DBLP:conf/icml/Wang-GST_Tacotron] use a piece of reference speech audio to specify the expected style. This approach means a single textual string can be synthesized in numerous ways (a one to many mapping). Controlling the attributes of images [@choi2017stargan; @attribute] is somewhat similar to this, ideally we would want to control the domain of the synthesized data. Approach ======== Neuroscience has revealed that the brain forms unified abstract representations from cross-sensory modalities [@pietrini2004beyond; @giard1999auditory]. For instance, the same mirror neurons are shown to fire when primates observe actions and hear sounds associated with them [@kohler2002hearing; @giard1999auditory]; visual and tactile recognition uses similar processes [@pietrini2004beyond]. While there is much we do not understand about the nature of how the brain processes multi-sensory signals, we were inspired by this property in the design of our network. Considering the differences between human senses and perception, our approach consists of a set of modality sub-networks, designed to convert input data into unified representations somewhat analogous to the senses, and a unified computing body to transfer these representations, inspired by the abstracted notion of perception. Given two data distributions, source $\mathbf{S}$ and target $\mathbf{T}$, we aim to learn a mapping that is robust to different data types (multi-modal) and to find diverse and plausible $t \in \mathbf{T}$ that correspond to $s \in \mathbf{S}$ (multi-domain). We denote the modality subnets as $ \mathbf{M}^j = \{\mathbf{M}_{in}, \mathbf{M}_{out}\}^j$, $j$ = {text, image, speech}. $\mathbf{M}_{in}^j$ are prenets, where samples from different modalities are converted to unified representations, $\mathbf{M}_{in}(\mathbf{S}) \to S$ and $\mathbf{M}_{in}(\mathbf{T}) \to T$. The unified computing body learns to relate and transfer the representations from the $S$ to $T$. $\mathbf{M}_{out}^j$ is a modality-specific generator, where the outputs from the unified computing body are fed into it and are translated to the desired data modality. During training, we have a collection of paired samples ${(s \in \mathbf{S}, t \in \mathbf{T})}$. However, the training dataset usually contains only one such pair. Therefore, we introduce a universal attention module $E_{att}$ to encode diverse domain information from the target distribution into a latent space. At test time, M$^{3}$D-GAN can produce diverse results by either explicitly providing a reference sample as a conditioning variable or randomly sampling from the latent space. For a simple notation, in this paper, we denote the domain-specific generator $\mathbf{M}_{out}^j$ as $G$. ![An illustration of our universal attention module.[]{data-label="fig:atten_module"}](attention_module.png){width="1\linewidth"} Universal Attention Module {#sec:attention} -------------------------- Figure \[fig:atten\_module\] shows an illustration of our universal attention module. We aim to model a variety of domain information from the target distribution; during training, rather than starting only from a source sample (e.g. pix2pix [@pix2pix]), we also take samples from the target distribution as references, denoted by $R$, i.e. $\{S,R\} \to \hat{T}$. When training on paired data, e.g. {text, image}, $R$ is the ground-truth target. For testing, $R$ can be any sample from $\mathbf{T}$ and specifies the domain we wish to synthesize. To this end, we design $E_{att}$ that consists of a reference encoder ($Enc_r$) and a universal token layer ($UTL$). They are jointly trained with the whole model and do not require any explicit labels. Given a reference $R$, the encoder $Enc_r$ compresses domain information into a fixed-length vector; we call this a *reference embedding*. This embedding is used as a query vector in the universal token layer $UTL$, which consists of a bank of *token embedding* and an attention layer, where the token embedding are randomly initialized. We use the attention layer to learn the similarity between the reference embedding and each of the tokens. This produces a set of weights that represent the contribution of each token. The weighted sum of these token embedding, which we call *domain embedding*, is used as the encoded latent code $z$ for generation. The bank of token embedding is shared across all training sequences. Note that this whole process can be understood as information bottleneck [@tishby2000information], which allows our model to learn a highly structured latent space. Training Objectives {#sec:objective} ------------------- **cVAE-GAN**: \[sec:gan\] When taking both source $S$ and reference $R$ samples as input, it is natural to use conditional VAE-GANs as a learning objective: [@mirza2014conditional]: $$\begin{split} \mathcal{L}_{GAN}=\mathbb{E}_{S,T \in p(S,T)}[log(D(S, T))] \\ + \mathbb{E}_{S, T \in p(S,T), z \sim E_{att}(R)}[log(1-D(S, G(S, z)))] \end{split}$$ A reconstruction loss is also adopted between the output and the ground truth: $$\mathcal{L}_{rec}=\mathbb{E}_{S,T \sim p(S,T), z \sim E_{att}(R)} \left\| \mathbf{T}, G(S,z) \right\|$$ Further, we encourage the latent distribution, encoded by $Enc_{r}$, to be close to a random Gaussian. This allows us to randomly sample a latent code as the reference embedding when there is no references $R$ at testing time. $$\mathcal{L}_{kl}=E_{R \sim p(R)} \mathcal{D}_{kl}Enc_r(R) \vert \vert \mathcal{N}(0, I)$$ The full objective for conditional VAE-GAN is: $$\mathcal{L}_{GAN}^{VAE}= \mathop{\min}\limits_{G, Enc_{att}}\mathop{\max}\limits_{D} {\mathcal{L}_{GAN} + \lambda_{rec}\mathcal{L}_{rec} + \lambda_{kl}\mathcal{L}_{kl}}$$ This process is shown in Figure \[fig:pipeline\] (yellow lines). {width="1.0\linewidth"} **Latent Regression**: We find that, by training the model only on the conditional VAE-GAN objective, the synthesized samples tend to be very similar even with different references. In other words, the latent codes output from $Enc_r$ are ignored by the model. In this way, when training jointly with the whole model but without any constraints, the universal attention module (includes $Enc_r$ and $UTL$) is hard to optimize. This issue has been pointed out in [@VQ-VAE], and named “posterior collapse”. We assume that if using $Enc_r$ to encode a synthesized sample $\hat{T}$, then the output latent code $\hat{z}$ should be highly correlated with the one used to synthesize itself, i.e. $ (S, z) \to \hat{T} \to \hat{z}$, $\|z, \hat{z}\|$. Thus, a solution would be to directly apply regression on the latent code $z$ that is encoded from reference $R$, i.e. $R \to z, (S, z) \to \hat{T} \to \hat{z}$, $\|z \sim Enc_r(R), \hat{z}\|$. However, if $Enc_r$ collapses, i.e., it encodes any input to the same or similar latent codes, the reconstructed code will still be the same and $Enc_r$ will not be trained in an optimal fashion. Therefore, rather than using a latent code that is encoded from the reference sample, i.e. $z \sim Enc_r(R)$, we start the latent regression process via a randomly drawn latent code $z \sim \mathcal{N}(0, I)$ and attempt to recover it, i.e. $z \sim \mathcal{N}(0, I), (S, z) \to \hat{T} \to \hat{z}$, $\|z \sim \mathcal{N}(0, I), \hat{z}\|$. We denote the randomly sampled latent code as $z_s$, and the one encoded from $R$ as $z_r$. The samples generated by $z_s$ as $\hat{T}_{sam}$, the ones generated by $z_r$ as $\hat{T}_{enc}$. Thus the latent regression loss $\mathcal{L}_{lat}$ is: $$\mathcal{L}_{lat} = \mathbb{E}_{S \sim p(S), z_s \sim \mathcal{N}(0, I)} \| z_s-Enc_r(G(S,z_s))\|_1$$ It is natural to also use the discriminator loss on $\hat{T}_{sam}$. The full loss can be written as: $$\mathcal{L}_{GAN}^{lat}=\mathop{\min}\limits_{G, Enc_{att}}\mathop{\max}\limits_{D}{\mathcal{L}_{GAN} + \lambda_{lat}\mathcal{L}_{lat}}$$ This process is shown in Figure \[fig:pipeline\] (green lines). **Distance Regularization** To further enforce the model to produce effective latent codes and prevent the model from collapsing, we propose a regularization on the generator that directly penalizes the mode collapse behavior. Our regularization is inspired by [@distance-GAN] which showed that the distance between a pair of samples should be highly related before and after translation. Thus, we design our regularization as: $$\label{eq:distance} \begin{split} \mathop{\max}\limits_{G}\mathcal{L}_{dist}=\mathbb{E}_{z_s,z_r}\left[ \frac{\left\|G(S, z_s)-G(S, z_r)) \right\| }{\left\| z_s-z_r \right\|} \right ] \end{split}$$ where $\left \| \cdot \right \|$ indicates a norm. This regularization will penalize when the generator collapses into few modes, and thus force the generator not to ignore the latent code and produce diverse outputs. Our full objective function can be written as: $$\mathcal{L}_{all}=\mathop{min}\limits_{G, E_{att}}\mathop{max}\limits_D \lambda_{1}\mathcal{L}_{GAN}^{VAE} + \lambda_{2} \mathcal{L}_{GAN}^{lat} + \lambda_{3} \mathcal{L}_{dist}$$ Network Architecture ==================== **Reference Encoder** $Enc_r$ $\&$ **universal token layer** $UTL$: The input to $Enc_r$ is first passed through a stack of four 2-D convolutional layers with \[64, 64, 128, 128\] channels. To adapt to sequential input, we make the output tensor to preserve the time resolution and feed it into a single-layer 128-unit GRU. The last GRU state serves as the reference embedding, which is then fed to the universal token layer. We use 10 tokens in our experiments, which we found sufficient to represent a small but rich variety of domain information in the training data. We use multihead attention [@vaswani2017attention] to compute the attention weights. It uses a softmax activation to output a set of combination weights over the tokens; the resulting weighted combination of these tokens is the domain embedding. **modality subnets (Prenet)** $\mathbf{M}_{in}$: We process speech data via mel-spectrograms which enables our model to take both images and speech via a two-layer fully 2D-convolutional network with dimension \[32, 32\]. For text input, we feed a sequence of 128-D character level embeddings into two fully-connected layers with \[256, 128\] units. The output is fed into a CBHG unit [@Tacotron] which has a Conv1D bank with 16 layers, and each layer has 128 units. After the residual connection, there are four 128 unit fully-connected layers with ReLU. The final Bidirectional GRU has 128 cells. **modality subnets (Generator)** $\mathbf{M}_{out}$: As speech and text are both sequential signals, we design the same generator architecture for them. It takes combined $S$ and $z$ as input. To match the dimension, we first replicate the single $z$ T times, and concatenat it with the input source sequence. The Attention RNN consists of 2-layer residual GRUs with 256 cells. The Decoder RNN has a 256 cell 1-layer GRU. When synthesizing speech, this generator directly predicts the mel-spectrogram. We use Griffin-Lim [@griffin-lim] as a vocoder. To process image data, six residual conv2D layers compress the input into a low dimension representation, and then six residual deconv2D layers are used for decoding. Finally, a deconv2D layer outputs 3D RGB images. Figure \[fig:network-architecture\] shows detailed diagrams for each module. We train our M$^{3}$D-GAN from scratch using the Adam-optimizer. For the task of image $\to$ image translation, we train our model for 30 epochs with a batch size of 1. For other tasks, we train our model with a batch size of 32. For text $\to$ image, and image $\to$ text, we train our model for 300 epochs. Text $\to$ speech and text $\to$ text are trained for 200k steps. Details can be found in the supplementary material. Experiments =========== We test the proposed approach on six synthesis tasks and compare (quantitatively and qualitatively) to baseline methods in each case. The datasets we used, and baseline methods we compared against, are listed in Table \[tab:tasks\]. Dataset Baseline Methods --------- --------------------------------- CNN-RNN-coco DeepSpeech2 [@DBLP:DeepSpeech] JointRNN [@machien-translation] : The datasets and baseline methods for each task. \[tab:tasks\] Qualitative Evaluation {#sec:quality} ---------------------- **Image $\to$ Image**: We test our model under two scenarios shown in Figure \[fig:pipeline\]. The first is to provide a reference image $R$, i.e., $(S,R) \to \hat{T}$ and synthesize samples that exhibit the content of $S$ and the style of the reference sample. Figure \[fig:paris\] shows the result: When given different references (first row), the synthesized samples preserve the content from the source image (the Eiffel Tower), and show different styles taken from references. The second scenario is to provide a random noise vector $z$ and synthesize a new sample by taking $S$ combined with $z$, i.e. $(S,z) \to \hat{T}$. Figure \[fig:paris\] (bottom row) shows the results: When adding randomly sampled noise vector $z$, we obtain images of the same content with different styles. More results are shown in Figure \[fig:shoes\_noise\] and Figure \[fig:img2img\]. Figure \[fig:shoes\_noise\] shows diverse images are produced when we provide a random noise vector to our model, which suggests that our universal attention module can alleviate the mode collapse problem. Figure \[fig:img2img\], on the other hand shows our model correctly capturing domain information from any given references. For example, when generating shoes from sketches (set (7)), we can see that the sneaker’s white stripes are correctly generated, while the shoes body colors are varied. Also, when given a reference that has a grey strip pattern (the fourth reference in this set), our model synthesizes a white body color with grey stripes. In another set (set (8)), the outside of the high heel shoes are correctly changed to the color of the reference, while the inside material remains the same. These results suggests that the model does take the reference’s domain information, while successfully preserving the content information from the source image. ![Edges $\to$ shoes generation by combining with randomly sampled noise vectors $z$ at testing time. For each row, the first column is the source sketch image.[]{data-label="fig:shoes_noise"}](edges2shoes_noise.jpg){width="0.8\linewidth"} {width="1\linewidth"} ![Text $\to$ Images. For each sentence, we randomly sampled 6 noise vectors for generating.[]{data-label="fig:text2img"}](txt2img.png){width="1\linewidth"} **Text $\to$ Image**: We demonstrate our model’s ability to perform cross-modal translation. We use the CUB-200-2011 dataset [@CUB_200_2011], where each image is paired with 10 sentences describing a bird in the image. We resize input images to have a 256 $\times$ 256 pixel resolution for better visual quality. During training, we randomly sample a sentence and pair it with an image. At inference, we provide a random noise vector along with the input sentence. Figure \[fig:text2img\] shows the results. Given a sentence description, our model produces images that correctly preserve certain characteristics of a bird described by the text (e.g., bird shape) while showing diversity in other attributes (e.g., pose). **Text $\to$ Speech**: We test our model’s ability to generate stylized speech from text inputs. In conventional TTS systems, training on a dataset of (text, audio) pairs, the model is supposed to synthesize an audio clip from a given textual string. The audio needs to convey the correct content of the text, and possess characteristics of the human voice. For adding styles to TTS, at inference, we combine an additional reference audio with a textual string and feed them into the system, where the text provides content, and the reference audio provides the desired style. For example, when synthesizing a sentence ‘The cat is laying on the table’. Given a reference audio signal which has happy emotional tone, the synthesized sample should say the sentence in happy tone. Note that, the reference audio signal can contain any textual content. To this end, the most challenging part is in disentangling the content and style information from the reference audio. We perform experiments to validate that our model can successfully capture domain (style) information from the reference audio. We compare our model to Tacotron-GST [@DBLP:conf/icml/Wang-GST_Tacotron] in how well they model auditory styles. We use an in-house dataset, dubbed EMT-4, that consists of 22,377 American English audio-text samples, with a total of 24 hours. All the audio samples are read by a single speaker in four emotion categories: happy, sad, angry and neutral. For each text sample, there is only one audio sample paired with one of the four emotion styles. We randomly select 15 sentences and 4 references from 4 emotion categories. Each sentence is paired with four reference audio samples for synthesis, producing 60 audio samples. We put the audio results on our [demo](https://researchdemopage.wixsite.com/cvpr-tts-demo/style-transfer) page. To see whether our model can capture domain information from references, we also conduct a “content-style swap” experiment where samples are synthesized by permuting text (content) and audio (style) from four (text, audio) pairs, one from each emotion category. We put the results on our [demo](https://researchdemopage.wixsite.com/cvpr-tts-demo/style-and-content-swapping) page; each column has the same content with different styles, each row has the same style with different content. The results suggests that our model has successfully disentangled content and style components. [.5]{} [rcc]{} Method & ------------------ Realism (acc) svhn $\to$ mnist ------------------ & ------------------- Diversity(LPIPS) edges $\to$ shoes ------------------- \ pix2pix [@pix2pix] & 89% & .013$\pm$.000\ cVAE-GAN [@bao2017cvae] & 86% & .096$\pm$.001\ Bicycle-GAN [@Bicycle-GAN] & 91% & .110$\pm$.002\ **Ours($z$)** & **91%** & **.112$\pm$.001**\ **Ours($r$)** & – & **.115$\pm$.002**\ Rand. Real Imgs. & 98% & .262$\pm$.007\ \[tab:image-to-image\] [.5]{} ----------------------------------------- -------- -------- Method BLEU-1 BLEU-4 CNN+RNN-coco  66.7 23.8 Ours 65.2 23.8 JointRNN (En-Fr) [@machien-translation] – 28.4 Ours – 22.2 ----------------------------------------- -------- -------- \[tab:image-to-text\] [.4]{} Method Inception $\#$miss ---------------------- ----------------- ---------- StackGAN [@stackgan] 3.7$\pm$0.4 36.0 DA-GAN [@DA-GAN] 5.6$\pm$0.4 19.0 Ours **5.2$\pm$0.4** **16.0** \[tab:text-to-image\] [.3]{} Method WER Acc ----------------------------------------- ---------- --------- Tac. [@Tacotron] 10.6 68% GST [@DBLP:conf/icml/Wang-GST_Tacotron] **10.2** 77% Ours **10.2** **80%** \[tab:text-to-speech\] [.3]{} Metric WER -------------------------------- ------ -- DeepSpeech2 [@DBLP:DeepSpeech] 5.15 Ours 7.3 \[tab:speech-to-text\] Quantitative Evaluation {#sec:quantity} ----------------------- **Image $\to$ Image**: We evaluate our model in terms of **realism** and **diversity** when synthesizing images from images. We train a classifier on the MNIST dataset [@mnist] and employ it on the translated samples from SVHN [@svhn]. Table \[tab:image-to-image\] shows that our model achieves comparable results in realism with the state-of-the-art. To evaluate diversity, we compute the diversity-score using the LPIPS metric [@LPIPS]. Taking 100 shoe sketches, we randomly generate 2000 images using each model by adding randomly sampled latent codes ($z$). The average distance between the 2000 samples for each method are reported in Table \[tab:image-to-image\]. Our model (Ours($z$)) produces the highest diversity score. While pix2pix [@pix2pix] has the lowest diversity, potentially due to the mode collapse issue in conditional GANs. We also generate samples by adding references. The diversity score is further improved (see Ours($R$)), illustrating that explicitly controlling the domain variable leads to better results. **Text $\to$ Image**: We use two metrics: the Inception Score [@improved-techniques-for-training-gan] and the number of missing modes (denoted as $\#$ miss). To compute the inception score, we finetune VGG-19 on CUB-200-2011 [@CUB_200_2011]. Table \[tab:text-to-image\] shows the results. Comparing with DA-GAN [@DA-GAN] which requires labels for training, our model produces comparable results even without using any label. For a more rigorous validation, we also adopt the missing mode metric ($\#$ miss) that represents the classifier reported number of missing modes. Comparing with the other two methods, our model misses the least number of modes, which further suggests the superiority of our model in alleviating the mode collapse issue. **Image $\to$ Text (image captioning):** We compare our model with CNN-RNN-COCO, a state-of-the-art image caption method based on . Table \[tab:image-to-text\] shows the results on MS-COCO; our model achieves BLEU-4 in 0.238 for both. Comparing with the model that specifically designed for this task, our model produces comparable results. **Text $\to$ Speech (TTS)** We evaluate two factors in the TTS task, i.e. fidelity (the synthesized speech should contain the desired content in a clearly audible form) and domain transfer accuracy (when transferring across domains, the synthesized speech should correctly correspond to the reference’s domain). We compare with two state-of-the-art stylized TTS models, Tacotron and GST, where Tacotron is an RNN-CNN auto-regressive model trained only on reconstruction loss, and GST improves on Tacotron by incorporating the global style token layers. To validate fidelity, we assess the performance of synthesized samples in a speech recognition task. We use a pre-trained ASR model based on WaveNet [@WaveNet] to compute the Word Error Rate (WER) for the samples synthesized by each model. Table \[tab:text-to-speech\] shows our model performing comparably with the state-of-the-art approaches. As our domains (emotions) are all categorical, we evaluate the performance in domain transfer by means of classification. To this end, we train a classifier on EMT-4, which shows a 98$\%$ accuracy. We then select 1000 samples synthesized from the test set of EMT-4. Table \[tab:text-to-speech\] shows that our model performs the best in terms of the domain transfer accuracy. **Speech $\to$ Text (speech recognition) and Text $\to$ Text (machine translation)** We test our model for the speech recognition task on LibriSpeech corpus constructed from audio books [@LibriSpeech]. For text-to-text, we test our model on the WMT’14 English-French (En-Fr) dataset. We remove sentences longer than 175 words, resulting in 35.5M sentence pairs for training. The source and target vocabulary is based on 40k BPE types. We evaluate with tokenized BLEU on the corpus-level. As seen in Table \[tab:image-to-text\] and Table \[tab:speech-to-text\], comparing with the state-of-the-art speech recognition method, i.e. DeepSpeech [@DBLP:DeepSpeech] and machine translation method, our model did not show outstanding results. We suspect this is because when translating text to text, the reference (text) will be processed as a single domain embedding, which is a global representation, and thus lose essential sequential information. Even though our generator decodes information by time step and works well for other modalities (e.g., speech), we suspect that our text encoding method could be suboptimal. Future work may focus on designing a more robust network architecture for the text modality. Ablation Study {#sec:ablation_study} -------------- [rcc]{} Model & ------------------ Realism (acc) svhn $\to$ mnist ------------------ : Evaluation for each component of M$^{3}$D-GAN. & ------------------- Diversity(LPIPS) edges $\to$ shoes ------------------- : Evaluation for each component of M$^{3}$D-GAN. \ cVAE-GAN [@bao2017cvae] & 86% & .096$\pm$.001\ $\mathcal{L}_{GAN}^{VAE}$ & 87% & .098$\pm$.001\ $\mathcal{L}_{GAN}^{VAE}$ + $\mathcal{L}_{GAN}^{lat}$ & 90% & .113$\pm$.000\ $\mathcal{L}_{all}$ w$\setminus$o Att & 89% & .111$\pm$.001\ $\mathcal{L}_{all}$ & **91 %** & **.115$\pm$.002**\ Rand. Real Imgs. & 98% & .262$\pm$.007\ \[tab:ablation\_study\] We conduct an ablation study on the image-to-image task to analyze each component in our model. We again use the classification accuracy and diversity score as the metric. The results are shown in Table \[tab:ablation\_study\]. - cVAE-GAN [@bao2017cvae] vs. $\mathcal{L}^{VAE}_{GAN}$: When training our model with the conventional conditional VAE-GANs objective [@bao2017cvae], we see improvements in terms of both the realism and diversity. It shows the contribution of our attention module in producing a highly structured latent space, which helps the model generate more realistic and diverse results. - $\mathcal{L}^{VAE}_{GAN}$ vs. $\mathcal{L}^{VAE}_{GAN} + \mathcal{L}^{lat}_{GAN}$: When we add the latent regression, the performance is further improved. - $\mathcal{L}^{VAE}_{GAN} + \mathcal{L}^{lat}_{GAN}$ vs. $\mathcal{L}_{all}$: Comparing with our model with all the proposed losses, the realism of the $\mathcal{L}^{VAE}_{GAN} + \mathcal{L}^{lat}_{GAN}$ baseline is slightly impaired, and the diversity is also lower. It shows that, the distance regularization $L_{dist}$ plays an important role in controlling the learned latent code. - $\mathcal{L}_{all}$ w$\setminus$o Att vs. $\mathcal{L}_{all}$: When discarding the attention module, both realism and diversity are lower than the proposed model. It illustrates that the attention mechanism does help in producing better results. Conclusion ========== We present M$^3$D-GAN for cross-modal cross-domain translation, which consists of modality-specific subnets with a universal attention module that learns to encode modality/domain information in a unified way. We show how the same architecture can be applied to a wide variety of tasks including text-to-image, image-to-text, image-to-image, text-to-speech, speech-to-text and text-to-text translation. The universal attention module we propose can learn a highly structured latent space by means of information bottleneck [@tishby2000information], which allows for explicit control of specific attributes. Leveraging the same architecture for different problems is advantageous as it allows us to develop shared representations in the abstracted unified space. We conduct a comprehensive set of experiments, both qualitative and quantitative, to show that our model achieves strong results across most of the tasks, especially for speech and image. Appendix {#appendix .unnumbered} ======== Implementation details for each task ==================================== The inputs and outputs for each task during the training stage and testing stage are listed in Table \[tab:task\_io\]. - Image$\to$Image: In this task, the source and target are images drawn from two different domains (e.g. day$\to$night, edges$\to$photos, etc.). During training, the references are images drawn from a target distribution, and are paired with the input (i.e. ground truth image). While at testing time reference images, which are used for indicating the desired style, can be provided. - Text$\to$Image: In this task, the source and reference are paired textual strings and images. At testing time, we do not use references to change the images’ style or attribute. Because a randomly sampled image may change the content that is given by the text. For example. when synthesizing an image from a sentence ‘This is a little yellow bird’, we expect the synthesized image to correctly capture the content of this description. If we were to randomly provide an image with a little white bird, the synthesized birds will have a white appearance. This would cause obvious problems when evaluating the synthesized samples. Therefore, in this task, we only perform synthesis by adding a noise vector, which does not harm the content from the input sentence. For a convenient implementation, we directly use preprocessed char-CNN-RNN text embeddings [@reed2016generative] for the CUB-200-2011 dataset. This is a method that was also adopted by StackGAN [@stackgan]. Note that, the preprocessed text embeddings are only used for this task. - Image$\to$Text: In this task, the source and reference are paired images and texts. To process text as a reference, we follow the method used in [@DBLP:journals/corr/ReedASL16] that allows us to obtain a global sentence text embedding. The input to the reference encoder $Enc_r$ is the average hidden unit activation over the sequence, i.e. $\phi(t)=1 \setminus {L\sum_{L}^{i=1}h_i}$, where $h_i$ is the hidden activation vector for the $i-th$ frame and $L$ is the sequence length. - Text$\to$Speech: In this task, we process the speech audio signal as a mel-spectrogram, our model also predicts the mel-spectrogram directly. The predicted mel-spectrogram can be synthesized directly to speech using either the WaveNet vocoder [@WaveNet] or the Griffin-Lim vocoder [@griffin-lim]. In our experiments, we use the Griffin-Lim for fast waveform generation. At testing time, the reference speech audio can be from any person (i.e. it does not necessarily need to be sampled from a subject in our training dataset). In our experiments, we test our model by synthesizing audio from reference audio sampled from our dataset, and audio from web (i.e., not from our dataset). It turns out that, the results synthesized by either these references are equivalent in terms of quality and style consistency. We highly recommend readers to hear some of our results on our [demo](https://researchdemopage.wixsite.com/cvpr-tts-demo/style-transfer) page. - Speech$\to$Text: In this task, the speech audio is also processed as a mel-spectrograms. We use the same method as in the Image$\to$Text task to process the input reference text as a global sentence embedding. For synthesis, the generator takes a mel-spectrogram sequence combined with the domain embedding obtained from text as input. The output is the predicted text, produced as one character at each time step. - Text$\to$Text: In this task, the source and reference are textual strings from two languages (e.g., English and French). The reference text is processed as a global sentence embedding from the prenet and then fed into a universal attention module to obtain the domain embedding. While the source text is output directly from the prenet and is not processed as a global embedding. To combine the domain embedding and the source feature, i.e. $(z_r, S)$, the domain embedding $z_r$ is first replicated with $T$ time steps and then concatenated with $S$, where $T$ is the total time step of sequence $S$. A detailed illustration of this process can be seen in Figure 5(b) in our paper. [|c|c|c|c|c|c|]{} & &\ & & & &\ & & & & $T_{enc}$ & $T_{sam}$\ Image$\to$Image & ------------------------- S: Image R: Image (ground truth) ------------------------- & --------------------- (S,R)$\to$$T_{enc}$ (S,z)$ \to$$T_{sam}$ --------------------- & --------------------------- S:Image R:Image (randomly sample) --------------------------- & $\checkmark$ & $\checkmark$\ Text$\to$Image & ------------------------ S:Text R:Image (ground truth) ------------------------ & ------------- $\setminus$ ------------- & -------- S:Text R:None -------- & $\times$ & $\checkmark$\ Image$\to$Text & ----------------------- S:Image R:text (ground truth) ----------------------- & ------------- $\setminus$ ------------- & --------- S:Image R:None --------- & $\times$ & $\checkmark$\ Text$\to$Speech & ---------------------------- S:Text R:Speech (mel-spectrogram) ---------------------------- & ------------- $\setminus$ ------------- & --------------------------- S:Text R:Speech(randomly sample) --------------------------- & $\checkmark$ & $\checkmark$\ Speech$\to$Text & ---------- S:Speech R:Text ---------- & ------------- $\setminus$ ------------- & ---------- S:Speech R:None ---------- & $\times$ & $\checkmark$\ Text$\to$Text & ----------------------- S:Text R:Text (ground truth) ----------------------- & ------------- $\setminus$ ------------- & -------- S:Text R:None -------- & $\times$ & $\checkmark$\ \[tab:task\_io\] Discussion ========== $\bullet$ Why we use the modality subnet for multiple tasks, and why this makes it easy to add additional tasks. To use the modality sub-net for multiple tasks aims to avoid designing different networks for each task. For example, when we conduct the task of image-to-image and image-to-text translation, the input modality for both these tasks are images. In this case, a single image-subnet could have the potential to produce effective latent code from all image data. When we want to add more tasks which start from image modal, this image-subnet can consequently be used other than grasp a new network. $\bullet$ Differences with Bicycle-GAN [@Bicycle-GAN] Comparing the task of image-to-image translation with Zhu et al [@Bicycle-GAN]. Our key difference is that UTL allows our model to encode a large variety of domain information in a latent space. In addition, the reference encoder learns to produce representative style code from any given references. In such a way, we can explicitly control any desired style for synthesizing by our model, but not just randomly generate diverse results. Detailed descriptions can be found in the testig stage of our paper.

ArXiv

--- abstract: | We study convex polyhedra in three-space that are inscribed in a quadric surface. Up to projective transformations, there are three such surfaces: the sphere, the hyperboloid, and the cylinder. Our main result is that a planar graph $\Gamma$ is realized as the $1$–skeleton of a polyhedron inscribed in the hyperboloid or cylinder if and only if $\Gamma$ is realized as the $1$–skeleton of a polyhedron inscribed in the sphere and $\Gamma$ admits a Hamiltonian cycle. Rivin characterized convex polyhedra inscribed in the sphere by studying the geometry of ideal polyhedra in hyperbolic space. We study the case of the hyperboloid and the cylinder by parameterizing the space of convex ideal polyhedra in anti-de Sitter geometry and in half-pipe geometry. Just as the cylinder can be seen as a degeneration of the sphere and the hyperboloid, half-pipe geometry is naturally a limit of both hyperbolic and anti-de Sitter geometry. We promote a unified point of view to the study of the three cases throughout. address: - 'Department of Mathematics, University of Texas at Austin' - 'Department of Mathematics, Brown University' - 'Department of mathematics, University of Luxembourg' author: - Jeffrey Danciger - Sara Maloni - 'Jean-Marc Schlenker' bibliography: - 'sample2.bib' - 'adsbib.bib' - 'sample1.bib' title: Polyhedra inscribed in a quadric --- [^1] Introduction and results ======================== Polyhedra inscribed in a quadric -------------------------------- According to a celebrated result of Steinitz (see e.g. [@ziegler:lectures Chapter 4]), a graph $\Gamma$ is the $1$–skeleton of a convex polyhedron in $\R^3$ if and only if $\Gamma$ is planar and $3$–connected. Steinitz [@ste_iso] also discovered, however, that there exists a $3$–connected planar graph which is not realized as the $1$–skeleton of any polyhedron inscribed in the unit sphere $S$, answering a question asked by Steiner [@ste_sys] in 1832. An understanding of which polyhedral types can or can not be inscribed in the sphere remained elusive until Hodgson, Rivin, and Smith [@hodgson1992characterization] gave a full characterization in 1992. This article is concerned with realizability by polyhedra inscribed in other quadric surfaces in $\R^3$. Up to projective transformations, there are two such surfaces: the hyperboloid $H$, defined by $x_1^2 + x_2^2 - x_3^2 = 1$, and the cylinder $C$, defined by $x_1^2 + x_2^2 = 1$ (with $x_3$ free). \[def:inscribed\] A convex polyhedron $P$ is [*inscribed*]{} in the hyperboloid $H$ (resp. the cylinder $C$) if $P\cap H$ (resp. $P \cap C$) is exactly the set of vertices of $P$. If a polyhedron $P$ is inscribed in the cylinder $C$, then $P$ lies in the solid cylinder $x_1^2 + x_2^2 \leq 1$ (and $x_3$ free), with all points of $P$ except its vertices lying in the interior. A polyhedron $P$ inscribed in the hyperboloid $H$ could lie in (the closure of) either complementary region of $\RR^3 \setminus H$. However, after performing a projective transformation, preserving $H$ and exchanging the two complementary regions of $\RR^3 \setminus H$, we may (and will henceforth) assume that all points of $P$, except its vertices, lie in the interior of the solid hyperboloid $x_1^2 +x_2^2 - x_3^2 \leq 1$. Recall that a [*Hamiltonian cycle*]{} in is a closed path visiting each vertex exactly once. We prove the following. \[thm:main\] Let $\Gamma$ be a planar graph. Then the following conditions are equivalent: 1. $\Gamma$ is the $1$–skeleton of some convex polyhedron inscribed in the cylinder. 2. $\Gamma$ is the $1$–skeleton of some convex polyhedron inscribed in the hyperboloid. 3. $\Gamma$ is the $1$–skeleton of some convex polyhedron inscribed in the sphere and $\Gamma$ admits a Hamiltonian cycle. The ball $x_1^2 + x_2^2 + x_3^2 < 1$, thought of as lying in an affine chart $\RR^3$ of $\RP^3$, gives the projective model for hyperbolic space $\HH^3$, with the sphere $S$ describing the ideal boundary $\partial_\infty \HH^3$. In this model, projective lines and planes intersecting the ball correspond to totally geodesic lines and planes in $\HH^3$. Therefore a convex polyhedron inscribed in the sphere is naturally associated to a *convex ideal polyhedron* in the hyperbolic space $\HH^3$. Following the pioneering work of Andreev [@Andreev; @Andreev-ideal], Rivin [@riv_ach] gave a parameterization of the deformation space of such ideal polyhedra in terms of dihedral angles. As a corollary, Hodgson, Rivin and Smith [@hodgson1992characterization] showed that deciding whether a planar graph $\Gamma$ may be realized as the $1$–skeleton of a polyhedron inscribed in the sphere amounts to solving a linear programming problem on $\Gamma$. To prove Theorem \[thm:main\], we show that, given a Hamiltonian path in $\Gamma$, there is a similar linear programming problem whose solutions determine polyhedra inscribed in either the cylinder or the hyperboloid. The solid hyperboloid $x_1^2 + x_2^2 - x_3^2 < 1$ in $\RR^3$ gives a picture of the projective model for *anti-de Sitter* (AdS) geometry in an affine chart. Therefore a convex polyhedron inscribed in the hyperboloid is naturally associated to a convex ideal polyhedron in the anti-de Sitter space $\AdS^3$, which is a Lorentzian analogue of hyperbolic space. Similarly, the solid cylinder $x_1^2 + x_2^2 < 1$ (with $x_3$ free) in an affine chart $\RR^3$ of $\RP^3$ gives the projective model for *half-pipe* (HP) geometry. Therefore a convex polyhedron inscribed in the cylinder is naturally associated to a convex ideal polyhedron in the half-pipe space $\HP^3$. Half-pipe geometry, introduced by Danciger [@dan_geo; @dan_age; @dan_ide], is a transitional geometry which, in a natural sense, is a limit of both hyperbolic and anti-de Sitter geometry. In order to prove Theorem \[thm:main\] we study the deformation spaces of ideal polyhedra in both $\AdS^3$ and $\HP^3$ concurrently. By viewing polyhedra in $\HP^3$ as limits of polyhedra in both $\HH^3$ and $\AdS^3$, we are able to translate some geometric information between the three settings. In fact we are able to give parameterizations (Theorems \[thm:main-AdS-angles\], \[thm:main-AdS-metrics\] and Theorem \[thm:main-HP\]) of the spaces of ideal polyhedra in both $\AdS^3$ and $\HP^3$ in terms of geometric features of the polyhedra. This, in turn, describes the moduli of convex polyhedra inscribed in the hyperboloid and the moduli of convex polyhedra inscribed in the cylinder, where polyhedra are considered up to projective transformations fixing the respective quadric. It is these parameterizations which should be considered the main results of this article; Theorem \[thm:main\] will follow as a corollary. Rivin’s two parameterizations of ideal polyhedra in $\HH^3$ {#sec:rivin_param} ----------------------------------------------------------- Rivin gave two natural parameterizations of the space of convex ideal polyhedra in the hyperbolic space $\HH^3$. Let $P$ be a convex ideal polyhedron in $\HH^3$, let $P^*$ denote the Poincaré dual of $P$, and let $E$ denote the set of edges of the $1$–skeleton of $P^*$ (or of $P$). Then the function $\theta \in \RR^E$ assigning to each edge $e^*$ of $P^*$ the dihedral angle at the corresponding edge $e$ of $P$ satisfies the following three conditions: 1. $0 < \theta(e^*) < \pi$ for all edges $e^*$ of $P^*$. 2. If $e_1^*, \ldots, e_k^*$ bound a face of $P^*$, then $\theta(e_1^*) + \cdots + \theta(e_k^*) = 2 \pi$. 3. If $e_1^*, \ldots, e_k^*$ form a simple circuit which does not bound a face of $P^*$, then $\theta(e_1^*) + \cdots + \theta(e_k^*) > 2 \pi$. Rivin [@riv_ach] shows that, for an abstract polyhedron $P$, any assignment of weights $\theta$ to the edges of $P^*$ that satisfy the above three conditions is realized as the dihedral angles of a unique (up to isometries) non-degenerate ideal polyhedron in $\HH^3$. Further the map taking any ideal polyhedron $P$ to its dihedral angles $\theta$ is a homeomorphism onto the complex of all weighted planar graphs satisfying the above linear conditions. This was first shown by Andreev [@Andreev-ideal] in the case that all angles are acute. The second parameterization [@MR1280952] characterizes an ideal polyhedron $P$ in terms of the geometry intrinsic to the surface of the boundary of $P$. The path metric on $\partial P$, called the *induced metric*, is a complete hyperbolic metric on the $N$-times punctured sphere $\Sigma_{0,N}$, which determines a point in the Teichmüller space $\Teich_{0,N}$. Rivin also shows that the map taking an ideal polyehdron to its induced metric is a homeomorphism onto $\Teich_{0,N}$. Two parameterizations of ideal polyhedra in $\AdS^3$ {#sec:intro-AdS-param} ---------------------------------------------------- Anti-de Sitter geometry is a Lorentzian analogue of hyperbolic geometry in the sense that the anti-de Sitter space $\AdS^n$ has all sectional curvatures equal to $-1$. However, the metric is Lorentzian (meaning indefinite of signature $(n-1, 1)$), making the geometry harder to work with than hyperbolic geometry, in many cases. For our purposes, it is most natural to work with the projective model of $\AdS^3$ (see Section \[ads\_background\]), which identifies $\AdS^3$ with an open region in $\RP^3$, and its ideal boundary $\partial_\infty \AdS^3$ with the boundary of that region. The intersection of $\AdS^3$ with an affine chart is the region $x_1^2 + x_2^2 - x_3^2 < 1$ bounded by the hyperboloid $H$. The ideal boundary $\partial_\infty \AdS^3$, seen in this affine chart, is exactly $H$. Let $P$ be a convex ideal polyhedron in $\AdS^3$ with $N$ vertices. That $P$ is ideal means that the closure of $P$ in $\AdS^3 \cup \partial_\infty \AdS^3$ is a polyhedron whose intersection with $\partial_\infty \AdS^3$ is precisely its vertices. That $P$ is convex means that after removing a space-like plane in its complement, $P$ is geodesically convex. Alternatively, $P$ is convex if and only if it is convex in some affine chart of $\RP^3$. Unlike in the hyperbolic setting, there are restrictions (Proposition \[prop:cyclic-order\]) on the positions of the $N$ vertices. Some choices of $N$ vertices on the ideal boundary $\partial_\infty \AdS^3$ do not determine an ideal polyhedron. Roughly, this is because the hyperboloid $H$ has mixed curvature and the convex hull of a collection of vertices on $H$ may contain points both inside and outside of $H$. All facets of $P$ are *spacelike*, meaning the restriction of the AdS metric is positive definite. Therefore, by equipping $\AdS^3$ with a time-orientation, we may sort the faces of $P$ into two types, those whose normal is future-directed, and those whose normal is past-directed. The future-directed faces unite to form a disk (a bent ideal polygon), as do the past-directed faces. The edges which separate the past faces from the future faces form a Hamiltonian cycle, which we will refer to as the *equator* of $P$. A *marking* of $P$ will refer to an identification, up to isotopy, of the equator of $P$ with the standard $N$-cycle graph so that the induced ordering of the vertices is positive with respect to the orientation and time orientation of $\AdS^3$. We let $\AdSPoly = \AdSPoly_N$ denote the space of all marked, non-degenerate convex ideal polyhedra in $\AdS^3$ with $N$ vertices, considered up to orientation and time-orientation preserving isometries of $\AdS^3$. The term *ideal polyhedron in* $\AdS^3$ will henceforth refer to an element of this space. Let $\Sigma_{0,N}$ denote the $N$-punctured sphere. Fix an orientation on $\Sigma_{0,N}$, a simple loop $\gamma$ visiting each puncture once and label the punctures in order along the path. We call the polygon on the positive side of $\gamma$ the *top* and the polygon on the negative side the *bottom* of $\Sigma_{0,N}$. Then, each ideal polyhedron $P$ is naturally identified with $\Sigma_{0,N}$ via the (isotopy class of the) map taking each ideal vertex to the corresponding puncture and the equator to $\gamma$. This identifies the union of the future faces of $P$ with the top of $\Sigma_{0,N}$ and the past faces with the bottom. See Figure \[fig:marking\]. We let $\Graph(\Sigma_{0,N}, \gamma)$ denote the collection of three-connected graphs embedded in $\Sigma_{0,N}$, up to isotopy, each of whose edges connects two distinct punctures and whose edge set contains the edges of $\gamma$. Via the marking, any ideal polyhedron $P$ *realizes* the edges of a graph $\Gamma \in\Graph(N, \gamma)$ as a collection of geodesic lines either on the surface of or inside of $P$. Consider a space-like oriented piece-wise totally geodesic surface in $\AdS^3$ and let $T$ and $T'$ be two faces of this surface meeting along a common edge $e$. We measure the *exterior dihedral angle* at $e$ as follows. The group of isometries of $\AdS^3$ that point-wise fix the space-like line $e$ is a copy of $\OO(1,1)$, which should be thought of as the group of *hyperbolic rotations* or *Lorentz boosts* of the time-like plane orthogonal to $e$. By contrast to the setting of hyperbolic (Riemannian) geometry, $\OO(1,1)$ has two non-compact components. Therefore there are two distinct types of dihedral angles possible, each of which is described by a real number rather than an element of the circle. Let $\varphi$ be the amount of hyperbolic rotation needed to rotate the plane of $T'$ into the plane of $T$. The sign of $\varphi$ is defined as follows. The light-cone of $e$ locally divides $\AdS^3$ into four quadrants, two of which are space-like and two of which are time-like. If $T$ and $T'$ lie in opposite space-like quadrants, then we take $\varphi$ to be non-negative, if the surface is convex along $e$, and negative, if the surface is concave along $e$. If $T$ and $T'$ lie in the same space-like quadrant, we take $\varphi$ to be non-positive, if the surface is convex at $e$, and positive, if the surface is concave at $e$. Therefore, the dihedral angles along the equator of a convex ideal polyhedron $P$ are negative, while the dihedral angles along the other edges are positive. Note that this definition of angle, and in particular the sign convention, agrees with a natural alternative definition in terms of cross-ratios (see Section \[sec:models\]). Let $P^*$ and $E$ be as before. We will show (Proposition \[prop:imageinA-AdS\]) that the function $\theta \in \RR^E$ assigning to each edge $e^*$ of $P^*$ the dihedral angle at the corresponding edge $e$ of $P$ satisfies the following three conditions: (i) $ \theta(e^*) < 0$ if $e$ is an edge of the equator $\gamma$, and $\theta(e^*) > 0$ otherwise. (ii) If $e_1^*, \ldots, e_k^*$ bound a face of $P^*$, then $\theta(e_1^*) + \cdots + \theta(e_k^*) = 0$. (iii) If $e_1^*, \ldots, e_k^*$ form a simple circuit which does not bound a face of $P^*$, and such that exactly two of the edges are dual to edges of the equator, then $\theta(e_1^*) + \cdots + \theta(e_k^*) > 0$. Let $\Gamma \in\Graph(N, \gamma)$. Then, thinking of $\Gamma$ as the $1$–skeleton of an abstract polyhedron $P$, we define $\Angles_{\Gamma}$ to be the space of all functions $\theta \in \RR^E$ which satisfy the above three conditions. Define $\AdSPoly_{\Gamma}$ to be the space of ideal polyhedra in $\AdS^3$ with $1$–skeleton identified with $\Gamma$, and let $\PsiAdS_\Gamma: \AdSPoly_\Gamma \to \Angles_\Gamma$ denote the map assigning to an ideal polyhedron its dihedral angles. All of the maps $\PsiAdS_{\Gamma}$ may be stitched together into one. Let $\Angles$ denote the disjoint union of all $\Angles_{\Gamma}$ glued together along faces corresponding to common subgraphs. Then, we show: \[thm:main-AdS-angles\] The map $\PsiAdS: \AdSPoly \to \Angles$, defined by $\PsiAdS(P) = \PsiAdS_{\Gamma}(P)$ if $P \in \AdSPoly_{\Gamma}$, is a homeomorphism. The equivalence of conditions (H) and (S) in Theorem \[thm:main\] follows directly from this theorem and from Rivin’s theorem (see Section \[sec:rivin\_param\]). Indeed, it is an easy exercise in basic arithmetic to convert any weight function $\theta \in \Angles_{\Gamma}$ into one that satisfies conditions (1), (2), and (3) of Rivin’s theorem. To convert any weight function on the edges of a graph $\Gamma$ that satisfies Rivin’s conditions into a weight function satisfying our conditions (i), (ii), and (iii) (which define $\Angles_{\Gamma}$) is also easy, provided there is a Hamiltonian cycle $\gamma$ in the $1$–skeleton. See Section \[sec:proof\_main\] for the detailed proof. We also give a second parameterization of ideal polyhedra in terms of the geometry intrinsic to their boundaries. Here we parameterize the space $\overline{\AdSPoly}_N = \AdSPoly_N \cup \poly_N$ of all marked polyhedra with $N$ vertices including both the non-degenerate polyhedra $\AdSPoly_N$ and the degenerate (or collapsed) polyhedra, parameterized by the space $\poly_N$ of marked ideal polygons in $\HH^2$ with $N$ vertices. Any space-like plane in $\AdS^3$ is isometric to the hyperbolic plane $\HH^2$. Therefore similar to the setting of hyperbolic $3$-space, the path metric on the surface of $P$ is a complete hyperbolic metric on the $N$-times punctured sphere $\Sigma_{0,N}$ determining a point in the Teichmüller space $\Teich_{0,N}$, again called the *induced metric*. We show the following result: \[thm:main-AdS-metrics\] The map $\Phi: \overline{\AdSPoly}_N \to \Teich_{0, N}$, taking a convex ideal polyhedron $P$ in $\AdS^3$ to the induced metric on $\partial P$, is a diffeomorphism. The (weaker) local version of this theorem is a crucial ingredient in proving Theorem \[thm:main-AdS-angles\]. Before continuing on to half-pipe geometry and the cylinder, let us make two remarks about potential generalizations of Theorems \[thm:main-AdS-angles\] and \[thm:main-AdS-metrics\]. In the proofs of Theorems \[thm:main-AdS-angles\] and \[thm:main-AdS-metrics\], many of our techniques should apply in the setting of *hyperideal* polyhedra, i.e. polyhedra whose vertices lie outside of the hyperboloid, but all of whose edges pass through the hyperboloid. We believe that similar parameterization statements may hold in this setting. The statements of Theorems \[thm:main-AdS-angles\] and \[thm:main-AdS-metrics\] bear close resemblance to a conjecture of Mess [@mes_lor] in the setting of globally hyperbolic Cauchy compact AdS space-times. Mess conjectured, by analogy to a related conjecture of Thurston in the setting of quasifuchsian groups, that such a spacetime should be determined uniquely by the bending data or by the induced metric on the boundary of the convex core inside the spacetime. There are existence results known in both cases, due to Bonsante–Schlenker [@bon_fix] and Diallo [@diallo2013] respectively, but no uniqueness or parameterization statement is known in this setting. Ultimately, Theorems \[thm:main-AdS-angles\] and \[thm:main-AdS-metrics\] on the one hand and Mess’s conjecture on the other hand boil down to understanding the connection between the geometry of a subset of $\partial_\infty \AdS^3$ and the geometry of its convex hull in $\AdS^3$. It is natural to ask whether Mess’s conjecture and our theorems on ideal polyhedra might naturally coexist as part of some larger universal theory relating the geometry of a convex spacetime in $\AdS^3$ to its asymptotic geometry at the ideal boundary. A parameterization of ideal polyhedron in $\HP^3$ {#sec:HP-param} ------------------------------------------------- Half-pipe (HP) geometry is a transitional geometry lying at the intersection of hyperbolic and anti-de Sitter geometry. Intuitively, it may be thought of as the normal bundle of a codimension one hyperbolic plane inside of either hyperbolic space or anti-de Sitter space. In [@dan_age; @dan_ide], the first named author constructs paths of three-dimensional projective structures on certain manifolds which transition from hyperbolic geometry to AdS geometry passing through an HP structure. In our setting, it is informative to imagine families of polyhedra in projective space whose vertices lie on a quadric surface evolving from the sphere to the hyperboloid passing through the cylinder. Indeed, the notion of transition is also useful for proving several key statements needed along the way to the main theorems. Half-pipe geometry is a homogeneous $(G,X)$–geometry. The projective model $X = \HP^3$ for half-pipe space is simply the solid cylinder $x_1^2 + x_2^2 < 1$ in the affine $x_1$-$x_2$-$x_3$ coordinate chart $\RR^3$. There is a natural projection $\varpi: \HP^3 \to \HH^2$, seen, in this model, as the projection of the solid cylinder to the disk. The projection is equivariant taking projective transformations which preserve the cylinder to isometries of the hyperbolic plane. The projection also extends to take the ideal boundary $\partial_\infty\HP^3 = C$ to the ideal boundary $\partial_\infty\HH^2$ of the hyperbolic plane. The structure group $G$ is the codimension one subgroup of all projective transformations preserving the cylinder which preserves a certain length function along the fibers of this projection. By pullback, the projection $\varpi$ determines a metric on $\HP^3$ which is degenerate along the fiber direction. In this metric, all non-degenerate $2$-planes are isometric to the hyperbolic plane. Let $P$ be a convex ideal polyhedron in $\HP^3$ with $N$ vertices. That $P$ is ideal means that the closure of $P$ in $\RP^3$ is a polyhedron contained in $\HP^3 \cup \partial_\infty \HP^3$ whose intersection with $\partial_\infty \HP^3$ is precisely its vertices. Since $\HP^3$ is contained in an affine chart, the notion of convexity is defined to be the same as in affine space. Then the $N$ vertices project to $N$ distinct points on the ideal boundary of the hyperbolic plane (else one of the edges of $P$ would be contained in $\partial_\infty\HP^3$, which we do not allow). Therefore $P$ determines an ideal polygon $p = \varpi(P)$ in the hyperbolic plane. Further, all facets of an ideal polyhedron in $\HP^3$ are non-degenerate; in particular the faces of $P$ are transverse to the fibers of $\varpi$. By equipping $\HP^3$ with an orientation of the fiber direction, we may sort the faces of $P$ into two types, those for which the outward pointing fiber direction is positive, and those for which it is negative. We call such faces *positive* or *negative*, respectively. The positive faces form a disk (a bent polygon) as do the negative faces. The edges of $P$ which separate a positive face from a negative face form a Hamiltonian cycle in the $1$–skeleton of $P$, again called the *equator*. As in the AdS setting, we let $\HPPoly = \HPPoly_N$ denote the space of all marked non-degenerate convex ideal polyhedra in $\HP^3$ with $N$ vertices, up to orientation preserving and fiber-orientation preserving transfomations. Again, the boundary of each ideal polyhedron $P$ is naturally identified with $\Sigma_{0,N}$ via the (isotopy class of) map taking each ideal vertex to the corresponding puncture and the equator to $\gamma$. Under this identification, the union of the positive faces (resp. the union of the negative faces) is identified with the top (resp. bottom) disk of $\Sigma_{0,N}$. Via the marking, any ideal polyhedron $P$ realizes the edges of a graph $\Gamma \in\Graph(N, \gamma)$ as a collection of geodesic lines either on the surface of or inside of $P$. The angle measure between two non-degenerate planes in $\HP^3$ can be defined in terms of the length function on the fibers. Alternatively, one should think of a non-degenerate plane in $\HP^3$ as an infinitesimal deformation of some fixed central hyperbolic plane in $\HH^3$ or $\AdS^3$. As such, the angle between two intersecting planes in $\HP^3$ should be thought of as an infinitesimal version of the standard angle measure in $\HH^3$ or $\AdS^3$. As in the AdS setting, we must distinguish between two types of dihedral angles: two non-degenerate half-planes meeting along a non-degenerate edge $e$ either lie on opposite sides of or the same side of the degenerate plane (which is the union of all degenerate lines) passing through $e$. As in the AdS setting, we take the convention that the dihedral angles along the equator of a convex ideal polyhedron $P$ are negative, while the dihedral angles along the other edges are positive. Let $\Gamma$ be the $1$–skeleton of $P$ with $\gamma$ subgraph corresponding to the equator. Let $P^*$ be the Poincaré dual of $P$. A simple argument in HP geometry (Section \[sec:proof-main-HP\]) shows that the function $\theta$ assigning to each edge $e^*$ of $P^*$ the dihedral angle at the corresponding edge $e$ of $P$ satisfies the same three conditions (i), (ii), and (iii) of the previous section; in other words $\theta \in \Angles$. Define $\HPPoly_{\Gamma}$ to be the space of ideal polyhedra in $\HP^3$ with $1$–skeleton identified with $\Gamma \in\Graph(N, \gamma)$ and let $\PsiHP_{\Gamma}: \HPPoly_{\Gamma} \to \Angles_{\Gamma}$ be the map assigning to an ideal polyhedron its dihedral angles. Then all of the maps $\PsiHP_{\Gamma}: \HPPoly_{\Gamma} \to \Angles_{\Gamma}$ may be, again, stitched together into one. We show: \[thm:main-HP\] The map $\PsiHP: \HPPoly \to \Angles$, defined by $\PsiHP(P) = \PsiHP_{\Gamma}(P)$, if $P \in \HPPoly_{\Gamma}$, is a homeomorphism. The equivalence of conditions (C) and (H) in Theorem \[thm:main\] follows from Theorem \[thm:main-HP\] and Theorem \[thm:main-AdS-angles\]. Note that there is no direct analogue of Theorem \[thm:main-AdS-metrics\] in the half-pipe setting. Indeed the induced metric on a ideal polyhedron in $\HP^3$ is exactly the double of the ideal polygon $\varpi(P)$ and the space of such doubles is a half-dimensional subset of $\Teich_{0,N}$. Intuitively, the induced metric does not determine $P$ because, as a polyhedron in $\HH^3$ (or $\AdS^3$) collapses onto a plane, the induced metric only changes to second order: the path metric on a plane bent by angle $\theta$ differs from the ambient metric only to second order in $\theta$. Strategy of the proofs and organization --------------------------------------- There is a natural relationship between bending in $\AdS^3$ and earthquakes on hyperbolic surfaces. We describe this relationship, in our context of interest, in Section \[sec:models\]. Here is a synopsis. Via the product structure on the ideal boundary $\partial_\infty \AdS^3 \cong \RP^1 \times \RP^1$, an ideal polyhedron $P \in \AdSPoly_N$ is determined by two ideal polygons $p_L$ and $p_R$ in the hyperbolic plane, each with $N$ labeled vertices (see Section \[sec:ideal\_poly\]). The two metrics $m_L, m_R \in \Teich_{0,N}$ obtained by doubling $p_L$ and $p_R$ respectively are called the *left metric* and *right metric* respectively. Given weights $\theta$ on a graph $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$, the pair $p_L, p_R$ determine an ideal polyhedron $P$ with bending data $\theta$ if and only if the left and right metrics satisfy: $$\label{eqn:earthquake-diagram-intro} m_R = E_{2\theta} m_L,$$ where $E_\theta$ is the shear map defined by shearing a surface along the edges of $\Gamma$ according to the weights given by $\theta$ (where a positive weight means shear to the left, and a negative weight means shear to the right). Directly solving for $p_L$ and $p_R$ given $\theta$ is very difficult. However, the infinitesimal version of this problem is more tractable; this is the relevant problem in the setting of half-pipe geometry. An ideal polyhedron $P \in \HPPoly_N$ is determined by an $N$-sided ideal polygon $p$ in the hyperbolic plane and an infinitesimal deformation $V$ of $p$ (see Section \[sec:models\]). Doubling yields an element $m$ of the Teichmüller space $\Teich_{0,N}$ and an infinitesimal deformation $W$ of $m$ which is tangent to the sub-space of doubled ideal polygons. The data $p, V$ determine an ideal polyhedron $P \in \HPPoly$ with bending data $\theta$ if and only if the infinitesimal deformation $W$ is obtained by infinitesimally shearing $m$ along the edges of $\Gamma$ according to the weights $\theta$. In Section \[sec:lengthfunctions\], we show how to solve for the polygon $p$ given $\theta \in \Angles_{\Gamma}$ by minimizing an associated length function. In Section \[sec:proof-main-HP\], we apply the results of Section \[sec:lengthfunctions\] to directly prove Theorem \[thm:main-HP\], that $\PsiHP$ is a homeomorphism, after first proving: \[prop:imageinA-HP\] The map $\PsiHP_{\Gamma}$ taking an ideal polyhedron $P \in \HPPoly_{\Gamma}$ to its dihedral angles $\theta$ has image in $\Angles_{\Gamma}$. In other words, $\theta$ satisfies conditions (i), (ii), and (iii) of Section \[sec:intro-AdS-param\]. The proof of this proposition is a simple computation in half-pipe geometry, which uses (among other things) an infinitesimal version of the Gauss–Bonnet theorem for polygons. In the AdS setting constructing inverses for the maps $\PsiAdS$ and $\Phi$ is too difficult, so we proceed in the usual next-best way: we prove each map is a proper, local homeomorphism, and then argue via topology. Because Teichmüller space $\Teich_{0,N}$ is a ball and because $\overline{\AdSPoly}_N$ is connected and has dimension equal to that of $\Teich_{0,N}$ (Proposition \[topol\_poly\]), Theorem \[thm:main-AdS-metrics\] is implied by the following two statements. \[lem:Phi-proper\] The map $\Phi: \overline{\AdSPoly}_N \to \Teich_{0,N}$ is proper. \[lem:Phi-rigidity\] The map $\Phi: \overline{\AdSPoly}_N \to \Teich_{0,N}$ is a local immersion. Lemma \[lem:Phi-proper\] is proved in Section \[sec:properness\] by directly studying the effect of degeneration of the left and right metrics $m_L, m_R$ of $P$ on the induced metric $\Phi(P)$ via Equation . Lemma \[lem:Phi-rigidity\] is deduced in Section \[sec:rigidity\] from a similar rigidity statement in the setting of convex Euclidean polyhedra using an *infinitesimal Pogorelov map*, which is a tool that translates infinitesimal rigidity questions form one constant curvature geometry to another. Next, to prove Theorem \[thm:main-AdS-angles\], we need the relevant local parameterization and properness statements in the setting of dihedral angles. Note that in the following lemmas, we consider each $\PsiAdS_\Gamma$ as having image in $\RR^E$, where again $E$ is the set of edges of the graph $\Gamma \in\Graph(N, \gamma)$. The first lemma is a properness statement for $\PsiAdS$. \[lem:PsiAdS-proper\] Consider a sequence $P_n \in \AdSPoly_{\Gamma}$ going to infinity in $\AdSPoly$ such that the dihedral angles $\theta_n = \PsiAdS_{\Gamma}(P_n)$ converge to $\theta_\infty \in \RR^E$. Then $\theta_\infty$ fails to satisfy condition (iii) of Section \[sec:intro-AdS-param\]. Lemma \[lem:PsiAdS-proper\] is proven in Section \[sec:properness\] together with Lemma \[lem:Phi-proper\]. In the next lemma, we assume $\Gamma$ is a triangulation (i.e. maximal) and extend the definition of $\Psi_\Gamma$ to all of $\AdSPoly$. Indeed, for $P \in \AdSPoly$, each ideal triangle of $\Gamma$ is realized as a totally geodesic ideal triangle in $P$. Therefore, the punctured sphere $\Sigma_{0,N}$ maps into $P$ as a bent (but possibly not convex) totally geodesic surface with $1$–skeleton $\Gamma$ and we may measure the dihedral angles (with sign) along the edges. \[lem:PsiAdS-rigidity\] Assume $\Gamma$ is a triangulation of $\Sigma_{0,N}$, with $E$ denoting the set of $3N - 6$ edges of $\Gamma$. If the $1$–skeleton of $P \in \AdSPoly$ is a subgraph of $\Gamma$, then $\PsiAdS_{\Gamma}: \AdSPoly \to \RR^E$ is a local immersion near $P$. Lemma \[lem:PsiAdS-rigidity\] is obtained as a corollary of Lemma \[lem:Phi-rigidity\] via a certain duality between metric data and bending data derived from the natural pseudo-complex structure on $\AdSPoly$. See Section \[sec:pseudo\] and Section \[sec:rigidity\]. The next ingredient for Theorem \[thm:main-AdS-angles\] is: \[prop:imageinA-AdS\] The map $\PsiAdS_{\Gamma}$ taking an ideal polyhedron $P \in \AdSPoly_{\Gamma}$ to its dihedral angles $\theta$ has image in $\Angles_{\Gamma}$. The content of this proposition is that $\PsiAdS_\Gamma(P)$ satisfies condition (iii) of Section \[sec:intro-AdS-param\] (conditions (i) and (ii) are automatic). This will be proven directly in Section \[sec:necess\] by a computation in $\AdS$ geometry. See Appendix \[sec:transitional-proof\] for an alternative indirect proof using transitional geometry. In Section \[sec:topology\], we explain why Lemmas \[lem:PsiAdS-proper\] and \[lem:PsiAdS-rigidity\], and Proposition \[prop:imageinA-AdS\] imply that $\PsiAdS$ is a covering onto $\Angles$. We then argue that $\Angles$ is connected and simply connected when $N \geq 6$ using Theorem \[thm:main-HP\], and we prove Theorem \[thm:main-AdS-angles\] (treating the cases $N=4,5$ separately). We also deduce Theorems \[thm:main\] from Theorem \[thm:main-AdS-angles\], \[thm:main-HP\] and Rivin’s theorem. Some of this work was completed while we were in residence together at the 2012 special program on Geometry and analysis of surface group representations at the Institut Henri Poincaré; we are grateful for the opportunity to work in such a stimulating environment. Our collaboration was greatly facilitated by support from the GEAR network (U.S. National Science Foundation grants DMS 1107452, 1107263, 1107367 “RNMS: GEometric structures And Representation varieties”). Hyperbolic, anti-de Sitter, and half-pipe geometry in dimension 3 {#sec:models} ================================================================= This section is dedicated to the description of the three-dimensional geometries of interest in this paper, and to the relationship between these geometries. We prove a number of basic but fundamental theorems, some of which have not previously appeared in the literature as stated. Of central importance is the interpretation of bending data in these geometries in terms of shearing deformations in the hyperbolic plane (Theorem \[thm:earthquakes\] and \[thm:HP-bend\]). In [@dan_age], the first named author constructs a family of model geometries in projective space that transitions from hyperbolic geometry to anti-de Sitter geometry, passing though half-pipe geometry. We review the dimension-three version of this construction here. Each model geometry $\XX = \XX(\BB)$ is associated to a real two-dimensional commutative algebra $\BB$. Let $\BB = \mathbb R + \mathbb R \kappa$ be the real two-dimensional, commutative algebra generated by a non-real element $\kappa$ with $\kappa^2 \in \RR$. As a vector space $\BB$ is spanned by $1$ and $\kappa$. There is a conjugation action: $ \overline{(a + b \kappa)} := a - b \kappa, $ which defines a square-norm $$|a + b\kappa|^2 := (a + b\kappa)\overline{(a + b\kappa)} = a^2 -b^2 \kappa^2 \ \in \ \mathbb R.$$ Note that $|\cdot|^2$ may not be positive definite. We refer to $a$ as the *real part* and $b$ as the *imaginary part* of $a + b\kappa$. If $\kappa^2 = -1$, then our algebra $\mathcal B = \CC$ is just the complex numbers, and in this case we use the letter $i$ in place of $\kappa$, as usual. If $\kappa^2 = +1$, then $\mathcal B$ is the *pseudo-complex (or Lorentz) numbers* and we use the letter $\tau$ in place of $\kappa$. In the case $\kappa^2 = 0$, we use the letter $\sigma$ in place of $\kappa$. In this case $\mathcal B = \Rsigma$ is isomorphic to the tangent bundle of the real numbers. Note that if $\kappa^2 < 0$, then $\mathcal B \cong \CC$, and if $\kappa^2 > 0$ then $\mathcal B \cong \Rtau$. Now consider the $2\times2$ matrices $M_2(\BB)$. Let $$\Herm(2, \BB) = \{ A \in M_2(\BB) : A^* = A\}$$ denote the $2\times 2$ Hermitian matrices, where $A^*$ is the conjugate transpose of $A$. As a real vector space, $\Herm(2,\BB) \cong \mathbb R^4$. We define the following (real) inner product on $\Herm(2,\BB)$: $$\left\langle \bminimatrix{a}{z}{\bar{z}}{d}, \bminimatrix{e}{w}{\bar{w}}{h} \right\rangle = -\frac{1}{2}tr\left( \bminimatrix{a}{z}{\bar{z}}{d} \bminimatrix{h}{-w}{-\bar{w}}{e}\right).$$ We will use the coordinates on $\Herm(2,\BB)$ given by $$\begin{aligned} \label{coordinates-on-Herm} X &= \bminimatrix{x_4+x_1}{x_2 - x_3 \kappa}{x_2 + x_3\kappa}{x_4 - x_1}.\end{aligned}$$ In these coordinates, we have that $$\langle X, X\rangle = -\text{det}(X) = x_1^2 + x_2^2 - \kappa^2 x_3^2 - x_4^2,$$ and we see that the signature of the inner product is $(3,1)$ if $\kappa^2 < 0$, or $(2,2)$ if $\kappa^2 > 0$. The coordinates above identify $\Herm(2,\BB)$ with $\RR^4$. Therefore we may identify the real projective space $\RP^3$ with the non-zero elements of $\Herm(2,\BB)$, considered up to multiplication by a real number. We define the region $\XX$ inside $\RP^3$ as the negative lines with respect to $\langle \cdot, \cdot\rangle$: $$\XX = \left\{ X \in \Herm(2,\BB) : \langle X, X\rangle < 0 \right\} / \RR^*.$$ Note that in the affine chart $x_4 = 1$, our space $\XX$ is the standard round ball if $\kappa = i$, the standard solid hyperboloid if $\kappa = \tau$, or the standard solid cylinder if $\kappa = \sigma$. Next, define the group $\PGL^+(2,\mathcal B)$ to be the $2\times 2$ matrices $A$, with coefficients in $\BB$, such that $|\det(A)|^2 > 0$, up to the equivalence $A \sim \lambda A$ for any $\lambda \in \mathcal B^{\times}$. The group $\PGL^+(2, \BB)$ acts on $\XX$ by orientation preserving projective linear transformations as follows. Given $A \in \PGL^+(2,\BB)$ and $X \in \XX$: $$A \cdot X := A X A^*.$$ \[rem:plane\] The matrices with real entries determine a copy of $\PSL(2,\RR)$ inside of $\PGL^+(2, \BB)$, which preserves the set $\plane$ of negative lines in the $x_1$-$x_2$-$x_4$ plane (in the coordinates above). The subspace $\plane$ of $\XX$ is naturally a copy of the projective model of the hyperbolic plane. We think of $\plane$ as a common copy of $\HH^2$ contained in every model space $\XX = \XX(\BB)$ independent of the choice of $\kappa^2$. Note that if $\BB = \CC$, then $\PGL^+(2,\BB) = \PSL(2,\CC)$ and $\XX$ identifies with the usual projective model for *hyperbolic space* $\XX = \HH^3$. In this case, the action above is the usual action by orientation preserving isometries of $\HH^3$, and gives the familiar isomorphism $\PSL(2,\CC) \cong \PSO(3,1)$, If $\BB = \Rtau$, with $\tau^2 = +1$, then $\XX$ identifies with the usual projective model for *anti-de Sitter space* $\XX = \AdS^3$. Anti-de Sitter geometry is a Lorentzian analogue of hyperbolic geometry. The inner product $\langle \cdot, \cdot \rangle$ determines a metric on $\XX$, defined up to scale. We choose the metric with constant curvature $-1$. Note that the metric on $\AdS^3$ has signature $(2,1)$, so tangent vectors are partitioned into three types: *space-like*, *time-like*, or *light-like*, according to whether the inner product is positive, negative, or null, respectively. In any given tangent space, the light-like vectors form a cone that partitions the time-like vectors into two components. Thus, locally there is a continuous map assigning the name *future pointing* or *past pointing* to time-like vectors. The space $\AdS^3$ is *time-orientable*, meaning that the labeling of time-like vectors as future or past may be done consistently over the entire manifold. The action of $\PGL^+(2,\Rtau)$ on $\AdS^3$ is by isometries, thus giving an embedding $\PGL^+(2,\Rtau) \hookrightarrow \PSO(2,2)$. In fact, $\PGL^+(2,\Rtau)$ has two components, distinguished by whether or not the action on $\AdS^3$ preserves time-orientation, and the map is an isomorphism. Lastly, we discuss the case $\BB = \Rsigma$, with $\sigma^2 = 0$. In this case, $\XX = \HP^3$ is the projective model for *half-pipe geometry* (HP), defined in [@dan_age] for the purpose of describing a geometric transition going from hyperbolic to AdS structures. The algebra $\mathbb R + \mathbb R \sigma$ should be thought of as the tangent bundle of $\mathbb R$: Letting $x$ be the standard coordinate function on $\mathbb R$, we think of $a + b \sigma$ as a path based at $a$ with tangent $b \frac{\partial}{\partial x}$. More appropriately, one should think of $\Rsigma$ as the bundle of imaginary directions in $\CC$ (resp. $\Rtau$) restricted to the subspace $\RR$. See Section \[subsec:polyhedra-HP\]. In each case, the orientation reversing isometries are also described by $\PGL^+(2, \BB)$ acting by $X \mapsto A \overline{X} A^*$. Although, we focus on dimension three, there are projective models for these geometries in all dimensions. Generally, the $n$-dimensional hyperbolic space $\HH^n$ (resp. the $n$-dimensional anti-de Sitter space $\AdS^n$) may be identified with the space of negative lines in $\RP^n$ with respect to a quadratic form of signature $(n,1)$ (resp. of signature $(n-1,2)$); the isometry group is the projective orthogonal group with respect to this quadratic form, isomorphic to $\PO(n,1)$ (resp. $\PO(n-1,1)$). The $n$-dimensional half-pipe space $\HP^n$ identifies with the space of negative lines with respect to a degenerate quadratic form with $n-1$ positive eigenvalues, one negative eigenvalue, and one zero eigenvalue. The structure group, as in the three-dimensional case, is a codimension one subgroup of all projective transformations preserving this set. See Section \[s:HP3-geometry\]. [**The ideal boundary.**]{} The ideal boundary $\partial_\infty\XX$ is the boundary of the region $\XX$ in $\RP^3$. It is given by the null lines in $\Herm(2,\BB)$ with respect to $\langle \cdot, \cdot \rangle$. Thus $$\partial_\infty\XX = \left\{ X \in \Herm(2,\BB) : \det(X) = 0, X \neq 0\right\}/\RR^*$$ can be thought of as the $2\times 2$ Hermitian matrices of rank one. We now give a useful description of $\partial_\infty \XX$ that generalizes the identification $\partial_\infty \HH^3 = \CP^1$. Any rank one Hermitian matrix $X$ can be decomposed (up to $\pm$) as $$\label{eqn:decomposition} X= \pm v v^*,$$ where $v \in \BB^2$ is a two-dimensional column vector with entries in $\BB$, unique up to multiplication by $\lambda \in \BB$ with $|\lambda|^2 = 1$ (and $v^*$ denotes the transpose conjugate). This gives the identification $$\partial_\infty \mathbb X \cong \mathbb P^1 \BB = \left\{ v \in \BB^2 : v v^* \neq 0 \right\} / \sim,$$ where $v \sim v \lambda \text{ for } \lambda \in \BB^{\times}.$ The action of $\PGL^+(2, \BB)$ on $\mathbb P^1 \BB$ by matrix multiplication extends the action of $\PGL^+(2,\BB)$ on $\XX$ described above. We note also that the metric on $\XX$ determines a compatible conformal structure on $\partial_{\infty} \XX = \mathbb P^1 \BB$. Restricted to $\BB \subset \mathbb P^1 \BB$, this conformal structure is exactly the conformal structure induced by the square-norm $| \cdot |^2$. In particular, it is Euclidean if $\kappa^2 < 0$, Lorentzian if $\kappa^2 > 0$, or degenerate if $\kappa^2 = 0$. We use the square-bracket notation $\btwovector{x}{y}$ to denote the equivalence class in $\mathbb P^1 \BB$ of $\begin{pmatrix}x \\ y \end{pmatrix} \in \BB^2$. Similarly, a $2\times 2$ square-bracket matrix $\bminimatrix{a}{b}{c}{d}$ denotes the equivalence class in $\PGL^+(2,\BB)$ of the matrix $\minimatrix{a}{b}{c}{d} \in \GL^+(2,\BB)$. Throughout, we will identify $\BB$ with its image under the injection $\BB \hookrightarrow \mathbb P^1 \BB$ given by $z \mapsto \btwovector{z}{1}$. In the case $\kappa^2 \geq 0$, the condition $v v^* \neq 0$ in the definition of $\mathbb P^1 \BB$ is *not* equivalent to the condition $v \neq 0$, because $\BB$ has zero divisors. The inclusion $\RR \hookrightarrow \BB$ induces an inclusion $\RP^1 \hookrightarrow \mathbb P^1 \BB$. This copy of $\RP^1$ is precisely the ideal boundary of the common hyperbolic plane $\plane$ contained in all model spaces $\XX$ (independent of the choice of $\kappa^2$). Recall that a subset $P$ of projective space is called *convex* if $P$ is contained in an affine chart and is convex in that affine chart. In the notation introduced here, the fundamental objects of this article are defined as follows: A *convex ideal polyhedron* in $\XX$ is a convex polyhedron $P$ in projective space such that the vertices of $P$ lie in $\partial_\infty \XX$ and the rest of $P$ lies in $\XX$. An *ideal triangle* in $\XX$ is a convex ideal polyhedron with three vertices. An *ideal simplex* or *ideal tetrahedron* is a convex ideal polyhedron with four vertices. Ideal simplices and their moduli will play an important role in this article. We review some of the basic theory, referring the reader to [@dan_ide] for a more detailed account. Let $Z_1, Z_2, Z_3, Z_4 \in \Herm(2,\BB)$ have rank one, and let $z_1, z_2, z_3, z_4$ denote the corresponding elements of $\mathbb P^1 \BB$. Assume that $Z_1, Z_2, Z_3$ determine an ideal triangle in $\XX$. There is a unique $A \in \PGL^+(2,\BB)$ such that $A z_1 = \infty := \btwovector{1}{0}, Az_2 = 0:= \btwovector{0}{1}$, and $A z_3 = 1:= \btwovector{1}{1}$. Then $$(z_1,z_2; z_3,z_4) := A z_4$$ is an invariant of the ordered ideal points $z_1,\ldots,z_4$, which will be referred to as the *cross ratio* of the four points, since it generalizes the usual cross ratio in $\CP^1$. It is straighforward to check that $z_1,z_2,z_3,z_4$ define an ideal tetrahedron in $\mathbb X$ if and only if $z = (z_1,z_2;z_3,z_4)$ (is defined and) lies in $\mathcal B \subset \mathbb P^1 \mathcal B$ and satisfies: $$|z|^2, |1-z|^2 > 0. \label{spacelike}$$ In this case $z$ is called the *shape parameter* of the ideal tetrahedron (with ordered vertices $z_1, z_2, z_3, z_4$). Using the language of Lorentzian geometry, we say that $z$ and $z-1$, as in , are *space-like*. In fact, all facets of an ideal tetrahedron are space-like and totally geodesic with respect to the metric induced by $\langle \cdot, \cdot \rangle$ on $\mathbb X$. The shape parameter $z$ is a natural geometric quantity associated to the edge $e = z_1z_2$ of the tetrahedron in the following sense, described in Thurston’s notes [@thu_the §4] in the hyperbolic case. Change coordinates (using an element of $\PGL^+(2,\BB)$) so that $z_1 = \infty$, and $z_2 = 0$. Then the subgroup $G_e$ of $\PGL^+(2,\BB)$ that preserves $e$ is given by $$G_e = \left\{ A = \bminimatrix{\lambda}{0}{0}{1}: \lambda \in \BB, |\lambda|^2 > 0 \right\}.$$ The number $\lambda = \lambda(A)$ associated to $A \in G_e$ is called the *exponential $\BB$-length* and generalizes the exponential complex translation length of a loxodromic element of $\PSL(2,\CC)$. Let $A \in G_e$ be the unique element so that $A z_3 = z_4$. Then the shape parameter is just the exponential $\BB$-length of $A$: $z = \lambda(A)$. There are shape parameters associated to the other edges as well. We may calculate them as follows. Let $\pi$ be any *even* permutation of $\{1,2,3,4\}$, which corresponds to an orientation preserving diffeomorphism of the standard simplex. Then $(z_{\pi(1)}, z_{\pi(2)}; z_{\pi(3)}, z_{\pi(4)})$ is the shape parameter associated to the edge $e' = z_{\pi(1)} z_{\pi(2)}$. This definition a priori depends on the orientation of the edge $e'$. However, one easily checks that $(z_2,z_1;z_4,z_3) = (z_1,z_2; z_3,z_4).$ Figure \[fig:shape-params\] summarizes the relationship between the shape parameters of the six edges of an ideal tetrahedron, familiar from the hyperbolic setting. Hyperbolic geometry in dimension three -------------------------------------- Let $\kappa^2 = -1$, so that $\mathcal B = \mathbb C$ is the complex numbers. In this case, the inner product $\langle \cdot, \cdot \rangle$ on $\Herm(2,\mathbb C)$ is of type $(3,1)$ and $\mathbb X$ is the unit ball in the affine chart $x_4 = 1$, known as the projective model for $\mathbb H^3$. A basic understanding of hyperbolic geometry, although not the main setting of interest, is very important for many of the arguments in this article. We will often use intuition from the hyperbolic setting as a guide, and so we assume the reader has a basic level of familiarity. Let us recall some basic facts here and present an important theorem, whose analogue in the AdS setting will be crucial. The ideal boundary $\partial_\infty \HH^3$ identifies with $\mathbb P^1 \BB = \CP^1$. Since the ball is strictly convex, any $N$ distinct points $z_1,\ldots, z_N$ determine an ideal polyhedron $P$ in $\HH^3$. In the case $N = 4$, the ideal simplex $P$ is determined by the shape parameter $z = (z_1, z_2; z_3, z_4) \in \CC$. Indeed, Condition  gives the well-known fact that the shape parameter $z$ may take any value in $\mathbb C\setminus \{0,1\}$. Consider the two faces $T = \Delta z_1 z_2 z_3$ and $T' = \Delta z_2 z_1 z_4$ of $P$, each oriented compatibly with the outward pointing normal, meeting along the edge $e = z_1 z_2$. Then, writing $z = e^{s + i \theta}$, the quantity $s$ is precisely the amount of *shear* along $e$ between $T$ and $T'$, while $\theta$ is precisely the dihedral angle at $e$. An infinitesimal deformation of an ideal polyhedron $P$ is given by a choice $V = (V_1, \ldots, V_N)$ of tangent vectors to $\CP^1$ at each of the vertices $z_1, \ldots, z_N$ of $P$. Such a deformation is considered trivial if $V_1, \ldots, V_N$ are the restriction of a global Killing field on $\HH^3 \cup \CP^1$ to the vertices $z_1, \ldots, z_n$. If necessary, augment the $1$–skeleton of $P$ so that it is an ideal triangulation $\Gamma$ of the surface of $P$. Then the map $z_\Gamma$, taking an ideal polyhedron $P$ to the collection of $3N-6$ cross ratios associated to the edges of $\Gamma$, is holomorphic and the following holds: \[thm:H3-duality\] An ideal polyhedron $P$ is infinitesimally rigid with respect to the induced metric if and only if $P$ is infinitesimally rigid with respect to the dihedral angles. Since the induced metric is determined entirely by the shear coordinates with respect to $\Gamma$, we have that the infinitesimal deformation $V$ does not change the induced metric to first order if and only if $d\log z_\Gamma (V)$ is pure imaginary. On the other hand, $V$ does not change the dihedral angles to first order if and only if $d\log z_\Gamma (V)$ is real. Therefore $V$ does not change the induced metric if and only if $i V$ does not change the dihedral angles. Theorem \[thm:H3-duality\] is a simpler version of Bonahon’s argument [@bon_she] that a hyperbolic three-manifold is rigid with respect to the metric data on the boundary of the convex core if and only if it is rigid with respect to bending data on the boundary of the convex core. In this setting of polyhedra, Bonahon’s shear-bend cocycle is replaced by a finite graph $\Gamma$ with edges labeled by the relevant shape parameters $z$ (or $\log z$). Anti-de Sitter geometry in dimension three {#ads_background} ------------------------------------------ Let $\mathcal{B}$ be the real algebra generated by an element $\tau$, with $\tau^2 = +1$, which defines $\mathbb X = \AdS^3$, the anti-de Sitter space. Let us discuss some important properties of the algebra $\mathcal B = \mathbb R + \mathbb R \tau$, known as the pseudo-complex numbers. [**The algebra $\BB = \mathbb R + \mathbb R \tau$ of pseudo-complex numbers.**]{} First, note that $\mathcal{B}$ is not a field as, for example, $(1+\tau)\cdot(1-\tau) = 0.$ The square-norm defined by the conjugation operation $|a+b\tau|^2 = (a + b \tau) \overline{(a + b\tau)} = a^2 - b^2$ comes from the $(1,1)$ Minkowski inner product on $\mathbb R^2$ (with basis $\{1, \tau\}$). The space-like elements of $\mathcal{B}$ (i.e. square-norm $ > 0$), acting by multiplication on $\mathcal{B}$, form a group and can be thought of as the similarities of the Minkowski plane that fix the origin. Note that if $|a + b\tau|^2 = 0$, then $b=\pm a$, and multiplication by $a + b\tau$ collapses all of $\mathcal{B}$ onto the light-like line spanned by $a + b\tau$. The elements $\frac{1+\tau}{2}$ and $\frac{1-\tau}{2}$ are two spanning idempotents which annihilate one another: $$\left(\frac{1\pm\tau}{2}\right)^2 = \frac{1\pm\tau}{2}, \ \text{ and } \ \left(\frac{1+\tau}{2}\right) \cdot \left(\frac{1-\tau}{2}\right) = 0.$$ Thus $\mathcal{B} \cong \mathbb R \oplus \mathbb R$, as $\mathbb R$–algebras, via the isomorphism $$(\varpi_L, \varpi_R): a\left(\frac{1-\tau}{2}\right) + b \left(\frac{1+\tau}{2}\right) \longmapsto (a,b). \label{isomorphism}$$ Here $\varpi_L$ and $\varpi_R$ are called the left and right projections $\BB \to \RR$. These projections extend to left and right projections $\mathbb P^1 \BB \to \RP^1$ which give the isomorphism $\mathbb P^1 \BB \cong \RP^1 \times \RP^1$. Indeed, $\mathbb P ^1 \mathcal B$ is the *Lorentz compactification* of $\mathcal B = \left\{ \btwovector{x}{1} : \ x \in \mathcal B \right\}$. The added points make up a wedge of circles, so that $\mathbb P^1 \mathcal B$ is topologically a torus. The square-norm $|\cdot|^2$ on $\mathcal B$ induces a flat conformal Lorentzian structure on $\mathbb P^1 \mathcal B$ that is preserved by $\PGL^+(2, \mathcal B)$. We refer to $\PGL^+(2, \mathcal B)$ as the *Lorentz Möbius transformations*. With its conformal structure $\mathbb P^1 \mathcal B$ is the $(1+1)$-dimensional Einstein universe $\text{Ein}^{1,1}$ (see e.g. [@barbot-1; @primer-einstein] for more about Einstein space). The splitting $\BB \cong \RR \oplus \RR$ determines a similar splitting $M_2 (\mathcal B) \cong M_2 \RR \oplus M_2 \RR$ of the algebra of $2 \times 2$ matrices which respects the determinant in the following sense: $\forall A \in M_2 (\mathcal B)$ $$(\varpi_L \det A, \varpi_R \det A) = (\det \varpi_L(A), \det \varpi_R(A)),$$ where, by abuse of notation, $\varpi_L$ and $\varpi_R$ also denote the extended maps $M_2 (\mathcal B) \to M_2 (\R)$. The orientation preserving isometries $\operatorname{Isom}^+ \AdS^3 = \PGL^+(2,\mathcal B)$ correspond to the subgroup of $\PGL(2,\mathbb R) \times \PGL(2,\mathbb R)$ such that the determinant has the same sign in both factors. The identity component of the isometry group (which also preserves time orientation) is given by $\PSL(2,\RR) \times \PSL(2,\RR)$. Note also that the left and right projections $\varpi_L, \varpi_R: \mathbb P^1 \BB \to \RP^1$ respect the cross ratio: $$(z_1, z_2; z_3, z_4) = \dfrac{1-\tau}{2} (\varpi_L z_1 , \varpi_L z_2; \varpi_L z_3, \varpi_L z_4) + \dfrac{1+\tau}{2} (\varpi_R z_1, \varpi_R z_2; \varpi_R z_3, \varpi_R z_4),$$ where on the right-hand side $(\cdot,\cdot;\cdot,\cdot)$ denotes the usual cross ratio in $\RP^1$. Ideal Polyhedra in $\AdS^3$ {#sec:ideal_poly} --------------------------- Consider an ideal polyhedron $P$ in $\AdS^3$ with $N$ vertices $z_1, \ldots, z_N \in \mathbb P^1 \BB$. For each $i = 1, \ldots, N$, let $x_i = \varpi_L(z_i)$ and $y_i = \varpi_R(z_i)$ be the left and right projections of $z_i$. Then, all of the $x_i$ (resp. all of the $y_i$) are distinct. Otherwise, the convex hull of the $z_i$ (in any affine chart) will contain a full segment in the ideal boundary. \[prop:cyclic-order\] The vertices $z_1, \ldots, z_N \in \mathbb P^1 \BB$ determine an ideal polyhedron $P$ in $\AdS^3$ if and only the left projections $x_1, \ldots, x_N$ and right projections $y_1, \ldots, y_N$ are arranged in the same cyclic order on the circle $\RP^1$. In general, a closed set $\Omega$ in $\RP^M$ is convex if and only any $M+1$ points of $\Omega$ span a (possibly degenerate) simplex contained in $\Omega$. Therefore the $z_1, \ldots, z_N$ define an ideal polyhedron if and only if any four vertices $z_{i_1}, z_{i_2}, z_{i_3}, z_{i_4}$ span an ideal simplex. This is true if and only if the cross ratio $z = (z_{i_1}, z_{i_2}; z_{i_3}, z_{i_4})$ is defined and satisfies that $|z|^2, |1-z|^2 > 0$. Since $z = \frac{1-\tau}{2}x + \frac{1+\tau}{2}y$, where $x = (x_{i_1}, x_{i_2}; x_{i_3}, x_{i_4})$ and $y = (y_{i_1}, y_{i_2}; y_{i_3}, y_{i_4})$, we have that $|z|^2 = xy$ and $|1-z|^2 = (1-x)(1-y)$. So $|z|^2, |1-z|^2 > 0$ if and only if $x$ and $y$ have the same sign and $(1-x)$ and $(1-y)$ have the same sign. Hence, $z_{i_1}, z_{i_2}, z_{i_3}, z_{i_4}$ span an ideal simplex if and only if the two four-tuples of vertices $(x_{i_1}, x_{i_2}, x_{i_3}, x_{i_4})$ and $(y_{i_1}, y_{i_2}, y_{i_3}, y_{i_4})$ are arranged in the same cyclic order on $\RP^1$. The proposition follows by considering all subsets of four vertices. We denote by $p_L = \varpi_L(P)$ (resp. $p_R = \varpi_R(P)$) the ideal polygon in the hyperbolic plane with vertices $x_1, \ldots, x_N$ (resp. $y_1,\ldots, y_N$). Let us quickly recall the definitions and terminology from Section \[sec:intro-AdS-param\]. We fix, once and for all, a time orientation on $\AdS^3$. Since all faces of an ideal polyhedron $P$ are space-like, the outward normal to each face is time-like and points either to the future or to the past. This divides the faces into two groups, the future (or top) faces, and the past (or bottom) faces. The union of the future faces is a bent polygon, as is the union of the past faces. The edges dividing the future faces from the past faces form a Hamiltonian cycle, called the *equator*, in the $1$–skeleton of $P$. We may project $P$ combinatorially to the left and right ideal polygons $p_L$ and $p_R$ respectively. Each face of $P$ is isometric to an ideal polygon in the hyperbolic plane. Therefore the *induced metric* on the boundary of $P$ is naturally a hyperbolic metric $m$ on the $N$-punctured sphere; it is a complete metric. Further, the labeling of the vertices, the equator, and the top and bottom of $P$ determine an identification (up to isotopy) of the surface of $P$ with the $N$-puncture sphere $\Sigma_{0,N}$, making $m$ into a point of the Teichmüller space $\Teich_{0,N}$. The marking also identifies the $1$–skeleton of $P$ with a graph $\Gamma$ on $\Sigma_{0,N}$ with vertices at the punctures. The edges of the equator project to exterior edges of $p_L$ (resp. $p_R$) and top/bottom edges project to interior edges of $p_L$ (resp. $p_R$). We may assume the $1$–skeleton is a triangulation by adding additional top/bottom edges as needed. Consider an edge $e = z_1z_2$ adjacent to two faces $T = \Delta z_1 z_2 z_3$ and $T' = \Delta z_4 z_1 z_2$, each oriented so that the normal points out of $P$. Then the cross ratio $z = (z_1, z_2; z_3, z_4)$ contains the following information: The edge $e$ is an equatorial edge if and only if $z = a + b \tau$ has real part $a > 0$. Since $z$ is space-like, we may express it as $$z = \pm e^{s + \tau \theta} := \pm e^s(\cosh \theta + \tau \sinh \theta).$$ By convexity of $P$, the imaginary part of $z$ is always positive. Hence, either $z = +e^{s + \tau \theta}$ with $\theta > 0$, or $z = -e^{s+\tau \theta}$ with $\theta < 0$. In the former case, the edge $e$ is a top/bottom edge and in the latter case, $e$ is an equatorial edge. In either case, $s = s(e)$ is precisely the *shear coordinate* of the induced metric $m$ along the edge $e$, and $\theta$ is the exterior *dihedral angle* at the edge $e$. We now give the fundamentally important relationship between shearing and bending in the setting of ideal polyhedra. Let $m_L$ (resp. $m_R$) denote the double of $p_L$ (resp. $p_R$). Since the vertices of $P$, and its projections $p_L$ and $p_R$, are labeled, we may regard $m_L$ and $m_R$ as points of the Teichmüller space $\Teich_{0,N}$; we call $m_L$ the *left metric* and $m_R$ the *right metric*. Recall the definition of $\AdSPoly_N$ given in Section \[sec:intro-AdS-param\]. \[thm:earthquakes\] Let $m_L, m_R, m \in \Teich_{0,N}$ be the left metric, the right metric, and the induced metric defined by $P \in \AdSPoly_N$, and let $\theta$ denote the dihedral angles. Then the following diagram holds: $$\label{eqn:earthquake-diagram} m_L \xrightarrow[]{E_\theta} m \xmapsto[]{E_\theta} m_R,$$ where $E_\theta$ denotes shearing along $\Gamma$ according to the weights $\theta$ (a positive weight means shear to the left). Further, given the left and right metrics $m_L$ and $m_R$ (any two metrics obtained by doubling two ideal polygons $p_L$ and $p_R$), the induced metric $m$ and the dihedral angles $\theta$ are the unique metric and weighted graph on $\Sigma_{0,N}$ (with positive weights on the top/bottom edges) such that  holds. Let $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$ represent the $1$–skeleton of $P$. By adding extra edges if necessary, we may assume $\Gamma$ is a triangulation. As above we associate the shape parameter $z = \varepsilon e^{s(\alpha)+ \tau \theta(\alpha)}$ to a given edge $\alpha$ of $\Gamma$, where $\varepsilon = \pm 1$. Then, $$\begin{aligned} z &= \varepsilon e^{s(\alpha)} (\cosh \theta(\alpha) + \tau \sinh \theta(\alpha))\\ &= \varepsilon e^{s(\alpha)} \left( \frac{1-\tau}{2} e^{-\theta(\alpha)} + \frac{1+\tau}{2} e^{\theta(\alpha)}\right)\\ &= \frac{1-\tau}{2}\varepsilon e^{s(\alpha) - \theta(\alpha)} + \frac{1+\tau}{2}\varepsilon e^{s(\alpha)+\theta(\alpha)} \end{aligned}$$ Therefore the shear coordinates in the left metric $m_L$ are given by $s_L = s- \theta$ and the shear coordinate in the right metric $m_R$ are $s_R = s + \theta$. Equation  follows. The uniqueness statement also follows from this calculation. Indeed, given $m_L$, $m_R$ and any graph $\Gamma \in \Graph(\Sigma_{0,N})$ we may solve for the shear coordinates $s$, determining a metric $m$, and the weights $\theta$ needed to satisfy . Specifically, $s = (s_R + s_L)/2$ and $\theta = (s_R - s_L)/2$, where now $s_L$ and $s_R$ denote the shear coordinates with respect to $\Gamma$. We may construct a polyhedral embedding of $\Sigma_{0,N}$ whose induced metric is $m$ and whose (exterior) bending angles are $\theta$ as follows. Begin with the polyhedral embedding of $\Sigma_{0,N}$ into a space-like plane given by doubling $p_L$. Then bend this embedding along the edges of $\Gamma$ according to the weights $\theta$; note that this can be done consistently because $\theta$ satisfies condition (ii) in the definition of $\Angles$ (Section \[sec:intro-AdS-param\]) (because the shear coordinates for $m_L$ and $m$ satisfy that condition). If $\theta$ is positive on the top/bottom edges and negative on the equatorial edges, this polyhedral embedding is convex; it is the boundary of a convex ideal polyhedron $P$. By the definition of $s$ and $\theta$, we have that the left and right metrics of $P$ are precisely $m_L$ and $m_R$. The uniqueness statement follows because $P$ is uniquely determined by $m_L$ and $m_R$. As a corollary we obtain a version of Thurston’s earthquake theorem for ideal polygons in the hyperbolic plane. A *measured lamination* on the standard ideal $N$-gon is simply a pairwise disjoint collection of diagonals with *positive* weights. We denote by $\ML_N$ the complex of these measured laminations. A function $\theta \in \Angles_{\Gamma}$ determines two measured laminations $\theta_+$ and $\theta_-$ by restriction to the top edges of $\Gamma$ and to the bottom edges. \[cor:earthquake\] Let $p_L, p_R \in \poly_N$ be two ideal polygons. Then there exists unique $\theta_+, \theta_- \in \ML_N$ such that $p_R = E_{\theta_+} p_L$ and $p_L = E_{\theta_-} p_R$, where again $E_\lambda$ denotes shearing according to the edges of $\lambda \in \ML_N$ according to the weights of $\lambda$. Let $x_1, \ldots, x_N$ be the ideal vertices of $p_L$ and let $y_1, \ldots, y_N$ be the ideal vertices of $p_R$. Then, the vertices $z_i = \frac{1-\tau}{2}x_i + \frac{1+\tau}{2} y_i$ define an ideal polyhedron $P \in \AdSPoly_N$ such that $\varpi_L(P) = p_L$ and $\varpi_R(P) = p_R$. We think of $\Sigma_{0,N}$ as the double of the standard ideal $N$-gon, meaning that the top hemisphere is identified with the standard ideal $N$-gon and the bottom hemisphere is identified with the standard ideal $N$-gon but with orientation reversed. The left metric $m_L$ (resp. $m_R)$ is obtained from $p_L$ (resp. $p_R$) by doubling. This means that the restriction of $m_L$ to the top hemisphere of $\Sigma_{0,N}$ is $p_L$ and the restriction of $m_L$ to the bottom hemisphere is $\overline{p_L}$, the same ideal polygon but with opposite orientation. Similarly, the restriction of $m_R$ to the top and bottom hemispheres of $\Sigma_{0,N}$ is $p_R$ and $\overline{p_R}$. Let $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$ denote the $1$–skeleton of $P$ and let $\theta \in \RR^{E(\Gamma)}$ denote the dihedral angles. Theorem \[thm:earthquakes\] implies that $m_R = E_{2 \theta} m_L$. Restricting to the top hemisphere, we have that $p_R = E_{\theta_+} p_L$ where $\theta_+ \in \ML_N$ is twice the restriction of $\theta$ to the top hemisphere. Restricting to the bottom hemisphere, we have that $\overline{p_R} = E_{\theta_-} \overline{p_L}$, where $\theta_-$ is the restriction of $\theta$ to the bottom hemisphere. This implies that $p_R = E_{-\theta_-} p_L$, or equivalently $p_L = E_{\theta_-} p_R$. Uniqueness of $\theta_+, \theta_-$ follows from uniqueness of $\theta$ in Theorem \[thm:earthquakes\]. In the setting of closed surfaces, it is known [@bon_fix] that given two filling measured laminations $\theta_+$ and $\theta_-$, there exists two hyperbolic surfaces $\rho_L$ and $\rho_R$ such that $\rho_R$ is obtained from $\rho_L$ by left earthquake along $\theta_+$ and also by right earthquake along $\theta_-$, and it is conjectured [@mes_lor] that $\rho_L$ and $\rho_R$ are unique. The analogous question, in the context of Corollary \[cor:earthquake\], of whether a given $\theta_+$ and $\theta_-$ are realized by some $p_L$ and $p_R$, and whether they are realized uniquely, is an interesting one. A necessary condition is that $\theta_+$ and $\theta_-$ be filling, which means that any lamination intersects $\theta_+$ or $\theta_-$ transversely; this is equivalent to the statement that the graph $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$, obtained by placing the support of $\theta_+$ on the top hemisphere and the support of $\theta_-$ on the bottom hemisphere, is three-connected. It will follow from Theorem \[thm:main-AdS-angles\] that in the case $N$ is odd, the polygons $p_L, p_R$ are unique, given the measured laminations $\theta_+, \theta_-$. This is because $\theta \in \Angles$ is determined entirely by its restrictions $\theta_+$ and $\theta_-$ to the top and bottom edges. However, there are examples of filling measured laminations $\theta_+, \theta_-$ such that there is no element $\theta \in \Angles$ whose restriction to the top edges is $\theta_+$ and whose restriction to the bottom edges is $\theta_-$ (see Appendix \[sec:transitional-proof\]). The situation is even worse in the case $N$ is even. There is a one dimensional family of pairs $p_L, p_R$ for which the laminations $\theta_+, \theta_-$ turn out to be the same. This is because for any $\theta \in \Angles$, there is a one parameter family of deformations of $\theta$ which leave $\theta_+, \theta_-$ unchanged: simply add and subtract the same quantity from the weights of alternating edges on the equator. Further, in the case $N$ even, only a codimension one subspace of filling laminations $\theta_+, \theta_-$ are realized in Corollary \[cor:earthquake\]. It is an interesting problem to determine this codimension one subspace. The pseudo-complex structure on $\AdSPoly_N$ {#sec:pseudo} -------------------------------------------- The space of marked ideal polyhedra $\AdSPoly_N$ naturally identifies with a subset of $(\RR + \RR \tau)^{N-3}$, by transforming each ideal polyhedron so that its first three vertices are respectively $0, 1, \infty \in \mathbb P^1 \BB$. The marking on each polyhedron $P \in \AdSPoly_N$ identifies $P$ with the standard $N$-punctured sphere $\Sigma_{0,N}$. So, given a triangulation $\Gamma$ on $\Sigma_{0,N}$ with vertices at the punctures and edge set denoted $E$, we may define the map $z_\Gamma: \AdSPoly_N \to (\RR + \RR \tau)^E$ which associates to each edge $e$ of a polyhedron $P$ the cross ratio of the four points defining the two triangles adjacent at $e$. This map is pseudo-complex holomorphic, meaning that the differential is $(\RR+ \RR\tau)$–linear. This observation allows us to prove the following analogue of Theorem \[thm:H3-duality\]. \[thm:duality\] A polyhedron $P \in \AdSPoly_N$ is infinitesimally rigid with respect to the induced metric if and only if $P$ is infinitesimally rigid with respect to the dihedral angles. Let $V \in T_P \AdSPoly_N \cong (\Rtau)^{N-3}$. Let $\Gamma$ be a triangulation obtained from the $1$–skeleton of $P$ by adding edges in the non-triangular faces if necessary. Since the induced metric is determined entirely by the shear coordinates with respect to $\Gamma$, we have that $V$ does not change the induced metric to first order if and only if $d\log z_\Gamma (V)$ is pure imaginary. On the other hand, $V$ does not change the dihedral angles to first order if and only if $d\log z_\Gamma (V)$ is real. Therefore $V$ does not change the induced metric if and only if $\tau V$ does not change the dihedral angles. Half-pipe geometry in dimension three {#s:HP3-geometry} ------------------------------------- We give some lemmas useful for working with $\HP^3$. Recall the algebra $\RR + \RR\sigma$, with $\sigma^2 = 0$. The half-pipe space is given by $$\HP^3:= \mathbb X = \left\{ X + Y \sigma : X,Y \in M_2(\RR), X^T = X, \det(X) > 0, Y^T = -Y\right\}/\sim,$$ where $(X + Y \sigma) \sim \lambda (X + Y \sigma)$ for $\lambda \in \RR^\times$. There is a projection $\varpi : \HP^3 \to \HH^2$, defined by $\varpi(X + Y \sigma) = X$, where we interpret the symmetric matrices $X$ of positive determinant, considered up to scale, as a copy of $\HH^2$. The fibers of this projection will be referred to simply as *fibers*. The projection can be made into a diffeomorphism $\mathbb X \to \HH^2 \times \RR$ (not an isometry) given in coordinates by $$\begin{aligned} \label{product-coords-dim3} X + Y \sigma \mapsto (X, L),\end{aligned}$$ where the length $L$ along the fiber is defined by the equation $$\begin{aligned} \label{fiber-length-dim3} Y &= L \sqrt{\det X}\begin{pmatrix} 0 & -1\\ 1 & 0 \end{pmatrix}. \end{aligned}$$ The ideal boundary $\partial_\infty \XX_0$ identifies with $\mathbb P^1 (\Rsigma)$, which identifies with the tangent bundle $T \RP^1$ via the natural map $T \RR^2 \to (\RR + \RR\sigma)^2$ sending a vector $v \in \RR^2$ and a tangent vector $w \in T_v \RR^2 = \RR^2$ to $v + \sigma w$. It will be convenient to think of an ideal vertex as an infinitesimal variation of a point on $\RP^1 \cong \partial_\infty \HH^2$. In this way, a convex ideal polyhedron $P$ in $\HP^3$ defines an infinitesimal deformation $V = V(P)$ of the ideal polygon $p = \varpi(P)$ in $\HH^2$. We restrict to the identity component of the structure group, which is given by $$\begin{aligned} \mathbb G_0 &= \PSL(2, \RR + \RR\sigma)\\ &= \{ A + B \sigma : A \in \SL(2,\RR), \text{ and } B \in T_A \SL(2,\RR) \} / \pm.\end{aligned}$$ The structure group identifies with the tangent bundle $T \PSL(2,\RR)$, and it will be convenient to think of its elements as having a finite component $A\in \PSL(2,\RR)$ and an infinitesimal component $a \in \mathfrak{sl}(2,\RR)$, via the isomorphism $$\begin{aligned} \PSL(2,\RR) \ltimes \mathfrak{sl}(2,\RR) &\to \mathbb G_0\\ (A,a) &\mapsto A + Aa \sigma,\end{aligned}$$ where $Aa \in T_A \PSL(2,\RR)$. (This is the usual isomorphism $G \ltimes \mathfrak g \to T G$ for a Lie group $G$ with Lie algebra $\mathfrak g = T_1 G$, gotten by left translating vectors from the identity.) The identification $\mathbb G_0 \cong T \PSL(2,\RR)$ is compatible with the identification $\partial_\infty \HP^3 \cong T \RP^1$. Thinking of $a \in \mathfrak{sl}(2,\RR)$ as an infinitesimal isometry of $\HH^2$, recall that at each point $X \in \HH^2$ we may decompose $a$ into its translational ($X$-symmetric) and rotational ($X$-skew) parts: $$\begin{aligned} a &= a_{X\text{-sym}} + a_{X\text{-skew}} \\ &:= \frac{1}{2}\left( a + Xa^T X^{-1} \right) + \frac{1}{2} \left( a - Xa^T X^{-1} \right),\end{aligned}$$ where the rotational part $a_{X\text{-skew}}$ is a rotation centered at $X$ of infinitesimal angle $\rot(a,X)$ defined by $$\sqrt{X}^{-1} a_{X\text{-skew}} \sqrt{X} = \rot(a,X) \begin{pmatrix} 0 & -1/2 \\ 1/2 & 0\end{pmatrix}.$$ The action of an element of $\mathbb G_0$ in the fiber direction depends on the rotational part of the infinitesimal part of that element. \[lem:HP3-action\] The action of a pure infinitesimal $1 + a\sigma$ on the point $X + Y\sigma \in \mathbb X$ is by translation in the fiber direction by amount equal to the rotational part $\rot(a, X)$ of the infinitesimal isometry $a$ at the point $X \in \HH^2$. In the product coordinates (\[product-coords-dim3\]): $$1+ a \sigma : (X,L) \mapsto (X, L + \rot(a,X)).$$ More generally, the action of $A + Aa\sigma$ is given by $$A + Aa\sigma : (X,L)\mapsto (A\cdot X, L + \rot(a,X)).$$ $$\begin{aligned} (1+a\sigma)\cdot (X + Y\sigma) &= (1+a\sigma) (X+ \sigma Y) (1 - a^T\sigma)\\ &= X + \sigma Y + \sigma(a X - X a^T)\\ &= X + \sigma Y + \sigma \ 2 a_{X\text{-skew}} X\\ &= X + \sigma Y + \sigma \ 2 \rot(a,X) \sqrt{X} \begin{pmatrix} 0 & -1/2\\ 1/2 & 0\end{pmatrix} \sqrt{X}\\ &= X + \sigma Y + \sigma \ \rot(a,X) \det(\sqrt{X}) \begin{pmatrix} 0 & -1\\ 1 & 0\end{pmatrix},\end{aligned}$$ and the first statement now follows from Equation (\[fiber-length-dim3\]). The second more general formula follows easily after left multiplication by $A$. Let $a \in \mathfrak{sl}_2 \RR$ be an infinitesimal translation of length $t$ along an oriented geodesic $\ell$ in $\HH^2$. Then, for any oriented geodesic $\wt{\ell}$ in $\HP^3$ that projects to $\ell$, the element $1+ a \sigma$ is called an *infinitesimal rotation* about the axis $\wt{\ell}$ of infinitesimal angle $t$. Thinking of the fiber direction in $\HP^3$ as the direction of infinitesimal unit length normal to $\HH^2$ into either $\HH^3$ or $\AdS^3$, the definition is justified by the previous lemma. In fact, the amount of translation in the fiber direction is $t$ times the signed distance to $\widetilde \ell$. Ideal polyhedra in $\HP^3$ {#subsec:polyhedra-HP} -------------------------- There are several important interpretations of a convex ideal polyhedron $P$ in $\HP^3$. As described in the previous section, $P$ defines an infinitesimal deformation $V = V(P)$ of the ideal polygon $p = \varpi(P)$ in $\HH^2$. Alternatively, $P$ may be interpreted as an infinitesimally thick polyhedron in $\HH^3$ or $\AdS^3$. Multiplying the tangent vector $V$ by $i$ (resp. $\tau$) describes an infinitesimal deformation $iV$ (resp. $\tau V$) of the polygon $p$ into $\HH^3$ (resp. $\AdS^3$). The polyhedron $P$ in $\HP^3$ is a rescaled limit of a path of hyperbolic (resp. anti-de Sitter) polyhedra collapsing to $p$ and tangent to $iV$ (resp. $\tau V$) in the following sense. Consider the path of algebras $\BB_t$ generated by $\kappa_t$ such that $\kappa_t^2 = -t|t|$. Then the geometries $\XX(\BB_t)$ associated to these algebras are conjugate to $\XX(\BB_1) = \XX(\CC) = \HH^3$ for all $t > 0$, or to $\XX(\BB_{-1}) = \XX(\Rtau) = \AdS^3$ for $t < 0$. For $t > 0$, the map $\mathfrak a_t : \CC \to \BB_t$ defined by $i \mapsto \kappa_t/|t|$ is an isomorphism of algebras. For $t < 0$, the map $\mathfrak a_t : \Rtau \to \BB_t$ defined by $\tau \mapsto \kappa_t/|t|$ is an isomorphism. Each of these maps defines a projective transformation, again denoted $\mathfrak a_t$, taking the standard model of hyperbolic space $\HH^3 = \XX(\BB_1)$ (resp. the standard model of anti-de Sitter space $\AdS^3 = \XX(\BB_{-1})$) to the conjugate model $\XX(\BB_t)$. \[prop:limit\] Consider a smooth family $Q_t$ of ideal polyhedra in $\HH^3$ (resp. $\AdS^3$), defined for $t > 0$ (resp. for $t < 0$). Assume that $Q_0 = p$ is an ideal polygon contained in the central hyperbolic plane $\plane$ bounded by $\RP^1$ and $Q'_0 = U + iW$ (resp. $Q_0' = U + \tau W$), where $U, W$ are infinitesimal deformations of $p$ as an ideal polygon in $\HH^2$. Then the limit of $\mathfrak a_t (Q_t)$ as $t \to 0$ is an ideal polyhedron $P$ in $\XX(\BB_0) = \HP^3$ which satisfies $\varpi(P) = Q_0$ and $V(P) = W$. The interplay between these two interpretations leads to Theorem \[thm:HP-bend\] below, which is a fundamental tool for studying half-pipe geometry. Before stating the theorem, let us recall the terminology introduced in Section \[sec:HP-param\] and state a proposition. We fix an orientation of the fiber direction once and for all. Every convex ideal polyhedron in $\HP^3$ has a top, for which the outward pointing fiber direction is positive, and a bottom, for which the outward pointing fiber direction is negative. The edges naturally sort into three types: an edge is called a *top edge* if it is adjacent to two top faces or a *bottom edge* if it is adjacent to two bottom faces, or an *equatorial edge* if it is adjacent to both a top and bottom face. The union of the top faces is a bent polygon which projects down to the ideal polygon $p = \varpi(P)$ in $\HH^2$. The union of the bottom faces also projects to $p$. The *infinitesimal dihedral angle* at an edge is measured in terms of the infinitesimal rotation angle needed to rotate one face adjacent to the edge into the same plane as the other. The dihedral angle at a top/bottom edge will be given a positive sign, while the dihedral angles at an equatorial edge will be given a negative sign. This sign convention is justified by the following (see [@dan_age §4.2]): \[prop:angles-limit\] The infinitesimal dihedral angle along an edge of $P$ is simply the derivative of the dihedral angle of the corresponding edge of $Q_t$, where $Q_t$ is as in Proposition \[prop:limit\]. Alternatively, dihedral angles may also be measured using the cross ratio. Indeed, if two (consistently oriented) ideal triangles $T = \Delta z_1 z_2 z_3$ and $T' = \Delta z_4 z_1 z_2$ meet at a common edge $\alpha = z_1 z_2$, then the cross ratio $z = (z_1, z_2; z_3, z_4)$ satisfies that $z = \varepsilon e^{s+ \sigma \theta} = \varepsilon e^s(1+ \sigma \theta)$, where $s$ is the shear between $T$ and $T'$, where $\theta$ is the dihedral angle, and where $\varepsilon$ is $+1$ if $\alpha$ is an edge of the equator and $-1$ if $\alpha$ is a top/bottom edge. We consider the bending angles on the top (resp. bottom) edges of an ideal polyhedron $P$ as a (positive) measured lamination on the ideal polygon $p = \varpi(P)$. The following theorem is the infinitesimal version of Theorem \[thm:earthquakes\] about the interplay between earthquakes and AdS geometry. \[thm:HP-bend\] Let $P$ be an ideal polyhedron in $\HP^3$ and let $\theta_{+}$ (resp. $\theta_{-}$) be the measured lamination on $p = \varpi(P)$ describing the bending angles on top (resp. on bottom). Then the infinitesimal deformation $V = V(P)$ of $p$ defined by $P$ is equal to $e_{\theta_{+}}(p)$, where $e_{\theta_{+}}$ is the infinitesimal left earthquake along $\theta_{+}$. Similarly, $V = -e_{\theta_{-}}(p)$ is obtained by right earthquake along $\theta_{-}$. Let $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$ represent the $1$–skeleton of $\partial P$. By adding extra edges if necessary, we may assume $\Gamma$ is a triangulation. As above we associate the shape parameter $z(\alpha) = \varepsilon e^{s(\alpha)+ \sigma \theta(\alpha)}$ to any given edge $\alpha$ of $\Gamma$. Note that the map taking four points on $\RP^1$ to their cross ratio is smooth and that the isomorphism $T \RP^1 \cong \mathbb P^1(\RR + \RR \sigma)$ commutes with the cross ratio operation. Therefore the shear coordinate of $p = \varpi(P)$ at $\alpha$ is $s(\alpha)$ and the infinitesimal variation of the shear coordinate at $\alpha$ under the deformation $V(P)$ is $\theta(\alpha)$. The result follows. Half-pipe geometry in dimension two ----------------------------------- The structure group $\mathbb G$ for $\HP^3$ acts transitively on degenerate planes, i.e. the planes for which the restriction of the metric on $\HP^3$ is degenerate. These are exactly the planes that appear vertical in the standard picture of $\HP^3$ (as in Figure \[fig:HP-action\]); they are the inverse image of lines (copies of $\HH^1$) in $\HH^2$ under the projection $\varpi$. Each degenerate plane is a copy of two-dimensional half-pipe geometry $\HP^2$. For the purposes of the following discussion, we will fix one degenerate plane in $\HP^3$ as our model: $$\HP^2 := \left\{ \begin{pmatrix} x & 0 \\ 0 & x^{-1} \end{pmatrix} + \sigma\begin{pmatrix} 0 & y \\ -y & 0 \end{pmatrix} \right\}.$$ Here we describe two important facts about $\HP^2$. The first is (reasonably) named the *infinitesimal Gauss-Bonnet formula*. See [@dan_age §3] for details about half-pipe geometry in arbitrary dimensions. There is an invariant notion of area in $\HP^2$. As above, let $L$ denote the length function along the fiber direction. Then the area of a polygon $p$ (or a more complicated body) is the integral of the length $L(\varpi^{-1}(x) \cap p)$ of the segment of $p$ above $x$, over all $x \in \varpi(p) \subset \HH^1$. Alternatively, if $p$ is the limit as $t \to 0$ of $\mathfrak a_t p_t$, where $p_t$ is a smooth family of collapsing polygons in $\HH^2$, then the area of $p$ is simply derivative at $t=0$ of the area of $p_t$. \[prop:inf-GB\] Let $p$ be a polygon in $\HP^2$ whose edges are each non-degenerate. Then the area of $p$ is equal to the sum of the exterior angles of $p$. In particular, the sum of the exterior angles of any polygon is positive. Let $p_t$ be a smooth family of collapsing polygons in $\HH^2$ so that $p$ is the limit as $t \to 0$ of $\mathfrak a_t p_t$. Then the area of $p$ is the derivative of the area of $p_t$ at $t=0$. Each exterior angle of $p$ is the derivative of the corresponding angle of $p_t$ at $t= 0$. The proposition follows from the usual Gauss-Bonnet formula for polygons in $\HH^2$. Secondly, we give a bound on the dihedral angle between two non-degenerate planes in terms of the angle seen in the intersection with a degenerate plane $H \cong \HP^2$. This will be used in the proof of Proposition \[prop:imageinA-HP\]. \[prop:HP-angle-bound\] Let $P, Q$ be two non-degenerate planes in $\HP^3$ which intersect at dihedral angle $\theta$. Let $H$ be a degenerate plane so that the lines $H \cap P$ and $H \cap Q$ intersect at angle $\vartheta$ in $H \cong \HP^2$. Then $\operatorname{sign}(\vartheta) = \operatorname{sign}(\theta)$ and $|\vartheta| \leq |\theta|$ with equality if and only if $H$ is orthogonal to the line $P \cap Q$. We may change coordinates so that $P = \plane$ (recall that $\plane$ is a copy of $\HH^2$ common to all of the models $\XX(\BB)$ in projective space, see Remark \[rem:plane\]). The second plane $Q$ is the limit as $t \to 0$ of $\mathfrak a_t Q_t$, where $Q_t$ is a smoothly varying family of planes in $\HH^3$ with limit $Q_0 = \plane$. We may choose the path $Q_t$ so that the line $L = Q_t \cap \plane$ is constant for all $t > 0$. The dihedral angle between $Q$ and $\plane$ is the derivative at $t= 0$ of the dihedral angle $\theta_t$ between $Q_t$ and $\plane$, now thought of as a plane in $\HH^3$. The degenerate plane $H$ defines a plane $H'$ in (the projective model of) $\HH^3$ which is orthogonal to $\plane$. Let $\vartheta_t$ be the angle formed by $Q_t \cap H'$ and $\plane \cap H'$ in $H' \cong \HH^2$. Then, because $H'$ and $\plane$ are orthogonal, we have that $$\tan \vartheta_t = \tan \theta_t \sin \varphi$$ where $0 < \varphi \leq \frac{\pi}{2}$ is the angle between the line $L = Q_t \cap \plane$ and $H'$. The proposition now follows since $\vartheta = \frac{d}{dt} \big|_{t=0} \vartheta_t$, $\theta = \frac{d}{dt} \big|_{t=0} \theta_t$ and $\theta_0 = \vartheta_0$ (both are either zero or $\pi$). Length functions and earthquakes {#sec:lengthfunctions} ================================ We prove Theorem \[thm:main-HP\] by showing that each ideal polyhedron in $\HP^3$ is realized as the unique minimum of a certain length function defined in terms of its dihedral angles. Our strategy is inspired by a similar one used by Series [@ser_ker], and later Bonahon [@bon_kle], in the setting of quasifuchsian hyperbolic three-manifolds with small bending. Shear and length coordinates on the Teichmüller space of a punctured sphere --------------------------------------------------------------------------- Consider an ideal triangulation $\Gamma$ of the $N$-times punctured sphere $\Sigma_{0,N}$. Let $\alpha_1, \ldots, \alpha_n$ denote the $n=3N-6$ edges of $\Gamma$. There are two natural coordinate systems on the Teichmüller space $\Teich_{0,N}$ of complete hyperbolic metrics on $\Sigma_{0,N}$ (see [@penner:decorated; @thurston:minimal]): - Let $s_1,\cdots, s_n$ denote the shear coordinates along the edges of $\Gamma$. The sum of the shear coordinates over edges adjacent to a particular vertex is always zero. Under this condition, the shears along the edges provide global coordinates on $\Teich_{0,N}$. - We may define length coordinates $\ell_1, \ldots, \ell_n$ on $\Teich_{0,N}$ as follows. In any hyperbolic structure, choose a horocycle around each cusp, and let $\ell_i$ denote the (signed) length of the segment of $\alpha_i$ connecting the two relevant horocycles. By abuse, we call $\ell_i$ the length of $\alpha_i$. Changing a horocycle at a particular cusp corresponds to adding a constant to the lengths of all edges going into that cusp. The lengths $\ell_1, \ldots, \ell_n$ are only well-defined up to this addition of constants, making these coordinates elements of $\mathbb R^n/\mathbb R^N$. It is well-known [@penner:decorated; @thurston:minimal] that both the shears and the lengths give global coordinate systems for Teichmüller space. It is quite simple to go from length coordinates to shear coordinates, in fact the map sending lengths to shears is linear. To describe this coordinate transformation more precisely, let us establish some notation. The orientation of the surface determines a cyclic order on the edges of any triangle. Given any two edges $\alpha_i, \alpha_j$, let $\epsilon_{ij} = -\epsilon_{ji}$ be the number of positively oriented triangles $T$ of $\Gamma$ such that $\alpha_i, \alpha_j$ are distinct edges of $T$ counted with a positive sign if $\alpha_j$ follows $\alpha_i$ in the cyclic order on the edges of $T$, and with negative sign if $\alpha_i$ follows $\alpha_j$. By definition, $(\epsilon_{ij})_{1\leq i,j\leq n}$ is an anti-self adjoint matrix with entries in $\{-1,0,1\}$. It is straightforward to check the following: \[lm:penner\] Given a hyperbolic metric $h\in \Teich_{0,N}$ with length coordinates $(\ell_i)$, the corresponding shear coordinates are defined by $$s_i =\frac{1}{2} \sum_j \epsilon_{ij}\ell_j~.$$ Note that the right-hand side is independent of the horocycles chosen to define the $\ell_i$. Let $\omega$ denote the anti-symmetric bilinear form on $\Teich_{0,N}$, defined by $$\label{eqn:omega1} \omega = \frac{1}{2}\sum_{i,j} \epsilon_{ij} d \ell_i \otimes d \ell_j.$$ Note that, by Lemma \[lm:penner\], we may also express $\omega$ as $$\label{eqn:omega} \omega = \sum_{i} d \ell_i \otimes d s_i.$$ It follows that $\omega$ is well-defined (independent of the ambiguity in the definition of $d\ell_i$) because for any tangent vector $Y$, $d s_i(Y)$ is a *balanced function* on the set $E = E(\Gamma)$ of edges, meaning it is a function whose values sum to zero on those edges incident to any vertex. From the second expression for $\omega$, we can see that it is a symplectic form, i.e. it is non-degenerate. In fact, we mention that $\omega$ is nothing other than (a multiple of) the Weil-Petersson symplectic form (see Wolpert [@wolpert:products] and Fock-Goncharov [@fock-goncharov-1]), though we will not need this fact. It is straight-forward to check directly that $\omega$ does not depend on the particular triangulation used in its definition. The gradient of the length function ----------------------------------- Given a function $f:\Teich_{0,N}\to \R$, we denote by $D^\omega f$ its symplectic gradient with respect to $\omega$, defined by the following relation: for any vector field $X$ on $\Teich_{0,N}$, $$\omega(D^\omega f,X) = df(X).$$ Let $\theta = (\theta_1, \ldots, \theta_n)$ be any balanced assignments of weights to the edges of $\Gamma$. Then one may define the corresponding *length function* $\ell_\theta$ as a function on $\Teich_{0,N}$: for any hyperbolic metric $h\in \Teich_{0,N}$, with length coordinates $(\ell_i)_{1\leq i\leq n}$, set $$\ell_\theta(h) = \sum_i \theta_i \ell_i~.$$ The function $\ell_\theta$ does not depend on the choice of horocycles at the cusps precisely because $\theta$ is balanced. We let $e_\theta$ denote the vector field on $\Teich_{0,N}$ defined by $ds_i(e_\theta) = \theta_i$, in other words $e_\theta$ shears along each edge according to the weights $\theta$. It follows immediately from that: \[lm:gradient\] Let $\theta = (\theta_1, \ldots, \theta_n)$ be balanced weights on the edges of $\Gamma$. Then $$D^\omega \ell_\theta = -e_\theta.$$ The space of doubles is Lagrangian ---------------------------------- We assume, from here on, that our graph $\Gamma$ admits a Hamiltonian cycle $\gamma$. Then cutting $\Sigma_{0,N}$ along $\gamma$ yields two topological ideal polygons, one of which we label *top* and the other *bottom*. There is an orientation reversing involution $\iota$ on $\Sigma_{0,N}$ which exchanges top with bottom and point-wise fixes $\gamma$. We let $\doubles$ denote the half-dimensional subspace of $\Teich_{0,N}$ which is fixed by the action of $\iota$, i.e. those hyperbolic metrics which are obtained by doubling a hyperbolic ideal polygon and marking the surface in such a way that the boundary of the polygon identifies with $\gamma$. \[pr:doubles\] The space of doubles $\doubles$ is a Lagrangian subspace of $\Teich_{0,N}$ with respect to $\omega$. We may compute $\omega$ with respect to a symmetric triangulation $\Gamma$ (one which is fixed under the involution $\iota$). For $h \in \Teich_{0,N}$, the shear coordinates $(s_i(h))$ are anti-symmetric, in the sense that, if $\iota(\alpha_i) = \alpha_j$, then $s_i(h) = -s_j(h)$. (So, in particular, $s_i(h) = 0$, if $\alpha_i$ is an edge of $\gamma$.) On the other hand, the lengths $(\ell_i(h))$ are symmetric, in the sense that, if $\iota(\alpha_i) = \alpha_j$, then $\ell_i(h) = \ell_j(h)$. The proposition follows immediately from the second expression for $\omega$ above. Convexity of the length function -------------------------------- We now show a form of convexity for the restriction of the length function $\ell_\theta$ to the space of doubles $\doubles$ in $\Teich_{0,N}$. It will sometimes be convenient to identify the space of doubles $\doubles$ with the space $\poly = \poly_N$ of marked ideal polygons in the hyperbolic plane, and to think of (the restriction of) $\ell_\theta$ as a function on $\poly$. The graph $\Gamma$ on $\Sigma_{0,N}$, then, projects to each polygon $p$ in $\poly$, with $\gamma$ identified to the perimeter edges of $p$ and all other edges of $\Gamma$ identified with diagonals of $p$. The following proposition is the analog, in the (simpler) setting of ideal polygons, of a theorem of Kerckhoff [@ker_lin] which played a key role in Series’s analysis of quasi-Fuchsian manifolds with small bending [@ser_ker]. In a similar way, the proposition is crucial for Theorem \[thm:main-HP\]. \[pr:convex\] For all $\theta\in \Angles_\Gamma$, the length function $\ell_\theta:\poly_N\to \R$ is proper and admits a unique critical point which is a non-degenerate minimum. The proof is based on two lemmas. \[lm:proper\] If $\theta\in \Angles_\Gamma$, then $\ell_\theta:\poly\to \R$ is proper. \[lm:path-convex\] The function $\ell_\theta$ is convex and non-degenerate on earthquake paths in $\poly$. Let $\theta\in \Angles_\Gamma$. Since $\ell_\theta$ is proper by Lemma \[lm:proper\], it has at least one minimum in $\poly$. Moreover Lemma \[lm:path-convex\] shows that any critical point is a non-degenerate minimum. Let $p,p'\in \poly_N$ be two minima of $\ell_\theta$. There is, by Corollary \[cor:earthquake\], a unique measured lamination $\lambda$ on $p$ such that $E_\lambda(p)=p'$. Then Lemma \[lm:path-convex\] shows that the function $t\mapsto \ell_\theta(E_{t\lambda}(p))$ is convex and non-degenerate, so it cannot have critical points both at $t=0$ and at $t=1$, a contradiction. So $\ell_\theta$ has a unique critical point on $\poly_N$. We now turn to the proofs of the two lemmas. Let $(p_n)_{n\in \N}$ be a sequence of ideal polygons with $N$ vertices, which degenerates in $\poly_N$. Then, after taking a subsequence, if necessary, there is a finite collection of segments $a_1, \ldots, a_p$ on the polygon such that: - $a_i$ and $a_j$ are disjoint, if $i\neq j$, - for all $n$, each $a_i$ is realized as a minimizing geodesic segment connecting two non-adjacent edges of $p_n$, - for all $i\in \{ 1,\cdots, p\}$, the length of $a_i$ in $p_n$ goes to zero, as $n \to\infty$, - any two edges of $p_n$ that can be connected by a segment disjoint from the $a_i$ remain at distance at least $\epsilon$, for some $\epsilon>0$ independent of $n$. After taking a further subsequence, the $p_n$ converge to the union of $p+1$ ideal polygons $p_\infty^{(1)}, \ldots, p_\infty^{(p+1)}$, which, topologically, is obtained by cutting the original polygon along each $a_i$ and then collapsing each (copy of each) segment $a_i$ to a new ideal vertex. Recall that given $r > 0$ and a geodesic line $\alpha$ in $\HH^2$, the $r$-neighborhood of $\alpha$ is called a *hypercycle* neighborhood of $\alpha$. We may choose horoballs at each ideal vertex and disjoint hypercycle neighborhoods $N_{i,n}$ of the (geodesic realization in $p_n$ of) $a_i$, with radii $r_{i,n} \to \infty$, which converge to a system of horoballs for the limiting ideal polygons $p_\infty^{(1)}, \ldots, p_\infty^{(p+1)}$. Our function $\theta$ is naturally defined on the limiting polygons, since all edges of the limit correspond to edges of the original polygon. However, $\theta$ is no longer balanced at the new ideal vertices of $p_\infty^{(1)}, \ldots, p_\infty^{(p+1)}$; instead the sum of the $\theta$ values along the edges going into one of the new vertices is strictly positive, since $\theta$ satisfies assumption $(iii)$ of Section \[sec:intro-AdS-param\]. Now, we may split $\ell_\theta$ into two pieces $$\ell_\theta= \ell_\theta \big |_{\cup N_{i,n}} + \ell_\theta \big |_{(\cup N_{i,n})^c},$$ corresponding to the weighted length contained in the union of the neighborhoods $N_{i,n}$ and the weighted length outside of those neighborhoods. The former is always positive, since $\theta$ satisfies condition $(iii)$ of Section \[sec:intro-AdS-param\], and since the arcs with positive weight crossing $a_i$ have length at least $2 r_{i,n}$ in $N_{i,n}$, while the two arcs with negative weight crossing $a_i$ have length exactly $2 r_{i,n}$ in $N_{i,n}$. The later converges to the $\theta$-length function $\ell_\theta(p_\infty^{(1)})+ \cdots + \ell_\theta(p_\infty^{(p+1)})$ of the limiting polygons with respect to the limiting horoballs. However, by altering the radii of the neighborhoods $N_{i,n}$, we may arrange for the limiting horoball around each of the new vertices to be arbitrarily small (i.e. far out toward infinity), making $\ell_\theta(p_\infty^{(1)})+ \cdots + \ell_\theta(p_\infty^{(p+1)})$ arbitrarily large. It follows that $\ell_\theta(p_n) \to +\infty$. Let $p\in \poly_N$, and let $\lambda$ be a measured lamination on $p$, that is, a set of disjoint diagonals $\beta_1, \cdots, \beta_q$ each with a weight $\lambda_i>0$. We need to prove that the function $t\to \ell_\theta(E_{t\lambda}p)$ is convex with strictly positive second derivative. To prove this, we prove an analogue of the Kerckhoff–Wolpert formula in this setting, specifically: $$\label{eqn:KW} \frac d{dt} \ell_\theta(E_{t \lambda}p) = \sum \theta_i \lambda_j \cos(\varphi_{ij}) + K,$$ where $\varphi_{ij} \in (0,\pi)$ is the angle at which the edge $\alpha_i$ of $\Gamma$ crosses the edge $\beta_j$ of the support of $\lambda$, the sum is taken over all $i,j$ so that $\alpha_i$ intersects $\beta_j$ non-trivially, and $K := K(\theta, \lambda)$ is independent of $p$ and $t$. The lemma follows from this formula by a standard argument about earthquakes (see [@ker_nie Lemma 3.6]): each angle $\varphi_{ij}$ of intersection strictly decreases with $t$ because, from the point of view of the edge $\beta_j$, the endpoints at infinity of $\alpha_i$ are moving to the left. It suffices to prove the formula  in the case that the lamination $\lambda$ is a single diagonal $\beta$ with weight equal to one. We choose horocycles $h_{v,t}$ at each vertex $v$ and at each time $t$ along the earthquake path as follows. Begin at time $t = 0$ with any collection of horocycles $\{h_{v,0}\}$. For a vertex $v$ that is not an endpoint of $\beta$, we simply apply the earthquake $E_{t \lambda}$ to $h_{v,0}$: Define $h_{v,t} = E_{t\lambda} h_{v,0}$. If $w$ is an endpoint of $\beta$, then the earthquake breaks the horocycle $h_{w,0}$ into two pieces. We define $h_{w,t}$ to be the horocycle equidistant from these two pieces. An easy calculation in hyperbolic plane geometry shows that, for $\alpha_i$ an edge of $\Gamma$ crossing $\beta$, we have $$\frac d{dt} \ell(\alpha_i) = \frac{d}{dt} \operatorname{dist}(h_{v,t}, h_{v',t}) = \cos(\varphi_{i}),$$ where $\varphi_{i}$ is the angle at which $\alpha_i$ crosses $\beta$, where $v$ and $v'$ are the endpoints of $\alpha_i$, and where $\operatorname{dist}(\cdot, \cdot)$ denotes the signed distance between horocycles. Further $\ell(\beta)$ remains constant along the earthquake path. Finally, for any edge $\alpha_k$ which shares one endpoint $v$ with $\beta$, we have that $\frac{d}{dt} \ell(\alpha_k) = \pm 1/2$ is independent of $p$ and $t$; the sign depends on whether $\alpha_k$ lies on one side of $\beta$, or the other. Proof of Theorem \[thm:main-HP\] {#sec:proof-main-HP} -------------------------------- We now have tools to prove Theorem \[thm:main-HP\]. First, however, we must prove Proposition \[prop:imageinA-HP\]. We must prove that the dihedral angles $\theta = \Psi_{\Gamma}(P)$ of any ideal polyhedron $P \in \HPPoly_{\Gamma}$ satisfies the three conditions (i), (ii), (iii) required for $\theta \in \Angles_{\Gamma}$, as described in Section \[sec:intro-AdS-param\]. Condition (i) is simply our convention of labeling the dihedral angles of equatorial edges with negative signs. So, we must prove that $\theta$ satisfies (ii) and (iii). That $\theta$ satisfies condition (ii) follows from the fact that the sum of the dihedral angles at a vertex of an ideal polyhedron in $\HH^3$ is constant (equal to $2\pi$). By Proposition \[prop:angles-limit\], the dihedral angles of $P$ are simply the derivatives of the (exterior) dihedral angles of $Q_t$, where $Q_t$ is a path of ideal polyhedra in hyperbolic space (or anti-de Sitter space), as in Proposition \[prop:limit\]. Now, let us prove that $\theta$ satisfies (iii). Consider a path $c$ on $P$ normal to the $1$–skeleton $\Gamma$ and crossing exactly two edges of the equator. Then, without affecting the combinatorics of the path, we deform so that $c$ is precisely $P \cap H$ for some vertical (degenerate) plane $H$ that is orthogonal to *both* edges of the equator crossed by $c$. Note that the angle between a non-degenerate line $\alpha$ and a degenerate plane $H$ is precisely the angle formed between the lines $\varpi(\alpha)$ and $\varpi(H)$ in $\HH^2$ and therefore we can indeed achieve that $H$ is orthogonal to both edges of the equator simultaneously (by contrast to the analogous situation in $\HH^3$ or $\AdS^3$). The plane $H$ is isomorphic to a copy of two-dimensional half-pipe geometry $\HP^2$. Inside $H$, the edges of $c$ are non-degenerate, forming a polygon with exterior angles *bounded above* by the corresponding dihedral angles of $P$; indeed if $\theta_i$ is the dihedral angle between two faces in $\HP^3$ and $\vartheta_i$ is the angle formed by those faces when intersected with $H$, then by Proposition \[prop:HP-angle-bound\], $\operatorname{sign}(\vartheta_i) = \operatorname{sign}(\theta_i)$ and $|\vartheta_i| \leq |\theta_i|$ with equality if and only if $H$ is orthogonal to the line of intersection between the faces; therefore the exterior angle in $H$ at each of the two points where $c$ intersects the equator is equal to the exterior dihedral angle along that equatorial edge (and both are negative) while the exterior angle at any other vertex of $c$ is strictly less that the exterior dihedral angle of $P$ at the corresponding edge (and both are positive). By the infinitesimal Gauss-Bonnet formula in $\HP^2$ (Proposition \[prop:inf-GB\]), the sum of the exterior angles of $c$ is positive and so it follows that the sum of the exterior dihedral angles over the edges of $P$ crossed by $c$ is also positive. The map $F: \HPPoly \to \poly \times \Angles$, taking an HP ideal polyhedron to its projection to $\HH^2$, an ideal polygon, and to its dihedral angles, has a continuous left inverse. Let $G: \poly \times \Angles \to \HPPoly$ be the map that takes $p \in \poly$ and bends according to the *top angles* $\theta_+$ of $\theta \in \Angles$, ignoring the rest of the information in $\theta$ (the bottom and equatorial dihedral angles). Then $G \circ F$ is the identity. Hence, to show that $\Psi = \PsiHP$ is a homeomorphism, we need only show that there is a continuous map $H: \Angles \to \poly$ such that $G(H(\Psi(P)), \Psi(P)) = P$. The existence of such a continuous map $H$ is guaranteed by Proposition \[pr:convex\] and a simple application of the Implicit Function Theorem as follows. For $\theta \in \Angles$, define $H(\theta)$ to be the unique minimum in $\poly$ of $\ell_\theta$ given by Proposition \[pr:convex\]. That $H$ is continuous (in fact differentiable on all strata of $\Angles$) follows from the convexity of $\ell_\theta$, thought of as a function on $\poly$. Now, recall that the space of ideal polygons $\poly$ identifies with the space of doubles $\doubles$ in $\Teich_{0,N}$. Hence, because $H(\theta)$ minimizes $\ell_\theta$ over $\poly$, the restriction of $d \ell_\theta$ (now thought of as a one-form on all of $\Teich_{0,N}$) to $\doubles$ is zero at (the double of) $H(\theta)$. It then follows that the infinitesimal shear $e_\theta$ on $\Teich_{0,N}$ is tangent to the subspace of doubles $\doubles$ at (the double of) $H(\theta)$ because $e_\theta$ is dual to $\ell_\theta$ (Lemma \[lm:gradient\]) and the space of doubles $\doubles$ is Lagrangian (Proposition \[pr:doubles\]). Therefore $e_\theta$ determines a well-defined infinitesimal deformation of the polygon $H(\theta)$ and the pair $p = H(\theta), V = e_\theta(H(\theta))$ determines an HP polyhedron $P$ such that $F(P) = (H(\theta), \theta)$ as in the discussion in Section \[subsec:polyhedra-HP\]. The formula $G(H(\Psi(P)),\Psi(P)) = P$ follows, and this completes the proof of Theorem \[thm:main-HP\]. Properness {#sec:properness} ========== In this section we will prove the two properness lemmas needed for the proofs of the main results. Lemma \[lem:Phi-proper\] states that the map $\Phi$, sending an ideal polyhedron in $\AdS^3$ to its induced metric, is proper. Lemma \[lem:PsiAdS-proper\], when combined with Proposition \[prop:imageinA-AdS\], will imply properness of the map sending a polyhedron of fixed combinatorics to its dihedral angles. Properness for the induced metric (Lemma \[lem:Phi-proper\]) ------------------------------------------------------------ To prove Lemma \[lem:Phi-proper\], we consider a compact subset $\mathcal K \subset \Teich_{0,N}$. We must show that the set $\Phi^{-1}(\mathcal K)$ is a compact subset of $\overline \AdSPoly$. In other words, if $P$ is a polyhedron with $m = \Phi(P) \in \mathcal K$, we must show that $P$ lies in a compact subset of $\overline \AdSPoly$. Since there are finitely many triangulations of the disk with $N$ vertices, we may consider polyhedra $P$ with fixed combinatorics, that is the graph $\Gamma$ is fixed. We may assume $\Gamma$ is a traingulation by adding edges if necessary. ![The polyhedron $P$ with combinatorics given by $\Gamma$. The red edge is $e$, and $q^+ = \pi_R(v^+; e)$, $q^- = \pi_R(v^-; e)$.[]{data-label="Fig:hep"}](Fig_hep.pdf){height="5cm"} Recall that the induced metric $m$ on $P$ is related to the left and right metrics $m_L$ and $m_R$ by the diagram in Theorem \[thm:earthquakes\]: $m_R = E_\theta(m)$ and $m = E_\theta(m_L)$, where $\theta: \Gamma \to \RR$ is the assignment of exterior dihedral angles to the edges of $P$ and $E_\theta$ is the shear amp associated to $\theta$. Also, recall that $m_L$ and $m_R$ are cusped metrics on the sphere that come from doubling the metric structures on the ideal polygon obtained by projecting the vertices of $P$ to the left and right foliations of $\partial_\infty \AdS^3$. To show that $P$ lies in a compact set, we must show that $m_L$ and $m_R$ lie in compact sets. It is enough to show that $\theta$ remains bounded over $\Phi_\gamma^{-1}(\mathcal K)$. Although we have not yet proved Proposition \[prop:imageinA-AdS\], we will use here that $\theta$ satisfies conditions (i) and (ii) in the definition of $\Angles_\Gamma$ (Section \[sec:intro-AdS-param\]). That these conditions are satisfied is essentially trivial, see Section \[sec:necess\]. Consider an edge $e$ of the equator $\gamma$ of $\Gamma$, and recall that $\theta(e) < 0$ (condition (i) of the definition of $\Angles_\Gamma$). Let $s_L(e), s_R(e), s(e)$ denote the shear coordinate along $e$, with respect to $\Gamma$, of the left metric $m_L$, the right metric $m_R$, and the induced metric $m$. Then, by Theorem \[thm:earthquakes\], we have: $$s_R(e) - s(e) = \theta(e) = s(e) - s_L(e)~.$$ Now the edge $e$ belongs to a unique triangle of $\Gamma$ in the top hemisphere of $\Sigma_{0,N}$, the third vertex of which we denote by $v^+$. On the bottom hemisphere, the edge $e$, again, belongs to a unique triangle, whose third vertex we denote by $v^-$. There are two cases to consider. Recall that we fixed an orientation of the equator $\gamma$. Imagining that we view $\Sigma_{0,N}$ from above, it is intuitive to call the positive direction *left* and the negative direction *right*. First suppose $v^+$ lies to the left of $v^-$ when viewed from $e$. The restriction of the right metric $m_R$ to the top hemisphere of $\Sigma_{0,N}$ is a marked hyperbolic ideal polygon $p_R$, in which the vertex $v^+$ again lies to the left of $v^-$. Since $m_R$ is the double of $p_R$, we may calculate the shear coordinate $s_R(e)$ of $s_R$ by the simple formula: $$s_R(e) = \pi_R(v^+; e) - \pi_R(v^-; e),$$ where $\pi_R(v; e)$ denotes the projection of $v$ onto the edge $e$ in $p_R$, see Figure \[Fig:hep\]. Then we have $s_R(e) > 0$ and so $s_L(e) > s(e) > s_R(e) > 0$. In particular, $$\theta(e) = s_R(e) - s(e) > -s(e).$$ In the case that $v^+$ lies to the right of $v^-$, we examine the left metric $m_L$. In the restriction $p_L$ of $m_L$ to the top hemisphere, the vertex $v^+$ again lies to the right of $v^-$ and so, by a similar calculation as above, $s_L(e) < 0$ and so $s_R(e) < s(e) < s_L(e) < 0$. Therefore $$\theta(e) = s(e) - s_L(e) > s(e)~.$$ In either case, $\theta(e)$ is bounded, because the shears $s(e)$ are bounded, as $m$ varies over the compact set $\mathcal{K}$. We have shown that all of the edges $e$ for which $\theta(e) < 0$ have $\theta(e)$ bounded. It then follows that the other edges $e'$, for which $\theta(e') \geq 0$, also have $\theta(e')$ bounded, since the sum of all positive and negative angles along edges coming into any vertex of $P$ must be zero (condition (ii) of the definition of $\Angles_\Gamma$). Therefore $\Phi^{-1}(\mathcal K)$ is compact. Proof of Lemma \[lem:PsiAdS-proper\] ------------------------------------ Let $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$. We consider a sequence $(P_n)_{n\in \N}$ going to infinity in $\AdSPoly_N$ such that the dihedral angles $\theta_n = \PsiAdS_\Gamma(P_n)$ converge to $\theta_\infty \in \RR^E$, where $E = E(\Gamma)$ denotes the edges of $\Gamma$ as usual. We msut show that $\theta_\infty$ fails to satisfy condition (iii) of Section \[sec:intro-AdS-param\]. For each $n$, let $p^L_n = \varpi_L(P_n)$ and $p^R_n = \varpi_R(P_n)$ be the ideal polygons whose ideal vertices are the left and right projections of the ideal vertices of $P_n$ (as in Section \[sec:ideal\_poly\]). Let $v^L_{1,n},\cdots, v^L_{N,n}$ denote the vertices in $\RP^1$ of $p_n^L$, and similarly let $v^R_{1,n},\cdots, v^R_{N,n}$ denote the vertices of $p_n^R$. By applying an isometry of $\AdS^3$, we may assume that the first three vertices of $P_n$ are $(0,0)$, $(1,1)$ and $(\infty, \infty)$ independent of $n$, so that $v^L_{1,n} = v^R_{1,n} = 0$, $v^L_{2,n} = v^R_{2,n} = 1$ and $v^L_{3,n} = v^R_{3,n} = \infty$ for all $n$. Since $\theta_n$ converges to the limit $\theta_\infty$ and the polyhedra $P_n$ diverge, the sequence of ideal polygons $(p_n^L)_{n\in \N}$ diverges (in the space of ideal $N$-gons up to equivalence). Reducing to a subsequence, we may assume all of the vertices converge to well-defined limits $v^L_{i,n} \to v^L_{i, \infty} \in \mathbb{RP}^1$. However, for some indices $i$, we will have $v^L_{i, \infty} = v^L_{i+1, \infty}$. Now, since the right polygon $p^R_n$ is obtained from $p^L_n$ by an earthquake of bounded magnitude, it follows that each vertex $v^R_{i,n}$ also converges to a well-defined limit $v^R_{i, \infty}$ and that $v^L_{i, \infty} = v^L_{i+1, \infty}$ if and only if $v^R_{i, \infty} = v^R_{i+1, \infty}$. In other words the polyhedra $P_n$ converge to a convex ideal polyhedron $P_\infty$ of strictly fewer vertices. The combinatorial structure of $P_\infty$ is obtained from $\Gamma$ by collapsing vertices and the corresponding edges and faces in the obvious way: if two vertices that span an edge collapse together, then that edge disappears. If that edge bounded a triangle, then that triangle collapses to an edge, and so on. Let $\Gamma_\infty$ denote the $1$-skeleton of $P_\infty$, and let $\Gamma_\infty^*$ denote the dual graph. Consider a simple path $c_\infty$ in $\Gamma_\infty^*$. We may lift $c_\infty$ to a path $c$ in the dual graph $\Gamma^*$ fof $\Gamma$ in the obvious way: an edge of $c_\infty$ is dual to an edge $e$ of $P_\infty$. Under the collapse $\Gamma \to \Gamma_\infty$, $e$ lifts to a collection of consecutive edges in $\Gamma$ which determines a path of adjacent edges in $\Gamma^*$. The sum of the dihedral angles assigned by $\theta_n$ to the path $c$ converges to the sum of the dihedral angles of $P_\infty$ over the edges of $c_\infty$. Now consider an ideal vertex of $P_\infty$ which is the limit of two or more vertices of the $P_n$, and let $c_\infty$ denote the path of edges bounding the face of $\Gamma_\infty^*$ dual to this vertex. Of course, the sum of the angles over the edges of $c_\infty$ is zero (by condition (ii) in the definition of $\Angles_{\Gamma_\infty}$). It therefore follows that $\theta_\infty$ assigns angles that sum to zero around the edges of the path $c$. Therefore $\theta_\infty$ violates condition (iii) in the definition of $\Angles_\Gamma$, since $c$ does not bound a face in $\Gamma^*$, and the proof is complete. Rigidity {#sec:rigidity} ======== This section is dedicated to the local versions of Theorems \[thm:main-AdS-angles\] and  \[thm:main-AdS-metrics\], which are Lemma \[lem:Phi-rigidity\] and Lemma \[lem:PsiAdS-rigidity\]. The Pogorelov map for $\AdS^n$ {#sec:Pogorelov} ------------------------------ We recall here the definition and main properties of the infinitesimal Pogorelov map, which turns infinitesimal rigidity problems for polyhedra (or submanifolds) in constant curvature pseudo-Riemannian space-forms into similar infinitesimal rigidity problems in flat spaces, where they are easier to deal with. These maps, as well as their non-infinitesimal counterparts, were discovered by Pogorelov [@Pogorelov Chapter 4] (in the Riemannian case). Another account and some geometric explanations of the existence of these maps can be found in Schlenker [@shu Prop. 5.7] or in Fillastre [@fillastre3 Section 3.3]. See also Labourie–Schlenker [@iie Cor. 3.3] or Izmestiev [@izmestiev:projective]. We follow here mostly the presentation given in [@fillastre3 Section 3.3], and refer to this paper for the proofs. Although we will return to dimension three shortly, we describe the Pogorelov map in any dimension $n$. Consider the complement $U$ in $\AdS^n$ of a spacelike totally geodesic hyperplane $H_0$, dual to a point $x_0 \in \AdS^n$. Here duality means that $H_0$ is defined by the equation $\langle x_0, x\rangle = 0$, where $\langle \cdot, \cdot \rangle$ is the inner product of signature $(n-1,1)$ defining $\AdS^n$. Then $U$ is naturally the intersection of $\AdS^n$ with an affine chart $\RR^n$ of projective space, and we may take $\iota(x_0) = 0$ to be the origin of this affine chart, where $\iota : U \hookrightarrow \RR^n$ denotes the inclusion. The union of all light-like geodesics passing through $x_0$ is called the light cone $C(x_0)$. We equip $\RR^n$ with a flat Lorentzian metric, making it into a copy of Minkowski space $\RR^{n-1,1}$. We may choose this metric so that the inclusion $\iota$ is an isometry at the tangent space to $x_0$. This has the pleasant consequence that $\iota(C(x_0))$ is precisely the light cone of $\iota(x_0)$ in $\RR^{n-1,1}$. We now define a bundle map $\Upsilon: TU \to T \RR^{n-1,1}$ over the inclusion $\iota : U \hookrightarrow \RR^{n-1,1}$ as follows: $\Upsilon$ agrees with $d \iota$ on $T_{x_0} U$. For any $x \in U \setminus C(x_0)$, and any vector $v \in T_x U$, write $v = v_r + v_{\perp}$, where $v_r$ is tangent to the radial geodesic passing through $x_0$ and $x$, and $v_{\perp}$ is orthogonal to this radial geodesic, and define $$\label{eqn:Pogorelov} \Upsilon(v) = \sqrt{\frac{\| \hat r \|^2}{\| d\iota(\hat r) \|^2}} d\iota(v_r) + d\iota( v_{\perp}),$$ where $\hat r$ is the unit radial vector (so $\|\hat r\|^2 = \pm 1$) and the norm $\| \cdot \|$ in the numerator of the first term is the AdS metric, while the norm in the denominator is the Minkowski metric. Note that a radial geodesic of $U$ (passing through $x_0$) is taken by $\iota$ to a radial geodesic in $\RR^{n-1,1}$ (passing through the origin) of the same type (space-like, light-like, time-like), although the length measure along the geodesic is not preserved. Hence the quantity under the square-root in is always positive. The key property of the infinitesimal Pogorelov map is the following (the proof is an easy computation in coordinates, see [@fillastre3 Lemma 3.4]). \[lm:pogorelov\] Let $Z$ be a vector field on $U \setminus C(x_0) \subset \AdS^n$. Then $Z$ is a Killing vector field if and only if $\Upsilon(Z)$ (wherever defined) is a Killing vector field for the Minkowski metric on $\R^{n-1,1}$. In fact, the lemma implies that the bundle map $\Upsilon$, which so far has only been defined over $U \setminus C(x_0)$, has a continuous extension to all of $U$. The bundle map $\Upsilon$ is called an *infinitesimal Pogorelov map*. Next, the bundle map $\Xi: T\R^{n-1,1}\to T\R^{n}$ over the identity, which simply changes the sign of the $n$-th coordinate of a given tangent vector, has the same property: it sends Killing vector fields in $\R^{n-1,1}$ to Killing vector fields for the Euclidean metric on $\R^{n}$. Hence the map $\Pi = \Xi \circ \Upsilon$ is a bundle map over the inclusion $U \hookrightarrow \RR^n$ with the following property: \[lm:pogorelov-Euclidean\] Let $Z$ be a vector field on $U \subset \AdS^n$. Then $Z$ is a Killing vector field if and only if $\Pi(Z)$ is a Killing vector field for the Euclidean metric on $\RR^n$. The bundle map $\Pi$ is also called an infinitesimal Pogorelov map. Henceforth we return to the setting of three-dimensional geometry. Rigidity of Euclidean polyhedra ------------------------------- In order to make use of the infinitesimal Pogorelov map defined above, we recall some elementary and well-known results about the rigidity of convex Euclidean polyhedra. It has been known since Legendre [@legendre] and Cauchy [@Cauchy] that convex polyhedra in Euclidean three-space $\RR^3$ are globally rigid. In fact, given two polyhedra $P_1, P_2$, if there is map $\partial P_1 \to \partial P_2$ which respects the combinatorics and is an isometry on each face, then the map is the restriction of a global isometry of Euclidean space. Later Dehn [@dehn-konvexer] proved that convex Euclidean polyhedra are also infinitesimally rigid. In fact, he showed that any first-order deformation $V$ of a polyhedron $P$ that preserves the combinatorics and the metric on each face is the restriction of a global Killing vector field. Here $V$ is not allowed, for example, to deform the polyhedron so that a quadrilateral face becomes two triangular faces. Still later, A.D. Alexandrov [@alexandrov] proved a stronger version of this statement: \[thm:alexandrov\] Let $P$ be a convex polyhedron in $\R^3$, and let $V$ be an infinitesimal deformation of $P$ (that might or might not change the combinatorics). Then, if the induced metric on each face is fixed, at first order, under $V$, the deformation $V$ is the restriction to $P$ of a global Euclidean Killing field. Proof of Lemma \[lem:Phi-rigidity\] (and Lemma \[lem:PsiAdS-rigidity\]) ----------------------------------------------------------------------- We first prove Lemma \[lem:Phi-rigidity\]. Lemma \[lem:PsiAdS-rigidity\] then follows from it and Theorem \[thm:duality\]. Let $P\in \AdSPoly_N$. We argue by contradiction and suppose that $\Phi$ is not a local immersion at $P$. This means that there exists a tangent vector $V$ to $\AdSPoly$ at $P$ such that $d\Phi(V)=0$. In other terms, there is a first-order deformation $V$ of $P$, as an ideal polyhedron in $\AdS^3$, which does not change the induced metric. Now, $V$ is described by tangent vectors $V_i \in T_{z_i} \partial_\infty \AdS^3$ at each ideal vertex $z_i$. Since $P$ is convex, it is contained in the complement $U \subset \AdS^3$ of a spacelike totally geodesic plane. We wish to use the Pogorelov map $\Pi$ defined in Section \[sec:Pogorelov\] above. However, $\Pi$ is *not* defined over the ideal boundary, so we need to be slightly careful. We may assume that the $1$–skeleton $\Gamma$ of $P$ is a triangulation. If not, we simply add diagonals to all of the non-triangular faces as needed. Consider a triangular face $T = \Delta z_{i_1} z_{i_2} z_{i_3}$. The tangent vectors $V_{i_1}, V_{i_2}, V_{i_3}$ determine a unique Killing field $X$, which defines the motion of the points of $T$ under the deformation. The deformation vectors for the vertices of an adjacent triangle $T' = \Delta z_{i_2} z_{i_1} z_{i_4}$ similarly determine a Killing field $X'$, which determines the motion of the points of $T'$. In general, $X$ and $X'$ might not agree on the common edge $e = z_1 z_2$. However, because $d\Phi(V) = 0$, the shear coordinate along $e$ does not change to first order, and therefore $X$ and $X'$ do agree along the edge $e$. It follows that $V$ defines a vector field $W$ on $\partial P$ whose restriction to any face agrees with a Killing field of $\AdS^3$. We now apply the Pogorelov map to obtain $\Pi(W)$, a vector field on the boundary of a convex polyhedron $\iota(P)$ in Euclidean space $\RR^3$. By Lemma \[lm:pogorelov-Euclidean\], the restriction of $\Pi(W)$ to each face of $\iota(P)$ agrees with a Euclidean Killing field. By Theorem \[thm:alexandrov\], $\Pi(W)$ must be the restriction of a global Euclidean Killing field $Y$. Hence, again, using Lemma \[lm:pogorelov-Euclidean\], we see that $W$ was the restriction of a global Killing field $\Pi^{-1}(Y)$ of $\AdS^3$ and therefore $V$ represents the trivial deformation in $\AdSPoly_N$. This completes the proof of Lemma \[lem:Phi-rigidity\]. Necessary conditions on the dihedral angles: proof of Proposition \[prop:imageinA-AdS\] {#sec:necess} ======================================================================================= In this section we prove Proposition \[prop:imageinA-AdS\], which states that the map $\PsiAdS_\Gamma$, taking an ideal polyhedron $P$ in $\AdS$ with $1$–skeleton $\Gamma$ to its dihedral angles $\theta = \PsiAdS_\Gamma(P)$, has image in the convex cone $\Angles_\Gamma$; in other words $\theta$ satisfies conditions (i), (ii), and (iii) of Section \[sec:intro-AdS-param\]. That $\theta$ satisfies (i) is just our sign convention for dihedral angles. That the dihedral angles $\theta$ satisfy (ii) follows exactly as in the hyperbolic setting: The intersection of $P$ with a small “horo-torus" centered about an ideal vertex of $P$ is a convex polygon in the Minkowski plane, whose exterior angles are equal to the corresponding exterior dihedral angles on $P$. The difficult part of Proposition \[prop:imageinA-AdS\] is to prove that $\theta$ satisfies condition (iii), and the remainder of this section is dedicated to this claim. Consider a simple cycle $e_0^*, e_1^*, \ldots, e_n^* = e_0^*$ in $\Gamma^*$ such that $\theta(e_j^*) < 0$ for exactly two edges $j = 1, r$. Let $f_i^*$ be the vertex of $\Gamma^*$, dual to a two-dimensional face $f_i$ of $P$, which is an endpoint of $e_i^*$ and $e_{i+1}^*$. In other words, the face $f_i$ of $P$ contains the edges $e_i$ and $e_{i+1}$. We must prove that the sum $\theta(e_1^*) + \cdots + \theta(e_n^*) > 0$. We now define a polyhedron $Q$ by “extending" the faces $f_1, \ldots, f_n$ and forgetting about the other faces of $P$. More rigorously: Since $P$ is contained in an affine chart of $\RP^3$, a lift $\tilde P$ of $P$ to the three-sphere $S^3$ is a convex polyhedron contained in an open half-space of $S^3$. Define $\tilde Q$ to be the intersection of the half-spaces defined by the lifts of $f_1, \ldots, f_n$. Then generically $\tilde Q$ will be contained in an open half-space, in which case $\tilde Q$ projects to a compact polyhedron $Q$ in some affine chart of $\RP^3$. We will, in a sense, reduce the generic case to the easier case that $\tilde Q$ is *not* contained in an open half-space, which we treat first. In this case, the combinatorial structure of $\tilde Q$ is very simple in that $\tilde Q$ has exactly two antipodal vertices. The projection $Q$ of $\tilde Q$ to $\RP^3$ has one vertex, which is contained in every face $f_1, \ldots, f_n$ and edge $e_1, \ldots, e_n$ of $Q$. Therefore $f_1, \ldots, f_n$ and $e_1, \ldots, e_n$ are orthogonal to the time-like plane $A$ dual to that vertex. As in Section \[sec:Pogorelov\], duality is defined with respect to the inner product of signature $(2,2)$ that defines $\AdS^3$. The intersection $q = A \cap Q$ is a convex compact polygon lying in $A \cong \AdS^2$ whose exterior angles are equal to the exterior dihedral angles of $Q$. That (iii) holds in this case now follows from: \[claim:AdS2-angles\] The sum of the exterior angles of a compact, convex, space-like polygon $q$ in $\AdS^2$ is strictly positive. This follows directly from the Gauss-Bonnet formula for Lorentzian polygons (see [@bir_gau]). Alternatively, one may easily prove the claim directly for triangles and then argue by induction. Before continuing to the general case, it is useful to examine the dual picture in this simple case. Let $q^*$ denote the dual convex polygon in $\RP^2$ (where $\RP^2$ is identified with its dual via the signature $(2,1)$ inner product defining $\AdS^2$). Since all edges of $q$ are space-like, the vertices of $q^*$ are contained in $\AdS^2$. If $v$ is a vertex of $q$ with positive exterior angle, then the dual edge $v^*$ in $q^*$ is a space-like edge contained in $\AdS^2$. However, if $v$ is a vertex of $q$ with negative exterior angle, then the dual edge $v^*$ of $q^*$ is begins and ends in $\AdS^2$ but contains a segment outside of $\AdS^2$. Conversely, the dual of any convex polygon $q^*$ in $\RP^2$ having the properties just described is a convex compact polygon in $\AdS^2$. Note that the length of an edge in $q^*$ is equal to the dihedral angle at the corresponding vertex of $q$, with the two edges which leave $\AdS^2$ having negative length. Therefore Claim \[claim:AdS2-angles\] is equivalent to: \[claim:AdS2-dual-lengths\] Let $q^*$ be a convex polygon in $\RP^2$ with vertices in $\AdS^2$ and with space-like edges, all but two (non-adjacent) of which are contained in $\AdS^2$. Then the sum of the lengths of the edges of $q^*$ is positive. This dual point of view will be useful in the general case, which we turn to now. Consider the generic case that the polyhedron $Q$ is compact in an affine chart of $\RP^3$. In this case, $Q$ will have extra edges, in addition to $e_1, \ldots, e_n$, which are not contained in $\AdS^3$; these edges may be either space-like or time-like. Let $Q^*$ denote the dual polyhedron in $\RP^3$, where we identify $\RP^3$ with its dual via the inner product of signature $(2,2)$ that defines $\AdS^3$. By perturbing a small amount if necessary, we may assume that all vertices of $Q$ lie outside of the closure of $\AdS^3$, so that the faces of $Q^*$ are each time-like. The vertices of $Q^*$, dual to the space-like faces $f_1, \ldots, f_n$, lie in $\AdS^3$. The dual edges $e_1^*, \ldots, e_n^*$ are space-like and form a Hamiltonian cycle in $\partial Q^*$ dividing it into two convex polyhedral surfaces $(\partial Q^*)_1$ and $(\partial Q^*)_2$. We need only work with one of these surfaces, say $(\partial Q^*)_1$. The surface $(\partial Q^*)_1$ is a polygon, bent along some interior edges. Note that two of the perimeter edges $e_1^*$ and $e_r^*$ of $(\partial Q^*)_1$ each contain a segment outside of $\AdS^3$, while $e_2^*, \ldots, e_{r-1}^*$ and $e_{r+1}^*, \ldots, e_n^*$ are contained in $\AdS^3$. We will show: \[lem:intrinsically-convex\] The surface $(\partial Q^*)_1$ is intrinsically locally convex. The lemma says that when $(\partial Q^*)_1$ is “un-folded" onto a time-like plane (a copy of $\AdS^2$), it is convex and therefore isomorphic to some $q^*$ as in Claim \[claim:AdS2-dual-lengths\] above. Therefore condition (iii) will follow from the lemma. Before embarking on the proof, we draw on some intuition from the Riemannian setting. To show that a developable polyhedral surface $S$ in a Riemannian space ($\RR^3$ say) is intrinsically locally convex, one must simply show that the total angle of $S$ at each vertex is less than $\pi$. Equivalently, one examines the *link* of each vertex $v$ of $S$, which is naturally a polygonal path in the unit sphere in the tangent space at $v$: $S$ is locally convex at $v$ if and only if the length of this polygonal path is less than $\pi$. We show that $(\partial Q^*)_1$ is locally convex in much the same way, by examining the link of each vertex of $(\partial Q^*)_1$ and measuring how long it is. However the space of rays emanating from a point in a Lorentzian space is not the Riemannian unit sphere, but rather what is called the *HS sphere* or $\HS^2$. The geometry of the HS sphere ----------------------------- *HS geometry*, introduced in [@shu; @cpt] and used recently in [@colI; @colII], is the natural local geometry near a point in a Lorentzian space-time such as $\AdS^3$. In those papers, HS-structures with cone singularities occur naturally as the induced geometric structures on the boundary of polyhedra or, in a related manner, on the links of vertices of the singular graph in Lorentzian 3-manifolds with cone singularities. Here we will use comparatively simpler notions without cone singularities. The tangent space at a point of $\AdS^3$ is a copy of the three-dimensional Minkowski space $\RR^{2,1}$. The *HS sphere* $\HS^2$ is the space of rays based at the origin in $\RR^{2,1}$. It admits a natural decomposition into five subsets: - Let $\HH^2_+$ (respectively $\HH^2_{-}$) denote the future oriented (resp. past oriented) time-like rays. Both $\HH^2_+$ and $\HH^2_-$ are copies of the Klein model for the hyperbolic plane and are equipped with the standard hyperbolic metric in the usual way. - Let $\dS^2$ denote the space-like rays, equipped with the standard de Sitter metric. - The light-like rays form two circles, $\partial\HH^2_+$ and $\partial\HH^2_{-}$, which are the boundaries of $\HH^2_+$ and $\HH^2_{-}$ respectively. The group $\operatorname{SO}_0(2, 1)$ of time-orientation and orientation preserving linear isometries of $\R^{2,1}$ acts naturally (and projectively) on $\HS^2$, preserving this decomposition. The geodesic $\sigma_{x,y}$ between two (non-antipodal) points $x,y \in \HS^2$ is defined to be the positive span of the two rays $x,y$. The space $\HS^2$ is equipped with a (partially defined) signed distance function $d(\cdot, \cdot)$ as follows. - If $x, y \in \HH^2_+$ or $x, y \in \HH^2_-$ then $d(x,y)$ is the usual hyperbolic distance, equal to the hyperbolic length of $\sigma_{x,y}$. - Let $x, y \in \dS^2$. We will only be interested in the case that $\sigma_{x,y}$ is time-like, meaning the plane in $\RR^{2,1}$ spanned by $\sigma_{x,y}$ has mixed signature. If $\sigma_{x,y}$ is contained in $\dS^2$, then $d(x,y)$ is defined to be the de Sitter length of $\sigma_{x,y}$, taken to be a negative (rather than imaginary) number. Note that in this case $d(x,y) = - d(x^*, y^*)$, where $x^*$ (resp. $y^*$) denotes the geodesic line dual to $x$ (resp. $y$) in $\HH^2_+$ (equal to the intersection with $\HH^2_+$ of the orthogonal complement of $x$ (resp. $y$)). In the case that $\sigma_{x,y}$ passes through $\HH^2_+$ (or $\HH^2_-$), we define $d(x,y) = +d(x^*, y^*)$. - Let $x \in \HH^2_+$ and $y \in \dS^2$. Then we define $d(x,y) = +d(x,y^*)$ if $x$ and $y$ lie on opposite sides of $y^*$ or $d(x,y) = -d(x,y^*)$ if $x$ and $y$ lie on the same side of $y^*$. We note that this distance function may be similarly defined in terms of the Hilbert distance (cross ratios) with respect to $\partial \HH^2_+$ or $\partial \HH^2_-$. Let $\sigma$ be a polygonal path in $\HS^2$ with endpoints $x,y \in \dS^2$ and call $\sigma$ *time-oriented* if $\sigma$ is the concatenation of three polygonal subpaths: a path crossing from from $x$ to $\HH^2_+$ which is future-oriented, followed by a path in $\HH^2_+$, followed by a path from $\HH^2_+$ back into $\dS^2$ which is past oriented. The length $\mathscr L(\sigma)$ is defined to be the sum of the lengths of the geodesic segments comprising $\sigma$. It is important to note that $\mathscr L(\sigma)$ is well-defined under sub-division. The crucial ingredient in the proof of Lemma \[lem:intrinsically-convex\] is the following substitute for the triangle inequality. \[cl:HS\] Let $\sigma$ be a time-oriented polygonal path with endpoints $x,y \in \dS^2$ and suppose further that $\sigma_{x,y}$ is time-like and crosses through $\HH^2_+$. Then $\mathscr L(\sigma) \geq \mathscr L(\sigma_{x,y})$. Let $x= x_0, x_1,\ldots, x_n = y$ be the ordered vertices of $\sigma$, with $x_0, \ldots, x_i$ lying in $\dS^2$, $x_{i+1}, \ldots, x_j$ lying in $\HH^2_+$, and $x_{j+1}, \ldots, x_n$ lying in $\dS^2$. Then, $$\begin{aligned} \mathscr L(\sigma) &= -\sum_{k=0}^{i-1} d(x_k^*, x_{k+1}^*) \ \ \ +\epsilon_1 d_(x_i^*, x_{i+1})\\ & + \sum_{k=i+1}^j d(x_{k}, x_{k+1}) \ \ \ +\epsilon_2 d_(x_j, x_{j+1}^*) -\sum_{k = j+1}^{n-1} d(x_{k}^*, x_{k+1}^*)\end{aligned}$$ where $\epsilon_1, \epsilon_2 = \pm 1$. We may assume, by sub-dividing, that $x_i$ and $x_{i+1}$ are on the same side of $x_i^*$ and that $x_j$ and $x_{j+1}$ are on the same side of $x_{j+1}^*$, so that $\epsilon_1 = \epsilon_2 = -1$. Therefore all of the dual lines $x_0^*, \ldots, x_i^*$ and $x_{j+1}^*, \ldots, x_n^*$ lie in between $x_{i+1}$ and $x_j$ in $\HH^2_+$. In fact, the dual lines are arranged, in order from closest to $x_{i+1}$ to closest to $x_j$, as follows: $x_i^*, x_{i-1}^*, \ldots, x_0^*, x_n^*, x_{n-1}^*, \ldots, x_{j+1}^*$. See Figure \[Fig:HS\]. Therefore, we have, by the triangle inequality in $\HH^2$, that $$\begin{aligned} d(x_{i+1}, x_{j}) \geq & \ \ d(x_{i+1}, x_i^*) + \sum_{k=1}^i d(x_k^*, x_{k-1}^*) \\ & + d(x_0^*, x_n^*) + \sum_{k=j+2}^ {n} d(x_k^*, x_{k-1}^*) + d(x_{j+1}^*, x_j)\end{aligned}$$ since the line connecting $x_{i+1}$ to $x_j$ crosses each of the dual lines in the above equation. Again by the triangle inequality in $\HH^2$, we also have $$d(x_{i+1}, x_{j}) \leq \sum_{k=i+1}^{j-1} d(x_{k}, x_{k+1}) .$$ It follows that $\mathscr L(\sigma_{x,y}) = d(x_0^*, x_n^*) \leq \mathscr L(\sigma)$. Proof of Lemma \[lem:intrinsically-convex\] ------------------------------------------- To complete the proof of Proposition \[prop:imageinA-AdS\], we now prove Lemma \[lem:intrinsically-convex\] which states that the convex pleated polygon $(\partial Q^*)_1$ is intrinsically locally convex. Consider a vertex $f_i^*$ of $(\partial Q^*)_1$. We consider the link $\sigma$ at $f_i^*$ of $(\partial Q^*)_1$, a polygonal path in the space of rays in $T_{f_i^*} \AdS^3$ which is naturally a copy of $\HS^2$. The endpoints $x$ and $y$ of $\sigma$ correspond to two consecutive dual edges $e_i^*$ and $e_{i+1}^*$ in the perimeter of $(\partial Q^*)_1$. Since the $e_j^*$ are space-like, $x,y$ lie in $\dS^2 \subset \HS^2$. By assumption, the edges $e_i$ and $e_{i+1}$ intersect outside of $\AdS^3$ (at a point which is positive with respect to the $(2,2)$ form), and therefore the plane containing $e_i^*$ and $e_{i+1}^*$ (which is dual to this point) is time-like, thus so is $\sigma_{x,y}$. By convexity of $Q$, the geodesic $\sigma_{x,y}$ passes through a hyperbolic region of $\HS^2$, which without loss in generality we take to be $\HH^2_+$. Further, by convexity of $Q$ and the fact that each of the faces of $\partial Q^*$ is time-like, the link $\sigma$ at $f_i^*$ of $(\partial Q^*)_1$ is time-oriented in the sense defined in the previous section. Therefore, it follows from Claim \[cl:HS\] that $\mathscr L(\sigma) \geq \mathscr L(\sigma_{x,y}) > 0$. Lemma \[lem:intrinsically-convex\] now follows because $\mathscr L(\sigma)$ is a complete invariant of the local geometry of $(\partial Q^*)_1$ at $f_i^*$; the development of $(\partial Q^*)_1$ onto a copy of $\AdS^2$ is convex at this vertex if and only if the length of the link is positive. This completes the proof of Proposition \[prop:imageinA-AdS\]. Topological arguments {#sec:topology} ===================== The topology of the space of ideal polyhedra {#sec:topolo_poly} -------------------------------------------- \[topol\_poly\] If $N \geq 3$, the space $\overline{\AdSPoly}_N$ is connected. If $N \geq 6$, then $\AdSPoly_N$ is connected and simply connected. By Proposition \[prop:cyclic-order\], the space $\overline{\AdSPoly}_N$ identifies with the space $\poly_N \times \poly_N$ of pairs $(p_L, p_R)$ of marked ideal $N$-gons in the hyperbolic plane considered up to the action of $\PSL_2 \RR \times \PSL_2\RR$. The space $\AdSPoly_N$ is obtained from $\overline{\AdSPoly}_N$ by removing all pairs $(p_L, p_R)$ such that $p_L$ and $p_R$ are isometric. Using the action of $\PSL_2 \RR \times \PSL_2\RR$ we may, in a unique way, put $p_L$ and $p_R$ into standard position so that the first three vertices of each polygon are $\infty,0$ and $1$. The remaining vertices of $p_L$ form an increasing sequence of $N-3$ points $x_4<\cdots < x_N$ in $(1,\infty)$. Similarly, the remaining vertices of $p_R$ also form an increasing sequence $ y_4 < \cdots < y_N$ in $(1,\infty)$ and $p_L$ is isometric to $p_R$ if and only if $(x_4, \ldots, x_N) = (y_4, \ldots, y_N)$. It follows that $\overline{\AdSPoly}_N$ is homeomorphic to $\RR^{N-3} \times \RR^{N-3}$ and $\AdSPoly_N$ is homeomorphic to $\RR^{N-3} \times \RR^{N-3}$ minus the diagonal. Therefore $\AdSPoly_N$ is homotopy equivalent to the sphere of dimension $N-4$. If $N \geq 6$, then $\HPPoly_N$ is connected and simply connected. Recall from Section \[subsec:polyhedra-HP\] that the space $\HPPoly_N$ identifies with the space of pairs $(p, V)$ where $p$ is a marked ideal $N$-gon in the hyperbolic plane and $V$ is a non-trivial infinitesimal deformation of $p$ considered up to the action of $T \PSL_2 \RR$. Using this action we may, in a unique way, place $(p, V)$ in standard position so that the first three vertices of $p$ are $x_1 = \infty, x_2 = 0, x_3 = 1$ and so that $V(x_1) = 0$, $V(x_2) = 0$, and $V(x_3) = 0$. The remaining $N-3$ tangent vectors are not all zero and their basepoints form an increasing sequence in $(1, \infty)$. It follows that $\HPPoly_N$ is homeomorphic to $T\RR^{N-3}$ minus the zero section. Therefore $\HPPoly_N$ is homotopy equivalent to the sphere of dimension $N-4$. As a corollary of Theorem \[thm:main-HP\] and this proposition we have: \[cor:topology\] The space of angle assignments $\Angles$ is connected and simply connected whenever the number of vertices $N \geq 6$. $\PsiAdS$ is a local homeo -------------------------- Lemma \[lem:PsiAdS-rigidity\] says that for each triangulation $\Gamma \in\Graph(N, \gamma)$, the map $\Psi_\Gamma: \AdSPoly \to \RR^E$ is a local immersion at any ideal polyhedron $P$ whose $1$–skeleton is contained in $\Gamma$. We now deduce the following result. \[lem:local-diffeo\] $\PsiAdS: \AdSPoly \to \Angles$ is a local homeomorphism. Given any $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$, we must first show that the dimension of $\Angles_\Gamma$ (if non-empty) is $2N -6$. The dimension of the convex cone $\Angles_\Gamma$ is determined by the rank of the $N$ equations of condition (ii). Assume first that $N$ is odd. Then these equations may be used to eliminate the the $N$ weights on the equator. Indeed if $\mathcal E_i$ denotes the equation of (ii) determined by the vertex $v_i$ of $\Gamma$, then treating indices cyclically we find that $$\mathcal E_{j+1} - \mathcal E_{j+2} + \cdots - \mathcal E_{j-1} + \mathcal E_j$$ is an equation which depends on (the weight at) the edge $e_j$ with endpoints $v_j$ and $v_{j+1}$ but on no other edge of the equator. This shows that the equations $\mathcal E_1, \ldots, \mathcal E_N$ have rank $N$ and the dimension of $\Angles_\Gamma$ is therefore $3N -6 - N = 2N-6$. Next if $N$ is even, we may only eliminate $N-1$ of the weights on the equator because all equatorial weights cancel in the alternating sum: $$\mathcal E_1 - \mathcal E_2 + \cdots +\mathcal E_{N-1} - \mathcal E_N.$$ However, note that this sum is not trivial since it depends non-trivially on (the weight at) any edge whose two endpoints are an even number of edges apart along the equator. Since $\Gamma$ is a triangulation, there must exist some such edge. So the equations defined by condition (ii) in the definition of $\Angles_\Gamma$ have rank $N$ in this case as well. Next, for each triangulation $\Gamma$, let $V_\Gamma \subset \RR^{E(\Gamma)}$ be the subspace satisfying the equations of condition (ii). Since $V_\Gamma$ has dimension $2N-6$, as shown above, each of the maps $\PsiAdS_\Gamma$ is a local diffeomorphism at any polyhedron $P$ whose $1$–skeleton is a subgraph of $\Gamma$. The map $\PsiAdS$, pieced together from the $\PsiAdS_\Gamma$over all $\Gamma$, is an open map by the definition of the topology of the complex $\Angles$. Further, since each $\PsiAdS_\Gamma$ is a local diffeomorphism in a neighborhood of any point in the *closure* of the stratum of $\AdSPoly$ defined by $\Gamma$, we have that $\PsiAdS$ is a local bijection to $\Angles$. It follows that $\PsiAdS$ is a local homeomorphism. Lemma \[lem:local-diffeo\] and Lemma \[lem:PsiAdS-proper\] imply that $\PsiAdS$ is a covering. Since for $N \geq 6$, $\AdSPoly$ is connected and $\Angles$ is connected and simply connected, we conclude that Theorem \[thm:main-AdS-angles\] holds when $N \geq 6$. The cases $N=4,5$ ----------------- Although the topology of $\Angles$ is slightly more complicated when $N=4,5$, the proof of Theorem \[thm:main-AdS-angles\] is straightforward in these cases. In the case $N=4$, the space $\AdSPoly$ is the space of marked (non-degenerate) ideal tetrahedra and has two components and the map $\PsiAdS$ is easily seen to be a homeomorphism. Indeed, an ideal tetrahedra in $\AdS^3$ is determined by its shape parameter (see Section \[ads\_background\]); its dihedral angles may be determined directly from the shape parameter. Conversely, the shape parameter is determined by any two angles along edges emanating from a common vertex. Therefore an ideal tetrahedron is determined entirely by the local geometry near any ideal vertex. In the case $N=5$, both $\AdSPoly$ and $\Angles$ are homotopy equivalent to the circle. To show that the map $\PsiAdS$ is a homeomorphism, rather than some non-trivial covering, consider an ideal polyhedron $P$. We may cut $P$ into two ideal tetrahedra $T, T'$ along some interior triangular face $\Delta$. The tetrahedron $T$ is determined by the angles along the three edges emanating from any vertex of $T$, in particular the vertex not belonging to $\Delta$. These three angles are dihedral angles of $P$ as well, so it follows that the geometry of $T$ is determined by the dihedral angles of $P$. Similarly, the geometry of $T'$ is determined by the dihedral angles of $P$. Since there is exactly one way to glue $T$ and $T'$ back together (with the correct combinatorics), the geometry of $P$ is determined by its dihedral angles, i.e. $\PsiAdS$ is injective, and is therefore a homeomorphism. Proof of Theorem \[thm:main\] {#sec:proof_main} ----------------------------- Finally, we prove Theorem \[thm:main\]. The equivalence of (C) and (H) is immediate from Theorems \[thm:main-HP\] and \[thm:main-AdS-angles\]. We now show the equivalence of (H) and (S) using Theorem \[thm:main-AdS-angles\] and Rivin’s Theorem, discussed in Section \[sec:rivin\_param\]. Let $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$, and as usual let $E = E(\Gamma)$ denote the edges of $\Gamma$. First suppose $P\in \AdSPoly_\Gamma$, and let $\theta = \PsiAdS(P)\in \Angles_{\Gamma}$. For any $t > 0$, the weights $t \theta$ are also in $\Angles_\Gamma$. We choose $t > 0$ so that: - for all edges $e \in E$, $t\theta(e)\in (-\pi,\pi)\setminus\{0\}$. - for all of the finitely many simple cycles $c$ in $\Gamma^*$, the sum of the values of $t \theta$ along $c$ is greater than $-\pi$. Note that any simple cycle $c$, as in (B) above, crosses the equator $\gamma$ at least twice. If $c$ crosses the equator $\gamma$ exactly twice, then this sum will either be zero, if $c$ bounds a face of $\Gamma^*$ (condition (ii) in the definition of $\Angles_\Gamma$), or positive if not (condition (iii)). Noting that $t \theta(e) \in (-\pi, 0)$ if $e \in \gamma$ and $t \theta(e) \in (0, \pi)$ if not, we let $\theta': E \to (0,\pi)$ be defined by $$\theta'(e) = \left\{ \begin{array}{ll} t\theta(e) & \mbox{if $e$ is not an edge of $\gamma$,} \\ \pi+t\theta(e) & \mbox{if $e$ is an edge of $\gamma$.} \end{array} \right.$$ Then $\theta'$ satisfies the three conditions of Rivin’s Theorem and is therefore realized as the dihedral angles of some ideal polyhedron $P'$ in $\HH^3$. In the projective model for $\HH^3$, $P'$ is a polyhedron inscribed in the sphere with $1$–skeleton $\Gamma$. Conversely, suppose $P'$ is an ideal polyhedron in $\HH^3$ with $1$–skeleton $\Gamma$. Then the dihedral angles $\theta':E \to (0,\pi)$ of $P'$ satisfy the three conditions of Rivin’s Theorem. We define $\theta: E \to \RR$ by $$\theta(e) = \left\{ \begin{array}{ll} \theta'(e) & \mbox{if $e$ is not an edge of $\gamma$,} \\ \theta'(e)-\pi & \mbox{if $e$ is an edge of $\gamma$.} \end{array} \right.$$ Then $\theta$ is easily seen to satisfy the three conditions in the definition of $\Angles_\Gamma$ and so by Theorem \[thm:main-AdS-angles\], $\theta = \PsiAdS(P)$ for some $P \in \AdSPoly$. In the projective model for $\AdS^3$, $P$ is a polyhedron inscribed in the hyperboloid with $1$–skeleton $\Gamma$. This completes the proof of Theorem \[thm:main\]. Let $\Gamma$ be a planar graph and suppose $\Gamma$ is realized as the $1$–skeleton of some ideal polyhedron inscribed in the sphere. Note that $\Gamma$ may contain many different Hamiltonian cycles. Applying the above to each Hamiltonian cycle $\gamma$ shows the following: The components of the space of realizations of $\Gamma$ as the $1$–skeleton of a polyhedron inscribed in the hyperboloid (or similarly, the cylinder) is in one-one correspondence with the Hamiltonian cycles in $\Gamma$. Ideal polyhedra with dihedral angles going to zero {#sec:transitional-proof} ================================================== We outline an alternative proof of Proposition \[prop:imageinA-AdS\] using transitional geometry ideas. The argument uses Lemmas \[lem:PsiAdS-proper\] and \[lem:PsiAdS-rigidity\] to produce deformation paths of polyhedra with dihedral angles going to zero in a prescribed manner. Here is the basic idea: starting from an ideal polyhedron $P \in \AdSPoly$ with dihedral angles $\theta$, we deform $P$ so that the dihedral angles are proportional to $\theta$ and decrease toward zero. An appropriate rescaled limit of these collapsing polyhedra yields an ideal polyhedron $P_\infty'$ in $\HP^3$ whose (infinitesimal) dihedral angles are precisely $\theta$; we then conclude, via Proposition \[prop:imageinA-HP\], that $\theta$ was in $\Angles$ to begin with. The main ingredient is the following proposition. Recall the projective transformations $\mathfrak a_t$ of Section \[subsec:polyhedra-HP\], which when applied to (the projective model of) $\AdS^3$ yield $\HP^3$ in the limit as $t \to 0$. \[prop:flat-limit\] Let $\Gamma \in\Graph(\Sigma_{0,N}, \gamma)$ and consider weights $\theta \in \RR^{E(\Gamma)}$ that satisfy conditions (i), (ii), and the following weaker version of (iii): - If $e_1^*, \ldots, e_n^*$ form a simple circuit that does not bound a face of $\Gamma^*$, and such that exactly two of the edges are dual to edges of the equator, then $\theta(e_1^*) + \cdots + \theta(e_n^*) \neq 0$. Let $P_k$ be a sequence in $\AdSPoly_{\Gamma}$ with dihedral angles $t_k \theta$ such that $t_k \to 0$. Then: 1. $P_k$ converges to an ideal $N$-gon $P_\infty$ in the hyperbolic plane. 2. $\mathfrak a_{t_k} P_k$ converges to an ideal polyhedron $P_\infty'$ in $\HP^3$ with $1$–skeleton $\Gamma$ and infinitesimal dihedral angles $\theta$. We briefly mention an analogue of the proposition in setting of quasi-Fuchsian hyperbolic three-manifolds. The first conclusion of the proposition can be seen as an analogue of Series’ theorem [@ser_lim], which states that when the bending data of a sequence of quasi-Fuchsian representations goes to zero in a controlled manner, the convex cores collapse to a Fuchsian surface. The second part is the analogue of work of Danciger–Kerckhoff [@dan_tra] showing that after application of appropriate projective transformations (in our notation, the $\mathfrak a_t$), the collapsing convex cores of such quasi-Fuchsian representations converge to a convex core in half-pipe geometry. We adapt the proof of Lemma \[lem:PsiAdS-proper\] (properness of the map $\PsiAdS$). As in that proof, we may again assume that the ideal vertices $(v^L_{1,k}, v^R_{1,k}), \ldots, (v^L_{N,k}, v^R_{N,k})$ of $P_k$ satisfy that: - $v^L_{1,k} = v^R_{1,k} = 0$, $v^L_{2,k} = v^R_{2,k} = 1$, $v^L_{3,k} = v^R_{3,k} = \infty$; - For each $i \in \{1,\ldots N\}$, $v^L_{i,k} \to v^L_{i,\infty}$ and $v^R_{i,k} \to v^R_{i,\infty}$; and - $v^L_{i, \infty} = v^L_{i+1, \infty}$ if and only if $v^R_{i, \infty} = v^R_{i+1, \infty}$. Therefore, we again find that the limit $P_\infty$ of $P_k$ (in this normalization) is a convex ideal polyhedron in $\AdS^3$, possibly of fewer vertices, and possibly degenerate (i.e. lying in a two-plane). The dihedral angle at an edge $e$ of $P_\infty$ is again the sum of $\theta_\infty(e')$ over all edges $e'$ of $\Gamma$ which collapse to $e$, where in this case $\theta_\infty = 0$. Therefore all dihedral angles of $P_\infty$ are zero and we have that $P_\infty$ is an ideal polygon lying in the hyperbolic plane $\plane$ containing the ideal triangle $\Delta_0$ spanned by $(0,0), (1,1)$, and $(\infty, \infty)$. To prevent collapse, we apply the projective transformations $\mathfrak a_{t_k}$ to the $P_k$. Up to taking a subsequence (in fact not necessary), the vertices $\mathfrak a_{t_k} v_{i,k}$ converge to points $v_{i,\infty}'$ in the ideal boundary $\partial_\infty \HP^3$. This can be seen from the following simple compactness statement, which may be verified by induction: Given $M \geq 1$ and $\Theta > 0$, there exists two smooth families of space-like planes $\mathcal Q_+(t)$ and $\mathcal Q_-(t)$, defined for $t \geq 0$, such that - $\mathcal Q_+(0) = \mathcal Q_-(0) = \plane$. - $\mathcal Q_+(t)$ and $\mathcal Q_-(t)$ are disjoint for $t > 0$ and their common perpendicular is a fixed time-like line $\alpha$ (independent of $t$). - The time-like distance (along $\alpha$) between $\mathcal Q_+(t)$ and $\mathcal Q_-(t)$ is $\operatorname{O}(t)$. - Any space-like convex connected ideal polygonal surface in $\AdS^3$ for which $\Delta_0$ is (contained in) a face, which has at most $M$ faces, and all of whose dihedral angles are bounded by $t \Theta$ lies to the past of $\mathcal Q_+(t)$ and to the future of $\mathcal Q_-(t)$. The first three conditions above imply that the limit of $\mathfrak a_t \mathcal Q_+(t)$ and $\mathfrak a_t \mathcal Q_-(t)$ as $t \to 0$ are two disjoint non-degenerate planes $\mathcal Q_+'$ and $\mathcal Q_-'$ in $\HP^3$. Therefore, the limit of $\mathfrak a_{t_k} P_k$ must, after extracting a subsequence if necessary, converge to some polyhedron in $\HP^3 \cup \partial_\infty \HP^3$ lying below $\mathcal Q_+'$ and above $\mathcal Q_-'$. As in the proof of Lemma \[lem:PsiAdS-proper\], the limit of $\mathfrak a_{t_k} P_k$ is the convex hull $P_\infty'$ of $v_{1,\infty}', \ldots, v_{n, \infty}'$ in $\HP^3$. The $1$–skeleton $\Gamma'$ of $P_\infty'$ is obtained from the original $1$–skeleton $\Gamma$ by collapsing some edges to vertices and some faces to edges or vertices. A simple argument in HP geometry gives that: Given $e' \in \Gamma'$, the infinitesimal dihedral angle $\theta_\infty'(e')$ of $P_\infty'$ at $e'$ is the sum of $\theta(e) = \frac{d}{dt} t \theta(e)\big|_{t=0}$ over all edges $e$ which collapse to $e'$. Next, consider the projection $\varpi: \operatorname{HP}^3 \to \plane$. Note that $\varpi(v_{i,\infty}') = v_{i,\infty}$. Let $\mathcal H$ denote the HP horo-cylinder which is the inverse image under $\varpi$ of a small horocycle in $\plane$ centered at a vertex $v_{i, \infty}$ of $P_\infty$. The metric on $\mathcal H$ inherited from $\HP^3$ is flat and degenerate; it is the pull-back under $\varpi$ of the metric on a horocycle. The intersection of $\mathcal H$ with $P_\infty'$ is a convex polygon $q$ in $\mathcal H$. The infinitesimal angles at vertices of $q$ are the same as the infinitesimal dihedral angles of the corresponding edges of $P_\infty'$. Note that the vertices of $q$ are the intersection with $\mathcal H$ of all edges emanating from the ideal points $v_{j, \infty}'$ such that $v_{j, \infty} = v_{i,\infty}$. A simple calculation in this degenerate plane shows that: The infinitesimal angles of $q$ sum to zero. Now, suppose, for contradiction, that $v_{i+1, \infty} = v_{i, \infty}$. Then, the vertices of $q$ correspond to a path $c'$ of edges of $\Gamma'$ whose inverse image under the collapse is a path $c$ of edges in $\Gamma$ which do not bound a face of $\Gamma^*$. It follows from the above that the sum of $\theta(e)$ over the edges $e$ of the path $c$ is zero, contradicting the condition (iii’). This argument also works in the context of hyperbolic ideal polyhedra with dihedral angles going to zero and $\pi$ at a controlled rate. Assuming the stronger condition (iii) on $\theta$, the limiting ideal polygon $P_\infty$ must be the unique minimum of the length function $\ell_\theta$ over the space $\poly_N$ of marked ideal polygons. See the proof of Theorem \[thm:main-HP\]. Let $\Gamma \in \Graph(\Sigma_{0,N}, \gamma)$ and suppose $P \in \AdSPoly_\Gamma$ such that the dihedral angles $\theta = \PsiAdS(P) \in \RR^{E(\Gamma)}$ violate condition (iii) in the definition of $\Angles_\Gamma$. We argue by contradiction. First we show that there are nearby weights $\theta'$ satisfying conditions (i), (ii), as well as condition (iii’) of Proposition \[prop:flat-limit\] above and so that at least one of the angle sum expressions of (iii’) is strictly *negative*. This may already be the case for $\theta$. If not, then there is at least one angle sum expressions as in (iii) which evaluates to zero, and we will perturb. In the case that $\Gamma$ is a triangulation, it is simple to verify that that none of the angle sum expressions in condition (iii) is locally constant when the equations of condition (ii) are satisfied, and therefore a nearby $\theta'$ exists as desired, since (iii’) consists of only finitely many conditions . If $\Gamma$ is not a triangulation, then it could be the case that an angle sum expression as in condition (iii) is constant equal to zero on the entire space of weights satisfying (ii) (see Figure \[fig:strange-example\] for an example!). A quick study of the possible ways such degenerate behavior can happen reveals that it is always possible to add a small number of edges (at most one for each angle sum expression of (iii) which evaluates to zero for $\theta$) with very small positive weights, while perturbing the other weights slightly, to produce $\theta'$ as desired. Next, by Lemma \[lem:local-diffeo\] (which was a simple consequence of Lemma \[lem:PsiAdS-rigidity\], independent of Proposition \[prop:imageinA-AdS\]), there is an ideal polyhedron $P' \in \AdSPoly$, close to $P$, so that $\PsiAdS(P') = \theta'$. Now, consider the path of weights $t\theta'$, defined for $t > 0$. Lemma \[lem:local-diffeo\] implies that there exists a path $P_t$ in $\AdSPoly$ such that $\PsiAdS(P_t) = t \theta'$, defined at least for $t$ close to one. In fact, the path $P_t$ may be defined for all $1 \geq t > 0$. Indeed if the limit as $t \to T > 0$ of $P_t$ failed to exist, then the proof of Lemma \[lem:PsiAdS-proper\] would imply that $\PsiAdS(P_t)$ either goes to infinity or limits to an element of $\RR^{E(\Gamma)}$ for which some angle sum expression as in (iii) is exactly zero, impossible since the limit as $t \to T$ of $\PsiAdS(P_t)$ is, of course, equal to $T \theta$. Hence, we may apply Proposition \[prop:flat-limit\] to the path $P_t$. The result is an ideal polyhedron $P_\infty' \in \HPPoly$ whose infinitesimal dihedral angles are precisely $\theta'$. This contradicts Proposition \[prop:imageinA-HP\] since $\theta'$ does not satisfy (iii). [^1]: The first author was partially supported by the National Science Foundation under the grant DMS 1103939. The second author was partially supported by the European Research Council under the [*European Community*]{}’s seventh Framework Programme (FP7/2007-2013)/ERC [*grant agreement*]{}

ArXiv

--- abstract: | Recent papers published in the last years contributed to resolve the enigma on the hypothetical Be nature of the hot pulsating star $\beta$ Cep. This star shows variable emission in the H$\alpha$ line, typical for Be stars, but its projected rotational velocity is very much lower than the critical limit, contrary to what is expected for a typical Be star. The emission has been attributed to the secondary component of the $\beta$ Cep spectroscopic binary system. In this paper, using both ours and archived spectra, we attempted to recover the H$\alpha$ profile of the secondary component and to analyze its behavior with time for a long period. To accomplish this task, we first derived the atmospheric parameters of the primary: T$_{\rm eff}$=24000$\pm$250 K and $\log g$=3.91$\pm$0.10, then we used these values to compute its synthetic H$\alpha$ profile and finally we reconstructed the secondary’s profile disentangling the observed one. The secondary’s H$\alpha$ profile shows the typical two peaks emission of a Be star with a strong variability. We analyzed also the behavior versus time of some line width parameters: equivalent width, V/R, FWHM, peaks separation and radial velocity of the central depression. Projected rotational velocity ($v \sin i$) of the secondary and the dimension of the equatorial surrounding disk have been estimated, too. date: 'Accepted 2008 March 27. Received 2008 March 27; in original form 2008 February 12' title: 'H$\alpha$ long term monitoring of the Be star $\beta$ Cep Aa' --- \[firstpage\] stars: emission-line, Be – stars: individual: $\beta$ Cephei – (stars:) binaries: spectroscopic Introduction ============ The star $\beta$ Cephei (HD205021, V=3.2) is well known to be the prototype of a class of hot pulsating stars. Several studies reported in the recent literature have been outlined a star oscillating with at least five frequencies: $f_1$=5.2497104 c d$^{-1}$, $f_2$=5.385 c d$^{-1}$ and $f_3$=4.920 c d$^{-1}$ [@aerts94] and $f_4$=5.083 c d$^{-1}$ and $f_5$=5.417 c d$^{-1}$ [@telting97]. Further, the star ($\beta$ Cep A) is actually a complicated multiple system. In fact, beside to have a visual companion ($\beta$ Cep B, V=7.9) at a distance of 13.4”, it is also a member of a spectroscopic system whose second star ($\beta$ Cep Aa) was discovered using speckle interferometry by @gezari72 at a distance of $\approx$0.25”. The parameters of the close binary orbit have been determined later by @pigu92 combining speckle interferometry and the variations in the pulsation period, due to the so-called light time effect. Recent speckle measurements by @hartkopf01 place the position of the companion to about 0.1” from the primary and @balega02 found a magnitude difference of 1.8 between the two components in a red filter centered on 810 nm and 60 nm wide. More observations are still necessary to improve the orbital solution. Another characteristic that gives more interest to this object is the variable emission observed in the H$\alpha$. Since @karpov33, the presence of this emission had been reporting by several authors: @wilson56 or @kaper95, who observed a decreasing emission from 1990 to 1995. In the same years, $\beta$ Cep has been observed with the coudé spectrograph of the 2.6 meter telescope of the Crimean Astrophysical Observatory by @panko97. Over the entire period covered by their observations, the H$\alpha$ was in emission with a pronounced two component structure and a small intensity above the continuum level. The same authors stated that in November 1987 the line showed an absorption profile without any sign of emission. The kind of variable emission discussed so far is typically the characteristic of Be stars. Be stars are rapidly rotating B-type stars that lose mass in an equatorial, circumstellar disk (see the recent review by @porter03). A general characteristic of Be stars is that they rotate very fast, typically at about 70$\%$-80$\%$ of their critical limit (v$_{\rm crit}$=$\sqrt{GM_*/R_*}$) corresponding to a several hundred of km s$^{-1}$ [@slett82]. On the contrary, the projected rotational velocity of $\beta$ Cep A (26 km s$^{-1}$, @morel06) is very much lower than this limit and this questions its real nature. @hadrava96 separated emission and absorption components of the H$\alpha$ profile from their 1996 spectra and found that they move in antiphase. They speculated on two possible scenarios: [*i)*]{} the observed emission arises from reprocessing of the light originating in the stellar photosphere in an outer envelope, [*ii)*]{} the star is not a pulsating object but a rapidly rotating oblique rotator with a magnetic dipole geometry. Recently, @schnerr06 argued for the first time that the H$\alpha$ emission observed from the $\beta$ Cep system is not related to the slowly rotating primary star, but to the secondary, being the latter a classical Be star. Unfortunately, they obtained only one spectrum at the [*Nordic Optical Telescope*]{} and, then, no conclusion about the variability of the emission could be drawn. In this paper we try to recover the recent emission history of $\beta$ Cep Aa. To achieve this goal, we firstly disentangled the spectra of the two components from the observed one. This has been done using a synthetic profile for the H$\alpha$ of $\beta$ Cep A, for which T$_{\rm eff}$ and $\log g$ have been estimated as described in Sect. \[primary\]. Then, we analyzed the behavior with time of some H$\alpha$ width parameters like its equivalent width, the full width at half maximum, the peaks separation and the radial velocity of the central depression. Finally, we attempted to estimate the $v \sin i$ of $\beta$ Cep Aa and the dimension of its surrounding disk. Observation and data reduction {#observ} ============================== We observed $\beta$ Cep for more than one year from July 2005 to November 2006. To extend our coverage to the past years, we queried various astronomical archives on the internet. At the end, we collected a number of 72 spectra covering the period from September 1993 to November 2006, for a total of more than 13 years. In the following we detailed each single observatory and instrument: - 42 spectra have been taken from the archive of the [*Ritter Observatory*]{}, University of Toledo, OH USA, equipped with 1-m Ritchey-Chretien reflector. The echelle spectrograph is connected to the Cassegrain focus of the telescope by fiber optic cable of 200 $\mu$m diameter. The detector is a CCD manufactured by EEV with 1200 x 800 pixels of 22.5 $\mu$m size. The spectra used in this paper have been observed with R$\approx$26000; - 4 spectra have been downloaded from the [*Isaac Newton Telescope*]{} archive, three of them have been acquired with IDS spectrograph and one with MUSICOS. The lines of the Th-Ar lamp show that for all the spectra the resolution is R$\approx$33000; - 7 spectra have been downloaded from the archive of the Elodie@OHP 1.93 m telescope. Spectra have been reduced by the standard pipeline procedure, described in @baranne96. Resolving power of these spectra is $\approx$42000; - the 91 cm telescope of the [*INAF-Osservatorio Astrofisico di Catania*]{}, has been used by us to carried out 19 spectra of our target. The telescope is fiber linked to a REOSC echelle spectrograph, which allows to obtain R=20000 spectra in the range 4300-6800 [Å]{}. The resolving power has been checked using emission lines of the Th-Ar calibration lamp. Spectra were recorded on a thinned, back-illuminated (SITE) CCD with 1024 x 1024 pixels of 24 $\mu$m size, typical readout noise of 6.5 e$^{-}$ and gain of 2.5 e$^-$/ADU. These spectra have been extracted from a more complete set of observations [@catanzaro08], in such a way to choose the one with the best SNR for each observing night. The stellar spectra, calibrated in wavelength and with the continuum normalized to a unity level, were obtained using standard data reduction procedures for spectroscopic observations within the NOAO/IRAF package, that is: bias frame subtraction, trimming, scattered light correction, flat-fielding, fitting traces and orders extraction and, finally, wavelength calibration. IRAF package [*rvcorrect*]{} has been used to include the velocity correction due to the Earth’s motions, all the spectra were then reduced into the heliocentric rest of frame. Atmospheric parameters of the primary {#primary} ===================================== The first task we had to asses in our study was to compute the synthetic H$\alpha$ profile of the primary component that we used in the reconstruction process of the secondary’s profile. This calculation is possible if one knows the effective temperature and surface gravity of the primary component of the system. The approach we used in this study to determine T$_{\rm eff}$ and $\log g$ was to compare the observed and theoretical profiles of H$\gamma$ and H$\beta$ lines by minimizing the goodness-of-fit parameter:\ $\chi^2 = \frac{1}{N} \sum \left(\frac{I_{\rm obs} - I_{\rm th}}{\delta I_{\rm obs}}\right)^2$ where $N$ is the total number of points, $I_{\rm obs}$ and $I_{\rm th}$ are, respectively, the intensities – normalized to the continuum – of the observed and computed profiles and $\delta I_{\rm obs}$ is the photon noise. Errors in T$_{\rm eff}$ and $\log g$ are estimated as the variation in the parameters which increases the $\chi^2$ of a unit. ![Comparison between observed (black) and synthetic (red) profile of H$\gamma$ and H$\beta$ lines. Atmospheric parameters used for this calculation are: T$_{\rm eff}$=24000$\pm$250 K and $\log g$=3.91$\pm$0.10. For each Balmer line we reported also the residuals (O-C) expressed in percentage. Note the strong oxygen lines present in the wings of the H$\gamma$ line.[]{data-label="balmer"}](balmer.ps){width="9cm"} The observed profiles presented in Fig. \[balmer\] have been calculated by averaging a number of 25 spectra collected by us during the Julian Day 2554041.45 [@catanzaro08]. Doppler correction has been applied to remove any radial velocity shift due to pulsations. This average spectrum has been used in order to improve the signal-to-noise ratio of each single exposures, in particular we obtained SNR$\approx$250 in the continuum close to the H$\gamma$ and SNR$\approx$300 in the continuum close to the H$\beta$. Theoretical profiles have been computed with SYNTHE[^1] [@kur81] on the basis of ATLAS9 [@kur93] atmosphere models. To reduce the number of free parameters, the rotational velocity of $\beta$ Cep A has been determined fitting the line profiles of the Si[iii]{} triplet at $\lambda$4552, 4567 and 4574 [Å]{}. The value obtained of $\rm v \sin i$=30 $\pm$ 1 km s$^{-1}$ is almost coincident with the line broadening of 29 km s$^{-1}$ measured by @morel06 in their analysis of high resolution spectra (R=50000). The best fit was achieved for the following adopted parameters: T$_{\rm eff}$=24000$\pm$250 K and $\log g$=3.91$\pm$0.10. In Fig. \[balmer\] we showed the comparison between the observed and computed lines profiles together with their residuals (O - C). It is interesting to note the two symmetrical bumps that appear both on the blue and red side of H$\beta$, approximately at $\pm$150 km s$^{-1}$ from the core of the line. They are located at the same velocity of the peaks observed in the H$\alpha$ of $\beta$ Cep Aa (see Fig. \[example\]), then they could be connected to the H$\beta$ of the secondary component. On the other side, $\beta$ Cep A is a non-radial pulsator, thus the peaks could be the result of the average process we undertook to recover the mean profiles used for T$_{\rm eff}$ and $\log g$ determination. This distortion, whatever its origin may be, could led to a maximum error on the parameters of the order of 2 $\%$. ![In this figure we illustrated the method used to recover the H$\alpha$ profile of the secondary. With bold solid line we represented the observed composite spectrum, with red dashed line we represented the synthetic primary component and with the normal line we indicated the reconstructed secondary profile.[]{data-label="example"}](halpha_ex.ps){width="9cm"} {width="18cm"} {width="18cm"} In the particular case of our target, this procedure could be considered valid under the initial hypothesis that H$\gamma$ and H$\beta$ are not influenced by the secondary component of the binary system. As a check on the consistency of our T$_{\rm eff}$ and $\log g$, we estimated the abundance of silicon from lines of different ionization states. In particular we used the Si[iii]{} triplet and the Si[iv]{} $\lambda$4654 [Å]{}. The abundance has been computed by means of spectral synthesis analysis using the microturbulence found by @gies92, that is: 7.6$\pm$1.5 km s$^{-1}$. We derived $-$4.67$\pm$0.08 from Si[iii]{} and $-$4.65$\pm$0.10 from Si[iv]{}. The consistency between these two determinations convinced ourselves about the accuracy of our atmospheric parameters. As adopted Si abundance we computed the weighted average between the two previous values, that is $\log$N(Si)/N$_{\rm tot}$=$-$4.66$\pm$0.06. This abundance, that could be expressed as $\log \epsilon(Si)$=7.38$\pm$0.06[^2], is fully consistent with typical abundances for B-type field stars reported in various literature study: for instance, @dufton90 gave $\log \epsilon(Si)$=7.5$\pm$0.2 for both the OB associations $\it h$ and $\chi$ Per and Cep OB3, @gies92 found $\log \epsilon(Si)$=7.58 from their sample of B stars and @cunha94 found 7.40$\pm$0.15 in a sample of B stars in the Orion association. For the sake of comparison, we searched for other determinations of T$_{\rm eff}$ and $\log g$ in the literature. In particular, @heyn94 found T$_{\rm eff}$=24550 and $\log g$=3.772 and @niec05 derived T$_{\rm eff}$=24150$\pm$350 K and $\log g$=3.69. As the reader can see, despite effective temperatures are perfectly consistent with our value, our gravity is $\approx$0.2 dex greater than previous determinations. Anyway, since other authors did not report their errors on $\log g$, no conclusion could be drawn about the consistency between our and literature values. Extraction of the secondary spectrum ==================================== The procedure we undertook to recover the H$\alpha$ profiles of the secondary component could be summarized as follows: - first, we computed synthetic H$\alpha$ profiles for the primary component using the atmospheric parameters derived in the previous section. Even for this line we used ATLAS9 and SYNTHE; - then, we recovered the secondary H$\alpha$ line profiles from the following formula: $$F_{\rm tot}\,=\,\frac{F_{\rm prim}~r~+~F_{\rm sec}}{1~+~r} \label{ftot}$$ where $F_{\rm tot}$ is the observed profile, $F_{\rm prim}$ is the synthetic H$\alpha$ profile of the primary, $r$ is the luminosity ratio between the two components computed from the magnitude difference $\Delta\,M$=1.8 [@balega02] opportunely scaled at the H$\alpha$ wavelength. Then from Eq. \[ftot\] it follows: $$F_{\rm sec}\,=\,(1~+~r)~F_{\rm tot}~-~F_{\rm prim}\,r$$ Since the primary component is a pulsating star with an amplitude of about 30 km s$^{-1}$ (@aerts94, @telting97, @catanzaro08), it has been necessary to correct the synthetic spectrum for the radial velocity corresponding to the pulsating phase. For this aim we used the nearby C[ii]{} doublet at $\lambda \lambda$ 6578, 6582 [Å]{}, both belonging to $\beta$ Cep A. For the sake of clarity, we show in Fig. \[example\] an example of our procedure. The profiles of $\beta$ Cep Aa, recovered as described before and converted in the heliocentric velocity scale, are showed in Fig. \[profiles1\] and in Fig. \[profiles2\]. Inspection of those two figures revealed the strong variability of $\beta$ Cep Aa, even though no transition between shell into normal B spectrum has been observed. It always shows a double-peak symmetrical profile, with V/R $\approx$ 1. Spectra are ordered by Julian Date. ![Variability of some quantities characteristics of the H$\alpha$ emission profile plotted in function of the Heliocentric Julian Day. From bottom to top panel: equivalent width expressed in [Å]{}, V/R, peaks separation $\Delta$V and radial velocities of the central depression, all expressed in km s$^{-1}$. Meaning of the symbols is: [*squares*]{} - data from Ritter Observatory, [*triangles*]{} - data from INT, [*asterisks*]{} - data from Elodie and [*circles*]{} - data from Catania Astrophysical Observatory. In the EW panel, the H$\alpha$ emission increase going up; the dotted line in the V/R panel represents the V=R symmetry. The top axis is labeled in fraction of years.[]{data-label="variab"}](variab.ps){width="13cm" height="17cm"} H$\alpha$ Variability and disk properties ========================================= In this section we analyzed the behavior with time of some line width parameters of the H$\alpha$ emission profile, i.e.: equivalent width (EW), the full width at half-maximum (FWHM), the peaks separation ($\Delta V$), the radial velocity of the central depression (RV$_{\rm CD}$) and the ratio of blue to red peak intensities (V/R) computed after subtraction of the continuum level[^3]. In order to better measure positions and flux levels of blue and red peaks, we fitted each profile with a function defined as the sum of the two gaussians centered respectively on the blue and the red peak. A similar fit has been considered to compute the radial velocity of the central depression. All this data with their relative errors have been reported in Tab. \[journal\]. In Fig. \[variab\] we plotted all these quantities versus time. As a general consideration, we observed two different regimes of variations: a smoothed one between Julian Date 2449250 and 2453000 and a more rapid variation starting from JD$>$2453000. ![Correlation between V/R, expressed in logarithm, and radial velocity of the central depression. []{data-label="vr_rv"}](vr_vs_rv.ps){width="9cm"} In the bottom panel of Fig. \[variab\] we reported the equivalent width of H$\alpha$ versus time. By definition, it is straightforward that lines with emission profiles will have negative EW. In the plot we inverted the directions of y axis in such a way to have ascending curve in correspondence of greater emission. EW decreases from September 1993 to May 1999, then a new strong emission phenomenon occurs in July 2002 that remains constant till August 2003. Again EW decreases, it reaches a minimum during September 2005 and it increases again until November 2006, date of last spectrum of our series. As concerns the FWHM we found its value almost constant in time, about 400 $\pm$ 30 km s$^{-1}$, at least at the resolving power of our spectra. The peaks separation reveled an increase from September 1993 to May 1999, then it decreased reaching a minimum in June 2005. After that date, the separation begins to increase again (third panel from bottom of Fig. \[variab\]). The fourth panel shows the behavior of the radial velocity of the central depression. It shows a progressive red shift till May 1999. Then, we observed a blue shift till June 2005. After this date velocities present much more scatter, probably due to a worse quality of data. In general, from the variations of the radial velocity, one could draw some useful indication about the state of the disk: expanding, contracting or stationary. This is true provided that the radial velocity of the underlying star is well known. Unfortunately, since this is not possible with our data, we could not say anything about the equilibrium state of the envelope. A similar trend has been observed for the V/R ratio, but with an exceptional increase (V/R $>$ 1.5) for three julian dates: 51144.261 (Elodie), 51155.338 (INT) and 51325.682 (INT). Cyclic variability of V/R has already been observed in a number of Be stars [@hanu95] and this evidence has been theoretically explained in terms of one-armed density waves precessing around the central star, the so-called Okazaki’s model [@okazaki91]. When the dense part is on the approaching side of the disk, the V peak is higher than R. On the contrary, when dense part is in the receding part of the disk, R peak will be enhanced. It is well known from literature (see for example @hubert87), that in addition to the V/R variation Be stars are always accompained by profile shifts: blueward when red component is stronger (V/R$<$1), redward when blue component is stronger (V/R$>$1). This feature is again explained in term of the one-armed density model as proposed by @okazaki96. In Fig. \[vr\_rv\] we show the correlation between RV and V/R. It is clear that when V/R$>$1 the profile is redshifted (RV$>$0) and, on the contrary, the profile is blueshifted (RV$<$0) when V/R$<$1. For all these quantities, we attempted to search for periodic variability using the [*Phase Dispersion Method (PDM)*]{} [@stellin78] as coded in the NOAO/IRAF package, but no succesfully results have been achieved. @dachs86 and @hanu86, studying the kinematics of the disks surrounding Be stars, found some useful correlations between line width parameters (in particular, FWHM and peak separation) and stellar projected rotational velocity, $v \sin i$. Later, the existence of those correlations have been confirmed by @hanu88. Using those correlations, we attempted to derive the $v \sin i$ of the underlying star. In particular, @hanu88 gave equations that linked each other FWHM, peaks separation and $v \sin i$ for objects whit EW of the H$\alpha$ emitting disk region less than 15 [Å]{}. From our measurements, we derived a velocity equal to $\approx$230 km s$^{-1}$. The rotation law in Be star circumstellar disk is usually written as $v(r_d)~\propto~r_d^{-j}$, where the index [*j*]{}=0.5 if for keplerian rotation and [*j*]{}=1 for conservation of angular momentum. According to former author, the observed normalized peaks separation is a function of the outer radius of the disk, expressed in term of stellar radius, via the formula: $$\frac{\Delta V}{2\,v \sin i}\,=\,r_d^{-j} \label{vel}$$ In their paper, @hanu88 concluded that for their sample of Be stars the value of [*j*]{} is closer to 1 than to 0.5. On the contrary, @hummel00 concluded that the formation of H$\alpha$ shell profiles in Be stars generally requires that [*j*]{}$\approx$0.5, i.e. the keplerian rotation is a valid approximation. As for our object, according to the Okazaki’s model, the correlation between RV and V/R showed in Fig. \[vr\_rv\] is a strong argument for the validity of keplerian rotation. Thus, using Eq. \[vel\], we estimated the radius of the outer disk surrounding $\beta$ Cep Aa to be 3.38 R$_*$ ([*j*]{}=0.5). Conclusion ========== In this paper we analyzed the last 13 years of the H$\alpha$ emission history of $\beta$ Cep Aa. Spectroscopic data, for a total of 72 spectra, have been obtained both from [*INAF - Osservatorio Astrofisico di Catania*]{} and from archives of other three observatories as described in Sec. \[observ\]. We used H$\beta$ and H$\gamma$ profiles to derive effective temperature and surface gravity of the primary. By spectral synthesis we obtained: T$_{\rm eff}$=24000$\pm$250 K and $\log g$=3.91$\pm$0.10. Regarding the secondary, we conclude that $\beta$ Cep Aa is a Be star which shows a high degree of variability in H$\alpha$ emission intensity as well as in the other line width parameters here considered. From 1993 to 1999 we observed a decrease in the emission and a subsequent increase till July 2002, that remains constant till August 2003. After this long cycle extended for 10 years, we again observed a decline toward a minimum and a raise toward a new phase of strong emission, in a shorter period of about 3 years. As concerns, V/R ratio, peaks separation and radial velocity of central depression, we also observed a long 10 years cycle out-of-phase with respect the EW and a more rapid variability occured during the last 3 years. As for the FWHM, we did not observe any appreciable change, at least at the resolving power of our data. Making use of some useful literature correlations between FWHM, peaks separation and stellar projected rotational velocities, we estimated for the first time the $v \sin i$ of $\beta$ Cep Aa equal to $\approx$230 km s$^{-1}$ and the outer radius of the surrounding disk, $r_d$=3.38 R$_*$. Acknowledgments {#acknowledgments .unnumbered} =============== This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. The author wish to thanks the anonymous referee for his/her suggestions useful to improve the scientific impact of the manuscript. A warm thanks to Anna for her contribution in improving the english form of my original manuscript. [99]{} Aerts C., Mathias P., Gillet D., Waelkens C. 1994, A&A, 286, 109 Balega I. I., Balega Y. Y., Hofmann K.-H., et al., 2002, A&A, 385, 87 Baranne A., Queloz D., Mayor M., et al., 1996, A&AS, 119, 373 Catanzaro G., Leone F., Busá I., Romano P., 2008, NewA, 13, 113 Cunha K., Lambert D. L., 1994, ApJ, 466, 170 Dachs J., Hanuschik R., Kaiser D., Rohe D., 1986, A&A, 159, 276 Dufton P. L., Brown P. J. F., Fitzsimmons A., Lennon D. J., 1990, A&A, 232, 431 Gezari D. Y., Labeyrie A., Stachnik,R. V. 1972, ApJ, 173, L1 Gies D. R., Lambert D. L., 1992, ApJ, 387, 673 Hanuschik R. W., Hummel W., Dietle O., Sutorius E., A&A, 300, 163 Hanuschik R. W., Kozok J. R., Kaiser D., 1988, A&A, 189, 147 Hanuschik R. W., 1986, A&A, 166, 185 Hadrava P., Harmanec P. 1996, A&A, 315, L401 Hartkopf W., Mason B., Wycoff G., McAlister H. 2001, Fourth Catalog of Interferometric Measurements of Binary Stars Heynderickx D., Waelkens C., Smeyers P., 1994, A&AS, 105, 447 Hummel W., Vrancken M., 2000, in: The Be Phenomenon in Early-Type Stars, IAU Coll. 175, ASP Conf. Series, Vol. 214, M. A. Smith, H. F. Henrichs, J. Fabregat, eds. Hubert A. M., Floquet M., Chambon M. T., 1987, A&A, 186, 213 Kaper L., Mathias P., 1995, in: Astrophysical Applications of Stellar Pulsations, ASP Conference Series, Vol. 83, 295 Karpov B. G., 1933, Lick Observatory Bull., 16, No. 457, 167 Kurucz R.L., 1993, A new opacity-sampling model atmosphere program for arbitrary abundances. In: Peculiar versus normal phenomena in A-type and related stars, IAU Colloquium 138, M.M. Dworetsky, F. Castelli, R. Faraggiana (eds.), A.S.P Conferences Series Vol. 44, p.87 Kurucz R.L., Avrett E.H., 1981, SAO Special Rep., 391 Morel T., Butler K., Aerts C., Neiner C., Briquet M. 2006, A&A, 457, 651 Niecmzura E., Daaszy’nska-Daszkiewwicz J., 2005, A&A, 433, 659 Okazaki A. T., 1996, PASJ, 48, 305 Okazaki A. T., 1991, PASJ, 43, 750 Pan’ko E. A., Tarasov E. A., 1997, Astronomy Letters, 23, 545 Pigulsky A., Boratyn D. A. 1992, A&A, 253, 178 Porter J. M., Rivinius T., 1993, PASP, 115, 1153 Sbordone L., Bonifacio P., Castelli F., Kurucz R., 2004, Mem. S.A.It. Suppl. Vol. 5, 93 Schnerr R. S., Henrichs H. F., Oudmaijer R. D., Telting J. H. 2006, A&A, 459, L21 Slettebak A., ApJS, 50, 55 Stellingwerf R. F., 1978, ApJ, 224, 953 Telting J. H., Aerts C., Mathias P. 1997, A&A, 322, 493 Wilson R., Seddon H., 1956, The Observatory, 76, 145 Journal of observations ======================= \[lastpage\] [^1]: All the Kurucz codes (ATLAS9 and SYNTHE) have been used in the Linux version implemented by @sbord04 [^2]: This conversion has been calculated considering normal helium abundance. [^3]: Since our spectra have been normalized to the unity, we computed this ratio as (V-1)/(R-1).

ArXiv

--- abstract: 'We propose a realization of the Haldane model using active mechanical metamaterials. The Haldane model employs complex valued bonds to break time reversal symmetry in a two-band quantum system. An analogous two-band mechanical system, i.e. a bipartite lattice with one degree of freedom per lattice site implies complex valued directional couplings between the sites, which is inconsistent with Newtonian dynamics. Here we subject a conventional mass-spring lattice to external forces that are processed through a closed loop electronic controller, which is pre-programmed to generate the desired couplings in real-time. The resulting metamaterial acts as an active medium for acoustic waves, and supports strictly unidirectional mechanical wave propagation along its boundaries. Our feedback-based approach opens a new direction of autonomous programmed metamaterials that can create arbitrary lattice parameters on the same platform, including non-local and nonlinear couplings, time dependent potentials and more.' author: - Lea Sirota - Roni Ilan - Yair Shokef - Yoav Lahini bibliography: - 'HaldaneModelbib.bib' title: 'Feedback-based Mechanical Haldane Model' --- The discovery of topologically protected wave phenomena in quantum physics [@thouless1982quantized; @haldane1988model; @kane2005quantum; @bernevig2006quantum] with their exceptional immunity to back-scattering has recently inspired the search for analogies in classical systems, substituting the electronic band structure with acoustic dispersion relations. As systems governed by Newtonian dynamics do not naturally exhibit topological phenomena by classical waves (e.g. acoustic or elastic waves), artificially designing a material in order to shape its dispersion relation is required. This became available with the notion of engineered materials, or metamaterials [@liu2008three; @seo2012acoustic; @dubois2017observation; @zhu2017double]. Indeed, numerous realizations of metamaterials supporting topologically protected wave propagation have been reported over the last decade in photonics [@rechtsman2013photonic; @chen2014experimental; @mousavi2015topologically], acoustics [@khanikaev2015topologically; @he2016acoustic; @zhang2017topological; @yves2017topological], and mechanics [@susstrunk2015observation; @pal2017edge; @chaunsali2018subwavelength; @wang2015topological; @nash2015topological]. The particular quantum topological effect that is being considered for the classical analogy plays a very important role. The Quantum Spin Hall Effect [@kane2005quantum; @bernevig2006quantum] or the Quantum Valley Hall Effect [@pan2014valley], which are based on breaking spatial symmetry in a lattice and support bi-directional edge waves are the natural path to classical analogies, as they can be implemented with purely passive mechanical components. Indeed, the vast majority of reports implement these effects e.g. by designing the spacing of steel bars [@zhang2017topological] and bottle-like Helmholtz resonators [@yves2017topological] in an acoustic waveguide, the spacing of resonators on a plate [@chaunsali2018subwavelength], the spring constants in a mass-spring lattice [@zhou2018quantum], a pendulum array with sophisticated couplings [@susstrunk2015observation], to name a few. On the other hand, implementing analogy of the Quantum Hall Effect, i.e. a Chern insulator, which requires breaking time reversal symmetry (TRS), is considerably more involved, as passive design becomes insufficient, and consequently there have been only few associated reports; focusing on mechanical and acoustic systems, a Chern insulator was emulated in [@wang2015topological; @nash2015topological] by attaching gyroscopes to a mass-spring lattice, whereas the same was achieved in [@khanikaev2015topologically] with circulating fluids. These methods of TRS breaking have one principle in common: they require *in-plane* degrees of freedom (DOFs) in two-dimensional systems, i.e. actual physical in-plane rotation of masses or fluids, which results in a four-band dispersion diagram. However, it would be advantageous to break TRS via out-of-plane DOFs alone, i.e. one DOF per mass, due to several reasons: first, it will enable reproduction of quantum effects that are associated with two-band systems, such as the Haldane model [@haldane1988model]. Second, it appears more feasible to realize such systems experimentally, as out-of-plane DOFs imply a scalar field, such as acoustic pressure or transverse plate deflection. However, when only out-of-plane DOFs are allowed, breaking TRS in classical systems requires lattice couplings that are inconsistent with classical physics laws, including complex-valued directional, or non-reciprocal couplings, as discussed below. In this Letter we present a method of breaking TRS in mechanical systems using only *out-of-plane* DOFs, with two DOFs per unit cell in a discrete lattice. We employ a feedback-based design, which includes application of external control forces to the masses in the out-of-plane direction, and operating the forces in real time according to a target feedback control scheme. Employing active control in the design of metamaterials has recently attracted considerable interest [@chen2017hybrid; @popa2015active; @cheer2017feedforward; @zangeneh2019active; @sirota2019tunable; @sirota2020modeling]. In particular, the effect of this design in emulating quantum topological phenomena is being explored [@darabi2019experimental; @rivet2018constant; @hofmann2019reciprocal; @lee2019anatomy; @scheibner2020non; @rosa2020dynamics; @brandenbourger2019non]. In our system, an autonomous pre-programmed controller in each unit cell receives measurements of displacements and velocities of masses in neighboring lattice sites, processes them and feeds back to the control forces. The control operation therefore determines, in real-time, the dynamic response of the masses. Since the particular couplings are solely defined by the algorithm that we program into the controller, the feedback-based metamaterial is able to sustain any couplings (within hardware limitations), including those that are otherwise physically not achievable, such as directional or non-reciprocal couplings. Furthermore, a single system is not limited to emulate a particular quantum effect, but can be programmed to any other functionality. We analytically and numerically demonstrate our feedback-based design method by realizing an exact mechanical implementation of the quantum Haldane Model [@haldane1988model]. This revolutionary model proved that the Quantum Hall Effect can be obtained without an external magnetic field, but rather by breaking TRS. The Haldane model comprises two interlacing triangular sub-lattices with principal lattice vectors $\{\textbf{a}_1,\textbf{a}_2\}$, exhibiting two sites per unit cell, $A$ and $B$, as illustrated in Fig. \[Scheme\](a). In our classical mechanical implementation the circles are identical masses $m_0$ that can vibrate only along the vertical axis $\textbf{a}_3$. We denote the lattice constant by $a$. The grey bars indicate nearest neighbor couplings, which are equivalent to Hookean springs of constant $t_1>0$ connecting the masses. When only the $t_1$ springs exist, the lattice is analogous to Graphene. The quantum Haldane Model assumes additional next nearest neighbor bonds of a complex strength $t_2e^{\pm i\phi}$ in the directions $\textbf{v}_1,\textbf{v}_2,\textbf{v}_3$, indicated by the dashed black lines in Fig. \[Scheme\](a). In a mechanical context, such a bond represent a *non-reciprocal* coupling, $t_2e^{+i\phi}$ (red arrow) towards one mass and $t_2e^{-i\phi}$ (blue arrow) towards the other connected mass, which violates Newton’s third law and is therefore non-physical. Such couplings cannot be implemented neither with passive devices such as springs, lever arms etc., nor with active devices like gyroscopes that rely on in-plane DOFs. We realize these couplings using active closed loop control. Unlike what might seem, the complex value is not a non-physical property, though, as in the time harmonic regime it relates to a velocity rather than to displacement. The full form of the Haldane Model is captured by the Bloch Hamiltonian $H(\textbf{k})=\sum_{l=0}^3\mathcal{H}_l(\textbf{k})\sigma_l$, where $\textbf{k}$ is the wave vector, $\sigma_l$ are the Pauli matrices, and $$\label{eq:H} \begin{split} &\mathcal{H}_0=\beta+2t_2\cos\phi\textstyle{\sum_{m=1}^3}\cos(\textbf{k}\cdot\textbf{v}_m),\\ &\mathcal{H}_1=-t_1\left(1+\cos(\textbf{k}\cdot\textbf{a}_1)+\cos(\textbf{k}\cdot\textbf{a}_2)\right),\\ &\mathcal{H}_2=t_1\left(\sin(\textbf{k}\cdot\textbf{a}_1)+\sin(\textbf{k}\cdot\textbf{a}_2)\right),\\ &\mathcal{H}_3=M-2t_2\sin\phi\textstyle{\sum_{m=1}^3}\sin(\textbf{k}\cdot\textbf{v}_m). \end{split}$$ In a quantum system, $\beta=0$. In the classical-mechanical analogy of Graphene $\beta=3t_1$, indicating the restoring $t_1$ force from the three nearest neighbor springs, and is not related to Haldane’s next nearest neighbors bonds. The constant $M$ accounts for a possible spatial inversion symmetry breaking in addition to the TRS breaking provided by the $t_2$ bonds. --------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- **(a)** ![Feedback control scheme for generating the exact Haldane Model in a mechanical metamaterial. (a) A honeycomb lattice in the $\{\textbf{a}_1,\textbf{a}_2\}$ space comprising identical masses (black and white circles) connected to nearest neighbors by Hookean springs (grey bars). The masses can move in the $\textbf{a}_3$ (out-of-plane) direction only, implying two DOFs per unit cell of $A$ and $B$ sites. The red and blue arrows indicate out-of-plane displacement and velocity measurements of the next nearest neighbors (in $\{\textbf{v}_1,\textbf{v}_2,\textbf{v}_3\}$ directions). Dashed lines indicate the Haldane model bonds that are created in real-time when control is turned on. (b) The measurement scheme detailed for the $A$ site in the $\{i,j\}$ unit cell. The black arrow indicates the total control force $f^A=f^++f^-+f^0$. Red and blue arrows distinguish between measurements that are fed to controller gains $t_2e^{+i\phi}$ and $t_2e^{-i\phi}$ (red and blue cubes in the figure), applied through the $f^+$ and $f^-$ components, respectively. (c) Feedback control scheme of the entire $\{i,j\}$ unit cell, including all the measured signals. []{data-label="Scheme"}](Scheme_3 "fig:"){width="7.0"} **(b)** ![Feedback control scheme for generating the exact Haldane Model in a mechanical metamaterial. (a) A honeycomb lattice in the $\{\textbf{a}_1,\textbf{a}_2\}$ space comprising identical masses (black and white circles) connected to nearest neighbors by Hookean springs (grey bars). The masses can move in the $\textbf{a}_3$ (out-of-plane) direction only, implying two DOFs per unit cell of $A$ and $B$ sites. The red and blue arrows indicate out-of-plane displacement and velocity measurements of the next nearest neighbors (in $\{\textbf{v}_1,\textbf{v}_2,\textbf{v}_3\}$ directions). Dashed lines indicate the Haldane model bonds that are created in real-time when control is turned on. (b) The measurement scheme detailed for the $A$ site in the $\{i,j\}$ unit cell. The black arrow indicates the total control force $f^A=f^++f^-+f^0$. Red and blue arrows distinguish between measurements that are fed to controller gains $t_2e^{+i\phi}$ and $t_2e^{-i\phi}$ (red and blue cubes in the figure), applied through the $f^+$ and $f^-$ components, respectively. (c) Feedback control scheme of the entire $\{i,j\}$ unit cell, including all the measured signals. []{data-label="Scheme"}](Scheme_3D "fig:"){width="6.5"} **(c)** ![Feedback control scheme for generating the exact Haldane Model in a mechanical metamaterial. (a) A honeycomb lattice in the $\{\textbf{a}_1,\textbf{a}_2\}$ space comprising identical masses (black and white circles) connected to nearest neighbors by Hookean springs (grey bars). The masses can move in the $\textbf{a}_3$ (out-of-plane) direction only, implying two DOFs per unit cell of $A$ and $B$ sites. The red and blue arrows indicate out-of-plane displacement and velocity measurements of the next nearest neighbors (in $\{\textbf{v}_1,\textbf{v}_2,\textbf{v}_3\}$ directions). Dashed lines indicate the Haldane model bonds that are created in real-time when control is turned on. (b) The measurement scheme detailed for the $A$ site in the $\{i,j\}$ unit cell. The black arrow indicates the total control force $f^A=f^++f^-+f^0$. Red and blue arrows distinguish between measurements that are fed to controller gains $t_2e^{+i\phi}$ and $t_2e^{-i\phi}$ (red and blue cubes in the figure), applied through the $f^+$ and $f^-$ components, respectively. (c) Feedback control scheme of the entire $\{i,j\}$ unit cell, including all the measured signals. []{data-label="Scheme"}](Control_1 "fig:"){width="6.2"} --------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- The goal of our embedded control system is to create a classical-mechanical metamaterial, whose Newtonian dynamics in lattice momentum space $\textbf{k}$ is given by $$\label{eq:EQ} \omega^2\textbf{p}(\textbf{k})=H(\textbf{k})\textbf{p}(\textbf{k}),$$ where $H(\textbf{k})$ is the Bloch Hamiltonian of the Haldane model, and $\textbf{p}(\textbf{k})$ is the complex amplitude vector of the $A$ and $B$ sites in momentum space. Our starting point is the Graphene-like lattice ($t_1$ springs only), in which we denote the DOFs of each $\{i,j\}$ unit cell by $\textbf{u}_{i,j}(t)=(\begin{array}{c c}u^A_{i,j}(t), & u^B_{i,j}(t)\end{array})^T$. The time domain unit cell dynamics, including external mechanical control forces $f^A_{i,j},f^B_{i,j}$ that are applied to the masses in the $a_3$ direction, reads $$\label{eq:Motion_OL} \begin{split} &\ddot{u}^A_{i,j}=3t_1u^A_{i,j}-t_1\left(u^B_{i,j}+u^B_{i+1,j}+u^B_{i,j+1}\right)+f^A_{i,j}, \\ &\ddot{u}^B_{i,j}=3t_1u^B_{i,j}-t_1\left(u^A_{i,j}+u^A_{i-1,j}+u^A_{i,j-1}\right)+f^B_{i,j}. \end{split}$$ The control forces are decomposed into $f^A_{i,j}=f^{A+}_{i,j}+f^{A-}_{i,j}+f^{A0}_{i,j}$ and $f^B_{i,j}=f^{B+}_{i,j}+f^{B-}_{i,j}+f^{B0}_{i,j}$. The $f^+$ and $f^-$ components are responsible for generating the $t_2e^{+i\phi}$ and the $t_2e^{-i\phi}$ couplings, respectively. $f^0$ is responsible for generating $M$. As depicted in Fig. \[Scheme\](b), e.g. for the $A$ site, the $f^{A+}$ and $f^{A-}$ components receive measured signals of displacements and velocities of the $u^A_{i+1,j},u^A_{i-1,j+1},u^A_{i,j-1}$ and $u^A_{i-1,j},u^A_{i+1,j-1},u^A_{i,j+1}$ DOFs, as indicated by the red and blue arrows, respectively. These arrows are also shown on the multi-cell lattice segment in Fig. \[Scheme\](a). The $f^{A0}$ component is not depicted in the figure. The measurements are processed in real-time by corresponding controllers, indicated by red and blue cubes. The control action is illustrated in Fig. \[Scheme\](c) for the $\{i,j\}$ unit cell. For each site $A$ and $B$ (the superscripts are omitted in the following), the control forces are related to the measured signals as $$\label{eq:C_law} \left( \begin{array}{c c c} f^+_{i,j} & f^-_{i,j} & f^0_{i,j} \end{array} \right)^T=C \left( \begin{array}{c c c c c} y^+_{i,j} & v^+_{i,j} & y^-_{i,j} & v^-_{i,j} & u_{i,j}\end{array} \right)^T,$$ where $$\label{eq:Measurements} \begin{split} &y^+_{i,j}=u_{i+1,j}+u_{i-1,j+1}+u_{i,j-1}, \\ &v^+_{i,j}=\dot{u}_{i+1,j}+\dot{u}_{i-1,j+1}+\dot{u}_{i,j-1}, \\ &y^-_{i,j}=u_{i-1,j}+u_{i+1,j-1}+u_{i,j+1}, \\ &v^-_{i,j}=\dot{u}_{i-1,j}+\dot{u}_{i+1,j-1}+\dot{u}_{i,j+1}. \end{split}$$ The control matrix $C$ at each $\{i,j\}$ unit cell is given by $$\label{eq:C} C=\left( \begin{array}{ccccc} t_2\cos \phi & \frac{t_2}{\omega}\sin \phi & 0 & 0 & 0 \\ 0 & 0 & t_2\cos \phi & -\frac{t_2}{\omega}\sin \phi & 0 \\ 0 & 0 & 0 & 0 & \pm M \end{array} \right),$$ where the sign of $M$ in is positive (negative) for the $A$ ($B$) sites. Since in frequency domain, the velocity $v$ is related to the displacement $u$ as $v=i\omega u$, the velocity coupling controllers are compensated by the actuation signal frequency $\omega$. The resulting time domain evolution of the control forces generates the next nearest neighbors complex non-reciprocal bonds that are required in the Haldane Model Hamiltonian . This closed loop system is internally stable and the control signals amplitudes do not exceed the input signal amplitude. Since a two-band system implies a scalar dynamical field, it significantly reduces the complexity of experimental realization compared to higher-band systems. For example, the out-of-plane displacement may be realized by an acoustic pressure field, created by a two-dimensional array of small loudspeakers. The feedback control system in - can be then realized by equipping each speaker with an autonomous micro-controller (which is essentially a small electronic chip [@analog]) that processes the pressure measured by microphones at corresponding next nearest neighbor locations. Next we demonstrate that the mechanical system governed by the classical analogy of the Haldane Model Hamiltonian , which we created with the control system -, reproduces all the known dynamic properties of the quantum Haldane Model. This is not obvious a-priori, since differs from the quantum Hamiltonian in its $\sigma_0$ term, and the complex valued couplings are reached only with the convergence of the control loops. ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- **(a)** **(b)** ![Infinite lattice dispersion relations of the Haldane model mechanical analogy. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Dispersion diagram is plotted over the entire Brillouin zone (a),(b) and through the high symmetry points (c),(d). (a),(b) Case $1$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. (c),(d) Case $2$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=3\sqrt{3}t_2\sin\phi\}$.[]{data-label="BulkDisp"}](3D_Disp_M0 "fig:"){width="4.3"} ![Infinite lattice dispersion relations of the Haldane model mechanical analogy. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Dispersion diagram is plotted over the entire Brillouin zone (a),(b) and through the high symmetry points (c),(d). (a),(b) Case $1$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. (c),(d) Case $2$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=3\sqrt{3}t_2\sin\phi\}$.[]{data-label="BulkDisp"}](1D_Disp_M0 "fig:"){width="4.3"} **(c)** **(d)** ![Infinite lattice dispersion relations of the Haldane model mechanical analogy. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Dispersion diagram is plotted over the entire Brillouin zone (a),(b) and through the high symmetry points (c),(d). (a),(b) Case $1$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. (c),(d) Case $2$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=3\sqrt{3}t_2\sin\phi\}$.[]{data-label="BulkDisp"}](3D_Disp_M3 "fig:"){width="4.3"} ![Infinite lattice dispersion relations of the Haldane model mechanical analogy. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Dispersion diagram is plotted over the entire Brillouin zone (a),(b) and through the high symmetry points (c),(d). (a),(b) Case $1$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. (c),(d) Case $2$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=3\sqrt{3}t_2\sin\phi\}$.[]{data-label="BulkDisp"}](1D_Disp_M3 "fig:"){width="4.3"} ![Infinite lattice dispersion relations of the Haldane model mechanical analogy. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Dispersion diagram is plotted over the entire Brillouin zone (a),(b) and through the high symmetry points (c),(d). (a),(b) Case $1$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. (c),(d) Case $2$ with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=3\sqrt{3}t_2\sin\phi\}$.[]{data-label="BulkDisp"}](colorbar_jet "fig:"){width="3.2"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- The first property that we analyze is the band structure, here acoustic dispersion, of an infinite-periodic lattice. Since represents a classical-mechanical system, whose dynamics is described by Newtonian equations of motion, the eigenvalues are squared frequencies. The frequencies are kept real and positive due to the constant shift of the dispersion curves by the addition of $\beta=3t_1$ to the $\sigma_0$ term in . Since this addition does not change the eigenvectors, the topological properties of the original quantum Haldane Model are preserved. We consider two combinations of $\phi$ and $M$: $\{\phi=\pi/3,M=0\}$ in case $1$, and $\{\phi=\pi/3,M=3\sqrt{3}t_2\sin\phi\}$ in case $2$. In both cases $t_2=0.2t_1$. The corresponding band structures of these two cases are depicted in Fig. \[BulkDisp\]. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Case $1$ falls within the non-trivial topological regime of Chern number $n=+1$ [@haldane1988model]. Similarly to the quantum system, the figure shows a gap at $K_T$ but not at $K$. Due to TRS breaking, the band structure is not symmetric between the $\Gamma-M-K-\Gamma$ and the $\Gamma-M_T-K_T-\Gamma$ trajectories. Case $2$ falls exactly on the border between the topologically trivial ($n=0$) and nontrivial ($n=+1$) regimes. Due to TRS breaking, a gap is opened at the $K$ point but not at the $K_T$ point (for $M=-3\sqrt{3}t_2\sin\phi$ it is the other way around with $n=-1$). ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- **(a)** **(b)** **(c)** ![Edge modes of the mechanical Haldane model. (a) Dispersion diagram of a lattice infinite in the $\textbf{x}$ direction and of eight honeycomb cells in the $\textbf{y}$ direction, with parameters of Case $1$, $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Black lines indicate bulk states. The red (blue) line indicates top (bottom) edge state with negative (positive) group velocity. (b),(c) Corresponding eigenmodes of the top and bottom edge states.[]{data-label="EdgeModeDisp"}](Edge_disp_new "fig:"){height="4.8"} ![Edge modes of the mechanical Haldane model. (a) Dispersion diagram of a lattice infinite in the $\textbf{x}$ direction and of eight honeycomb cells in the $\textbf{y}$ direction, with parameters of Case $1$, $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Black lines indicate bulk states. The red (blue) line indicates top (bottom) edge state with negative (positive) group velocity. (b),(c) Corresponding eigenmodes of the top and bottom edge states.[]{data-label="EdgeModeDisp"}](mode1 "fig:"){height="4.6"} ![Edge modes of the mechanical Haldane model. (a) Dispersion diagram of a lattice infinite in the $\textbf{x}$ direction and of eight honeycomb cells in the $\textbf{y}$ direction, with parameters of Case $1$, $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Black lines indicate bulk states. The red (blue) line indicates top (bottom) edge state with negative (positive) group velocity. (b),(c) Corresponding eigenmodes of the top and bottom edge states.[]{data-label="EdgeModeDisp"}](colorbar_hot_leg "fig:"){height="2.6"} ![Edge modes of the mechanical Haldane model. (a) Dispersion diagram of a lattice infinite in the $\textbf{x}$ direction and of eight honeycomb cells in the $\textbf{y}$ direction, with parameters of Case $1$, $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. The frequency scale is normalized by $\omega_0=\sqrt{t_1/m_0}$. Black lines indicate bulk states. The red (blue) line indicates top (bottom) edge state with negative (positive) group velocity. (b),(c) Corresponding eigenmodes of the top and bottom edge states.[]{data-label="EdgeModeDisp"}](mode2 "fig:"){height="4.6"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Next we verify that our feedback-based topological metamaterial reproduces the edge mode dispersion of the quantum Haldane model. We consider a lattice that is infinite in the $\textbf{x}$ direction and finite in the $\textbf{y}$ direction, with parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. The dispersion diagram of an eight honeycomb cells strip is plotted in Fig. \[EdgeModeDisp\](a). As expected for a Chern insulator, an edge mode emerges inside the bulk bandgap, corresponding to top edge propagation with negative group velocity ($S_n$ point) and to bottom edge propagation with positive group velocity ($S_p$ point). The associated eigenmodes are depicted in Fig. \[EdgeModeDisp\](b),(c). The wave localization on the lattice edge is guaranteed by the topological property of the band-structure. -- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Time domain simulation of the mechanical Haldane Chern insulator defined by - with Case $1$ parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. A time harmonic force excitation $F(t)=F_0e^{i\omega t}$, indicated by a blue arrow, is applied to a $8\times 30$ honeycomb lattice in the $\textbf{a}_3$ direction. The displacement responses of the masses, in $\textbf{a}_3$ as well, are depicted at the normalized time instances $\widehat{T}_1=50$, $\widehat{T}_2=100$, $\widehat{T}_3=150$ and $\widehat{T}_4=280$.[]{data-label="Sim"}](T1 "fig:"){width="7.0"} ![Time domain simulation of the mechanical Haldane Chern insulator defined by - with Case $1$ parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. A time harmonic force excitation $F(t)=F_0e^{i\omega t}$, indicated by a blue arrow, is applied to a $8\times 30$ honeycomb lattice in the $\textbf{a}_3$ direction. The displacement responses of the masses, in $\textbf{a}_3$ as well, are depicted at the normalized time instances $\widehat{T}_1=50$, $\widehat{T}_2=100$, $\widehat{T}_3=150$ and $\widehat{T}_4=280$.[]{data-label="Sim"}](T2 "fig:"){width="7.0"} ![Time domain simulation of the mechanical Haldane Chern insulator defined by - with Case $1$ parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. A time harmonic force excitation $F(t)=F_0e^{i\omega t}$, indicated by a blue arrow, is applied to a $8\times 30$ honeycomb lattice in the $\textbf{a}_3$ direction. The displacement responses of the masses, in $\textbf{a}_3$ as well, are depicted at the normalized time instances $\widehat{T}_1=50$, $\widehat{T}_2=100$, $\widehat{T}_3=150$ and $\widehat{T}_4=280$.[]{data-label="Sim"}](T3 "fig:"){width="7.0"} ![Time domain simulation of the mechanical Haldane Chern insulator defined by - with Case $1$ parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. A time harmonic force excitation $F(t)=F_0e^{i\omega t}$, indicated by a blue arrow, is applied to a $8\times 30$ honeycomb lattice in the $\textbf{a}_3$ direction. The displacement responses of the masses, in $\textbf{a}_3$ as well, are depicted at the normalized time instances $\widehat{T}_1=50$, $\widehat{T}_2=100$, $\widehat{T}_3=150$ and $\widehat{T}_4=280$.[]{data-label="Sim"}](T4 "fig:"){width="7.0"} ![Time domain simulation of the mechanical Haldane Chern insulator defined by - with Case $1$ parameters $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. A time harmonic force excitation $F(t)=F_0e^{i\omega t}$, indicated by a blue arrow, is applied to a $8\times 30$ honeycomb lattice in the $\textbf{a}_3$ direction. The displacement responses of the masses, in $\textbf{a}_3$ as well, are depicted at the normalized time instances $\widehat{T}_1=50$, $\widehat{T}_2=100$, $\widehat{T}_3=150$ and $\widehat{T}_4=280$.[]{data-label="Sim"}](colorbar_hot "fig:"){width="5.0"} -- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- We now demonstrate that the feedback-based mechanical Haldane Chern insulator indeed supports uni-directional edge wave propagation. We perform time domain simulations of a finite size metamaterial ($8\times 30$ honeycomb net), which is operated in a real-time feedback loop according to -. The simulation results are given in Fig. \[Sim\]. Open boundary conditions are assumed. The system is excited by a time-harmonic force $F(t)=F_0e^{i\omega t}$ in the $\textbf{a}_3$ direction at the middle of the top lattice edge, indicated by a blue arrow in the figure. The actuation frequency $\omega=1.55\omega_0$ lies inside the bulk bandgap in Fig. \[EdgeModeDisp\](a) and corresponds to the point $S_n$ there. The parameters are set to $\{t_2=0.2t_1,\phi=\pi/3,M=0\}$. Closed loop time responses of the masses out-of-plane displacements $u_{i,j}(t)$ (normalized by $F_0$) are shown at different normalized time instances $\widehat{T}=T/T_0$ with $T_0=\sqrt{m_0/t_1}$. One clearly sees that a wave propagates along the lattice edges strictly in the counterclockwise direction, i.e. with a negative group velocity, as expected for the top edge mode. Due to topological protection, the wave circumvents the sharp lattice corners without any back-scattering. To summarize, we demonstrated an analogy of a two-band Chern insulator, specifically the quantum Haldane model, in a classical mechanical metamaterial. Such a model requires complex valued directional couplings between masses, where only out-of-plane DOFs are allowed. As this two-band system implies a scalar field, it is much more feasible for experimental realization (e.g. with acoustic pressure field) compared to higher-band systems. However, the non-physical couplings could not be obtained with existing devices. We employed a feedback-based design, in which these couplings were generated via real-time autonomous control of a Graphene-like lattice. External control forces were applied to the masses, based on measured displacements and velocities of next-nearest-neighbor sites. The required non-physical couplings were encoded into the entries of the controller matrix. We demonstrated that the resulting system has all the band properties of the quantum Haldane model and supports strictly uni-directional, topologically protected, wave propagation along the metamaterial edges. Our feedback-based approach has many advantages in condensed matter research, as it offers generation of arbitrary couplings on a single platform only by means of changing a control program. In particular, it has the ability to break TRS with minimal number of bands, break spatial symmetry, add gain or loss, implement long range hopping or non-local interactions, correct for imperfections, reduce dissipation and to change lattice parameters in real time (faster than the acoustic waves propagation). We thank Moshe Goldstein and Daniel Sabsovich for fruitful discussions. This research was supported in part by the Israel Science Foundation Grant No. 968/16 and by the Israeli Ministry of Science and Technology.

ArXiv

--- abstract: 'Vacuum solutions to the Einstein equations can be viewed as the interplay between the geometry and the gravitational wave energy content. The constraints on initial data reflect this interaction. We assume we are looking at cosmological solutions to the Einstein equations so we assume that the 3-space is compact, without boundary. In this article we investigate, using both analytic and numerical techniques, what happens when the extrinsic curvature is increased while the background geometry is held fixed. This is equivalent to trying to magnify the local gravitational wave kinetic energy on an unchanged background. We find that the physical intrinsic curvature does not blow up. Rather the local volume of space expands to accommodate this attempt to increase the kinetic energy.' author: - Shan Bai - 'Niall [Ó]{} Murchadha' title: Scaling up the extrinsic curvature in gravitational initial data --- Introduction ============ Initial data for the Einstein equations is usually constructed by the conformal method. One is given ‘free’ data and rescales it to get the physical initial data. This is necessary because the Einstein initial data has constraints. A comprehensive discussion of the constraints can be found in [@CB], especially in Chapter VII. Interesting physics tends to occur at the boundaries of the space of free data: one gets at the very least some insight into the limitations of the conformal method. The free data consists of a ‘base’ metric, a Riemannian 3-metric, a divergence-free, trace-free symmetric tensor ( a TT tensor), and a scalar which is the trace of the extrinsic curvature. If the domain is a compact manifold without boundary we often choose the scalar as a constant. We then use a conformal transformation to solve the constraints. If neither the TT tensor nor the constant vanish, we can always find the appropriate conformal factor [@OMY]. Parts of the boundary of the free data are easily accesible. We can scale any one of the three parts by multiplying it by a constant and letting the constant become either large or small. What effect has this on the physics? For example, does the solution just cease to exist, does the volume of the spacetime blow up (or shrink to zero), do apparent horizons (which may be interpreted as cosmological horizons) appear? Maxwell’s equations also have constrained initial data. By choosing the Maxwell free data as a pair of 3-vectors, $(\vec{A}, \vec{F})$, a parallel can be seen between the two fields. The magnetic and electric fields can be generated via $\vec{B} = \nabla\times\vec{A}$ and $\vec{E} = \vec{F} - \vec{\nabla} V$, where $V$ is a scalar function chosen to satisfy $\nabla^2V = \nabla_iF^i$. It is clear that if we multiply either $\vec{A}$ or $\vec{F}$ by any constant, $\vec{B}$ or $\vec{E}$ will be multiplied by the same amount. As a result, this kind of rescaling can be used to increase the electromagnetic energy density $(E^2 + B^2)$ without limit. Can we perform such a rescaling in General Relativity? Can we increase the gravitational wave energy at will by multiplying any part of the free data by a large constant? Because the gravitational wave energy interacts in a very nonlinear way with the geometry, it is not clear what happens with the Einstein equations. In this article, we discuss one such rescaling of the gravitational free data. The initial data consists of a 3-metric, $g_{ij}$, and a symmetric tensor, $K^{ij}$ which is the extrinsic curvature of the slice. This is essentially the velocity of the 3-metric. $K^{ij}$ is generated from the TT tensor and the constant. The constant is the Hubble constant, it represents a uniform expansion or contraction of the space, while the TT tensor (which has only 2 degrees of freedom per space point) can be interpreted as the gravitational wave velocity. We multiply the term that represents the gravitational wave velocity by a large number and track the solution. It turns out that the local energy density remains constant, while the volume of the universe expands. This article is a combination of analytic and numerical works. Although we study a highly nonlinear elliptic equation, nevertheless we obtain some interesting analytical results. In addition, we can accurately solve the equation numerically. This numerical work not only confirms the analytic results, where available, but also indicates the behaviour of solutions in situations where we cannot prove anything. In Section II we confine ourselves to describing the conformal method of solving the constraints. In Section III we show that if the extrinsic curvature vanishes nowhere on a compact manifold and we increase it, the conformal factor uniformly blows up. However, when the extrinsic curvature vanishes somewhere, the analysis in Section III is no longer valid. To investigate this special case, we revert to a spherically symmetric toy model, deriving some analytic results in Section IV and showing the numerical work in Section V. We distinguish between data sets where the extrinsic curvature vanishes in finite regions and when it only vanishes at an isolated point. Here we supply strong evidence that we do not get blow-up in regions of vanishing extrinsic curvature, and that we get slow blow-up when the extrinsic curvature vanishes at an isolated point. We conclude with a summary and an outline of future work. Solving the Einstein Constraints ================================ Initial data for the Einstein equations consists of two parts: the first part is a manifold equipped with a Riemannian 3-metric $g_{ij}$, and the second is a symmetric tensor $K^{ij}$ on the same manifold. $K^{ij}$ is the extrinsic curvature of the 3-slice, i.e., the time derivative of the 3-geometry. The metric and extrinsic curvature cannot be chosen arbitrarily. They must satisfy two constraints: the first is the Hamiltonian constraint, $$R-K_{ij}K^{ij} + K^2= 0, \label{H}$$ where $R$ is the 3-scalar curvature of $g_{ij}$ and $K$ is the trace of $K^{ij}$, i.e., $K = g_{ij}K^{ij}$. The second, the so-called momentum constraint, is $$\nabla_iK^{ij} - \nabla^jK = 0.\label{M}$$ The terminology and notation comes from [@ADM]. The standard way of generating solutions to these equations is by means of a conformal transformation. On any given manifold, it is easy to construct TT tensors [@JY]. These are tensors that are both tracefree and divergencefree, $g_{ij}K^{ij}_{TT} = 0; \nabla_iK^{ij}_{TT} = 0$. Such $TT$ tensors are conformally covariant. If one multiplies the given base metric $g_{ij}$ by an arbitrary positive function $\phi$ to construct a new metric , $\bar{g}_{ij} = \phi^4g_{ij}$, then $\bar{K}^{ij}_{TT} = \phi^{-10}K^{ij}_{TT}$ is TT with respect to $\bar{g}$. [@JY] Any $K^{ij}$ which is the sum of a $TT$ part and a constant trace, i.e., $K^{ij} = K^{ij}_{TT} + \frac{1}{3}Kg^{ij}$, where $K$ is a constant, satisfies the momentum constraint Eq.(\[M\]). If we make a conformal tranformation of the metric, $\bar{g}_{ij} = \phi^4g_{ij}$, we find that $\bar{K}^{ij} = \phi^{-10}K^{ij}_{TT} + \frac{1}{3}K\bar{g}^{ij}$ again satisfies the momentum constraint with respect to the new metric. Note that $K$ is not transformed; rather, it remains a given constant. In the conformal method, one uses the fact that we can freely choose the conformal factor to solve the Hamiltonian constraint. If we conformally transform the metric, $\bar{g}_{ij} = \phi^4g_{ij}$, we find that the scalar curvature transforms as $$\bar{R} = \phi^{-4}R - 8 \phi^{-5}\nabla^2\phi. \label{barR}$$ We want the final metric and the final extrinsic curvature to satisfy the Hamiltonian constraint Eq.(\[H\]), $\bar{R} - \bar{K}^{ij}\bar{K}_{ij} + K^2 = \bar{R} - \bar{K}^{ij}_{TT}\bar{K}^{TT}_{ij} + \frac{2}{3}K^2 = 0$. This reduces to the famous Lichnerowicz-York equation [@L; @JY] $$\begin{aligned} \nabla^2\phi -\frac{R}{8}\phi + \frac{1}{8} K^{TT}_{ij}K^{ij}_{TT} \phi^{-7} - \frac{K^2}{12}\phi^5&=& 0\\ \nabla^2\phi -\frac{R}{8}\phi + \frac{1}{8} A^2 \phi^{-7} - \frac{K^2}{12}\phi^5&=&0\label{LY}\end{aligned}$$ where $A^2= K^{TT}_{ij}K^{ij}_{TT}$. This equation is very well behaved [@OMY]. In this article we focus on the situation where the topology of the 3-manifold is compact and without boundary. It can be shown that Eq.(\[LY\]) has a unique positive solution if $K\ne 0$ and if $K^{ij}_{TT}$ is not identically zero. It need not be non-zero everywhere; it is enough that it not vanish somewhere [@OMY]. Let us remind the reader that $K$ is a constant, while $K^{ij}_{TT}$ is a function. In the special cases, where either $K = 0$ or $K^{ij}_{TT} \equiv 0$, we have an extra condition related to the sign of the scalar curvature. However, in the general case no such restriction applies. This existence result does not depend either on the metric, other than it be uniformly elliptic, or on the topology of the 3-manifold. To recapitulate: we start with a triplet, i.e., the free data, $(g_{ij}, K^{ij}_{TT}, K)$, and construct a new set $(\bar{g}_{ij}, \bar{K}^{ij}_{TT}, K) = (\phi^4g_{ij}, \phi^{-10}K^{ij}_{TT}, K)$ that satisfies the constraints. We would like to investigate the boundary of the set of free data. There exist parts of the boundary that are easy to reach. We can change the original triplet in a simple way by multiplying any one of them by a constant without touching the other two, because, for example, multiplying a TT tensor by a constant does not change its TT’ness with respect to a fixed metric. Of course, such a transformation will change the solution, $\phi$, of Eq.(\[LY\]), and thus the data satisfying the constraints will be different. In this article we will consider such a rescaling. We pick one triplet $(g_{ij}, K^{ij}_{TT}, K)$, and use it to construct a family of free data of the form $(g_{ij}, \alpha^{12}K^{ij}_{TT}, K)$, where $\alpha$ is a running parameter. In particular, we wish to focus attention on the behaviour of $\phi$ as $\alpha$ becomes large in order to see what happens to the physical initial data that emerges. This is a particularly interesting part of the boundary. One can regard $K^{ij}_{TT}$ as the velocity of the gravitational waves coded into the initial data. This rescaling seems to push up the gravitational wave energy content of the free data. Therefore, we are trying to increase the gravitational wave energy while controlling the rest of the geometry as much as we can. In the special case where $K^{ij}_{TT}$ vanishes nowhere, we can show that the conformal factor uniformly blows up. We demonstrate this in the next section. a Harnack type inequality for the conformal factor ================================================== We want to look at Eq.(\[LY\]) $$\nabla^2\phi -\frac{R}{8}\phi + \frac{1}{8} A^2 \phi^{-7} - \frac{K^2}{12}\phi^5 = 0,$$ where $A^2= K^{TT}_{ij}K^{ij}_{TT}$. We are interested in the situation where we scale $K^{TT}_{ij}K^{ij}_{TT}$ by a constant $\alpha^{12}$. Therefore we look at $$\nabla^2\phi -\frac{R}{8}\phi + \alpha^{12}\frac{1}{8} A^2 \phi^{-7} - \frac{K^2}{12}\phi^5 = 0,\label{LY1}$$ and we wish to show that $\phi$ scales linearly with $\alpha$ as $\alpha$ becomes large. We write $\tilde\phi = \phi/\alpha$ and then Eq.(\[LY1\]) becomes $$\nabla^2\tilde\phi -\frac{R}{8}\tilde\phi + \alpha^{4}(\frac{1}{8} A^2 \tilde\phi^{-7} - \frac{K^2}{12}\tilde\phi^5) = 0.\label{LY2}$$ We wish to solve the family of equations on a compact manifold without boundary. It turns out that the sign of the scalar curvature plays a minor role in the behaviour of the solutions. We can always set the scalar curvature to a constant value because of the Yamabe theorem [@Y], which tells us that any Riemannian metric on a compact manifold can be conformally transformed to a metric of constant scalar curvature (this really only makes sense on a compact manifold). The key quantity is the Yamabe number $$Y = \inf \frac{\int[(\nabla \theta)^2 + \frac{1}{8}R\theta^2]dv}{[\int \theta^6 dv]^{1/3}},$$ where the infimum is taken over all smooth functions, $\theta$. The sign of the Yamabe number fixes the sign of the constant scalar curvature. Since Eq.(\[LY2\]) is conformally covariant, and since conformal transformations form a group under composition, we can set $R$ to a constant value without losing any generality. However, we do need to handle the three separate cases, $Y > 0, R > 0, \ \ Y < 0, R < 0,$ and $Y = 0, R = 0$ independently. In each case we will set the value of $K^2 = 9$. This choice does not change in any fundamental way the behaviour of the solution. $Y > 0,\ \ R > 0$ -------------------- We assume that we are in the positive Yamabe class and set the scalar curvature $R = +24$; the specific number can be chosen freely. Now Eq.(\[LY2\]) reduces to $$\nabla^2\tilde\phi - 3\tilde\phi + \alpha^{4}(\frac{1}{8} A^2 \tilde\phi^{-7} - \frac{3}{4}\tilde\phi^5) = 0.\label{LY3}$$ Eq.(\[LY3\]), because it is just a rescaled version of the original Lichnerowicz-York equation Eq.(\[LY\]), which is extremely well behaved, has a regular positive solution. Let us look at what happens at the maximum of $\tilde{\phi}$, which we shall assume occurs at a point $r = r_{max}$. The first two terms in Eq.(\[LY3\]) will be negative at $r = r_{max}$ so we get $$\left[\frac{1}{8}A^2\tilde{\phi}^{-7} - \frac{3}{4}\tilde{\phi}^5\right]_{r_{max}} > 0\label{rmax1}.$$ This becomes $$[\max\tilde{\phi}]^{12} < \frac{1}{6}A^2|_{r_{max}} \le \frac{1}{6}\max A^2. \label{maxA}$$ The second inequality is needed because the location of $r_{max}$ may well depend on $\alpha$ and, in general, does not coincide with the maximum of $A^2$. However, Eq.(\[maxA\]) gives us a uniform upper bound on $\tilde{\phi}$ as a constant independent of $\alpha$. We get a uniform lower bound by looking at Eq.(\[LY3\]) when $\tilde{\phi}$ is a minimum, which we shall assume occurs at $r = r_{min}$. At $r = r_{min}$ we get $$\frac{1}{6}\min A^2 \le \frac{1}{6} A^2|_{r_{min}} \le 4\frac{[\min\tilde{\phi}]^{8}}{\alpha^4} + [\min\tilde{\phi}]^{12}. \label{minphi}$$ Consider the cubic equation $$x^3 + 4\frac{x^2}{\alpha^4} - \frac{1}{6}\min A^2 = 0.\label{x}$$ This will have a positive root that increases as $\alpha$ increases. Fix $\alpha$, $\alpha = \alpha_0$, and find the positive root of Eq.(\[x\]) for $\alpha = \alpha_0$. This number is a lower bound for $[\min\tilde{\phi}]^4$ that is independent of $\alpha$ for all $\alpha > \alpha_0$. In turn, this means that $\min\phi$ diverges at least as fast as $\alpha$ as $\alpha$ becomes large. Using the bounds on both $\min\phi$ and $\max\phi$, we have shown that there exists a universal constant $C_0$ independent of $\alpha$ such that $$\frac{\min\phi}{\max\phi} > C_0 \gtrsim \left[\frac{\min A^2}{\max A^2}\right]^{1/3}.$$ The maximum and minimum of $\phi$ both increase together proportional to $\alpha $ so that their ratio remains bounded independent of $\alpha$. This can be regarded as a version of the Harnack inequality [@GT] for the non-linear equation Eq.(\[LY3\]). $Y < 0,\ \ R < 0$ -------------------- We assume that we are in the negative Yamabe class, and set the scalar curvature $R = -24$. Now Eq.(\[LY2\]) reduces to $$\nabla^2\tilde\phi + 3\tilde\phi + \alpha^{4}(\frac{1}{8} A^2 \tilde\phi^{-7} - \frac{3}{4}\tilde\phi^5) = 0.\label{LY4}$$ Let us look again at what happens at the maximum of $\tilde{\phi}$. We again assume that this occurs at a point $r = r_{max}$. The first term in Eq.(\[LY4\]), $\nabla^2\tilde\phi$, will be negative at $r = r_{max}$ so we get $$\left[\frac{1}{8}A^2\tilde{\phi}^{-7} - \frac{3}{4}\tilde{\phi}^5 + \frac{3\tilde\phi}{\alpha^4}\right]_{r_{max}} > 0\label{rmax2}.$$ This becomes $$[\max\tilde{\phi}]^{12} - \frac{4[\max\tilde{\phi}]^8}{\alpha^4}< \frac{1}{6}A^2|_{r_{max}} < \frac{1}{6}\max A^2. \label{maxA1}$$ Consider the cubic equation $$x^3 - \frac{4x^2}{\alpha^4} - \frac{1}{6}\max A^2 = 0.\label{x2}$$ This equation has a positive root which is an upper bound for $[\max\tilde{\phi}]^{4}$ for any given $\alpha$. This root decreases with increasing $\alpha$. An easy way to confirm this is to differentiate Eq.(\[x2\]) with respect to $\alpha$. We get $$(3x^2 - \frac{8x}{\alpha^4})\frac{dx}{d\alpha} + \frac{16x^2}{\alpha^5} = 0.$$ Using Eq.(\[maxA1\]), we get $3x^2 - 8x/\alpha^4 = 4x/\alpha^4+ \max A^2/6x > 0.$ This means that $dx/d\alpha < 0$, as required. As in the positive Yamabe case, pick a value for $\alpha$, $\alpha = \alpha_0$, find the root of Eq.(\[x2\]), and it will be a uniform upper bound of $\max\tilde{\phi}^4$ for all $\alpha > \alpha_0$. We get a uniform lower bound by looking at Eq.(\[LY4\]) at the point where $\tilde{\phi}$ is a minimum. We assume that this occurs at $r = r_{min}$. The first two terms in Eq.(\[LY4\]) are positive at the minimum. Therefore we get $$\frac{1}{6}\min A^2 \le \frac{1}{6} A^2|_{r_{min}} \le [\min\tilde{\phi}]^{12}. \label{minphi1}$$ This inequality Eq.(\[minphi1\]) gives the desired lower bound for $[\min\tilde{\phi}]$, which is independent of $\alpha$. In turn, this means that $\min\phi$ diverges at least as fast as $\alpha$, as $\alpha$ becomes large. Using the bounds on both $\min\phi$ and $\max\phi$, we have shown that there again exists a universal constant $C_0$ independent of $\alpha$ such that $$\frac{\min\phi}{\max\phi} > C_0 \gtrsim \left[\frac{\min A^2}{\max A^2}\right]^{1/3},$$ and we again recover a Harnack inequality, but now for Eq.(\[LY4\]). $Y = 0,\ \ R = 0$ ----------------- Now Eq.(\[LY2\]) can be reduced to $$\nabla^2\tilde\phi + \alpha^{4}(\frac{1}{8} A^2 \tilde\phi^{-7} - \frac{3}{4}\tilde\phi^5) = 0.\label{LY5}$$ Showing the existence of uniform bounds in this case is even easier. We immediately get $$[\max\tilde{\phi}]^{12} < \frac{1}{6}A^2|_{r_{max}} < \frac{1}{6}\max A^2. \label{maxA2}$$ and $$[\min\tilde{\phi}]^{12} > \frac{1}{6}A^2|_{r_{min}} > \frac{1}{6}\min A^2. \label{minA2}$$ Again we get $$\frac{\min\phi}{\max\phi} > C_0 > \left[\frac{\min A^2}{\max A^2}\right]^{1/3},$$ Discussion ---------- The Harnack inequalities that have been derived in the last three subsections are clearly only valid when $\min(A^2) = \min(K^{ij}_{TT}K^{TT}_{ij}) \ne 0$. Since $TT$ tensors are usually constructed by a decomposition method, one might think that it would be difficult to find $TT$ tensors which vanish either at points or in regions. Interestingly, we can construct such $TT$ tensors [@BaOM] and they cannot be ignored. It has not yet been possible to derive general results for such special TT tensors. Therefore we revert to considering only a spherically symmetric toy model. It turns out, in the spherical situation, that the case where $A^2$ vanishes in a region is easier to analyse than the case where $A^2$ vanishes at a point. When $A^2$ vanishes in a region, we can prove that the minimum of $\phi$ has an upper bound and does not scale with $\alpha$. In the region where $A^2 \ne 0$ we get the standard linear scaling with $\alpha$. We use a combination of analytical and numerical techniques in the next two sections deal with this spherical model. A Spherical Toy Model: Analytical results ========================================= We will restrict our attention to the case where the base 3-metric is spherically symmetric, and we replace the position-dependent function $A^2$ by an arbitrary spherically symmetric function because we know there exists no regular spherically symmetric $TT$ tensor on flat space or on a spherical compact manifold without boundary[@BOM1]. For convenience, we set $K = 3$. We start with a round 3-sphere of constant scalar curvature $R_0$ (a natural choice is $R_0 = 24$) and seek solutions to $$\begin{aligned} \nabla^2\phi - \frac{1}{8}R_0\phi + \frac{1}{8}\alpha^{12}A^2\phi^{-7} - \frac{2K^2}{24}\phi^5 &=& 0\\ \nabla^2\phi - 3\phi + \frac{1}{8}\alpha^{12}A^2\phi^{-7} - \frac{3}{4}\phi^5 &=& 0\label{1}\end{aligned}$$ for varying $\alpha$. It is clear that when either $\alpha \rightarrow 0$ or $\alpha \rightarrow \infty$ strange things happen. In keeping with the focus of this article, we will only consider here the case where $\alpha \rightarrow \infty$. Looking at Eq.(\[1\]), it is clear that $\phi$ cannot remain regular as $\alpha \rightarrow \infty$, because the third term would diverge while the others remain regular. The results derived in Section III remain valid so we know that $\phi$ is linearly proportional to $\alpha$ as long as $A^2$ is nowhere zero. We need to deal with the situation where $A^2$ vanishes somewhere. In particular, let us consider the situation where $A^2$ vanishes in a spherical region around the north pole. When dealing with spherical symmetry, we are free to take advantage of the fact that a round 3-sphere can be decompactified to flat 3-space, and that conformal transformations form a group under composition. In this picture the equation we wish to study is $$\nabla^2\hat\phi + \frac{1}{8}\alpha^{12}\hat A^2\hat\phi^{-7} - \frac{3}{4}\hat\phi^5 = 0,\label{2}$$ with $\hat\phi \rightarrow 0$ at infinity. The conformal factor that maps flat space into a round sphere of scalar curvature equalling 24 is $\theta = \sqrt{b}/\sqrt{b^2 + r^2}$ for any $b > 0$. The mapping that brings us from Eq.(\[1\]) to Eq.(\[2\]) requires that $A^2$ and $\phi$ be transformed. The transformation is that $\hat A^2 = \theta^{-12}A^2$ and $\hat\phi = \theta^{-1}\phi$. In other words, $\phi$ will be finite at the ‘point at infinity’ in the compact manifold ,while $\hat\phi \approx \sqrt{b}/r$ at the corresponding infinity in $R^3$. A spherical region around the north pole corresponds to a disc $0 \le r < r_1$, where $r$ is the standard conformally flat coordinate radius, on which $A^2 = 0$. It is now convenient to switch to the asymptotically flat picture. The equation we consider is Eq.(\[2\]), and, in the disc $0\le r < r_1$ it reduces to $$\nabla^2\hat\phi - \frac{3}{4}\hat\phi^5 = 0.\label{2d}$$ We can write down the solution of this equation explicitly. It is $$\hat\phi = \frac{\sqrt{a}}{\sqrt{a^2 - r^2}},\label{2e}$$ where $a$ is a parameter. These are the functions that map one from flat space to the round hyperboloid of constant negative scalar curvature. These functions blow up at $r = a$, and since we know that the total solution is regular, we know that the blow-up must occur outside the range of validity of these functions. While we do not know [*a priori*]{} the value of $a$, and it will change with $\alpha$, we do know that we have a lower bound for $a$, i.e., $a \ge r_1$. The minimum value of $\hat\phi$ for these special solutions occurs at the origin. There we get $$\min\hat\phi = \hat\phi(r = 0) = \sqrt{\frac{1}{a}} \le \sqrt{\frac{1}{r_1}}.$$ Therefore $\min\phi$ does not blow up like $\alpha$ but reaches some limit, while $\max\phi$ becomes unboundedly large. A similar argument holds when $A^2$ vanishes near the south pole. We can repeat this argument when $A^2$ vanishes on some belt $r_1 < r < r_2$. The equation, Eq.(\[2d\]), is still the same, but now is valid for $r_1 < r < r_2$. We cannot write down a set of explicit solutions, but we have much information about the solutions. Fix the location of the minimum, in this case in the interval $(r_1, r_2)$, and fix the value of $\hat\phi$ at the minimum. This uniquely determines the solution. The solution is ‘U’ shaped, blowing up twice at $r_A$ and $r_B$. The bigger the $\min\hat\phi$, the narrower the ‘U’, i.e., $(r_B - r_A)(\min\hat\phi)^2$ is bounded [@CHOM]. However, we know that $(r_A < r_1)$ and $(r_B > r_2)$ because the blow-up cannot occur in the range of validity of Eq(\[2d\]). Therefore $r_B - r_A > r_2 - r_1$, and so the value of $\min\hat\phi$ is bounded above. The more interesting case is when $A^2$ vanishes at a point rather than in a region. To repeat: we can show that when $A^2$ vanishes nowhere, we have uniform blow-up over the entire domain; when $A^2$ vanishes in a region, the minimum saturates. We conjecture that we will have behaviour which is ‘halfway’ between the two situations dealt with above. More precisely, we conjecture that the minimum will blow up with $\alpha$, but at a rate which is slower than linear. The details will depend on the rate that $A^2$ goes to zero at the point. We have done some numerical modeling, both to confirm the analytic results and to investigate those situations where we can prove nothing concrete. These models will be discussed in the next section. A spherical toy model: numerical results ======================================== In this section, the equation we deal with is a one-dimensional elliptic equation with a highly nonlinear source term, $$\nabla^2\phi + S(\phi) =0.$$ We use a one-dimensional pseudo spectral method to solve this equation in two computational domains [@Ansorg; @Ansorg2; @Pfeiffer]. We have a coordinate $y \in [0, 2)$ replacing $r \in [0, \infty)$. In the interior domain we have $y = r, y \in [0, 1], r \in [0, 1]$, while in the exterior domain we have $y = 2 - 1/r, y \in (1, 2), r \in (1, \infty)$. Therefore, infinity can be included when we put the computational grid at the point $y=2$. We use Chebyshev polynomials as basis functions and the collocation points are $$\begin{aligned} y_{i} &=& \cos [ \frac{\pi i}{2n} ]\ \ {\rm in}\ \ [0, 1]\\ y_{j} &=& 1 + \cos [ \frac{\pi j}{2n} ]\ \ {\rm in}\ \ \ [1, 2]\end{aligned}$$ where $i=0, 1 ...n$ and $j=0,1, 2...n$. The two domains meet at $i=0$ on the interior domain and $j=n$ on the exterior domain. The grid point $j=0$ touches the infinity of coordinate $r$. We require the solution to be $C^1$. This means that $\phi$ and its normal derivative must match at the interface between the domains. The discrete equations are solved by the Newton-Raphson method. The resolution is taken as 100 on each domain. $A^2 = 1$ --------- We here consider the simplest case, where we choose the base metric to be a round sphere with scalar curvature equalling 24, and pick $A^2$ to be a global constant equalling 1. Then Eq.(\[1\]) $$\nabla^2\phi - 3\phi + \frac{1}{8}\alpha^{12}A^2\phi^{-7} - \frac{3}{4}\phi^5 = 0,$$ reduces to an algebraic equation $$\phi^{12} + 4\phi^8 - \frac{1}{6}\alpha^{12} = 0,$$ or in terms of the normalised $\tilde{\phi} = \phi/\alpha$, $$\tilde{\phi}^{12} + \frac{8\tilde{\phi}^8}{\alpha^4} - \frac{1}{6} = 0.$$ As $\alpha \rightarrow \infty$, we expect $\tilde{\phi}^{12} \rightarrow \frac{1}{6}$ from below, or $\tilde{\phi} \rightarrow 0.8612992$. The round base metric can be written as $$g_{ij} = \frac{1}{(1 + r^2)^2}\delta_{ij}.$$ The computation is all done in the flat space, so we have to solve $$\nabla^2\hat\phi + \frac{1}{8}\alpha^{12}\hat A^2\hat\phi^{-7} - \frac{3}{4}\hat\phi^5 = 0,\label{2c}$$ where $\hat A^2 = \frac{1}{(1 + r^2)^6}$. At the risk of confusion we introduce a normalized $\hat\varphi = \hat\phi/\alpha$, then Eq.(\[2c\]) becomes $$\nabla^2\hat\phi' + \frac{1}{8}\alpha^{4}\hat A^2\hat\varphi^{-7} - \frac{3}{4}\alpha^4\hat\varphi^5 = 0.\label{2b}$$ The relationship between $\tilde{\phi}$ and $\hat\varphi$ is $\tilde{\phi}= \sqrt{1 + r^2}\hat\varphi$. We solve Eq.(\[2b\]) numerically with $\hat A^2 = \frac{1}{(1 + r^2)^6}$. We see that the solution is, as expected, of the form $\hat\phi' = C/\sqrt{1 + r^2}$ with $C$ depending on $\alpha$. In Figure 1 we present $\tilde{\phi}$ on the compact manifold (by multiplying $\hat\phi'$ by $\sqrt{1 + r^2}$). We clearly see that we get a family of constant functions that asymptote to a fixed function as $\alpha$ becomes large. {width="7cm" height="5cm"} We list the values of $C$ w.r.t. $\alpha$ as the table I. $\alpha$ C ------------- -------------------- 1 0.6680900040565467 2.154434690 0.8387341686204405 4.641588834 0.8601789420392659 10 0.8612470753749010 21.54434690 0.8612968154854488 46.41972069 0.8612991245657523 : It is a list of the scaling parameter $\alpha$ and the constancy of the standard solution $\hat\phi'$. It is clear that, as expected, $C$ asymptotes to 0.8612992. $A^2$ vanishing at a single point --------------------------------- A simple choice of $\hat A^2$ that corresponds to $A^2$ vanishing at the south pole is to pick $\hat A^2 = \frac{1}{(1 + r^2)^\beta}$ with $\beta > 6$. We restrict our attentions to $\beta=7$ and $\beta=10$. We present $\hat\phi$ for several values of $\alpha$ with $\beta=7$ and $\beta=10$ in Figure 2 and 3, where $\alpha = 1,2,3,4,5,10$ gives rise to the red, green, blue, pink, light blue, and yellow lines respectively. ![$\hat\phi$ on $S^3$ with $\beta=7$ []{data-label=""}](2.eps){width="7cm" height="5cm"} ![ $\hat\phi$ on $S^3$ with $\beta=10$[]{data-label=""}](3.eps){width="7cm" height="5cm"} It is clear that, in both cases, the maximum of $\hat\phi$, at the north pole, settles to a constant value independent of $\alpha$. This shows that the maximum of $\phi$ grows linearly with $\alpha$. On the other hand, the minimum of $\hat\phi$, at the south pole, where $A^2 = 0$, decreases with increasing $\alpha$. Hence the minimum of $\phi$ does not grow linearly with $\alpha$. To analyse the behaviour of the minimum of $\phi$ as a function of $\alpha$, we consider each of the two cases, i.e., $\beta = 7$ and $\beta=10$. We conjecture that $\phi(\theta = \pi)$ scales with some power of $\alpha$, for large $\alpha$. We plot $\ln\phi/\ln\alpha$ versus $\alpha$ for each of the two choices of $\beta$. These are Figures 4 and 5. Each of the two curves flattens out for large $\alpha$. These show that $\min\phi$ grows like $\alpha^n$ with $n \approx 0.71$ when $\beta = 7$, and $\min\phi$ grows like $\alpha^n$ with $n \approx 0.405$ when $\beta = 10$. {width="7cm" height="5cm"} {width="7cm" height="5cm"} If $\beta = 6$, $A^2$ does not vanish at the south pole and we get linear growth of $\phi$ there. The limit $\beta = \infty$ corresponds to $A^2$ vanishing in a region near the south pole, and we get no growth at all. It is nice to see that the growth rate seems to diminish smoothly as we move from $\beta = 6$ to $\beta = \infty$. $A^2$ vanishing in a patch -------------------------- We start by finding, on the round sphere, a smooth positive spherical function that vanishes in a region. We will use this as $A^2$. We first find a cut function $\eta(x)$ which is defined on $[0,2]$ by $$\eta(x) = \left\{ \begin{array}{cc} 0 & x\in[0,0.5) \\ 209-2240x+10080x^2&\\ -24640x^3 +35280x^4-29568x^5& \\ +13440x^6-2560x^7& x\in[0.5,1.0) \\ 1&x \in[1.0,\infty) \end{array} \right.$$ This function smoothly interpolates between 0 at $x = 0.5$ and 1 at $x = 1$ and is $C^3$ at each end. In terms of the rescaled coordinate function $y$ which we introduced at the beginning of this section, we construct the following function $$f(y) = \left\{ \begin{array}{cc} \eta(1-y) & y\in[0,0.5) \\ 0 & \ \ \ y\in[0.5,1.25) \\ \eta(2y-2)& \ \ \ y\in[1.25,1.5)\\ 1& y\in[1.5,2) \end{array} \right.$$ This is a function that is 1 at the origin, goes to zero at $y = r = 0.5$, is zero in $y \in (0.5, 1.25)$, rises to 1 in $y \in (1.25, 1.5)$, and stays equal to unity out to $y = 2, r = \infty$. We choose $\hat{A}^2 = f(y)/(1 + r^2)^6$. This corresponds to a smooth $A^2$ on the sphere, which equals 1 at both the north and south poles but equals zero in a central region. Figure 6 shows $A^2$ on the sphere. ![$A^2$ on $S^3$[]{data-label=""}](6.eps){width="7cm" height="5cm"} We solve the equation for a range of parameters $\alpha$. We display the various solutions in Figure 7, showing the normalized $\hat\phi = \phi/\alpha$ on $S^3$. The parameters used are $\alpha = 1.000$, $2.154$, $4.642$, $10.00$, $21.54$, $53.13$, $79.37$, and $100.0$. These correspond to the red, green, blue, pink, light blue, yellow, dark blue, and mauve lines respectively on Figure 7. {width="7cm" height="5cm"} One can clearly see that the value of $\hat\phi$, on the support of $A^2$, tends to a stationary limit, while $\hat\phi$ collapses off the support of $A^2$. This shows that $\phi$ scales linearly with $\alpha$, on the support of $A^2$, while in the region where $A^2$ is zero, $\hat\phi$ continues to diminish so that $\phi$ approaches a stationary value. We wish to show that $\min\phi = \alpha\min\hat\phi$ increases with $\alpha$ but approaches some fixed upper bound. The minimum occurs around $\theta = 1.35$. In Figure 8 we have plotted $\ln \phi/\ln \alpha$ versus $\alpha$ at $\theta = 1.35$. One should compare this graph with Figures 4 and 5. This looks like a graph which is going to an asymptotic value of 0, which indicates that $\phi$, at $\theta = 1.35$, is heading for a fixed number, independent of $\alpha$. {width="7cm" height="5cm"} Conclusions =========== We have shown that if we scale $K^{ij}_{TT}K^{TT}{ij}$ by $\alpha^{12}$, we find that the conformal factor, in general, scales like $\alpha$. However, the physical $\bar{K}^{ij}_{TT}\bar{K}^{TT}_{ij} = \phi^{-12}\alpha^{12}K^{ij}_{TT}K^{TT}_{ij} = \tilde\phi^{-12}K^{ij}_{TT}K^{TT}_{ij}$, and as $\alpha$ becomes larger and larger $\tilde\phi$ remains finite. This means that $\bar{K}^{ij}_{TT}\bar{K}^{TT}_{ij}$ remains finite. Hence the velocity part of the gravitational wave energy density remains bounded even though the corresponding ‘free’ data blows up. On the other hand, the volume of space becomes unboundedly large, because $\sqrt{\bar{g}} = \phi^6\sqrt{g} = \alpha^6\tilde\phi^6\sqrt{g}$ blows up. Therefore the total gravitational wave energy in a coordinate sphere becomes larger and larger while the local energy density remains bounded. We see that the gravitational wave energy inside any coordinate sphere increases like $\alpha^6$ while the surface area increases like $\alpha^4$. This indicates that this family of initial data will eventually contain horizons. The numerics, when combined with the analytic calculations, show a coherent picture. If $K^{ij}_{TT}K^{TT}_{ij}$ has no zeros, then the conformal factor blows up uniformly. If the $K^{ij}_{TT}K^{TT}_{ij}$ is zero on a patch, we expect no blow up on this patch, but, nevertheless, we continue to get the standard blowup elsewhere. If $K^{ij}_{TT}K^{TT}_{ij}$ has an isolated zero, we will get blowup at this point, but at a rate slower than in the rest of the space. The rate of blowup is not universal in this case, but depends on how quickly $K^{ij}_{TT}K^{TT}_{ij}$ moves away from zero at that point. There seems to be a smooth transition between the ‘slow blow-up’ case and the ‘no blow-up’ case. There are a number of obvious extensions to this work. To find a real $TT$ tensor, we need to abandon spherical symmetry and, at the very least, work with axially symmetric data. We would expect that if we have a real $K^{ij}_{TT}K^{TT}_{ij}$, which either vanishes at a point or vanishes in a region, to get behaviour similar to the spherical model. We would be surprised if anything could be proven analytically; we are going to have to depend on numerical modeling. Work is in progress in this direction. It would be interesting to repeat this analysis in the asymptotically flat case. We would probably want to work with maximal initial data, i.e., $K = 0$, and just have a metric and a $TT$ tensor as free data. It is clear that one can change the metric so that the ADM mass becomes unboundedly large and trapped surfaces appear [@BOM]. What happens if we blow-up the extrinsic curvature on a fixed background metric? Will we get the same behaviour? Preliminary investigations indicate that we do: the ADM mass diverges and trapped surfaces appear. We intend to investigate this further. SB and NÓM were supported by Grant 07/RFP/PHYF148 from Science Foundation Ireland. SB would like to thank Marcus Ansorg and José Luis Jaramillo for helpful conversations about the spectral method. Choquet-Bruhat, Y., [*General Relativity and the Einstein Equations,*]{} (Oxford, OUP, 2009). Ó Murchadha, N., and York, J.W., J. Math. Phys. [**14**]{}, 1551 (1973). Arnowitt, R., Deser, S., and Misner, C., in [*Gravitation: an introduction to current research,*]{} ed. L. Witten (Wiley, New York, 1962). York, J.W., J. Math. Phys. [**14**]{}, 456(1973); Ann. Inst. Henti Poincaré, [**21**]{}, 319 (1974). Lichnerowicz, A., J. Math. Pures Appl. [**23**]{}, 39 (1944). Yamabe, H., Osaka Math J. [**12**]{}, 21 (1960); Schoen, R., J. Diff. Geom. [**20**]{}, 479 (1984). see e.g. Gilbarg, D., and Trudinger, N., [*Elliptic Partial Differential Equations of the Second Order*]{}, (Springer, Berlin 1983). Beig, R., and Ó Murchadha, N., Phys. Rev. D[**57,**]{} 4728 (1998). Conboye, R., Hilliard, R., and Ó Murchadha, N., to be published. Bai, S., and Ó Murchadha, N., to be published. Beig, R., and Ó Murchadha, N., Phys. Rev. Lett.[**66,**]{} 2421 (1991); Class. Quantum Grav. [**64**]{}, 419 (1994); Class. Quantum Grav. [**13**]{}, 739 (1996); Ó Murchadha, N. and Xie, N. to be published. Ansorg, M., Bruegmann,B.,and Tichy, W., Phys.Rev. D72, 024018 (2005). Ansorg, M., Phys.Rev.D70, 064011 (2004). Pfeiffer, H., Kidder, L., Scheel, M., and Teukolsky, S., Comput.Phys.Commun. 152 (2003) 253-273.

ArXiv

--- abstract: 'We theoretically investigate the impact of the electrothermal flow on the dielectrophoretic separation of single-walled carbon nanotubes (SWNT). The electrothermal flow is observed to control the motions of semiconducting SWNTs in a sizeable domain near the electrodes under typical experimental conditions, therefore helping the dielectrophoretic force to attract semiconducting SWNTs in a broader range. Moreover, with the increase of the surfactant concentration, the electrothermal flow effect is enhanced, and with the change of frequency, the pattern of the electrothermal flow changes. It is shown that under some typical experimental conditions of dielectrophoresis separation of SWNTs, the electrothermal flow is a dominating factor in determining the motion of SWNTs.' author: - Yuan Lin$^1$ - Junichiro Shiomi$^2$ - Shigeo Maruyama$^2$ - 'Gustav Amberg$^{1,}$' bibliography: - 'DEPbib.bib' title: 'Electrothermal flow in Dielectrophoresis of Single-Walled Carbon Nanotubes' --- Introduction ============ Single-walled carbon nanotubes are key materials in nanotechnology as potential candidates for diverse applications owning to their extraordinary mechanical, thermal, optical and electrical properties [@Saito]. On exploring the utility of the electrical properties, one of the current critical challenges is the separation of metallic (m-SWNTs) and semiconducting SWNTs (s-SWNTs). Among various post-synthesis separation methods devised and applied [@krupke2003; @chen; @Chattopadhyay; @Zheng; @Maeda], dielectrophoresis separation has been demonstrated to be possible with high selectivity and simplicity [@krupke2004]. Furthermore, DEP also allows us to deposit SWNTs to selected sites to construct electric devices [@zhibin1]. While the complexity in the optical measurements gives rise to difficulties in interpretation of the obtained spectra [@Baik], deeper understanding in the transport dynamics of SWNTs under DEP operation is of a great importance. It also allows us to discuss the efficiency and possibility for optimization of system designs and working conditions. Here, numerical simulations should serve as a powerful tool since the dynamical observation of SWNTs transported in the suspension is extremely challenging in experiments. On modeling a SWNT as a prolate ellipsoid, the DEP force on the SWNT in an inhomogeneous AC electric field $\bf E$ is calculated as [@krupke2004], $$\mathbf F_{\text{DEP}}=\frac{\pi ab^2\epsilon_\text{m}}{12}\alpha\nabla|\mathbf E|^2, \label{eqn1}$$ where $\alpha=\displaystyle Re[\frac{\epsilon_p^*-\epsilon_\text{m}^*}{\epsilon_\text{m}^*+(\epsilon_p^*-\epsilon_\text{m}^*)L_p}]$, $\quad\epsilon_{m,p}^*=\displaystyle\epsilon_{m,p}-i\frac{\sigma_{\text{m,p}}}{\omega}$ , $\omega$ is the frequency of the AC field and $\epsilon_{\text{m,p}}^*$ is the complex dielectric permittivity. Here the subscripts $m$ and $p$ denote the medium and particle. The constants $a$ and $b$ denote half the tube-length and tube-radius. The depolarization factor $L_p=\displaystyle\frac{b^2}{2a^2e^3}[ln(\frac{1+e}{1-e})-2e]$, $e=\displaystyle\sqrt{1-\frac{b^2}{a^2}}$. On calculating the DEP force of SWNTs using parameters from electrical measurements of suspended pure SWNTs [@Dai], the force on m-SWNTs becomes several orders higher than that on s-SWNTs, indicating that the separation should be easy. Furthermore, numerical calculations suggested that DEP forces of both m-SWNTs and s-SWNTs are insensitive to variations in frequency of the AC electric field and surfactant concentration within the accessible range in experiments [@dimaki]. This does not agree with the experimental results where the success of DEP separation strongly depended on the frequency and the surfactant concentration [@krupke2004]. Experiments showed that the separation was successful only above some certain frequencies below which both m-SWNTs and s-SWNTs were found on the electrodes. Although the complete picture of the transport dynamics is yet to be revealed, Krupke [*et al*]{} [@krupke2004] attributed the discrepancy to the electrical conduction through the surrounding surfactant. They have represented the surface conduction effect with the effective electrical conductivity which was obtained by fitting the solution of Eq. (1) to the experimental results. Previous numerical simulations on DEP separation of SWNTs have usually considered only the DEP forces and the Brownian motion assuming that the bulk flow velocity is zero. Although the basic concept of DEP-separation in principle is simple, the system involves effects which may cause bulk flow motions such as electroosmosis, thermal convection, electrothermal flow. The electroosmosis effect is only important at low frequencies (less than $10^4$ Hz, which is well below the frequency applied in experiments) [@morgan], and the thermal convection is expected to be negligible in micro and nanoscales. On the other hand, the electrothermal effect is known to be substantial in micro systems [@morgan; @ramos]. Electrothermal flow is driven by a body force caused by electric field acting on gradients in permittivity and/or conductivity due to a non-uniform temperature field [@morgan]. In this paper, we investigate the impact of the electrothermal effect on the DEP-separation of SWNTs by formulating a dynamical model of the integrated system. Here we mainly discuss the transport of s-SWMT, to which electrothermal force has non-trivial effects. We demonstrate that for a commonly used surfactant and electric field with practical magnitude, electrothermal flows can be sufficiently large to have a substantial impact on the separation efficiency of the DEP method. It is shown that electrothermal flows can significantly weaken the DEP-separation by driving the s-SWNTs towards the electrodes, which is consistent with the experimental observations. Mathematical modeling ===================== Using data for a commonly used material in experiments [@krupke2004], the system under consideration consists of SWNTs with a diameter of 1.4 nm and a length of 1 $\mu$m dispersed in aqueous solutions of sodium dodecylbenzene sulfonate (SDBS). For the electrical conductivity, we adopt the effective conductivity (0.35 S/m) suggested by Krupke [*et al*]{} [@krupke2004]. While the influence of surfactant on the electrical conduction of SWNTs surrounded by SDBS is not clear, we adopt the empirically effective value. Similarly, the permittivity of s-SWNTs was set to be $5\epsilon_0$ [@krupke2004]. Although the magnitude of permittivity of s-SWNTs is arguable, a variation within one order of magnitude should not affect the current analysis. Electrical conductivity of SDBS solutions $\sigma_\text{m}$ was estimated as 4 mS/m, 29 mS/m and 230 mS/m for concentration of $0.01\%$, $0.1\%$ and $1\%$, respectively [@krupke2004]. The permittivity of the SDBS solutions $\epsilon_\text{m}$ at room temperature is about $80\epsilon_0$ and has negligible dependence on the concentration. It is straightforward to calculate the DEP force factor for s-SWNTs and the results show that it decreases about one order when the concentration of SDBS increases by one order. We solve the heat conduction equation with a heat source due to an electric field [@ramos], $k\nabla^2T+\sigma_\text{m} E^2=0$, where the thermal conductivity $k$=0.6 Wm$^{-1}$K$^{-1}$. As a typical system geometry in experiments [@krupke2004], a pair of 20 $\mu$m-long and 1 $\mu$m-wide electrodes were located with a distance of 10 $\mu$m at the bottom of the calculation domain as shown in Fig. \[fig1a\]. The boundary conditions are: $T_{\text{wall}}=300$ K at the surrounding walls, and $\displaystyle\frac{dT}{dz}=0$ at the bottom, which means the substrate and electrodes are thermally insulated. Figure \[fig1b\] shows the temperature profiles for various SDBS concentrations above the tip of the electrode ($x,y$)=($0~\mu$m$,5~\mu$m), with the applied AC potential $\phi=20$ V (peak to peak). Maximum temperature differences are 0.3 K, 1.6 K and 11.8 K on the electrode (z=0) for 0.01%, 0.1% and 1% SDBS solutions, respectively. The electrothermal body force acting on the bulk fluid due to the AC electric field is expressed as [@ramos] $$\mathbf f=-\frac{1}{2}[(\frac{\nabla\sigma_\text{m}}{\sigma_\text{m}}-\frac{\nabla\epsilon_\text{m}}{\epsilon_\text{m}})\cdot\mathbf E\frac{\epsilon_\text{m}\mathbf E}{1+(\omega\tau)^2}+\frac{1}{2}|\mathbf E|^2\nabla\epsilon_\text{m}], \label{eqn:3}$$ where $\tau=\epsilon_\text{m}/\sigma_\text{m}$ is the charge relaxation time of the solution. The first term on the right hand side expresses the Coulomb force and the second the dielectric force. For a typical electrolyte solution, $(1/\sigma_\text{m})/(\partial \sigma_\text{m}/\partial T)=0.02$ K$^{-1}$ and $(1/\epsilon_\text{m})/(\partial \epsilon_\text{m}/\partial T)=-0.004$ K$^{-1}$ [@morgan]. The first (second) term is dominant in the low (high) frequency regime (the crossover frequency is $\omega_c=1/\tau$, which is 5.6 MHz and 41 MHz for 0.01% and 0.1% SDBS concentrations). In the calculation, we adopted 300 kHz and 1 GHz, which represent the low and high frequency regime, respectively. For low Reynolds numbers, the transport term of the Navier-Stokes (N-S) equations can be neglected. Hence, the N-S equations become $$\rho_m\frac{\partial \mathbf u_\text{f}}{\partial t}=\eta\nabla^2\mathbf u_\text{f}-\nabla p+\mathbf f \qquad \nabla\cdot\mathbf u_\text{f}=0. \label{eqn:4}$$ For low Reynolds numbers, the inertial effect can be neglected [@morgan]. The terminal velocity of s-SWNTs is calculated as $\mathbf u=\mathbf u_\text{d}+\mathbf u_\text{f}+\mathbf u_\text{b}$, here $\mathbf u_\text{d}=\mathbf F_{\text{DEP}}/f_\text{t}$ with the translational friction coefficient for a prolate ellipsoid given by $\displaystyle f_\text{t}=\frac{6\pi\eta a}{ln(2a/b)}$ [@morgan]. Characteristic variables of electrothermal flow are $U^*=\displaystyle\frac{\partial \sigma_{\textrm m}}{\partial T}\cdot\frac{25\epsilon_\text{m}\phi^4}{k(1+(\omega\tau)^2)\eta L}$, $T^*=\displaystyle\frac{\sigma_\text{m}\phi^2}{k}$, $L$=1 $\mu$m and $t^*=\displaystyle\frac{\partial T}{\partial \sigma_{\textrm m}}\cdot\frac{k(1+(\omega\tau)^2)\eta }{25\epsilon_\text{m}\phi^4}$. Here $\phi$ is the applied peak-to-peak AC voltage. From the Einstein relation, the Brownian velocity can be derived as $\mathbf u_b=\sqrt{6D/\ dt}$ with the diffusion coefficient $D=(k_BT/f_\text{t})$, and $dt$ is the time interval of observation. We define dimensionless variables $\mathbf{\tilde u}=\mathbf u/U^*$, $\tilde T=(T-T_{\text{wall}})/T^*$, $\tilde x=x/L$, $\tilde{dt}=dt/\ t^*$ and $\mathbf{\tilde E}=\mathbf EL/\phi$, dimensionless parameters $ P_1=\displaystyle\frac{(1+(\omega\tau)^2)}{12}$, $ P_2=\displaystyle0.0463\frac{\partial T}{\displaystyle\partial \sigma_{\textrm m}}\frac{ b^2 \alpha ln(\frac{ a}{ b})k(1+(\omega\tau)^2)}{L^2\phi^2}$ and $ P_3=\displaystyle\sqrt{\frac{\partial T}{\partial \sigma_{\textrm m}}\frac{k_B\sigma_\text{m} ln(\frac{2a}{ b})\cdot(1+(\omega\tau)^2)}{0.6\cdot\epsilon_\text{m}\phi^2\pi a}}$. Consequently, we obtain a set of system equations in a non-dimensional form, $$\begin{aligned} \nabla\cdot\mathbf{\tilde E}&=&0, \quad \nabla^2\tilde T=-|\mathbf{\tilde E^2}|,\nonumber\\ \frac{\partial \mathbf{\tilde u_\text{f}}}{\partial\tilde t}&=&\nabla^2\mathbf{\tilde u_\text{f}}-\nabla\tilde p+(\nabla\tilde T\cdot \mathbf{\tilde E}\mathbf){\tilde E}+P_1|\mathbf{\tilde E^2}|\nabla\tilde T, \nonumber \\ \mathbf{\tilde u_{b}}&=&\tilde r\sqrt{ (\tilde T+T_{\text{wall}}/\ T^*)/\ \tilde{dt}},\quad \mathbf{\tilde u_\text{d}}=\nabla|\mathbf{\tilde E}|^2 \nonumber \\ \mathbf{\tilde u}&=&P_2\mathbf{\tilde u_\text{d}}+\mathbf{\tilde u_\text{f}}+P_3\mathbf{\tilde u_\text{b}}, \label{eqn:6}\end{aligned}$$ where $\tilde r$ denotes a random number with normal distribution whose mean is 0 and standard deviation is 1. Equation (4) is solved using finite element method toolbox femLego[@gustav]. Force estimations ================= Since the inertial effect is neglected, the dimensionless velocities are equivalent to the dimensionless forces. Taking the characteristic timescale $t^*$ as the time interval $dt$, the dependence of the relative magnitudes of DEP, Brownian and electrothermal forces on the applied electric potential was examined. Figure \[fig2\] shows the dimensionless forces at ($x,y,z$)=(0 $\mu $m, 14 $\mu$m, 1 $\mu$m) for two different surfactant concentrations 0.01% and 0.1%. The $y$ position was chosen to be in the regime with the positive electrothermal force that attracts SWNTs towards the electrode and $z$=1 $\mu$m equals to the length of SWNTs. Here, the frequency is 300 kHz, which is within the frequency range commonly used in experiments [@krupke2004; @zhibin1]. Figure \[fig2\] shows a clear dependence of the locally dominant force on the applied voltage and surfactant concentration. The change of the Brownian motion due to voltage and surfactant concentration is subtle and relatively negligible. The voltage-dependence of the strength of the electrothermal force is proportional to $\phi^4$ and that of the DEP force is proportional to $\phi^2$. As a result, for both concentrations 0.01% and 0.1%, the electrothermal force overcomes the DEP forces at the high voltage limit since the electrothermal force increases with the voltage much faster than the DEP force. As for the influence of the surfactant concentration, the electrothermal force increases with the concentration while the DEP force takes the opposite trend. When surfactant concentration is 0.01%, the dominant force is the Brownian force for low voltage ($\phi<6$ V), the DEP force for intermediate voltage (6 V$<\phi<19$ V) and the electrothermal force for high voltage ($\phi>19$ V). On the other hand, when the concentration is 0.1%, the local DEP force around the selected location is never dominant for the entire range of $\phi$. This implies that for SDBS concentration higher than 0.1%, the positive DEP force on s-SWNTs plays minor role and the transport is governed mainly by the electrothermal force. The result suggests that there is a crossover between the electrothermal and DEP force with respect to the SDBS concentration. This means that beyond the crossover, an attempt to enhance the DEP-separation efficiency by increasing the surfactant concentration, [*i.e.*]{} by reducing the magnitude of the positive DEP on s-SWNTs may result in enhancing the electrothermal force, which reduces the efficiency. ![Forces change with $\phi$ at ($x,y,z$)=($0~\mu$m, $14~\mu$m, $1~\mu$m). The Brownian force profiles overlap on each other[]{data-label="fig2"}](fig2.jpg){width="8.2cm"} Figure \[fig3\] shows the magnitudes of the dimensionless forces with the vertical coordinate $z$ for the same ($x$, $y$) locations as in Fig. \[fig2\]. Following typical experiments, $\phi$=20 V was applied to 0.1% and 0.01% SDBS solutions. Figure \[fig3\] shows that in 0.1% SDBS, the Brownian motion surpasses the DEP force except very close to the electrode (less than 1 $\mu$m). Here we observe that, in a domain larger than that where the DEP force is dominant, s-SWNTs are carried by the electrothermal flow for both SDBS concentrations. As a result, the electrothermal flow overcomes the Brownian motion in a relatively large domain and transport s-SWNTs onto the electrodes. To demonstrate this, we will verify the direction of the electrothermal flow and the actual traces of s-SWNTs under the action of the resultant force. ![Forces change with z, ($x$, $y$)=(0 $\mu$m, 14 $\mu$m), $\phi$=20 V. The Brownian force profiles overlap on each other[]{data-label="fig3"}](fig3.jpg){width="8.2cm"} Transport simulations of semiconducting SWNTs ============================================= Figures \[fig4a\] and \[fig4b\] show streamlines of the electrothermal flows in 0.1% SDBS solution for two different frequencies. They are cross-sectional views of the 3D simulation results at the plane of $x=0$. The thick line drawn from ($x, y$)=(0 $\mu$m, 5 $\mu$m) to (0 $\mu$m, 25 $\mu$m) marks one of the two electrodes \[Fig. \[fig1a\]\]. The evident dependence of the flow pattern on $\omega$ is highlighted by comparing the flow pattern for $\omega=300$ kHz \[Fig. \[fig4a\]\] corresponding to a typical working frequency with that for $\omega=1 $ GHz \[Fig. \[fig4b\]\] corresponding to the upper limit of frequency explored in experiments. When $\omega=300$ kHz, the vortex core appears close to the gap of the electrodes, whereas when $\omega=1$ GHz, it is located close the domain boundary ($y$=25 $\mu$m). The pattern of the electrothermal flow varies with frequency due to variation in the balance of the Coulomb force and dielectric force. The flow patterns for $\omega=300$ kHz and $\omega=1$ GHz resemble the flow patterns for low and high frequency limits, where the Coulomb force and dielectric force become dominant, respectively. The flow pattern also strongly depends on the geometry of the system. In the current 3D system, independently of frequency, the electrothermal force gives rise to upward flows on the gap of the electrodes and downward flows on domain boundary ($x=0~\mu$m, $y=25~\mu$m). This differs from the 2D case with an infinitesimal gap, where the direction of the flow circulation at low frequency limit is opposite from that at high frequency limit [@ramos; @morgan]. \ After solving Eq. (\[eqn:6\]), we obtain the actual trajectories of the s-SWNTs as plotted in Fig. \[fig4c\] and Fig. \[fig4d\]. The simulation times are 33 ms and 24 ms respectively. The s-SWNTs are seen to follow the streamlines of the electrothermal flow, especially well in the region close to the electrodes. This agrees with the local force comparisons shown in Fig. \[fig3\], where the electrothermal force appeared to be dominant in a broad region close to the electrode. This means that the variation of the flow patterns observed in Fig. \[fig4a\] and Fig. \[fig4b\] results in different spatial trapping distributions of SWNTs. When $\omega=300$ kHz, s-SWNTs are attracted broadly on the electrode, while when $\omega=1$ GHz, the s-SWNTs are mainly attracted to a relatively small region close to the domain boundary. Let us briefly discuss the influence of the electrothermal effect on m-SWNTs. We have also performed the simulations for m-SWNTs with $\sigma_{p}=10^8$ S/m and $\epsilon_{p}=-10^4 \epsilon_0$ [@dimaki]. The magnitude of the positive DEP force for m-SWNTs is much larger than that for s-SWNTs. Consequently, the impact of both electrothermal and Brownian forces on the overall transport is much smaller for m-SWNTs than for s-SWNTs. However, the electrothermal force is far from negligible, especially for higher SDBS concentrations. In fact, in the gap region of the electrode pair where the electrothermal force is directed upwards ($y$=0 $\mathrm\mu$m to 5 $\mathsf\mu$m in Fig. \[fig4\]), the electrothermal force overcomes the DEP force and repels m-SWNTs away from the electrodes when the concentration of SDBS is larger than 0.1%. The impact on the collection of m-SWNTs may be limited since the m-SWNTs will be circulated along the vortex and eventually transported onto the electrodes after a certain time as seen in Fig. \[fig4\] for s-SWNTs. However, the effect should certainly influence the distribution of yielded m-SWNTs on the electrode, which is important for deposition of SWNTs to selected sites for electric device constructions. Conclusions =========== We formulated a dynamical system to simulate the DEP motion of s-SWNTs, which is different from previous simulations mainly by taking the electrothermal flow into account. To realize the DEP separation, the applied electric potential needs to be high enough so that the DEP force on m-SWNTs overcomes the Brownian motion. This results in a high temperature gradient, which creates a substantial electrothermal flow. This electrothermal flow will bring the s-SWNTs close to the electrodes where the weaker DEP force on s-SWNTs can collect them. Thus, the collaborative action of electrothermal flow and DEP restricts the allowable range of potentials where DEP separation may be possible. The higher the concentration of the surfactant is, the stronger the electrothermal flow is. Also, when the frequency increases, the pattern (direction and magnitude) of the electrothermal flow changes. Therefore the electrothermal flow would very likely increase the difficulty both in separating s-SWNTs from m-SWNTs and in depositing SWNTs to a certain position. The main conclusion of this paper is that when designing DEP separation of SWNTs, it is necessary to consider the possible influence of the electrothermal flow. The Swedish Research Council (VR) is greatly acknowledged for financial support. The authors thank Z.-B. Zhang and S.-L. Zhang for stimulating discussions.

ArXiv

--- abstract: 'Nuclear inflows of metal–poor interstellar gas triggered by galaxy interactions can account for the systematically lower central oxygen abundances observed in local interacting galaxies. Here, we investigate the metallicity evolution of a large set of simulations of colliding galaxies. Our models include cooling, star formation, feedback, and a new stochastic method for tracking the mass recycled back to the interstellar medium from stellar winds and supernovae. We study the influence of merger–induced inflows, enrichment, gas consumption, and galactic winds in determining the nuclear metallicity. The central metallicity is primarily a competition between the inflow of low–metallicity gas and enrichment from star formation. An average depression in the nuclear metallicity of $\sim0.07$ is found for gas–poor disk–disk interactions. Gas–rich disk–disk interactions, on the other hand, typically have an enhancement in the central metallicity that is positively correlated with the gas content. The simulations fare reasonably well when compared to the observed mass–metallicity and separation–metallicity relationships, but further study is warranted.' author: - Paul Torrey - 'T. J. Cox' - Lisa Kewley - Lars Hernquist bibliography: - 'NMpaper.bib' title: The Metallicity Evolution of Interacting Galaxies --- Introduction ============ The nuclear metallicities of star-forming galaxies are characterized by a mass-metallicity relation [hereafter, MZ @Lequeux1979; @Rubin1984; @Tremonti2004]. Contributing to the scatter in the MZ relation are interacting galaxies, which are consistently lower in central metallicity than non–interacting systems of equivalent mass, as first noted by @KGB06, and later confirmed for ultraluminous infrared galaxies (ULIRGs) [@Rupke2008], close pairs in SDSS [@SloanClosePairs; @MichelDansac2008; @Peeples2009; @SolA2010], and local, low mass systems [@EktaChengalur2010]. Observations such as these are most naturally explained as the result of vigorous merger–induced inflows of gas [@BH91; @BH96] which rearrange the initial metallicity gradient [@Shields1990; @BelleyRoy1992; @Z94; @MA04] and “dilute” the nuclear metallicity [@KGB06]. In this sense, the same gas that drives nuclear starbursts [@MH94a; @MH96; @Iono04] and triggers central AGN activity and black hole growth [@DSH05; @SDMHModel; @HH06; @Hopkins07], also results in suppressed nuclear metallicity. Simulations of merger-driven inflows of gas naturally predict a flattening of the initial metallicity gradient [@Perez2006; @Rupke2010; @PerezScoop], which has now been observed in a number of colliding systems [@KewleyGradients2010; @RupkeGradients2010]. However, more than hydrodynamics are at play in determining the nuclear metallicity evolution. To further our understanding, we should consider ongoing star formation with associated chemical enrichment, feedback from star formation and AGN activity, and the interchange of material between the stellar and gaseous phases as stars are born and later return material to the interstellar medium. The four main processes responsible for the evolution of the nuclear metallicity are gas inflows, chemical enrichment from star formation, gas consumption, and galactic outflows. These effects compete with one another to influence the nuclear metallicity, making it difficult to determine their relative importance a priori. Numerical simulations have only recently been used to quantify the detailed impact of these effects on metallicity gradients and nuclear metallicities of interacting pairs. @Rupke2010 simulated colliding galaxies without chemical enrichment to explore dynamically induced changes in metallicity gradients, finding that a drastic flattening can occur, accompanied by a drop in the nuclear metallicity. @DilutionPeak2010 and @PerezScoop performed simulations with star formation and chemical enrichment and found similar results to @Rupke2010, but noted that the dip in the nuclear metallicity values can be counteracted by chemical enrichment from star formation. These simulations have refined our knowledge of the depressed MZ relation. In this paper, we explore the evolution of the nuclear metallicity during mergers using numerical simulations which include cooling, star formation, stellar feedback, and black hole growth and AGN feedback. Our approach allows us to systematically investigate the importance of gas inflows, chemical enrichment from star formation, gas consumption as a result of star formation, and galactic outflows. In order to unambiguously determine the role of metal enrichment, we have developed a stochastic method to recycle stellar particles back to the interstellar medium without requiring inter-particle mass mixing. The stochastic gas recycling method, which is designed by analogy to the widely used stochastic star formation method, has attributes that are distinct from kernel weighted mass return, making it particularly useful for our study. We consider a range of situations with systematically varied parameters to enhance our understanding of metallicity evolution as a natural extension of the merger process. We find that, for systems modeled after those in the local Universe, the main driver behind changes to the nuclear metallicity is the flow of low metallicity gas to the nuclear regions, or metallicity dilution. However, we identify two previously under-appreciated effects that influence the strength of this dilution: locking of gas and metals in the stellar phase, and gas and metal removal via stellar-driven winds. We also find that, for systems modeled after high redshift galaxies, the main driver of the nuclear metallicity shifts to chemical enrichment. In these simulations, the nuclear metallicity increases, contrary to what is observed in the local Universe, suggesting that the interacting galaxy MZ relation may evolve differently at high redshifts. We compare our simulations directly to observations by synthesizing a population of progenitor galaxies with properties consistent with observed samples and show that the empirical depression in the close-pair mass-metallicity relation can be reproduced while accounting for chemical enrichment. We find good agreement between our simulations and observations for both the interacting galaxy mass-metallicity relation and separation-metallicity relation. Methods {#sec:MergerSimulations} ======= We employ a library of merger simulations carried out using the N-body/Smooth Particle Hydrodynamics (SPH) code [Gadget-2]{} [@Gadget2]. [Gadget-2]{} is based on a formulation of SPH [@SHEntropy] which conserves both energy and entropy simultaneously (when appropriate). In addition to accounting for gravity and hydrodynamics, our simulations also include a sub-grid two-phase model of star formation, supernova feedback, radiative cooling of gas, and star formation driven winds [@KatzCooling; @SHMultiPhase hereafter SH03]. These features give a working description of the multiphase nature of the interstellar medium [e.g. @MO77] without explicitly resolving the various phases. Supermassive black hole sink particles are included in our simulations and have both accretion and thermal feedback associated with them, but their presence does not influence any results presented in this paper. The simulations presented here also include a novel method for returning gas to the interstellar medium (ISM) from (prompt) supernovae and asymptotic giant branch (AGB) winds. While the main body of our simulations code has been used for numerous other studies [see, e.g., @Cox2006b; @Cox2006; @Robert06; @Robert06a; @Robert06b], the inclusion of time–delayed gas recycling is a new feature which is described. The Basic Model: SH03 {#sec:SH03outline} --------------------- The primary astrophysical processes modeled in our simulations – including cooling, star formation, feedback, and galactic winds – are included as in the SH03 model. Because these processes are discussed extensively in that paper we review them only briefly here. The instantaneous rate of star formation for an SPH particle is given by $$\frac{dM_{*}}{dt} = \frac{M _{c}}{t_{*}} \label{eqn:sfr}$$ where $M _{*}$ is the stellar mass, $M _{c}$ is the mass of cold gas (i.e. some fraction of the SPH particle’s mass that is in the cold phase as determined via the sub-grid model prescription of SH03), and $t_{*}$ is a characteristic star formation timescale. We adopt a value $t_* = 4.5$ Gyrs so that our simulations are consistent with the Kennicutt-Schmidt relation [@KS1; @KS2; @Cox2006]. Star formation is assumed to only take place in gas which has densities above a set threshold, which in our case is 0.5 cm$^{-3}$. Equation \[eqn:sfr\] is used as the basis for a Monte Carlo method for actually converting SPH particles to star particles. Each SPH particle is assigned a probability of turning into a star particle as follows: $$\label{eqn:Psfr} p_* = \left[ 1 - \exp \left( - \frac{\Delta t}{t _{\mathrm{SFR}}} \right) \right]$$ where $\Delta t$ is the current simulation time step, and $t_{\mathrm{SFR}} = M_{gas} / (dM_* / dt)$ is the star formation timescale. A particle is converted if a random number drawn between 0 and 1 is less than Eq. \[eqn:Psfr\]. The instantaneous rate of mass launched in star formation driven winds is given by $$\frac{dM_{w}}{dt} = \gamma \frac{dM _{*}}{dt} \label{eqn:winds}$$ where $dM _{*}/dt$ is the local star formation rate and $\gamma$ is the mass entrainment efficiency. We typically take $\gamma=0.3$, which is the mean value for luminous infrared galaxies measured by @RupkeOutflows2005, and the resulting winds are ejected at $v_{\rm w}= 242$ km/sec. Equation (\[eqn:winds\]) is used as the basis for a Monte Carlo method for wind generation. The associated probability of launching a particle in a wind is given by $$p_{w} = \left[ 1 - \exp \left( - \frac{\Delta t}{t_{w}} \right) \right]$$ where $t_w = M_{gas} / (dM_w /dt)$. Particles put into winds receive a velocity kick and are not allowed to interact with the surrounding medium while they leave their host environment. In principle, gas particles launched by winds may still contribute to the nuclear gas phase metallicity. However, in practice, these particles very quickly leave the nuclear regions, and no longer contribute to the nuclear metallicity. Gas Recycling {#sec:GasRecycling} ------------- While models for star formation and feedback are ubiquitous in the current generation of numerical simulations, it is much less common to include the recycling of mass and metals which are lost from evolving stellar populations, a fundamental process in the enrichment of the interstellar and intergalactic mediums. The few models that do include this “gas recycling” typically do so in a continuous fashion, during which, at every time step, a small fraction of mass and metals is transferred from a single stellar particle to a set of nearest SPH neighbors [e.g. @SteinmetzGasReturn; @KawataGasReturn; @KawataGibsonGasReturn; @OkamotoSGR; @ScannapiecoGasReturn; @StinsonGasReturn; @TornatoreGasReturn; @WiersmaGasReturn; @DilutionPeak2010; @PerezScoop]. This approach can accurately track the spatial and temporal distribution of mass and metals at a reasonable computational cost, but unfortunately, because particles can exchange mass with many neighbors, at many different times, it is impossible to trace the exact origin of any particular mass or metal element. Given a desire to unambiguously track the origin and evolution of interstellar metals, the work presented here adopts a stochastic approach to gas recycling [e.g. @LiaSGR; @MartinezSerranoSGR]. This less-used approach is designed to mirror the stochastic star formation scheme that is typically implemented in numerical simulations. While stochastic star formation probabilistically converts SPH particles into a fixed number of collisionless stellar particles based upon the star-formation rate, our stochastic gas recycling converts stellar particles into SPH particles based upon theoretical mass return rates for a simple stellar population. When used in conjunction, these methods can accurately track the temporal and spatial interchange of baryonic material between stellar populations and the interstellar medium, a necessary element to any study of the evolution of interstellar metals. In our stochastic approach, the probability of turning an individual star particle into an SPH particle at any given time step is given by $$p_{recy} = \left[ 1 - \exp \left( - \frac{\Delta t}{t_{recy}} \right) \right] \label{eq:pgas}$$ where $\Delta t$ is the current simulation time step, and $$t_{recy} = \frac{ M_* (t) }{ dM_{gas}/dt }. \label{eq:tgas}$$ Here, $M_*(t) $ is the expected stellar mass at time $t$, and $dM_{gas}/dt$ is the mass return rate at the same time. While the physical mass of the stellar particle in the simulation remains constant, the expected stellar mass identifies the amount of mass that would have been returned to the ISM if we implemented continuous gas recycling. We should make clear that there are no hybrid particles in the simulation, so the expected stellar mass has no influence on the simulations dynamics. It is necessary, however, for Eq. \[eq:tgas\] to use the expected stellar mass in order for the stochastic recycling routine to reproduce the desired Starburst99 gas recycling rate in the simulations. In practice, the simulation selects a random number between 0 and 1, and performs the particle conversion if this number is less than $p_{recy}$. We calculate the mass return rate, $dM_{gas}/dt$, directly from Starburst99 synthetic population models [@Starburst99_1; @Starburst99_2; @Starburst99_3]. These models include gas recycling contributions from core collapse supernova and asymptotic giant branch (AGB) winds, but do not include contributions from Type Ia supernovae. At each time step, stellar particles are assigned a mass return rate determined by their age, metallicity, and an initial mass function ($\Upsilon$). For the simulations presented here, we choose a Kroupa initial mass function with solar metallicity. The resulting mass return rate, and mass return fraction are demonstrated in Figure \[fig:massReturnRate\]. While it will be fruitful to explore variations in the mass return according to varying IMF assumptions, real–time metallicity information, or even easily parameterized functional forms of mass return that include Type Ia supernovae  [see, e.g., @JungwiertGasReturn; @Kravtsov] we choose to allow variations in mass return through a simple scale factor to the return rate shown in Figure \[fig:massReturnRate\]. Specifically, we define a scalar, $\zeta$, and employ a mass return rate normalized as such $$\frac{dM_{gas}}{dt} = \left( \frac{ \zeta }{\frac{1}{M_*}\int _0 ^\infty f_{sb99}(t , Z, \Upsilon) dt } \right) f_{sb99}(t , Z, \Upsilon), \label{eqn:gas_recycling_rate}$$ where $f_{sb99}(t , Z, \Upsilon)$ is the mass return rate from Starburst99 as shown in Figure \[fig:massReturnRate\]. The expected stellar mass in Eq. \[eq:tgas\] is then calculated as $$M_* (t) = M_0 - \int_0 ^{t } \frac{dM_{gas}}{dt'} dt',$$ where $M_0$ is the original mass of the stellar particle. We note that for large values of $\zeta$, and for old stellar ages, $M_*(t)$ can become negative, yielding unphysical negative values for $t_{recy}$ and consequently a negative probability $p_{gas}$. This situation is non-catastrophic, however, because, in practice, it leads to an immediate conversion of this stellar particle to SPH. The parameter $\zeta$ allows us to control the timescale and efficiency of gas recycling. For example, the Kroupa IMF we employ dictates that 37% of the stellar mass will be returned to the ISM after 10 Gyr. Setting $\zeta=0.37$ makes the normalization unity and the recycling therefore tracks exactly what is shown in Fig. \[fig:massReturnRate\]. Setting $\zeta=1$ will produce complete (i.e., 100%) recycling of the stellar mass within 10 Gyr, while $\zeta=100$ effectively yields the instantaneous recycling approximation with all the stellar mass being returned within $\sim3\times10^6$ Yr. Finally, and perhaps obviously, setting $\zeta=0$ turns off recycling altogether giving us the flexibility to study the impacts of varying amounts of recycling on metallicity evolution. When a star particle is returned to the gas phase, the particle type is instantly converted from stellar to gas at the end of the current time step. The fields which are defined for both SPH particles and star particles, such as the position, velocity, mass, metallicity, etc., remain unchanged during the conversion. For concreteness, the metallicity of a particle remains unchanged during the transition to or from the stellar state. The metallicity increases while in the gas phase (according to Eq \[eqn:MetalEnrichment\]), but remains fixed while in – or making transitions to and from – the stellar state. SPH quantities must be initialized for the newly converted gas particle. In principle, we could set these fields to properly reflect the physical state of the gas being returned and explore the feedback implications that naturally result from this gas recycling model. However, in practice, we are resolution-limited, we already include feedback according to the SH03 model, and our primary concern is representing the overall mass-budget faithfully. To this end, we give a newly formed SPH particle properties that will allow it to quickly homogenize into the surrounding medium. In particular, the entropy is set equal to that of gas at a temperature of $T = 50,000$ K and a density (chosen to be the star formation critical density) of $\rho = 0.5 \;\mathrm{cm ^{-3}}$. It is important to note that our simulations are nearly invariant to our choice of these values. Changing the initial entropy by an order of magnitude in either direction yields no obvious differences in the simulations. In effect, this implementation of gas return serves to provide a passive source of gas for our evolving system, without providing strong feedback. During both star formation and gas recycling events, no particle splitting occurs. As such, the total number of baryon particles (stars and gas) is conserved throughout the simulation. Effectively, baryon particles are permitted to flip back and forth between the gas and stellar phases, according to the star formation and gas recycling rates. Because baryon particle number is conserved and no mass is transferred between particles, mapping a particle’s initial position to a final position a trivial task, regardless of the number of times it was it was turned into a star particle or returned to the gas phase. Using this stochastic gas recycling method provides the distinct advantage that we can retrace all mass to a unique initial position regardless of the star formation or gas recycling history. Metallicity Enrichment {#sec:MetallicityEnrichment} ---------------------- We implement a method for calculating the metal enrichment based on the instantaneous star formation rate to determine the metallicity enrichment rate. Unlike star formation, galactic winds, and gas return, metal enrichment is carried out in a continuous fashion, where the metal formation rate is given by $$\frac{dM_{Z}}{dt} = y \frac{dM_{*}}{dt} \label{eqn:MetalEnrichment}$$ and $y$ is the metal yield and $dM_{*}/dt$ is the instantaneous star formation rate. For our simulations, we use a fixed yield of $y=0.02$. The metallicity will increase wherever there is ongoing star formation, as determined via the SH03 star formation model. As such, diffuse gas will not be star forming, and will therefore not enrich (see SH03 for details). The scalar metallicity is updated at each time step using an Eulerian integration scheme. This scalar metallicity does not track independent species, but instead provides a single metallicity value proportional to the integrated star formation rate. This metal enrichment scheme is strictly independent of our gas recycling routine. The metallicity enrichment does not affect the evolution of the simulation because there are no metallicity dependent dynamical processes included, such as metal line cooling. Because no mixing is allowed, nearby particles may have large metallicity variations. This is not problematic, as long as we interpret the metallicity at a given location to be a kernel weighted average of nearby particles, as is traditional for determining fluid quantities in SPH. Mixing would homogenize the individual particle metallicities, while leaving the average value of the kernel weighted metallicity unchanged. Using this enrichment scheme, we are able to cleanly decompose the metallicity of an SPH particle into contributions from its initial metallicity and star formation history. More important, we are able to scale our metal yield and modify our initial metallicity setup in our post-processing analysis, without requiring additional simulations. This is possible because all of the gas content of an SPH particle has an unambiguous and unique initial location and star formation history, which is not true when particle mass mixing is used. In later sections, our ability to arbitrarily modify initial metallicity gradients of our progenitor galaxies is critical in allowing us to thoroughly sample the range of metallicity gradients of our progenitor galaxies, without large computational requirements. Isolated Galaxies {#sec:IsolatedGalaxyMetallicityEvolution} ================= In our analysis, we take the nuclear region of a galaxy to be a sphere of radius 1 kpc about the galactic center. We determine the nuclear metallicity from the star formation rate weighted average of all gas particles inside this sphere. A 1 kpc spherical region is used for consistency with the observations of @KGB06. However, it should also be noted that our results would not fundamentally change for slightly larger or smaller definitions of the nuclear region. We use star formation rate weighted averages to mimic observations of HII regions, which naturally select star forming gas. Unless otherwise stated, all nuclear metallicities quoted in this paper are gas-phase and star formation rate weighted. In particular, the central depressions in gas-phase metallicity presented here should not be confused with enhancements in stellar metallicity seen in late-stage mergers and relaxed elliptical galaxies. These increases in stellar metallicity are relics from the merger-driven starbursts that leave behind central stellar cusps in merger remnants [@MH94b; @Hopkins08a] and elliptical galaxies [@Hopkins09a; @Hopkins09b]. These starbursts occur at late stages in a merger, following coalescence, from enriched, star-forming gas, yielding a central population of young, metal-enhanced stars [see, e.g. Figs. 27 and 28 in @Hopkins09a]. The star formation rate weighted nuclear metallicity is defined as $$\bar{Z} = \frac{\int _V \dot \rho_* \left( \vec r \right) Z \left( \vec r \right) dV}{\int_V \dot \rho_* \left( \vec r \right) dV }$$ where $\dot \rho_* \left( x,y,z \right)$ is the star formation rate density, $Z \left(x,y,z \right)$ is the fraction of gas-mass in metals, and the integral is performed over the 1 kpc spherical region about the galaxy’s center. The star formation rate density is calculated as $$\dot \rho_* \left( \vec r \right) = \sum _j \frac{dM_{*,j}}{dt} W \left( \left| \vec r - \vec r _j \right| , h_j \right)$$ where $W \left( \left| \vec r - \vec r _j \right| , h_j \right)$ is a smoothing kernel function, $h$ is a smoothing length, and $dM_{*}/dt$ is a particle’s instantaneous star formation rate (discussed in §\[sec:SH03outline\]). For our analysis, we use a galaxy’s central black hole to define the galactic center; however, our results are unchanged if we use the potential minimum, or any other reasonable measure of the galactic center. This approach neglects projection effects, but instead provides information about the “true” nuclear region. The metallicity of each SPH particle consists of two separate contributions: an initial metallicity and an enriched metallicity. The initial metallicity of an SPH particle is completely determined by the particle’s initial radial position (discussed in §\[sec:GalaxySetup\]) and is unchanging in time, while the enriched metallicity grows owing to star formation. The metallicity of an SPH particle as a function of time is given by $$\label{eqn:z_t} Z_{gas}\left(t \right) = Z_{init} + \frac{y}{M_{gas}} \int_{0} ^{t} \frac{d M_{*}(t')}{dt'} dt'$$ where $d M_{*} (t')/dt' $ is the particle’s complete time dependent star formation rate history. The integration of the rightmost term in Eq. \[eqn:z\_t\] should be carried out starting at the beginning of the simulation (i.e. $t=0$), regardless of whether a particle has been recycled or not. Since $dM_{*} (t') /dt'$ is non-zero only when a particle is actively star forming (i.e. in the gas state and above the star formation threshold density), periods of time when a particle is in the stellar state or below the star formation threshold density (see §\[sec:SH03outline\]) will not contribute to Eq. \[eqn:z\_t\]. The metallicity of the nuclear region is calculated similarly to the metallicity of an individual SPH particle, with an additional sum over an ensemble of particles weighted by their star formation rates. We explicitly break the nuclear metallicity into two components $$\label{eqn:z_components} \bar Z = \frac{\sum_{i} Z_{i,init} \dot M_{*,i}}{\sum_{i} \dot M_{*,i}} + \frac{\sum_i Z_{i,enrich} \dot M_{*,i}}{\sum_i \dot M_{*,i}}$$ where the sum is performed over the particles that fall within the nuclear region. The first term contains the contribution from the initial metallicity gradient of our progenitor galaxies while the second term contains the contribution from star formation induced metallicity enrichment. For clarity throughout, we call the first term the “dynamical metallicity”, as it is a dynamical result of the initially assumed metallicity gradient, while we call the second term the “enriched metallicity”, as it is a product of metal enrichment. For concreteness, the metallicity gradients for one of our “C” isolated galaxies evolved in isolation for 2 Gyrs is shown in Figure \[fig:isoGal\] with contributions from the dynamical and enriched components explicitly shown (see §\[sec:IsoGalEvo\]). ----------------- ----------------------- ----------------------- -------------- --------- ------------ ----------- ------------- Disk Identifier Total Halo Initial Stellar Initial Disk $h$ $N_{Halo}$ $N_{gas}$ $N_{stars}$ Mass $ [M_{\odot}]$ Mass $[M_{\odot}]$ Gas Fraction \[kpc\] A $2.3 \times 10^{11}$ $1.1 \times 10^{10}$ $10\%$ 2.4 532,500 30,000 145,500 B $5.1 \times 10^{11}$ $2.4 \times 10^{10}$ $10\%$ 3.2 532,500 30,000 145,500 C $9.5 \times 10^{11}$ $4.4 \times 10^{10}$ $10\%$ 3.9 532,500 30,000 145,500 D $13.5 \times 10^{11}$ $6.2 \times 10^{10}$ $10\%$ 4.4 532,500 30,000 145,500 C2 $9.5 \times 10^{11}$ $4.0 \times 10^{10}$ $20\%$ 3.9 532,500 40,000 135,500 C3 $9.5 \times 10^{11}$ $ 3.2 \times 10^{10}$ $ 40\% $ 3.9 532,500 60,000 115,500 C4 $9.5 \times 10^{11}$ $2.4 \times 10^{10}$ $60\%$ 3.9 532,500 80,000 95,500 ----------------- ----------------------- ----------------------- -------------- --------- ------------ ----------- ------------- Isolated Galaxy Setup {#sec:GalaxySetup} --------------------- The isolated galaxies used in this paper are modeled following the analytical work of @MMWdisk employing the procedure outlined in @SDMHModel. Our fiducial galaxy consists of a dark matter halo, an embedded rotationally supported exponential disk, a stellar bulge, and a central supermassive black hole. We construct a set of four isolated galaxies with total system masses ranging from $\sim 10^{11} M_{\odot}$ to $\sim 10^{12} M_{\odot}$, as outlined in Table \[MergerProperties\]. Although varied in mass, all systems are constructed to be self-similar in order to isolate the effects of mass from other quantities that may correlate with mass in observed systems. The total mass of the disk (stars and gas combined) is chosen to be a constant fraction (4.0%) of the halo mass for all progenitor galaxies, with four settings for the initial gas fraction (8%, 20%, 40%, and 60%). All halos have spin parameters equal to $\lambda =0.05$, which effectively sets the disk radial scale length. The initial stellar disk scale height is set to a fixed fraction (0.2) of the initial disk scale length, while the gaseous scale height is determined by satisfying hydrostatic equilibrium [@SDMHModel]. Stellar bulges are included as part of the fiducial galaxy setup and are given a fixed fraction (1%) of the system mass. The resulting disk setup has a Toomre Q parameter that varies as a function of galactocentric radius, but is everywhere greater than unity – indicating stability against axisymmetric perturbations. The dark matter halos follow  @H90 profiles with concentration indices of $c=10$. The stellar and gaseous disk components are modeled with exponential surface density profiles $$\Sigma _{*,g} \left( r\right) = \frac{M_{*,g}}{2 \pi h^2} \exp \left( -\frac{r}{h} \right) \label{eqn:sigma}$$ where $M_{*,g}$ is the total mass of stars and gas in the disk, respectively, and $h$ is their common scale length set by the disk’s angular momentum  [@MMWdisk; @SpringelWhiteDisk]. Stellar bulges are taken to be spherical  @H90 profiles, where the bulge scale length is given in terms of the halo scale length, $a_b = 0.2 a$, and the bulge mass fraction to be 1% of the halo mass. The initial metallicity profile of our progenitor galaxies enters as an assumption. We model the metallicity profile of the disk as an exponential $$Z \left( r \right) = Z_0 \; \exp \left( - \frac{r }{ h_z } \right) \label{eqn:initZ}$$ where the central metallicity, $Z_0$, and the metallicity scale length, $h_z$, are chosen to be consistent with observations. Isolated galaxies show dispersion in their radial metallicity gradients. We could include this dispersion in our simulations by adding in a Gaussian error term to Eq. \[eqn:initZ\]; However, no dispersion is used when setting the initial metallicity profile for our isolated galaxies because this would add an extra element of uncertainty into our models and add noise to our results, without yielding systematic changes or further insight. We assume that Oxygen makes up 30% of the metals by mass. The constant Oxygen mass fraction is a necessary assumption of our models, because we track only one scalar metallicity value. We note that our results are not particularly sensitive to our choice of the Oxygen mass fraction, since this is merely a normalization of the overall metallicity. We fix the central metallicity values to the observed relation of @Tremonti2004 so that the central metallicity is given in terms of the stellar mass $$12+\mathrm{log} \left( O/H\right) = -1.492 + 1.847 x - 0.08026 x^2 \label{eqn:MZTremonti}$$ where $x=\mathrm{log}_{10} M_*$. Initially, we use the exact value given by equation (\[eqn:MZTremonti\]) when discussing the process of nuclear metallicity evolution in §\[sec:MergerSetup\]. However, in §\[sec:ComparisonToObservations\] we pick the central metallicity of each progenitor galaxy from a Gaussian distribution with a standard deviation, $ \sigma = 0.1$ dex, taken from @Tremonti2004. Similarly, we use the average result from @Z94 of $h_z = h/0.2$ in §\[sec:MergerSetup\]. However, in §\[sec:ComparisonToObservations\] we pick the slope of the metallicity gradient from a Gaussian distribution with a standard deviation of $\sigma = 0.3 h_z$. Isolated Galaxy Evolution {#sec:IsoGalEvo} ------------------------- While the metallicity evolution of isolated galaxies is not the focal point of this paper, we take a moment in this section to demonstrate that our galaxy models are indeed stable, and evolve very little in isolation. We specifically demonstrate that the galaxy models do not develop strong bars since these structures can lead to significant radial inflows of gas. Such processes obscure the true origin of any central metallicity evolution, and what is specifically a result of the merger. {height="2.0in"} The evolved gas surface density, seen in Figure \[fig:gasPanel\], demonstrates that our disks are both stable and free of strong bars over at least 2 Gyrs, the typical duration of the galaxy major mergers we study. As a result, the metallicity gradients of our isolated systems evolve very little, as seen in Figure \[fig:isoGal\]. We further demonstrate the characteristics of our isolated disks in Figure \[fig:IsolatedDisksExample\] where we show the star formation rates and gas fractions for the C2, C3, and C4 systems. To understand where the small amount of metallicity evolution does come from, Figure \[fig:isoGal\] also shows the evolution of the initial metallicity gradient, or the “dynamical” metallicity (middle panel), and the contribution from ongoing star formation, or the “enriched” metallicity (right panel). From this information it is easy to see that the small increase in central metallicity and gradient is a result of chemical enrichment from ongoing star formation, while the initial metallicity gradient remains nearly unchanged. Motivation to emphasize the stability of our initial disk models stems from the fact that the disks used in previous studies [e.g. @DilutionPeak2010] appear to be bar unstable over short timescales (see their Figure 2 in their Appendix). These bars result in strong radial mixing and a flattening of their radial metallicity profiles (see Figure 1 in their Appendix). While bar induced gas inflows are indeed a valid physical mechanisms for modifying the nuclear metallicity, we wish to eliminate this complication from our study to make our results easier to interpret. The evolution of the nuclear metallicity of the four progenitor disks, described in Table \[MergerProperties\], are demonstrated in Figure \[fig:IsolatedDisksStats\]. The nuclear metallicity of all isolated disks is well described by a mild and monotonic increase over at least 2 Gyrs. The steady behavior of the isolated galaxy nuclear metallicity ensures that any strong changes during the merger simulations are a product of the merger process, rather than disk instability. While exploring the behavior of isolated disks, we can also demonstrate the effects and capabilities of stochastic gas recycling. When evolving the “C” systems (i.e. C2, C3, and C4 from Table \[MergerProperties\]) we vary the gas recycling parameter and examine the resulting gas fraction and star formation rate evolution. The results, shown in Figure \[fig:IsolatedDisksExample\], demonstrate that the gas is locked into stars more efficiently when the gas recycling parameter is small. As a result, the star formation rates, which largely depend on the amount of gas available for star formation, are consistently larger for increasing values of the gas recycling parameter. Simulations that use efficient gas recycling maintain larger gas fractions and star formation rates throughout the simulation. In the limit of very large gas recycling parameters, all newly formed stellar material is quickly returned to the ISM. As a result, the gas fraction for large gas recycling parameters stays nearly constant. The corresponding star formation rates slowly increase as gas naturally cools and condenses over time. Merging Galaxies {#sec:MergerSetup} ================ While keeping all of the conventions for nuclear metallicity defined in §\[sec:IsolatedGalaxyMetallicityEvolution\], we now consider the evolution of galactic nuclear metallicity for merging systems. Since our interacting systems are spatially extended and overlap substantially during the interaction, they do not follow Keplerian trajectories. However, we use the terminology of Keplerian orbits to clearly describe the merger setup. To completely specify our merging setup, we must specify the properties of the merger orbit (2 parameters), and the relative orientation of each galaxy with respect to the orbit (2 parameter for each galaxy). We start by specifying that the systems will be on zero energy orbits (i.e. eccentricity value of unity for Keplerian objects). The orbital angular momentum is then set by picking a value for the impact parameter, assuming Keplerian trajectories (the “real” impact parameter found in the simulations will be larger than this value). We hold the impact parameter fixed (5 kpc) for all simulations. The spin angular momentum vector of each galaxy is varied with respect to the orbital angular momentum vector according to the orientations detailed in Table \[OrbitalOrientations\] as visually depicted in Figure \[fig:diskOrientation\]. The parameter space of mergers is quite large. To explore this parameter space, without using an excessively large number of simulations, we first perform a detailed exploration of a single case. We then systematically vary merger parameters, such as the disk orientation, progenitor disk mass, merger mass ratio, orbital angular momentum, and so forth. While this does not necessarily cover all of merger parameter space, it does enable us to paint a coherent picture detailing the relation between the evolution of the merger and the evolution of the nuclear metallicity. We select our fiducial galaxy merger to be two identical “B” disks merging on the “e” trajectory (see Tables \[MergerProperties\] and \[OrbitalOrientations\]). This setup is neither average nor special, and is simply selected to demonstrate in detail the relationship between the merger state and evolution of the nuclear metallicity. While §\[sec:CaseStudy\] is restricted to understanding a single merger evolution, §\[sec:MergerOrientation\]-\[sec:MergerGasFraction\] show that much of the nuclear metallicity evolution can be understood in terms of generic merger properties that will continue to drive the nuclear metallicity evolution as we move beyond this fiducial setup. ------------- ---------------- ------------- --------------- ------------- Orientation $\theta _ {1}$ $\phi _{1}$ $\theta _{2}$ $\phi _{2}$ Identifier \[deg\] \[deg\] \[deg\] \[deg\] a 90 90 0 0 b 180 0 0 0 c 180 0 180 0 d 90 0 0 0 e 30 60 -30 45 f 60 60 150 0 g 150 0 -30 45 h 0 0 0 0 i 0 0 71 -30 j -109 90 71 90 k -109 -30 71 -30 l -109 30 180 0 m 0 0 71 90 n -109 -30 71 30 o -109 30 71 -30 p -109 90 180 0 ------------- ---------------- ------------- --------------- ------------- : Orbital Orientations[]{data-label="OrbitalOrientations"} ![Schematic representation of the merger setup. Our coordinate system is defined by the plane of the merger orbit, and each disk’s orientation is independently adjusted with respect to this plane.[]{data-label="fig:diskOrientation"}](DiskOrientation.eps){height="2.7in"} Metallicity Evolution in Merging Systems {#sec:CaseStudy} ---------------------------------------- In stark contrast to the isolated systems presented in §\[sec:IsoGalEvo\], the nuclear metallicity of interacting galaxies is a complicated function of time, as shown in Figure \[fig:ZTcasestudy\]. Moreover, there is a clear change in the evolutionary behavior of the nuclear metallicity following pericenter passage when the disks stop behaving like isolated systems, and enter a period of evolution dominated by the merger dynamics. Before delving further into the merger details, we note that there is a characteristic “double dip” shape to the nuclear metallicity evolution. This is caused by the ongoing competition between metallicity dilution and chemical enrichment. While the exact evolutionary track shown here is specific to this particular merger setup, the characteristic double dip in the metallicity will appear repeatedly in many other cases. In general, galactic close passages lead to gas inflows (yielding inflows of low metallicity gas) which are quickly followed by star formation (yielding chemical enrichment). During the period of time between first pericenter passage and final coalescence, the nuclear metallicity evolution is driven by: 1) Radial inflows of low metallicity gas caused by the tidal interaction, 2) Chemical enrichment caused by star formation activity, 3) Locking of metals and gas in the stellar phase, and 4) Removal of gas and metals via galactic winds. The combined effect of these four mechanisms, which are discussed in detail in the following sub-sections, determines the relative importance of metallicity dilution and chemical enrichment, which ultimately decide when and where the nuclear metallicity will be diluted or enhanced. ### Metallicity Dilution and Chemical Enrichment {#sec:DilutionAndEnrichment} The influence of radial gas inflows and chemical enrichment are demonstrated in Figure \[fig:ZTCenterPiece\], which shows the galactic separation, rate of change of nuclear metallicity, nuclear star formation rate, and nuclear gas inflow rate. In terms of basic merger properties, there is an influx of gas following first pericenter passage and periods of gas inflow associated with each additional close passage. These gas inflows give rise to high nuclear star formation rates. While these previous points have been studied extensively in other papers [@BH91; @BH96; @MH94a; @MH96; @Iono04], we instead focus here on the influence that these generic merger properties have on the evolution of the nuclear metallicity. Specifically, times of strong gas inflow correspond to periods of nuclear metallicity depression, while high star formation activity aligns with nuclear metallicity enhancement. These qualitative relationships remain true as the merger parameters are varied. Previous studies have found that the depression in the nuclear metallicity is correlated with the mass of gas that migrated to the nuclear region [@Rupke2010]. This result is reproduced in our simulations when we neglect star formation (similar to the red line in Figure \[fig:ZTcasestudy\]) and is a clear-cut demonstration of metallicity dilution. However, when we also consider contributions from chemical enrichment and the consumption of gas via star formation, this correlation disappears. Inspection of Figure \[fig:ZTCenterPiece\] shows that, in general, periods of strong gas inflow occur simultaneously with times of ongoing nuclear metallicity depression. Similarly, periods of high star formation give rise to enhancements in the nuclear metallicity. Interestingly, neither the nuclear star formation rate nor the nuclear mass inflow rate correlate well with the rate of change of the nuclear metallicity because they exchange roles playing the dominate driver of the nuclear metallicity throughout the merger. The reason is that metallicity dilution, chemical enrichment, locking of gas in the stellar phase, and galactic winds all influence metallicity evolution. While more involved parameters (e.g. the difference between the star formation and nuclear mass inflow rate) yield better, albeit imperfect, correlations, these trends can be misleading and they tend to over-complicate a fairly simple point. In particular, changes to the nuclear metallicity at any time during the interaction can be understood by examining the nuclear star formation and gas inflow rates. Furthermore, the nuclear star formation and gas inflow rates are naturally explained via the well studied merger process. From this single merger example, it seems plausible that the role of metallicity dilution and chemical enrichment are natural consequences of the merger process. However, by inspection of Figure \[fig:ZTcasestudy\], neither effect is overwhelmingly dominant. Hence, in the following analysis we pay attention not only to metallicity dilution and chemical enrichment, but also effects such as the locking of gas in stars and the launching of galactic winds that work to modulate the efficiency of these processes. Instead of searching for correlations between the changes in the metallicity and these distinct processes, we vary parameters in our simulations to test their influence on the evolution of the nuclear metallicity. ### Gas Consumption {#sec:GasConsumption} Any process that can modify the gas reservoir in the nuclear region can affect the metallicity measurement. Here, we consider the influence of locking gas in the stellar phase when determining the evolution of the nuclear metallicity. Star formation lowers the reservoir of gas in the nuclear region, and hence amplifies the effect of metallicity dilution that occurs when low metallicity gas floods the central region. While this is definitely a physical effect, its magnitude will be misjudged if the simulations do not properly account for gas recycling. Using the stochastic gas recycling method outlined in §\[sec:GasRecycling\], we are able to modulate the efficiency with which gas becomes locked in the stellar phase without changing our star formation efficiency or feedback prescriptions. Instead, we change the gas recycling parameter, allowing gas to be returned from the stellar state back to the ISM. We first consider a simulation where newly formed stellar material is instantaneously returned to the ISM, such that no gas will be trapped in the stellar phase. At the opposite extreme, we perform a simulation with star formation, but no gas recycling (i.e. $\zeta = 0.0$), which would trap the largest amount of gas in the stellar phase. In between these extremes, we consider three values for the gas recycling fraction, $\zeta$, that show how the nuclear metallicity evolution changes as the efficiency with which gas is locked in the stellar state is varied. When the gas recycling parameter is set to very large values, the resulting enriched metallicity becomes unphysically high. This occurs when many generations of star formation are allowed to occur over very short timescales as the gas is quickly recycled into the ISM repeatedly. However, the resulting dynamical metallicity remains physical and meaningful, as it describes the change in the nuclear metallicity that will occur as a function of the efficiency with which material is locked in the stellar phase. In fact, dynamical metallicity given by our simulation with instantaneous gas recycling (see the black line in the center panel of Figure \[fig:ZTpanel\]) is very similar to the results of @Rupke2010. These simulations capture changes to the nuclear metallicity caused by dilution without any contributions or contamination from the creation of stars. In this case, the nuclear metallicity is well-correlated with the mass of gas in the nuclear region, as noted by @Rupke2010 [see their Fig. 2]. This indicates that the initially high nuclear metallicity is diluted by inflows of low metallicity gas. However, as the gas recycling parameter is dialed back to lower values, gas is more efficiently trapped in the stellar phase which increases the effect of metallicity dilution. The simulation with star formation and no gas recycling brackets the upper end for the amount of gas that can be trapped in the stellar phase. In this case, the dynamical metallicity (red dotted line in the center plot of Figure \[fig:ZTpanel\]) is lower than the dynamical metallicities of all other simulations and falls $\sim$0.2 dex below that of the nuclear metallicity with instantaneous recycling. This difference is caused purely by the effect of locking gas in the stellar state. When no gas recycling is used, a substantial fraction of the inner gaseous reservoir is converted into the stellar state, allowing influx of low metallicity gas to have stronger influences. In addition, the enriched metallicity is lower than any other simulation because the resulting star formation rates are lower when gas is efficiently locked in the stellar state. The resulting total metallicity is below the total metallicity where gas recycling is used. As gas recycling is included at intermediate values, both the dynamical metallicity and enriched metallicity increase. The increase in the dynamical metallicity for larger gas return fractions is well-motivated by the simple physical arguments presented above. Specifically, increasing the gas recycling fraction lowers the efficiency with gas is depleted from the nuclear region into the stellar state. Hence, as the gas recycling fraction is increased, the mass of gas in the inner reservoir increases, and the influence of metallicity dilution is decreased. The increase in the enriched metallicity for larger gas return fractions is similarly explained. Gas that resides in the nuclear region will tend to remain there for longer periods of time (on average) and achieve higher instantaneous and integrated star formation rates (hence, higher metallicities). While this sheds light on the influence of gas recycling, we would like to use physically motivated values for the gas recycling fraction in what follows. We take the unscaled Starburst99 mass return fraction as the fiducial mass return fraction used for the remainder of this paper. {height="2.3in"} {height="2.3in"} {height="2.3in"} {height="2.3in"} {height="2.3in"} {height="2.3in"} ### Galactic Winds {#sec:GalacticWinds} Galactic winds are thought to play a leading role in the evolution of the distribution of metals throughout cosmic time. Starburst–driven, or AGN–driven winds are a ubiquitous phenomena that act to transport metals from the depths of galaxies into the interstellar, or intergalactic medium [see, e.g., @VeilleuxARAA for a comprehensive review]. Galactic winds may also play a prominent role in shaping the mass–metallicity relationship [see, e.g., @DaveMZ2006] and because galaxy mergers induce episodes of enhanced star formation, they may also induce periods of significant feedback and galactic outflows thereby altering the central metallicity evolution of galaxy mergers. In the context of this paper, galactic winds can remove gas from central regions and modify the reservoir of material contributing to the nuclear metallicity [@PerezScoop]. In this capacity, the influence of galactic winds is identical to the locking of gas in the stellar state. To demonstrate the impact of galactic winds on the nuclear metallicity, we present a set of simulations where we vary the mass entrainment, $\gamma$. In these simulations we do include gas recycling at the fiducial level, but note that the our conclusions are unchanged if this effect is turned off. The dynamical metallicity is not substantially affected by galactic winds. Since star formation is already fairly efficient at locking gas in the stellar state, galactic winds provide only an incremental modification to this effect. However, notable changes occur in the enriched metallicity. Specifically, the enriched metallicity is depressed for large values of the mass entrainment because gas with high star formation rates is either quickly converted into the stellar phase or launched in a wind. Therefore, the integrated star formation rates and enriched metallicity values for gas particles remains lower, on average, for higher mass entrainment values. The resulting nuclear metallicity, shown in Figure \[fig:WindFraction\], is depressed as mass entrainment is increased. This change is primarily caused by changes to the enriched metallicity. However, there is one additional effect that distinguishes galactic winds from locking of material in the stellar phase. While gas that is permanently locked in the stellar state will forever be unable to contribute to the nuclear metallicity or star formation, gas that has been launched in a wind will ultimately fall back onto the disk, or even into the nuclear region. We see this effect having an influence in the enriched metallicities of the highest mass entrainment value simulations at late times. Specifically, as gas inflow from previously ejected wind material falls upon the nuclear region, the star formation rates and enriched metallicities are pushed toward higher values. It should be noted that the galactic wind launching prescription used in this work does not describe a situation where the galactic wind predominately ejects enriched gas. In other words, the metallicity of the wind is the same as the ambient metallicity of the star forming region from which the wind was launched. This prescription could underestimate the efficiency with which metals are pushed out of a galaxy and into the intergalactic medium. However, this wind launching prescription is straightforward to understand and has the direct and notable effect of removing gas from the nuclear region. Galactic winds do little to change the radial rearrangement of gas which attends a galaxy merger, as shown by the relative insensitivity of the dynamical metallicity to the efficiency of the winds. The only appreciable effects are on the enriched metallicity. In the following analysis, we take $\gamma = 0.3$ as the fiducial mass entrainment value [@RupkeOutflows2005]. Disk Spin Orientation {#sec:MergerOrientation} --------------------- While the previous analysis focused on a single merger setup, the effect of merger orientation can be tested using a set of identical “B” galaxies merging from 16 orientations (a-p, as detailed in Table \[OrbitalOrientations\]). All orientations maintain unique metallicity evolutions, as the specifics governing the magnitude and timing of the gas inflows and starbursts are orientation dependent. Here we will consider which aspects of our fiducial merger’s metallicity evolution are found to hold for various merger orientations. To test this, we first examine the metallicity evolution for the four orientations shown in Figure \[fig:ZTorientations\]. {height="2.3in"} {height="2.3in"} {height="2.3in"} As is evident from this figure, three of the four disks show strong depressions in their nuclear metallicity following pericentric passage. This is caused by the strong tidal interaction and resulting gas inflows which are a natural byproduct of the merger process. The disk not experiencing a post pericentric nuclear metallicity depression is identified as being on a nearly retrograde orientation. More concretely, the induced gas inflow following pericenter passage is found to be small compared to the other systems. The idea that some orientations, specifically those on retrograde orbits, have suppressed responses during pericentric passages has been studied previously [@TT72; @Donghia10], and we note that the metallicity evolution is affected accordingly in the cases where strong post-pericenter passage responses are not found. Subsequent depressions in the nuclear metallicity, can be identified as occurring during periods of strong gas inflow caused by close encounters. For the “e” and “h” orientations, only one additional close encounter occurs before proceeding to coalescence. However, for the “f” orientation, three metallicity dips can be seen. These metallicity dips are natural products of this particular merger orientation, which leads to three distinct pericenter passages before the galaxies ultimately coalesce. While a hard and fast rule governing the metallicity depression’s dependence on merger orientation cannot be given, two statements can be made: most orientations undergo a metallicity depression following pericentric passages (unless the tidal interaction is sufficiently suppressed), and most orientations show metallicity depressions immediately preceding final coalescence. Both of these statements are demonstrated in figure \[fig:ZTorientations\], where three of four (e,f, and h; not k) orientations show post-pericentric dips in their nuclear metallicity and three of four (e, k, and h; not f) orientations show dips in their nuclear metallicities preceding final coalescence. If we identify the merging time, $t_m$, to be the time between pericenter passage and final coalescence, we can average the metallicity evolution for several orientations to get an orientation averaged progression of the nuclear metallicity during the merger. Figure \[fig:ZTavg\] shows the mean nuclear metallicity evolution for all 16 orientations, with the solid region indicating the 1$\sigma$ variations. Despite the variation that occurs in the nuclear evolutionary tracks for individual orientations, the mean result has a well-defined pattern with reasonably small dispersion and a clear interpretation. Specifically, there are two notable metallicity depressions caused by gas inflows, one following pericentric passage and one preceding black hole coalescence, with a chemical enrichment induced peak separating in between. Also shown in Figure \[fig:ZTavg\] is the average dynamical metallicity. Here we can see that simulations without chemical enrichment substantially under-predict the resulting nuclear metallicity. However, even when chemical enrichment is included, the mean nuclear metallicity is depressed throughout the interaction. Progenitor Masses {#sec:MergerMass} ----------------- We examine the effect that system mass has on the nuclear metallicity evolution using a set of major mergers (i.e. 1:1 mass ratios) with progenitor galaxies of varying mass. As detailed in §\[sec:GalaxySetup\], the four galaxies used in this paper are constructed to be self-similar. Hence, the mass study performed here isolates the influence of merger mass from other parameters that may scale with mass (e.g. gas fraction, galaxy structure, etc.). We use four orientations in each mass bin (the “e”, “f”, “h”, and “k” orientations), and average the resulting metallicity evolutionary tracks. The result, shown in Figure \[fig:ZTmasses\], is that the shape of the evolutionary track is nearly preserved, while the magnitude of the depression depends weakly on mass. We note that the preserved shape of the metallicity evolution supports our argument in the previous sections that the metallicity evolutionary track is a natural byproduct of the generic merger process. This indicates that, within the limited range of galaxy masses that are used here, the system mass is not a major consideration in determining the nuclear metallicity evolution. Gas Fraction {#sec:MergerGasFraction} ------------ Initial gas fraction is an important parameter to consider in nuclear metallicity evolution. For higher gas fraction galaxies that, by definition, have lower initial stellar mass, two basic changes occur: The initial nuclear metallicity of the system, as prescribed by the MZ relation, is lowered (decreasing the importance of the dynamical metallicity), and higher star formation rates occur (increasing the importance of chemical enrichment). The combined result, as discussed in this section, is that the nuclear metallicities of interacting systems at higher gas fractions tend to evolve to larger values, rather than decreasing, in contrast to their lower gas fraction counterparts. We first consider the nuclear metallicity evolution for a series of disks with varied initial gas fractions (8, 20, 40, and 60%). We merge together these systems on the “e“, “f”, “h”, and “k” orientations. If we initialize the metallicity properties of these disks using the local MZ relation, an assumption that we will relax later in this section, we can track the metallicity evolution and plot the change in the nuclear metallicity, as in Figure \[fig:GasFractionHist\]. This shows that the depressions in the nuclear metallicity that occurred for the low initial gas fractions considered in the previous sections shift toward less negative, or even positive values, as the initial gas fraction is raised. This transition is caused by the increasing dominance of the enriched metallicity over the dynamical metallicity, primarily owing to the higher star formation rates in the higher gas fraction systems. These highest gas fraction systems, however, are not natural analogs to galaxies in the local Universe. Therefore, our choice to initialize the metallicity properties of these systems using the local MZ relation comes into question. Instead, we should consider an MZ relation better suited for higher redshifts. A full theoretical understanding of the origin of the MZ relation has not yet been obtained, however, cosmological simulations indicate that the shape of the MZ relation is achieved at early times (i.e. z$>$6) and that galaxies move along this MZ relation toward higher stellar mass, while the amplitude of the MZ relation slowly increases over time [@DaveMZ2006], in broad consistency with observations [@ErbMZ2006; @Maiolino2008]. Without committing ourselves to a single choice for the high redshift MZ relation, we can re-initialize the metallicity properties of our galaxies using an MZ relation that is offset by a constant fixed value, $\delta$, toward lower metallicities. Since our simulations have no metallicity dependence, re-normalizing the MZ relation has no effect on the enriched metallicity, but lowers the impact of the dynamical metallicity. The result, shown in the right panel of Figure \[fig:GasFractionHist\], is that as the the MZ relation is shifted to lower metallicities, the change in the nuclear metallicity is expected to become increasingly positive. For all choices for the initial MZ relation, the higher gas fraction systems have a median change in metallicity that is more positive than their low gas fraction counterparts. Using the local MZ relation (i.e. the black line in Figure \[fig:GasFractionHist\]) for systems with relatively low gas fractions, we find that negative changes to the nuclear metallicity are expected in a merger. This is consistent with the observed lowering of the MZ relation. However, as we move to higher gas fractions, or MZ relations better suited for higher redshifts, we find that increases to the nuclear metallicity are expected. This result is caused by shifting the dominant driver of the nuclear metallicity from metallicity dilution to chemical enrichment. Moreover, if the gas content is sufficiently high, the galactic stellar component cannot efficiently torque down the gas, resulting in moderated early stage inflows [@Hopkins09c]. The same merger fundamentals apply to low and high gas fraction disks, independent of their initial metallicities, but the relative increase in the role of chemical enrichment with respect to the dynamical metallicity causes higher redshift or gas fraction systems to increase their nuclear metallicities during the merger process. Without being overly restrictive about our choice for initial gas fraction or offset from the local MZ relation, high-redshift, gas-rich galaxies are expected to have median increases in their nuclear metallicity when undergoing a merger. Based on this evidence, we predict that the depression that occurs in the MZ relation during galaxy mergers and interactions in the local Universe would not occur at higher redshift. Instead, increases in the MZ relation are expected. Comparison with Observations {#sec:ComparisonToObservations} ============================ Much work in recent years has gone into observationally classifying the metallicity of interacting or close pair galaxies in the nearby Universe  [@KGB06; @SloanClosePairs; @Peeples2009; @SolA2010]. In this section we compare our results directly with these observations. To facilitate this comparison, we parameterize the evolution of the nuclear metallicity in terms of the galactic projected separation. We define the projected separation as $s = r \cos (\xi)$ with $\xi$ being a uniformly distributed random number between $0$ and $2\pi$. Since all of our systems are, by definition, interacting systems, this projected separation neglects contributions from truly non-interacting systems. We present the mass-metallicity (MZ) and separation-metallicity (SZ) relations for a simulated population of interacting galaxies. For these comparisons, we generate a population of progenitor galaxies that contain the metallicity properties of observed field galaxies [@Shields1990; @BelleyRoy1992; @Z94; @MA04]. We use a Gaussian distribution for the nuclear metallicity selection with a mean value consistent with the MZ relation of [@Tremonti2004] with a constant standard deviation across mass bins of $\sigma = 0.1$ dex. The metallicity scale length is taken from a Gaussian distribution with a mean gradient consistent with @Z94 of $h_z = h / 0.2$ with $\sigma = 0.3h_z$. The initial metallicity properties of each galaxy in each snapshot are assigned based on the above criteria. Our ability to pick metallicity properties of the progenitor disks in our post-processing analysis is a consequence of not having a dynamical dependence on metals or allowing mixing. This lets us sample the parameter space of initial disk metallicity properties thoroughly, without re-running simulations. The Mass-Metallicity Relation for Interacting Galaxies ------------------------------------------------------ The MZ relation for our simulated interacting galaxies is shown in Figure \[fig:MZ\]. We select galaxies with projected separations less than 30 kpc – the same as was used in @SloanClosePairs. We plot the MZ relation in Figure \[fig:MZ\]. The result is a depressed MZ relation with an average depression of 0.07 dex, which is in agreement with the observed depression of 0.05-0.10 dex [@SloanClosePairs]. Unlike luminosity, which can increase dramatically during merger induced starburst activity, the stellar mass of an interacting galaxy will change only by a modest factor. Therefore, the evolution of galaxies on the MZ relation is driven by changes to the nuclear metallicity, not stellar mass. In the simulation analysis, we are able to quickly determine a system’s stellar mass. The stellar masses for the close pair systems included in the @SloanClosePairs sample have had their stellar masses determined via optical and IR colors [@KauffmannSDSS; @Tremonti2004] which have been shown to compare well with the spectrally-determined stellar masses [@DroryMasses]. Assuming that the observationally determined stellar masses are not strongly affected by ongoing starburst activity, the depression in both the observed and simulated MZ relation is being caused entirely by changes to the nuclear metallicity – not by changes in the galactic luminosity or stellar mass. The depression in our simulated MZ relation would be slightly smaller if we included a set of field galaxies that lie on the MZ relation, but appear close in projected separation (i.e. if we included a population of interlopers). This result gives a comprehensive reproduction of the depression in the MZ relation for interacting galaxies that accounts for all relevant physics and the intrinsic scatter in the initial metallicity properties of the progenitor galaxies, while covering a reasonable portion of merger phase space. Separation vs Metallicity {#sec:SZ} ------------------------- Figure \[fig:OneSZ\] shows an example of two galaxies (“C” galaxies) from one merger simulation (the “e” orientation) moving in separation–metallicity (SZ) space as they approach coalescence. The evolution can be broken down into four distinct stages, as marked by arrows in Figure \[fig:OneSZ\]. First, the galaxies approach first pericenter passage at relatively constant nuclear metallicities. Second, the galaxies separate as gas inflows dilute their nuclear metallicity. Third, the galaxies begin falling back together, with star formation causing a significant amount of chemical enrichment. And, finally, the galaxies undergo strong gas inflows that further dilute their nuclear metallicities as they approach coalescence. The details of this SZ evolution will change substantially with the merger orientation. However, we can gather from Figure \[fig:ZTorientations\] that all four stages of this evolution will be somewhat generic. Namely, galaxies will present constant nuclear metallicities as they approach each other for the first time, and strongly depressed nuclear metallicities as they approach final coalescence, with mildly depressed nuclear metallicities in between. Using the evolutionary track of \[fig:OneSZ\], we can develop some expectations for the distribution of galaxies in SZ space. Clearly, we expect close pair galaxies to show depressed nuclear metallicities. In addition, we can expect depressed metallicities out to large galactic separations. While the largest nuclear metallicity depressions are expected around close separations, widely separated interacting galaxies show mildly depressed nuclear metallicities after interacting with their companion. The observability of these trends can be assessed by using our entire merger library and converting real-separations into projected separations. The distribution of all simulated galaxies in SZ space is shown in Figure \[fig:SZ\] along with data from @KGB06. The distribution of our simulated galaxies, shown with the colored contours in the lower half of Figure \[fig:SZ\], coincide with many of the observed systems, which is demonstrated in the histogram in the top half of Figure \[fig:SZ\]. There is a trend toward lower metallicities with close separations. However, this trend is very mild and the most-extreme low-separation low-metallicity objects from the observed sample are not reproduced by our models. Despite this difference, it should be noted that our simulations agree with the majority of the observed objects from @KGB06, and, there is a systematic but subtle drop in the simulated nuclear metallicity as the galaxies decrease their projected separation. We note that the overall distribution of simulated galaxies in SZ space is relatively featureless, as evidenced by the contours in Figure \[fig:SZ\]. The nuclear metallicity depressions present at both small and large separations in Figure \[fig:OneSZ\] are washed out as a result of projecting the physical separations, the wide variety of metallicity dilution versus separations that result from different spin–orbit couplings, and the scatter in the initial nuclear metallicities as prescribed by the mass-metallicity relation. In addition, while late stage mergers (e.g., stage 4 in Figure \[fig:OneSZ\]) present the largest depressions in nuclear metallicity, they only last for a short period of time compared to the overall merger process. Since we place equal weight on every timestep for every simulation when determining the distribution of simulated galaxies in Figure \[fig:SZ\], the highly-depressed late-stage mergers will naturally have a less dramatic impact. Several factors can influence the distribution of our model galaxies in SZ space. In particular, the (in)completeness of our merger library and the metal enrichment scheme will influence the resulting SZ relation. We discuss both of these factors in the following subsections, so that we may better understand the disagreement between our models and the observations at low separations and low metallicities. However, we also note that additional observations would help clarify how common these low-separation, low-metallicity objects are. ### Merger Library Completeness Although the merger library used in this work was fairly comprehensive, there is a significant portion of merger parameter space that has not been explored. Perhaps this could explain why we do not reproduce the low-separation, low-metallicity objects. We can describe our exploration of merger parameter space with two components: Variations of initial isolated galaxies models and merger configuration parameters. Additional merger configurations (i.e. variations to the orbital orientation and angular momentum) are unlikely to yield surprising results. Out of all of our merger simulations, a maximum depression of about 0.3 dex was found, with many depressions being must less drastic. To achieve the lowest-metallicity objects found in @KGB06, depressions of $\sim$ 0.6-0.7 dex would be required. It seems unlikely that the isolated galaxy models used in this study will reproduce very low metallicity objects, even with a more extensive exploration of merger orientations. On the other hand, galaxy models that are significantly different from the ones used in this study, such as low mass galaxies, could fill out a new region in SZ space. For example, galaxies with initial stellar masses of $M_{s} \sim 10^9 M_{\odot}$ will have initial nuclear metallicities of $12+ \log (O/H) \sim 8.6$ and could have their nuclear metallicities depressed to $\sim$ 8.3 or 8.4 during the merger. More generally, it seems likely that the lowest metallicity objects of @KGB06 are likely merging systems that had low metallicities prior to the onset of the merger which may be a reasonable possibility given the low B-band luminosities associated with these objects. These systems may be reproduced by our simulations when we include a more extensive set of initial galaxy models. ### Approximating Delayed Mixing Although oxygen abundances are well-approximated by an instantaneous enrichment model, this does not necessarily imply that metals produced in SNe instantaneously migrate to HII regions. Specifically, we can relax our assumption that the metallicity of an SPH particle should be instantaneously identical to the metallicity of an HII region. Our analysis calculated the nuclear metallicity from the current metallicities of gas particles. While the metals produced in core collapse supernovae are quickly returned to the ISM ($\sim 10^7$ years) and supernovae remnants cool and fade into the ambient ISM over similar timescales ($\sim 10^6$ years), there could then be some additional delay before this gas is sufficiently mixed with the surrounding medium and ends up in HII regions. In particular, we expect the metals produced by SNe to affect the *next* generation of HII regions. As a first approximation, this does not require a change to our simulations, but rather a change to our post-processing analysis. We can define an effective HII region metallicity that is given by the metallicity that the SPH particle had one generation of star formation ago. If the number of generations of star formation is given by $$N_* (t) = \frac{1}{M} \int _0 ^t \frac{dM_*}{ dt^\prime} dt ^\prime$$ where $M$ is the particle mass, then the effective HII region metal content is given by $$M_{Z} ^\prime (t) = \left\{ \begin{array}{l l} M_{Z,init} + y M \left( N_* - 1 \right) & \quad \text{if $N_* \ge 1$ }\\ & \\ M_{Z,init} & \quad \text{if $N_* < 1$ }\\ \end{array} \right.$$ which reflects the particle’s metallicity one generation of star formation in the past. This delayed chemical enrichment scheme gives an SZ relation, shown in Figure \[fig:MultiContourSFR\], which has a low metallicity tail that is qualitatively similar to the observed galaxies. This does not imply that our previous calculation of the nuclear metallicity using the non-delayed chemical enrichment values were incorrect, but rather that the assumption of instantaneous transition of metals into HII regions may be overly restrictive. This is a plausible explanation, provided that the associated delay-times are reasonable estimates for the timescale over which SN ejecta would mix and diffuse into the ISM. If we parameterize an effective HII region metallicity in terms of a delay-time, rather than an integrated star formation rate, we can get an estimate for the required delay time. Rather than neglecting the enrichment from the last generation of star formation, we neglect the enrichment that occurred over the past $t_d$ years. The effective metal content of a particle is then given by $$M_{Z} ^\prime (t) = \left\{ \begin{array}{l l} M_{Z,init} + y \int _0 ^ {t - t_d} \frac{dM_*}{dt^\prime} dt ^\prime & \quad \text{if $t \ge t_d$ }\\ & \\ M_{Z,init} & \quad \text{if $t < t_d$ }\\ \end{array} \right.$$ where $t_d$ is the assumed time delay. This time delayed chemical enrichment scheme gives an SZ relation, shown in Figure \[fig:MultiContour\], where larger delay times lead to systematic decreases in the inferred metallicity of close pairs. This is similar to those seen in Figure \[fig:MultiContourSFR\], but with an explicit time dependence. The delay times required for our simulations to match observations is of the order of a few $10^8$ years, which is comparable to estimates for the metal mixing timescale of newly produced metals into the ISM [@OeyMetalMixing; @ScaloElmegreenMetalMixing]. We conclude that delaying the transfer of metals from SN remnants to HII regions is important to consider; however, it is unclear whether these long mixing times are appropriate for the extreme environments in merging galaxy nuclei where external forces, gas inflows, high gas densities, and large star formation rates are present. Discussion {#sec:Discussion} ========== Predictions for Galaxies with High Gas Content ---------------------------------------------- The results of § \[sec:MergerGasFraction\] demonstrate that the merger–induced metallicity evolution is heavily dependent upon gas content, specifically, higher gas fraction disks produce less nuclear metallicity depression than lower gas fraction disks (see, e.g., Fig. \[fig:GasFractionHist\]). In the extreme, high gas content mergers may even yield significant [*enhancements*]{} of central metallicity, a result that emerges both because of the increased impact of enrichment from star formation as well as because nuclear gas inflows are suppressed when the gas fraction is very high [@Hopkins09c]. As a consequence of these trends, we expect that the nuclear metallicities of gas–rich merging galaxies will be similar to, or greater than, an appropriately matched quiescent sample. This is in contrast to the depression of nuclear metallicity observed in low gas–content mergers which was the focus of this paper. Such considerations may be applicable at high redshift where current observations suggest that galaxies contain significantly higher gas fractions [@Tacconi2010; @ErbGasFraction] and suppressed metallicity at fixed galaxy mass [@ErbMZ2006; @Maiolino2008] or for less–massive gas–rich galaxies in the local universe [@Cat10]. The predicted metallicity enhancements in gas–rich merging systems was also noted in a recent paper by @PerezScoop using similar methods. While both works observed similar metallicity enhancements, the timing was not identical, and purported cause of these effects were very different. The @PerezScoop model shows rapid metallicity enhancement during early stages of the merger primarily as a result of rapid star formation owing to disk instabilities in the gas–rich system (see, e.g., their Figs. 8 and 9). Such effects occur prior to, or during the early stages of the merger and are therefore decoupled from the final coalescence stage. In our model, the gas–rich disks are designed to be stable a priori, and therefore the metallicity enhancements are driven by intense star formation triggered by the merger. While the end result is the same, namely metallicity enhancements, the relative timing and strength of the vigorous star formation, the metallicity dilution owing to inflows, and the metallicity enhancement owing to star formation are fundamentally different, and therefore potentially discernible via observations. We caution, however, that the galaxy merger simulations carried out here may require some generalization before they are directly applicable at high redshifts. Effects such as the higher UV background, higher merger rate, and direct fueling of low metallicity gas via the surrounding intergalactic medium at high redshifts may be of significant influence in the observed metallicity of galaxies. And numerical uncertainties, such as the proper treatment of feedback from star formation and accreting black holes will invariable need to be explored. Future Considerations: An Improved Chemical Enrichment Model {#sec:future} ------------------------------------------------------------ By comparing our models of interacting galaxies to the observed mass-metallicity (MZ) and separation–metallicity (SZ) relations we have refined our understanding of the factors at play during the metallicity evolution of galaxies. It is crucial to point out (yet rarely done so), however, that numerical results such as ours are based upon the ability to robustly track a number of complicated and interrelated astrophysical processes. While we consider our sub–resolution models to be well motivated and well tested, we admit that there are straightforward improvements to these models which will allow for more stringent tests of our physical picture. We now take a moment to outline these future improvements to our models. The chemical enrichment model employed here tracks a global metallicity through a single scalar variable that is continuously updated assuming that the entire solar yield (assumed to be a mass fraction of 0.02) is instantaneously recycled within a single star–forming fluid element (i.e., an SPH particle). We admittedly note the simplicity of this model, but consider the ease with which it is implemented, the straightforward analysis it allows, and its inherent self–consistency as distinct advantages of this approach. The instantaneous recycling approximation is also justified because we compare to oxygen derived metallicities, and oxygen is thought to be recycled on a short ($\sim 10^7$ years) timescale through core–collapse supernovae. Moreover, in a practical sense, similar models have been employed in nearly every SPH code over the last two decades which makes our results easy to gauge within the existing theoretical framework. There are, however, avenues in which distinct progress can be made to our chemical enrichment model. Efforts toward this end have been pioneered by a number of authors over the last $\sim$5-10 years [see, e.g., @KawataGasReturn; @TornatoreGasReturn; @ScannapiecoGasReturn; @OppenheimerandDave; @KimAbel to name a few] by employing two common features: metallicity–dependent cooling, and more complex enrichment schemes that explicitly track various species and their time–dependent release. Our gas recycling scheme is particularly well-suited for including mass return by a single stellar population. As individual star particles are converted back to interstellar gas, they can straightforwardly carry with them the appropriate enrichment from Type I and II supernovae as well as mass–loss from massive young stars and asymptotic giant branch stars. Such an approach is necessary to study trends in abundance ratios, such as those observed in the centers of elliptical galaxies. Additional accuracy may also be gained by including sub–resolution models for the turbulent diffusion of metals [see, e.g., @WadsleyDiffusion; @ShenDiffusion]. Coupled to an advanced enrichment algorithm, a physically motivated procedure for diffusing metals will allow enrichment to occur locally but still have global influences if conditions warrant. Such methods have been shown to be important for tracking metallicity in cosmological simulations [e.g., @ShenDiffusion] and we intend to assess their influence in isolated and merging galactic systems. Finally, the comparisons between the models and the observations can be significantly improved by actually tracking the detailed line emission using 3D radiative transfer [@JonssonCoxGroves]. As efforts to produce data cubes with full spatial and spectral information are developed [e.g. @RichKewley2010], full metallicity maps can be produced and directly compared to our simulations. Radiative transfer in our simulations will shed light on any additional information that may be encoded in the line profiles and allow for a more even–handed comparison. Conclusion ========== Using numerical simulations, we have investigated the impact that mergers have on galactic nuclear metallicity. Our models include the capability of describing star formation, chemical enrichment, gas recycling, and star– formation driven winds. The analysis performed here relies upon the ability to accurately track the spatial and temporal return of metals to the ISM enabled by our stochastic gas–recycling algorithm within GADGET (see § \[sec:GasRecycling\]). One of the primary results of this work is to reinforce the notion that galaxy mergers, and their attending gravitational tidal forces, generate significant inflows of gas shown explicitly in Figures \[fig:ZTorientations\] and \[fig:ZTavg\]  [see also @Rupke2010; @DilutionPeak2010; @PerezScoop]. These inflows transport metal–poor gas into the nuclear region and reinforce the generic connection between close galaxy interactions and nuclear metallicity dilution – no matter how many close passages occur. Merger–induced gaseous inflows are, however, not the only factor influencing the nuclear metallicity, and this paper quantifies the competing effects of enrichment from star formation (§ \[sec:DilutionAndEnrichment\]), gas consumption (§ \[sec:GasConsumption\]), and galactic winds (§ \[sec:GalacticWinds\]). Our models provide a physical template to understand nuclear metallicity evolution and lead us to conclude that merger–induced metallicity dilution and chemical enrichment from star formation drive the nuclear metallicity, while gas consumption and galactic winds play a secondary role to modulate the efficiency of the primary processes. We have made a distinct effort in this work to directly compare our models to observations, allowing us to validate the physical evolutionary model that we have presented and associate observable trends with the various stages of the merger evolution. We find that the merger–induce depressions of nuclear metallicity are typically $\sim0.07$ dex, an offset similar to that observed to the mass–metallicity relation for close pairs in SDSS [@SloanClosePairs]. The models also demonstrate that central depressions of metallicity are time–dependent owing to the specific merger orbit, the galaxies’ spin–orbit coupling, and varying structural properties of the galaxies. When compared to the observed separation–metallicity relation of @KGB06, the models generally show a similar trend to produce lower central metallicities at smaller separations. The precise role of interactions is difficult to discern, however, because of the significant scatter in the relations (see \[fig:SZ\] and the discussion in  §\[sec:SZ\]). Future comparisons in which galaxies are morphologically classified according to merger stage, or separated according to IR luminosity or star–formation rate may be better suited to isolated the true effects of the interaction. This work also demonstrates that the progenitor gas content can have a profound influence on the resulting merger–induced metallicity dilution. Mergers between gas–rich progenitors can yield systematic metallicity enhancement, as opposed to metallicity dilution (see § \[sec:MergerGasFraction\], and specifically Fig. \[fig:GasFractionHist\]). In our picture, central metallicity enrichment comes from vigorous merger–induced star formation that can compensate for, and eventually overcome the merger–induced inflow of metal–poor gas. This view is slightly different than put forth in recent work by @PerezScoop, which argues that disk instabilities and clump formation drive rapid star formation and subsequent metal enrichment. But once these processes occur, which is typically early in the merger process, the story becomes similar to when the galaxies have low gas content. By studying the detailed correlations between merger stage, pair separation, central metallicity, and metallicity gradients, in both observations and the models, future work will allow for a better understanding of the (merger and enrichment) processes at work.

ArXiv

--- abstract: | We compute the mid-rapidity densities of pions, kaons, baryons and antibaryons in $Au$-$Au$ collisions at $\sqrt{s} =$ 130 GeV in the Dual Parton Model supplemented with final state interaction (comovers interaction). The ratios $B/n_{part}$ ($\overline{B}/n_{part}$) increase between peripheral ($n_{part} = 18$) and central ($n_{part} = 350)$ collisions by a factor 2.4 (2.0) for $\Lambda$’s, 4.8 (4.1) for $\Xi$’s and 16.5 (13.5) for $\Omega$’s. The ratio $K^-/\pi^-$ increases by a factor 1.3 in the same centrality range. A comparison with available data is presented. --- =23 truecm = 15 truecm =-0.5 truecm =-2 truecm \#1 \[\#1\]$$\begin{aligned} } \def\emini{\end{aligned}$$ewcommand[.]{} \#1[\[\[\#1\]\]]{} \#1\#2[\[\[\#1\],\[\#2\]\]]{} \#1\#2\#3[\[\[\#1\],\[\#2\],\[\#3\]\]]{} \#1\#2[\[\[\#1\]–\[\#2\]\]]{} \#1\#2\#3\#4[\[\[\#1\],\[\#2\],\[\#3\]–\[\#4\]\]]{} \#1\#2\#3[\[\[\#1\],\[\#2\]-\[\#3\]\]]{} [**Strange particle production at RHIC in the Dual Parton Model**]{}\ 8 truemm [**A. Capella$^{\rm a)}$, C.A. Salgado$^{\rm b)}$, D. Sousa$^{\rm c)}$**]{}\ 5 truemm $^{\rm a)}$ Laboratoire de Physique Théorique[^1]\ Université de Paris XI, Bâtiment 210, F-91405 Orsay Cedex, France\ 3 truemm $^{\rm b)}$ Theory Division, CERN, CH-1211 Geneva 23, Switzerland\ 3 truemm $^{\rm c)}$ ECT\*, Trento, Italy 1 truecm LPT Orsay 02-09 CERN-TH/2002-090 ECT\* 02-10 February 2002 The enhancement of the ratio of yields of strange baryons and antibaryons per participant nucleon, observed at CERN-SPS, [@1r] [@2r], is one of the main results of the Heavy Ion CERN program. A description of these data has been given in [@3r] in the framework of the Dual Parton Model DPM [@4r], supplemented with final state interaction. The net baryon yield is computed in the same framework taking into account the mechanism of baryon stopping, associated with baryon junction transfer in rapidity . We use its implementation in [@3r], which describes the SPS data. In the absence of nuclear shadowing, the rapidity density of a given type of hadron $h$ produced in $AA$ collisions at fixed impact parameter, is given by [@4r] [@9r] \[1e\] [dN\^[AAh]{} dy]{}(y,b) &=& n\_A(b)\ &&+ ( n(b) - n\_A(b) ) 2k   N\_[h,(b)]{}\^[q\_s-\_s]{} (y)  . Here \[2e\] n(b) = \_[pp]{} A\^2 d\^2s T\_A(s)  T\_B(b-s)/\_[AA]{}(b) is the average number of binary collisions and \[3e\] n\_A(b) = A d\^2s  T\_A(s) / \_[AA]{}(b)  , is the average number of participant pairs at fixed impact parameter $b$. $P$ and $T$ denote projectile and target nuclei. $k$ is the average number of inelastic collisions in $pp$ and $\mu(b) = k \nu (b)$ with $\nu (b) = n(b) /n_A(b)$ is the average total number of collisions suffered by each nucleon. At $\sqrt{s} = 130$ GeV we have $k = 2$ [@9r]. The $N_{h,\mu (b)}(y)$ in eq. (\[1e\]) are the rapidity distributions of hadron $h$ in each individual string. In DPM they are given by convolutions of momentum distribution and fragmentation functions[^2]. The first term in (\[1e\]) is the rapidity distribution in one $NN$ collision of $AA$, resulting from the superposition of $2k$ strings, multiplied by the average number of participant pairs. Since in DPM there are two strings per inelastic collision, the second term, consisting of strings stretched between sea quarks and antiquarks, makes up for the total average number of strings $2k$ $n(b)$. It was shown in [@9r] that eq. (\[1e\]), supplemented with shadowing corrections, leads to values of charged multiplicities at mid-rapidities as a function of centrality in agreement with data, both at SPS and RHIC. Here we use the same shadowing corrections as in ref. [@9r] – leading to the lower edge of the shaded area in Fig. 4 of [@9r].\ . Let us now consider the net baryon production $\Delta B = B-\overline{B}$. In the standard version of DPM [@4r] (or QGSM [@11r]) the leading baryon results from the fragmentation of a valence diquark. This component will be called diquark preserving (DP). The stopping observed in $Pb$ $Pb$ collisions at SPS has led to the introduction of a new mechanism based on the transfer in rapidity of the baryon junction . Here we follow the formalism in [@3r] which describes the SPS data. In an $AA$ collision, this component, called diquark-breaking (DB), gives the following rapidity distribution of the two net baryons in a single $NN$ collision of $AA$ [@3r] \[4e\] ( [dN\_[DB]{}\^[B]{} dy]{} (y) )\_[(b)]{} = C\_[(b)]{} where $Z_{\pm} = \exp (\pm y - y_{max})$ and $\nu(b) = n(b)/n_A(b)$. $C_{\nu(b)}$ is determined from the normalization to two at each $b$. The net baryon rapidity distribution in $AA$ collisions is then given by \[5e\] [dN\^[AA B]{} dy]{}(y,b) = n\_A(b)  . The physical content of eq. (\[5e\]) is as follows. Each nucleon interacts in average with $\nu(b)$ nucleons of the other nucleus. It has been argued in [@3r] that in only one of these collisions the string junction, carrying the baryon number, follows a valence diquark, which fragments according to the DP mechanism. In the $\nu (b) - 1$ others, the string junction is freed from the valence diquark and net baryon production takes place according to the DB mechanism, eq. (\[4e\]). In order to conserve baryon number, we have to divide by $\nu (b)$ and multiply by the number of participating nucleons. We obtain in this way eq. (\[5e\])[^3]. This equation gives the total net baryon density[^4]. In order to get the relative densities of each baryon and antibaryon species we use simple quark counting rules [@3r]. Denoting the strangeness suppression factor by $S/L$ (with $2L+ S = 1$), baryons produced out of three sea quarks (which is the case for pair production) are given the relative weights $$I_3 = 4L^3 : 4L^3 : 12L^2S : 3LS^2 : 3LS^2 : S^3 \ .$$ for $p$, $n$, $\Lambda + \Sigma$, $\Xi^0$, $\Xi^-$ and $\Omega$, respectively. The various coefficients of $I_3$ are obtained from the power expansion of $(2L + S)^3$. In order to take into account the decay of $\Sigma^*(1385)$ into $\Lambda \pi$, we redefine the relative rate of $\Lambda$’s and $\Sigma$’s using the empirical rule $\Lambda = 0.6(\Sigma^+ + \Sigma^-$) – keeping, of course, the total yield of $\Lambda$’s plus $\Sigma$’s unchanged. In this way the normalization constants of all baryon species in pair production are determined from one of them. This constant, together with the relative normalization of $K$ and $\pi$, are determined from the data for very peripheral collisions. In the calculations we use $S = 0.1$ $(S/L = 0.22)$. For net baryon production two possibilities have been considered. The first one is that the behaviour in $Z^{1/2}$, eq. (\[4e\]), is associated to the transfer of the string junction without valence quarks . In this case the net baryon is made out of three sea quarks and the relevant weights are given by $I_3$. In the second one, eq. (\[4e\]) is a pre-asymptotic term associated to the transfer of the baryon junction plus one valence quark [@6r]. In this case the relevant weights are given by $I_2$, i.e. from the various terms in the expansion of $(2L + S)^2$. This second possibility is favored by the data[^5]. Since the normalization of the total net baryon yield is determined from baryon number conservation, there is no extra free normalization constant. Moreover, the total net baryon yield is not affected by final state interaction.\ . The hadronic densities obtained above will be used as initial conditions in the gain and loss differential equations which govern final state interactions. In the conventional derivation [@12r] of these equations, one uses cylindrical space-time variables and assumes boost invariance. Furthermore, one assumes that the dilution in time of the densities is only due to longitudinal motion, which leads to a $\tau^{-1}$ dependence on the longitudinal proper time $\tau$. These equations can be written as [@12r] [@3r] \[6e\] = \_[k]{} \_[k]{}  \_k   \_ - \_k \_[ik]{}  \_i  \_k  . The first term in the r.h.s. of (\[6e\]) describes the production (gain) of particles of type $i$ resulting from the interaction of particles $k$ and $\ell$. The second term describes the loss of particles of type $i$ due to its interaction with particles of type $k$. In eq. (\[6e\]) $\rho_i = dN_i/dy d^2s(y,b)$ are the particles yields per unit rapidity and per unit of transverse area, at fixed impact parameter. They can be obtained from the rapidity densities (\[1e\]), (\[5e\]) using the geometry, i.e. the $s$-dependence of $n_A$ and $n$, eqs. (\[2e\]), (\[3e\]). The procedure is explained in detail in [@13r]. $\sigma_{k\ell}$ are the corresponding cross-sections averaged over the momentum distribution of the colliding particles. Equations (\[6e\]) have to be integrated from initial time $\tau_0$ to freeze-out time $\tau_f$. They are invariant under the change $\tau \to c \tau$ and, thus, the result depends only on the ratio $\tau_f/\tau_0$. We use the inverse proportionality between proper time and densities and put $\tau_f/\tau_0 = (dN/dyd^2s(y,b))/\rho_f$. Here the numerator is given by the DPM particles densities. We take $\rho_{f} = [3/\pi R_p^2](dN^-/dy)_{y^*\sim 0} = 2$ fm$^{-2}$, which corresponds to the charged density per unit rapidity in a $pp$ collision at $\sqrt{s} = 130$ GeV. This density is about 70 % larger [@9r] than at SPS energies. Since the corresponding increase in the $AA$ density is comparable, the average duration time of the interaction will be approximately the same at CERN SPS and RHIC – about 5 to 7 fm. Next, we specify the channels that have been taken into account in our calculations. They are \[7e\] N K () , () K  , K We have also taken into account the strangeness exchange reactions \[8e\] () K N , K ()  , K  . as well as the channels corresponding to (\[7e\]) and (\[8e\]) for antiparticles[^6]. We have taken $\sigma_{ik} = \sigma = 0.2$ mb, i.e. a single value for all reactions in (\[7e\]) and (\[8e\]) – the same value used in ref. [@3r] to describe the CERN SPS data.\ . All our results refer to mid-rapidities. The calculations have been performed in the interval $-0.35 < y^* < 0.35$. In Fig. 1a-1d we show the rapidity densities of $B$, $\overline{B}$ and $B - \overline{B}$[^7] versus $h^- = dN^-/d\eta = (1/1.17) dN/dy$ and compare them with available data . We see that, in first approximation, $p$, $\overline{p}$, $\Lambda$ and $\overline{\Lambda}$ scale with $h^-$. Quantitatively, there is a slight decrease with centrality of $p/h^-$ and $\overline{p}/h^-$ ratios, a slight increase of $\Lambda /h^-$ and $\overline{\Lambda}/h^-$ and a much larger increase for $\Xi$ ($\overline{\Xi})/h^-$ and $\Omega$ ($\overline{\Omega})/h^-$. In Fig. 2a and 2b we plot the yields of $B$ and $\overline{B}$ per participant normalized to the same ratio for peripheral collisions versus $n_{part}$. The enhancement of $B$ and $\overline{B}$ increases with the number of strange quarks in the baryon. This increase is comparable to the one found at SPS between pA and central PbPb collisions – somewhat larger for antibaryons. The ratio $K^-/\pi^-$ increases by 30 % in the same centrality range, between 1.1 and 1.4 in agreement with present data [@14r]. The ratios $\overline{B}/B$ have a mild decrease with centrality of about 15 % for all baryon species – which is also seen in the data [@18r]. Our values for $N^{ch}/N_{max}^{ch} = 1/2$ are : $\overline{p}/p = 0.69$, $\overline{\Lambda}/\Lambda = 0.72$, $\overline{\Xi}/\Xi = 0.79$, $\Omega/\overline{\Omega} = 0.83$[^8], to be compared with the measured values [@15r] : $$\overline{p}/p = 0.63 \pm 0.02 \pm 0.06 \quad , \quad \overline{\Lambda}/\Lambda = 0.73 \pm 0.03 \quad , \quad \overline{\Xi}/\Xi = 0.83 \pm 0.03 \pm 0.05 \ .$$ The ratio $K^+/K^- = 1.1$ and has a mild increase with centrality, a feature also seen in the data. As explained above only one parameter has been adjusted in order to determine the absolute yields of baryons and antibaryons. Since both the PHENIX [@14r] and the STAR data [@16r] are not corrected for feed-down from weak decays, this free parameter has to be re-determined after these corrections are known. It will then be possible to compare its value with the one obtained from other sets of data, in particular $pp$. Likewise, the corrected yields of net protons will determine the exact amount of stopping and should allow to decide whether or not there is “anomalous” stopping in $AA$, i.e. an excess as compared to an extrapolation from $pp$ and $pA$. At present there is no clear indication of anomalous stopping.\ After completion of this work, the PHENIX collaboration [@21r] has published the yields of $p$ and $\overline{p}$ for the 5 % most central events corrected for feed-down.The corrections are of 30 %.The systematic error is 20 %.Also, C.Roy (STAR collaboration) communicated to us a recent STAR preliminary result on the mid-rapidity densities of $\Xi^-$ and $\Xi^+$ for the 14 % most central events. These values are 60 to 70 % higher than our results. Their systematic error is 20 %. . Before final state interactions, all ratios $K/h^-$, $B/h^-$ and $\overline{B}/h^-$ decrease slightly with increasing centrality. This effect is rather marginal at RHIC energies and mid-rapidities. The final state interactions (\[7e\]), (\[8e\]) lead to a gain of strange particle yields. The reason for this is the following. In the first direct reaction (\[7e\]) we have $\rho_{\pi} > \rho_K$, $\rho_N > \rho_{\Lambda}$, $\rho_{\pi} \rho_N \gg \rho_K \rho_{\Lambda}$. The same is true for all direct reaction (\[7e\]). In view of that, the effect of the inverse reactions (\[7e\]) is small. On the contrary, in all reactions (\[8e\]), the product of densities in the initial and final state are comparable and the direct and inverse reactions tend to compensate with each other. Baryons with the largest strange quark content, which find themselves at the end of the chain of direct reactions (\[7e\]) and have the smallest yield before final state interaction, have the largest enhancement. Moreover, the gain in the yield of strange baryons is larger than the one of antibaryons since $\rho_B > \rho_{\overline{B}}$. Furthermore, the enhancement of all baryon species increases with centrality, since the gain, resulting from the first term in eq. (\[6e\]), contains a product of densities and thus, increases quadratically with increasing centrality. Although the inverse slopes (“temperature”) have not been discussed here, let us note that in DPM they are approximately the same for all baryons and antibaryons both before and after final state interaction – the effect of final state interaction being rather small [@14rnew] [@20r]. In conclusion, we would like to emphasize the fact that in DPM (before final state interaction) the rapidity density of charged particle per participant increases with centrality. This increase is larger for low centralities [@9r]. This has an important effect on both the size and the pattern of strangeness enhancement in our results. It explains why the departure from a linear increase of $\Xi$’s and $\Omega$’s (concave shape) seen in Figs. 1c and 1d is also more pronounced for lower centralities. It leads to the convexity in the centrality dependence of the yields of hyperons and antihyperons per participant in Figs. 2 (Note, however, a change of curvature for very peripheral collisions where the effect of final state interaction is negligible). This centrality pattern is a distinctive feature as well as a firm prediction of our approach.\ It is a pleasure to thank N. Armesto, A. Kaidalov, K. Redlich and Yu. Shabelski for discussions and C. Roy (STAR) and M.J. Tannenbaum (PHENIX) for information on the data. C.A.S. is supported by a Marie Curie Fellowship of the European Community program TMR (Training and Mobility of Researchers), under the contract number HPMF-CT-2000-01025.\ [**References**]{} \[1r\] WA97 coll., E. A. Andersen et al, Phys. Lett. [ **B433**]{}, 209 (1998) ; Phys. Lett. [**B449**]{}, 401 (1999).\ NA57 coll., N. Carrer, Nucl. Phys. [**A698**]{}, 118c (2002). \[2r\] NA49 coll., H. Appelshäuser et al, Phys. Lett. [ **B433**]{}, 523 (1998) ; Phys. Lett. [**B444**]{}, 523 (1998) ; Eur. Phys. J. [**C2**]{}, 661 (1998). \[3r\] A. Capella and C. A. Salgado, New Journal of Physics [**2**]{}, 30.1-30.4 (2000) ; Phys. Rev. [**C60**]{}, 054906 (1999).\ A. Capella, E. G. Ferreiro and C. A. Salgado, Phys. Lett. [**B459**]{}, 27 (1999) ; Nucl. Phys. [**A661**]{}, 502 (1999). \[4r\] A. Capella, U. Sukhatme, C.-I. Tan and J. Tran Thanh Van, Phys. Lett. [**81B**]{}, 68 (1979) ; Phys. Rep. [**236**]{}, 225 (1994). \[5r\] G. C. Rossi and G. Veneziano, Nucl. Phys. [**B123**]{}, 507 (1977). \[6r\] B. Z. Kopeliovich and B. G. Zakharov, Sov. J. Nucl. Phys. [**48**]{}, 136 (1988) ; Z. Phys. [**C43**]{}, 241 (1989). \[7r\] D. Kharzeev, Phys. Lett. [**B378**]{}, 238 (1996). \[8r\] A. Capella and B. Z. Kopeliovich, Phys. Lett. [**B381**]{}, 325 (1996). \[9rnew\] S. E. Vance and M. Gyulassy, Phys. Rev. Lett. [ **83**]{}, 1735 (1999). \[9r\] A. Capella and D. Sousa, Phys. Lett. [**B511**]{}, 185 (2001). \[10r\] G. H. Arakelyan, A. Capella, A. Kaidalov and Yu. M. Shabelski, hep-ph/0103337, to be published in Z. Phys. C. \[11r\] A. B. Kaidalov, Phys. Lett. [**116B**]{}, 459 (1982). \[13rnew\] E769 coll., E. M. Aitala et al., Phys. Lett. [**B496**]{}, 9 (2000). \[12r\] B. Koch, U. Heinz and J. Pitsut, Phys. Lett. [ **B243**]{}, 149 (1990). \[13r\] A. Capella, A. B. Kaidalov and D. Sousa, nucl-th/0105021, to be published in Phys. Rev. C. \[14rnew\] N. S. Amelin, N. Armesto, C. Pajares and D. Sousa, Eur. Phys. J. C [**22**]{}, 149 (2001). \[14r\] PHENIX coll., K. Adcox et al, nucl-ex/0112006. \[15r\] STAR coll., C. Roy, nucl-ex/111017. \[16r\] STAR coll., C. Adler et al, Phys. Rev. Lett. [**87**]{}, 262302 (2001). \[17r\] STAR coll., H. Caines, Nucl. Phys. [**A698**]{}, 112c (2002). \[18r\] STAR coll., C. Adler et al, Phys. Rev. Lett. [**86**]{}, 4778 (2001). \[20newr\] STAR coll., J. Castillo, Proc. Strangeness in Quark Mater 2001, Frankfurt (Germany) to be published. \[20r\] J. Ranft, A. Capella and J. Tran Thanh Van, Phys. Lett. [**B320**]{}, 346 (1994). \[21r\] PHENIX coll., K.Adcox et al, nucl-ex/0204007. 2 truecm **Figure Caption :** \(a) Calculated values of rapidity densities of $p$ (solid line), $\overline{p}$ (dashed line), and $p - \overline{p}$ (dotted line) at mid rapidities, $|y^*| < 0.35$, are plotted as a function of $dN^-/d\eta$, and compared with PHENIX data [@14r] ; (b) same for $\Lambda$ and $\overline{\Lambda}$ compared to preliminary STAR data [@16r] ; (c) same for $\Xi^-$ and $\overline{\Xi}^+$ ; (d) same for $\Omega$ and $\overline{\Omega}$.\ Calculated values of the ratios $B/n_{part}$ (a) and $\overline{B}/n_{part}$ (b), normalized to the same ratio for peripheral collisions ($n_{part} = 18$), are plotted as a function of $n_{part}$. **Fig. 1** **Fig. 2.a** **Fig. 2.b** [^1]: Unité Mixte de Recherche UMR n$^{\circ}$ 8627 - CNRS [^2]: For pions, we use the same fragmentation functions given in [@9r]. For simplicity, the same form is used for kaons. For $p\overline{p}$ pair production we take [@10r] $x D_{qq}^p(x) = x D_{qq}^{\overline{p}}(x) \sim (1 - x)^5$ and $x D_q^p(x) = x D_q^{\overline{p}}(x) \sim (1 - x)^3$. For the other baryon species an extra $\alpha_{\rho}(0) - \alpha_{\phi}(0) = 1/2$ is added to the power of $(1 - x)$ for each strange quark in the baryon [@10r]. [^3]: In the numerical calculations we neglect the first term of (\[5e\]) since the $DP$ component gives a very small contribution at $y^* \sim 0$ and RHIC energies – about 5% of the $DB$ one for the most central bin where its effect is maximal. [^4]: In order to conserve strangeness locally, we have added an extra $1/2 K^+$ and $1/2K^0$ to each produced net $\Lambda$ (plus $\Sigma$’s), an extra $K^+$ and $K^0$ to each net $\Xi$ and an extra $3/2 K^+$ and $3/2 K^0$ to each net $\Omega$. [^5]: Note, however, that a non-zero value of net omegas has been observed in $hA$ collisions [@13rnew]. This requires a non-vanishing contribution proportional to $I_3$. However, its effect in $AA$ collisions is presumably small since, in this case, the net omegas are almost entirely due to final state interactions. [^6]: To be precise, of all possible charge combinations in reactions (\[7e\]), we have only kept those involving the annihilation of a light $q$-$\overline{q}$ pair and production of an $s$-$\overline{s}$ in the $s$-channel. The other reactions, involving three quarks in the $t$-channel intermediate state, have substantially smaller cross-sections and have been neglected. All channels involving $\pi^0$ have been taken with cross-section $\sigma /2$ since only one of the $u\overline{u}$ and $d\overline{d}$ components of $\pi^0$ can participate to a given charge combination. For details see the second paper of [@3r]. [^7]: A Monte Carlo calculation in a similar framework with string fusion can be found in [@14rnew]. A net proton rapidity density of about 10 for central $Au$ $Au$ collisions at mid-rapidities at RHIC was first predicted in [@8r] using a stopping mechanism similar to the one considered here. [^8]: In ref. [@3r], the relative weights of net baryons were given by the factors $0.5(I_2 + I_3)$ – instead of $I_2$. In this case the values of the ratios are $\overline{p}/p = 0.70$, $\overline{\Lambda}/\Lambda = 0.71$, $\overline{\Xi}/\Xi = 0.76 $ and $\overline{\Omega}/\Omega = 0.78$. Their increase with the number of strange quarks in the baryon is smaller.

ArXiv

--- address: | ^1^MOE Key Lab of Computational Linguistics, School of EECS, Peking University\ ^2^Peng Cheng Laboratory, Shenzhen, China ^3^Beijing University of Posts and Telecommunications\ [email protected], zheng\[email protected], [email protected], [email protected] author: - 'Tianyu Liu^1^, Fuli Luo^1^, Pengcheng Yang^1^, Wei Wu^1^, Baobao Chang^1,2^' - | Zhifang Sui^1,2^\ ^1^MOE Key Lab of Computational Linguistics, School of EECS, Peking University\ ^2^Peng Cheng Laboratory, Shenzhen, China\ [{tianyu0421, luofuli, yang\_pc, wu.wei, chbb, szf}@pku.edu.cn]{} bibliography: - 'lrec2020.bib' title: 'HypoNLI: Exploring the Artificial Patterns of Hypothesis-only Bias in Natural Language Inference' --- Introduction ============ Natural language inference (NLI) (also known as recognizing textual entailment) is a widely studied task which aims to infer the relationship (e.g., *entailment*, *contradiction*, *neutral*) between two fragments of text, known as *premise* and *hypothesis* [@dagan2006pascal; @DBLP:series/synthesis/2013Dagan]. NLI models are usually required to determine whether a hypothesis is true (*entailment*) or false (*contradiction*) given the premise, or whether the truth value can not be inferred (*neutral*). A proper NLI decision should apparently rely on both the premise and the hypothesis. However, some recent studies [@gururangan2018annotation; @poliak2018hypothesis; @tsuchiya2018performance] have shown that it is possible for a trained model to identify the true label by only looking at the hypothesis without observing the premise. The phenomenon is referred to as annotation artifacts [@gururangan2018annotation], statistical irregularities [@poliak2018hypothesis] or partial-input heuristics [@feng2019misleading]. In this paper we use the term *hypothesis-only bias* [@poliak2018hypothesis] to refer to this phenomenon. Such hypothesis-only bias originates from the human annotation process of data collection. In the data collection process of many large-scale NLI datasets such as SNLI and MultiNLI , human annotators are required to write new sentences (hypotheses) based on the given premise and a specified label among *entailment*, *contradiction* and *neutral*. Some of the human-elicited hypotheses contain patterns that spuriously correlate to some specific labels. For example, 85.2% of the hypothesis sentences which contain the phrase *video games* were labeled as *contradiction*. The appearance of *video games* in hypothese can be seen as a stronger artificial indicator to the label *contradiction*. To get a deeper understanding of the specific bias captured by NLI models in the training procedure, we try to extract explicit surface patterns from the training sets of SNLI and MultiNLI, and show that the model can easily get decent classification accuracy by merely looking at these patterns. After that, we derive hard (adversarial) and easy subsets from the original test sets. They are derived based on the indication of the artificial patterns in the hypotheses. The gold labels of easy subsets are consistent with such indication while those of hard subsets are opposite to such indication. The model performance gap on easy and hard subsets shows to what extend a model can mitigate the hypothesis-only bias. After analyzing some competitive NLI models, including both non-pretrained models like Infersent [@conneau2017supervised], DAM [@parikh2016decomposable] and ESIM [@chen2016enhanced] and popular pretrained models like BERT [@devlin2018bert], XLNet [@yang2019xlnet] and RoBERTa [@liu2019roberta], we find that the hypothesis-only bias makes NLI models vulnerable to the adversarial (hard) instances which are against such bias (accuracy $<$ 60% on InferSent), while these models get much higher accuracy (accuracy $>$ 95% on InferSent) on the easy instances. This is an evidence to show that the NLI models might be over-estimated as they benefit a lot from the hints of artificial patterns. A straightforward way is to eliminate these human artifacts in the human annotation process, such as encouraging human annotators to use more diverse expressions or do dataset adversarial filtering [@zellers2018swag] and multi-round annotation [@nie2019adversarial]. However in this way, the annotation process would inevitably become more time-consuming and expensive. To this end, this paper explores two ways based on the derived artificial patterns to alleviate the hypothesis-only bias *in the training process*: down-sampling and adversarial debiasing. We hope they would serve as competitive baselines for other NLI debiasing methods. Down-sampling aims at reducing the hypothesis-only bias in NLI training sets by removing those instances in which the correct labels may easily be revealed by artificial patterns. Furthermore, we exploring adversarial debiasing methods [@belinkov2019don; @belinkov2018mitigating] for the sentence vector-based models in NLI [@yang2016hierarchical; @lin2017structured; @wu2018phrase; @luo2018leveraging]. The experiments show that the guidance from the derived artificial patterns can be helpful to the success of sentence-level NLI debiasing. Datasets ======== In this section, we identify the artificial patterns from the hypothesis sentences which highly correlate to specific labels in the training sets and then derive hard, easy subsets from the original test sets based on them. Artificial Pattern Collection {#extraction} ----------------------------- ‘Pattern’ in this work refers to (maybe nonconsecutive) word segments in the hypothesis sentences. We try to identify the ‘artificial patterns’ which spuriously correlate to a specific label due to certain human artifacts. We use $\mathrm{H}(M,t,\lambda)$ to represent a set of artificial patterns. $M$ and $t$ denotes the max length of the pattern and the max distance between two consecutive words in a pattern, respectively. For a artificial pattern $b\in \mathrm{H}(M,t,\lambda)$, there exists a specific label $l$ for $b$ that the conditional probability $\mathrm{p}(l|b)= count(b,l) / count(b)>\lambda$. For example, for the underlined pattern ‘A \# \# outdoors .’ in Table \[showpatterns\], the length of this pattern is 3, and the distance between the consecutive words ‘A’ and ‘outdoors’ is 2. Its conditional probability with the label *entailment* is 95.9%. Notably, all the recognized artificial patterns in our paper appear in at least 50 instances of the training sets to avoid misrecognition[^1]. In the rest of paper, unless otherwise specified, we set $M=3,t=3$ [^2]. By doing so, we only tune the hyper-parameter $\lambda$ in $\mathrm{H}$(3,3,$\lambda$) to decide using a rather strict (smaller $\lambda$) or mild (bigger $\lambda$) strategy while deriving the artificial patterns. Analysis of Hypothesis-only Bias {#causeofbias} -------------------------------- Previous work [@gururangan2018annotation; @poliak2018hypothesis] trained a sentence-based hypothesis-only classifier which achieves decent accuracy. Different from them, we show that in Table \[classifier\] the classifier which merely uses the artificial patterns as features achieves comparable performance with the fasttext [@Joulin2016FastText] classifier. Table \[classifier\] shows the classifier based on multi-word patterns with the default $M$ and $t$ (see Footnote \[fn:mandt\], $\mathrm{H}$(3,3,0.5)) achieves much higher accuracy than that based on only unigram patterns ($\mathrm{H}$(1,1,0.5)). -------------------------------------- -- -- -- **& **&\ & & Matched & Mismatched\ majority class & 34.3 & 35.4 & 35.2\ `fasttext` & 67.2 & 53.7 & 52.5\ Unigram & 60.2 & 49.8 & 49.6\ Pattern & 64.4 & 52.9 & 52.9\ **** -------------------------------------- -- -- -- : The accuracies of the hypothesis-only classifiers on SNLI test and MultiNLI dev sets. We train a MLP classifier with unigrams (Unigram) or multi-words patterns (Pattern) as features. Details in Sec \[causeofbias\].[]{data-label="classifier"} We also compare the test accuracies on the easy and hard sets (Sec \[hardandeasy\]) of the baseline models (I-9,D-9,E-9) in Table \[tb:train\_on\_snli\] and \[tb:train\_on\_mnli\]. Empirically we find that the NLI models achieve very high accuracy on the easy sets while performing poorly on the hard sets. We also observe the same tendency in the models trained on the randomly downsampled training sets (e.g. I-1, I-3, I-5, I-7, etc.). It shows that NLI models fit the artificial patterns in the training set very well, which makes them fragile to the adversarial examples (hard set) which are against these patterns. Thus we assume the artificial patterns contributes to the hypothesis-only bias. Hard and Easy Subsets {#hardandeasy} --------------------- Some instances contain artificial patterns that are strong indicators to the specific labels. We treat the instances in the test sets which are consistent with such indication as ‘easy’ ones and those instances which are against such indication as ‘hard’ ones. For easy subsets, the labels of *all* the artificial patterns in the specific hypotheses must be consistent with the gold labels. We show an easy instance below: the artificial patterns ‘The dogs are \# on’ and ‘bed .’ (bed is the last word of the sentence) are strong indicators to the correct classification. ------------------------------------------------------------------------------------------------------- **Premise**: Two cats playing on the bed together . **Hypothesis**: *The dogs are* playing *on* the `bed .` **Gold Label**: <span style="font-variant:small-caps;">Contradiction</span> **Artificial patterns**: (`bed .`, <span style="font-variant:small-caps;">Contradiction</span>, 83.2% ); (*The dogs are \# on*, <span style="font-variant:small-caps;">Contradiction</span>, 82.9%) ------------------------------------------------------------------------------------------------------- : Examples for easy and hard instances. The indications of artificial patterns are consistent with the gold label in the easy case while they are against the gold label in the hard case. The triple $(\mathbf{P}, l, p)$ show the related label indications $l$ for specific artificial patterns $\mathbf{P}$ and their conditional probability $p$. \ (a) An easy instance ----------------------------------------------------------------------------------------------------- **Premise**: A bare-chested man fitting his head and arm into a toilet seat ring while spectators watch in a city. **Hypothesis**: A gentleman with `no` chest hair , wrangles *his way* through a toilet seat . **Gold Label**: <span style="font-variant:small-caps;">Entailment</span> **Artificial patterns**: (`no`, <span style="font-variant:small-caps;">Contradiction</span>, 82.7%) (*his way*, <span style="font-variant:small-caps;">Neutral</span>, 82.4%) ----------------------------------------------------------------------------------------------------- : Examples for easy and hard instances. The indications of artificial patterns are consistent with the gold label in the easy case while they are against the gold label in the hard case. The triple $(\mathbf{P}, l, p)$ show the related label indications $l$ for specific artificial patterns $\mathbf{P}$ and their conditional probability $p$. \ (b) A hard instance For the hard subsets, on the other side, the indications of the artificial patterns should be *all* different from gold labels. We also show a hard instance below: in this situation, the artificial patterns ‘no’[^3] and ‘his way’ may misguide the NLI models to the wrong answers. Notably we do not put instances with conflicting indications (e.g. an instance with 2 artificial patterns, one of which has the same label with the gold label while the other does not) into easy or hard subsets to build more challenging adversarial examples. The sizes of hard and easy sets actually depend on how we harvest artificial pattern, i.e. $\lambda$ in $\mathrm{H}(M=3,t=3,\lambda)$ (Sec \[extraction\] and Footnote \[fn:mandt\]). For the sake of simplicity, we utilize $\lambda=0.8$ and $\lambda=0.7$ [^4] as the thresholds to derive easy and hard subsets for SNLI and MultiNLI respectively in the following experiments, as adopting a relatively bigger $\lambda$ can choose the instances which largely accord with the artificial patterns and are thus eligible to serve as adversarial examples. The sizes of easy and hard sets in SNLI test set, MultiNLI-matched dev set and MultiNLI-mismatched dev set are 327/1760; 410/1032; 371/1085 respectively. [^5] The performance of an ideally unbiased NLI model on the easy and hard sets should be close to each other. Besides we should not see huge gap between the model accuracy on the easy and hard subsets. Baselines {#sec:baseline} --------- We set up both pretrained and non-pretrained model baselines for the proposed adversarial datasets. We rerun their public available codebase with the default hyper-parameter and optimizer settings, including InferSent[^6], DAM[^7], ESIM[^8], BERT (uncased base), XLNet (cased base) and RoBERTa (base)[^9]. For BERT, XLNet and RoBERTa, we concatenate the premise sentence and hypothesis sentence with \[SEP\] token as the input. For output classifier, we use a linear mapping to transform the vector at the position of \[CLS\] token at the last layer of these pretrained models to a normalized 3-element vector (using softmax) which represents the scores for each label. We report the test accuracies on easy, hard subsets and the UW+CMU hard subsets [@gururangan2018annotation] which are derived from a hypothesis-only classifier. From Table \[tab:big\_baseline\], we can tell that the proposed hard sets are more challenging than UW+CMU hard subsets. Exploring Debiasing Methods =========================== Full Easy Hard UW+CMU ----------- ------ ------ ------ -------- InferSent 84.5 97.2 58.9 69.3 DAM 85.8 97.8 62.1 72.0 ESIM 87.6 97.7 68.2 75.2 BERT 90.5 98.2 71.2 80.3 XLNet 90.9 98.0 73.6 80.7 RoBERTa 91.7 98.9 75.8 82.7 : Model baselines on the proposed hard and easy test sets. For MultiNLI, we trained the models using matched dev sets as the valid set and reported the test accuracies on mismatched dev sets. ’UW+CMU’ refers to the adversarial set (Sec \[sec:baseline\]) detected by a neural based hypothesis-only classifier.[]{data-label="tab:big_baseline"} \ (a) Models trained on SNLI Full Easy Hard ----------- ------ ------ ------ -- InferSent 70.4 92.7 54.4 DAM 70.5 92.0 55.1 ESIM 76.7 93.9 65.6 BERT 83.4 95.2 75.0 XLNet 86.5 96.3 78.2 RoBERTa 87.2 96.5 81.4 : Model baselines on the proposed hard and easy test sets. For MultiNLI, we trained the models using matched dev sets as the valid set and reported the test accuracies on mismatched dev sets. ’UW+CMU’ refers to the adversarial set (Sec \[sec:baseline\]) detected by a neural based hypothesis-only classifier.[]{data-label="tab:big_baseline"} \ (b) Models trained on MultiNLI Down-sampling Baselines ----------------------- $\lambda$ No. Mode Full Easy Hard $\Delta^{Hard}_{Easy}(\downarrow)$ ----------- ----- -------- ---------- ---------- ---------- ------------------------------------ I-1 Rand **76.4** **93.8** 48.7 45.1 I-2 Debias 66.9 64.6 **56.0** 8.6 I-3 Rand **81.1** **96.1** 54.1 42.0 I-4 Debias 76.9 79.8 **58.0** 21.8 I-5 Rand **82.8** **96.9** 56.0 40.9 I-6 Debias 80.9 86.4 **** 24.9 I-7 Rand **83.5** **96.9** 56.6 40.3 I-8 Debias 82.9 90.4 **60.0** 30.4 1.0 I-9 All 58.9 38.3 : Model performance on the SNLI test set. We report the average scores of multiple independent runs. $\Delta^{Hard}_{Easy}$ represents the gap between hard and easy test sets (lower is better). $\lambda$ is the debiasing threshold. We use 2 ‘modes’ to down-sample the training sets, namely biased instances removing (‘Debias’) and randomly downsampling (‘Rand’), the latter has the same training size and label distribution with with ‘Debias’ mode for a fair comparison. The downsampled training sizes are 4.0%, 19.8%, 43.8%, 67.4% and 100% of the whole training size (549867) for $\lambda\in\{0.5,0.6,0.7,0.8,1.0\}$ respectively. Note that when $\lambda$=1.0, we use the whole training set without any downsampling. []{data-label="tb:train_on_snli"} \ (a) InferSent trained on SNLI $\lambda$ No. Mode Full Easy Hard $\Delta^{Hard}_{Easy}(\downarrow)$ ----------- ----- -------- ---------- ---------- ---------- ------------------------------------ D-1 Rand **74.1** **92.3** 46.7 45.6 D-2 Debias 67.5 67.8 **53.3** 14.5 D-3 Rand **82.4** **96.3** 56.7 39.6 D-4 Debias 79.3 84.4 **62.1** 21.3 D-5 Rand **84.4** **97.3** 59.4 37.9 D-6 Debias 83.0 89.6 26.5 D-7 Rand **85.3** 60.1 37.7 D-8 Debias 84.6 93.5 **62.6** 30.9 1.0 D-9 All 62.1 35.7 : Model performance on the SNLI test set. We report the average scores of multiple independent runs. $\Delta^{Hard}_{Easy}$ represents the gap between hard and easy test sets (lower is better). $\lambda$ is the debiasing threshold. We use 2 ‘modes’ to down-sample the training sets, namely biased instances removing (‘Debias’) and randomly downsampling (‘Rand’), the latter has the same training size and label distribution with with ‘Debias’ mode for a fair comparison. The downsampled training sizes are 4.0%, 19.8%, 43.8%, 67.4% and 100% of the whole training size (549867) for $\lambda\in\{0.5,0.6,0.7,0.8,1.0\}$ respectively. Note that when $\lambda$=1.0, we use the whole training set without any downsampling. []{data-label="tb:train_on_snli"} \ (b) DAM trained on SNLI $\lambda$ No. Mode Full Easy Hard $\Delta^{Hard}_{Easy}(\downarrow)$ ----------- ----- -------- ---------- ---------- ---------- ------------------------------------ E-1 Rand **76.8** **94.6** 48.9 45.7 E-2 Debias 65.3 62.5 **53.8** 7.7 E-3 Rand **83.6** **96.6** 62.2 34.4 E-4 Debias 78.6 79.4 **63.6** 15.8 E-5 Rand **85.9** **97.2** 64.2 35.0 E-6 Debias 83.8 88.2 19.4 E-7 Rand **86.9** **97.3** 67.9 29.4 E-8 Debias 86.2 92.1 **70.9** 21.3 1.0 E-9 All 68.2 29.4 : Model performance on the SNLI test set. We report the average scores of multiple independent runs. $\Delta^{Hard}_{Easy}$ represents the gap between hard and easy test sets (lower is better). $\lambda$ is the debiasing threshold. We use 2 ‘modes’ to down-sample the training sets, namely biased instances removing (‘Debias’) and randomly downsampling (‘Rand’), the latter has the same training size and label distribution with with ‘Debias’ mode for a fair comparison. The downsampled training sizes are 4.0%, 19.8%, 43.8%, 67.4% and 100% of the whole training size (549867) for $\lambda\in\{0.5,0.6,0.7,0.8,1.0\}$ respectively. Note that when $\lambda$=1.0, we use the whole training set without any downsampling. []{data-label="tb:train_on_snli"} \ (c) ESIM trained on SNLI $\lambda$ No. Mode Full Easy Hard $\Delta^{Hard}_{Easy}(\downarrow)$ ----------- ----- -------- ---------- ---------- ---------- ------------------------------------ I-1 Rand **67.5** **92.6** 50.9 41.7 I-2 Debias 64.2 76.0 **56.1** 19.9 I-3 Rand **69.0** **92.7** 53.4 39.3 I-4 Debias 67.5 80.9 21.2 I-5 Rand **69.1** 52.2 40.8 I-6 Debias 68.3 84.6 **57.1** 27.5 I-7 Rand 69.2 **92.5** 52.5 40.0 I-8 Debias **69.6** 91.4 **53.6** 37.8 1.0 I-9 All 92.7 54.4 38.3 : Models performance on MultiNLI mismatched dev set. We tune the models on MultiNLI matched dev set. The training sizes are 24.4%, 53.1%, 68.0%, 81.2% and 100% of the whole training size (392702) for $\lambda\in\{0.5,0.6,0.7,0.8,1.0\}$ respectively. Note that we do not report the scores on MultiNLI test sets as they are unable to access. The **bold** numbers mark higher scores between ‘Rand’ and ‘Debias’ mode for each $\lambda$. The numbers highlight the highest scores in each column.[]{data-label="tb:train_on_mnli"} \ (a) InferSent trained on MultiNLI $\lambda$ No. Mode Full Easy Hard $\Delta^{Hard}_{Easy}(\downarrow)$ ----------- ----- -------- ---------- ---------- ---------- ------------------------------------ D-1 Rand **64.4** **91.9** 46.2 45.7 D-2 Debias 61.9 77.4 **52.0** 22.4 D-3 Rand **68.3** **91.8** 52.5 39.3 D-4 Debias 65.8 79.0 21.0 D-5 Rand **69.6** 52.2 40.4 D-6 Debias 68.0 83.9 **57.5** 25.4 D-7 Rand **70.0** **92.4** 53.8 38.6 D-8 Debias 69.6 91.6 **54.9** 26.7 1.0 D-9 All 92.0 55.1 36.9 : Models performance on MultiNLI mismatched dev set. We tune the models on MultiNLI matched dev set. The training sizes are 24.4%, 53.1%, 68.0%, 81.2% and 100% of the whole training size (392702) for $\lambda\in\{0.5,0.6,0.7,0.8,1.0\}$ respectively. Note that we do not report the scores on MultiNLI test sets as they are unable to access. The **bold** numbers mark higher scores between ‘Rand’ and ‘Debias’ mode for each $\lambda$. The numbers highlight the highest scores in each column.[]{data-label="tb:train_on_mnli"} \(b) DAM trained on MultiNLI $\lambda$ No. Mode Full Easy Hard $\Delta^{Hard}_{Easy}(\downarrow)$ ----------- ----- -------- ---------- ---------- ---------- ------------------------------------ E-1 Rand **70.4** **92.8** 59.0 33.8 E-2 Debias 67.0 75.2 **63.6** 11.6 E-3 Rand **73.9** **93.8** 61.6 22.2 E-4 Debias 73.2 84.8 18.4 E-5 Rand 74.6 **92.8** 64.5 28.3 E-6 Debias **74.9** 89.1 **65.9** 23.2 E-7 Rand 75.7 64.4 29.6 E-8 Debias **75.8** 93.4 **65.0** 28.4 1.0 E-9 All 93.9 65.6 28.3 : Models performance on MultiNLI mismatched dev set. We tune the models on MultiNLI matched dev set. The training sizes are 24.4%, 53.1%, 68.0%, 81.2% and 100% of the whole training size (392702) for $\lambda\in\{0.5,0.6,0.7,0.8,1.0\}$ respectively. Note that we do not report the scores on MultiNLI test sets as they are unable to access. The **bold** numbers mark higher scores between ‘Rand’ and ‘Debias’ mode for each $\lambda$. The numbers highlight the highest scores in each column.[]{data-label="tb:train_on_mnli"} \ (c) ESIM trained on MultiNLI Sec \[causeofbias\] verifies that the artificial patterns lead to correct hypothesis-only classification, which motivates us to remove such patterns in the training sets by down-sampling. Specifically we down-sample the training sets of SNLI and MultiNLI and retrain 3 prevailing NLI models: InferSent, DAM and ESIM. ### Downsampling Details We down-sampled the training sets by removing the biased instances (‘Debias’ mode) that contain the artificial patterns. **Choosing down-sampling threshold $\lambda$**: The threshold $\lambda$ is exactly the same $\lambda$ defined in Sec \[extraction\]. We consider a training instance as a biased one even if it contains only one artificial pattern. When adopting smaller $\lambda$, we harvest more artificial patterns as described in Sec \[extraction\]. Accordingly more training instances would be treated as biased ones and then filtered. In a word, smaller $\lambda$ represents more strict down-sampling strategy in terms of filtering the artificial patterns. $\lambda=0.5$ serves as the lower bound because the highest pattern-label conditional probability ($\mathrm{p}(l|b)$ in Sec \[extraction\]) for premises, which aren’t observed the same bias as hypotheses, is less than 0.5 in both SNLI and MultiNLI training set.. **Ruling out the effects of training size**: The model performance might be highly correlated with the size of training set. To rule out the effects of training size as much as possible, we set up randomly down-sampled training sets (‘rand’ mode) with the same size as the corresponding ‘debias’ mode under different $\lambda$ for a fair comparison. **Keeping the label distribution balanced**: After removing the biased instances in the training set by different $\lambda$ (‘debias’ mode), suppose we get $n_1, n_2, n_3$ ($n_1 \ge n_2 \ge n_3$) instances for the 3 pre-defined labels of NLI in the down-sampled training set. Then we down-sample the subsets with $n_1, n_2$ instances to $n_3$ instances and get a dataset with $3n_3$ instances. For the corresponding ‘rand’ mode, we sample $n_3$ instances for each pre-defined label from training set. **Convincing scores of multiple runs**: To relieve the randomness of randomly down-sampling and model initialization, for the ‘rand’ mode in Table \[tb:train\_on\_snli\] and \[tb:train\_on\_mnli\], we firstly randomly down-sample the training set (with the label distribution balanced) according to different $\lambda$ for 5 times and get 5 randomly down-sampled training sets for each $\lambda$. Then for each down-sampled training set, we run 3 independent experiments with random model initialization under the same experimental settings. So each score in the ‘rand’ mode of Table \[tb:train\_on\_snli\] and \[tb:train\_on\_mnli\] comes from 15 independent runs. The scores in the ‘debias’ mode of Table \[tb:train\_on\_snli\] and \[tb:train\_on\_mnli\] come from 5 independent runs with random model initialization. ### Discussions {#implement} From table \[tb:train\_on\_snli\] and \[tb:train\_on\_mnli\], we observe that: 1) The NLI models fit the bias patterns in the hypotheses very well even in the small-scale randomly down-sampled training sets (I-1, D-1 and E1) which only accounts for 4.0% of the original training set (SNLI), as the performance gaps between easy and hard subsets in these settings are still huge ($>$40% for SNLI in Table \[tb:train\_on\_snli\]). 2\) Under the same $\lambda$, the proposed ‘debias’ down-sampling not only outperforms its ‘rand’ counterpart in terms of hard subsets, but also greatly reduce the performance gap on easy and hard sets. 3\) The gains on hard sets on MultiNLI are smaller than those on SNLI as MultiNLI is less biased regarding the pattern-label conditional probability (Table \[showpatterns\]). Down-sampling achieves larger gains on more biased datasets. In SNLI, the ‘debias’ down-sampling even outperforms the baseline models (I-8 vs I-9, D-8 vs D-9, E-8 vs E-9), which is really impressive as the training size of I-8, I-8 and E-8 is only 67.5% of the baseline models. [@gururangan2018annotation] expressed concerns upon down-sampling (DS) methods: 1) Will removing the artificial patterns cause new artifacts? (e.g. removing the word ‘no’, which is a strong indicator for *contradiction* may leave the remaining dataset with this word mostly appearing in the *neutral* or *entailment* classes thus create new artifact) and 2) Will DS methods prevent the models to learn specific inference phenomena (e.g. ‘animal’ is a hypernym of ‘dog’)? First of all, different from [@gururangan2018annotation] which only considered unigram patterns, our artificial patterns are mostly multi-word patterns rather than unigram patterns as the former usually has larger concurrent probability $\mathrm{p}(l|b)$ as shown in Table \[showpatterns\]. Our intention is to use the multi-word patterns to capture the specific ways of expression (human artifacts), rather than single words, of the human annotators. For the first concern, instead of filtering the unigram ‘no’, we prefer removing multi-word patterns which contain ‘no’, such as ‘There are no’ or ‘no \# on’ for MultiNLI as shown in Table \[showpatterns\]. For the hypernym mentioned in the second concern, as we prefer filtering multi-word patterns like ‘The dogs are \# on’, we would not deliberately filter the unigram ‘dog’ unless adopting very aggressive DS strategy ($\lambda=0.5$) in both SNLI and MultiNLI. Adversarial Debiasing --------------------- Since the hypothesis-only bias comes solely from the hypothesis sentence, we wonder if it is possible to get rid of these biases via debiasing the hypothesis sentence vector. More specifically, we focus on the ‘sentence vector-based models’ [^10] category as defined on SNLI’s web page[^11]. Notably the idea of debiasing NLI via adversarial training has been proposed before [@belinkov2019don; @belinkov2018mitigating]. We hereby briefly introduce how we implement our adversarial training and how we incorporate instance reweighting method in this framework. In the following experiments, we use the full training sets without any down-sampling. We use the InferSent [@conneau2017supervised] (biLSTM with max pooling) model as the benchmark sentence encoder. ![The illustration of the sentence-level debiasing framework, which is elaborated in Sec \[sec:adver\_intro\]. []{data-label="adversarial"}](adversarial.pdf){width="0.8\linewidth"} ### Adversarial Debiasing Framework {#sec:adver_intro} As shown in Fig \[adversarial\], given the outputs $s_h=\mathrm{E}_h(x_h), s_p=\mathrm{E}_s(x_s)$ of hypothesis and premise encoders $\mathrm{E}_h, \mathrm{E}_p$, we are interested in predicting the NLI label $y$ using a classifier $\mathrm{C}$, $\mathrm{p}_{\mathrm{C}}(y|s_h=\mathrm{E}_h(x_h), s_p=\mathrm{E}_s(x_s))$. In addition, we train a hypothesis-only discriminator trying to predict the correct label $y$ solely from the hypothesis sentence representation $s_h$ by modeling $\mathrm{p}_{\mathrm{D}}(y|s_h=\mathrm{E}_h(x_h))$. We formulate the training process in the adversarial setting by a min-max game. Specifically we train the discriminator $\mathrm{D}$ to predict the label using only hypothesis sentence vector. Additionally we train the sentence encoder $\mathrm{E}_h$, $\mathrm{E}_p$ and the classifier $\mathrm{C}$ to fool the discriminator $\mathrm{D}$ without hurting inference ability. $\gamma$ is a hyper-parameter which controls the degree of debiasing. $$\begin{aligned} \min_{\mathrm{E}_h,\mathrm{E}_p,\mathrm{C}} \max_{\mathrm{D}} \mathbbm{E}_{x_h,x_p,y \sim \mathrm{p}(X,Y)}[\gamma \log \mathrm{p}_\mathrm{D}(y|\mathrm{E}_h(x_h)) - \\ \log \mathrm{p}_\mathrm{C}(y|\mathrm{E}_h(x_h), \mathrm{E}_p(x_p))] \label{eq:minmax} \end{aligned}$$ We train the encoders, discriminator and classifier in Eq \[eq:minmax\] together with a gradient reversal layer [@ganin2016domain] as shown in Fig \[adversarial\]. We negate the gradients from the discriminator $\mathrm{D}$ (red arrow in Fig \[adversarial\]) to push the hypothesis encoder $\mathrm{E}_h$ to the opposite direction while update its parameters. The usage of gradient reversal layer makes it easier to optimize the min-max game in Eq \[eq:minmax\] [@xie2017controllable; @chen2018adversarial] than training the two adversarial components alternately like Generative Adversarial Nets (GANs) [@goodfellow2014generative]. We update the model parameters $\theta$ by gradient descending ($m$ is the batch size): $$\small \theta_{\mathrm{D}}^{new} = \theta_{\mathrm{D}}^{old} - \frac{1}{m} \sum_{i=1}^m \nabla_{\theta_{\mathrm{D}}} [\log \mathrm{p}_{\mathrm{D}}(y^i|\mathrm{E}_h(x_h^i))] \label{eq:d}$$ $$\small \theta_{\mathrm{C}}^{new} = \theta_{\mathrm{C}}^{old} - \frac{1}{m} \sum_{i=1}^m \nabla_{\theta_{\mathrm{C}}}[\log \mathrm{p}_{\mathrm{C}}(y^i|\mathrm{E}_h(x_h^i),\mathrm{E}_p(x_p^i))] \label{eq:c}$$ $$\small \begin{aligned} \theta_{\mathrm{E}_h}^{new} = \theta_{\mathrm{E}_h}^{old} - \frac{1}{m} \sum_{i=1}^m \nabla_{\theta_{\mathrm{E}_h}}[\log \mathrm{p}_{\mathrm{C}}(y^i|\mathrm{E}_h(x_h^i),\mathrm{E}_p(x_p^i))] \\ \underbrace{+ \frac{\gamma}{m} \sum_{i=1}^m \nabla_{\theta_{\mathrm{E}_h}}[\log \mathrm{p}_{\mathrm{D}}(y^i|\mathrm{E}_h(x_h^i))] }_{\textbf{gradient reverse}} \end{aligned} \label{eq:theta_eh}$$ ### Guidance from Artificial Patterns {#sec:guidance} The artificial patterns turns out to be useful guidances for both the discriminator $\mathrm{D}$ and the classifier $\mathrm{C}$ as they indicate whether an instance is biased or not. We thus reweight the training instances in the training set based on the division of ‘biased’ and ‘unbiased’ training subsets. **Guidance for Discriminator**: During the adversarial process, we optimize the discriminator $\mathrm{D}$ by maximizing the log likelihood loss like Eq \[eq:d\]. We find increasing the weights of the biased instances in the training set is of great help to the adversarial debiasing model. Because in this way, the discriminator can learn more from the biased instances to better fit the hypothesis-only bias. The whole adversarial debiasing training process could benefit from a stronger hypothesis-only discriminator. Formally, we replace negative log likelihood loss function in Eq \[eq:d\] with a weighted loss function: $$\small \begin{aligned} \frac{1}{m} \sum_{i=1}^{m} [\mathbbm{1}\{(x^i,y^i) \in \mathcal{D}_{unbias}\} \log \mathrm{p}_{\mathrm{D}}(y^i|\mathrm{E}_h(x_h^i)) + \\ \alpha_1 * \mathbbm{1}\{(x^i,y^i) \in \mathcal{D}_{bias}\} \log \mathrm{p}_{\mathrm{D}}(y^i|\mathrm{E}_h(x_h^i))] \end{aligned} \label{eq:d-reweight}$$ , where $\mathcal{D} = \mathcal{D}_{unbias} \cup \mathcal{D}_{bias}$ denotes the whole training corpus. The division of biased and unbiased training subsets depends on the debiasing threshold $\lambda$ (just like the down-sampling threhold in Table \[tb:train\_on\_snli\] and \[tb:train\_on\_mnli\]). $\alpha_1 \geq 1$ is a hyper-parameter which reflects the attention on biased instances for the hypothesis-only discriminator. **Guidance for Classifier**: Similar to the re-weighting method in Eq \[eq:d-reweight\], we also apply the re-weighting strategy on the parameter learning for the inference classifier in Eq \[eq:c\]. We hope the classifier can capture the concrete semantics in NLI instead of over-fitting the artificial patterns in the hypotheses. Thus we increase the weights of the unbiased training subset in the loss function of Eq \[eq:c\]. $$\small \begin{aligned} \frac{1}{m} \sum_{i=1}^{m} [\mathbbm{1}\{(x^i,y^i) \in \mathcal{D}_{bias}\} \log \mathrm{p}_{\mathrm{C}}(y^i|\mathrm{E}_h(x_h^i),\mathrm{E}_h(x_p^i)) + \\ \alpha_2 * \mathbbm{1}\{(x^i,y^i) \in \mathcal{D}_{unbias}\} \log \mathrm{p}_{\mathrm{C}}(y^i|\mathrm{E}_h(x_h^i),\mathrm{E}_h(x_p^i))] \end{aligned} \label{eq:c-reweight}$$ , where $\alpha_2 \geq 1$ is a threshold to control the attention the models pay on the unbiased instances. ### Training Details {#sec:traindetails} Apart from the weighted loss functions guided by the artificial patterns, we also investigate the following two techniques in the adversarial training process. **Multiple discriminators**: The min-max game in Eq \[eq:minmax\] could benefit from stronger discriminators. So we try $k \in \{1,2,3\}$ discriminators to enhance its ability to do hypothesis-only classifications. In our experiments, we find that $k=2$ is the best configuration for the discriminator. **Dynamic reweighting**: For hyper-parameter $\alpha$ ($\alpha_1$ and $\alpha_2$ in Eq \[eq:d-reweight\] and Eq \[eq:c-reweight\] respectively), we find it useful to adjust $\alpha$ dynamically in the training process. $\alpha^{0}$ and $\alpha^t$ represents the initial values we set before training and its value after $t$ training iterations respectively. Additionally we set up a hyper-parameter $\phi$ to control the gap of model accuracies $\delta$ on the easy and hard subsets in the dev set. $$\alpha^{t+1}= \left\{ \begin{array}{lr} \max(\alpha^t+\epsilon, \alpha^0), & \delta \geq \phi \\ \max(\alpha^t-\epsilon, \alpha^0), & \delta < \phi \\ \end{array} \right.$$ where $\epsilon$ is a hyper-parameter set as 0.5 for models trained on both datasets. Besides, we set $\phi$ as 0.15 and 0.10 for SNLI and MultiNLI respectively. Notably although we update the hyper-parameters $\alpha_1$ and $\alpha_2$ dynamically in different iterations based on $\phi$, we still select the model which has the best performance on the dev sets as the best model in each run. **Parameter settings**: We use grid search to find the best hyper-parameter settings: $\alpha_1,\alpha_2 \in \{1,3,5,10\}$, $\gamma \in \{0.5,1,3,5\}$ in Eq \[eq:d-reweight\], \[eq:c-reweight\] and Eq \[eq:minmax\]. We also try $\lambda \in \{0.5,0.6,0.7,0.8\}$ as the threshold to split $\mathcal{D}_{bias}$ and $\mathcal{D}_{unbias}$ in Eq \[eq:d-reweight\] and \[eq:c-reweight\]. Specifically, we treat the instances which contain the artificial patterns in $\mathrm{H}$(3,3,$\lambda$) (Sec \[extraction\], Footnote \[fn:mandt\]) as $\mathcal{D}_{bias}$, and the remaining instances as $\mathcal{D}_{unbias}$. For the results in Table \[tb:train\_on\_adv\], we set $\gamma=3$ and $\gamma=1$ for SNLI and MultiNLI respectively. For both datasets, we set $\alpha_1=5,\alpha_2=5$ as well as $\lambda=0.7$ as the threshold for separating the biased and unbiased subsets in Eq \[eq:d-reweight\] and \[eq:c-reweight\]. For a fair comparison, we do not tune any hyper-parameter in the InferSent encoder, the learning rate and the optimizer setting. The results of ‘dInferSent’ and its variations in Table \[tb:train\_on\_adv\] comes from 5 independent runs with random initialization. **Model** **Full** **Easy** **Hard** $\Delta^{Hard}_{Easy}(\downarrow)$ ----------------------------- ---------- ---------- ---------- ------------------------------------ InferSent 58.9 38.3 InferSent+DS($\lambda$=0.8) 82.9 90.4 60.0 30.4 InferSent+Guidance 84.1 95.5 **61.7** 33.8 dInferSent 81.6 **92.5** 59.9 32.6 +Guidance **82.2** 86.9 63.3 23.6 +Guidance+Reweight 80.9 78.2 10.9 : The comparison of InferSent (baseline), InferSent+DS (downsampling) and dInferSent (adversarial debiasing) on SNLI test set and MultiNLI mismatched dev set respectively. We choose the down-sampling (DS) method with $\lambda=0.8$ because it performs best on the hard subsets. The ‘Guidance’ and ‘Reweight’ methods are elaborated in Sec \[sec:guidance\] and Sec \[sec:traindetails\] respectively. []{data-label="tb:train_on_adv"} \ (a) InferSent trained on SNLI **Model** **Full** **Easy** **Hard** $\Delta^{Hard}_{Easy}(\downarrow)$ ----------------------------- ---------- ---------- ---------- ------------------------------------ InferSent 54.4 38.3 InferSent+DS($\lambda$=0.8) 69.9 91.4 53.6 37.8 InferSent+Guidance 70.1 92.1 **54.9** 37.2 dInferSent **68.8** **91.1** 54.7 36.4 +Guidance 68.0 87.9 55.3 32.6 +Guidance+Reweight 66.5 79.4 20.6 : The comparison of InferSent (baseline), InferSent+DS (downsampling) and dInferSent (adversarial debiasing) on SNLI test set and MultiNLI mismatched dev set respectively. We choose the down-sampling (DS) method with $\lambda=0.8$ because it performs best on the hard subsets. The ‘Guidance’ and ‘Reweight’ methods are elaborated in Sec \[sec:guidance\] and Sec \[sec:traindetails\] respectively. []{data-label="tb:train_on_adv"} \ (b) InferSent trained on MultiNLI ### Discussions {#discussions} From Table \[tb:train\_on\_adv\], we observe that although the performance gap between the easy and hard subsets is reduced to some extent by the vanilla dInferSent models in both SNLI and MultiNLI. The model still does not reach our expectation to lower the gap between hard and easy sets. We assume this is because the denoising discriminator in Fig \[adversarial\] somewhat impedes the inference ability of the NLI models as it may disturb the hypothesis sentence encoder especially when the sentences do not contain hypothesis-only bias. The explicit guidance (‘+Guidance’) from the artificial patterns alleviates this issue in both datasets as in this way the discriminator pays more attention on the potentially biased instances thus has smaller influence on the hard instances in the training procedure. These models achieve higher accuracies on the hard subset than the baseline models in both datasets. The ‘reweight’ trick in Sec \[sec:traindetails\] greatly reduces the performance gap between the easy and hard sets as it dynamically adjusts the debiasing strategies (i.e. the weight of training instances in Eq \[eq:d-reweight\] and \[eq:c-reweight\]). Related Work ============ The bias in the data annotation exists in many tasks, e.g. lexical inference [@levy2015supervised], visual question answering [@goyal2017making], ROC story cloze [@DBLP:conf/acl/CaiTG17] etc. The NLI models are shown to be sensitive to the compositional features in premises and hypotheses [@nie2018analyzing], data permutations [@schluter2018data; @wang2018if] and vulnerable to adversarial examples [@iyyer2018adversarial; @minervini2018adversarially; @GlocknerSG18] and crafted stress test [@geiger2018stress; @DBLP:conf/coling/NaikRSRN18]. [@RudingerMD17] showed hypothesis in SNLI has the evidence of gender, racial and religious stereotypes, etc. [@sanchez2018behavior] analysed the behaviour of NLI models and the factors to be more robust. [@feng2019misleading] discussed how to use partial-input baseline (hypothesis-only classifier in NLI) in future dataset creation. [@clark2019don] uses an ensemble-based method to mitigate known bias. The InferSent model, which served as an important baseline in this paper, are found to achieve superb performance on SNLI by word-level heuristics [@dasgupta2018evaluating]. [@maccartney2009natural] first revealed the difficulties of natural language inference model with bag-of-words models. Different from the artificial patterns we used in this paper, other artifact evidence includes sentence occurrence [@zhang2019selection], syntactic heuristics between hypotheses and premises [@mccoy2019right] and black-box clues derived from neural models [@gururangan2018annotation; @poliak2018hypothesis; @he2019unlearn]. The adversarial debiasing training proposed in this paper is inspired by the success of Generative Adversarial Networks (GANs) [@goodfellow2014generative]. Several works on learning encoders which are invariant to certain properties of text and image [@chen2018adversarial; @zhang2017aspect; @xie2017controllable; @moyer2018evading; @jaiswal2018unsupervised] in the adversarial settings. Conclusion ========== In this study, we show that the hypothesis-only bias in trained NLI models mainly comes from unevenly distributed surface patterns, which could be used to identify hard and easy instances for more convincing re-evaluation on currently overestimated NLI models. The attempts to mitigate the bias are meaningful as such bias not only makes NLI models fragile to adversarial examples. We try to mitigate this bias by removing those artificial patterns in the training sets, with experiments showing that it is a feasible way to alleviate the bias under proper down-sampling methods. We also show that adversarial debiasing with the guidance from the harvested artificial patterns is a feasible way to mitigate the hypothesis-only bias for sentence vector-based NLI models. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank the anonymous reviewers for their valuable suggestions. This work is supported by the National Science Foundation of China under Grant No. 61751201, No. 61772040, No. 61876004. The corresponding authors of this paper are Baobao Chang and Zhifang Sui. Bibliographical References {#main:ref} ========================== Language Resource References {#lr:ref} ============================ [^1]: Suppose a pattern only appears once in a training instance, its $\mathrm{p}(l|b)$ always equals 1 for the label in that instance. [^2]: We also tried larger $M$ and $t$, e.g. 4 or 5, but did not observe considerable changes of the artificial patterns, e.g. 95.4% patterns in H(5,5,0.5) are covered by H(3,3,0.5).\[fn:mandt\] [^3]: ‘no’ is different from ‘No’ shown in Table \[showpatterns\] as the latter indicates the word appears in the beginning of the sentence. [^4]: MultiNLI’s pattern-label conditional possibilities are generally smaller than those of SNLI as shown in Table \[showpatterns\]. So we use smaller $\lambda$ to ensure the size of derived subsets. [^5]: The datasets used in this paper can be found in https://tyliupku.github.io/publications/ [^6]: https://github.com/facebookresearch/InferSent [^7]: https://github.com/harvardnlp/decomp-attn [^8]: https://github.com/coetaur0/ESIM [^9]: https://github.com/huggingface/transformers [^10]: It would be more challenging to manipulate the gradients in the non-sentence vector-based models, e.g. models which contain interactions between hypothesis and premise sentence encoders like [@chen2017neural]. We leave this to the future work. [^11]: https://nlp.stanford.edu/projects/snli/

ArXiv

--- abstract: 'We consider self-similar solutions with finite mass to Smoluchowski’s coagulation equation for rate kernels that have homogeneity zero but are possibly singular such as Smoluchowski’s original kernel. We prove pointwise exponential decay of these solutions under rather mild assumptions on the kernel. If the support of the kernel is sufficiently large around the diagonal we also proof that $\lim_{x \to \infty} \frac{1}{x} \log \Big( \frac{1}{f(x)}\Big)$ exists. In addition we prove properties of the prefactor if the kernel is uniformly bounded below.' author: - 'B. Niethammer[^1] and J. J. L. Velázquez[^2]' bibliography: - '../coagulation.bib' title: 'Exponential tail behaviour of self-similar solutions to Smoluchowski’s coagulation equation' --- Introduction {#S.introduction} ============ In this article we study the decay properties of self-similar solutions to Smoluchowski’s classical coagulation equation [@Smolu16]. This model describes the binary collision of clusters of size $\xi >0$ where the rate of coagulation of clusters of size $\xi$ and $\eta$ is governed by a nonnegative and symmetric rate kernel $K=K(\xi,\eta)$. Then the number density $n(t,\xi)$ of clusters of size $\xi$ at time $t$ is determined by the rate equation $$\label{smolu1} \partial_t n(t,\xi)=\frac 1 2 \int_0^\xi K(\eta,\xi-\eta) n(t,\xi-\eta)n(t,\eta)\,d\eta - \int_0^{\infty} K(\xi,\eta) n(t,\xi)n(t,\eta)\,d\eta \,.$$ The precise form of $K$ is determined by the microscopic details of the coagulation process. In [@Smolu16] Smoluchowski derived the kernel $K(\xi,\eta)=K_0\big(\xi^{1/3} + \eta^{1/3}\big) \big( \xi^{-1/3} + y^{-1/3}\big)$ in the case that clusters are approximately balls in three dimensions, diffuse by Brownian motion when they are well separated and coagulate quickly when they get close. The coagulation equation has since been used in many different application areas, such as aerosol physics, astronomy or population dynamics, and correspondingly one finds a large variety of different kernels in the respective literature (see e.g. [@Drake72; @Aldous99]). It is well-known by now that the behaviour of solutions to depends on the growth behaviour of $K$ for large clusters. More precisely, solutions to only conserve mass for all times if the kernel $K$ grows at most linearly at infinity. If $K$ grows faster than linearly at infinity, the phenomenon of gelation occurs, that is, roughly speaking, infinitely large clusters are created at finite time and as a consequence, there is a finite time after which the first moment of $n$ decreases [@McLeod62a; @Jeon98; @EMP02]. In this article we are not concerned with the gelation phenomenon, but will consider kernels that grow slower than linearly. In this case it has been proved for a large class of kernels [@Norris99; @LauMisch02; @LauMisch04] that for initial data $n_0$ with finite first moment there exists a unique global solution to that conserves the mass for all times. A topic of particular interest in the analysis of coagulation equations is the long-time behaviour of solutions. Since most kernels $K$ are homogeneous one expects due to a corresponding scale invariance of equation , that solutions of converge to a unique self-similar solution as $t \to \infty$. This so-called scaling hypothesis is however so far only well-understood for the so-called solvable kernel, that is for $K=const., K=\xi+\eta$ and $K=\xi\eta$, for which explicit solution formulas are available. A complete analysis of the large-time behaviour of solutions for those kernels has been given in [@MePe04]. More precisely, it has been shown that has a unique self-similar solution with finite mass and in addition a family of self-similar solutions with infinite mass, so-called fat-tail solutions. The domains of attraction of each of these self-similar solutions has been characterized. Roughly speaking, initial data with a certain decay behaviour at infinity produce solutions that converge to the self-similar solution with the same decay behaviour. For non-solvable kernels much less is understood about the large-time behaviour and even to understand self-similar solutions turns out to be a formidable task. Only quite recently, the first existence results of self-similar solutions for homogeneous kernels with degree $\gamma<1$ have become available [@FouLau05; @EMR05] and certain properties of the corresponding self-similar profiles have been established [@FouLau06a; @EsMisch06; @CanMisch11; @NV11b]. Also, the existence of self-similar solutions with fat tails has been proved for a range of kernels [@NV11a; @NV12a], but one of the major open problems is still the uniqueness of self-similar solutions either with finite mass, or, in the case of fat tails, with a given power-law decay behaviour. One idea to prove uniqueness has been to characterize the asymptotics of self-similar solutions for small cluster sizes in order to be able to reformulate the uniqueness problem as an initial value problem. However, at least to our knowledge, this strategy has not yet been successful. In recent work we were able to prove the first uniqueness result for self-similar solutions with finite mass for kernels that are in a certain sense close to the constant one [@NV13b]. One key ingredient of the proof is a precise understanding of the decay behaviour of solutions at infinity which is one of the results of the present paper. In addition, we also prove here more detailed estimates for self-similar solutions (for kernels with homogeneity zero) that are of interest by itself but might also be helpful in proving uniqueness of self-similar solutions for kernels that are not necessarily close to the constant one. In order to formulate our results we first recall the equation for self-similar solutions to with finite mass. For kernels with homogeneity zero such solutions are of the form $n(t,\xi)=t^{-2} f(x)$ with $x=\xi/t$, and the so-called self-similar profile $f$ satisfies $$\label{eq1} -xf'(x) - 2f(x) = \tfrac 1 2 \int_0^x K(x-y,y) f(x-y)f(y)\,dy - f(x) \int_0^{\infty} K(x,y) f(y)\,dy \,$$ with $$\label{eq2} \int_0^{\infty} x f(x)\,dx =M\,.$$ In general however, due to possible divergences of $f$ as $x \to 0$, the integrals in are not necessarily finite such that one has to go over to a weak formulation. Under the assumption that $\lim_{x\to 0} x^2 f(x)=0$ one can integrate to obtain that $f$ satisfies $$\label{eq1b} x^2f(x) = \int_0^x \int_{x-y}^{\infty} K(y,z) y f(z)f(y)\,dz\,dy\,.$$ In [@vanDoErnst88] the authors gave a self-consistent analysis of solutions to , . More precisely, under the assumption that a nontrivial, nonnegative solution exists, they derived detailed properties of the respective profiles for large and small $x$ for a large range of different kernels. As indicated above the first rigorous existence results for , appeared much later. In [@FouLau05] the authors prove for a large range of kernels of homogeneity $\gamma<1$, that there exists a weak self-similar profile, that is a nonnegative function $f \in L^1(x^2\,dx)$ with $xyK(x,y)f(x)f(y) \in L^1((0,\infty)^2)$ and such that is satisfied for almost all $x$. In addition, the self-similar profiles satisfy certain moment estimates that are in agreement with the expectation that profiles with finite mass decay exponentially. Another strategy was employed in [@EMR05] to establish a similar result but for a smaller class of kernels. In further work, regularity properties of self-similar profiles were derived, such as in [@FouLau06a] for the sum kernel $x^{\lambda}+y^{\lambda}$ with $\lambda \in (0,1)$. More precisely, it was shown that every self-similar profile $f$ satisfies $f \in C^1(0,\infty)$ and decays exponentially. For kernels of the form $K(x,y)=x^ay^b+x^by^a$ with $-1 \leq a\leq b<1$ and $a+b \in (-1,1)$ it was proved later in [@CanMisch11] that each self-similar profile is infinitely differentiable. In what follows we assume that there exists a function $f \in L^1_{loc}(0,\infty)$ that is nonnegative, $xf \in L^1(0,\infty)$ and satisfies for almost all $x>0$. Furthermore, we assume for our convenience that $f \in C^0(0,\infty)$. Strictly speaking, such a result has been proved only for a smaller class of kernels but it is not difficult to derive, once one has established certain properties of the profile as $x \to 0$ such as Lemma \[L.negativemoment\] below. One can then proceed as in Chapter 4.2 of [@NV12a] to first prove by an iterative argument that $f \in L^{\infty}_{loc}(0,\infty)$ from which continuity follows rather easily. Our goal in this paper is to prove several results on the details of the decay behaviour of such self-similar profiles. The more detailed the results are, the stronger assumptions we need on the kernel $K$. We emphasize, however, that all but the last result, Theorem \[T4\], hold in particular for kernels such as Smoluchowski’s original one. We feel that this is important, since many kernels that appear in the applied literature, are of a similar form. Throughout this paper we assume that $K$ is homogeneous of degree zero and satisfies for some $\alpha \in [0,1)$ that $$\label{Kassump1} K(x,y) \leq K_0 \Big(\Big(\frac{x}{y}\Big)^{\alpha} + \Big(\frac{y}{x}\Big)^{\alpha}\Big) \qquad \mbox{ for all } x,y \in (0,\infty)$$ and for some $\kappa\in (0,1]$ that $$\label{Kassump2} \min_{|x-y|\leq \kappa(x+y)} K(x,y) \geq k_0>0\,.$$ Our first main result establishes pointwise exponential decay of self-similar profiles. \[T1\] Let $f$ be a solution to -. Then there exist positive constants $C_1,C_2,\alpha_1,\alpha_2$ such that $$C_1 e^{-\alpha_1 x} \leq f(x) \leq C_2 e^{-\alpha_2 x}\qquad \mbox{ for all } x \geq 1\,.$$ Theorem \[T1\] is proved in Section \[S.bounds\], more precisely Lemma \[L.ublargex\] gives the upper bound, while Lemma \[L.lowerbound\] provides the lower bound. If we write $$f(x)= e^{-xa(x)},$$ then Theorem \[T1\] gives that $0<\alpha_1\leq a(x)\leq \alpha_2$. Our next goal is to prove that $\lim_{x \to \infty} a(x)$ exists. For this result we need further regularity of the self-similar solution $f$, which is a nontrivial issue since a standard bootstrap argument does not work for equation . Some regularity results are available (cf. [@FouLau06a; @CanMisch11]) and while the proofs might also apply to the kernels that we consider here, we find it convenient to provide a self-consistent proof that $f \in BV_{loc}(0,\infty)$ in Section \[S.bv\]. For this result we need that $K$ is sufficiently regular. We assume that $K$ is differentiable and $$\label{Kdiff} \Big | \frac{\partial}{\partial x} K(x,y) \Big| \leq C \frac{1}{x} \Big( \Big(\frac{x}{y}\Big)^{\alpha} + \Big(\frac{y}{x}\Big)^{\alpha}\Big) \qquad \mbox{ for all } x,y>0\,.$$ In order to show that $a$ has a limit as $x \to \infty$, we also need that the region in which $K$ is bounded below is sufficiently large. More precisely, we have the following result (cf. Proposition \[P.limit\]). \[T2\] Let $K$ satisfy , with $\kappa >1/3$, as well as . Then $\lim_{x \to \infty} a(x)$ exists. We then turn out attention to the prefactor. To that aim we define the function $u\colon (0,\infty) \to (0,\infty)$ via $$\label{udef} f(x)=u(x) e^{-a^*x}\qquad \mbox{ where } \qquad a^*=\lim_{x \to \infty} a(x)\,.$$ Under the additional assumption that $K$ is uniformly bounded below we prove the following result in Section \[S.preex\]. \[T3\] Let $K$ satisfy , with $\kappa=1$, as well as . Define $u_{R}(x):=u(Rx)$. Then there exists a sequence $R_j \to \infty$ and a nonnegative measure $\mu$ such that $u_{R_j} \wto \mu$ on $(0,\infty)$ as $j \to \infty$. The limit $\mu$ is nontrivial and satisfies the equation $$\label{muequation} x \mu(x) = \frac{1}{2a^*} \int_0^x K(y,x-y)\mu(x-y)\mu(y)\,dy$$ in the distributional sense (cf. ). Ideally, we would like to know that $\mu$ is uniquely determined (and hence a constant). However, to prove such a result turns out to be difficult and we are currently only able to derive such a uniqueness result (as for the full coagulation equation) only for kernels that are close to constant in the sense outlined below. \[T4\] Let $K$ satisfy in addition to - that $$\label{Kassump4} -\eps \leq K-2 \leq \eps \Big( \Big(\frac{x}{y}\Big)^{\alpha} + \Big(\frac{y}{x}\Big)^{\alpha} \Big)$$ and $$\label{Kassump5} \Big | \frac{\partial}{\partial x} K(x,y) \Big| \leq C\eps \frac{1}{x} \Big( \Big(\frac{x}{y}\Big)^{\alpha} + \Big(\frac{y}{x}\Big)^{\alpha}\Big) \qquad \mbox{ for all } x,y>0\,.$$ Then, if $\eps>0$ is sufficiently small, the only nontrivial solution to is the constant one, i.e. $\mu \equiv 2a^*/\int_0^1 K(s,1-s)\,ds$. The proof of Theorem \[T4\] is contained in Section \[S.preuni\]. Estimates for large $x$ {#S.bounds} ======================= An upper bound -------------- \[L.ubsmallxnew\] There exists a constant $C>0$ such that any solution of , satisfies $$\label{ubsmallx1} \sup_{R>0}\frac{1}{R} \int_{R/2}^Rx f(x)\,dx\leq C\,.$$ The proof follows exactly as in Lemma 2.1 of [@NV11b]. \[L.negativemoment\] For any $\gamma\in[0,1)$ there exists a constant $C>0$ such that any solution of , satisfies $$\label{ubsmallx} \int_0^{\infty} x^{1{-}\gamma} f(x)\,dx \leq C\,.$$ The estimate in $(0,1)$ follows from via a dyadic decomposition, an argument we will also use repeatedly in this paper. Indeed, $$\begin{aligned} \int_0^1 x^{1{-}\gamma} f(x)\,dx & =\sum_{n=0}^{\infty} \int_{2^{-(n{+}1)}}^{2^{-n}} x^{1{-}\gamma} f(x)\,dx\\ & \leq 2 \sum_{n=0}^{\infty} 2^{n\gamma} \int_{2^{-(n{+}1)}}^{2^{-n}} x f(x)\,dx \leq C \sum_{n=0}^{\infty} 2^{n(\gamma{-}1)} \leq C\,,\end{aligned}$$ while the estimate in $(1,\infty)$ is a consequence of . \[L.ublargex\] There exist constants $C,a>0$ such that any solution of , satisfies $$f(x) \leq C e^{-ax} \qquad \mbox{ for all } x\geq 1\,.$$ We denote for $\gamma \geq 1$ $$\label{ublargex2} M(\gamma):=\int_0^{\infty} x^{\gamma} f(x)\,dx \,.$$ Our goal is to show inductively that $$\label{ublargex3} M(\gamma) \leq \gamma^{\gamma} e^{A\gamma}$$ for some (large) constant $A$. To that aim we first multiply by $x^{\gamma-2}$ with some $\gamma>1$ and after integrating we obtain $$(\gamma{-}1) M(\gamma) = \tfrac 1 2 \int_0^{\infty}\int_0^{\infty} K(x,y) f(x) f(y)\big( (x+y)^{\gamma} - x^{\gamma} - y^{\gamma} \big)\,dx\,dy\,.$$ By symmetry we also find $$\begin{aligned} M(\gamma)&= \frac{1}{\gamma{-}1} \int_0^{\infty} \int_0^x K(x,y) f(x) f(y) \Big( \big(x+y\big)^{\gamma} - x^{\gamma}\Big) \,dy\,dx\\ & = \int_0^1\,dx \int_0^x \,dy \dots + \int_1^{\infty} \,dx \int_0^{x/\gamma} \,dy \dots + \int_1^{\infty} \,dx \int_{x/\gamma}^x \,dy \dots\,.\end{aligned}$$ Due to we have $$\begin{aligned} \int_0^1\int_0^x K(x,y) f(x) f(y)& \Big( \big(x+y\big)^{\gamma} - x^{\gamma}\Big)\,dy\,dx \leq C \int_0^1 \int_0^x K(x,y) f(x) f(y) x^{\gamma}\,dy\,dx\label{ublargex4} \\ & \leq C\int_0^1 x^{\alpha +\gamma}f(x)\int_0^x y^{1{-}\alpha} f(y)\,dy\,dx \leq C\,.\nonumber\end{aligned}$$ Using and $(x+y)^{\gamma} - x^{\gamma} \leq c x^{\gamma-1}y$ for $ y \leq \frac{x}{\gamma}$, we find that $$\label{ublargex5} \begin{split} \int_1^{\infty} \int_0^{x/\gamma} K(x,y) f(x) f(y)& \Big( \big(x+y\big)^{\gamma} - x^{\gamma}\Big)\,dy\,dx\\ & \leq \int_1^{\infty} \int_0^{x/\gamma} K(x,y) y x^{\gamma{-}1} f(x)f(y)\,dy\,dx\\ &\leq \int_1^{\infty} \int_0^{x/\gamma} x^{\gamma+\alpha-1} y^{1-\alpha} f(x)f(y)\,dy\,dx \\ &\leq C M(\gamma+\alpha -1)\,, \end{split}$$ so that for the sum of both terms we can prove by induction that it is smaller than $1/2 \gamma^{\gamma} e^{A\gamma}$. It remains to estimate $$\begin{aligned} \frac{C}{\gamma-1}&\int_1^{\infty} \,dx \int_{x/\gamma}^x\,dy K(x,y) f(x) f(y) \big(x+y\big)^{\gamma}\\ & \leq \frac{C}{\gamma} \int_1^{\infty} \,dx \int_{x/\gamma}^x \,dy f(x) f(y) \big(x+y\big)^{\gamma} \Big( \frac{x}{y}\Big)^{\alpha}=:(*)\,.\end{aligned}$$ In the following $\{\zeta_n\} \subset (0,1]$, $\zeta_0=1$, will be a decreasing sequence of numbers that will be specified later. Then we define a corresponding sequence of numbers $\kappa_n$ such that given a sequence $\{\theta_n\} \subset (0,1)$, also to be specified later, we have $$\label{kappandef} \big(x+y\big)^{\gamma} \leq \kappa_n^{\gamma} x^{\gamma(1-\theta_n)} y^{\gamma \theta_n} \qquad \mbox{ for } \frac{y}{x} \in [\zeta_{n+1},\zeta_n]\,.$$ Equivalently we have $$\label{kappadef1} \kappa_n = \max_{\zeta \in [\zeta_{n{+}1},\zeta_n]} \Big( \frac{1+\zeta}{\zeta^{\theta_n}}\Big)\,.$$ With these definitions we have $$(*) \leq \frac{C}{\gamma} \sum_{n=0}^{n_0(\gamma)} \kappa_n^{\gamma} \zeta_{n{+}1}^{-\alpha} M(\gamma(1-\theta_n)) M(\gamma \theta_n)\,,$$ where $n_0(\gamma)$ is such that $\zeta_{n_0(\gamma)} = \frac{1}{\gamma}$. We now choose $\theta_n$ such that for $\psi_{\theta_n}(\zeta):= \log(1+\zeta)-\theta_n \log \zeta$ we have $$\min_{\zeta \in [\zeta_{n+1},\zeta_n]} \psi_{\theta_n}(\zeta) = \log(1+\zeta_n) - \theta_n \log(\zeta_n)\,.$$ This is equivalent to $$\label{thetandef} \theta_n = \frac{\zeta_n}{1+\zeta_n}\,.$$ We want to prove by induction over $\gamma$ that $(*) \leq \frac{1}{2} \gamma^{\gamma} e^{A\gamma}$. Inserting the corresponding hypothesis, this reduces to showing that $$\frac{C}{\gamma} \sum_{n=0}^{n_0} \zeta_{n{+}1}^{-\alpha} \exp \Big( \gamma \big( \max_{\zeta \in [\zeta_{n+1},\zeta_n]} \psi_{\theta_n}(\zeta) + \theta_n \log \theta_n + (1-\theta_n) \log (1-\theta_n)\big) \Big)\leq \frac 1 2\,.$$ By definition we have $$\begin{aligned} &\max_{\zeta \in [\zeta_{n+1},\zeta_n]} \psi_{\theta_n}(\zeta) + \theta_n \log \theta_n + (1-\theta_n) \log (1-\theta_n) \\ & = \min_{\zeta \in [\zeta_{n+1},\zeta_n]}\psi_{\theta_n}(\zeta) + (\max-\min)_{\zeta \in [\zeta_{n+1},\zeta_n]} \psi_{\theta_n}(\zeta) + \theta_n \log \theta_n + (1-\theta_n) \log (1-\theta_n) \\ & = (\max-\min)_{\zeta \in [\zeta_{n+1},\zeta_n]} \psi_{\theta_n}(\zeta)\,.\end{aligned}$$ Thus we need to investigate $$\begin{aligned} (\max-\min)_{\zeta \in [\zeta_{n+1},\zeta_n]} \psi_{\theta_n}(\zeta) &= \psi_{\theta_n}(\zeta_{n{+}1}) - \psi_{\theta_n}(\zeta_n)\\ & = \log \Big( \frac{1+\zeta_{n{+}1}}{1+\zeta_n} \Big) - \theta_n \log \Big( \frac{\zeta_{n{+}1}}{\zeta_n} \Big)\\ & = \log \Big( 1 + \frac{\zeta_{n{+}1}-\zeta_n}{1+\zeta_n}\Big) - \frac{\zeta_n}{1+\zeta_n} \log \Big( 1 + \frac{\zeta_{n{+}1}-\zeta_n}{\zeta_n}\Big)\,.\end{aligned}$$ Expanding the nonlinear terms we find $$W:=(\max-\min)_{\zeta \in [\zeta_{n+1},\zeta_n]} \psi_{\theta_n}(\zeta) \leq C \Big( |\zeta_{n{+}1}-\zeta_n|^2 + \frac{(\zeta_{n{+}1}-\zeta_n\big)^2}{ \zeta_n}\Big)\,.$$ We now split $\{1,2,\cdots,n_0\}$ into a finite number of sets $\{1,2,\cdots,N_1\}$, $\{N_1+1, \cdots, N_2\}$, $\cdots$, $\{N_{k-1}+1, \cdots,N_k=n_0\}$ in the following way. We first define $$\zeta_0 =1 \,, \quad \eta_0= 1+ \frac{1}{\sqrt{\gamma}}\,, \quad \zeta_n = \eta_0^{-n} \zeta_0\,, \quad \mbox{ for all } n \leq N_1$$ where $N_1$ is such that $ \zeta_{n} \geq \frac{1}{\sqrt{\gamma}}$, that is we can choose $N_1 \sim \sqrt{\gamma} \log \gamma$. With these definitions we find $$\Big| \frac{\zeta_{n{+}1} - \zeta_n}{\zeta_n}\Big| \leq \frac{C}{\sqrt{\gamma}} \qquad \mbox{ for all } 1\leq n\leq N_1$$ and thus $$\frac{1}{\gamma} \sum_{n=0}^{N_1} \zeta_{n{+}1}^{-\alpha} \exp \big( \gamma W\big) \leq \frac{C N_1}{\gamma} \gamma^{\alpha/2} \sim \gamma^{(\alpha-1)/2} \log \gamma \to 0 \quad \mbox{ as } \gamma \to \infty\,.$$ Next, for $n \in (N_1,N_2]$ we define $$\zeta_n=\eta_1^{-(n-N_1)} \zeta_{N_1}\,, \qquad \eta_1=1+\frac{1}{\gamma^{1/4}}\,.$$ Then $\zeta_n \leq \frac{1}{\sqrt{\gamma}}$, such that $|\zeta_{n{+}1}-\zeta_n|^2 \leq \frac{C}{\gamma}$ and $$\gamma W \leq C \gamma \Big( \zeta_n |\eta_1-1|^2 + \frac{1}{\gamma}\Big) \leq C\,.$$ We need to determine $N_2$ such that $$(N_2-N_1) \frac{1}{\zeta_{N_2}^{\alpha} \gamma} \to 0 \qquad \mbox{ as } \gamma \to \infty\,.$$ Making the ansatz $\zeta_{N_2} = \gamma^{-\sigma}$, that is $N_2 \sim \gamma^{1/4} \log \gamma$, this implies that we need $$\frac{\gamma^{1/4} \gamma^{\alpha \sigma} \log \gamma}{\gamma} \ll 1\,$$ as $\gamma \to \infty$ and this needs $\alpha \sigma < \frac 3 4$. If we can choose $\sigma =1$ we are done. Otherwise we need to iterate the argument. Thus, assume that $3\leq 4\alpha$ and set $\sigma_1=\frac{3}{8\alpha}$. We define for $k \geq 1$ the sequence $\sigma_{k+1} = \frac{1}{4\alpha} (1+\sigma_k)$. Then, we define $\eta_{k+1}= 1+\frac{1}{\gamma^{(1-\sigma_{k+1})/2}}$ and $\zeta_n = \eta_{k+1}^{-(n-N_{k+1})} \zeta_{N_{k+1}}$ for $n \in (N_{k+1},N_{k+2}]$ with $\zeta_{N_{k+1}}=\gamma^{-\sigma_k}$. Then $N_{k+1}-N_k = \gamma^{(1-\sigma_k)/2} \log \gamma $ and we find that our sum is controlled by $C \gamma^{(1-\sigma_k)/2 -1 + \alpha \sigma_{k+1}} \ll 1$ by our definition of $\sigma_{k+1}$. Since $\alpha<1$ we find after a finite number of steps that $\sigma_k \geq 1$, and then we can stop the iteration. It remains to show that implies the pointwise estimate for $f$. Indeed, implies for $R>0$ that $$R^{\gamma} \int_{R}^{2R} f(x)\,dx \leq \int_{R}^{2R} x^{\gamma} f(x)\,dx \leq \gamma^{\gamma} e^{A\gamma}$$ and thus $$\int_R^{2R} f(x)\,dx \leq \exp\Big( \gamma (\log(\gamma)+ \log(R)) + A \gamma\Big)\,.$$ The minimum of $\psi(\gamma):= \gamma (\log(\gamma) +\log(R)) + A \gamma$ is obtained for $\gamma = e^{-(A+1)} R$ and thus $$\int_R^{2R} f(x)\,dx \leq \exp\Big( - e^{-(A+1)} R\Big)$$ and obviously it follows that there exists (another) $A>0$ such that $$\label{ublargex20} \int_R^{2R} f(x)\,dx \leq \exp\Big( - A R\Big)\,.$$ Notice that also implies that $ \int_R^{\infty} f(x)\,dx \leq \exp\Big( - A R\Big)$ for some $A>0$. To deduce now the pointwise bound we first get from that $$\begin{split} x^2 f(x) &=\int_0^x\,dy \int_{x-y}^{\infty} \,dz\cdots \\ &\leq \int_0^1\,dy\int_{x/2}^{\infty} \,dz \cdots + \int_0^1\,dz\int_{x-z}^x\,dy \cdots + \int_1^x\,dy \int_{\max(1,x-y)}^{\infty}\,dz\cdots\\ &=:(I)+(II)+(III)\,. \end{split}$$ Next, we argue via a dyadic argument (as in the proof of Lemma \[L.negativemoment\]), using , that $\int_{x/2}^{\infty} z^{\alpha}f(z)\,dz\leq Ce^{-\frac{A}{2}x}$. With this estimate and we obtain that $$|(I)| \leq C \int_0^1\int_{x/2} \Big(\frac{z}{y}\Big)^{\alpha} y f(y)f(z)\,dz\,dy \leq C e^{-\frac{A}{2}x}\,.$$ Furthermore, given $x-y$ we choose $M$ such that $2^{-M}\leq x-y \leq 2^{-M+1}$ and obtain with that $$\int_{x-y}^1 z^{-\alpha} f(z)\,dz = \sum_{n=0}^M \int_{2^{-(n{+}1)}}^{2^{-n}} z^{-\alpha} f(z)\,dz \leq C2^{M\alpha} \leq C(x-y)^{-\alpha}\,.$$ Hence $$\begin{split} |(II)| &\leq C \int_0^1 \int_{x-z}^x\Big(\frac{y}{z}\Big)^{\alpha}y f(y) f(z)\,dy\,dz \\ & = C\int_{x-1}^x y f(y) \int_{x-y}^1z^{-\alpha}f(z)\,dz\,dy \leq x \int_{x-1}^x (x-y)^{-\alpha} f(y)\,dy\,. \end{split}$$ Finally, we can estimate term $(III)$, using as above dyadic arguments and as well as , to obtain $$\begin{aligned} |(III)| &= \int_1^x\int_{\max(1,x-y)}^\infty K(y,z)y f(y) f(z)\,dz\,dy\\ &=\int_{y \leq z}\cdots + \int_{y>z} \cdots\\ &\leq C\int_1^x\int_{\max(1,x-y)}^{\infty}\Big( \frac{z}{y}\Big)^{\alpha}yf(y) f(z) \,dz\,dy +C\int_1^x\int_{\max(1,x-y)}^{\infty}\Big( \frac{y}{z}\Big)^{\alpha}yf(y) f(z) \,dz\,dy\\ & \leq C \int_{x/2}^{\infty} z^{\alpha} f(z)\,dz + \int_{x/2}^{\infty} y^{1+\alpha}f(y)\,dy\\ & \leq C e^{-\frac{A}{2} x}\,.\end{aligned}$$ In summary we obtain $$\label{ublargex21} f(x)\leq Ce^{-Ax}+\frac{C}{x} \int_{x-1}^x\frac{f(y)}{(x-y)^{\alpha}} \,dy$$ with a (possibly new) constant $A>0$. Iterating gives, using Fubini, $$f(x) \leq Ce^{-Ax} +C^2 \int_{x-1}^x\frac{e^{-Ay}}{(x-y)^{\alpha}}\,dy + C^2 \int_{x-2}^x f(y_2) \int_{y_2}^{y_2+1} \frac{1}{(y_1-y_2)^{\alpha}(x-y_1)^{\alpha}}\,dy_1\,dy_2\,.$$ Now $$\int_{x-1}^x \frac{e^{-Ay}}{(x-y)^{\alpha}}\,dy\leq B e^{-Ax}$$ and $$\int_{y_2}^{y_2+1} \frac{1}{(y_1-y_2)^{\alpha}(x-y_1)^{\alpha}}\,dy_1 \leq \frac{C}{(x-y_2)^{2\alpha-1}} \,.$$ Hence, if $2\alpha<1$ we can finish the proof using Gronwall’s inequality. If $2\alpha>1$ we iterate $L$ times to find $$\begin{aligned} f(x)&\leq C^{L+1}B^{L+1} e^{-Ax}\\ &+ C^{L+2} \int_{x-(L+1)}^x f(y_{L+1}) \int_{y_{L+1}}^{y_{L+1}+1} \frac{dy_L}{(y_l-y_{L+1})^{\alpha}} \cdots \frac{C}{(x-y_2)^{2\alpha-1}}dy_{L+1}\\ & \leq C^{L+1}B^{L+1} e^{-Ax} + C^{L+2} \int_{x-(L+1)}^x f(y_{L+1}) \frac{C}{(x-y_2)^{(L+1)\alpha-L}}\,dy_{L+1}\,.\end{aligned}$$ If $L$ satisfies $(L+1)\alpha < L$ then we can also use Gronwall’s inequality to conclude the proof. A lower bound ------------- \[L.lowerbound\] There exists $c,a>0$ such that $$f(x)\geq c e^{-ax} \mbox{ for all } x\geq 1\,.$$ We start with a lower bound on the integral $I(R):=\int_R^{R+1} f(x)\,dx$. From equation we deduce for large $R$ that $$\label{lowerbound1} \begin{split} I(R) & \geq \frac{k_0}{R-1}\int_{R}^{R+1} \int_{\frac{(1{-}\kappa)x}{2}}^{\frac{1{+}\kappa}{2}x} f(y)\int_{\max(x-y,\frac{1{-}\kappa}{1{+}\kappa}y)}^{ \frac{1{+}\kappa}{1{-}\kappa}y} f(z)\,dz\,dy\,dx\\ & \geq \frac{Ck_0}{R} \Big[ I\Big( \frac{R+1}{2}\Big)^2 + I \Big( \frac{R+1}{2}+1\Big) I\Big( \frac{R+1}{2}-1\Big) + \cdots \\ & \qquad \quad + I\Big(\frac{R+1}{2} + \Big \lfloor \frac{\kappa R}{4}\Big \rfloor \Big) I\Big(\frac{R+1}{2} - \Big \lfloor \frac{\kappa R}{4}\Big \rfloor \Big)\Big]\,. \end{split}$$ We can deduce from Lemma \[L.auxiliary\] below that there exists $R_0>0$ such that $I(R) > \frac{8 e^a}{Ck_0\kappa} e^{-aR}$ for all $R \in [R_0,3R_0]$ for a sufficiently large $a>0$. Then implies with $K=\frac{8 e^a}{Ck_0\kappa}$ that $$I(R) \geq \frac{Ck_0}{R} \Big \lfloor \frac{\kappa R}{2}\Big\rfloor^2 K^2 e^{-aR} > K e^{-aR)}\qquad \mbox{ for all } R \in \Big[ \frac{2R_0-1}{1-\kappa/2},\frac{2(3R_0)-1}{1+\kappa/2}\Big]\,.$$ Choosing $\kappa$ smaller if necessary, we can assume that t $\frac{2R_0-1}{1-\kappa/2} \leq 3R_0$ and $\frac{2R-1}{1+\kappa/2}\geq \frac{3}{2} R$ for all $R \geq R_0$. This means that we have the estimate in $[R_0,\frac{3}{2} 3 R_0]$. In the next step, we obtain the estimate in $[R_0,\frac{3^2}{2^2} 3 R_0]$ and iterating this argument we obtain $$\label{lowerbound2} I(R) \geq Ke^{-aR} \qquad \mbox{ for all } R \geq R_0\,.$$ The pointwise lower bound now follows from using in , first for all $x \geq R_0$, and then by adapting the constants we obtain it for all $x \geq 1$. \[L.auxiliary\] There exists $\tilde R>0$ such that $f(x)>0$ for all $x \geq \tilde R$. We first claim that there exists $x_0>0$ such that $$\label{aux1} \int_{B_{\frac{\kappa x_0}{4}(x_0)}}x f(x)\,dx >0\,.$$ This follows from the fact that $$(0,\infty) \subset \cup_n B_{\frac{\kappa x_n}{4}(x_n)} \qquad \mbox{ with } \qquad x_n=\Big (1+ \frac{\kappa}{8}\Big)^n\,.$$ Then $$M = \int_0^{\infty} x f(x)\,dx \leq \sum_n \int_{B_{\frac{\kappa x_n}{4}(x_n)} }xf(x)\,dx$$ and hence there exists $n_0 \in \N$ such that is true for $x_0:=x_{n_0}$. Then, since $$B_{\frac{\kappa x_0}{4}}(x_0) \times B_{\frac{\kappa x_0}{4}}(x_0) \subset \{ (y,z)\,:\, |y-z| \leq \kappa(y+z)\}\,,$$ equation implies that $$x^2 f(x) \geq k_0 \Big( \int_{B_{\frac{\kappa x_0}{4}(x_0)} }y f(y)\,dy \Big)^2 >0 \qquad \mbox{ for all } x \in \Big( x_0 \big( 1+ \frac{\kappa}{4}\big), 2x_0 \big(1-\frac{\kappa}{4}\big)\Big)\,.$$ We can then iterate this argument to obtain $f(x)>0$ for all $x \geq x_0 \big( 1+ \frac{\kappa}{4}\big)$. BV regularity {#S.bv} ============= \[L.bv\] If $K$ satisfies , and if $f$ is a solution to and , then $f \in BV_{\mbox{loc}}(0,\infty)$. In the following we let $\phi \in C^1_b(0,\infty)$ with $\mbox{supp}(\phi) \subset (0,\infty]$ and $\mbox{supp}(\phi') \subset (0,\infty)$. Our goal is to show that $$\label{bv2} \Big|\int_0^{\infty} f(x) \phi'(x) \,dx\Big|\leq C \|\phi\|_{L^{\infty}(0,\infty)}\,.$$ In a first step we take $\phi$ such that $\mbox{supp}(\phi)\subset [a, \infty]$ and $\mbox{supp}(\phi') \subset [a,b]$ for some $a,b>0$. We divide by $x^2$, multiply with $\phi'$, integrate and change the order of integration to obtain $$\label{bv3} \begin{split} \int_0^{\infty} &f(x) \phi'(x)\,dx= \int_0^{\infty} \int_0^{\infty} K(y,z) y f(y) f(z) \Big( \int_y^{y{+}z} \frac{\phi'(x)}{x^2}\,dx\Big)\,dz\,dy\\ & = \int_0^{\infty} \int_0^{\infty} \,K(y,z) y f(y) f(z) \Big( 2 \int_y^{y{+}z} \frac{\phi(x)}{x^3}\,dx + \frac{\phi(y{+}z)}{(y{+}z)^2} - \frac{\phi(y)}{y^2}\Big)\,dz\,dy\,. \end{split}$$ We split $\int_0^{\infty} dz = \int_0^{a/2} dz \dots + \int_{a/2}^{\infty} dz\dots$. Using the assumption on the support of $\phi$ and recalling Lemma \[L.negativemoment\] and Lemma \[L.ublargex\] we can estimate $$\label{bv4} \begin{split} \Big| \int_0^{\infty} \int_{a/2}^{\infty} &\,K(y,z) y f(y) f(z) \Big( 2 \int_y^{y{+}z} \frac{\phi(x)}{x^3}\,dx + \frac{\phi(y{+}z)}{(y{+}z)^2} - \frac{\phi(y)}{y^2}\Big)\,dz\,dy\Big|\\ & \leq C \int_{a/2}^{\infty} y^{-2} f(y) \int_{a/2}^{\infty}\Big( \Big(\frac{y}{z}\Big)^{\alpha}+\Big(\frac{z}{y}\Big)^{\alpha}\Big) f(z) \,dz\,dy \|\phi\|_{L^{\infty}} \\ &\leq C \|\phi\|_{L^{\infty}}\,. \end{split}$$ Here and in the following the constants depend on $a$, but we will not write this explicitly. Next, we write $$\label{bv4b} \begin{split} 2 \int_y^{y{+}z} \frac{\phi(x)}{x^3}\,dx & + \frac{\phi(y{+}z)}{(y{+}z)^2} - \frac{\phi(y)}{y^2} = \\ & 2 \int_y^{y{+}z} \frac{\phi(x)}{x^3} \,dx + \frac{\phi(y{+}z)}{(y{+}z)^2} - \frac{\phi(y{+}z)}{y^2} + \frac{1}{y^2} \big( \phi(y{+}z)-\phi(y)\big) \end{split}$$ and estimate for $y \geq a/2$ that $$\Big | 2 \int_y^{y{+}z} \frac{\phi(x)}{x^3}\,dx + \frac{\phi(y{+}z)}{(y{+}z)^2} - \frac{\phi(y{+}z)}{y^2} \Big| \leq C \|\phi\|_{L^{\infty}}\, z\,.$$ With this last estimate we find, recalling again Lemmas \[L.negativemoment\] and \[L.ublargex\] that $$\label{bv5} \begin{split} \Big| \int_0^{\infty} \int_0^{a/2} K(y,z) y f(y) f(z) &\Big(2 \int_y^{y{+}z} \frac{\phi(x)}{x^3} \,dx+ \frac{\phi(y{+}z)}{(y{+}z)^2} - \frac{\phi(y{+}z)}{(y{+}z)^2} \Big)\,dz\,dy\Big|\\ & \leq C \|\phi\|_{L^{\infty}}\,. \end{split}$$ Thus, with , and , we have obtained $$\label{bv6} \Big|\int_0^{\infty} f(x) \phi'(x) \,dx\Big|\leq C \|\phi\|_{L^{\infty}} + \Big| \int_0^{\infty} \int_0^{a/2} K(y,z) y f(y) f(z) \frac{1}{y^2} \big( \phi(y{+}z)-\phi(y)\big)\,dz\,dy\Big|$$ and it remains to estimate the last term on the right hand side of . Our strategy is to write this term as $\int f(y) (T\phi)'(y)$ with $$(T\phi)'(y) = \frac{1}{y} \int_0^{a/2} f(z) K(y,z) \big( \phi(y{+}z)-\phi(y)\big)\,dz$$ and $$(T \phi)(y) = \int_0^y \int_0^{a/2} f(z)K(\xi,z) \frac{1}{\xi} \big( \phi(\xi{+}z)-\phi(\xi)\big)\,dz\,d\xi$$ and iterate the previous estimates for the function $T\phi$. We first need to verify that $T\phi$ is an admissible test function. We easily check that if $\mbox{supp} (\phi) \subset [a,\infty]$ and $\mbox{supp} (\phi') \subset [a,b]$ that then $\mbox{supp} (T\phi) \subset [a/2,\infty]$ and $\mbox{supp}( (T \phi)') \subset [a/2,b]$ and hence we can use $T\phi$ as test function. Estimate then implies $$\label{bv7} \Big |\int_0^{\infty} f(y) \big(T\phi\big)'(y)\,dy\Big| \leq C \|T\phi\|_{L^{\infty}} + \Big| \int_0^{\infty} f(y) \big( T^2\phi\big)'(y)\,dy\Big|\,.$$ Thus, our first task is to estimate $\|T\phi\|_{L^{\infty}}$. Notice first that implies $$\label{Kdiff1} \Big | \frac{K(x-y,y)}{x-y} - \frac{K(x,y)}{x}\Big| \leq C y^{1-\alpha} \qquad \mbox{ for } x \geq \frac{a}{2}\;,\; y \in (0,a/2]\,.$$ Using , $\mbox{supp}(\phi) \subset [a,\infty)$ and Lemma \[L.negativemoment\], we find $$\label{bv8} \begin{split} (T\phi)(y)& = \int_0^{a/2} f(z) \int_{a/2}^{y} \frac{K(\xi,z)}{\xi} \phi(\xi{+}z)\,d\xi\,dz - \int_0^{a/2} f(z) \int_a^y \frac{K(\xi,z)}{\xi} \phi(\xi)\,d\xi\,dz\\ &= \int_{0}^{a/2} f(z) \Big [ \int_{a/2}^y \frac{K(\xi,z)}{\xi} \phi(\xi+z)\,d\xi - \int_a^y \frac{K(\xi,z)}{\xi} \phi(\xi)\,d\xi\Big]\,dz\\ &= \int_0^{a/2} f(z) \int_y^{y+z} \frac{K(\xi-z,z)}{\xi-z} \phi(\xi)\,d\xi\,dz \\ &\quad + \int_0^{a/2} f(z) \int_{a/2+z}^{y}\ \Big( \frac{K(\xi-z,z)}{\xi-z} - \frac{K(\xi,z)}{\xi}\Big)\phi(\xi)\,d\xi\,dz \\ & \leq C \|\phi\|_{L^{\infty}} \int_0^{a/2} f(z) z^{1-\alpha} \,dz+ C \int_0^{a/2} f(z) \int_{a/2+z}^y z^{1-\alpha} \phi(\xi)\,d\xi\,dz\\ & \leq C \|\phi\|_{L^{\infty}}\,. \end{split}$$ It remains to estimate $\Big| \int_0^{\infty} f(y) \big( T^2\phi\big)'(y)\,dy\Big|$. To that aim recall $$\label{bv9} \big( T^2\phi\big)'(y) = \frac{1}{y} \int_0^{a/2} f(z)K(y,z) \Big( \big(T\phi\big)(y{+}z) - \big(T\phi\big)(y) \Big)\,dz$$ and we write $$\label{bv10} \begin{split} (T\phi)(y+z)&-(T\phi)(y) = \int_y^{y+z} \int_0^{a/2} f(\eta) K(\xi,\eta) \frac{1}{\xi}\big( \phi(\xi+\eta)-\phi(\xi)\big)\,d\eta\,d\xi\\ &=\int_0^{a/2} f(\eta) \Big( \int_{y+\eta}^{y+\eta+z} \frac{K(\xi-\eta,\eta)}{\xi-\eta} \phi(\xi)\,d\xi - \int_y^{y+z}\frac{K(\xi,\eta)}{\xi} \phi(\xi)\,d\xi\Big)\,d\eta\\ &= \int_0^{a/2} f(\eta) \int_{y+\eta}^{y+\eta+z} W(\xi,\eta)\phi(\xi)\,d\xi\,d\eta\\ &\quad + \int_0^{a/2} f(\eta) \Big(\int_{y+z}^{y+z+\eta} \frac{K(\xi,\eta)}{\xi}\phi(\xi) \,d\xi- \int_y^{y+\eta} \frac{K(\xi,\eta)}{\xi}\phi(\xi)\,d\xi\Big)\,d\eta\\ &=: (T_1\phi)(y+z)-(T_1\phi)(y) + (T_2\phi)(y+z)-(T_2\phi)(y)\,, \end{split}$$ where $$W(\xi,\zeta)= \frac{K(\xi- \zeta,\xi)}{\xi-\zeta} - \frac{K(\xi,\zeta)}{\xi}\,.$$ Invoking again we find $$\label{bv13} \begin{split} \Big| (T_1\phi)(y{+}z)-(T_1\phi)(y)\Big| &= \Big| \int_0^{a/2} f(\zeta) \int_y^{y{+}z} W(\xi,\zeta) \phi(\xi)\,d\xi\,d\zeta\Big|\\ &\leq C \|\phi\|_{L^{\infty}} z \int_0^{a/2} f(\zeta) \zeta^{1-\alpha} \,d\zeta \leq C \|\phi\|_{L^{\infty}}z\,. \end{split}$$ If we could obtain the same estimate for $T_2$ we would obtain from together with Lemma \[L.negativemoment\] and the properties of $\mbox{supp} (\phi)$ that $\|T^2\phi\|_{0,1} \leq C \|\phi\|_{L^{\infty}}$ and the proof would be finished. Unfortunately, we cannot in general expect an estimate for $T_2\phi$ as in . The reason is that, using , we obtain $$\big|(T_2\phi)(y+z)-(T_2\phi)(y)\big| \leq C_a \|\phi\|_{L^{\infty}} z^{1-\alpha} \int_0^{a/2}f(\eta) \,d\eta \,,$$ but in general we do not know whether the integral $\int_0^{a/2} \,d\eta f(\eta)$ is finite and in addition the factor $z^{1-\alpha}$ causes problems in if $\alpha \geq 1/2$. Thus, in order to obtain the desired estimates we need to iterate again. More precisely, we define $L \in \N$ such that $L>1+\frac{1}{1-(\alpha+\eps)}$ with some fixed $\eps \in (0,(1-\alpha)/2)$. We are going to show that $$\label{bv15} \|T^L\phi\|_{0,1} \leq C \|\phi\|_{L^{\infty}}\,.$$ We have seen that $\mbox{supp}(\phi) \subseteq [a,\infty]$ implies $\mbox{supp}(T^L\phi) \subseteq [\frac{a}{2^L},\infty]$, whereas $\mbox{supp}(\phi') \subseteq [a,b]$ implies $\mbox{supp}((T^L\phi)') \subseteq [\frac{a}{2^L},b]$. Hence, $T^L\phi$ is an admissible test function and we obtain from , and that $$\Big| \int_0^{\infty} f(y) (T^L\phi)'(y)\,dy\Big| \leq C \|T^L\phi\|_{L^{\infty}} + \Big| \int_0^{\infty} f(y) (T^L\phi)'(y)\,dy\Big| \leq C \|\phi\|_{L^{\infty}}\,,$$ which finishes the proof of Lemma \[L.bv\]. Thus, it remains to prove . We split the integral in the definition of $T_2$ as $\int_0^{a/2}\,d\eta \dots = \int_0^z \,d\eta \cdots + \int_z^{a/2} \,d\eta \dots$ and write $$\label{bv14} \begin{split} H(y&,z,\eta):= \int_{y+z}^{y+z+\eta} \frac{K(\xi,\eta)}{\xi} \phi(\xi) \,d\xi- \int_y^{y+\eta} \frac{K(\xi,\eta)}{\xi}\phi(\xi)\,d\xi\\ &= \int_y^{y+\eta} \Big( \frac{K(\xi+z,\eta)}{\xi+z} - \frac{K(\xi,\eta)}{\xi}\Big) \phi(\xi+z)\,d\xi + \int_y^{y+\eta} \frac{K(\xi,\eta)}{\xi} \big(\phi(\xi+z)-\phi(\xi)\big)\,d\xi\\ &=:(I)+(II)\,. \end{split}$$ Assumption implies that $$|(I)| \leq C_a \|\phi\|_{L^{\infty}} \eta^{1-\alpha} z\,,$$ while gives $$|(II)| \leq C_a \|\phi\|_{0,\gamma} \eta^{1-\alpha} z^{\gamma}$$ for $\gamma \in [0,1)$ such that $$\label{bv16} \Big| \int_0^z f(\eta) H(y,z,\eta) \,d\eta\Big| \leq C\Big( \|\phi\|_{L^{\infty}} z \int_0^z \eta^{1-\alpha} f(\eta)\,d\eta + \|\phi\|_{0,\gamma} z^{\gamma} \int_0^z \eta^{1-\alpha} f(\eta)\,d\eta\,.$$ A dyadic argument, as in the proof of Lemma \[L.negativemoment\], implies that $\int_0^z \eta^{1-\alpha} f(\eta) \,d\eta\leq C z^{1-\alpha}$ such that $$\label{bv17} \Big| \int_0^z f(\eta) H(y,z,\eta)\,d\eta \Big| \leq C\Big( \|\phi\|_{L^{\infty}} z^{2-\alpha} + \|\phi\|_{0,\gamma} z^{1-\alpha+\gamma}\Big) \,.$$ Next, we notice that in the case that $\eta >z$ it holds $$H(y,z,\eta)= \int_{y+\eta}^{y+\eta+z} \,d\xi \dots + \int_y^{y+z} \,d\xi \dots\,.$$ Hence we can derive, using , that $$\label{bv18} \begin{split} \Big| \int_{y+\eta}^{y+\eta+z} & \frac{K(\xi,\eta)}{\xi} \phi(\xi)\,d\xi - \int_y^{y+z} \frac{K(\xi,\eta)}{\xi} \phi(\xi)\,d\xi\Big| \\ & = \Big| \int_y^{y+z} \Big[ \Big( \frac{K(\xi+\eta,\eta)}{\xi+\eta} - \frac{K(\xi,\eta)}{\xi} \Big) \phi(\xi+\eta) + \frac{K(\xi,\eta)}{\xi} \big( \phi(\xi+\eta) - \phi(\xi)\big)\Big]\,d\xi \Big|\\ & \leq C \|\phi\|_{L^{\infty}} z \eta^{1-\alpha} + C \|\phi\|_{0,\gamma} z \eta^{\gamma-\alpha}\,. \end{split}$$ Using estimate we obtain $$\label{bv19} \Big|\int_z^{a/2} f(\eta) H(y,z,\eta)\,d\eta\Big| \leq C \|\phi\|_{L^{\infty}} z + C \|\phi\|_{0,\gamma} z \int_z^{a/2} \eta^{\gamma-\alpha} f(\eta)\,d\eta\,.$$ We assume that $\gamma \leq \alpha$, since eventually we want to use the estimate also for $\gamma =0$. We obtain with Lemma \[L.negativemoment\] $$\label{bv20} \int_z^{a/2} \eta^{\gamma-\alpha} f(\eta) \,d\eta\leq C_{\eps} z^{\gamma-\alpha - \eps}\,.$$ In summary, , , and imply $$\Big| (T\phi)(y+z) - (T\phi)(y)\Big| \leq C\big( \|\phi\|_{L^{\infty}} z + \|\phi\|_{0,\gamma} z^{1-\alpha-\eps + \gamma} \big)$$ and consequently $$\label{bv21} \|T\phi\|_{0,1-\alpha-\eps+\gamma} \leq C \|\phi\|_{0,\gamma}\,.$$ We can iterate to obtain $$\label{bv22} \|T^L\phi\|_{0,1-\alpha-\eps+\gamma_L} \leq C \|T^{L-1}\phi\|_{0,\gamma_L} =C \|T^{L-1}\phi\|_{0,1-\alpha-\eps +\gamma_{L-1}} \leq \dots \leq C \|\phi\|_{0,\gamma_1}\,.$$ With the definitions $\gamma_\ell:=\min(1,(\ell-1)(1-\alpha-\eps))$, $\ell=1,\dots,L$ and the choice of $L$ we have $\gamma_L \geq 1$ and $\gamma_1=0$ which finishes the proof of . Asymptotics as $x \to \infty$ {#S.asymptotics} ============================= A lower bound on the changes of $\log \frac{1}{f(x)}$ ----------------------------------------------------- Writing $$f(x)=e^{-x a(x)}$$ our first goal is to derive a lower bound on $a'(x)$. Notice that since $f \in C^0 \cap BV_{loc}$ and $f>0$, we also have that $a \in BV_{loc}(0,\infty)$ (see e.g. [@AmDalMa90]). Our first goal will be to derive a lower bound on $a'$. For that purpose we need the following Lemma. \[L.lowerasymp1\] Suppose that $a\colon \R\to \R$ such that $$\label{lowerasymp1a} 0<\alpha_1 \leq a(x) \leq \alpha_2< \infty \qquad \mbox{ for all } x \geq 1\,.$$ Then, given an $\eps\in(0,1)$ and $R\geq 1$ there exists $\bar x \in I_R:=\Big( R\big(1+\frac{\eps}{4}\big), R\big(1+\frac{3\eps}{4}\big)\Big)$ such that $$a(y) \leq a(\bar x) + \frac{C(\eps)}{\bar x} \qquad \mbox{ for all } y \in [\bar x-1,\bar x] \cap I_R$$ with $C(\eps)= \frac{4(\alpha_2-\alpha_1)}{\eps}$. Consider first the case that $R \eps \geq 2$. Then we have $$a(y) \leq \alpha_2 = \alpha_1+\frac{C(\eps)\eps}{4} \leq \alpha_1 + \frac{C(\eps)}{2R} \leq \alpha_1 + \frac{C(\eps)}{2} \frac{\big(1+\frac{3\eps}{4}\big)}{\bar x} \leq a(\bar x) + \frac{C(\eps)}{\bar x}$$ for all $\bar x \leq R\big( 1+\frac{3\eps}{4}\big)$. Now assume that $R\eps >2$. Suppose there exists $R \geq \frac{2}{\eps}$ such that for all $\bar x \in I_R$ there exists $y \in [\bar x-1,\bar x]\cap I_R$ with $a(y)>a(\bar x) + \frac{C(\eps)}{\bar x}$. We define a sequence $(x_n)$ as follows: we set $x_0=R(1+\frac{3\eps}{4})$. By assumption there exists $x_1 \in [x_0-1,x_0]\cap I_R$ such that $a(x_1)>a(x_0)+ \frac{C}{x_0}$. Iteratively we obtain $x_n \in [x_{n-1}-1,x_{n-1}]\cap I_R$ such that $a(x_n) \geq a(x_0) + C \big( \frac{1}{x_0} + \dots + \frac{1}{x_{n-1}}\big) \geq a(x_0) + \frac{Cn}{x_0}$. For $n=\lfloor \frac{R\eps}{2}\rfloor$ we obtain $$a(x_n) \geq \alpha_1 + \frac{4 (\alpha_2-\alpha_1)}{\eps} \frac{R\eps}{2 R\big( 1+\frac{3\eps}{4}\big)} =\alpha_1 + \frac{2 (\alpha_2-\alpha_1)}{ 1+ \frac{3\eps}{4}} > \alpha_2$$ which gives a contradiction and finishes the proof of the Lemma. \[L.lowerasymp2\] There exists $R>0$ and $b>0$ such that $$\int_{x}^{x+1} (a')_{-}(x)\,dx \leq \frac{C}{x}+ C_R e^{-b x}\qquad \mbox{ for all } x \geq R\,.$$ Here we use the Jordan decomposition $a'(x)=(a')_{+}(x)-(a')_{-}(x)$ with $(a')_{+}\geq 0$ and $(a')_{-}\geq 0$. In order to make the idea of the proof clear, we first present the formal derivation of the result, that is we ignore for the moment that some of the integrals are not well-defined. From we obtain $$\label{lowerasymp1} \begin{split} f'(x)& = - \frac{2}{x}f(x) + \frac{1}{x} \int_0^1 \big( K(x,y) f(x) - K(x{-}y,y) f(x{-}y) \big) f(y) \,dy \\ & \quad + \frac{1}{x} \Big( \int_1^{\infty} K(x,y) f(x)f(y) \,dy - \int_1^{x/2} K(x{-}y,y) f(x{-}y)f(y)\,dy\Big)\,. \end{split}$$ Using Lemma \[L.ublargex\] to conclude that $\frac{1}{x}\int_1^{\infty} dy\, K(x,y) f(y)\leq Cx^{\alpha-1}$, we find $$\label{lowerasymp2} f'(x) \leq \frac{C}{x^{1-\alpha}} f(x) + \frac{f(x)}{x} \int_0^1 K(x{-}y,y) f(y) \Big(\frac{ K(x,y)}{K(x{-}y,y)} - \frac{f(x{-}y)}{f(x)}\Big) f(y)\,dy\,.$$ Using and Lemma \[L.negativemoment\] we can estimate for $x \geq 1$ that $$\begin{split} \Big | \frac{f(x)}{x} \int_0^1 K(x{-}y,y) f(y) \Big( \frac{ K(x,y)}{K(x{-}y,y)} -1\Big)\,dy \Big| & \leq \frac{f(x)}{x} C \int_0^1 \Big( \frac{x}{y}\Big)^{\alpha} \frac{y}{x} f(y)\,dy \\ &\leq \frac{f(x)}{x^{2-\alpha}} \leq \frac{f(x)}{x^{1-\alpha}} \end{split}$$ and thus we can absorb this error term into the first term of the right hand side of . In terms of $a$ we have found the inequality $$\label{lowerasymp3} \begin{split} \big( x a(x)\big)' & \geq - \frac{C}{x^{1-\alpha}} + \frac{1}{x} \int_0^1 K(x{-}y,y)f(y) \Big( e^{-(x{-}y) a(x{-}y) + xa(x)} -1\Big) \,dy\\ & = - \frac{C}{x^{1-\alpha}} + \frac{1}{x} \int_0^1 K(x{-}y,y) f(y)\Big( e^{ya(x{-}y)} e^{x(a(x)-a(x{-}y))} -1\Big) \,dy \\ & \geq - \frac{C}{x^{1-\alpha}} + \frac{1}{x} \int_0^1 K(x{-}y,y)f(y)\Big(e^{x(a(x)-a(x{-}y))} -1\Big) \,dy\,. \end{split}$$ Estimate implies that $$\label{lowerasymp4} x (a')_{+}(x)-x (a')_{-}(x)+a(x)\geq - \frac{C}{x^{1-\alpha}}+ \frac{1}{x} \int_0^1 K(x{-}y,y)f(y)\Big(e^{x(a(x)-a(x{-}y))} -1\Big)\,dy\,.$$ Furthermore, since $a$ is continuous, we can write and estimate $$a(x)-a(x-y) = \int_{x{-}y}^x a'(\xi)\,d\xi \geq - \int_{x{-}y}^x (a')_{-}(\xi) \,d\xi \,.$$ Thus we have, using $1-e^{-z} \leq z$ for $z \geq 0$ and a dyadic argument as in Lemma \[L.negativemoment\], that $$\label{lowerasymp6} \begin{split} x (a')_{-}(x) &\leq \alpha_2 +\frac{C}{x^{1-\alpha}} + \frac{1}{x} \int_0^1 K(x{-}y,y) f(y)\Big(1- e^{-x \int_{x{-}y}^x (a')_{-}(\xi) \,d\xi } \Big) \,dy\\ & \leq C + C \int_0^1 \int_{x{-}y}^x (a'(\xi))_{-}\,d\xi \frac{(x-y)^{\alpha}}{y^{\alpha}} f(y)\,dy\\ &\leq C + C \int_{x{-}1}^x (a')_{-}(\xi) \int_{x{-}\xi}^1 \frac{f(y)}{y^{\alpha}} (x-y)^{\alpha} \,dy \,d\xi\\ & \leq C + C x^{\alpha} \int_{x{-}1}^x (a')_{-}(\xi)\frac{1}{(x-\xi)^{\alpha}}\,d\xi\,. \end{split}$$ As a consequence, we obtain $$(a')_-(x) \leq \frac{C}{x} + \frac{C}{x^{1-\alpha}} \int_{x-1}^x (a')_-(\xi) \frac{1}{(x-\xi)^{\alpha}}\,d\xi$$ and with $F(x):= \int_x^{x+1} (a')_-(y)\,dy$ it follows $$\label{lowerasymp7} \begin{split} F(x)& \leq \frac{C}{x} + \frac{C}{x^{1-\alpha}} \int_{x}^{x+1} \int_{\xi-1}^{\xi} \frac{(a')_-(y)}{(\xi-y)^{\alpha}}\,dy\,d\xi\\ & \leq \frac{C}{x} + \frac{C}{x^{1-\alpha}} \int_{x-1}^{x+1} (a')_-(y) \int_{\max(x,y)}^{\max(x+1,y+1)} \frac{1}{(\xi-y)^{\alpha}}\,d\xi \,dy\\ & \leq \frac{C}{x} + \frac{C}{x^{1-\alpha}} F(x-1) + \frac{C}{x^{1-\alpha}} F(x)\,. \end{split}$$ If $x \geq R$ where $C/R^{1-\alpha} \leq \frac 1 2$ we have found that $$\label{lowerasymp8} F(x) \leq \frac{C}{x} + \frac 1 2 F(x-1) \qquad \mbox{ for all } x \geq R\,.$$ Since $a \in BV_{loc}$ we can assume that $F(R-1)<\infty$. Hence iterating gives the desired result. In order to derive rigorously we have to work with the weak formulation of . Multiplying with a test function $\phi \in C_0^1(R,\infty)$ with sufficiently large $R\geq 1$, we find $$\label{rigorous1} - \int_0^{\infty} x^2 f(x)\phi'(x)\,dx = \int_0^{\infty} \int_0^{\infty} K(y,z) y \big ( \phi(y)-\phi(y+z)\big) f(y) f(z)\,dy\,dz\,.$$ From [@AmDalMa90] we know that $$\label{rigorous2} - \int_0^{\infty} x^2 f(x)\phi'(x)\,dx = - \int_0^{\infty} x^2 \big( a(x) + x a'(x)\big) f(x)\phi(x)\,dx\,.$$ We are going to show that $$\label{rigorous3} - \int_0^{\infty} x^2 f(x)\phi'(x)\,dx \leq C \int_0^{\infty} x^{2{+}\alpha} \phi(x)f(x)\,dx + C \int_0^{\infty} \big( a'\big)_-(\xi)\int_{\xi}^{\xi+1} \frac{x^{2+\alpha} f(x)\phi(x)}{(x-\xi)^{\alpha}}\,dx\,d\xi\,.$$ Plugging into and absorbing the term $\int x^2 a(x) f(x)\phi(x)\,dx$ into the first term on the right hand side we obtain $$\label{rigorous4} \begin{split} -\int_0^{\infty} x^3 a'(x) f(x)\phi(x)\,dx &\leq C \int_0^{\infty} x^{2+\alpha} \phi(x)f(x)\,dx\\ & + C \int_0^{\infty} \big( a'\big)_-(\xi)\int_{\xi}^{\xi+1} \frac{x^{2+\alpha} f(x)\phi(x)}{(x-\xi)^{\alpha}}\,dx\,d\xi\,. \end{split}$$ Given a set $A \subset \R$ we take (after approximation) $\phi = \frac{\chi_A(x)}{f(x)}$, where the support of $A$ is such that $f>0$ on $A$ (cf. Lemma \[L.auxiliary\]). Defining $$\nu(A):= \int_0^{\infty} \big( a'\big)_-(\xi)\int_{\xi}^{\xi+1} \frac{x^{2+\alpha} \chi_A(x)}{(x-\xi)^{\alpha}}\,dx\,d\xi$$ we obtain $$\label{rigorous5} -x^3 \big( a'\big)(A)\leq C\int_{A} x^{2+\alpha} \,dx + \nu(A)\,.$$ We also know that given $A$ there exists $A_{\pm}$ such that $\big(a'\big)_+(A_-)=0$, $\big(a'\big)_-(A_+)=0$ and $\big(a'\big)_+(A \cap A_+)= \big(a'\big)_+(A)$, $\big(a'\big)_-(A\cap A_-)=\big(a'\big)_-(A)$. Then we deduce from that $$\label{rigorous6} x^3 \big(a'\big)_-(A) \leq C \int_A x^{2+\alpha}\,dx + \nu(A)$$ and by choosing $A=[\bar x,\bar x+1]$ we obtain $$\begin{split} \int_{\bar x}^{\bar x+1} \big(a'\big)_-(\xi)\,d\xi & \leq \frac{C}{\bar x} + \frac{C}{{\bar x}^3} \int_0^{\infty} \big( a'\big)_- (\xi) \int_{\max(\xi,\bar x)}^{\max(\xi+1,\bar x+1)} \frac{x^{2+\alpha}}{(x-\xi)^{\alpha}}\,dx\,d\xi\\ & \leq \frac{C}{\bar x} + \frac{C}{{\bar x}^{1-\alpha}} \int_{[\bar x-1,\bar x+1]} \big(a'\big)_-(\xi)\,d\xi\,, \end{split}$$ which is just . It remains to prove . To that aim we rewrite the right hand side of as $$\int_0^{\infty} \,dy \int_1^{\infty} \,dz \dots +\int_0^{\infty} \,dy \int_0^1\,dz \dots =: (I)+(II)\,.$$ Due to the properties of $\phi$ and $K$ we have $$|(I)| \leq C \int_0^{\infty}y^{1+\alpha} f(y)\phi(y)\,dy \leq C \int_0^{\infty}y^2f(y)\phi(y)\,dy\,.$$ We rewrite the second term as $$\begin{split} (II)&=\int_0^{\infty}\int_0^1 f(y)f(z) \big[y K(y,z) \phi(y) -(y-z)K(y-z,z)\phi(y)\big]\,dz\,dy\\ & \quad + \int_0^{\infty} \int_0^1 f(y)f(z) \big[(y-z)K(y-z,z)\phi(y) - yK(y,z)\phi(y+z)\big]\,dz\,dy\\ &=:(II)_a+(II)_b\,. \end{split}$$ Due to , and the properties of $\phi$ we can estimate $$\big|(II)_a\big| \leq C \int_0^{\infty} f(y) \phi(y) y^{\alpha} \int_0^1 f(z) z^{1-\alpha}\,dz\,dy \leq C \int_0^{\infty} y^2 f(y) \phi(y)\,dy\,.$$ As before in the formal argument we have the estimate $$\label{rigorous7} \frac{f(y)}{f(y+z)} = e^{z a(z)} e^{y(a(y+z)-a(y))} \geq \exp\Big(-y \int_y^{y+z} \big(a'\big)_-(\xi)\,d\xi\Big) \geq 1-y \int_y^{y+z} \big(a'\big)_-(\xi)\,d\xi\,.$$ After a change of variables the term $(II)_b$ can be estimated as $$\begin{split} (II)_b&= \int_0^{\infty} \int_0^1 f(z) \big( f(y+z)-f(y)\big) y K(y,z) \phi(y+z)\,dz\,dy\\ & \leq \int_0^{\infty} \int_0^1 f(z) f(y+z) y^2 K(y,z) \phi(y+z)\int_y^{y+z} \big( a'\big)_-(\xi)\,d\xi\,dz\,dy\\ &= \int_0^{\infty} \int_0^1 f(z) f(y) (y-z)^2 K(y-z,z)\phi(y) \int_{y-z}^y \big( a'\big)_-(\xi)\,d\xi\,dz\,dy\\ & = \int_0^{\infty} \big(a'\big)_-(\xi) \int_{\xi}^{\xi+1} f(y)\phi(y) \int_{y-\xi}^1 f(z) (y-z)^2 K(y-z,z)\,dz\,dy\,d\xi\\ & \leq C \int_0^{\infty} \big(a'\big)_-(\xi) \int_{\xi}^{\xi+1} y^{2+\alpha} (y-\xi)^{-\alpha} f(y) \phi(y)\,dy\,d\xi \end{split}$$ and this implies . Strictly speaking in the previous argument we have been adding and subtracting a term that is possibly infinity due to divergences as $z \to 0$. This difficulty can be removed, by integrating $z$ over the interval $(\eps,1)$ first, performing all the operations and finally let $\eps \to 0$. The limit $\frac{1}{x} \log \frac{1}{f(x)}$ exists -------------------------------------------------- \[L.limit1\] (Doubling Lemma) There exist $R,C>0$ such that for $ x \geq R$ we have $$a(X)-a(x) \leq \frac{\log\big( C(X +1)\big)}{X} \qquad \mbox{ for } X=2 x -2\,.$$ From Lemma \[L.lowerasymp1\] we have $$a(y) \leq a( x) + \frac{C}{x} \big( 1+ (x-y)\big) \qquad \mbox{ for } y \in [x/2, x]$$ and the same inequality for $z \in [ x/2, x]$. The lower bound on the kernel implies that for sufficiently large $x$ we have $K \geq k_0$ in $[x-1,x]^2$. Then we obtain, using , for $X=2x-2$ that $$\label{doubling1} \begin{split} 1& \geq \frac{C}{X^2} \int_{y \in [x-1,x]} dy \int_{z \in [x -1, x]} dz \,y e^{X a(X) - ya(y) - z a(z)}\\ & \geq \frac{C}{X} \int_{y \in [ x-1, x]} dy \int_{z \in [ x -1, x]} dz \,e^{Xa(X) - (y+z) a(x) - \frac{C}{x}(y+z)}\\ & \geq \frac{C}{X} \int_{y \in [x-1,x]} dy \int_{z \in [x -1,x]} dz\, e^{X (a(X) - a(x)) -C}\,. \end{split}$$ As a consequence, we find $$e^{X (a(X) - a(x)) -C} \leq CX$$ and the statement of the Lemma follows. We define $$M_{\delta}(x) = \max_{\delta x \leq y \leq x} a(y)\,.$$ We obviously have $$\limsup_{x \to \infty} a(x) = \limsup_{x \to \infty} M_{\delta}(x)\,.$$ For the following we assume that for some $\theta \in (0,1)$ $$\label{Kassump3} K(y,x-y) \geq k_0 \qquad \mbox{ for } y \in [\theta x,(1-\theta)x]\,.$$ Observe that this follows from for $\theta \geq (1-\kappa)/2$. \[L.limit2\] (Flatness Lemma) Given $\delta>0$ there exist for all $\eps>0$ and $\eta>0$ numbers $R$ and $\sigma$ such that the following holds: if $a(x) \geq M_{\delta}(x)-\sigma$ with $x \geq R$, then $a(y) \geq M_{\delta}(x)-\eta$ for all $y \in ((\theta+\eps) x, (1-(\theta+\eps))x]$. By definition we have that $a(y) \leq M_{\delta}(x)$ for all $y \in [\delta x, x]$. Then implies that $$\label{flatness1} 1\geq \frac{1}{x^2} \int_{\delta x}^x \int_{x{-}y}^{\infty} K(y,z) y e^{xa(x)-yM_{\delta}(x) - z a(z)}\,dz\,dy\,.$$ Assume that the statement of the Lemma is not true. Then, given $\eta>0$, there exists $z^* \in [(\theta+\eps) x, (1-(\theta+\eps))x]$ such that $a(z^*) < M_{\delta}(x)-\eta$. By Lemma \[L.lowerasymp1\] there exists $x^*\in [z^*(1+\eps/4),(1+3\eps/4)z^*]=:I$ such that $$a(z) \leq a(x^*) + \frac{C}{x^*} \leq a(z^*)+\frac{C}{x^*}(1+(z^*-x^*)) < M_{\delta}(x)-\eta + \frac{C}{x^*}$$ for all $z \in [x^*-1,x^*]\cap I$. Next, we notice that $[x^*-1,x^*] \subseteq [\theta x,(1-\theta)x]$ for sufficiently large $x$. This follows since $$x^* -1 \geq z^*(1+\eps/4) \geq(\theta+\eps)(1+\eps/4)x-1\geq\theta x$$ and $$x^* \leq z^*(1+3\eps/4) \leq (1-(\theta+\eps))(1+3\eps/4)x \leq(1-\theta)x$$ if $x$ is sufficiently large. In addition we have, if $z \geq x^*-1/2$ and if $x$ is sufficiently large that $x-z+1 \leq \frac{1+\kappa}{1-\kappa}z$. Thus, we can estimate the right hand side of further via $$\label{flatness2} \begin{split} 1 & \geq \frac{c}{x} \int_{x^*-1/2}^{x^*} \int_{ x-z}^{ x+1-z} e^{x(M_{\delta}(x)-\sigma) - y M_{\delta}(x) - z (M_{\delta}(x)-\eta)}\,dy\,dz \\ &\geq \frac{c}{x}\int_{x^*-1/2}^{x^*} e^{-x\sigma + \eta z}\int_{x-z}^{x+1-z} e^{M_{\delta}(x)(x-(y+z))}\,dy\,dz\\ & \geq \frac{c}{x} e^{-\alpha_2} e^{-x\sigma + \eta \theta x}\,. \end{split}$$ Choosing $\sigma=\eta \theta/2$ we obtain $e^{\eta \theta x/2}\leq Cx$ which gives a contradiction for sufficiently large $x$. \[P.limit\] If $\theta < 1/3$ or equivalently $\kappa >1/3$, the limit of $a(x)$ as $x \to \infty$ exists. We assume that $$b^*:=\limsup_{x \to \infty} a(x) > a^*:=\liminf_{x \to \infty} a(x)\,$$ and let $\eps=(b^*-a^*)/10$. [*Claim 1:*]{} There exists $\beta>0$ such that if $x_n=2+\beta 2^n$, we have $a(x_n) \leq a^* + 2\eps$. Indeed, the doubling Lemma \[L.limit1\] implies that we have $a(x_{n+1}) -a(x_n) \leq \frac{1}{\sqrt{x_n}}$. The definition of $a^*$ implies that there exists $x_0$ such that $a(x_0) =a(2+\beta) < a^*+\eps$. Then $a(x_n)-a(x_0) = \sum_{i=1}^n a(x_i)-a(x_{i-1}) \leq \sum_{i}^n \frac{1}{\beta 2^i} \leq \frac{2}{\beta}$ and thus $a(x_n) \leq a^* + \eps + \frac{C}{\sqrt{x_0}} \leq a^* + 2\eps$ for sufficiently large $\beta$ which proves Claim 1. There exists $({\bar x}_n)$ such that $a({\bar x}_n) \geq M_{\delta}({\bar x}_n) - \sigma$ with $M_{\delta}( {\bar x}_n) \geq b^* - \eps$. First notice that for sufficiently large $x$ we have $M_{\delta}(x) \leq b^*+ \frac{\sigma}{4}$. Next, since $a$ is continuous there exists $x_n^*$ with $a(x^*_n) < b^*-\frac{\sigma}{2}$. We define ${\bar x}_n:= \inf \{ x>x^*_n\,:\, a({\bar x}_n) = b^*-\frac{\sigma}{2}\}$. Then $$a({\bar x}_n) = b^*-\frac{\sigma}{2} > M_{\delta}({\bar x}_n) - \frac{\sigma}{4}-\frac{\sigma}{2} > M_{\delta}({\bar x}_n)- \sigma\,.$$ Since $M_{\delta}({\bar x}_n) \geq a({\bar x}_n)$ we have $M_{\delta}({\bar x}_n) \geq b^*- \frac{\sigma}{2} \geq b^*-\eps$. We can now apply the flatness Lemma \[L.limit2\] to conclude that $$a(y) \geq M_{\delta}({\bar x}_n) - \eta \geq M_{\delta}({\bar x}_n) - 2\eps \qquad \mbox{ for all } y \in [(\theta+\eps) {\bar x}_n, (1-(\theta+\eps)){\bar x}_n]\,.$$ If we can choose $m \in \N$ such that $x_m \in [(\theta+\eps) {\bar x}_n, (1-(\theta+\eps)){\bar x}_n]$ we obtain a contradiction. This is possible if $(1-(\theta+\eps)) {\bar x}_n > 2(\theta +\eps){\bar x}_n$. Since $\eps>0$ is arbitrary, this is satisfied for $1-\theta >2\theta$, hence for $\theta <1/3$. The prefactor ============== Existence of the prefactor {#S.preex} -------------------------- The function $u$ defined in satisfies the equation $$\label{uequation} x^2 u(x) = \int_0^x \int_{x{-}y}^{\infty} \, y K(y,z) u(y)u(z) e^{a^* (x-(y{+}z))}\,dz\,dy\,.$$ As explained in the introduction we need to assume from now on that $K$ is uniformly bounded below, that is $K$ satisfies with $\kappa=1$. We start with a consequence of the proof of Lemma \[L.ublargex\] that gives that $u$ has infinite first moment. \[L.unonintegrable\] The integral $\int_0^{\infty} x u(x)\,dx$ is not finite. This is just the observation that implies $$x^2 u(x) \leq \int_0^x \int_{x{-}y}^{\infty} \, y K(y,z) u(y)u(z)\,dz\,dy\,.$$ If we assume that $\int_0^{\infty} x u(x)\,dx< \infty$, we can apply the proof of Lemma \[L.ublargex\] to $u(x)$ and obtain exponential decay of $u$. But this contradicts the definition of $a^*$. \[L.uupperbound\] For any $A\geq 1$ there exists a constant $C_A>0$ such that $$\frac{1}{R} \int_1^{AR} u(x)\,dx \leq C_A\, \qquad \mbox{ for } R \geq \frac{2}{a^*}\,.$$ We denote $U(q)=\int_1^{\infty} e^{-qx} u(x)\,dx$ which is well-defined for all $q>0$. Then implies $$\label{utransformequation} -\partial_q U(q)=\int_0^{\infty} \int_0^{\infty} K(y,z) y u(y) u(z) e^{-a^*(y+z)}\Big\{ \int_{\max(1,y)}^{\max(1,y+z)} \frac{1}{x} e^{(a^*-q)x}\,dx \Big\}\,dy\,dz\,.$$ Symmetrizing and using the uniform lower bound on $K$ we find for $q \leq a^*/2$ that $$\label{uprimestimate} \begin{split} -\partial_q U(q) & \geq \int_1^{\infty} \int_1^{\infty}K(y,z) u(y) u(z) e^{-a^*(y+z)} \frac{y}{y+z} \frac{1}{a^*{-}q} \Big( e^{(a^*-q)(y+z)} - e^{(a^*-q)y}\Big)\,dy\,dz\\ & \geq c \int_1^{\infty} \int_1^{\infty} u(y) u(z) e^{-q(y+z)} \Big( 1 - \frac{y}{y+z} e^{-(a^*-q)z} - \frac{z}{y+z} e^{-(a^*-q)y}\Big)\,dy\,dz\\ & \geq c\, U^2(q)\,, \end{split}$$ where the last inequality follows since $ 1 - \frac{y}{y+z} e^{-(a^*-q)z} - \frac{z}{y+z} e^{-(a^*-q)y }\geq c>0$ for $y,z \geq 1$ and $q \in [0,a^*/2]$. Integrating inequality we find $$\frac{1}{U(q)} \leq \frac{1}{U(q_0)} + c(q-q_0) \qquad \mbox{ for } 0<q < q_0 \leq \frac{a^*}{2}\,.$$ If $1/U(q_0) - cq_0 <0$ it follows that there exists $0<q^*\leq q_0$ such that $1/U(q^*) \leq 0$ which gives a contradiction to the fact that $U$ is well-defined for all $q>0$. Since $q_0\leq \frac{a^*}{2}$ was arbitrary it follows that $U(q) \leq C/q$ for all $q\leq a^*/2$ and thus $$\int_1^{AR} e^{-qx} u(x)\,dx \leq \frac{C}{q} \qquad \mbox{ for all } q \in \Big (0,\frac{a^*}{2}\Big]\,.$$ Choosing $q=\frac{1}{R}$ the statement of the Lemma follows. \[C.urlimit\] Let $u_R(x):=u(Rx)$. Then there exists a sequence $R_j$ with $R_j \to \infty$ and a nonnegative measure $\mu \in {\cal M}^+(0,\infty)$ such that $u_{R_j} \wto \mu$ on any compact subset of $(0,\infty)$. Furthermore we have $$\label{muapriori1} \int_{(0,\infty)} e^{-qy}\mu(y)\,dy \leq \frac{C}{q} \qquad \mbox{ for all } q>0$$ and consequently $$\label{muapriori1b} \int_{(0,x]}\mu(y)\,dy \leq Cx \qquad \mbox{ for all } x>0\,.$$ \[L.munontrivial\] There exists a constant $C>0$ such that $\mu$ satisfies $$\label{muapriori1c} C x \leq \int_{(0,x]}\mu(x)\,dx \qquad \mbox{ for all } x >0\,.$$ Our goal is to derive a uniform lower bound on averages of $u$ via uniform lower bounds on $U$. To that aim we define $V(q):= \int_1^{\infty} x^{-\alpha} u(x) e^{-qx}\,dx$. [*Step 1:*]{} We claim for $U$ as in Lemma \[L.uupperbound\], that $$\label{munontrivial1} -\partial_q U(q) \leq K_1 \Big( U V^{\frac{1}{1{-}\alpha}} + U + 1\Big)\qquad \mbox{ for } q \in\Big(0, \frac{a^*}{2}\Big]\,.$$ We recall from Lemma \[L.uupperbound\] and split the integral on the right hand side into the following parts: $$\begin{split} \int_1^{\infty} \,dy \int_1^{\infty} \,dz \dots + &\int_0^1\,dy\int_1^{\infty}\,dz\dots + \int_1^{\infty} \,dy \int_0^1 \,dz \dots + \int_0^1\,dy\int_0^1\,dz\dots \\ & =: (I)+(II)+(III)+(IV)\,. \end{split}$$ Using Hölders inequality, we find $$\label{munontrivial2} \begin{split} |(I)|&\leq C \int_1^{\infty} \int_1^{\infty}\Big ( \Big(\frac{y}{z}\Big)^{\alpha} + \Big(\frac{z}{y}\Big)^{\alpha}\Big) u(y)u(z) e^{-q(y+z)} \,dy\,dz \\ &\leq C V(q) \int_1^{\infty} z^{\alpha} u(z) e^{-qz}\,dz \\ & \leq |\partial_q U|^{\alpha} |U|^{1{-}\alpha} |V| \leq \eps |\partial_q U| + C_{\eps} |U| |V|^{\frac{1}{1{-}\alpha}}\,. \end{split}$$ Next, using , we have $$\label{munontrivial3} \begin{split} |(II)| &\leq C \int_0^1 \int_1^{\infty} y^{1{-}\alpha} u(y) z^{\alpha} u(z)e^{-a^*(y+z)}\int_1^{y+z} \frac{e^{(a^*-q)x}}{x}\,dx \,dz\,dy\\ &\leq C \int_1^{\infty} z^{\alpha}u(z) \frac{e^{-qz}}{z}\,dz \leq C U\,. \end{split}$$ Similarly we estimate $$\label{munontrivial4} \begin{split} |(III)| & \leq \int_1^{\infty} \int_0^1 y^{1+\alpha} z^{-\alpha} u(y)u(z) e^{-a^*(y+z)}\int_y^{y+z} \frac{e^{(a^*-q)x}}{x}\,dx \,dz\,dy\\ & \leq C \int_1^{\infty} \int_0^1 y^{1{+}\alpha} z^{-\alpha} u(y)u(z) e^{-a^*(y+z)} \frac{z}{y} e^{(a^*-q)y}\,dz\,dy\\ & \leq C \int_1^{\infty} y^{\alpha} u(y) e^{-qy}\,dy\\ & \leq \eps |\partial_q U(q)| + C_{\eps} |U(q)|\,. \end{split}$$ Finally, we have $$\label{munontrivial5} \begin{split} |(IV)|& \leq C \int_0^1 \int_0^1 u(y) u(z) y^{1-\alpha} z^{-\alpha} \Big\{\int_1^{\max(1,y+z)} \frac{e^{(a^*-q)x}}{x}\,dx \Big\}e^{-a^*(y+z)} \,dy\,dz \\ & \leq C \int_0^1 \int_0^1 u(y) u(z) y^{1-\alpha} z^{1-\alpha} \,dy\,dz \leq C \end{split}$$ and follows from -, noticing that $|\partial_q U(q)|=-\partial_q U(q)$. There exists $K_2>0$ such that $$\label{vestimate} V(q) \leq \frac{K_2}{q^{1-\alpha}}\qquad \mbox{ for all } q \in \Big(0,\frac{a^*}{2}\Big]\,.$$ Indeed, the definition of $V$ implies that for $q>\hat q>0$ $$\label{vestimate1} \begin{split} |V(q)-V(\hat q)| & \leq \Big| \int_1^{\infty} y^{-\alpha} u(y) \Big( e^{-qy} - e^{-\hat q y}\Big)\,dy \Big|\\ & \leq |q-\hat q|^{\alpha} \Big| \int_1^{\infty} ((q-\hat q)y)^{-\alpha} \Big( 1 - e^{-y(q-\hat q)}\Big) e^{-\hat q y} u(y)\,dy\\ & \leq K_3 |q-\hat q|^{\alpha} U(\hat q)\,. \end{split}$$ Recall that we have proved in Lemma \[L.uupperbound\] that $U(q) \leq C/q$ for $q \in (0,a^*/2]$. Thus, we obtain for $\hat q=2^{-(n+1)}$ and $q=2^{-n}$ that $$V\big(2^{-(n+1)}\big) \leq V(2^{-n}) + C 2^n 2^{-(n+1)\alpha}$$ and iterating this formula we find $$V\big(2^{-(n+1)} \big)\leq C \Big ( 1 + \sum_{i=1}^n 2^{i(1-\alpha)}\Big) \leq C 2^{(n+1)(1-\alpha)}$$ and this proves . We claim that there exists $C>0$ such that $$\label{munontrivial6} U(q) \geq \frac{C}{q} \qquad \mbox{ for all } q \in \Big(0,\frac{a^*}{2}\Big]\,.$$ Suppose that is not true. Then, given an arbitrarily small $\eps>0$, there exists $q^*\in \Big(0,\frac{a^*}{2}\Big]$ such that $U(q^*) \leq \frac{\eps}{q^*}$. We can assume without loss of generality that $U \geq 1$. Integrating and using we find $$\label{mu2a} U\Big( \frac{q}{2}\Big) \leq U(q) \exp \Big( K_1 {K_2}^{\frac{1}{1-\alpha}} \ln 2\Big)$$ for all $q \in \Big(0,\frac{q^*}{2}\Big]$. For the following let $K:=\max\Big(K_1,K_2,K_3, \exp\big(K_1 K_2^{\frac{1}{1-\alpha}} \ln 2 \big)\Big)$. It follows in particular from that $$\label{mu2} U\Big( \frac{q^*}{2}\Big) \leq \frac{1}{\frac{q^*}{2}} \frac{\eps}{2} K\,.$$ Estimate also implies that $$\label{mu2b} V\Big( \frac{q^*}{2}\Big) \leq V(q^*) + K \big( \frac{q^*}{2}\big)^{\alpha} U\big( \frac{q^*}{2}\big) \leq \Big( \frac{q^*}{2}\Big)^{\alpha-1} \Big( \frac{K}{2^{1-\alpha}} + K^2 \frac{\eps}{2} \Big)\,.$$ We can find $\theta>1$ such that for sufficiently small $\eps$ $$\label{theta2} \frac{K}{2^{1-\alpha}} + \frac{\eps}{2} K^2 \exp \Big( K\sum_{i=0}^{\infty} \theta^{-\frac{i}{1-\alpha}} \Big) \leq \frac{K}{\theta}\,.$$ We can now prove by induction that $$\label{V1} V\Big(\frac{q^*}{2^{n}}\Big) \leq \frac{K}{\theta^n} \Big( \frac{q^*}{2^n}\Big)^{\alpha-1}$$ and $$\label{U1} U\Big( \frac{q^*}{2^n}\Big) \leq \Big( \frac{q^*}{2^n}\Big)^{-1} \frac{\eps}{2^n} \exp\Big( K\sum_{i=0}^{n} \theta^{-\frac{i}{1-\alpha}}\Big)\,.$$ In fact, for $n=1$ follows from and from and . To go from $n$ to $n+1$ we we use , and to obtain, assuming $\theta <2$, that $$\label{V2} \begin{split} V\Big(\frac{q^*}{2^{n+1}}\Big)&\leq V\Big(\frac{q^*}{2^n}\Big) + K\Big(\frac{q^*}{2^{n+1}}\Big)^{\alpha} U\Big(\frac{q^*}{2^{n}}\Big) \\ & \leq \frac{K}{\theta^n} \Big( \frac{q^*}{2^n}\Big)^{\alpha-1} + K^2 \Big( \frac{q}{2^{n+1}}\Big)^{\alpha-1} \frac{\eps}{2^{n+1}} \exp\Big( K \sum_{i=0}^{n} \theta^{-\frac{i}{1-\alpha}} \Big)\\ &= \Big( \frac{q^*}{2^{n+1}}\Big)^{\alpha-1} \frac{1}{\theta^n} \Big( \frac{K}{2^{1-\alpha}} + \frac{\eps}{2} K^2 \exp \Big( K \sum_{i=0}^n \theta^{-\frac{i}{1-\alpha}} \Big)\\ &\leq \Big( \frac{q^*}{2^{n+1}}\Big)^{\alpha-1} \frac{K}{\theta^n}\,. \end{split}$$ Going back with $\eqref{V2}$ to we find $$\label{U2} \begin{split} U\Big( \frac{q^*}{2^{n+1}}\Big) & \leq U\Big( \frac{q^*}{2^n}\Big) \exp \Big( K \theta^{-(n+1)/(1-\alpha)}\Big)\\ & \leq\Big( \frac{q^*}{2^n}\Big)^{-1} \frac{\eps}{2^{n+1}} \exp \Big( K \sum_{i=0}^{n+1} \theta^{-i/(1-\alpha)}\Big) \,. \end{split}$$ Thus, and prove and . In particular, we have a constant $C>0$ such that $U(q^*/2^n) \leq \frac{C}{q^*}$ which gives a uniform bound on $U$. As a consequence of we also get a uniform bound on $|\partial_q U|$ and thus a contradiction to Lemma \[L.unonintegrable\]. This finishes the proof of . We show that $$\label{munontrivial8} \int_1^{BR} u(x)\,dx \geq cR$$ for sufficiently large $B$ from which the claim of the Lemma follows. Indeed, choosing $q=\frac{1}{R}$ in gives $$R \leq \int_1^{BR} e^{-\frac{x}{R}} u(x)\,dx + \int_{BR}^{\infty} e^{-\frac{x}{R}} u(x)\,dx\,.$$ On the other hand, using Lemma \[L.uupperbound\], we have $$\begin{split} \int_{BR}^{\infty} e^{-\frac{x}{R}} u(x)\,dx &\leq \sum_{n=0}^{\infty} \int_{2^n BR}^{2^{n+1}BR} e^{-\frac{x}{R}} u(x)\,dx\\ & \leq \sum_{n=0}^{\infty} e^{-2^nB} \int_{2^n BR}^{2^{n+1}BR}u(x)\,dx \leq R \sum_{n=0}^{\infty} B 2^n e^{-2^nB}\,. \end{split}$$ Since $\sum_{n=0}^{\infty} B 2^n e^{-2^nB}<\frac 1 2$ for sufficiently large $B$ estimate follows. finally, the lower bound in is a consequence of . \[L.muequation\] $\mu$ is a weak solution to . Equation means that for $\phi \in C_0([0,\infty))$ it holds $$\label{weakmuequation} \int_0^{\infty} x^2 \phi(x) \mu(x)\,dx = \frac{1}{a^*} \int_0^{\infty} \int_0^{\infty} K(x,y) y \mu(x)\mu(y) \phi(x+y)\,dx\,dy\,.$$ We now take $\phi \in C_0^1([0,\infty))$. From we deduce that $u_R$ satisfies $$\label{urequation} \begin{split} x^2 u_R(x) &= \frac{1}{R^2} \int_0^{Rx} \int_{Rx-y}^{\infty} K(y,z) y u(y) u(z) e^{a^*(Rx-(y+z))} \,dz\,dy \\ & = R \int_0^x \int_{x-y}^{\infty} K(y,z) y u_R(y) u_R(z) e^{a^*R(x-(y+z))}\,dz\,dy\,. \end{split}$$ Hence, $$\label{urweakequation} \int_0^{\infty} x^2 u_R(x)\phi(x)\,dx = \int_0^{\infty} \int_0^{\infty} K(y,z) y u_R(y) u_R(z) \psi_R(y,z) \,dy\,dz$$ with $$\label{psidef} \psi_R(y,z) := R \int_y^{y+z} \phi(x) e^{a^*R(x-(y+z))}\,dx\,.$$ The function $\psi_R$ satisfies $$\label{psi1} |\psi_R(y,z)| \leq \|\phi\|_{L^{\infty}} R z\,$$ and $$\label{psi2} \psi_R(y,z) = \frac{1}{a^*}\int_0^{a^*Rz} \phi\Big( y+z - \frac{\xi}{a^*R}\Big)e^{-\xi}\,d\xi \to \frac{1}{a^*}\phi(y+z) \quad \mbox{ as } R \to \infty \mbox{ for all }y,z>0\,.$$ Furthermore $$\label{psi3} \begin{split} \Big| \psi_R(y,z) - \frac{1}{a^*}\phi(y+z)\Big| & \leq \Big| \frac{1}{a^*} \int_0^{a^*Rz} \Big[\phi\Big(y+z - \frac{\xi}{a^*R}\Big) -\phi(y+z)\Big] e^{-\xi}\,d\xi\Big|\\ & \qquad + \Big| \int_{a^*Rz}^{\infty} \frac{\phi(y+z)}{a^*} e^{-\xi}\,d\xi\Big|\\ & \leq\|\phi'\|_{L^{\infty}} \int_0^{a^*Rz} \frac{\xi}{a^*R} e^{-\xi}\,d\xi + \|\phi\|_{L^{\infty}} e^{-a^*Rz}\\ & \leq \|\phi'\|_{L^{\infty}} \frac{C}{R} + \|\phi\|_{L^{\infty}} e^{-a^*Rz}\,. \end{split}$$ For a large constant $L$ we split the integral on the right hand side of into the parts $$\int_0^{\infty}\,dy \int_0^{L/R} \,dz \dots + \int_0^{\infty}\,dy \int_{L/R}^{\infty}\,dz \dots =:(I)+(II)\,.$$ We recall that due to we have $\int_0^L x^{1-\alpha} u(x)\,dx \leq C_L$. Hence $$\label{psi4} \int_0^{L/R} x^{1-\alpha} u_R(x)\,dx = R^{\alpha-2} \int_0^L x^{1-\alpha} u(x)\,dx \leq \frac{C_L}{R^{2-\alpha}}\,.$$ Second, Lemma \[L.uupperbound\] implies $\int_{L/R}^A u_R(x)\,dx \leq C$. Hence $u_R \chi_{\{x \geq L/R\}} \wto \mu$ on $[0,A]$. Combining this with properties - and we find $$\label{urweak1} |(I)|\leq C R \int_0^{L/R} u_R(z) z^{1-\alpha} \,dz \leq \frac{C_L}{R^{1-\alpha}} \to 0 \qquad \mbox{ as } R \to \infty \mbox{ for fixed }L\,$$ and $$\label{urweak2} \limsup_{R\to\infty} \Big | (II) - \int_0^{\infty} \int_0^{\infty} K(y,z) y \mu(y)\mu(z)\frac{1}{a^*}\psi(y+z)\,dy\,dz\Big| \leq C e^{-a^*L}\,.$$ Since $L$ was arbitrary the desired result follows. Uniqueness of the prefactor {#S.preuni} --------------------------- In this section we are going to show that under additional assumptions on the kernel $K$ the solution to is unique. \[L.profunique\] Let $K$ satisfy the assumptions of Theorem \[T4\]. Then, if $\eps>0$ is sufficiently small, the only solution to that satisfies and is the constant one, i.e. $\mu \equiv \mu_0:=2a^*/\int_0^1 K(s,1-s)\,ds$. Our proof is somewhat similar in spirit, though simpler, as the proof of uniqueness of self-similar solutions to the coagulation equation with kernels satisfying and . We can assume without loss of generality that $2 a^*=1$ such that $|\mu_0-1|\leq C \eps$. We introduce the Laplace-transform of $\mu$ via $$\tilde U(q)=\int_{(0,\infty)} e^{-qx}\mu(x)\,dx\,, \qquad q>0\,.$$ Using the usual dyadic decomposition we can derive from that $$\label{muapriori2} \int_0^x y^{-\alpha} \mu(y)\,dy \leq C x^{1-\alpha}\qquad \mbox{ for all } x>0\,.$$ Furthermore, $\tilde U$ satisfies the equation $$\label{Uequation} -\partial_q\tilde U= {\tilde U}^2 + \int_0^{\infty} \int_0^{\infty} \Big(K(x,y)-1\Big) e^{-q(x+y)} \mu(x)\mu(y)\,dx\,dy=: {\tilde U}^2 +{\cal M}\,.$$ We first estimate ${\cal M}$. By symmetry, using assumption , we have $${\cal M} \leq C \eps \int_0^{\infty} \int_0^{x/2} \Big( \frac{x}{y}\Big)^{\alpha} \mu(x-y) \mu(y) e^{-qx}\,dy\,dx\,.$$ For any $R>0$ we estimate, using and , $$\begin{split} \int_R^{2R} \int_0^{x/2} \Big(\frac{x}{y}\Big)^{\alpha} \mu(x-y)\mu(y)\,dy\,dx & = \int_0^R \int_{\max(R,2y)}^{2R} \Big(\frac{x}{y}\Big)^{\alpha} \mu(x-y)\mu(y)\,dx\,dy\\ & \leq C R^{\alpha} \int_0^R \frac{\mu(y)}{y^{\alpha}}\,dy \int_{0}^{2R} \mu(x)\,dx \leq C R^2\,. \end{split}$$ and this implies $$\label{uni2} \begin{split} {\cal M}& \leq C \eps \sum_{n=-\infty}^{\infty} \int_{2^n}^{2^{n+1}} e^{-qx} \int_0^{x/2} \Big(\frac{x}{y}\Big)^{\alpha} \mu(x-y)\mu(y)\,dy\,dx\\ & \leq C\eps \sum_{n=-\infty}^{\infty} e^{-q2^n} \big( 2^n\big)^2 \leq \frac{C \eps}{q^2}\,. \end{split}$$ Next, we claim that $$\label{Wapriori} \Big|\tilde U(q)-\frac{1}{q}\Big| \leq \frac{C\eps}{q} \qquad \mbox{ for all } q>0\,.$$ Suppose otherwise. Then there exists for any arbitrarily large $K$ a number $q^*>0$ such that $|\tilde U(q^*)-\frac{1}{q^*}|> \frac{K\eps}{q^*}$. Define $G(x):=q^*U(q^*x)$ such that $$-\partial_x G(x) = G(x)^2 + O\Big(\frac{\eps}{x^2}\Big) \qquad \mbox{ and } |G(1)-1|\geq K\eps\,.$$ Let $\bar G$ be the solution of $-\partial_x \bar G = {\bar G}^2$ with $\bar G(1)=G(1)$, that is $\bar G (x) = \frac{G(1)}{1+G(1)(x-1)}$ and $\bar G(1/2)= \frac{2G(1)}{2-G(1)}$. By the standard theory of differential equations we find $|G(1/2) - {\bar G}(1/2)|\leq C\eps$ and thus $$\big| G(1/2) -2\big| \geq \Big| \frac{2G(1)}{2-G(1)} -2\Big| - C\eps = \frac{4}{|2-G(1)|}|G(1)-1| - C\eps\,.$$ If $G(1) \geq 2$ we get immediately a contradiction to the fact that $\tilde U(q)$ is defined for all $q>0$. On the other hand, we know that $G(1) \geq c_1>0$. Hence we obtain that $$\big| G(1/2) -2\big| \geq \theta |G(1)-1| \qquad \mbox{ for some } \theta>2\,,$$ if $K$ is sufficiently large. Thus, we have obtained $$\Big| U \Big( \frac{q^*}{2}\Big) - \frac{1}{\frac{q^*}{2}}\Big| >\frac{\theta}{2} \frac{K\eps}{\frac{q^*}{2}}\,.$$ We can iterate the argument to get $$\Big| U \Big( \frac{q^*}{2^n}\Big) - \frac{1}{\frac{q^*}{2^n}}\Big| > \Big(\frac{\theta}{2}\Big)^n\frac{K\eps}{\frac{q^*}{2^n}}\,,$$ which gives again a contradiction since $\theta/2>1$. Hence, we have proved . To proceed we introduce $$\nu(x):=\mu(x) - \mu_0 \qquad \mbox{ with } \qquad \mu_0 = \frac{1}{\int_0^1 K(s,1-s)\,dx}$$ and $$V(q)=\int_0^{\infty} e^{-qx}\nu(x)\,dx\,.$$ Since $$x\nu(x) = 2\mu_0 \int_0^x K(y,x-y)\nu(y)\,dy + \int_0^x K(y,x-y) \nu(x-y)\nu(y)\,dy$$ we find $$\begin{split} -\partial_q V& = \frac{2\mu_0}{q} V + 2 \mu_0 \int_0^{\infty} e^{-qx} \int_0^x \Big( K(y,x-y)-1\Big)\nu(y)\,dy\,dx\\ &\qquad + \int_0^{\infty} e^{-qx} \int_0^x K(y,x-y) \nu(x-y)\nu(y)\,dy\,dx\,. \end{split}$$ Integrating this equation gives $$\label{uni3a} \begin{split} q^{2\mu_0} V(q)& = -2 \mu_0 \int_0^q \eta^{2\mu_0} \int_0^{\infty} e^{-\eta x} \int_0^x \Big(\big( K(y,x -y) -1\big)\nu(y)\\ &\qquad\qquad + K(y,x-y) \nu(x-y)\nu(y)\Big) \,dy\,dx\,d\eta\,. \end{split}$$ We introduce the norm $$\label{normdef} \|V\|:=\sup_{q>0} | qV(q)|$$ such that $|\mu_0-1|\leq C\eps$ and imply $\|V\| \leq C\eps$. Next, we claim that $$\label{uni3} \sup_{q>0} \Big| q \int_0^q \Big(\frac{\eta}{q}\Big)^{2\mu_0} \int_0^{\infty} e^{-\eta x} \int_0^x \big( K(y,x-y) -1\big)\nu(y)\,dy\,dx\,d\eta \Big| \leq C \eps \|V\|$$ and $$\label{uni4} \sup_{q>0} \Big| q \int_0^q \Big( \frac{\eta}{q}\Big)^{2\mu_0} \int_0^{\infty} e^{-\eta x } \int_0^x K(y,x-y) \nu(x-y)\nu(y)\,dy\,dx\,d\eta \Big| \leq C \eps \|V\|^2\,.$$ Then, , and imply $$\| V\| \leq C \eps \|V\| + C \|V\|^2$$ and since $\|V\|\leq C\eps$ it follows that $V=0$ if $\eps$ is sufficiently small. It remains to prove and which we will do by contradiction. By scaling we can assume that $\eps=1$. To prove we thus assume that there exist sequences $(q_n), (\nu_n)$ and $(K_n)$ such that $|q_n V(q_n)| \to 0$ as $n \to \infty$ while $$\label{uni5} \Big| q_n \int_0^{q_n} \Big(\frac{\eta}{q_n}\Big)^{2\mu_0} \int_0^{\infty} e^{-\eta x} \int_0^x \big( K_n(y,x-y) -1\big)\nu_n(y)\,dy\,dx\,d\eta\Big| \geq 1\,.$$ By assumption we can assume that there exists $K=K(x,y)$ such that $K_n \to K$ locally uniformly on $(0,\infty)^2$ and $K$ satisfies the same bounds as $K_n$. We change variables in , such that $\hat x=q_n x$, $\hat y = q_n y$ and $\hat \eta = \eta/q_n$, define $\hat \nu_n(\hat y)=\nu_n(y)$ and obtain, dropping the hats in the notation, that turns into $$\label{uni6} \Big|\int_0^1 \eta^{2\mu_0} \int_0^{\infty} e^{-\eta x} \int_0^x \big(K_n(y,x-y)-1\big)\nu_n(y)\,dy\,dx\,d\eta\Big | \geq 1\,.$$ Changing the order of integration gives $$\label{uni7} \Big| \int_0^{\infty} \nu_n(y) \int_y^{\infty} \int_0^1 e^{-\eta x}\eta^{2\mu_0} \big( K_n(y,x-y)-1\big) \,d\eta \,dx\,dy\Big| \geq 1\,.$$ If we define $\phi(x):=\int_0^1 e^{-\eta x} \eta^{2\mu_0}\,d\eta$ we have that $\phi \in C^0(0,\infty) \cap L^{\infty}(0,\infty)$ with $\phi(x) \leq \frac{C}{x^{1+2\mu_0}}$ as $x \to \infty$. The bound then implies that for $\psi_n(y):=\int_y^{\infty} \phi(x) \big( K_n(y,x-y)-1\big) \,dx$ we have $\psi_n(y) \leq \frac{C}{y^{\alpha}}$ for $y \geq 1$, $\psi_n(y) \leq \frac{C}{y^{2\mu_0}}$ for $y \geq 1$ and that $\psi_n\to \psi$ locally uniformly on $(0,\infty)$. By assumption $q_n V(q_n) \to 0$. This means, after using the rescaling above, that $\nu_n \wto 0$ as $n \to \infty$. The bounds on $\nu_n$ and $\psi_n$ and the fact that $2 \mu_0 \geq 2-C\eps$ imply that $\int_0^{\infty} \psi_n(y)\nu_n(y)\,dy \to 0$ which gives a contradiction to and finishes the proof of . The proof of follows similarly and we omit the details here. The authors acknowledge support through the CRC 1060 [*The mathematics of emergent effects* ]{} at the University of Bonn, that is funded through the German Science Foundation (DFG). [^1]: Institute of Applied Mathematics, University of Bonn, Endenicher Allee 60, 53115 Bonn, Germany [^2]: Institute of Applied Mathematics, University of Bonn, Endenicher Allee 60, 53115 Bonn, Germany

ArXiv

--- abstract: | Interval routing is a space efficient method to realize a distributed routing function. In this paper, we show that every circular-arc graph allows a shortest path strict 2-interval routing scheme, i.e., a routing function that only implies shortest paths can be realized in every circular-arc graph, where at most two (strict) intervals are stored at the ends of every edge. Since circular-arc graphs do not allow shortest path 1-interval routing schemes in general, the result implies that the class of circular arc graphs has strict compactness 2, which was a hitherto open question. Further, we show that the constructed 2-interval routing scheme can be represented as an 1-interval routing scheme with at most one additional interval assigned for each vertex of the graph and we outline an algorithm to implement the routing scheme for $n$-vertex circular-arc graphs in $\mathcal{O}(n^2)$ time. [**Keywords:**]{} interval routing, compact routing, circular-arc graphs, cyclic orders author: - | Frank Gurski\ Heinrich-Heine-University Düsseldorf\ Institute of Computer Science\ D-40225 Düsseldorf\ `[email protected]` - | Patrick Gwydion Poullie[^1]\ University of Zurich\ Department of Informatics\ Communication Systems Group\ CH-8050 Zürich\ `[email protected]` bibliography: - 'bibfile.bib' title: 'Interval Routing Schemes for Circular-Arc Graphs' --- Introduction ============ Routing is an essential task that a network of processors or computers must be able to perform. Interval routing is a space-efficient solution to this problem. Sets of consecutive destination addresses that use the same output port are grouped into intervals and then assigned to the respective port. In this way, the required storage space is greatly reduced compared to the straight-forward approach, in which an output port for every destination address is stored in particular. Of course, the advantage in space efficiency depends heavily on the number of intervals assigned to the output ports, which, in turn, depends on the address and network topology. Interval routing was first introduced in [@Santoro82routingwithout; @Santoro85labelling]; for a discussion of interval routing we refer the reader to [@Leeuwen87interval; @FredericksonJ88; @Bakker91linearinterval] and for a detailed survey to [@Gavoille00asurvey]. When theoretical aspects of interval routing are discussed, the network is represented by a directed (symmetric) graph $G=(V,A)$. As with many routing methods, the problem is that space efficiency and provided path lengths are conflicting goals [@65953]: as shown in [@Guevremont98worstcase], for every $n\in\mathbb{N}$, there exists an $n$-vertex graph $G_n$ such that for every shortest path interval routing scheme (for short IRS) for $G_n$ the maximal number of intervals per arc is only bounded by $\Theta(n)$. Also, it is NP-hard to determine the most space efficient shortest path IRS for a given graph [@Eilam2002; @Flammini1997]. Other worst case results can be found in [@savio99onthespace]. On the other hand, there are many graph classes that are known to allow shortest path IRSs with a constantly bounded number of intervals on all arcs [@FG98]. If this number is a tight bound, it is denoted the compactness of the graph. The compactness of undirected graphs is defined as the compactness of their directed symmetric version[^2] and the compactness of a graph class as the smallest $k$, such that every graph in this class has at most compactness $k$, if such $k$ exists. In this paper we show that the class of circular-arc graphs has (strict) compactness 2, by giving an algorithm to construct a corresponding IRS. A first and different approach to realize space efficient, shortest path routing in circular-arc graphs can be found in [@f.dragan:new]. In [@GP08] an $\mathcal{O}(\log(|V|))$-bit distance labeling scheme is developed which allows in circular-arc graphs for each pair of vertices to compute their exact distance in $\mathcal{O}(1)$ time. In [@DYL06] an $\mathcal{O}(\log^2(|V|))$-bit routing labeling scheme is developed which allows in circular-arc graph for every vertex to make a routing decision in $\mathcal{O}(1)$ time. The resulting routing path has at most 2 more edges than the optimal (shortest) one. Nevertheless, our result is interesting, since interval graphs and unit circular-arc graphs are included in the class of circular-arc graphs and known to have compactness 1 [@interval; @FG98] while for the compactness of circular arc graphs hitherto only a lower bound of 2 was known. Rather than by intervals on a line (as with interval graphs) circular-arc graphs can be represented by arcs on a circle, that is to say, every circular-arc graph is the intersection graph of a set of arcs on a circle. The first polynomial-time (${\mathcal{O}}(|V|^3)$) algorithm to recognize circular-arc graphs and give a corresponding circular-arc model can be found in [@firstCAG]. An algorithm with linear runtime for the same purpose is given in [@springerlink:10.1007/s00453-003-1032-7]. A maximum clique in a circular-arc graph can be computed much faster than in an arbitrary graph: an according algorithm that operates in ${\mathcal{O}}(|E|+|V|\cdot\log(|V|))$ time or even in ${\mathcal{O}}(|E|)$ time, if the circular-arc endpoints are given in sorted order, is provided in [@Bhattacharya1997336]. A discussion of circular-arc graphs can be found in [@Hsu:1995:OAR:204695.204697]. Circular-arc graphs have, amongst others, applications in cyclic scheduling, compiler design [@Golumbic:2004:AGT:984029; @tucker:493], and genetics [@nla.cat-vn2675280]. The paper is organized as follows. In Section \[chap:pre\] we give the definitions of cyclic orders, circular-arc graphs, and IRSs. In Definition \[defiIRS\] we introduce a specialized version of the well known definition of interval routing schemes. Our definition allows to state Corollary \[tropicthunder\], a powerful local criterion to prove that a graph allows a shortest path $k$ strict IRS, for a fixed $k$. We apply this criterion in Section \[main\] to show Theorem \[mainTheo\], which is the main result of this paper and states that every circular-arc graph allows a shortest path strict 2-interval routing scheme. Therefore in Subsection \[constrCO\] we chose an arbitrary circular-arc graph $G=(V,A)$ and fix a cyclic order ${\mathscr{L}}$ on $V$. In Subsection \[furtherDefs\] we partition the vertices in $V-\{v\}$ in three sets $A_v,B_v,C_v$ in dependency on ${\mathscr{L}}$ and an arbitrary vertex $v$. In Subsection \[2-app\] we show that the vertices of the three sets $A_v,B_v,C_v$ can be assigned to the edges incident to $v$, such that the criterion is fulfilled. In Section \[improving\] we show that the shortest path 2-SIRS, that is implied by the proof of Theorem \[mainTheo\], utilizes a number of intervals that can be bounded close to the number that is allowed in an 1-IRS. In Section \[implement\] we outline how to implement the implied IRS in $\mathcal{O}(|V|^2)$ time. Preliminaries\[chap:pre\] ========================= Cyclic Orders {#cyclicOrders} ------------- Cyclic orders are relevant for the two main subjects of this paper, namely circular-arc graphs and non-linear interval routing schemes. Just like a permutation, a cyclic order implies an order on a set with the only difference being that there is no distinct first, second, or $n$-th element. A *cyclic order* on a set $A$ with $n$ elements is a function ${\mathscr{C}}\colon A\rightarrow A$ such that $\forall a\in A\colon\forall i\in\mathbb{N}^+\colon{\mathscr{C}}^i(a)= a ~\Leftrightarrow~ i\bmod n=0$.[^3] We call ${\mathscr{C}}(a)$ the *successor of $a$ (in ${\mathscr{C}}$)*, denoted by $a\succ_{\mathscr{C}} {\mathscr{C}}(a)$. Every cyclic order is a bijective mapping. Let ${\mathscr{C}}$ be a cyclic order on a set $A=\{a_1,a_2,\ldots,a_n\}$. One could imagine ${\mathscr{C}}$ as an arrangement of $A$ on a clock face, for an $n$-hour clock. *To go through ${\mathscr{C}}$ beginning at $a_i\in A$* means to consider the elements in $A$ as they appear in ${\mathscr{C}}$, beginning at $a_i$. Next we define subsets of $A$ that appear consecutive in ${\mathscr{C}}$; the subsequent corollary states that we can combine these [“intervals”]{}, if they appear consecutive in ${\mathscr{C}}$. \[interval\] For a cyclic order ${\mathscr{C}}$ on a set $A$ and $a,b\in A$, we recursively define the *ring-interval from $a$ to $b$ (in ${\mathscr{C}}$)* as $$[a,b]_{\mathscr{C}}\colonequals \begin{cases} \{a\}&\text{if }a=b\\ \{a\}\cup[{\mathscr{C}}(a),b]_{\mathscr{C}}&\text{else.}\\ \end{cases}$$ The *ring-sequence from $a$ to $b$ (in ${\mathscr{C}}$)*, denoted by ${{\mathcal{S}}}[a,b]_{\mathscr{C}}=(a,\ldots,b)$, is the order on $[a,b]_{\mathscr{C}}$ in which the elements are added to ring-interval $[a,b]_{\mathscr{C}}$ by its recursive definition. \[ringkombi\] Let ${\mathscr{C}}$ be a cyclic order on a set $A$ and $a,b,c,d\in A$. If $[a,b]_{\mathscr{C}}\cap[c,d]_{\mathscr{C}}=\emptyset$ and $b\succ_{\mathscr{C}}c$, we have $[a,b]_{\mathscr{C}}\cup[c,d]_{\mathscr{C}}=[a,d]_{\mathscr{C}}$. Circular-Arc Graphs {#sub:CAG} ------------------- We assume that the reader is familiar with basic graph theoretical definitions. The *intersection graph* of a family of sets is the graph where the vertices are the sets, and the edges are the pairs of sets that intersect. Every graph is the intersection graph of some family of sets. A graph is an *interval graph* if it is the intersection graph of a finite set of intervals (line segments) on a line and an *unit interval graph* if these intervals have unit length. A graph $G$ is a *circular-arc graph* if it is the intersection graph of a finite set of arcs on a circle; the latter we call an *arc model of $G$*. Since an interval graph is a special case of a circular-arc graph, namely a circular-arc graph that can be represented with a set of arcs that do not cover the entire circle, we define the set of *real circular-arc graphs* as the set of circular-arc graphs that are not interval graphs. It is obvious that every real circular-arc graph is connected. An *unit circular-arc graph* is a circular-arc graph that has an arc model in which the arcs have unit length. For a survey on circular-arc graphs see [@Lin20095618] and for the definition of further special graph classes see the book of Brandstädt et al. [@BLS99]. A real circular-arc graph is illustrated in Figure \[fig:animals\]. A difference between interval graphs and circular-arc graphs, that is worth mentioning, is that the maximal cliques of interval graphs can be associated to points of the [“interval model”]{} and therefore an interval graph can have no more maximal cliques than vertices. In contrast, circular-arc graphs may contain maximal cliques that do not correspond to points of some arc model [@Lin20095618]. In fact, just as in arbitrary graphs, the number of maximal cliques in circular-arc graphs can grow exponentially in the size of the graph [@firstCAG]. The following definition and the subsequent corollary formalize important graphic properties of circular arc graphs. \[lcrc\] Let $G=(V,E)$ be a circular-arc graph, $M_G$ be an arc model of $G$, and $C$ be the circle of $M_G$. To each point $p$ on $C$ corresponds a clique that contains the vertices whose corresponding arcs intersect $p$. Let ${\mathsf{X}}\subseteq{\mathsf{Y}}$ be the set of cliques that are maximal with respect to inclusion among the set ${\mathsf{Y}}$ of cliques that correspond to points on $C$. Pair $({\mathsf{X}},{\mathscr{C}})$, where ${\mathscr{C}}$ is the cyclic order on ${\mathsf{X}}$ that is implied by ordering the elements in ${\mathsf{X}}$ as their corresponding points appear (clockwise) on $C$, is called a *clique-cycle* for $G$. For every vertex $v\in V$ that has at least one neighbor exist two cliques ${\mathsf{a}},{\mathsf{b}}\in{\mathsf{X}}$ such that every clique ${\mathsf{x}}\in{\mathsf{X}}$ contains $v$ if and only if ${\mathsf{x}}\in[{\mathsf{a}},{\mathsf{b}}]_{\mathscr{C}}$. We call ${\mathsf{a}}$ the *left clique of $v$*, denoted by ${\alpha(v)}$, and ${\mathsf{b}}$ the *right clique of $v$*, denoted by ${\beta(v)}$. \[remarkDoms\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$ and $d\in V$ a dominating vertex. Then the left clique ${\alpha(d)}$ and the right clique ${\beta(d)}$ of $d$ are not unique, contrary to every non dominating vertex, but can be chosen arbitrarily with ${\beta(d)}\succ_{\mathscr{C}}{\alpha(d)}$ being the only constraint. When circular-arc graphs are discussed in this paper, the argumentation is based on their clique-cycles. Note that a clique-cycle can be easily deduced from an arc model of a circular-arc graph. Consider the arc model of a circular-arc graph and two intersecting arcs $v,w$. There are two kinds in which $v$ and $w$ can intersect. The intersection of $v$ and $w$ constitutes another arc. This is the case if arc $v$ is included in $w$ or vice versa, if they are congruent, or if exactly one endpoint of $v$ lies in $w$ and vice versa. The intersection of $v$ and $w$ constitutes two arcs (the corresponding vertices will be defined as counter vertices in the subsequent definition). The latter applies if $v$ covers one part of the circle and $w$ the other part and both endpoints of $v$ lie in $w$ and vice versa. Although the arcs cover very different areas of the circle, it is impossible to say which of the two arcs [“reaches further to the left or right”]{}, when the point of view is the middle of the circle. This problem is addressed in the next two definitions. \[reaches\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$. We call $w\in V$ a *counter vertex* of $v\in V$, if $v$ and $w$ are adjacent and $$\nexists {\mathsf{a}},{\mathsf{b}}\in{\mathsf{X}}\colon [{\mathsf{a}},{\mathsf{b}}]_{\mathscr{C}}=\{{\mathsf{c}}\in{\mathsf{X}}~|~v\in{\mathsf{c}}\wedge w\in{\mathsf{c}}\}.$$ We call $(v,w)$ a *pair of counter vertices* and denote the set of all counter vertices of $v$ by ${{\mathcal{C}}({v})}$.[^4] \[defi:reaches\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$, $v,w$ two adjacent vertices with $v\notin{{\mathcal{C}}({w})}$ and ${\mathsf{x}}\in {\mathsf{X}}$ a clique with $v,w\in {\mathsf{x}}$. We say, *$w$ reaches at least as far to the right as $v$* if $[{\mathsf{x}},{\beta(v)}]_{\mathscr{C}}\subseteq [{\mathsf{x}},{\beta(w)}]_{\mathscr{C}}$, and *$w$ reaches further to the right than $v$* if $[{\mathsf{x}},{\beta(v)}]_{\mathscr{C}}\subset [{\mathsf{x}},{\beta(w)}]_{\mathscr{C}}$. Analogous, we define the partial relations *$w$ reaches at least as far to the left as $v$* by $[{\alpha(v)},{\mathsf{x}}]_{\mathscr{C}}\subseteq [{\alpha(w)},{\mathsf{x}}]_{\mathscr{C}}$ and *$w$ reaches further to the left than $v$* by $[{\alpha(v)},{\mathsf{x}}]_{\mathscr{C}}\subset [{\alpha(w)},{\mathsf{x}}]_{\mathscr{C}}$. The next corollary follows directly from the arc model of a circular-arc graph. \[adjacentToVorCoV\] Let $G=(V,E)$ be a circular-arc graph and $v\in V$ be a vertex with ${{\mathcal{C}}({v})}\neq\emptyset$. Then every vertex in $V$ is adjacent to $v$ or to every vertex in ${{\mathcal{C}}({v})}$, in other words, $v$ and any of its counter vertices constitute a dominating set. Interval Routing Schemes {#IRS} ------------------------ In this section we describe interval routing when the underlying network is modeled by a directed graph $G = (V, A)$, where $V$ represents the nodes of the network and $A$ represents the links (bidirectional links are modeled by arcs in both directions). The straightforward approach to realize a distributed routing function in $G$ is to store for every arc $(v,w)\in A$ a set $D_{(v,w)}\subset V-\{v\}$ of destination addresses at vertex $v$; $v$ then forwards a package with destination address $u\in V-\{v\}$ via an arc $(v,w')\in A$ with $u\in D_{(v,w')}$. The drawback is that every vertex in the network has to store $|V|-1$ entries and therefore the overall storage space requirement is in $\Theta(|V|^2)$. Using interval routing the space requirement at individual vertices is compressed by introducing a global order on $V$ and assigning intervals in this order to the arcs, rather than subsets of $V$.[^5] Definition \[intervalle\] is needed for Definition \[defiIRScommon\], which comprehensively defines interval routing. Definition \[GinIRS\] classifies graphs, with respect to supported interval routing schemes. \[intervalle\] For $k\in\mathbb{N}$ and $a,b\leq k$ we define an *interval* as $$[a,b]= \begin{cases} \{a,a+1,\ldots,b\}&\text{ if }a\leq b,\\ [a,k]\cup [1,b]&\text{ if }a>b\text{.} \end{cases}$$ The interval $[a,b]$ is *linear* if $a\leq b$ and *cyclic* if $a>b$. For the sake of completeness we now give a definition of interval routing schemes that is more or less standard in literature. For the main result of this paper the more specialized version in Definition \[defiIRS\] is exploited. \[defiIRScommon\] Let $G=(V,A)$ be a directed graph. An *interval routing scheme for $G$* is a pair $R=({\mathcal{L}},{\mathcal{I}})$, where ${\mathcal{L}}\colon V\rightarrow \{1,2,\ldots,|V|\}$ is a bijective mapping and ${\mathcal{I}}$ maps every arc to a set of intervals $[a,b]$, $a,b\in\{1,2,\ldots,|V|\}$, such that, for every vertex $v\in V$, 1. ${\mathcal{I}}$ maps the outgoing arcs of $v$ to a set of pairwise disjoint intervals, 2. for every vertex $u\neq v$, one of these these intervals contains ${\mathcal{L}}(u)$ (one of these intervals also may contain ${\mathcal{L}}(v)$), and 3. \[sp\] there exists a directed path $(v=w_1,w_2,\ldots,w_n=u)$ such that ${\mathcal{L}}(u)$ is contained in an interval in ${\mathcal{I}}((w_{i},w_{i+1}))$, for $1\leq i<n$. We call $R$ a *linear interval routing scheme* (for short *LIRS*), if ${\mathcal{I}}$ only maps to linear intervals. If for every vertex $v\in V$ the outgoing arcs of $v$ are solely mapped to intervals, that do not contain ${\mathcal{L}}(v)$, $R$ is called a *strict interval routing scheme* (for short *SIRS*). If the paths implied by Constraint \[sp\] are all shortest paths, we denote $R$ a *shortest path IRS* and, if ${\mathcal{I}}$ maps every arc to at most $k$ intervals, a *$k$-IRS*. These notations may be combined. \[GinIRS\] The *compactness of a directed graph* $G$ is the minimal $k$ such that $G$ provides a shortest path $k$-IRS. The *compactness of an undirected graph* is the compactness of its directed symmetric version and the *compactness of a graph class* is the smallest $k$ such that every graph in this class has at most compactness $k$, if such $k$ exists. We define *$k$-IRC* as the class of graphs with compactness at most $k$. Strict, linear, and strict linear compactness are defined analogous. Since every cyclic interval can be replaced by two linear intervals and every non strict interval by two strict intervals, we obtain the following corollary, that we refer to as the *interval routing hierarchy*. For every $k\in\mathbb{N}^{>0}$, the following inclusions hold. \[irh\] $k$[-SLIRC]{.nodecor} [c]{}\ \ \ \ [c]{} $k$[-SIRC]{.nodecor}\ \ \ \ \ \ \ $k$[-LIRC]{.nodecor}\ [c]{}\ \ \ \ $k$[-IRC]{.nodecor} [c]{}\ \ \ \ [c]{} $(k+1)$[-SIRC]{.nodecor}\ \ \ \ \ \ \ $(k+1)$[-LIRC]{.nodecor}\ In general, it is NP-complete to decide if a graph has compactness $k$, for all variants of compactness, even if $k$ is fixed [@Eilam2002; @Flammini1997]. However, the compactness of many graph classes is known [@Gavoille00asurvey], as also reflected by the next theorem. (See [@Gavoille00asurvey] for references to proofs.) \[theo:bisher\] Paths, complete graphs, and $d$-dimensional grids have strict linear compactness 1 and trees, rings, outerplanar graphs, interval graphs, and unit circular-arc graphs have strict compactness 1. \[hithertoresults\] Since every wheel graph with six or more outer vertices has compactness 2 the class of circular-arc graphs is not in $1$-IRC. The compactness of the class of circular-arc graphs has not been determined until now, which is somehow surprising, since the class of circular-arc graphs is closely related to the class of interval graphs and unit circular-arc graphs and both classes are known to be in $1$-SIRC [@interval; @FG98]. In this paper we show that the class of circular-arc graphs is in 2-SIRC. In order to ease the proof, Definition \[defiIRS\] formalizes shortest path $k$-SIRSs alternatively and differs from Definition \[defiIRScommon\] in the following two aspects. Definition \[defiIRScommon\] represents the vertex order by mapping every vertex to a distinct number, which implies a heavy use of case-by-case analysis, in order to handle cyclic and linear intervals correctly; Definition \[defiIRS\] avoids this arduousness by exploiting cyclic orders. The third constraint in Definition \[defiIRS\] also differs from its equivalent in Definition \[defiIRScommon\], as Definition \[fe\] is exploited, to yield a constraint that can be verified easier. An equal constraint, that is more similar to the third constraint in Definition \[defiIRScommon\], would be to require a directed *shortest* path $(v=w_1,w_2,\ldots,w_n=u)$ such that $u$ is contained in an ring-interval in ${\mathcal{I}}((w_{i},w_{i+1}))$, for $1\leq i<n$. \[fe\] Let $G=(V,A)$ be directed graph and $u,v,w\in V$. Vertex $v\in V$ is a *first vertex from $u$ to $w$*, if there exists a shortest directed path $(u,v,\ldots, w)$, i.e., there is a shortest directed path from $u$ to $w$, that first traverses $v$. The set of first vertices from $u$ to $w$ is denoted by $${{\mathcal{F}}(u,w)}=\{v\in V~|~v\text{ is a first vertex from }u\text{ to }w\}.$$ \[defiIRS\] Let $G=(V,E)$ be directed graph. A *shortest path strict interval routing scheme for $G$* is a pair $R=({\mathscr{L}},{\mathcal{I}})$, where ${\mathscr{L}}$ is a cyclic order on $V$, denoted the *vertex order*, and ${\mathcal{I}}$, denoted the *arc-labeling*, maps every arc to a set of ring-intervals in ${\mathscr{L}}$ such that for every vertex $v\in V$, 1. ${\mathcal{I}}$ maps the outgoing arcs of $v$ to a set of pairwise disjoint ring-intervals in ${\mathscr{L}}$, 2. for every vertex $u\neq v$ one of these these ring-intervals contains $u$ (vertex $v$ must not appear in one of these intervals), and 3. if $u$ is contained in a ring-interval in ${\mathcal{I}}((v,w))$, then $w$ is a first vertex from $v$ to $u$. Let ${\mathscr{F}}$ be a vertex order. We say *$v\in V$ given ${\mathscr{F}}$ suffices a shortest path $k$-SIRS*, if there exists an arc-labeling for $G$ that maps every outgoing arc of $v$ to at most $k$ ring-intervals in ${\mathscr{F}}$, such that $v$ satisfies the three constraints above. The next corollary is evident and important for the argumentation in Chapter \[main\]. \[tropicthunder\] If for a given directed graph $G=(V,A)$ exists a vertex order ${\mathscr{F}}$ such that every vertex $v\in V$ suffices a shortest path $k$-SIRS given ${\mathscr{F}}$, then $G$ supports a shortest path $k$-SIRS. Main result\[main\] =================== In this chapter we prove the main result of this paper. \[mainTheo\] The class of circular-arc graphs is in 2-SIRC. Since every non-real circular-arc graph $G$ is an interval graph and therefore in 1-SIRC [@interval], we only need to consider real circular-arc graphs, when we prove the theorem. For the rest of this chapter let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for an arbitrary real circular-arc graph $G_u=(V,E)$ and $G=(V,A)$ be the directed symmetric version of $G_u$. To show that there exists a shortest path $2$-SIRS for $G$ and thus Theorem \[mainTheo\] holds, we proceed as follows. In Section \[constrCO\] we construct a cyclic order ${\mathscr{L}}$ on $V$. In Section \[furtherDefs\] we choose an arbitrary vertex $v\in V$ and partition the vertices in $V-\{v\}$ in three disjoint ring-intervals $A_v,B_v,C_v$ in ${\mathscr{L}}$ in dependency on $v$. In order to show that $v$ suffices a shortest path 2-SIRS given ${\mathscr{L}}$, we show how to define an arc-labeling ${\mathcal{I}}_v$ for the outgoing arcs of $v$ that satisfies the corresponding constraints given in Definition \[defiIRS\]. We consider each of the three ring-intervals $A_v,B_v,C_v$ in Section \[lvbisv\] and \[mvbislv\] and show how the vertices in the respective ring-interval can be sufficiently mapped to by ${\mathcal{I}}_v$. Since we have chosen $v$ arbitrarily, it follows that every vertex in $V$ suffices a shortest path $2$-SIRS given ${\mathscr{L}}$ and therefore, by Corollary \[tropicthunder\], $G$ supports a shortest path $2$-SIRS. Since $G_u$ is an arbitrary real circular-arc graph, this proves Theorem \[mainTheo\]. The following notations are straight forward but may be formalized for convenience. Let $(v,w)\in E$ be an arc with ${\mathcal{I}}_v((v,w))= S$, where $S$ is a set of ring-intervals in ${\mathscr{L}}$, and $R$ be a ring-interval in ${\mathscr{L}}$. *Assigning $R$ to arc $(v,w)$* means to define ${\mathcal{I}}_v\colon(v,w)\mapsto S\cup \{R\}$. Of course, when we start constructing ${\mathcal{I}}_v$, every arc is mapped to the empty set. Let $[a,b]_{\mathscr{L}}$ and $[c,d]_{\mathscr{L}}$ be two ring-intervals that can be joined to one ring-interval $[a,d]_{\mathscr{L}}$ by Corollary \[ringkombi\]. If $[a,b]_{\mathscr{L}}$ and $[c,d]_{\mathscr{L}}$ are assigned to the same arc $(v,w)$, we can redefine $${\mathcal{I}}_v\colon(v,w)\mapsto\{[a,b]_{\mathscr{L}},[c,d]_{\mathscr{L}}\} \text{ by } {\mathcal{I}}_v\colon(v,w)\mapsto\{[a,d]_{\mathscr{L}}\}$$ and therefore save one ring-interval on arc $(v,w)$. This redefinition is called *compressing the (two) ring-intervals on arc $(v,w)$ (to one ring-interval)*. Let $V'\subseteq V$ be a set of vertices, e.g. a ring-interval in ${\mathscr{L}}$. *To distribute $V'$ (over the outgoing arcs of $v$)* means to partition $V'$ into ring-intervals and assign these to outgoing arcs of $v$ (sufficiently). For convenience, we sometimes refer to ring-intervals as *intervals*. Definition of the Vertex Order {#constrCO} ------------------------------ In this section we construct a cyclic order ${\mathscr{L}}$ on $V$ that servers as the given vertex order. A pseudocode for this purpose is presented in Listing \[pseudoLabeling\] and explained next. By Corollary \[remarkDoms\] the left and right clique of a dominating vertex are not unique. Since it relieves the proof, if all dominating vertices have the same left and thus also the same right clique, these cliques are unified in Line 1. The dummy vertex introduced in Line 4 is needed to close the cyclic order ones all vertices are integrated in ${\mathscr{L}}$. The loop in Line 6 runs ones through all cliques in ${\mathsf{X}}$ (beginning at ${\mathsf{f}}$, that is arbitrarily chosen in Line 2) in the order defined by ${\mathscr{C}}$ (Line 13). For every visited clique ${\mathsf{n}}$ the loop in Line 8 integrates every vertex, whose left clique is ${\mathsf{n}}$, in ${\mathscr{L}}$. By Line 9, vertices with the same left clique are ordered with respect to their right clique. Fix $\mc{z}\in \mc{X}$ arbitrarily and choose $\lc{d}\colonequals \ri{C}(\mc{z})$ and $\rc{d}\colonequals \mc{z}$ for all $d\in \dom$. Choose $\mc{f}\in \mc{X}$ arbitrarily $\mc{n}\colonequals \mc{f}$ $V\colonequals V\cup \{t\}$ /* add dummy vertex $t$ */ $p\colonequals t$ while(true) $A\colonequals \{v\in V~|~\lc{v}=\mc{n}\}$ while($A\neq\emptyset$) Choose $a\in A$ such that every vertex in $A$ reaches at least as far to the right as $a$. $\ri{L}(p)\colonequals a$ $p\colonequals a$ $A\colonequals A-\{a\}$ $\mc{n}\colonequals \ri{C}(\mc{n})$ if($\mc{n}=\mc{f}$) break $\ri{L}(p)\colonequals\ri{L}(t)$ $V=V-\{t\}$ /* remove dummy vertex $t$ */ return $\ri{L}$ Note that for a fixed ${\mathsf{z}}$ in Line 1, the generated cyclic order is fully deterministic, except for the ordering of true twins.[^6] Notably, this determinism holds irrespective of the choice of clique ${\mathsf{f}}$ in Line 2. The following observations are crucial for the rest of this paper. 1. \[item:leftconsecutiv\] The vertices are ordered primary by their left clique, that is to say, vertices having the same left clique appear consecutive in $ {\mathscr{L}}$. 2. Vertices with the same left clique are ordered [“ascending”]{} by their right clique. 3. \[item:onemoreleft\] For two vertices $v,w\in V$ with $v\succ_{\mathscr{L}} w$, we know that either 1. ${\alpha(v)}={\alpha(w)}$ and $w$ reaches at least as far to the right as $v$, or 2. ${\alpha(v)}\succ_{\mathscr{C}}{\alpha(w)}$, in other words, $v$ reaches one clique further to the left than $w$. The next definition is based on Assertion \[item:leftconsecutiv\] of the preceding remark, that is to say, on the fact that, since vertices having the same left clique appear consecutive in $ {\mathscr{L}}$, for a given clique ${\mathsf{c}}\in{\mathsf{X}}$, we can find two vertices $v,w\in {\mathsf{c}}$ such that the vertices in ring-interval $[v,w]_{\mathscr{L}}$ are exactly those with left clique ${\mathsf{c}}$. Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for some circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], ${\mathsf{c}}\in{\mathsf{X}}$, and $v,w\in {\mathsf{c}}$ the unique vertices such that $[v,w]_{\mathscr{L}}={\mathsf{c}}$ ($v$ and $w$ are not necessarily distinct). We call $v$ the *head vertex of ${\mathsf{c}}$* and $w$ the *tail vertex of ${\mathsf{c}}$* and denote them by ${{\mathcal{H}}({\mathsf{c}})}$ and ${{\mathcal{T}}({\mathsf{c}})}$, respectively. The next corollary is obvious and links the preceding definition to ${\mathscr{C}}$. \[coro:tail vertexheal\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], and ${\mathsf{c}}\in{\mathsf{X}}$. The successor (in ${\mathscr{L}}$) of the tail vertex of ${\mathsf{c}}$ is the head vertex of the successor (in ${\mathscr{C}}$) of ${\mathsf{c}}$, or formally: ${\mathscr{L}}({{\mathcal{T}}({\mathsf{c}})})={{\mathcal{H}}({\mathscr{C}}({\mathsf{c}}))}$. Partition of the Vertex Order {#furtherDefs} ----------------------------- The cyclic order ${\mathscr{L}}$ on $V$ allows to state further definitions. We now fix $v\in V$ arbitrarily for the rest of the proof and show that $v$ suffices a shortest path $2$-SIRS given ${\mathscr{L}}$. If $v$ is a dominating vertex, we can define ${\mathcal{I}}_v\colon(v,w)\mapsto[w,w]_{\mathscr{L}}$ for every vertex $w\in V-\{v\}$ and have thereby shown that $v$ suffices a shortest path $1$-SIRS. Therefore, we now assume that $v$ is not a dominating vertex. \[de\_l\_v\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for some circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], and $v\in V$ not a dominating vertex. Let $L\subset V$ be the set of vertices that are adjacent to $v$ and reach further to the left than $v$ but are neither counter vertices of $v$ nor dominating vertices and the set of vertices in $[{{\mathcal{H}}({\alpha(v)})},{\mathscr{L}}(v)]_{\mathscr{L}}$, if $v\neq{{\mathcal{H}}({\alpha(v)})}$. If $L\neq\emptyset$, we define the *left vertex of $v$*, denoted by ${{l_{v}}}$, as the unique vertex in $L$ such that $$\label{eq_l_v} \forall w \in L\colon [w,v]_{\mathscr{L}}\subseteq[{{l_{v}}},v]_{\mathscr{L}} \text{ (or, equivalently, } L\subseteq[{{l_{v}}},v]_{\mathscr{L}}\text{)}.$$ \[no\_vertex\_further\_than\_lv\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], $v\in V$ not a dominating vertex, ${{l_{v}}}$ the left vertex of $v$, and $u$ be a vertex that reaches further to the left than $v$ but is neither a dominating vertex nor a counter vertex of $v$. Then ${{l_{v}}}$ reaches at least as far to the left as $u$. Assume $u$ reaches further to the left than ${{l_{v}}}$. Then by Definition \[defi:reaches\] we know that $[{\alpha({{l_{v}}})},{\alpha(v)}]_{\mathscr{C}}\subset[{\alpha(u)},{\alpha(v)}]_{\mathscr{C}}$, which implies that $${{\mathcal{S}}}[{\alpha(u)},{\alpha(v)}]_{\mathscr{C}}=({\alpha(u)},\ldots,{\alpha({{l_{v}}})},\ldots,{\alpha(v)}).$$ Since ${\mathscr{L}}$ is primary ordered by the left cliques of the vertices, it follows $u\succ_{\mathscr{L}}\ldots\succ_{\mathscr{L}} {{l_{v}}}\succ_{\mathscr{L}}\ldots\succ_{\mathscr{L}}v$ and therefore $[{{l_{v}}},v]_{\mathscr{L}}\subseteq[u,v]_{\mathscr{L}}$. This contradicts (\[eq\_l\_v\]) in Definition \[de\_l\_v\], since $u$ reaches further to the left than $v$ and is neither a dominating vertex nor a counter vertex of $v$. In Definition \[leftrightface\] we partition the vertices in $V-\{v\}$ into three ring-intervals in ${\mathscr{L}}$. The following argumentation gives the idea behind this partitioning and also proves Theorem \[Lvrl\_v\]. Consider we go through ${\mathscr{L}}$ beginning at $v$. Since the vertices in ${\mathscr{L}}$ are primary ordered by their left cliques, the left clique of the vertices first considered is ${\alpha(v)}$ (or ${\mathscr{C}}({\alpha(v)})$, if $v$ is the tail vertex of clique ${\alpha(v)}$). The next vertices traversed have left clique ${\mathscr{C}}({\alpha(v)})$, followed by vertices with left clique ${\mathscr{C}}^2({\alpha(v)})$, and so on. We eventually reach the set of vertices with left clique ${\beta(v)}$. The last vertex we come across in this set is ${{\mathcal{T}}({\beta(v)})}$ and fixed in Definition \[de\_m\_v\]. The vertices we came across so far (excluding $v$, including ${{\mathcal{T}}({\beta(v)})}$) will be defined as the [“vertices to the right of $v$”]{} in Definition \[leftrightface\] and are adjacent to $v$, as $v$ was contained in their left clique. \[de\_m\_v\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], and $v\in V$ not a dominating vertex. We define the *middle vertex of $v$*, denoted by ${{m_{v}}}={{\mathcal{T}}({\beta(v)})}$, as the tail vertex of the right clique of $v$. Assume we continue visiting the vertices after ${{m_{v}}}$ in the same manner. The next vertex $u$ we come across that is not a dominating vertex and adjacent to $v$ has to be ${{l_{v}}}$, as we could find a contradiction in the same manner as in the proof of Theorem \[no\_vertex\_further\_than\_lv\], if we had $u\neq{{l_{v}}}$. These vertices after ${{m_{v}}}$ and before ${{l_{v}}}$ are defined as the [“vertices face-to-face with $v$”]{} in Definition \[leftrightface\]. When we continue to go through ${\mathscr{L}}$, we eventually reach $v$ again. These vertices from ${{l_{v}}}$ to $v$ (excluding $v$) are defined as the [“vertices to the left of $v$”]{} in Definition \[leftrightface\]. \[leftrightface\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], $v\in V$ not a dominating vertex, ${{l_{v}}}$ the left vertex of $v$, if $l_v$ exists, and ${{m_{v}}}$ the middle vertex of $v$. We define 1. the *vertices to the right of $v$* 1. as $A_v=[{\mathscr{L}}(v),{{m_{v}}}]_{\mathscr{L}}$, if $v\neq{{m_{v}}}$ or else 2. as $A_v=\emptyset$, and 2. the *vertices face-to-face with $v$* 1. as $B_v=\emptyset$, if ${\mathscr{L}}({{m_{v}}})={{l_{v}}}$, or else 2. as $B_v=[{\mathscr{L}}({{m_{v}}}),{\mathscr{L}}^{-1}(v)]_{\mathscr{L}}$, if ${{l_{v}}}$ does not exist, or else 3. as $B_v=[{\mathscr{L}}({{m_{v}}}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$. 3. the *vertices to the left of $v$* 1. as $C_v=[{{l_{v}}},{\mathscr{L}}^{-1}(v)]_{\mathscr{L}}$, if ${{l_{v}}}$ exists, or else 2. as $C_v=\emptyset$, The next theorem follows by the argumentation between Theorem \[no\_vertex\_further\_than\_lv\] and Definition \[leftrightface\]. \[Lvrl\_v\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], and $v\in V$ but not a dominating vertex. 1. \[Lvrl\_v1\] Every vertex to the right of $v$ is adjacent to $v$. 2. Every vertex face-to-face with $v$ that is not a dominating vertex is not adjacent to $v$. Definition of the Arc-Labeling {#2-app} ------------------------------ Next we show that $v$ given ${\mathscr{L}}$ suffices a shortest path $2$-SIRS by constructing a mapping ${\mathcal{I}}_v$ from the set of outgoing arcs of $v$ to at most two ring-intervals in ${\mathscr{L}}$, according to Definition \[defiIRS\]. We investigate the vertices to the left of $v$ and the vertices to the right of $v$ in Section \[lvbisv\] and the vertices face-to-face with $v$ in Section \[mvbislv\]. For every vertex $u$ in the respective interval, we determine a first vertex $w$ from $v$ to $u$ such that $u$ and every vertex that is hitherto assigned to arc $(v,w)$ can be precisely embraced by at most two ring-intervals in ${\mathscr{L}}$. This is a simple task for the vertices in $A_v$ and $C_v$ and even for the vertices in $B_v$ the approach is straight forward, if there is a dominating vertex or a counter vertex of $v$. In fact, it only gets tricky, if there are no dominating vertices and no counter vertices. ### Vertices to the Left and Right {#lvbisv} By the first assertion of Theorem \[Lvrl\_v\] every vertex to the right of $v$ is adjacent to $v$, wherefore these vertices can be distributed by assigning $[w,w]_{\mathscr{L}}$ to arc $(v,w)$, for every vertex $w\in A_v$. Assume that we have $C_v\neq\emptyset$. In general, not every vertex in $C_v$ is adjacent to $v$. Although ${{l_{v}}}$ is adjacent to all vertices in $C_v$, it is generally not possible to assign $C_v$ to arc $(v,{{l_{v}}})$, since $C_v$ may contain vertices which must be assigned to [“their own arc”]{}, since they are adjacent to $v$. Let $v_i\in C_v$ be a vertex that is adjacent to $v$ and assume we start at $v_i$ to go through ${\mathscr{L}}$ until we come across the next vertex that is adjacent to $v$ or $v$. Denote this vertex $v_j$. We can show that $v_i$ is adjacent to every in $[{\mathscr{L}}(v_i),v_j]_{\mathscr{L}}$, wherefore we can assign $[v_i,{\mathscr{L}}^{-1}(v_j)]_{\mathscr{L}}$ to arc $(v,v_i)$. Therefore the vertices to the left of $v$ can be distributed by assigning every vertex $w\in C_v$ to arc $(v,w)$, if $w$ is adjacent to $v$, and else to arc $(v,w')$, where $w'$ is the first vertex that precedes $w$ in ${\mathscr{L}}$ and is adjacent to $v$. Theorem \[theototheleft\] proves the outlined idea formally. When distributing the vertices in sets $A_v$ and $C_v$ as just outlined, every outgoing arc of $v$ (that is not incident to a dominating vertex face-to-face with $v$) gets one ring-interval assigned. \[theototheleft\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], $v\in V$ but not a dominating vertex, and ${{l_{v}}}$ the left vertex of $v$ (then the vertices in $[{{l_{v}}},{\mathscr{L}}^{-1}(v)]$ are the vertices to the left of $v$). Let $({{l_{v}}}=v_1,v_2,\ldots,v_n=v)$ be the sequence of vertices that we obtain, when we order the vertices in $[{{l_{v}}},{\mathscr{L}}^{-1}(v)]$ that are adjacent to $v$ as they appear in ring-sequence ${{\mathcal{S}}}[{{l_{v}}},v]_{\mathscr{L}}$ and append $v$. For $1\leq i<n$ and $x\in[v_i,{\mathscr{L}}^{-1}(v_{i+1})]_{\mathscr{L}}$, vertex $v_i$ is a first vertex from $v$ to $x$. For some $i$, $1\leq i<n$, let $x\in[v_i,{\mathscr{L}}^{-1}(v_{i+1})]_{\mathscr{L}}$. Clearly the theorem holds if $x=v_i$. If $x\neq v_i$, $x$ does not appear in $(v_1,v_2,\ldots,v_n)$ and therefore is not adjacent to $v$. Since the vertices in ${\mathscr{L}}$ are primary ordered by their left clique and $v_i$ appears before $x$ in ${{\mathcal{S}}}[{{l_{v}}},v]_{\mathscr{L}}$, it follows that the left clique of $x$ is in $[{\alpha(v_i)},{\alpha(v)}]_{\mathscr{C}}$. Since $v_i$ is adjacent to $v$, $v_i$ is contained in every clique in $[{\alpha(v_i)},{\alpha(v)}]_{\mathscr{C}}$. This implies that $v_i$ is contained in the left clique of $x$ and thus $v_i$ and $x$ are adjacent. Therefore $(v,v_i,x)$ is a shortest path between $v$ and $x$ and $v_i$ is a first vertex from $v$ to $x$. ### Vertices Face-to-Face {#mvbislv} It is probably the common case, that neither counter nor dominating vertices exist (only then path lengths are unbounded). Therefore, in the next paragraph we first discuss the distribution of vertices face-to-face with $v$ under these assumptions. The subsequent paragraph will show that proving the sufficiency of $v$ to a shortest path 2-SIRS given ${\mathscr{L}}$ is much easier, if we assume that $v$ has counter vertices or the graph contains dominating vertices. The third paragraph is dedicated to the case in that $v$ has no counter vertices (and the graph does not contain dominating vertices) but there exists at least one pair of counter vertices. As one can see, every possible constellation is covered by one of these three cases. #### Neither Dominating nor Counter Vertices exist {#beideLeer} Let $w\in V$ be a vertex that is not adjacent to $v$, $L$ be the set of vertices adjacent to $v$ that reach farthest to the left and $R$ be the set of vertices adjacent to $v$ that reach farthest to the right ($L$ and $R$ might intersect). The arc model of $G$ gives reason to believe that every vertex in $L$ or every vertex in $R$ is a first vertex from $v$ to $w$. A formal argument for this assertion can be deduced from the proof of Theorem \[b\]. Since we have ${{l_{v}}}\in L$, we now fix a vertex adjacent to $v$ that reaches farthest to the right. \[de\_r\_v\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a real circular-arc graph $G=(V,E)$ without dominating vertices and without counter vertices, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], $v\in V$, ${{l_{v}}}$ the left vertex of $v$, ${{m_{v}}}$ the middle vertex of $v$, and $R\subset V$ be the set of vertices that are adjacent to $v$ and reach farthest to the right. We define the *right vertex of $v$*, denoted by ${{r_{v}}}$, as ${{l_{v}}}$, if ${{l_{v}}}\in R$, or as ${{m_{v}}}$, if ${{m_{v}}}\in R$, or else as an arbitrary vertex in $R$. In order to show that $v$ suffices a shortest path 2-SIRS given ${\mathscr{L}}$, we could choose ${{r_{v}}}$ arbitrarily in $R$. However, we are able to show that $v$ given ${\mathscr{L}}$ suffices a shortest path $1$-SIRS, if ${{l_{v}}}\in R$ or ${{m_{v}}}\in R$. Since we assume that $G$ is a real circular-arc graph without counter and dominating vertices, every vertex has a left and right vertex and therefore $L$ and $R$ in the definition above are not empty. The next definition redefines the notation of the left and right vertex as a function, in order to allow recursive usage to easily determine a first vertex from $v$ to every vertex. \[lvi\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a real circular-arc graph $G=(V,E)$ without dominating vertices and without counter vertices, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], and $v\in V$. We define ${{l_{}}^{}(v)}$ as the left vertex of $v$ and ${{r_{}}^{}(v)}$ as the right vertex of $v$.[^7] For $u\in V$, we define ${{\mathcal{L}}_v(u)}$ as the smallest $i$ such that $u$ is adjacent to ${{l_{}}^{i}(v)}$ and ${{\mathcal{R}}_v(u)}$ as the smallest $i$ such that $u$ is adjacent to ${{r_{}}^{i}(v)}$. The following theorem is somehow clear when illustrated but since it is very basic for the rest of this subsection, it may be proven in detail. \[b\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a real circular-arc graph $G=(V,E)$ without dominating vertices and without counter vertices, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], and $v\in V$. The following three assertions hold for every vertex $u\in V$ that is not adjacent to $v$. 1. \[jans\] If ${{\mathcal{L}}_v(u)}<{{\mathcal{R}}_v(u)}$, then ${{l_{v}}}$ is a first vertex from $v$ to $u$. 2. \[jans2\] If ${{\mathcal{L}}_v(u)}>{{\mathcal{R}}_v(u)}$, then ${{r_{v}}}$ is a first vertex from $v$ to $u$. 3. \[jans3\] If ${{\mathcal{L}}_v(u)}={{\mathcal{R}}_v(u)}$, then ${{l_{v}}}$ and ${{r_{v}}}$ are first vertices from $v$ to $u$. Let $p'=(v,w_1,w_2,\ldots,w_j,u)$, with $j\geq2$, be a shortest path in a circular-arc graph. For $1\leq i< j$, no vertex $w_{i+1}$ can reach further to the left *and* right than $w_{i}$, because then $w_{i+1}$ would be adjacent to every vertex that $w_i$ is adjacent to and $w_i$ therefore excrescent. Furthermore, if $w_2$ reaches further to the left than $w_1$, then $w_{i+1}$ reaches further to the left then $w_i$, and if $w_2$ reaches further to the right than $w_1$, then $w_{i+1}$ reaches further to the right than $w_i$, for $1\leq i< j$. Therefore, if $w_2\in{\alpha(w_1)}$, then $w_{i+1}\in{\alpha(w_i)}$ and, if $w_2\in{\beta(w_1)}$, then $w_{i+1}\in{\beta(w_i)}$, for $1\leq i< j$. We only prove Proposition \[jans\]. Suppose we have ${{\mathcal{L}}_v(u)}=k<{{\mathcal{R}}_v(u)}$. Then there is a path $p=(v,{{l_{}}^{1}(u)},{{l_{}}^{2}(u)},\ldots,{{l_{}}^{k}(u)},u)$ (of length $k+1$). Assume there is a shortest path $p'=(v,w_1,w_2,\ldots,w_j,u)$ with $1<j<k$.[^8] As argued above, we have \[fewo\] $w_{i+1}\in{\alpha(w_i)}$ or \[fewo2\] $w_{i+1}\in{\beta(w_i)}$, for $1\leq i< j$. Assume [(\[fewo\])]{} holds. Since there are neither dominating nor counter vertices in $G$, ${{l_{v}}}$ reaches at least as far to the left as $w_1$ and thus ${{l_{v}}}$ is also contained in ${\alpha(w_1)}$. Therefore $w_2$ and ${{l_{v}}}$ are adjacent and we can replace $w_1$ in $p'$ by ${{l_{v}}}$. Since $w_2$ as well as ${{l_{}}^{2}(v)}$ are adjacent to ${{l_{v}}}$, ${{l_{}}^{2}(v)}$ reaches at least as far to the left as $w_2$. This implies ${{l_{}}^{2}(v)}\in{\alpha(w_2)}$ and since the left clique of $w_2$ contains $w_3$, $w_2$ can be replaced by ${{l_{}}^{2}(v)}$ in $p'$. By sequentially applying this argument to $w_3,w_4,\ldots,w_{j-1}$, we obtain $(v,{{l_{}}^{1}(u)},{{l_{}}^{2}(u)},\ldots,{{l_{}}^{j-1}(u)},w_j,u)$ as a shortest path. Since also ${\alpha(w_j)}$ contains ${{l_{}}^{j}(v)}$, every clique in $[{\alpha(w_j)},{\alpha({{l_{}}^{j-1}(v)})}]_{\mathscr{C}}$ contains ${{l_{}}^{j}(v)}$. Because at least one of these cliques contains $u$, $u$ is adjacent to ${{l_{}}^{j}(v)}$. This contradicts Definition \[lvi\], since we have $j<k={{\mathcal{L}}_v(u)}$. If [(\[fewo2\])]{} holds, we obtain $(v,{{r_{}}^{1}(u)},{{r_{}}^{2}(u)},\ldots,{{r_{}}^{j}(u)},u)$ as a shortest path, which is a somehow [“stronger”]{} contradiction, since we have $j<k<{{\mathcal{R}}_v(u)}$. We conclude that if ${{\mathcal{L}}_v(u)}<{{\mathcal{R}}_v(u)}$ holds, ${{l_{v}}}$ is a first vertex from $v$ to $u$. Assertion \[jans\] is an implication (and not an equivalence), since ${{r_{v}}}$ might be adjacent to ${{l_{}}^{2}(v)}$, which implies that ${{r_{v}}}$ might be a first vertex from $v$ to $u$, even though we have ${{\mathcal{L}}_v(u)}<{{\mathcal{R}}_v(u)}$. Proposition \[jans2\] and \[jans3\] can be shown analogue. Because no vertex in $B_v$ is adjacent to $v$, the theorem implies that ${{l_{v}}}$ or ${{r_{v}}}$ is a first vertex from $v$ to every vertex in $B_v$. Since sequence ${{r_{}}^{1}(v)},{{r_{}}^{2}(v)},\ldots$ [“runs”]{} trough the clique-cycle as implied by ${\mathscr{C}}$ and sequence ${{r_{}}^{1}(v)},{{r_{}}^{2}(v)},\ldots$ [“runs”]{} through the clique-cycle as implied by ${\mathscr{C}}^{-1}$, that is to say, in the other direction, there exists an $i\in\mathbb{N}$ such that ${{l_{}}^{i}(v)}$ and ${{r_{}}^{i}(v)}$ [“meet”]{}, or formally, are adjacent. This number is fixed in Definition \[apex\]. Theorem \[a\] is needed for the proof of Theorem \[separator\], which shows that there exists a vertex ${s_v}\in B_v$ with ${{l_{}}^{i}(v)},{{r_{}}^{i}(v)}\in{\alpha({s_v})}$ such that ${{r_{v}}}$ is a first vertex from $v$ to every vertex in $B_r=[{\mathscr{L}}({{m_{v}}}),{s_v}]$ and vertex ${{l_{v}}}$ is a first vertex from $v$ to every vertex in $B_l=[{\mathscr{L}}({s_v}),{\mathscr{L}}^{-1}({{l_{v}}})]$. Since Theorem \[separator\] implies that the vertices in $B_v$ can be distributed by assigning $B_l$ to arc $(v,{{l_{v}}})$ and $B_r$ to arc $(v,{{r_{v}}})$, Theorem \[mainTheo\] holds for the case in that neither dominating vertices nor counter vertices exist. \[apex\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a real circular-arc graph $G=(V,E)$ without dominating vertices and without counter vertices, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], and $v\in V$. If we have ${\alpha({{l_{}}^{1}(v)})}={\beta({{r_{}}^{1}(v)})}$ or $[{\alpha(v)},{\beta(v)}]_{\mathscr{C}}\subset[{\beta({{r_{}}^{1}(v)})},{\alpha({{l_{}}^{1}(v)})}]_{\mathscr{C}}$ we set $i=1$, otherwise we define $i$ as the smallest number greater than 1 such that ${{l_{}}^{i}(v)}$ and ${{r_{}}^{i}(v)}$ are adjacent or equal. We denote $i$ the *apex number of $v$*. Let $i>2$ be the apex number of $v$. Vertices ${{r_{}}^{i}(v)}$ and ${{l_{}}^{i-1}(v)}$ are adjacent if and only if ${{l_{}}^{i}(v)}$ and ${{r_{}}^{i-1}(v)}$ are adjacent. Furthermore, for $2\leq j < i$ vertices ${{r_{}}^{j}(v)}$ and ${{l_{}}^{j}(v)}$ are not adjacent. \[a\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a real circular-arc graph $G=(V,E)$ without dominating vertices and without counter vertices, $v\in V$, and $i$ the apex number of $v$. No vertex $w$ with ${\alpha(w)}\in[{\alpha({\mathscr{L}}({{m_{v}}}))},{\mathscr{C}}^{-1}({\alpha({{l_{}}^{i-1}(v)})})]_{\mathscr{C}}$ is adjacent to ${{l_{}}^{i-2}(v)}$. Suppose there is a vertex $w$ with ${\alpha(w)}\in[{\alpha({\mathscr{L}}({{m_{v}}}))},{\mathscr{C}}^{-1}({\alpha({{l_{}}^{i-1}(v)})}]_{\mathscr{C}}$ adjacent to ${{l_{}}^{i-2}(v)}$. Then we have $w\in{\mathscr{C}}^{-1}({\alpha({{l_{}}^{i-1}(v)})})$ and ${{l_{}}^{i-1}(v)}\notin{\mathscr{C}}^{-1}({\alpha({{l_{}}^{i-1}(v)})})$. This implies that $w$ reaches further to left than ${{l_{}}^{i-1}(v)}$. Since ${{l_{}}^{i-1}(v)}$ is the left vertex of ${{l_{}}^{i-2}(v)}$ this contradicts Theorem \[no\_vertex\_further\_than\_lv\]. \[separator\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a real circular-arc graph $G=(V,E)$ without dominating vertices and without counter vertices, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], $v\in V$, $i$ the apex number of $v$, ${{l_{v}}}$ the left vertex of $v$, ${{m_{v}}}$ the middle vertex of $v$, and ${{r_{v}}}$ the right vertex of $v$. There exists a vertex ${s_v}\in[{\mathscr{L}}({{m_{v}}}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$ such that, for every vertex $w\in[{\mathscr{L}}({{m_{v}}}),{s_v}]_{\mathscr{L}}$, we have ${{r_{v}}}\in{{\mathcal{F}}(v,w)}$ and, if ${\mathscr{L}}({s_v})\neq{{l_{v}}}$, for every vertex $w'\in[{\mathscr{L}}({s_v}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$, we have ${{l_{v}}}\in{{\mathcal{F}}(v,w')}$. If $i=1$ the theorem holds for ${s_v}={\mathscr{L}}^{-1}({{l_{v}}})$, since every in $[{\mathscr{L}}({{m_{v}}}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$ is adjacent to ${{r_{v}}}$. The following argumentation is partially illustrated in Figure \[neu\]. ![ This picture illustrates the proof of Theorem \[separator\] and partially shows a circular-arc graph, i.e., the lines represent arcs. A line ends with a triangle, if it may reach not further to left or right, respectively. The upper two lines with a dotted end illustrate that an arc, that has its right end in the interval depicted by the rotated curly bracket, may reach no further to the left than the triangle of the respective line. We begin to search for $u$ at the right end of the lower curly bracket. The rotated square brackets on top mark the areas of the left cliques of the vertices in sets $A,B,C$. The case in that ${{l_{}}^{i-1}(v)}= u$ looks similar. []{data-label="neu"}](neu.pdf){width="60.00000%"} Let $i>1$, $$A=[{\mathscr{L}}({{m_{v}}}),{{\mathcal{T}}({\beta(r^{i-1}(v))})}]_{\mathscr{L}},$$ and $a\in A$. We have ${{\mathcal{R}}_v(a)}\leq i-1$, because $a$ is adjacent to a vertex in $\{{{r_{}}^{1}(v)},{{r_{}}^{2}(v)},\ldots,{{r_{}}^{i-1}(v)}\}$. Since the vertices in ${\mathscr{L}}$ are primary ordered by their left cliques, we have $${\alpha(a)}\in[{\alpha({\mathscr{L}}({{m_{v}}}))},{\beta(r^{i-1}(v))}]_{\mathscr{C}}\subseteq[{\alpha({\mathscr{L}}({{m_{v}}}))},{\mathscr{C}}^{-1}({\alpha({{l_{}}^{i-1}(v)})}]_{\mathscr{C}}$$ and therefore Theorem \[a\] implies ${{\mathcal{L}}_v(a)}\geq i-1$. Since we have ${{\mathcal{R}}_v(a)}\leq i-1\leq {{\mathcal{L}}_v(a)}$, Theorem \[b\] implies $$\label{eqd} \forall a\in A\colon{{r_{v}}}\in {{\mathcal{F}}(v,a)}.$$ Now we go through ${\mathscr{L}}$ beginning at ${\mathscr{L}}({{\mathcal{T}}({\beta({{r_{}}^{i-1}(v)})})}={{\mathcal{H}}({\mathscr{C}}({\beta({{r_{}}^{i-1}(v)})}))}$,[^9] in other words, we begin at the first vertex after the [“last”]{} vertex in $A$, until we eventually come across a vertex $u$ that either is ${{l_{}}^{i-1}(v)}$ or adjacent to ${{l_{}}^{i-1}(v)}$.[^10] We define $$B= \begin{cases} [{\mathscr{L}}({{\mathcal{T}}({\beta({{r_{}}^{i-1}(v)})})}),{\mathscr{L}}^{-1}(u)]_{\mathscr{L}} &\text{if } {\mathscr{L}}({{\mathcal{T}}({\beta({{r_{}}^{i-1}(v)})})}\neq u\\ \emptyset &\text{else,} \end{cases}$$ i.e. $B$ is the set of vertices after ${{\mathcal{T}}({\beta({{r_{}}^{i-1}(v)})})}$ and before $u$ in ${\mathscr{L}}$. Note that, by the choice of $u$, no vertex in $B$ is adjacent to ${{l_{}}^{i-1}(v)}$. Assume there is a vertex $x\in B$ that is not adjacent to ${{r_{}}^{i}(v)}$. When we go through ${\mathscr{C}}$ beginning at ${\alpha({{r_{}}^{i-1}(v)})}$ we eventually come across ${\beta({{r_{}}^{i-1}(v)})}$ (since ${{r_{}}^{i}(v)}$ and ${{r_{}}^{i-1}(v)}$ are adjacent, we have ${\alpha({{r_{}}^{i}(v)})}\in[{\alpha({{r_{}}^{i-1}(v)})},{\beta({{r_{}}^{i-1}(v)})}]_{\mathscr{C}}$). Before we come across ${\alpha(x)}$, we come across ${\beta({{r_{}}^{i}(v)})}$, since otherwise $x$ and ${{r_{}}^{i}(v)}$ would be adjacent. We do not come across ${\alpha({{l_{}}^{i-1}(v)})}$ until we came across ${\beta(x)}$, since otherwise $x$ and ${{l_{}}^{i-1}(v)}$ would be adjacent. It follows that ${{l_{}}^{i-1}(v)}$ and ${{r_{}}^{i}(v)}$ are not adjacent. Because ${{l_{}}^{i}(v)}$ and ${{r_{}}^{i}(v)}$ are adjacent, by the choice of $i$, ${\alpha({{l_{}}^{i}(v)})}$ must appear in $[{\alpha({{r_{}}^{i}(v)})},{\beta({{r_{}}^{i}(v)})}]_{\mathscr{C}}$. It cannot appear in $[{\alpha({{r_{}}^{i}(v)})},{\beta({{r_{}}^{i-1}(v)})}]_{\mathscr{C}}$, because then ${{l_{}}^{i}(v)}$ is adjacent to ${{r_{}}^{i-1}(v)}$, which contradicts Definition \[de\_r\_v\], since ${{l_{}}^{i}(v)}$ reaches further to the right than ${{r_{}}^{i}(v)}$ and ${{r_{}}^{i}(v)}$ is the right vertex of ${{r_{}}^{i-1}(v)}$. Therefore ${\alpha({{l_{}}^{i}(v)})}$ appears in $[{\mathscr{C}}({\beta({{r_{}}^{i-1}(v)})}),{\beta({{r_{}}^{i}(v)})}]_{\mathscr{C}}$ and we have $$\label{wtf} {\alpha({{l_{}}^{i}(v)})}\in[{\mathscr{C}}({\beta({{r_{}}^{i-1}(v)})}),{\mathscr{C}}^{-1}({\alpha(x)})]_{\mathscr{C}}.$$ Since the vertices in ${\mathscr{L}}$ are primary ordered by their left clique, Inclusion (\[wtf\]) implies $$\label{wtfwtf} {{l_{}}^{i}(v)}\in[{{\mathcal{H}}({\mathscr{C}}({\beta({{r_{}}^{i-1}(v)})}))},x]_{\mathscr{L}}.$$ Remember that we began at ${{\mathcal{H}}({\mathscr{C}}({\beta({{r_{}}^{i-1}(v)})}))}$ to go through ${\mathscr{L}}$ until we found a vertex $u$ that is adjacent to ${{l_{}}^{i-1}(v)}$. Inclusion [(\[wtfwtf\])]{} implies that we must have come across ${{l_{}}^{i}(v)}$ before $x$. But then we have $u={{l_{}}^{i}(v)}$. Since $$x\notin[{{\mathcal{H}}({\mathscr{C}}({\beta({{r_{}}^{i-1}(v)})}))},{\mathscr{L}}^{-1}({{l_{}}^{i}(v)})]=[{\mathscr{L}}({{\mathcal{T}}({\beta({{r_{}}^{i-1}(v)})})}),{\mathscr{L}}^{-1}(u)]_{\mathscr{L}}=B$$ contradicts the initial assumption, every vertex in $B$ is adjacent to ${{r_{}}^{i}(v)}$. Since by the choice of $u$ no vertex in $B$ is adjacent to ${{l_{}}^{i-1}(v)}$, we have $$\forall b\in B\colon{{\mathcal{L}}_v(b)}\geq i={{\mathcal{R}}_v(b)},$$ which, by Theorem \[b\], implies $$\label{eqc} \forall b\in B\colon{{r_{v}}}\in {{\mathcal{F}}(v,b)}.$$ We define ${s_v}={\mathscr{L}}^{-1}(u)$. Since the ring-intervals $A$ and $B$ appear consecutive in ${\mathscr{L}}$, by Corollary \[ringkombi\], we have $$\label{bbt} A\cup B=[{\mathscr{L}}({{m_{v}}}),{\mathscr{L}}^{-1}(u)]=[{\mathscr{L}}({{m_{v}}}),{s_v}].$$ From [(\[eqd\])]{}, [(\[eqc\])]{}, and [(\[bbt\])]{} follows $$\label{eqw1} \forall w\in[{\mathscr{L}}({{m_{v}}}),{s_v}]\colon{r_{v}}\in {{\mathcal{F}}(v,w)}.$$ Since the vertices in ${\mathscr{L}}$ are primary ordered by their left cliques, there is no vertex $w\in[{{\mathcal{H}}({\mathscr{C}}({\beta({{r_{}}^{i-1}(v)})}))},{{l_{}}^{i-1}(v)}]_{\mathscr{L}}$ which is adjacent to ${{r_{}}^{i-1}(v)}$ and therefore we have ${{\mathcal{R}}_v(w)}\geq i$. We define $$C= \begin{cases} [{\mathscr{L}}({s_v}),{\mathscr{L}}^{-1}({{l_{}}^{i-1}(v)})]_{\mathscr{L}}&\text{if }{\mathscr{L}}({s_v})\neq {{l_{}}^{i-1}(v)},\\ C=\emptyset&\text{else,} \end{cases}$$ i.e. $C$ is the set of vertices after ${s_v}$ and before ${{l_{}}^{i-1}(v)}$ in ${\mathscr{L}}$. Note that $$B,C\subseteq[{{\mathcal{H}}({\mathscr{C}}({\beta({{r_{}}^{i-1}(v)})}))},{{l_{}}^{i-1}(v)}]_{\mathscr{L}}.$$ We have $$\label{inca} \forall {\mathsf{c}}\in[{\alpha({\mathscr{L}}({s_v}))},{\alpha({{l_{}}^{i-1}(v)})}]_{\mathscr{C}}\colon{{l_{}}^{i}(v)}\in {\mathsf{c}}$$ since otherwise $u={\mathscr{L}}({s_v})$ would reach further to the left than ${{l_{}}^{i}(v)}$ although both are adjacent to ${{l_{}}^{i-1}(v)}$ and ${{l_{}}^{i}(v)}$ is the left vertex of ${{l_{}}^{i-1}(v)}$. This would contradict Theorem \[no\_vertex\_further\_than\_lv\]. It follows $$\forall c\in C\colon{{\mathcal{L}}_v(c)}\leq i\leq {{\mathcal{R}}_v(c)},$$ which, by Theorem \[b\], implies $$\label{eqa} \forall c\in C\colon{{l_{v}}}\in {{\mathcal{F}}(v,c)}.$$ Obviously we have $$\label{eqb} \forall w\in[{{l_{}}^{i-1}(v)},{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}\colon{{l_{v}}}\in {{\mathcal{F}}(v,w)}.$$ By Corollary \[ringkombi\], we have $C\cup[{{l_{}}^{i-1}(v)},{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}=[{\mathscr{L}}({s_v}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$ and therefore [(\[eqa\])]{} and [(\[eqb\])]{} imply $$\label{eqw2} \forall w\in[{\mathscr{L}}({s_v}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}\colon {{l_{v}}}\in {{\mathcal{F}}(v,w)}.$$ The proof follows from [(\[eqw1\])]{} and [(\[eqw2\])]{}. In Section \[lvbisv\] we assigned $[{{l_{v}}},x]_{\mathscr{L}}$ to arc $(v,{{l_{v}}})$ for some vertex $x$ to the left of $v$. If $[{\mathscr{L}}({s_v}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$ is assigned to arc $(v,{{l_{v}}})$ additionally (as implied by Theorem \[separator\]) the intervals on arc $(v,{{l_{v}}})$ obviously can be compressed. Therefore, if neither dominating vertices nor counter vertices exist, $v$ given ${\mathscr{L}}$ suffices a shortest path 2-SIRS, but only $(v,{{r_{v}}})$ actually needs to be mapped to two ring-intervals. Below Definition \[de\_r\_v\] we stated that $v$ given ${\mathscr{L}}$ suffices a shortest path 1-SIRS, if ${{r_{v}}}={{l_{v}}}$ or ${{r_{v}}}={{m_{v}}}$, because in these cases the ring-intervals assigned to arc $(v,{{r_{v}}})$ also can be compressed to one ring-interval. Another constellation where $v$ given ${\mathscr{L}}$ suffices a shortest path 1-SIRS is described by the following theorem, which also holds in the subsections below. \[theo:1sirs\] Let $({\mathsf{X}},{\mathscr{C}})$ be a clique-cycle for a circular-arc graph $G=(V,E)$, ${\mathscr{L}}$ given by Listing \[pseudoLabeling\], $v\in V$ not a dominating vertex, ${{l_{v}}}$ the left vertex of $v$, and ${{m_{v}}}$ the middle vertex of $v$. If ${{m_{v}}}$ is adjacent to every vertex face-to-face with $v$ that ${{l_{v}}}$ is not adjacent to, $v$ suffices a shortest path 1-SIRS. Since Theorem \[theo:1sirs\] is not essential for the proof of Theorem \[mainTheo\] but refines the classification of some circular-arc graphs in the interval routing hierarchy, we leave the proof to the interested reader. Note that the theorem implies that $v$ given ${\mathscr{L}}$ suffices a shortest path 1-SIRS, if ${{l_{v}}}$ or ${{m_{v}}}$ is adjacent to every vertex face-to-face with $v$. #### Dominating Vertices or Counter Vertices exist {#gyros} In this subsection we consider the case in that $v$ has counter vertices or the graph contains dominating vertices. We continue with a case-by-case analysis regarding the existence of dominating vertices that are face-to-face with $v$, since, by Theorem \[Lvrl\_v\], these are the only vertices in $B_v$ that are adjacent to $v$. ##### Dominating Face-to-Face Vertices do not exist Assume there are no dominating vertices in $B_v$. Then, by Theorem \[Lvrl\_v\], no vertex in $B_v$ is adjacent to $v$, which, by Corollary \[adjacentToVorCoV\], implies that every vertex in $B_v$ is adjacent to a possible counter vertex $c_v$ of $v$. It is easy to verify that $c_v$ is a vertex to the right of $v$, just as every possible dominating vertex is to the left or right of $v$. Let $u\in{{\mathcal{C}}({v})}\cup{{\mathcal{D}}(G)}$ and $w\in B_v$. Since $v$ and $w$ are not adjacent but $u$ is adjacent to $v$ as well as to $w$, $u$ is a first vertex from $v$ to $w$ and therefore we can distribute the remaining vertices by assigning ring-interval $B_v$ to arc $(v,u)$, which had hitherto one ring-interval assigned. It follows that $v$ given ${\mathscr{L}}$ suffices a shortest path 2-SIRS, whereat only arc $(v,u)$ is actually mapped to two ring-intervals. Note that, if ${{m_{v}}}\in{{\mathcal{C}}({v})}\cup{{\mathcal{D}}(G)}$, $v$ given ${\mathscr{L}}$ suffices a shortest path 1-SIRS, by Theorem \[theo:1sirs\]. ##### Dominating Face-to-Face Vertices exist In Section \[constrCO\] we took care that all dominating vertices have the same left clique and the same right clique, wherefore all dominating vertices appear consecutive in ${\mathscr{L}}$. Therefore we can find two dominating vertices $d_l,d_r$ such that $[d_l,d_r]_{ {\mathscr{L}}}$ is exactly the set of dominating vertices (just as $[{{\mathcal{H}}({\mathsf{c}})},{{\mathcal{T}}({\mathsf{c}})}]_{\mathscr{L}}$ is exactly the set of vertices with left clique ${\mathsf{c}}$, for some ${\mathsf{c}}\in{\mathsf{X}}$). Consider the ring-sequence ${{\mathcal{S}}}[{\mathscr{L}}({{m_{v}}}), x]_{ {\mathscr{L}}}$, with $B_v=[{\mathscr{L}}({{m_{v}}}), x]_{ {\mathscr{L}}}$. The dominating vertex which appears first in ${{\mathcal{S}}}[{\mathscr{L}}({{m_{v}}}), x]_{ {\mathscr{L}}}$ is $d_l$. Thus we assign $[ {\mathscr{L}}({{m_{v}}}),d_l]_{ {\mathscr{L}}}$ to arc $(v,d_l)$. The dominating vertex which appears last in ${{\mathcal{S}}}[{\mathscr{L}}({{m_{v}}}),x]_{ {\mathscr{L}}}$ is $d_r$ and thus we assign $[d_r,x]_{ {\mathscr{L}}}$ to arc $(v,d_r)$. For the remaining dominating vertices $d_i$, we assign $[d_i,d_i]_{\mathscr{L}}$ to arc $(v,d_i)$. It follows that $v$ given ${\mathscr{L}}$ suffices a shortest path 1-SIRS. #### Excluding Counter Vertices {#counteratall} Last we consider the case in that $G$ does not contain dominating vertices and the considered vertex $v$ has no counter vertices but there exists at least one pair $(w,c_w)$ of counter vertices. Then $v$ is adjacent to either $w$ or $c_w$ or to both, as discussed next by a case-by-case analysis. It is easy to see that, since $v$ has no counter vertices and the graph is a real circular-arc graph without dominating vertices, $v$ must have a left vertex ${{l_{v}}}$. Therefore we have $B_v=[{\mathscr{L}}({{m_{v}}}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$ and no vertex in $B_v$ is adjacent to $v$. ##### The Considered Vertex is adjacent to both Counter Vertices Assume $v$ is adjacent to $w$ as well as $c_w$. There is a clique ${\mathsf{c}}\in{\mathsf{X}}$ that contains $v,w$ and $c_w$. Assume we begin at ${\mathsf{c}}$ to go through ${\mathscr{C}}$. We eventually reach a clique ${\mathsf{d}}$ that just contains either $w$ or $c_w$. Without loss of generality, let $w$ be contained in ${\mathsf{d}}$. The first clique ${\mathsf{e}}$ to appear that contains $c_w$ again also contains ${{l_{v}}}$ (we might have $c_w={{l_{v}}}$), otherwise $c_w$ would reach further to the left than ${{l_{v}}}$, which contradicts Theorem \[no\_vertex\_further\_than\_lv\]. Note that ${\mathsf{e}}$ is the left clique of $c_w$. Since $w$ and $c_w$ are counter vertices, the right clique of $w$ cannot appear before the left clique of $c_w$ and therefore the right clique of $w$ cannot appear before ${\mathsf{e}}$. It follows $$\forall {\mathsf{a}}\in[{\mathsf{c}},{\mathsf{e}}]_{\mathscr{C}}\colon w\in{\mathsf{a}}.$$ Since the vertices in ${\mathscr{L}}$ are primary ordered by their left cliques, the left clique of every vertex in $B_v=[{\mathscr{L}}({{m_{v}}}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$ appears in $[{\alpha({{m_{v}}})},{\alpha({{l_{v}}})}]_{\mathscr{C}}$. Because we have $[{\alpha({{m_{v}}})},{\alpha({{l_{v}}})}]_{\mathscr{C}}\subseteq[{\mathsf{c}},{\mathsf{e}}]_{\mathscr{C}}$, $w$ is contained in the left clique of every vertex face-to-face with $v$ and therefore adjacent to every such vertex. Since ${\mathcal{I}}_v$ hitherto maps arc $(v,w)$ to one ring-interval, we can assign $B_v$ to arc $(v,w)$ and have thereby shown that $v$ given ${\mathscr{L}}$ suffices a shortest path 2-SIRS, whereat only arc $(v,w)$ is actually mapped to two ring-intervals. ##### The Considered Vertex is adjacent to one of the Counter Vertices Without loss of generality, let $v$ be adjacent to $w$ but not to $c_w$. Let $r$ be a vertex adjacent to $v$ that, similar to Definition \[de\_r\_v\], reaches farthest to the right. Since, by Theorem \[no\_vertex\_further\_than\_lv\], ${{l_{v}}}$ is one of the vertices adjacent to $v$ that reaches farthest to the left, ${{l_{v}}}$ as well as $r$ are adjacent to $c_w$. In fact, we might have $w={{l_{v}}}$ or $w=r$. Let ${m_{r}}$ be the middle vertex of $r$. By Theorem \[Lvrl\_v\], every vertex to the right of $r$ is adjacent to $r$. Let $A_r=[{\mathscr{L}}(r),{m_{r}}]_{\mathscr{L}}$ be the set of vertices to the right of $r$ ($A_r$ is not empty, since otherwise $G$ would be an interval graph), and note that ${{m_{v}}}$ is contained in $A_r$, if $r\neq{{m_{v}}}$. Consider we go through ${\mathscr{L}}$ beginning at ${\mathscr{L}}({{m_{v}}})$ until we reach ${m_{r}}$ or ${{l_{v}}}$. If we have ${{l_{v}}}={m_{r}}$, we consider it as reaching ${{l_{v}}}$. If we first reach ${{l_{v}}}$, we exactly traversed the vertices in $[{\mathscr{L}}({{m_{v}}}),{{l_{v}}}]_{\mathscr{L}}\subseteq B_v$. It follows that we have $B_v\subseteq A_r$. Therefore no vertex face-to-face with $v$ is adjacent to $v$ but to $r$. We assign $B_v=[{\mathscr{L}}({{m_{v}}}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$ to arc $(v,r)$ and have thereby shown that $v$ given ${\mathscr{L}}$ suffices a shortest path 2-SIRS, whereat only arc $(v,r)$ is actually mapped to two ring-intervals. Now assume we first reach ${m_{r}}$. By Theorem \[Lvrl\_v\], no vertex in $[{\mathscr{L}}({{m_{v}}}),{m_{r}}]_{\mathscr{L}}\subseteq A_r$ is adjacent $v$ but to $r$. Therefore we can assign $[{\mathscr{L}}({{m_{v}}}),{m_{r}}]_{\mathscr{L}}$ to arc $(v,r)$. If we have ${\mathscr{L}}({m_{r}})={{l_{v}}}$ we already have distributed every vertex face-to-face with $v$. Otherwise we consider the vertices in $X=[{\mathscr{L}}({m_{r}}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}\subset B_v$, which, by Theorem \[Lvrl\_v\], are not adjacent to $v$. Let $u\in X$ be some vertex adjacent to a vertex $u'$ that is adjacent to $v$. Since, by Theorem \[no\_vertex\_further\_than\_lv\], ${{l_{v}}}$ is a vertex adjacent to $v$ that reaches farthest to the left, $u$ is also adjacent to ${{l_{v}}}$. For every other vertex $b\in X$ sequence $(v,{{l_{v}}},c_w,b)$ is a shortest path, wherefore we can assign $X$ to arc $(v,{{l_{v}}})$. It follows that $v$ given ${\mathscr{L}}$ suffices a shortest path 2-SIRS, as we can distribute every vertex face-to-face with $v$, by assigning $[{\mathscr{L}}({{m_{v}}}),{m_{r}}]_{\mathscr{L}}$ to arc $(v,r)$ and $[{\mathscr{L}}({m_{r}}),{\mathscr{L}}^{-1}({{l_{v}}})]_{\mathscr{L}}$ to arc $(v,{{l_{v}}})$. Note that only arc $(v,r)$ has to be mapped to two ring-intervals, since the intervals assigned to arc $(v,{{l_{v}}})$ can be compressed. Furthermore, $v$ given ${\mathscr{L}}$ suffices a shortest path 1-SIRS, if ${{m_{v}}}$ is a vertex that reaches farthest to the right. Improving the Space Requirement {#improving} =============================== Although circular-arc graphs refuse shortest path 1-IRSs in general, we show in this section that the shortest path 2-SIRS, that is implied by the proof of Theorem \[mainTheo\], utilizes a number of intervals that can be bounded close to the number that is allowed in an 1-IRS. Let $G_u=(V,E)$ be a real circular-arc graph and $G_d=(V,A)$ be its directed symmetric version. To prove Theorem \[mainTheo\] a cyclic order ${\mathscr{L}}$ on $V$ was fixed and then intervals in ${\mathscr{L}}$ were assigned to the outgoing arcs of an arbitrary vertex $v$, in order to show that $v$ given ${\mathscr{L}}$ suffices a shortest path 2-SIRS. It turned out that in every case (discussed in Section \[mvbislv\]) there is at most one outgoing arc of $v$ that has to be assigned with two intervals. Since we have $2\cdot|E|=|A|$ and a 2-IRS allows up to two intervals per arc, a 2-IRS for $G_d$ permits up to $4\cdot|E|$ intervals in total. Obviously, we can obtain a shortest path 2-SIRS for $G_d$ by applying our results (of Section \[lvbisv\] and \[mvbislv\]) to every vertex in $V$ and labeling the arcs accordingly. It follows that a shortest path 2-SIRS for $G_d$ constructed corresponding to the proof only has one outgoing arc per vertex with two intervals assigned. Therefore the number of intervals in this 2-SIRS is bounded by $2\cdot|E|+|V|$. This result is enhanced by the fact that we have $|V|\leq|E|$, if $G$ is a *real* circular-arc graph. Furthermore, the equality only holds when $G$ is a ring and then our result yields a shortest path 1-SIRS. Since already a shortest path 1-IRS for $G_d$, that not necessarily exists, allows up to $2\cdot|E|$ intervals in total, we believe the estimate gives rise to calling the constructed shortest path 2-SIRS [“economic”]{}. Implementation {#implement} ============== Next we outline how the interval routing scheme, that is implied by the proof of Theorem \[mainTheo\], can be implemented for a given circular-arc graph with $n$ vertices and $m$ edges. We discuss how the vertex order ${\mathscr{L}}$ and the critical vertices ${{l_{v}}},{{m_{v}}},{{r_{v}}},{s_v}$ can be determined in ${\mathcal{O}}(n^2)$ time, for every vertex $v\in V$. It is assumed that not only a circular-arc graph $G=(V,E)$, with $|V|=n$ and $|E|=m$, is given, but also a clique-cycle $({\mathsf{X}},{\mathscr{C}})$ for $G$. Furthermore, every vertex $v$ has a pointer ${\alpha(v)}$ to its left clique and a pointer ${\beta(v)}$ to its right clique in ${\mathsf{X}}$. Since the proof required a similar [“input”]{}, these assumptions are justified. Note that ${\mathsf{X}}$ does not actually need to contain sets of vertices; plain elements that represent the cliques suffice to apply the presented algorithm. By numbering the elements in ${\mathsf{X}}$ as they appear in ${\mathscr{C}}$, beginning at an arbitrary element, we construct a bijective mapping ${\mathcal{N}}\colon{\mathsf{X}}\rightarrow\{1,2,\ldots,|{\mathsf{X}}|\}$. The *broadness of a vertex $v$* is defined as the number of cliques in ${\mathsf{X}}$ that contain $v$. By exploiting ${\mathcal{N}}$ the broadness of a vertex can be computed in $\mathcal{O}(1)$ time. Sorting the vertices ascending by their broadness requires $\mathcal{O}(n\cdot\log(n))$ time. We introduce a list ${\mathcal{Q}}({\mathsf{c}})$ for every clique ${\mathsf{c}}\in{\mathsf{X}}$. Next the vertices are traversed in the sorted order and every vertex $v$ is added to list ${\mathcal{Q}}({\alpha(v)})$. By concatenating the lists as their respective cliques appear in ${\mathscr{C}}$ we obtain ${\mathscr{L}}$ in $\mathcal{O}(n\cdot\log(n))$ time. The left vertex ${{l_{v}}}$ of a given vertex $v\in V$ can be determined as follows: we inspect all vertices that are adjacent to $v$ and store the vertex $w$ that reaches farthest to the left (for two adjacent vertices $w,w'$, we can determine in constant time which one reaches further to the left, by considering ${\mathcal{N}}({\alpha(w)})$ and ${\mathcal{N}}({\alpha(w')})$) and appears first in ${\mathcal{Q}}({\alpha(w)})$ (this can also be decided in constant time, by numbering the elements in each list consecutively). Vertex $w$ must be ${{l_{v}}}$. It follows that the set of left vertices can be computed in $\mathcal{O}(m)$ time. The right vertices can be computed in a similar way. The determination of the middle vertices is much easier and can be accomplished in $\mathcal{O}(n)$ for all vertices. For a given vertex $v$, ${{l_{}}^{i}(v)}$ and ${{r_{}}^{i}(v)}$ can be determined by traversing the sequences ${{l_{}}^{1}(v)},{{l_{}}^{2}(v)},\ldots$ and ${{r_{}}^{1}(v)},{{r_{}}^{2}(v)},\ldots$ synchronously; then the determination of ${s_v}$ is a simple task and can be realized in ${\mathcal{O}}(n)$ for a specific vertex and in ${\mathcal{O}}(n^2)$ for all vertices. It follows that the overall runtime of the algorithm is in $\mathcal{O}(n^2+m)=\mathcal{O}(n^2)$. The algorithm can easily be extended to handle the existence of dominating and counter vertices correctly without exceeding the time bound. Conclusions {#conc} =========== We have shown that the class of circular-arc graphs is in 2-SIRC, which is an optimal result, since it is not in 1-SIRC or even in 1-IRC. Although we classified circular-arc graphs with respect to non-linear interval routing schemes, the interval routing hierarchy allows to deduce that every circular-arc graph is in 3-LIRC, although we do not know if this bound is tight, that is to say, we do not know, if every circular-arc graph is in 2-LIRC. We highlighted some cases in that a real circular-arc graph allows a shortest path 1-SIRS though there are some cases uncovered, i.e. there exist circular-arc graphs that allow shortest path 1-SIRSs but the approach outlined in Section \[main\] only yields a shortest path 2-SIRS. A question that emerges in this context is, if there exist circular-arc graphs that are in 1-SIRC but the vertex labeling generated by Listing \[pseudoLabeling\] only permits a shortest path 2-SIRS, in other words, does already the vertex order prevent the generation of a shortest path 1-SIRS or only the improper choice of arc labelings? Although the class of interval graphs and unit circular-arc graphs is in 1-SIRC, the class of circular-arc graphs, which is a super-class of the former two, is not even in 1-IRC. However, we showed in Section \[improving\] that a shortest path 2-SIRS for every circular-arc graph only requires a number of intervals that is closer to the number of intervals allowed by an 1-IRS than by a 2-IRS. The graph in Figure \[fig:animals\] was found within the framework of this paper and is not in 1-IRC. In contrast to a wheel graph, which is somehow the classic example to show that the class of circular-arc graphs is not in 1-IRC, it is more complex and does not contain a dominating vertex. We think this is worth mentioning, since we initially had the suspicion that dominating or [“almost dominating”]{} vertices were the feature that circular-arc graphs need to provide in order to prohibit shortest path 1-IRSs. Note that the graph persists of two equal parts and could be easily extended by inserting more of these parts, while it would never allow a shortest path 1-IRS. [^1]: Corresponding author. [^2]: The *directed symmetric version of an undirected graph* $G_u=(V,E)$ is the directed graph $G_d=(V,A)$ with $A=\{(v,w),(w,v)~|~\{v,w\}\in E\}.$ [^3]: If the condition holds for one element in $A$, it holds for all elements in $A$. [^4]: Note that, $w\in{{\mathcal{C}}({v})}~\Leftrightarrow~v\in{{\mathcal{C}}({w})}$. [^5]: Intervals can be uniquely represented by their two marginal elements. [^6]: Two vertices in a graph are called *true twins* if they are adjacent to the same set of vertices and to each other. [^7]: It follows that ${{l_{}}^{2}(v)}$ is the left vertex of the left vertex of $v$ and ${{r_{}}^{2}(v)}$ is the right vertex of the right vertex of $v$, etc. [^8]: If we have $1=j<k$, we get a contradiction from the fact $[{\alpha(w_1)},{\beta(w_1)}]_{\mathscr{C}}\subseteq[{\alpha({{l_{v}}})},{\beta({{r_{v}}})}]_{\mathscr{C}}$. [^9]: This equality holds by Corollary \[coro:tail vertexheal\]. [^10]: This vertex $u$ may be $l^i(v)$. Also there may be vertices in $A$ that are adjacent to ${{l_{}}^{i-1}(v)}$, in other words, we may have [“passed by”]{} a vertex that is adjacent to ${{l_{}}^{i-1}(v)}$ already.

ArXiv

--- abstract: 'The linking integral is an invariant of the link-type of two manifolds immersed in a Euclidean space. It is shown that the ordinary Gauss integral in three dimensions may be simplified to a winding number integral in two dimensions. This result is then generalized to show that in certain circumstances the linking integral between arbitrary manifolds may be similarly reduced to a lower dimensional integral.' address: Drexel University author: - 'Daniel J. Cross' title: Linking Integral Projection --- Reduction of the Gauss Integral to the Winding Number Integral {#sec:wind} ============================================================== The linking number of two disjoint oriented closed curves in $\bbR^3$ is an integer invariant that in some sense measures the extent of linking between the curves. While there are many equivalent ways to compute this number[@Rol03], the most well-known is the linking integral of Gauss. In this section we show that this integral in 3-space may always be simplified to an integral in 2-space which is equivalent to a winding number integral. Given two disjoint immersed closed curves $s\mapsto \gamma_1(s)$ and $t\mapsto\gamma_2(t)$ in $\bbR^3$, the Gauss linking integral of the pair reduces to a sum of winding numbers of one curve about a sequence of points determined by the other, contained in some 2-dimensional hyperplane. The link of $\gamma_1$ and $\gamma_2$, $\operatorname{lk}(\gamma_1,\gamma_2)$, is given by the Gauss integral, $$\operatorname{lk}(\gamma_1,\gamma_2)=\frac{1}{4\pi}\int\det\left(r,{\displaystyle\frac{\partial \gamma_1}{\partial s}},{\displaystyle\frac{\partial \gamma_2}{\partial t}}\right)\frac{dsdt}{r^3},$$ where each term in the determinant is a column vector and $r=\gamma_2-\gamma_1$ is the relative position vector. Through a homotopy of the maps we may arrange $\gamma_1$ to lie in the plane $x_3=0$ with $\gamma_2$ intersecting the plane perpendicularly in a finite number of points $p_i$. This may be done so that $\gamma_1$ and $\gamma_2$ remain disjoint throughout. Since the Gauss integral is a homotopy invariant, $\operatorname{lk}(\gamma_1,\gamma_2)$ is preserved through this deformation. If the homotopy was merely continuous we may replace it with an arbitrarily close smooth homotopic approximation. Next, deform $\gamma_2$ near the intersection with the plane so that it becomes a straight line segment perpendicular to the plane in a neighborhood of each intersection point. Now deform it further by “stretching” it away from the plane so that the straight line segments are extended further away from the plane and the rest of $\gamma_2$ is pushed further away from the plane. In the limit that the stretching goes off to infinity, the denominator of the integral falls off sufficiently fast that its contribution goes to zero. We are left with a finite number of infinite line segments perpendicular to the plane and disjoint from $\gamma_1$. We assume each line is parameterized in the standard way, $t\mapsto \pm x_3$. We will now assume each line parameterized by $t\mapsto x_3$ so that $\partial\gamma_2/\partial t=e_3$, but introduce an orientation to each point $o(p_i)$ which is $\pm 1$ depending on the original parameterization of the corresponding line in an obvious way. We then see that the linking integral becomes $$\sum_io(p_i)\frac{1}{4\pi}\int\det\left(r,{\displaystyle\frac{\partial \gamma_2}{\partial s}},e_3\right)\frac{dsdt}{(\rho^2+t^2)^{3/2}},$$ where $\rho$ is the restriction of $r$ to the plane $x_3=0$. Notice that since $e_3=(0,0,1)^t$ the determinant reduces to that of the upper-left block, which is $\det(\rho_{12},\partial\gamma_2/\partial s))$ and is independent of $t$. Thus we may evaluate the integral $$\int_\bbR\frac{dt}{(\rho^2+t^2)^{3/2}}=\frac{2}{\rho^2},$$ and the linking integral becomes $$\sum_io(p_i)\frac{1}{2\pi}\int\det\left(\rho,{\displaystyle\frac{\partial \gamma_2}{\partial s}}\right)\frac{ds}{\rho^2},$$ which is easily seen to be the sum of the winding numbers of $\gamma_2$ about each point $p_i$ times the orientation of $p_i$. The construction in the proof also allows one to show the linking integral may also be given as an intersection number of $\gamma_1$ with a surface spanned by $\gamma_2$. Indeed, perturb $\gamma_2$ to an embedding and let $S$ be a Seifert surface constructed by Seifert’s algorithm[@Seifert34; @Rol03]. The number of Seifert discs above a $p_i$ is precisely the winding number of $\gamma_1$ about $p_i$ and the induced orientation of each Seifert disc is given by the orientation of the bounding curve. Finally, with $o(p_i)$ we have the signed intersection number of $\gamma_2$ with the Seifert disc, and the sum over all gives the signed intersection number of $\gamma_2$ with $S$. The General Linking Integral Projection ======================================= In this section the proposition of section \[sec:wind\] is generalized from curves to arbitrary compact boundaryless oriented manifolds $M^n$ and $N^n$ mapped disjointly into $\bbR^{p+1}$, $p=m+n$. In this case one may define a linking number by ${\operatorname{lk}}(M,N)=(-1)^{m}\deg \hat r$, where $\hat r$ is the unit relative position vector defined by $$\begin{aligned} \hat r:M\times N&\to S^p\\ (x,y)&\mapsto \frac{r}{||r||}=\frac{x-y}{||x-y||},\end{aligned}$$ where $x$ and $y$ are points in the images of $M$ and $N$ in $\bbR^{p+1}$ respectively. We will show that under certain conditions the linking number calculation reduces to a calculation in a hyperplane. We note that these expressions may be defined with differing sign conventions in which case the conclusion of the theorem will hold up to a sign. The present convention is most convenient for expressing the present result. Given $M$ and $N$ as above, suppose that there exists smooth homotopies of $M$ and $N$ maintaining disjointness and taking $M$ into an $m+n'+1$-dimensional hyperplane $H$, $0\leq n'\leq n$, and that $H$ intersects $N$ transversely in the submanifold $N'$. Then $\operatorname{lk}(M,N)=\operatorname{lk}(M,N')$, where the first linking integral is taken in $\bbR^{p+1}$ and the second in $H\cong\bbR^{p'+1}$, where $p=m+n$ and $p'=m+n'$. It is straightforward to show[@frankel06] that the degree of this map may be written explicitly as $$\label{eq:app_link} \deg\hat r = \frac{(-1)^m}{{\rm vol}S^p}\int_{M\times N} \det\left(r,{\displaystyle\frac{\partial x}{\partial s}},{\displaystyle\frac{\partial y}{\partial t}}\right) \frac{dsdt}{||r||^{p+1}},$$ where $s$ and $t$ represent oriented local coordinates $s_i$ and $t_j$ on $M$ and $N$ respectively, and the quantities in the determinant are column vectors. We now homotope $M$ into $H\simeq\bbR^{p'+1}$ and homotope $N$ so that it intersects $H$ transversely. The intersection $N'=\cup N'_i$ will be a finite disjoint union of closed oriented manifolds of dimension $n'$ (the codimension of the transverse intersection of two manifolds is the sum of their codimensions). We may actually assume that $N\perp H$ (in the Euclidean metric of $\bbR^{p+1}$) so that $N$ is locally of the form $N'\times\bbR^{n-n'}$ in some neighborhood of $H$. Now extend this local product decomposition by pushing the rest of $N$ off to $\infty$ as was done in section \[sec:wind\]. Since the integrand in Eq.\[eq:app\_link\] falls off sufficiently fast with distance, this contribution to the integral goes to zero, so we may make the replacement $N\to N'\times\bbR^{n-n'}$. Adapt the coordinates on $N$ with respect to the product decomposition so that the last $n-n'$ coordinates are Euclidean coordinates on $\bbR^{n-n'}$. The partial derivatives $\partial y/\partial t_i$ with respect to these coordinates are just $\pm e_i$, the $i$th unit vector, but the signs may vary on different $N'_i$. We may absorb these signs into the orientation of the components, considering their orientations reversed if necessary (rather than explicitly introducing an orientation function as was done in section \[sec:wind\]). The matrix in the integrand has a block structure with an $n-n'$ unit matrix in the lower right block and zero in the upper right block. Thus the determinant may be replaced with just that of the upper left block. Since this matrix is independent of the last $n-n'$ coordinates the distance function $r$ reduces to $\rho=r|_H$. It remains to evaluate the integral $$I=\int_{-\infty}^\infty \cdots\int_{-\infty}^\infty \frac{dt_{q+1}\cdots dt_n}{\left(\rho^2+\sum t_i^2\right)^{(p+1)/2}},$$ for $i=q+1,\ldots,n$. Write $t_n=z$ and $a=\rho^2+\sum t_i^2$, $i\neq n$ and then $$\int_{-\infty}^\infty \frac{dz}{\left(a+z^2\right)^{(p+1)/2}}= \frac{\sqrt{\pi}}{a^{p/2}}\frac{\Gamma(\frac{p}{2})}{\Gamma(\frac{p+1}{2})} =\frac{1}{a^{p/2}}\frac{{\rm vol}_{S^p}}{{\rm vol}_{S^{p-1}}},$$ using the well-known expression for the volume of a sphere. By progressively isolating the variables $t_i$ we obtain an integral of the same form but with $p$ decreasing by one each time. Proceeding by induction we obtain $$I=\frac{1}{||\rho||^{p'+1}}\frac{{\rm vol}_{S^p}}{{\rm vol}_{S^{p'}}},$$ where $p'=p-n+n'=m+n'$. Hence Eq.\[eq:app\_link\] becomes $$\begin{aligned} \deg\hat r &= \frac{(-1)^m}{{\rm vol}S^{p'}}\int_{M\times N'} \det\left(\rho,{\displaystyle\frac{\partial x}{\partial t}},{\displaystyle\frac{\partial y}{\partial s}}\right) \frac{dsdt}{||\rho||^{p'+1}}\\ &=(-1)^{m}\deg\hat\rho,\end{aligned}$$ which is $\operatorname{lk}(M,N')$. This result has been applied in [@Cross09a] to demonstrate that two solid tori ($D^2\times S^1$) embedded into $\bbR^4$ differing by a Dehn twist are non-isotopic. In fact, it is shown that the isotopy classes of embeddings of the solid torus into $\bbR^4$ are in bijective correspondence with the integers, the correspondence given by the number of applied Dehn twists.

ArXiv

[**Solving a Continuous Multifacility Location Problem by DC Algorithms**]{}\ [Anuj Bajaj]{}[^1], [Boris S. Mordukhovich]{}[^2], [Nguyen Mau Nam]{}[^3], and [Tuyen Tran]{}[^4].\ The paper presents a new approach to solve multifacility location problems, which is based on mixed integer programming and algorithms for minimizing differences of convex (DC) functions. The main challenges for solving the multifacility location problems under consideration come from their intrinsic discrete, nonconvex, and nondifferentiable nature. We provide a reformulation of these problems as those of continuous optimization and then develop a new DC type algorithm for their solutions involving Nesterov’s smoothing. The proposed algorithm is computationally implemented via MATLAB numerical tests on both artificial and real data sets.\ [**Key words.**]{} Mixed integer programming, multifacility location, difference of convex functions, Nesterov’s smoothing, the DCA\ [**AMS subject classifications.**]{} 49J52, 49J53, 90C31 \[section\] \[Theorem\][Proposition]{} \[Theorem\][Remark]{} \[Theorem\][Lemma]{} \[Theorem\][Corollary]{} \[Theorem\][Definition]{} \[Theorem\][Example]{} Introduction and Problem Formulation {#intro} ==================================== In the 17th century, Pierre de Fermat posed a problem of finding for a point that minimizes the sum of its Euclidean distances to three given points in the plane. The problem was soon solved by Evangelista Torricelli, and it is now known as the *Fermat-Torricelli problem*. This problem and its extended version that involves a finite number of points in higher dimensions are examples of continuous [*single facility*]{} location problems. Over the years several generalized models of the Fermat-Torricelli type have been introduced and studied in the literature with practical applications to facility location decisions; see [@HM2015; @Martini; @n2; @nars; @nh; @NamGiles] and the references therein. An important feature of single facility location problems and the problems studied in the aforementioned references is that only one *center/server* has to be found to serve a finitely many *demand points/customers*. However, numerous practical applications lead to formulations of facility location problems in which more than one center must be found to serve a finite number of demand points. Such problems are referred to as *multifacility location problems* (MFLPs). Given a finite number of demand points $a_1,\ldots,a_n$ in $\R^d$, we consider here the facility location in which $k$ centers $v_1,\ldots,v_k$ ($1\le k\le n$) in $\R^d$ need to be found to serve these demand points by assigning each of them to its nearest center and minimizing the total distances from the centers to the assigned demand points. In the case where $k=1$, this problem reduces to the generalized Fermat-Torricelli problem of finding a point that minimizes the sum of the distances to a finite number of given points in $\R^d$. Let us formulate the problem under consideration in this paper as the following problem of *mixed integer programming* with [*nonsmooth*]{} objective functions. It is convenient to use a variable $k\times d$-matrix $\mathbf V$ with $v_i$ as its $i$th row to store the centers to be found. We also use another variable $k\times n$-matrix $\mathbf U=[u_{i,j}]$ with $u_{i,j}\in\{0,1\}$ and $\sum_{i=1}^{k}u_{i,j}=1$ for $j=1,\ldots,n$ to assign demand points to the centers. The set of all such matrices is denoted by $\mathcal{U}$. Note that $u_{i,j}=1$ if the center $v_i$ is assigned to the demand point $a_j$ while $\sum_{i=1}^{k}u_{i,j}=1$ means that the demand point $a_j$ is assigned to only one center. Our goal is to solve the constrained optimization problem formulated as follows: $$\begin{aligned} \label{main problem} \begin{array}{ll} &\mbox{\rm minimize }\mathcal{F}(\mathbf U,\mathbf V):=\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\|a_{j}-v_{i}\|\\ &\mbox{\rm subject to }\mathbf U\in\mathcal{U}\;\;\mbox{\rm and }\;\mathbf V\in\R^{k\times d}. \end{array}\end{aligned}$$ Taking into account that $u_{i,j}\in\{0,1\}$, it is convenient to use $u^2_{i,j}$ instead of $u_{i,j}$ in the definition of the objective function $\mathcal{F}$; See Section 4. Note that a similarly looking problem was considered by An, Minh and Tao [@ha] for different purposes. The main difference between our problem and the one from [@ha] is that in [@ha] the squared Euclidean norm is used instead of the Euclidean norm in our formulation. From the point of applications this difference is significant; namely, using the Euclidean norm allows us to model the total distance in *supply delivery*, while using the squared Euclidean norm is meaningful in [*clustering*]{}. Mathematically these two problems are essentially different as well. In addition to the challenging discrete nature and nonconvexity that both problems share, the objective function of our multifacility location problem is [*nondifferentiable*]{} in contrast to [@ha]. This is yet another serious challenge from both theoretical and algorithmic viewpoints. Observe also that for $k=1$ our problem becomes the aforementioned generalized Fermat-Torricelli problem that does not have a closed-form solution, while the problem considered in [@ha] reduces to the standard problem of minimizing the sum of squares of the Euclidean distances to the demand points. The latter has a simple closed form solution given by the mean of the data points. In this paper we develop the following algorithmic procedure to solve the formulated nonsmooth problem of mixed integer programming: [**(i)**]{} Employ [*Nesterov’s smoothing*]{} to approximate the nonsmooth objective function in by a family of [*smooth*]{} functions, which are represented as [*differences of convex*]{} ones. [**(ii)**]{} Enclose the obtained smooth discrete problems into constrained problems of [*continuous DC optimization*]{} and then approximate them by unconstrained ones while using [*penalties*]{}. [**(iii)**]{} Solve the latter class of problems by developing an appropriate modification of the algorithm for minimizing differences of convex functions known as the [*DCA*]{}. As a result of all the three steps above, we propose a [*new algorithm*]{} for solving the class of multifacility local problems of type , verify its efficiency and implementation with MATLAB numerical tests on both artificial and real data sets. Recall that the early developments on the DCA trace back to the work by Tao in 1986 with more recent results presented in [@AnNam]–[@Tao], [@TA1], [@TA2], and the bibliographies therein. Nesterov’s smoothing technique was introduced in his seminal paper [@n83] and was further developed and applied in many great publications; see, e.g., [@n; @n18] for more details and references. The combination of these two important tools provides an effective way to deal with nonconvexity and nondifferentiability in many optimization problems encountered in facility location, machine learning, compressed sensing, and imaging. It is demonstrated in this paper in solving multifacility location problems of type . The rest of this paper is organized as follows. Section \[Def\] contains the basic definitions and some preliminaries, which are systematically employed in the text. In Section \[Survey\] we briefly overview two versions of the DCA, discuss their convergence, and present two examples that illustrate their performances. Section \[nest-smooth\] is devoted to applying Nesterov’s smoothing technique to the objective function of the multifacility location problem and constructing in this way a smooth approximation of the original problem by a family of DC ones. Further, we reduce the latter smooth DC problems of discrete constrained optimization to unconstrained problems by using an appropriate penalty function method. Finally, the obtained discrete optimization problems are enclosed here into the DC framework of unconstrained continuous optimization. In Section \[TheAlgorithm\] we proposed, based on the above developments, a new algorithm to solve the multifacility location problem by applying the updated version of the DCA taken from Section \[Survey\] to the smooth DC problems of continuous optimization constructed in Section \[nest-smooth\]. The proposed algorithm is implemented in this section to solving several multifacility problems arising in practical modeling. Section \[Conclusion\] summarizes the obtained results and discusses some directions of future research. Basic Definitions and Preliminaries {#Def} =================================== For the reader’s convenience, in this section we collect those basic definitions and preliminaries, which are largely used throughout the paper; see the books [@HUL; @bmn; @r] for more details and proofs of the presented results. Consider the difference of two convex functions $g-h$ on a finite-dimensional space and assume that $g\colon\R^d\to\oR:=(-\infty,\infty]$ is extended-real-valued while $h\colon\R^d\to\R$ is real-valued on $\R^d$. Then a general problem of [*DC optimization*]{} is defined by: $$\label{DCP} \mbox{\rm minimize }f(x):=g(x)-h(x),\quad x\in\R^d.$$ Note that problem is written in the unconstrained format, but—due to the allowed infinite value for $g$—it actually contains the domain constraint $x\in\dom(g):=\{u\in\R^d\;|\;g(u)<\infty\}$. Furthermore, the explicit constraints of the type $x\in\Omega$ given by a nonempty convex set $\Omega\subset\R^d$ can be incorporated into the format of via the indicator function $\delta_\Omega(x)$ of $\Omega$, which equals $0$ for $x\in\Omega$ and $\infty$ otherwise. The representation $f=g-h$ is called a [*DC decomposition*]{} of $f$. Note that the class of DC functions is fairly large and include many nonconvex functions important in optimization. We refer the reader to the recent book [@m18] with the commentaries and bibliographies therein for various classes of nonconvex optimization problems that can be represented in the DC framework . Considering a nonempty (may not be convex) set $\Omega\subset\R^d$ and a point $x\in\R^d$, define the *Euclidean projection* of $x$ to $\Omega$ by $$\label{proj} P(x;\Omega):=\big\{w\in\Omega\;\big|\;\|x-w\|=d(x;\Omega)\big\},$$ where $d(x;\Omega)$ stands for the *Euclidean distance* from $x$ to $\Omega$, i.e., $$\label{dist} d(x;\Omega):=\inf\big\{\|x-w\|\;\big|\;w\in\Omega\big\}.$$ Observe that $P(x;\Omega)\ne\emp$ for closed sets $\Omega$ while being always a singleton if the set $\Omega$ is convex. Given further an extended-real-valued and generally nonconvex function $\ph\colon\R^d\to\oR$, the [*Fenchel conjugate*]{} of $\ph$ is defined by $$\ph^*(v):=\sup\big\{\la v,x\ra-\ph(x)\;\big|\;x\in\R^d\big\},\quad v\in\R^d.$$ If $\ph$ is [*proper*]{}, i.e., $\dom(\ph)\ne\emp$, its Fenchel conjugate $\ph^*\colon\R^d\to\oR$ is automatically convex. The [*subdifferential*]{} of $\ph\colon\R^d\to\oR$ at $\ox\in\dom(\ph)$ is the set of subgradients given by $$\label{sub} \partial\ph(\ox):=\big\{v\in\R^d\;\big|\;\la v,x-\ox\ra\le\ph(x)-\ph(\ox)\;\;\mbox{\rm for all }\;x\in\R^d\big\}.$$ If $\ox\notin\dom(\ph)$, we let $\partial\ph(\ox):=\emp$. Recall that for functions $\ph$ differentiable at $\ox$ with the gradient $\nabla\ph(\ox)$ we have $\partial\ph(\ox)=\{\nabla\ph(\ox)\}$. The following proposition gives us a two-sided relationship between the Fenchel conjugates and subgradients of convex functions. \[characterization\] Let $\ph\colon\R^d\to\oR$ be a proper, lower semicontinuous, and convex function. Then $v\in\partial\ph^*(y)$ if and only if $$\label{e1} v\in\mbox{\rm argmin}\,\big\{\ph(x)-\la y,x\ra\;\big|\;x\in\R^d\big\}.$$ We have furthermore that $w\in\partial\ph(x)$ if and only if $$\label{e2} w\in\mbox{\rm argmin}\,\big\{\ph^*(y)-\la x,y\ra\;\big|\;y\in\R^d\big\}.$$ [**Proof.**]{} To verify the first assertion, suppose that is satisfied and then get $0\in \partial\psi(v)$, where $\psi(x):=\ph(x)-\la y,x\ra$ as $x\in\R^d$. It tells us that $$0\in\partial\ph(v)-y,$$ and hence $y\in\partial\ph(v)$, which is equivalent to $v\in\partial\ph^*(y)$ due to the biconjugate relationship $\ph^{**}=\ph$ valued under the assumptions made. In the opposite way, assuming $v\in\partial\ph^*(y)$ gives us by the proof above that $0\in\partial\psi(v)$, which clearly yields and thus justifies the first assertion. To verify the second assertion, suppose that holds and then get $0\in\partial\psi(w)$, where $\psi(y):=\ph^*(y)-\la x,y\ra$ as $y\in\R^d$. This clearly implies that $$0\in\partial\ph^*(w)-x,$$ and hence $x\in\partial\ph^*(w)$, which is equivalent to $w\in\partial\ph(x)$ due to the biconjugate relationship. The proof of the opposite implication in is similar to the one given above. $\h$ Finally in this section, recall that for a given variable matrix $\mathbf U\in\R^{k\times n}$ as in the optimization problem , the *Frobenius norm* on $\R^{k\times n}$ is defined by $$\label{frob} \|\mathbf U\|_{F}:=\sqrt{\sum_{i=1}^{k}\sum_{j=1}^{n}|u_{i,j}|^2}.$$ Overview of the DCA and Some Examples {#Survey} ===================================== In this section we first briefly overview two algorithms of the DCA type to solve DC problems while referring the reader to [@TA1; @TA2] for more details and further developments. Then we present numerical examples illustrating both algorithms.\ [**Algorithm 1: DCA-1**]{}. --------------------------------------- [INPUT]{}: $x_0\in\R^d$, $N\in\N$. [**for**]{} $l=1,\ldots,N$ [**do**]{} $y_{l-1}\in \partial h(x_{l-1})$. $x_{l}\in\partial g^*(y_{l-1})$. [**end for**]{} [OUTPUT]{}: $x_{N}$. --------------------------------------- Since the convex function $h\colon\R^d\to\R$ in is real-valued on the whole space $\R^d$, we always have $\partial h(x)\ne\emp$ for all $x\in\R^d$. At the same time, the other convex function $g\colon\R^d\to\oR$ in is generally extended-real-valued, and so the subdifferential of its conjugate $g^*$ may be empty. Let us present an efficient condition that excludes this possibility. Recall that a function $g\colon\R^d\to\oR$ is *coercive* if $$\lim_{\|x\|\to\infty}\frac{g(x)}{\|x\|}=\infty.$$ \[nonemp-sub\] Let $g\colon\R^d\to\oR$ be a proper, lower semicontinuous, and convex function. If in addition $g$ is coercive, then $\partial g^*(v)\ne\emp$ for all $v\in\R^d$. [**Proof.**]{} Since $g$ is proper, the conjugate function $g^*$ takes values in $(-\infty,\infty]$ being convex on $\R^d$. Taking into account that $g$ is also and lower semicontinuous and invoking the aforementioned biconjugate relationship, we find $w\in\R^d$ and $c\in\R$ such that $$\label{sub-est} c+\la w,x\ra\le g(x)\;\;\mbox{\rm for all }\;x\in\R^d.$$ The coercivity property of $g$ ensures the existence of $\eta>0$ for which $$\|x\|\big(\|w\|+1\big)\le g(x)\;\;\mbox{\rm whenever }\;\|x\|\ge\eta.$$ It follows furthermore that $$\sup\big\{\la v,x\ra-g(x)\;\big|\;\|x\|\ge\eta\big\}\le-\|x\|\;\;\mbox{\rm for any }\;v\in\R^d.$$ By using , we arrive at the estimates $$\sup\big\{\la v,x\ra-g(x)\;\big|\;\|x\|\le\eta\big\}\le\sup\big\{\la v,x\ra-\la w,x\ra-c\;\big|\;\|x\|\le\eta\big\}<\infty,\quad v\in\R^d.$$ This tells us that $g^*(v)<\infty$, and therefore $\dom(g^*)=\R^d$. Since $g^*$ is a convex function with finite values, it is continuous on $\R^d$ and hence $\partial g^*(v)\ne\emp$ for all $v\in\R^d$. $\h$ To proceed further, recall that a function $\ph\colon\R^d\to\oR$ is $\gamma$-*convex* with a given modulus $\gamma\ge 0$ if the function $\psi(x):=\ph(x)-\frac{\gamma}{2}\|x\|^2$ as $x\in\R^d$ is convex on $\R^d$. If there exists $\gamma>0$ such that $\ph$ is $\gamma-$convex, then $\ph$ is called *strongly convex* on $\R^d$. We also recall that a vector $\ox\in\R^d$ is a *stationary point* of the DC function $f$ from if $$\partial g(\ox)\cap\partial h(\ox)\ne\emp.$$ The next result, which can be derived from [@TA1; @TA2], summarizes some convergence results of the DCA. Deeper studies of the convergence of this algorithm and its generalizations involving the Kurdyka-Lojasiewicz (KL) inequality are given in [@AnNam; @TAN]. \[dsa-conv\] Let $f$ be a DC function taken from , and let $\{x_l\}$ be an iterative sequence generated by Algorithm [1]{}. The following assertions hold: 1. The sequence $\{f(x_l)\}$ is always monotone decreasing. 2. Suppose that $f$ is bounded from below, that $g$ is lower semicontinuous and $\gamma_1$-convex, and that $h$ is $\gamma_2$-convex with $\gamma_1+\gamma_2>0$. If $\{x_l\}$ is bounded, then the limit of any convergent subsequence of $\{x_l\}$ is a stationary point of $f$. In many practical applications of Algorithm 1, for a given DC decomposition of $f$ it is possible to find subgradient vectors from $\partial h(x_l)$ based on available formulas and calculus rules of convex analysis. However, it may not be possible to explicitly calculate an element of $\partial g^*(y_l)$. Such a situation requires either constructing a more suitable DC decomposition of $f$, or finding $x_{l+1}\in\partial g^*(y_l)$ by using the description of Proposition \[characterization\]. This leads us to the following modified version of the DCA. [**Algorithm 2: DCA-2**]{}. ---------------------------------------------------------------------------- [INPUT]{}: $x_0\in\R^d$,$N\in\N$ [**for**]{}$l=1,\ldots,N$ \[1ex\] Find $y_{l-1}\in \partial h(x_{l-1})$ \[1ex\] Find $x_{l}$ by solving the problem: \[1ex\] $\mbox{\rm minimize}\;\ph_l(x):=g(x)-\la y_{l-1},x\ra,\;x\in\R^d.$ \[1ex\] [**end for**]{} [OUTPUT]{}: $x_{N}$ ---------------------------------------------------------------------------- Let us now present two examples illustrating the performances of Algorithms 1 and 2. The first example concerns a polynomial function of one variable. \[1D\] Consider the function $f\colon\R\to\R$ given $$f(x):=x^4-2x^2+2x-3\;\;\mbox{\rm for }\;x\in\R.$$ This function admits the DC representation $f=g-h$ with $g(x):=x^4$ and $h(x):=2x^2-2x+3$. To minimize $f$, apply first the [*gradient method*]{} with constant stepsize. Calculating the derivative of $f$ is $f^\prime(x)=4x^3-4x+2$ and picking any starting point $x_0\in\R$, we get the sequence of iterates $$x_{l+1}=x_l-t(4x_l^3-4x_l+2)\;\;\mbox{\rm for }\;l=0,1,\ldots.$$ constructed by the gradient method with stepsize $t>0$. The usage of the DC [*Algorithm*]{} 1 (DCA-1) gives us $y_l=\nabla h(x_l)=4x_l-2$ and then $g^*(x)=3(x/4)^{4/3}$ with $\nabla g^*(x)=(x/4)^{1/3}$. Thus the iterates of DCA-1 are as follows: $$x_{l+1}=\nabla g^*(y_l)=\Big(\frac{y_l}{4}\Big)^{1/3}=\Big(\frac{4x_l-2}{4}\Big)^{1/3}=\Big(\frac{2x_l-1}{2}\Big)^{1/3},\quad l=0,1,\ldots.$$ ![Convergence of the gradient method and DCA-1.[]{data-label="convergencegraph"}](Ex33.eps) Figure \[convergencegraph\] provides the visualization and comparison between the DCA-1 and the gradient method. It shows that for $x_0=0$ and $t=0.01$ the DCA-1 exhibits much faster convergence. The next two-dimensional example illustrates the performance of the DCA-2. \[2D\] Consider the nonsmooth optimization problem defined by $$\mbox{\rm minimize }\;f(x_1,x_2):=x_1^4+x_2^2-2x_1^2-|x_2|\;\;\mbox{\rm over }\;x=(x_1,x_2)\in\R^2.$$ The graph of the function $f$ is depicted in Figure \[graph3D\]. Observe that this function has four global minimizers, which are $(1,0.5),(1,-0.5),(-1,0.5)$, and $(-1,-0.5)$. It is easy to see that $f$ admits a DC representation $f=g-h$ with $g(x_1,x_2):=x_1^4+x_2^2$ and $h(x_1,x_2):=2x_1^2+|x_2|$. We get the gradient $\nabla g(x)=[4x_1^3,2x_2]^T$ and the Hessian $\nabla^2g(x)= \begin{bmatrix} 12{x_1}^2&0\\ 0&2 \end{bmatrix}$, while an explicit formula to calculate $\partial g^*(y_l)$ is not available. Let us apply the [DCA]{}-2 to solve this problem. The subdifferential of $h$ is calculated by $$\partial h(x)=\big[4x_1,\mbox{\rm sign}(x_2)\big]^T\;\;\mbox{\rm for any }\;x=(x_1,x_2)\in\R^2.$$ Having $y_l$, we proceed with solving the subproblem $$\label{subpr} \mbox{\rm minimize }\;\ph_l(x):=g(x)-\la y_l,x\ra\;\;\mbox{\rm over }\;x\in\R^2$$ by the classical Newton method with $\nabla^2\ph_l(x)=\nabla^2g(x)$ and observe that the DCA-2 shows its superiority in convergence with different choices of initial points. Figure \[convergencegraph2D\] presents the results of computation by using the DCA-2 with the starting point $x_0=(-2,2)$ and employing the Newton method with $\epsilon=10^{-8}$ to solve subproblem . ![Convergence of the DCA-2.[]{data-label="convergencegraph2D"}](Fig34.eps) ![Convergence of the DCA-2.[]{data-label="convergencegraph2D"}](Ex34.eps) Smooth Approximation by Continuous DC Problems {#nest-smooth} ============================================== In this section we first employ and further develop Nesterov’s smoothing technique for the case of multifacility location problem . Then we enclose the family of DC mixed integer programs obtained in this way into a class of smooth DC problems of continuous optimization. The suggested procedures are efficiently justified by deriving numerical estimates expressed entirely via the given data of the original problem . We begin with recalling the following useful result on Nesterov’s smoothing related to the problem under consideration, which is taken from [@Nam2018 Proposition 2.2]. \[Prop3.1\] Given any $a\in\R^d$ and $\mu>0$, a Nesterov smoothing approximation of the function $f\colon\R^d\rightarrow\R$ defined by $$f(x):=\|x-a\|,\quad x\in\R^d,$$ admits the smooth DC representation $$f_\mu(x):=\dfrac{1}{2\mu}\|x-a\|^2-\dfrac{\mu}{2}\bigg[d\bigg(\dfrac{x-a}{\mu};\B\bigg)\bigg]^2.$$ Furthermore, we have the relationships $$\nabla f_\mu(x)=P\bigg(\dfrac{x-a}{\mu};\Bbb B\bigg)\;\;\mbox{\rm and }\; f_\mu(x)\le f(x)\le f_\mu(x)+\dfrac{\mu}{2},$$ where $\Bbb B\subset\R^d$ is the closed unit ball, and where $P$ stands for the Euclidean projection . Using Proposition \[Prop3.1\] allows us to approximate the objective function $\mathcal{F}$ in by a smooth DC function $\mathcal{F}_\mu$ as $\mu>0$ defined as follows: $$\begin{aligned} \begin{array}{ll} \mathcal{F}_\mu(\mathbf U,\mathbf V)&:=\disp\dfrac{1}{2\mu}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\|a_{j}-v_{i}\|^2-\disp\dfrac{\mu}{2}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\bigg[d\bigg(\dfrac{a_{j}-v_{i}}{\mu};\B\bigg)\bigg]^2 \\&=\mathcal{G}_\mu(\mathbf U,\mathbf V)-\mathcal{H}_\mu(\mathbf U,\mathbf V), \end{array}\end{aligned}$$ where $\mathcal{G}_\mu,\mathcal{H}_\mu\colon\R^{k\times n}\times\R^{k\times d}\to\R$ are given by $$\begin{aligned} \begin{array}{ll} &\mathcal{G}_\mu(\mathbf U,\mathbf V):=\disp\dfrac{1}{2\mu}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\|a_{j}-v_{i}\|^2,\\ &\mathcal{H}_\mu(\mathbf U,\mathbf V):=\disp\dfrac{\mu}{2}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\bigg[d\bigg(\dfrac{a_{j}-v_{i}}{\mu};\B\bigg)\bigg]^2 \nonumber. \end{array}\end{aligned}$$ This leads us to the construction of the following family of smooth approximations of the main problem defined by $$\begin{aligned} \label{discrete} \begin{array}{ll} &\mbox{\rm minimize }\;\mathcal{F}_\mu(\mathbf U,\mathbf V):=\mathcal{G}_\mu(\mathbf U,\mathbf V)-\mathcal{H}_\mu(\mathbf U,\mathbf V)\;\;\mbox{\rm as }\;\mu>0\\ &\mbox{\rm subject to }\;\mathbf U\in\mathcal{U}=\Delta^{n}\cap\{0,1\}^{k \times n}\;\;\mbox{\rm and }\;\mathbf V\in\R^{k\times d}, \end{array}\end{aligned}$$ where $\Delta^n$ is the the $n$th Cartesian degree of the $(k-1)$-simplex $\Delta:=\{y\in[0,1]^k\;|\;\sum_{i=1}^{k}y_{i}=1\}$, which is a subset of $\R^{k}$. Observe that for each $\mu>0$ problem is of [*discrete optimization*]{}, while our intention is to convert it to a family of problems of [*continuous optimization*]{} for which we are going to develop and implement a DCA-based algorithm in Section \[TheAlgorithm\]. The rest of this section is devoted to deriving two results, which justify such a reduction. The first theorem allows us to verify the [*existence*]{} of optimal solutions to the constrained optimization problems that appear in this procedure. It is required for having well-posedness of the algorithm construction. \[existence\] Let $(\Bar{\mathbf U},\Bar{\mathbf V})$ be an optimal solution to problem . Then for any $\mu>0$ we have $\Bar{\mathbf V}\in\mathcal{B}$, where $\mathcal{B}:=\prod_{i=1}^{k}B_{i}$ is the Cartesian product of the $k$ Euclidean balls $B_{i}$ centered at $0\in\R^d$ with radius $r:=\sqrt{\sum_{j=1}^{n}\|a_j\|^2}$ that contain the optimal centers $\bar v_i$ for each index $i=1,\ldots,k$. [**Proof.**]{} We can clearly rewrite the objective function in in the form $$\label{F} \mathcal{F}_\mu(\mathbf U,\mathbf V)=\dfrac{1}{2\mu}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}\|a_{j}-v_{i}\|^2-\dfrac{\mu}{2}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}\bigg[d\bigg(\dfrac{a_{j}-v_{i}}{\mu};\B\bigg)\bigg]^2$$ due to interchangeability between $u_{i,j}^2$ and $u_{i,j}$. Observe that $\mathcal{F}_\mu(\mathbf U,\mathbf V)$ is differentiable on $R^{k\times n}\times\R^{k\times d}$. Employing the classical Fermat rule in with respect to $V$ gives us $\nabla_{\mathbf V}\mathcal{F}_\mu(\Bar{\mathbf U},\Bar{\mathbf V})=0$. To calculate this partial gradient, we need some clarification for the second term in , which is differentiable as a whole while containing the nonsmooth distance function . The convexity of the distance function in the setting of allows us to apply the subdifferential calculation of convex analysis (see, e.g., [@bmn Theorem 2.39]) and to combine it with an appropriate chain rule to handle the composition in . Observe that the distance function square in is the composition of the nondecreasing convex function $\ph(t):=t^2$ on $[0, \infty)$ and the distance function to the ball $\B$. Thus the chain rule from [@bmn Corollary 2.62] is applicable. Thus, we can show that $d^2(\cdot; \B)$ is differentiable with $$\label{sdsub} \nabla d^2(x; \B)=2[x-P(x; \B)]\; \mbox{\rm for }x\in \R^d.$$ Using , we consider the following two cases: **Case 1:** $(a_{j}-\bar v_{i})/\mu\in\B$ for the fixed indices $i\in\{1,\ldots,k\}$ and $j\in\{1,\ldots,n\}$. Then $$\nabla d^2\bigg(\dfrac{a_{j}-\bar v_{i}}{\mu};\B\bigg)=\{0\},$$ which gives us $$\disp\frac{\partial\mathcal{F}_\mu}{\partial v_i}(\Bar{\mathbf U},\Bar{\mathbf V})=\dfrac{1}{\mu}\sum_{j=1}^{n}\bar u_{i,j}(\bar v_{i}-a_{j}),\quad i=1,\ldots,k,$$ for the corresponding partial derivatives of $\mathcal{F}_\mu$. [**Case 2:**]{} $(a_{j}-\bar v_{i})/\mu\notin\B$ for the fixed indices $i\in\{1,\ldots,k\}$ and $j\in\{1,\ldots,n\}$. In this case we have $$\begin{aligned} \disp\frac{\partial\mathcal{F}_\mu}{\partial v_i}(\Bar{\mathbf U},\Bar{\mathbf V})&=&\dfrac{1}{2\mu}\sum_{j=1}^{n}\bar u_{i,j}2(\bar v_{i}-a_{j})+\sum_{j=1}^{n}\bar u_{i,j}\bigg[\dfrac{a_{j}-\bar v_{i}}{\mu}-P\bigg(\dfrac{a_{j}-\bar v_{i}}{\mu};\B\bigg)\bigg]\\ &=&\dfrac{1}{\mu}\sum_{j=1}^{n}\bar u_{i,j}(\bar v_{i}-a_{j})+\disp\sum_{j=1}^{n} \bar u_{i,j}\bigg[\dfrac{a_{j}-\bar v_{i}}{\mu}-\bigg(\dfrac{a_{j}-\bar v_{i}}{\|a_{j}-\bar v_{i}\|}\bigg)\bigg]\\ &=&\disp\dfrac{1}{\|a_{j}-\bar v_{i}\|}\sum_{j=1}^{n}\bar u_{i,j}(\bar v_{i}-a_{j}).\end{aligned}$$ Thus in both cases above it follows from the stationary condition $\nabla_{\mathbf V}\mathcal{F}_\mu(\Bar{\mathbf U},\Bar{\mathbf V})=0$ that $$\begin{aligned} \bar v_{i}=\dfrac{\sum_{j=1}^{n}\bar u_{i,j}a_{j}}{\sum_{j=1}^{n}\bar u_{i,j}}\;\;\mbox{\rm for all }\;i=1,\ldots,k,\end{aligned}$$ since we have $\sum\limits_{j=1}^{n}\bar u_{i,j}>0$ due to the nonemptiness of the clusters. Then the classical Cauchy-Schwarz inequality leads us to the estimates $$\|\bar v_{i}\|^2 \le\dfrac{\bigg(\sum_{j=1}^{n}\bar u_{i,j}a_{j}\bigg)^2}{\bigg(\sum_{j=1}^{n}\bar u_{i,j}\bigg)^2}\le\sum_{j=1}^{n}\|a_{j}\|^2:=r^2,$$ which therefore verify all the conclusions of this theorem. $\h$ Our next step is to enclose each discrete optimization problem into the corresponding one of [*continuous optimization*]{}. For the reader’s convenience if no confusion arises, we keep the same notation ${\mathbf U}$ for [*all*]{} the $k\times n$-matrices without the discrete restrictions on their entries. Define now the function $\mathcal{P}\colon\R^{k\times n}\to\R$ by $$\mathcal{P}(\mathbf U):=\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}(1-u_{i,j})\;\;\mbox{\rm for all }\;{\mathbf U}\in\R^{k\times n}$$ and observe that this function is concave on $\R^{k\times n}$ with $\mathcal{P}(\mathbf U)\ge 0$ whenever $\mathbf U\in\Delta^n$. Furthermore, we have the representations $$\label{constr} \mathcal{U}=\big\{\mathbf U\in\Delta^n\;\big|\;\mathcal{P}(\mathbf U)=0\big\}=\big\{\mathbf U\in\Delta^n\;\big|\;\mathcal{P}(\mathbf U)\le 0\big\}$$ for the set of feasible $k\times n$-matrices $\mathcal{U}$ in the original problem . Employing further the standard [*penalty function*]{} method allows us to eliminate the most involved constraint on ${\mathbf U}$ in given by the function ${\mathcal P}$. Taking the penalty parameter $\alpha>0$ sufficiently large and using the smoothing parameter $\mu>0$ sufficiently small, consider the following family of continuous optimization problems: $$\begin{aligned} \label{discrete1} \begin{array}{ll} &\mbox{\rm minimize }\;\mathcal{F}_\mu(\mathbf U,\mathbf V)+\alpha\mathcal{P}(\mathbf U)=\mathcal{G}_\mu(\mathbf U,\mathbf V)-\mathcal{H}_\mu(\mathbf U,\mathbf V)+\alpha\mathcal{P}(\mathbf U)\\ &\mbox{\rm subject to }\;\mathbf U\in\Delta^n\;\;\mbox{\rm and }\;\mathbf V\in\mathcal{B}. \end{array}\end{aligned}$$ Observe that Theorem \[existence\] ensures the existence of feasible solutions to problem and hence optimal solutions to this problem by the Weierstrass theorem due to the continuity of the objective functions therein and the compactness of the constraints sets $\Delta^n$ and ${\mathcal B}$. Let us introduce yet another parameter $\rho>0$ ensuring a DC representation of the objective function in as follows: $$\begin{aligned} \begin{array}{ll} \mathcal{F}_\mu(\mathbf U,\mathbf V)+\alpha\mathcal{P}(\mathbf U)&=\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2-\bigg(\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2-\mathcal{F}_\mu(\mathbf U,\mathbf V)-\alpha\mathcal{P}(\mathbf U)\bigg)\\\nonumber &=\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^{2}-\bigg(\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^{2}-\mathcal{G}_\mu(\mathbf U,\mathbf V)+\mathcal{H}_\mu(\mathbf U,\mathbf V)-\alpha\mathcal{P}(\mathbf U)\bigg)\\ &=:\mathcal{G}(\mathbf U,\mathbf V)-\mathcal{H}(\mathbf U,\mathbf V), \end{array}\end{aligned}$$ where the function $\mathcal{G}(\mathbf U,\mathbf V):=\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2$ is obviously convex, and $$\mathcal{H}(\mathbf U,\mathbf V):=\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2-\mathcal{G}_\mu(\mathbf U,\mathbf V)+\mathcal{H}_\mu(\mathbf U,\mathbf V)-\alpha\mathcal{P}(\mathbf U).$$ Since $\mathcal{H}_\mu(\mathbf U,\mathbf V)-\alpha\mathcal{P}(\mathbf U)$ is also convex as $\alpha>0$, we are going to show that for any given number $\mu>0$ it is possible to determine the values of the parameter $\rho>0$ such that the function $\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2-\mathcal{G}_\mu(\mathbf U,\mathbf V)$ is convex under an appropriate choice of $\rho$. This would yield the convexity of $\mathcal{H}(\mathbf U,\mathbf V)$ and therefore would justify a desired representation of the objective function in . The following result gives us a precise meaning of this statement, which therefore verify the required reduction of to [*DC continuous optimization*]{}. \[convexity\] The function $$\begin{aligned} \label{g1} \mathcal{G}_1(\mathbf U,\mathbf V):=\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2-\mathcal{G}_\mu(\mathbf U,\mathbf V)\end{aligned}$$ is convex on $\Delta^n\times\mathcal{B}$ provided that $$\label{rho} \rho\ge\dfrac{n}{2\mu}\bigg[\bigg(1+\dfrac{1}{n}{\xi}^2\bigg)+\sqrt{\bigg(1+\dfrac{1}{n}{\xi}^2\bigg)^2+\dfrac{12}{n}{\xi}^2}\bigg],$$ where $\xi:=r+\underset{1\le j\le n}{\mbox{max}}\|a_{j}\|$ and $r:=\sqrt{\sum_{j=1}^{n}\|a_j\|^2}$. [**Proof.**]{} Consider the function $\mathcal{G}_1(\mathbf U,\mathbf V)$ defined in for all $(\mathbf U,\mathbf V)\in\Delta^n\times\mathcal{B}$ and deduce by elementary transformations directly from its construction that $$\begin{aligned} \mathcal{G}_1(\mathbf U,\mathbf V)&=&\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2-\mathcal{G}_\mu(\mathbf U,\mathbf V)\nonumber\\ &=&\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2 - \dfrac{1}{2\mu}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\|a_{j}-v_{i}\|^2\nonumber\\ &=&\dfrac{\rho}{2}\|\mathbf U\|^2+\dfrac{\rho}{2}\|\mathbf V\|^2-\dfrac{1}{2\mu}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\|a_{j}-v_{i}\|^2\nonumber\\ &=&\dfrac{\rho}{2}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2+\dfrac{\rho}{2n}\sum_{i=1}^{k}\sum_{j=1}^{n}\|v_{i}\|^2-\dfrac{1}{2\mu}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\|a_{j}-v_{i}\|^2\nonumber\\ &=&\sum_{i=1}^{k}\sum_{j=1}^{n}\dfrac{\rho}{2}u_{i,j}^2+\dfrac{\rho}{2n}\|a_{j}-v_{i}\|^2+\dfrac{\rho}{n}\langle a_j,v_i\rangle-\dfrac{\rho}{2n}\|a_{j}\|^2-\dfrac{1}{2\mu}u_{i,j}^{2}\|a_{j}-v_{i}\|^2.\end{aligned}$$ Next we define the functions $\gamma_{i,j}\colon\R\times\R^d\rightarrow\R$ for all $i=1,\ldots,k$ and $j=1,\ldots,n$ by $$\label{gamma} \gamma_{i,j}(u_{i,j},v_{i}):=\dfrac{\rho}{2}u_{i,j}^2+\dfrac{\rho}{2n}\|a_{j}-v_{i}\|^2-\dfrac{1}{2\mu}u_{i,j}^2\|a_{j}-v_{i}\|^2$$ and show that each of these functions is convex on the set $\{u_{i,j}\in[0,1],\,\|v_{i}\|\le r\}$, where $r>0$ is taken from Theorem \[existence\]. To proceed, consider the Hessian matrix of each function in given by $$J_{\gamma_{i,j}}(u_{i,j},v_{i}):= \begin{bmatrix} \rho-\dfrac{1}{\mu}\|a_{j}-v_{i}\|^2&&-\dfrac{2}{\mu}u_{i,j}(v_{i}-a_{j})\\ -\dfrac{2}{\mu}u_{i,j}(v_{i}-a_{j})&&\dfrac{\rho}{n}-\dfrac{1}{\mu}u^2_{i,j} \end{bmatrix}$$ and calculate its determinant $\det(J_{\gamma_{i,j}}(u_{i,j},v_{i}))$ by $$\begin{aligned} \label{det} \det(J_{\gamma_{i,j}}(u_{i,j},v_{i})) &:=& \bigg(\rho-\dfrac{1}{\mu}\|a_{j}-v_{i}\|^2\bigg)\bigg(\dfrac{\rho}{n}-\dfrac{1}{\mu}u^2_{i,j}\bigg)-\dfrac{4}{\mu^{2}}u^2_{i,j}(v_{i}-a_{j})^T(v_{i}-a_{j})\nonumber\\ &=&\dfrac{\rho^{2}}{n}-\rho\bigg(\dfrac{u^2_{i,j}}{\mu}+\dfrac{1}{n\mu }\|v_{i}-a_{j}\|^{2}\bigg)-\dfrac{3u^2_{i,j}}{\mu^2}\|v_{i}-a_{j}\|^2.\end{aligned}$$ It follows from the well-known second-order characterization of the convexity that the function $\gamma_{i,j}(u_{i,j},v_{i})$ is convex on $\{u_{i,j}\in[0,1],\,\|v_{i}\|\le r\}$ if $\det(J_{\gamma_{i,j}}(u_{i,j},v_{i}))\ge 0$. Using [@Tao Theorem 1] gives us the estimate $$\begin{aligned} \det\big(J_{\gamma_{i,j}}(u_{i,j},v_{i})\big)&\ge&\dfrac{\rho^2}{n}-\rho\bigg(\dfrac{1}{\mu}+\dfrac{1}{n\mu}\|v_{i}-a_{j}\|^2\bigg)-\dfrac{3}{\mu^2}\|v_{i}-a_{j}\|^2.\end{aligned}$$ Then we get from the construction of ${\mathcal B}$ in Theorem \[existence\] that $0<\|v_{i}-a_{j}\|\le\|v_{i}\|+\|a_{j}\|\le r+\underset{1\le j\le n}{\mbox{max}}\|a_{j}\|=:\xi$, and therefore $$\label{quadratic} \det\big(J_{\gamma_{i,j}}(u_{i,j},v_{i})\big)\ge\dfrac{\rho^2}{n}-\dfrac{\rho}{\mu}\bigg(1+\dfrac{1}{n}\xi^2\bigg)-\dfrac{3}{\mu^2}\xi^2.$$ It allows us to deduce from the aforementioned condition for the convexity of $\gamma_{i,j}(u_{i,j},v_{i})$ that we do have this convexity if $\rho$ satisfies the estimate . $\h$ Design and Implementation of the Solution Algorithm ===================================================  \[TheAlgorithm\] Based on the developments presented in the previous sections and using the established smooth DC structure of problem with the subsequent $\rho-$parameterization of the objective function therein as ${\mathcal G}({\mathbf U},{\mathbf V})-{\mathcal H}({\mathbf U},{\mathbf V})$, we are now ready to propose and implement a new algorithm for solving this problem involving both DCA-2 and Nesterov’s smoothing. To proceed, let us present the problem under consideration in the equivalent [*unconstrained*]{} format by using the infinite penalty via the indicator function: $$\begin{aligned} \label{unconstprob} \begin{array}{ll} \mbox{\rm minimize }&\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2-\mathcal{H}(\mathbf U,\mathbf V)+\delta_{\Delta\times\mathcal{B}}(\mathbf U,\mathbf V)\\ &\mbox{\rm subject to }\;(\mathbf U,\mathbf V)\in\R^{k\times n}\times\R^{k\times d}, \end{array}\end{aligned}$$ where $\mathcal{B}$, $\Delta$, and $\rho$ are taken from Section \[nest-smooth\]. We first explicitly compute the gradient of the convex function $\mathcal{H}(\mathbf U,\mathbf V)$ in . Denoting $$\begin{aligned} \begin{array}{ll} [\mathcal{Y},\mathcal{Z}]&:=\nabla \mathcal{H}(\mathbf U,\mathbf V)=\nabla\bigg(\dfrac{\rho}{2}\|(\mathbf U,\mathbf V)\|^2-\disp\dfrac{1}{2\mu}\disp\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\|a_{j}-v_{i}\|^2\\ &+\disp\dfrac{\mu}{2}\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}^2\bigg[d\bigg(\dfrac{a_{j}-v_{i}}{\mu};\B\bigg)\bigg]^2-\alpha\disp\sum_{i=1}^{k}\sum_{j=1}^{n}u_{i,j}(1-u_{i,j})\bigg), \end{array}\end{aligned}$$ we have $\mathcal{Y}=\nabla\mathcal{H}_{\mathbf U}(\mathbf U,\mathbf V)$ and $\mathcal{Z}=\nabla\mathcal{H}_{\mathbf V}(\mathbf U,\mathbf V)$. Thus for each $i=1,\ldots,k$ and $j=1,\ldots,n$ the $(j,i)-$entry of the matrix $\mathcal{Y}$ and the $i$th row of the matrix $\mathcal{Z}$ are $$\begin{aligned} \begin{array}{ll} \mathcal{Y}_{j,i}&:=\rho u_{i,j}-\dfrac{u_{i,j}}{\mu}\|a_{j}-v_{i}\|^2+\mu u_{i,j}\bigg[d\bigg(\dfrac{a_{j}-v_{i}}{\mu};\B\bigg)\bigg]^2+2\alpha u_{i,j}-\alpha,\\ \mathcal{Z}_{i}&:=\disp\rho v_{i}-\dfrac{1}{\mu}\sum_{j=1}^{n}u_{i,j}^{2}(v_{i}-a_{j})-\sum_{j=1}^{n}u_{i,j}^2\bigg[\dfrac{a_{j}-v_{i}}{\mu}-P\bigg(\dfrac{a_{j}-v_{i}}{\mu};\B\bigg)\bigg], \end{array}\end{aligned}$$ respectively. Let us now describe the proposed algorithm for solving the DC program and hence the original problem of multifacility location. The symbols $\mathcal{Y}^{l-1}_{[j,:]}$ and $\mathcal{Z}^{l-1}_{i}$ in this description represents the $j$th row of the matrix $\mathcal{Y}$ and the $i$th row of the matrix $\mathcal{Z}$ at the $l$th iteration, respectively. Accordingly we use the symbols $\mathbf U^{l}_{[:,j]}$ and $\mathbf V^{l}_{i}$. Recall also that the Frobenius norm of the matrices in this algorithm is defined in . [**Algorithm 3: Solving Multifacility Location Problems**]{}. [| l |]{} : $\mathbf X$ (the dataset), $\mathbf V^{0}$ (initial centers), ClusterNum (number of clusters), $\mu>0$,\ $\beta$ (scaling parameter) $>0$, $N\in\N$\ [INITIALIZATION]{}: $\mathbf U^{0}$, $\epsilon >0$, $\mu_{f}$ (minimum threshold for $\mu$) $>0$, $\alpha>0$, $\rho>0$,\ tol (tolerance parameter) $=1$\ [**while**]{} [tol $>\epsilon$ $\mu >\mu_f$]{}\ [**for**]{} $l=1,2,\ldots,N$\ For $1\le i\le k$ and $1\le j\le n$ compute\ $\mathcal{Y}^{l-1}_{j,i}:=\rho u^{l-1}_{i,j}-\dfrac{u^{l-1}_{i,j}}{\mu}\left\|a_{j}-v^{l-1}_{i}\right\|^{2}+\mu u^{l-1}_{i,j}\bigg[d\bigg(\dfrac{a_{j}-v^{l-1}_{i}}{\mu};\B\bigg)\bigg]^{2}+2\alpha u^{l-1}_{i,j}-\alpha,$\ \ $\mathcal{Z}^{l-1}_{i}:=\rho v^{l-1}_{i}-\dfrac{1}{\mu}\disp\sum_{j=1}^{n}(u^{l-1}_{i,j})^{2}(v^{l-1}_{i}-a_{j})-\sum_{j=1}^{n}(u^{l-1}_{i,j})^{2}\bigg[\dfrac{a_{j}-v^{l-1}_{i}}{\mu}-P\bigg(\dfrac{a_{j}-v^{l-1}_{i}}{\mu};\B\bigg)\bigg].$\ For $1\le i\le k$ and $1\le j\le n$ compute\ $\mathbf U^{l}_{[:,j]}:=P\bigg(\dfrac{\mathcal{Y}^{l-1}_{[j,:]}}{\rho};\Delta\bigg)$,\ $\mathbf V^{l}_{i}:=P\bigg(\dfrac{\mathcal{Z}^{l-1}_{i}}{\rho};B_{i}\bigg)= \begin{cases} \dfrac{\mathcal{Z}^{l-1}_{i}}{\rho} &\mbox{if }\;|\mathcal{Z}^{l-1}_{i}||\le\rho r,\\ \dfrac{r\mathcal{Z}^{l-1}_{i}}{||\mathcal{Z}^{l-1}_{i}||}&\mbox{if }\;||\mathcal{Z}^{l-1}_{i}||>\rho r. \end{cases} $\ [**end for**]{}\ [UPDATE:]{}\ $\text{tol}:=\left\|[\mathbf U^{l},\mathbf V^{l}]-[\mathbf U^{l-1},\mathbf V^{l-1}]\right\|_{F}$\ $\mu:=\beta\mu$.\ [**end while**]{}\ [OUTPUT]{}: $[\mathbf U^{N},\mathbf V^{N}]$.\ Next we employ Algorithm 3 to solve several multifacility location problems of some practical meaning. By trial and error we verify that the values chosen for $\mu$ determine the performance of the algorithm for each data set. It can be seen that very small values of the smoothing parameter $\mu$ may prevent the algorithm from clustering, and thus we gradually decrease these values. This is done via multiplying $\mu$ by some number $0<\beta<1$ and stopping when $\mu<\mu_f$. Note also that in the implementation of our algorithm we use the standard approach of choosing $\mathbf U^{0}$ by computing the distance between the point in question and each group center $\mathbf V^{0}$ and then by classifying this point to be in the group whose center is the closest to it by assigning the value of $1$, while otherwise we assign the value of $0$. Let us now present several numerical examples, where we compute the optimal centers by using Algorithm 3 via MATLAB calculations. Fix in what follows the values of $\mu=0.5$, $\beta=0.85$, $\epsilon=10^{-6}$, $\mu_{f} =10^{-6}$, $\alpha=30$, and $\rho=30$ unless otherwise stated. The objective function is the [*total distance*]{} from the centers to the assigned data point. Note that this choice of the objective function seems to be natural from practical aspects in, e.g., airline and other transportation industries, where the goal is to reach the destination via the best possible route available. This reflects minimizing the transportation cost. In the following examples we implement the standard $k$-means algorithm in MATLAB using the in-built function kmeans(). \[Eg4.1\] Let us consider a data set with $14$ entries in $\R^2$ given by $$\mathbf X:= \begin{bmatrix} 0 & 2 & 7 & 2 & 3 & 6 & 5 & 8 & 8 & 9 & 1 &7 & 0 &0 \\ 3 & 2 & 1 & 4 & 3 & 2 & 3 & 1 & 3 & 2 & 1 & 4 & 4&1 \end{bmatrix}^{T}$$ with the initial data defined by $$\begin{aligned} \begin{array}{ll} &\mathbf V^0\ := \left[ \begin{array}{lr} 7.1429 & 2.2857 \\ 1.1429 & 2.5714 \end{array}\right]\mbox{is obtained from the $k$-means algorithm; see Table~\ref{Table1}, }\\ &\mbox{ClusterNum}:=2. \end{array}\end{aligned}$$ Employing Algorithm 3, we obtain the optimal centers as depicted in Table \[Table1\] and Figure \[Figure 4.1\]. [l l l]{} **Method** & **Optimal Center ($\mathbf V^{N}$)** & **Cost Function**\ $k$-means & $\left[ \begin{array}{lr} 7.1429 & 2.2857 \\ 1.1429 & 2.5714 \end{array}\right]$ & 22.1637\ \ Algorithm 3 & $\left[ \begin{array}{lr} 7.2220 & 2.1802 \\ 1.1886 & 2.5069 \end{array}\right]$ & 22.1352\ \[Table1\] ![*MFLP with $14$ demand points and $2$ centers.*[]{data-label="Figure 4.1"}](Ex5552.eps) Table \[Table1\] shows that the proposed Algorithm 3 is marginally better for the given data in comparison to the classical $k$-means approach in terms of the objective function. \[balltests\] In this example we test our algorithm on a dataset $\mathbf X$ containing 10 distinct points on each boundary of 4 balls of radius $r=0.3$ centered at (2,2), (4,2), (4,4), and (2,4). We generate the points as follows: $$\Big\{C_i+{r_i}\Big(\mbox{cos}\Big(\dfrac{j\pi}{5}\Big),\mbox{sin}\Big(\dfrac{j\pi}{5}\Big)\Big)\;\Big|\;\;i=1,\ldots,4;\;j=1,\ldots,10\Big\},$$ where $C_i$ and $r_i$ are the center and radius of each ball respectively; see [@NamGiles]. Typically the centroids are the centers of the balls. Choosing a random point from the boundary of each ball for the initial centers, this algorithm converges to the optimal solution $$\mathbf V^{N}:=\begin{bmatrix} 2.0000&2.0000\\ 4.0000&2.0000\\ 2.0000&4.0000\\ 4.0000&4.0000 \end{bmatrix}.$$ Its visualization is shown in Figure \[Figure 6.2\]. ![*MFLP with $40$ demand points and $4$ centers.*[]{data-label="Figure 6.2"}](Ex53.eps) Note that a drawback in employing the random approach to choose the initial cluster $\mathbf V^{0}$ in Example 5.2 is the need of having prior knowledge about the data. Typically it may not be plausible to extract such an information from large unpredictable real life datasets. In the next example we choose the initial cluster by the process of [*random selection*]{} and see its effect on the optimal centers. Then the results obtained in this way by Algorithm 3 are compared with those computed by the $k$-means approach. \[Eg4.2\] Let $\mathbf X$ be $200$ standard normally distributed random datapoints in $\R^2$, and let the initial data be given by $$\begin{aligned} \begin{array}{ll} &\mathbf V^0:=\mbox{randomly permuting and selecting $2$ rows of $\mathbf X$},\\ &\mbox{ClusterNum}:=2. \end{array}\end{aligned}$$ We obtain the optimal centers as outlined in Table \[TableAlg\]. [l l l]{} **Method** & **Optimal Center ($\mathbf V^N$)** & **Cost Function**\ $k$-means & $\left[ \begin{array}{lr} 2.1016 & 1.2320 \\ -1.3060 & -1.0047 \end{array}\right]$ & 403.3966\ \ Algorithm 3 & $\left[ \begin{array}{lr} 1.4902 & 0.7406\\ -1.3464 & -1.0716 \end{array}\right]$ & 401.7506\ \[TableAlg\] Observe from Table \[TableAlg\] that the proposed Algorithm 3 is better for the given data in comparison to the standard $k$-means approach. In addition, our approach gives a better approximation for the optimal center as shown in Figure \[Figure 4.2\]. ![*MFLP with $200$ demand points and $2$ centers.*[]{data-label="Figure 4.2"}](Ex557.eps) Note that a real-life data may not be as efficiently clustered as in Example \[Eg4.2\]. Thus a suitable selection of the initial cluster $\mathbf V^0$ is vital for the convergence of the DCA based algorithms. In the next Example \[Eg4.3\] we select $\mathbf V^0$ in Algorithm 3 by using the standard $k$-means method. The results achieved by our Algorithm 3 are again compared with those obtained by using the $k$-means approach. \[Eg4.3\] Consider the dataset $\mathbf X$ consisting of the latitudes and longitudes of $50$ most populous cities in the USA[^5] with $$\begin{aligned} \begin{array}{ll} &\mathbf V^0\ := \left[ \begin{array}{lr} -80.9222 & 37.9882\\ -97.8273 & 35.3241 \\ -118.3121 & 36.9535 \end{array}\right] \mbox{is obtained from the $k$-means algorithm; see Table~\ref{IntCluster}},\\ &\mbox{ClusterNum}:=3. \end{array}\end{aligned}$$ By using Algorithm 3 we obtain the following optimal centers as given in Table \[IntCluster\]. [l l l ]{} **Method** & **Optimal Center ($\mathbf V^N$)** & **Cost Function**\ $k$-means & $\left[ \begin{array}{lr} -80.9222 & 37.9882\\ -97.8273 & 35.3241 \\ -118.3121 & 36.9535 \end{array}\right]$& 288.8348\ \ Algorithm 3 (combined with $k$-means)& $\left[ \begin{array}{lr} -81.0970 & 38.3092\\ -97.4138 & 35.3383 \\ -119.3112 & 36.5410 \end{array}\right]$& 286.6523\ \[IntCluster\] We see that Algorithm 3 (combined with $k$-means) in which the initial cluster $\mathbf V^0$ is selected by using $k$-means method performs better in comparison to the standard $k$-means approach (Table \[IntCluster\]). Moreover, it gives us optimal centers as depicted in Figure \[Figure 4.3\]. ![*MFLP with $50$ demand points and $3$ centers.*[]{data-label="Figure 4.3"}](Ex558.eps) In the next example we efficiently solve yet another multifacility location problem by using Algorithm 3. Consider the dataset $\mathbf X$ in $\R^2$ that consists of the latitudes and longitudes of $988$ US cities [@citydata] with $$\begin{aligned} \begin{array}{ll} &\mathbf V^0\ := \left[ \begin{array}{lr} -89.6747 & 41.1726\\ -88.4834 & 30.2475 \\ -118.4471 & 35.2843 \\ -75.7890 & 40.0329 \\ -114.1897 & 43.0798 \end{array}\right] \mbox{is obtained from the $k$-means algorithm},\\ &\mbox{ClusterNum}:=5. \end{array} \end{aligned}$$ The optimal centers illustrated in Figure \[Figure 4.5\] are given by $$\mathbf V^N:=\left[ \begin{array}{lr} -88.2248 & 40.7241\\ -88.0154 & 30.7041 \\ -120.0735 & 33.1941 \\ -74.5328 & 38.6996\\ -113.2341 & 42.3969 \end{array}\right].$$ The total transportation cost in this problem is 5089.5150. ![*MFLP with $988$ demand points and $5$ centers.*[]{data-label="Figure 4.5"}](Ex559.eps) In the last example presented in this section we efficiently solve a higher dimensional multifacility location problem by using Algorithm 3 and compare it’s value of the cost function with the standard $k$-means algorithm. Let $\mathbf X$ in $\R^{13}$ be the *wine dataset* from the UCI Machine Learning Repositiory [@wine] consisting of $178$ demand points. We apply Algorithm 3 with $$\begin{aligned} \begin{array}{ll} &\mathbf V^0 \ \mbox{is obtained from the $k$-means algorithm},\\ &\mbox{ClusterNum}:=3. \end{array} \end{aligned}$$ The total costs using Algorithm 3 and the $k$-means algorithm are obtained in Table \[CostFunc\] showing that former algorithm is better than the latter. [ll]{} **Method** & **Cost Function**\ $k$-means & 16556\ \ Algorithm 3 (combined with $k$-means)& 16460\ \[CostFunc\] Concluding Remarks ==================  \[Conclusion\] In this paper we develop a new algorithm to solve a class of multifacility location problems. Its implementation exhibits a better approximation compared to the classical $k$-means approach. This is demonstrated by a series of examples dealing with two-dimensional problems with nonnegative weights. Thus the verification and implementation of the proposed algorithm for real-life multifacility location problems in higher dimensions with arbitrary weights is a central direction of our future work. In addition, refining the initial cluster selection and the stopping criterion is an important area to explore. [99]{} L.T.H. An, M.T. Belghiti, and P.D. Tao, [*A new efficient algorithm based on DC programming and DCA for clustering*]{}, J. Glob. Optim. 37 (2007), pp. 593–608. N.T. An and N.M. Nam, [*Convergence analysis of a proximal point algorithm for minimizing differences of functions*]{}, Optim. 66 (2017), pp. 129–147. L.T.H. An, H.V. Ngai, and P.D. Tao, [*Convergence analysis of difference-of-convex algorithm with subanalytic data*]{}, J. Optim. Theory Appl. 179 (2018), pp. 103–126. L.T.H. An, L.H. Minh, and P.D. Tao, [*New and efficient DCA based algorithms for minimum sum-of-squares clustering*]{}, Pattern Recogn. 47 (2014), pp. 388–401. L.T.H. An and P.D. Tao, [*Minimum sum-of-squares clustering by DC programming and DCA*]{}, Int. Conf. Intell. Comp. (2009), pp. 327–340. J. Brimberg, [*The Fermat-Weber location problem revisited*]{}, Math. Program. 71 (1995), pp. 71–76. Z. Drezner, [*On the convergence of the generalized Weiszfeld algorithm*]{}, Ann. Oper. Res. 167 (2009), pp.  327–336. T. Jahn, Y.S. Kupitz, H. Martini, and C. Richter, [*Minsum location extended to gauges and to convex sets*]{}, J. Optim. Theory Appl. 166 (2015), pp. 711–746. J.-B. Hiriart-Urruty and C. Lemaréchal, [*Fundamentals of Convex Analysis*]{}, Springer, Berlin, 2001. H.W. Kuhn, [*A note on Fermat-Torricelli problem*]{}, Math. Program. 4 (1973), pp. 98–107. Y.S. Kupitz and H. Martini, [*Geometric aspects of the generalized Fermat-Torricelli problem*]{}, Bolyai Soc. Math. Stud. 6 (1997), pp. 55–128. H. Martini, K.J. Swanepoel, and G. Weiss, [*The Fermat-Torricelli problem in normed planes and spaces*]{}, J. Optim. Theory Appl. 115 (2002), pp. 283–314. B.S. Mordukhovich, [*Variational Analysis and Applications*]{}, Springer, Cham, Switzerland, 2018. B.S. Mordukhovich and N.M. Nam, [*Applications of variational analysis to a generalized Fermat-Torricelli problem*]{}, J. Optim. Theory Appl. 148 (2011), pp. 431–454. B.S. Mordukhovich and N.M. Nam, [*An Easy Path to Convex Analysis and Applications*]{}, Morgan & Claypool Publishers, San Rafael, 2014. N.M. Nam, N.T. An, R.B. Rector, and J. Sun, [*Nonsmooth algorithms and Nesterov’s smoothing technique for generalized Fermat-Torricelli problems*]{}, SIAM J. Optim. 24 (2014), pp. 1815–1839. N.M. Nam, W. Geremew, S. Reynolds, and T. Tran, [*Nesterov’s smoothing technique and minimizing differences of convex functions for hierarchical clustering*]{}, Optim. Lett. 12 (2018), pp. 455–473. N.M. Nam and N. Hoang, [*A generalized Sylvester problem and a generalized Fermat-Torricelli problem*]{}, J. Convex Anal. 20 (2013), pp. 669–687. N.M. Nam, R.B. Rector, and D. Giles, [*Minimizing differences of convex functions with applications to facility location and clustering*]{}, J. Optim. Theory Appl. 173 (2017), pp. 255–278. Yu. Nesterov, [*A method for unconstrained convex minimization problem with the rate of convergence $O(1/k^2)$*]{}, Soviet Math. Dokl. 269 (1983), pp. 543–-547. Yu. Nesterov, [*Smooth minimization of nonsmooth functions*]{}, Math. Program. 103 (2005), pp. 127–152. Yu. Nesterov, [*Lectures on Convex Optimization*]{}, 2nd edition, Springer, Cham, Switzerland, 2018. R.T. Rockafellar, [*Convex Analysis*]{}, Princeton University Press, Princeton, NJ, 1970. P.D. Tao and L.T.H. An, [*Convex analysis approach to D.C. programming: theory, algorithms and applications*]{}, Acta Math. Vietnam. 22 (1997), pp. 289–355. P.D. Tao and L.T.H. An, [*A D.C. optimization algorithm for solving the trust-region subproblem*]{}, [*SIAM J. Optim.*]{} 8 (1998), pp. 476–505. United States Cities Database, [*Simple Maps: Geographic Data Products*]{}, 2017, <http://simplemaps.com/data/us-cities>. E. Weiszfeld, [*Sur le point pour lequel la somme des distances de $n$ points donnés est minimum*]{}, T$\hat{\mbox{\rm o}}$hoku Math. J. 43 (1937), pp. 355–386. UCI Machine Learning Repository, <https://archive.ics.uci.edu/ml/datasets/wine>. [^1]: Department of Mathematics, Wayne State University, Detroit, Michigan 48202, USA ([email protected] ). Research of this author was partly supported by the USA National Science Foundation under grant DMS-1808978 and by the USA Air Force Office of Scientific Research grant \#15RT04. [^2]: Department of Mathematics, Wayne State University, Detroit, Michigan 48202, USA ([email protected]). Research of this author was partly supported by the USA National Science Foundation under grants DMS-1512846 and DMS-1808978, by the USA Air Force Office of Scientific Research grant \#15RT04, and by Australian Research Council under grant DP-190100555. [^3]: Fariborz Maseeh Department of Mathematics and Statistics, Portland State University, Portland, OR 97207, USA ([email protected]). Research of this author was partly supported by the USA National Science Foundation under grant DMS-1716057. [^4]: Fariborz Maseeh Department of Mathematics and Statistics, Portland State University, Portland, OR 97207, USA ([email protected]). [^5]: Available at https://en.wikipedia.org/wiki/List of United States cities by population

ArXiv

**Abstract** High-energy cosmic neutrinos can reveal new fundamental particles and interactions, probing energy and distance scales far exceeding those accessible in the laboratory. This white paper describes the outstanding particle physics questions that high-energy cosmic neutrinos can address in the coming decade. A companion white paper discusses how the observation of cosmic neutrinos can address open questions in astrophysics. Tests of fundamental physics using high-energy cosmic neutrinos will be enabled by detailed measurements of their energy spectrum, arrival directions, flavor composition, and timing.\ **Endorsers** High-Energy Cosmic Neutrinos {#high-energy-cosmic-neutrinos .unnumbered} ---------------------------- What are the fundamental particles and interactions of Nature? High-energy cosmic neutrinos are uniquely poised to explore them in an uncharted and otherwise unreachable energy and distance regime. They allow us to explore the cosmic and energy frontiers of particle physics, complementing current and future colliders that will explore the energy and intensity frontiers. Despite the spectacular success of the Standard Model (SM) of particle physics, we know that it must be extended to account for at least the existence of neutrino mass, dark matter, and dark energy. A common feature of many theories beyond the Standard Model (BSM) is that their effects are more clearly apparent the higher the energy of the process, where new particles, interactions, and symmetries, undetectable at lower energies, could make themselves evident. Yet, particle colliders have failed to find clear evidence of BSM physics up to TeV energies, the highest reachable in the lab. To access particle interactions beyond the TeV scale, we must use particle beams made by natural cosmic accelerators. They produce the most energetic neutrinos, photons, and charged particles known, with energies orders of magnitude higher than in man-made colliders. Cosmic neutrinos are especially fitting probes of fundamental physics beyond the TeV scale, as shown in Fig. \[fig:scales\]. First, cosmic neutrinos reach higher energies than neutrinos made in the Sun, supernovae, the atmosphere of Earth, particle accelerators, and nuclear reactors. Further, they reach Earth with energies higher than that of gamma rays and likely as high as ultra-high-energy (UHE) cosmic rays. Second, because most cosmic neutrinos come from extragalactic sources located at cosmological distances, even tiny BSM effects could accumulate up to observable levels as neutrinos travel to Earth, having crossed essentially the observable Universe. And, third, because the propagation of neutrinos from the sources to the detectors is well understood and predicted by the SM, BSM effects could be more easily spotted than in charged particles. Tests of fundamental physics using cosmic neutrinos are possible in spite of astrophysical and cosmological uncertainties. Yet this endeavor is not without challenges: the neutrino detection cross section is tiny and cosmic neutrino fluxes are expected to fall rapidly with neutrino energy. Nevertheless, we show below that these obstacles are either surmountable or can be planned for. ![\[fig:scales\]Tests of fundamental physics accessible with neutrinos of different energies.](scales_full.pdf){width="96.00000%"} Open Questions: What Can High-Energy Cosmic Neutrinos Test? {#open-questions-what-can-high-energy-cosmic-neutrinos-test .unnumbered} ----------------------------------------------------------- Figure \[fig:scales\] shows the wide breadth of important open questions in fundamental physics that cosmic neutrinos can address [@Ahlers:2018mkf; @Anchordoqui:2005is; @Marfatia:2015hva]. They complement questions tackled by neutrinos of lower energies. Cosmic neutrinos span a wide range in energy. In the TeV–PeV range, astrophysical neutrinos are regularly detected by IceCube [@Aartsen:2013bka; @Razzaque:2013uoa; @Aartsen:2013jdh; @Aartsen:2014gkd; @Aartsen:2015rwa; @Aartsen:2016xlq] from what are likely mainly extragalactic sources [@Ahlers:2013xia; @Anchordoqui:2013dnh; @Ahlers:2015moa; @Denton:2017csz; @Aartsen:2017ujz; @IceCube:2018dnn; @IceCube:2018cha]. At the EeV scale, cosmogenic neutrinos, produced by UHE cosmic rays interacting with photon backgrounds through the GZK effect [@Greisen:1966jv; @Zatsepin:1966jv], are predicted but have not yet been observed [@Zas:2017xdj; @Aartsen:2018vtx; @Gorham:2019guw]. See Ref. [@Ackermann:2019ows] for a discussion of astrophysics enabled by observations of cosmic neutrinos. **How do neutrino cross sections behave at high energies?** The neutrino-nucleon cross section in the TeV–PeV range was measured for the first time using astrophysical and atmospheric neutrinos [@Aartsen:2017kpd; @Bustamante:2017xuy; @Aartsen:2018vez], extending [@Berezinsky:1974kz; @Hooper:2002yq; @Hussain:2006wg; @Borriello:2007cs; @Hussain:2007ba] measurements that used GeV neutrinos from accelerators [@Conrad:1997ne; @Formaggio:2013kya; @Tanabashi:2018oca]. Fig. \[fig:cross\_section\] shows that the measurements agree with high-precision SM predictions [@CooperSarkar:2011pa]. Future measurements in the EeV range would probe BSM modifications of the cross section at center-of-momentum energies of 100 TeV [@Kusenko:2001gj; @AlvarezMuniz:2001mk; @Anchordoqui:2001cg; @Cornet:2001gy; @Kowalski:2002gb; @AlvarezMuniz:2002ga; @Anchordoqui:2005pn; @Marfatia:2015hva; @Ellis:2016dgb; @Anchordoqui:2018qom] and test the structure of nucleons [@Henley:2005ms; @Armesto:2007tg; @Enberg:2008te; @Illarionov:2011wc; @Bhattacharya:2015jpa; @Garzelli:2015psa; @Halzen:2016pwl; @Halzen:2016thi; @Bhattacharya:2016jce; @Benzke:2017yjn; @Giannini:2018utr] more deeply than colliders [@Anchordoqui:2006ta; @Bertone:2018dse]. **How do flavors mix at high energies?** Experiments with neutrinos of up to TeV energies have confirmed that the different neutrino flavors, $\nu_e$, $\nu_\mu$, and $\nu_\tau$, mix and oscillate into each other as they propagate [@Tanabashi:2018oca]. Figure \[fig:flavor\_ratios\] shows that, if high-energy cosmic neutrinos en route to Earth oscillate as expected, the predicted allowed region of the ratios of each flavor to the total flux is small, even after accounting for uncertainties in the parameters that drive the oscillations and in the neutrino production process [@Bustamante:2015waa]. However, at these energies and over cosmological propagation baselines [@Learned:1994wg], mixing is untested; BSM effects could affect oscillations, vastly expanding the allowed region of flavor ratios and making them sensitive probes of BSM [@Beacom:2003nh; @Pakvasa:2007dc; @Bustamante:2010bf; @Bustamante:2010nq; @Mehta:2011qb; @Bustamante:2015waa; @Arguelles:2015dca; @Shoemaker:2015qul; @Gonzalez-Garcia:2016gpq; @Rasmussen:2017ert; @Ahlers:2018yom]. **What are the fundamental symmetries of Nature?** Beyond the TeV scale, the symmetries of the SM may break or new ones may appear. The effects of breaking lepton-number conservation, or CPT and Lorentz invariance [@Colladay:1998fq], cornerstones of the SM, are expected to grow with neutrino energy and affect multiple neutrino observables [@Kostelecky:2003cr; @Hooper:2005jp; @Kostelecky:2008ts; @Kostelecky:2011gq; @Gorham:2012qs; @Borriello:2013ala; @Stecker:2014xja; @Anchordoqui:2014hua; @Tomar:2015fha; @Amelino-Camelia:2015nqa; @Liao:2017yuy; @Anchordoqui:2006wc]. Currently, the strongest constraints in neutrinos come from high-energy atmospheric neutrinos [@Aartsen:2017ibm]; cosmic neutrinos could provide unprecedented sensitivity [@AmelinoCamelia:1997gz; @Hooper:2005jp; @GonzalezGarcia:2005xw; @Anchordoqui:2005gj; @Bazo:2009en; @Bustamante:2010nq; @Kostelecky:2011gq; @Diaz:2013wia; @Stecker:2014oxa; @Stecker:2014xja; @Tomar:2015fha; @Ellis:2018ogq; @Laha:2018hsh]. Further, detection of ZeV neutrinos, well beyond astrophysical expectations, would probe Grand Unified Theories [@Sigl:1996gm; @Berezinsky:2009xf; @Berezinsky:2011cp; @Lunardini:2012ct; @Anchordoqui:2018qom]. **Are neutrinos stable?** Neutrinos are essentially stable in the SM [@Pal:1981rm; @Hosotani:1981mq; @Nieves:1983fk], but BSM physics could introduce new channels for the heavier neutrinos to decay into the lighter ones [@Chikashige:1980qk; @Gelmini:1982rr; @Tomas:2001dh], with shorter lifetimes. During propagation over cosmological baselines, neutrino decay could leave imprints on the energy spectrum and flavor composition [@Beacom:2002vi; @Baerwald:2012kc; @Shoemaker:2015qul; @Bustamante:2016ciw; @Denton:2018aml]. The associated sensitivity outperforms existing limits obtained using neutrinos with shorter baselines [@Bustamante:2016ciw]. Comparable sensitivities are expected for similar BSM models, like pseudo-Dirac neutrinos [@Beacom:2003eu; @Joshipura:2013yba; @Shoemaker:2015qul]. **What is dark matter?** Cosmic neutrinos can probe the nature of dark matter. Dark matter may decay or self-annihilate into neutrinos [@Feng:2010gw; @Beacom:2006tt; @Yuksel:2007ac; @Murase:2012xs], leaving imprints on the neutrino energy spectrum, [*e.g.*]{}, line-like features. Searches for these features have yielded strong constraints on dark matter in the Milky Way [@Adrian-Martinez:2015wey; @Aartsen:2016pfc; @Aartsen:2017ulx] and nearby galaxies [@Aartsen:2013dxa]. High-energy cosmic neutrinos can probe superheavy dark matter with PeV masses [@Feldstein:2013kka; @Esmaili:2013gha; @Higaki:2014dwa; @Rott:2014kfa; @Dudas:2014bca; @Ema:2013nda; @Zavala:2014dla; @Murase:2015gea; @Anchordoqui:2015lqa; @Boucenna:2015tra; @Dev:2016qbd; @Hiroshima:2017hmy; @Chianese:2017nwe] and light dark matter [@Higaki:2014dwa; @Fong:2014bsa; @Cohen:2016uyg; @Hiroshima:2017hmy; @Sui:2018bbh]. Multi-messenger constraints are crucial to assess dark matter explanations of the observed neutrino spectrum [@Ahlers:2013xia; @Bhattacharya:2014vwa; @Murase:2015gea; @Cohen:2016uyg; @Bhattacharya:2017jaw]. Further, anisotropies in the neutrino sky towards the Galactic Center can reveal dark matter decaying [@Bai:2013nga] or interacting with neutrinos [@Arguelles:2017atb]. **Are there hidden interactions with cosmic backgrounds?** High-energy cosmic neutrinos may interact with low-energy relic neutrino backgrounds via new interactions [@Lykken:2007kp; @Ioka:2014kca; @Ng:2014pca; @Blum:2014ewa; @Shoemaker:2015qul; @Altmannshofer:2016brv; @Barenboim:2019tux], with large-scale distributions of matter via new forces [@Bustamante:2018mzu], or with dark backgrounds [@Capozzi:2018bps], including dark energy [@Anchordoqui:2007iw; @Klop:2017dim]. These interactions may mimic the existence of neutrino mass, affect the neutrino flavor composition, and induce anisotropies in the high-energy neutrino sky. ![\[fig:flavor\_ratios\]Flavor composition at Earth of high-energy cosmic neutrinos, indicating the “theoretically palatable" [@Bustamante:2015waa] regions accessible with the Standard Model with massive neutrinos ($\nu$SM), with new physics similar to neutrino decay, and with new physics similar to Lorentz-invariance violation. The neutrino mixing parameters are generously varied within their uncertainties at $3\sigma$. The tilt of the tick marks indicates the orientation along which to read the flavor content.[]{data-label="fig:parameter_fit"}](cross_sections_compare_6yr_with_results_testable_regions.pdf){width="\linewidth"} ![\[fig:flavor\_ratios\]Flavor composition at Earth of high-energy cosmic neutrinos, indicating the “theoretically palatable" [@Bustamante:2015waa] regions accessible with the Standard Model with massive neutrinos ($\nu$SM), with new physics similar to neutrino decay, and with new physics similar to Lorentz-invariance violation. The neutrino mixing parameters are generously varied within their uncertainties at $3\sigma$. The tilt of the tick marks indicates the orientation along which to read the flavor content.[]{data-label="fig:parameter_fit"}](flavor_ratios_new_physics.png){width="\linewidth"} Neutrino Observables: What Do We Use to Probe Fundamental Physics? {#neutrino-observables-what-do-we-use-to-probe-fundamental-physics .unnumbered} ------------------------------------------------------------------ To probe fundamental physics, we look at four neutrino observables, individually or together [@Arguelles_review:in_prep]. [**Energy spectrum:**]{} The spectrum of neutrinos depends on their production processes, but BSM effects could introduce identifiable features, [*e.g.*]{}, peaks, troughs, and cut-offs. Present neutrino telescopes reconstruct the energy $E$ of detected events to within 0.1 in $\log_{10} (E/{\rm GeV})$ [@Aartsen:2013vja]. For TeV–PeV astrophysical neutrinos, the spectrum is predicted to be a featureless power law. IceCube data are consistent with that, but also with a broken power law [@Murase:2013rfa; @Chen:2013dza; @Chen:2014gxa; @Aartsen:2015knd; @Anchordoqui:2016ewn; @Vincent:2016nut]. For EeV cosmogenic neutrinos, the spectrum has a different but predictable shape [@Beresinsky:1969qj; @Berezinsky:1975zz; @Stecker:1978ah; @Hill:1983xs; @Yoshida:1993pt; @Engel:2001hd; @Anchordoqui:2007fi; @Takami:2007pp; @Ahlers:2009rf; @Ahlers:2010fw; @Kotera:2010yn; @Yoshida:2012gf; @Ahlers:2012rz; @Aloisio:2015ega; @Heinze:2015hhp; @Romero-Wolf:2017xqe; @AlvesBatista:2018zui; @Moller:2018isk; @vanVliet:2019nse; @Heinze:2019jou], so BSM effects, [*e.g.*]{}, modifications of neutrino-nucleon cross sections [@Kusenko:2001gj; @AlvarezMuniz:2001mk; @Anchordoqui:2001cg; @Cornet:2001gy; @Kowalski:2002gb; @AlvarezMuniz:2002ga; @Anchordoqui:2005pn; @Marfatia:2015hva; @Ellis:2016dgb; @Anchordoqui:2018qom], may also be apparent. [**Arrival directions:**]{} If the diffuse flux of cosmic neutrinos comes from an isotropic distribution of sources, then it should be isotropic itself. However, interactions with cosmic backgrounds might induce anisotropies. For instance, they could create a neutrino horizon, whereby high-energy neutrinos could only reach us from a few nearby sources [@Ioka:2014kca; @Ng:2014pca; @Cherry:2014xra]. Similarly, neutrino interactions with dark matter could introduce an anisotropy towards the Galactic Center [@Arguelles:2017atb]. Presently, the pointing resolution at neutrino telescopes is sub-degree for events initiated by $\nu_\mu$ — tracks — and of a few degrees for events initiated mainly by $\nu_e$ and $\nu_\tau$ — showers [@Aartsen:2013vja]. [**Flavor composition:**]{} At the neutrino sources, high-energy cosmic neutrinos are believed to be produced in the decay of pions, [*i.e.*]{}, $\pi^+ \to \mu^+ \nu_\mu$ followed by $\mu^+ \to e^+ \nu_e \bar{\nu}_\mu$. This results in an initial flavor composition of $\left( \nu_e : \nu_\mu : \nu_\tau \right) = \left( 1 : 2 : 0 \right)$, adding $\nu$ and $\bar{\nu}$. Upon reaching Earth, oscillations have transformed this into nearly $\left( 1 : 1: 1 \right)_\oplus$ [@Pakvasa:2008nx]. The detection of $\nu_\tau$ is minimally required for testing this standard oscillation scenario [@Palladino:2018qgi; @Parke:2015goa]. While there are variations on this canonical expectation [@Barenboim:2003jm; @Kashti:2005qa; @Lipari:2007su], the expected flavor ratios fall within a well-defined region [@Bustamante:2015waa]. However, numerous BSM models active during propagation may modify this [@Mehta:2011qb; @Rasmussen:2017ert], including neutrino decay and Lorentz invariance violation, as shown in Fig. \[fig:flavor\_ratios\]. A precise measurement of the flavor composition could distinguish between these two classes of models [@Bustamante:2015waa]. Presently, measuring flavor at neutrino telescopes is challenging, since the showers made by $\nu_e$ and $\nu_\tau$ look similar [@Aartsen:2015ivb; @Aartsen:2015knd], which makes the contours of allowed flavor composition in Fig. \[fig:flavor\_ratios\] wide. [**Timing:**]{} A violation of Lorentz invariance would modify the energy-momentum relation of neutrinos and photons [@AmelinoCamelia:2003ex; @Christian:2004xb; @Diaz:2014yva], causing them to have different speeds at different energies. This would manifest in neutrinos [@Diaz:2016xpw; @Murase:2019xqi], photons [@Longo:1987ub; @Wang:2016lne; @Wei:2016ygk; @Boran:2018ypz], and gravitational waves [@Baret:2011tk] emitted at the same time from transient sources arriving at Earth at different times. Presently, electronics in neutrino telescopes can timestamp events to within a few nanoseconds [@Aartsen:2016nxy]. Today, the strength of the tests performed using these observables is limited at PeV energies, where data is scant, but event statistics are growing and there are ongoing efforts to improve the reconstruction of neutrino properties. Once neutrinos of higher energies are detected, the same observables can be used to test fundamental physics in a new energy regime. ![\[fig:physics\_flow\_chart\][Observatory requirements to test fundamental physics with cosmic neutrinos.]{}](decadal_neutrino_astronomy_flow_chart_fundamental_physics_v3_long.pdf){width="\textwidth"} Observatory Requirements to Achieve the Science Goals {#observatory-requirements-to-achieve-the-science-goals .unnumbered} ----------------------------------------------------- Answering fundamental physics questions requires improving the precision with which neutrino observables are measured, which is currently limited by the low numbers of events. The statistics in the TeV–PeV energy range will grow using existing neutrino detectors and their planned upgrades. This will be supplemented by improved techniques to reconstruct neutrino energy, direction, and flavor. At the EeV scale, our ability to address fundamental physics questions is contingent on the discovery of neutrinos at these energies. In addition to emphasizing the importance of improved statistics, we highlight two measurements that can be improved in the coming decade: the neutrino cross section and flavor composition. Presently, the measurement of the TeV–PeV neutrino cross section in multiple energy bins is sorely statistics-limited [@Bustamante:2017xuy]. In this energy range, where the measured cross section is compatible with SM expectations, large BSM deviations are unlikely. But smaller deviations are still possible, especially close to PeV energies. To extract the cross section, Ref. [@Bustamante:2017xuy] used about 60 shower events collected by IceCube in six years across all energies. A detector that is five times larger [@Aartsen:2014njl] would collect 300 showers in the same time, reducing the statistical error in the extracted cross sections by a factor of $\sqrt{6}/6 \approx 0.4$ [@Anchordoqui:2019ufu]. At that point, the statistical and systematic errors would become comparable, with a size of about 0.2 in the logarithm of the cross section (in units of cm$^2$). At the EeV scale, measuring the cross section to within an order of magnitude could distinguish between SM predictions and BSM modifications; see Fig. \[fig:cross\_section\]. This target is achievable with tens of events in the PeV–EeV energy range. Detection will be challenging, since the flux is expected to decrease fast with energy and the cross section is expected to grow with energy, making the Earth opaque to neutrinos. Facing significant uncertainties in the predicted flux of cosmogenic neutrinos [@Kotera:2010yn; @Ahlers:2012rz; @Romero-Wolf:2017xqe; @AlvesBatista:2018zui; @Heinze:2019jou], we advocate for the construction of larger neutrino observatories to boost the chances of discovering and collecting a sufficiently large number of cosmogenic neutrinos. Flavor composition must be measured with a precision better than 40% to match the theoretical SM uncertainty band and identify BSM deviations, as shown in Fig. \[fig:flavor\_ratios\]. Reaching this target at TeV–PeV energies requires supplementing the larger event statistics with the detection of flavor-specific signals [@Glashow:1960zz; @Learned:1994wg; @Li:2016kra]. With 20% precision, we could distinguish between models similar to neutrino decay or to Lorentz invariance violation. Improved statistics will also permit searches for a potential energy dependence of mixing, which could point to the presence of BSM effects [@Mehta:2011qb; @Bustamante:2015waa]. In the EeV range, we advocate exploring new methods to measure flavor in existing and upcoming experiments ([*e.g.*]{}, Ref. [@Wang:2013bvz]). Some planned EeV detectors will be sensitive primarily [@Fargion:1999se] to $\nu_\tau$[@Olinto:2017xbi; @Sasaki:2017zwd; @BEACON_ARENA2018:Talk; @Liu:2018hux; @Alvarez-Muniz:2018bhp; @Otte:2018uxj], while others will be sensitive to all flavors [@Allison:2011wk; @Barwick:2014rca; @Aartsen:2014njl; @Gorham:2008dv; @Adrian-Martinez:2016fdl; @James:2017pvr], but might not be able to distinguish between them easily. Thus, we should consider combining data from the two types of experiments in order to infer at least the $\nu_\tau$ fraction. Further, with the available sub-degree pointing resolution, we can begin to probe anisotropies in the neutrino sky that may result, [*e.g.*]{}, from Lorentz-invariance violation [@Abbasi:2010kx] or BSM matter interactions [@Arguelles:2017atb]. Additionally, we can cull a set of neutrino events that are truly extragalactic, by using only those that point away from the Galactic Center, which allows us to make robust searches for BSM effects that are enhanced over cosmological distances ([*e.g.*]{}, Ref. [@Bustamante:2016ciw]). We advocate for a strategy for the coming decade that improves precision on flavor identification and improves statistics across a broad energy scale, from 10 TeV up to the EeV scale. While this strategy targets mainly cross section and flavor measurements, it will impact other neutrino observables and relentlessly test the predictions of the SM and of many BSM scenarios. [100]{} M. Ahlers, K. Helbing, and C. Pérez de los Heros, “[Probing Particle Physics with IceCube]{},” [*Eur. Phys. J. C*]{} [**78**]{} (2018) 924, [[1806.05696]{}](http://arxiv.org/abs/1806.05696). L. Anchordoqui and F. Halzen, “[IceHEP high energy physics at the south pole]{},” [*Annals Phys.*]{} [**321**]{} (2006) 2660, [[hep-ph/0510389]{}](http://arxiv.org/abs/hep-ph/0510389). D. Marfatia, D. W. McKay, and T. J. Weiler, “[New physics with ultra-high-energy neutrinos]{},” [*Phys. Lett. B*]{} [**748**]{} (2015) 113, [[1502.06337]{}](http://arxiv.org/abs/1502.06337). Collaboration, M. G. Aartsen [*et. al.*]{}, “[First observation of PeV-energy neutrinos with IceCube]{},” [*Phys. Rev. Lett.*]{} [**111**]{} (2013) 021103, [[1304.5356]{}](http://arxiv.org/abs/1304.5356). S. Razzaque, “[The Galactic Center Origin of a Subset of IceCube Neutrino Events]{},” [*Phys. Rev. D*]{} [**88**]{} (2013) 081302, [[1309.2756]{}](http://arxiv.org/abs/1309.2756). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Evidence for High-Energy Extraterrestrial Neutrinos at the IceCube Detector]{},” [ *Science*]{} [**342**]{} (2013) 1242856, [[ 1311.5238]{}](http://arxiv.org/abs/1311.5238). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Observation of High-Energy Astrophysical Neutrinos in Three Years of IceCube Data]{},” [ *Phys. Rev. Lett.*]{} [**113**]{} (2014) 101101, [[1405.5303]{}](http://arxiv.org/abs/1405.5303). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Evidence for Astrophysical Muon Neutrinos from the Northern Sky with IceCube]{},” [ *Phys. Rev. Lett.*]{} [**115**]{} (2015) 081102, [[1507.04005]{}](http://arxiv.org/abs/1507.04005). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Observation and Characterization of a Cosmic Muon Neutrino Flux from the Northern Hemisphere using six years of IceCube data]{},” [*Astrophys. J.*]{} [**833**]{} (2016) 3, [[1607.08006]{}](http://arxiv.org/abs/1607.08006). M. Ahlers and K. Murase, “[Probing the Galactic Origin of the IceCube Excess with Gamma-Rays]{},” [*Phys. Rev. D*]{} [**90**]{} (2014) 023010, [[1309.4077]{}](http://arxiv.org/abs/1309.4077). L. A. Anchordoqui [*et. al.*]{}, “[Cosmic Neutrino Pevatrons: A Brand New Pathway to Astronomy, Astrophysics, and Particle Physics]{},” [*JHEAp*]{} [ **1-2**]{} (2014) 1, [[1312.6587]{}](http://arxiv.org/abs/1312.6587). M. Ahlers, Y. Bai, V. Barger, and R. Lu, “[Galactic neutrinos in the TeV to PeV range]{},” [*Phys. Rev. D*]{} [**93**]{} (2016) 013009, [[1505.03156]{}](http://arxiv.org/abs/1505.03156). P. B. Denton, D. Marfatia, and T. J. Weiler, “[The Galactic Contribution to IceCube’s Astrophysical Neutrino Flux]{},” [*JCAP*]{} [**1708**]{} (2017) 033, [[1703.09721]{}](http://arxiv.org/abs/1703.09721). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Constraints on Galactic Neutrino Emission with Seven Years of IceCube Data]{},” [ *Astrophys. J.*]{} [**849**]{} (2017) 67, [[1707.03416]{}](http://arxiv.org/abs/1707.03416). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Multimessenger observations of a flaring blazar coincident with high-energy neutrino IceCube-170922A]{},” [*Science*]{} [**361**]{} (2018) eaat1378, [[1807.08816]{}](http://arxiv.org/abs/1807.08816). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Neutrino emission from the direction of the blazar TXS 0506+056 prior to the IceCube-170922A alert]{},” [*Science*]{} [**361**]{} (2018) 147, [[1807.08794]{}](http://arxiv.org/abs/1807.08794). K. Greisen, “[End to the cosmic ray spectrum?]{},” [*Phys. Rev. Lett.*]{} [**16**]{} (1966) 748. G. T. Zatsepin and V. A. Kuzmin, “[Upper limit of the spectrum of cosmic rays]{},” [*JETP Lett.*]{} [**4**]{} (1966) 78. \[Pisma Zh. Eksp. Teor. Fiz. 4, 114 (1966)\]. Collaboration, E. Zas, “[Searches for neutrino fluxes in the EeV regime with the Pierre Auger Observatory]{},” [*PoS*]{} [**ICRC2017**]{} (2018) 972. Collaboration, M. G. Aartsen [*et. al.*]{}, “[Differential limit on the extremely-high-energy cosmic neutrino flux in the presence of astrophysical background from nine years of IceCube data]{},” [*Phys. Rev. D*]{} [**98**]{} (2018) 062003, [[ 1807.01820]{}](http://arxiv.org/abs/1807.01820). Collaboration, P. W. Gorham [*et. al.*]{}, “[Constraints on the ultra-high energy cosmic neutrino flux from the fourth flight of ANITA]{},” [[1902.04005]{}](http://arxiv.org/abs/1902.04005). M. Ackermann [*et. al.*]{}, “[Astrophysics Uniquely Enabled by Observations of High-Energy Cosmic Neutrinos]{},” [*Bull. Am. Astron. Soc.*]{} [**51**]{} (2019) 185, [[1903.04334]{}](http://arxiv.org/abs/1903.04334). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Measurement of the multi-TeV neutrino cross section with IceCube using Earth absorption]{},” [ *Nature*]{} [**551**]{} (2017) 596, [[ 1711.08119]{}](http://arxiv.org/abs/1711.08119). M. Bustamante and A. Connolly, “[Extracting the Energy-Dependent Neutrino-Nucleon Cross Section Above 10 TeV Using IceCube Showers]{},” [ *Phys. Rev. Lett.*]{} [**122**]{} (2019) 041101, [[1711.11043]{}](http://arxiv.org/abs/1711.11043). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Measurements using the inelasticity distribution of multi-TeV neutrino interactions in IceCube]{},” [*Phys. Rev. D*]{} [**99**]{} (2019) 032004, [[1808.07629]{}](http://arxiv.org/abs/1808.07629). V. S. Berezinsky and A. [Yu]{}. Smirnov, “[Astrophysical upper bounds on neutrino-nucleon cross-section at energy $E \geq 3 \times 10^{17}$ eV]{},” [*Phys. Lett. B*]{} [**48**]{} (1974) 269. D. Hooper, “[Measuring high-energy neutrino nucleon cross-sections with future neutrino telescopes]{},” [*Phys. Rev. D*]{} [**65**]{} (2002) 097303, [[hep-ph/0203239]{}](http://arxiv.org/abs/hep-ph/0203239). S. Hussain, D. Marfatia, D. W. McKay, and D. Seckel, “[Cross section dependence of event rates at neutrino telescopes]{},” [*Phys. Rev. Lett.*]{} [**97**]{} (2006) 161101, [[ hep-ph/0606246]{}](http://arxiv.org/abs/hep-ph/0606246). E. Borriello, A. Cuoco, G. Mangano, G. Miele, S. Pastor, O. Pisanti, and P. D. Serpico, “[Disentangling neutrino-nucleon cross section and high energy neutrino flux with a km$^3$ neutrino telescope]{},” [*Phys. Rev. D*]{} [ **77**]{} (2008) 045019, [[0711.0152]{}](http://arxiv.org/abs/0711.0152). S. Hussain, D. Marfatia, and D. W. McKay, “[Upward shower rates at neutrino telescopes directly determine the neutrino flux]{},” [*Phys. Rev. D*]{} [ **77**]{} (2008) 107304, [[0711.4374]{}](http://arxiv.org/abs/0711.4374). J. M. Conrad, M. H. Shaevitz, and T. Bolton, “[Precision measurements with high-energy neutrino beams]{},” [*Rev. Mod. Phys.*]{} [**70**]{} (1998) 1341, [[hep-ex/9707015]{}](http://arxiv.org/abs/hep-ex/9707015). J. A. Formaggio and G. P. Zeller, “[From eV to EeV: Neutrino Cross Sections Across Energy Scales]{},” [*Rev. Mod. Phys.*]{} [**84**]{} (2012) 1307, [[1305.7513]{}](http://arxiv.org/abs/1305.7513). Collaboration, M. Tanabashi [*et. al.*]{}, “[Review of Particle Physics]{},” [*Phys. Rev. D*]{} [**98**]{} (2018) 030001. A. Cooper-Sarkar, P. Mertsch, and S. Sarkar, “[The high energy neutrino cross-section in the Standard Model and its uncertainty]{},” [*JHEP*]{} [ **08**]{} (2011) 042, [[1106.3723]{}](http://arxiv.org/abs/1106.3723). A. Kusenko and T. J. Weiler, “[Neutrino cross-sections at high-energies and the future observations of ultrahigh-energy cosmic rays]{},” [*Phys. Rev. Lett.*]{} [**88**]{} (2002) 161101, [[hep-ph/0106071]{}](http://arxiv.org/abs/hep-ph/0106071). J. Alvarez-Muniz, F. Halzen, T. Han, and D. Hooper, “[Phenomenology of high-energy neutrinos in low scale quantum gravity models]{},” [*Phys. Rev. Lett.*]{} [**88**]{} (2002) 021301, [[hep-ph/0107057]{}](http://arxiv.org/abs/hep-ph/0107057). L. A. Anchordoqui, J. L. Feng, H. Goldberg, and A. D. Shapere, “[Black holes from cosmic rays: Probes of extra dimensions and new limits on TeV scale gravity]{},” [*Phys. Rev. D*]{} [**65**]{} (2002) 124027, [[hep-ph/0112247]{}](http://arxiv.org/abs/hep-ph/0112247). F. Cornet, J. I. Illana, and M. Masip, “[TeV strings and the neutrino nucleon cross-section at ultrahigh-energies]{},” [*Phys. Rev. Lett.*]{} [**86**]{} (2001) 4235, [[ hep-ph/0102065]{}](http://arxiv.org/abs/hep-ph/0102065). M. Kowalski, A. Ringwald, and H. Tu, “[Black holes at neutrino telescopes]{},” [*Phys. Lett. B*]{} [**529**]{} (2002) 1, [[hep-ph/0201139]{}](http://arxiv.org/abs/hep-ph/0201139). J. Alvarez-Muniz, J. L. Feng, F. Halzen, T. Han, and D. Hooper, “[Detecting microscopic black holes with neutrino telescopes]{},” [*Phys. Rev. D*]{} [**65**]{} (2002) 124015, [[ hep-ph/0202081]{}](http://arxiv.org/abs/hep-ph/0202081). L. A. Anchordoqui, J. L. Feng, and H. Goldberg, “[Particle physics on ice: Constraints on neutrino interactions far above the weak scale]{},” [*Phys. Rev. Lett.*]{} [**96**]{} (2006) 021101, [[hep-ph/0504228]{}](http://arxiv.org/abs/hep-ph/0504228). J. Ellis, K. Sakurai, and M. Spannowsky, “[Search for Sphalerons: IceCube vs. LHC]{},” [*JHEP*]{} [**05**]{} (2016) 085, [[1603.06573]{}](http://arxiv.org/abs/1603.06573). L. A. Anchordoqui, “[Ultra-High-Energy Cosmic Rays]{},” [*Phys. Rep.*]{} [ **801**]{} (2019) 1, [[1807.09645]{}](http://arxiv.org/abs/1807.09645). E. M. Henley and J. Jalilian-Marian, “[Ultra-high energy neutrino-nucleon scattering and parton distributions at small $x$]{},” [*Phys. Rev. D*]{} [**73**]{} (2006) 094004, [[ hep-ph/0512220]{}](http://arxiv.org/abs/hep-ph/0512220). N. Armesto, C. Merino, G. Parente, and E. Zas, “[Charged current neutrino cross-section and tau energy loss at ultra-high energies]{},” [*Phys. Rev. D*]{} [**77**]{} (2008) 013001, [[ 0709.4461]{}](http://arxiv.org/abs/0709.4461). R. Enberg, M. H. Reno, and I. Sarcevic, “[Prompt neutrino fluxes from atmospheric charm]{},” [*Phys. Rev. D*]{} [**78**]{} (2008) 043005, [[0806.0418]{}](http://arxiv.org/abs/0806.0418). A. [Yu]{}. Illarionov, B. A. Kniehl, and A. V. Kotikov, “[Ultrahigh-energy neutrino-nucleon deep-inelastic scattering and the Froissart bound]{},” [*Phys. Rev. Lett.*]{} [**106**]{} (2011) 231802, [[1105.2829]{}](http://arxiv.org/abs/1105.2829). A. Bhattacharya, R. Enberg, M. H. Reno, I. Sarcevic, and A. Stasto, “[Perturbative charm production and the prompt atmospheric neutrino flux in light of RHIC and LHC]{},” [*JHEP*]{} [**06**]{} (2015) 110, [[1502.01076]{}](http://arxiv.org/abs/1502.01076). M. V. Garzelli, S. Moch, and G. Sigl, “[Lepton fluxes from atmospheric charm revisited]{},” [*JHEP*]{} [**10**]{} (2015) 115, [[1507.01570]{}](http://arxiv.org/abs/1507.01570). F. Halzen and L. Wille, “[Upper Limit on Forward Charm Contribution to Atmospheric Neutrino Flux]{},” [[ 1601.03044]{}](http://arxiv.org/abs/1601.03044). F. Halzen and L. Wille, “[Charm contribution to the atmospheric neutrino flux]{},” [*Phys. Rev. D*]{} [**94**]{} (2016) 014014, [[1605.01409]{}](http://arxiv.org/abs/1605.01409). A. Bhattacharya, R. Enberg, Y. S. Jeong, C. S. Kim, M. H. Reno, I. Sarcevic, and A. Stasto, “[Prompt atmospheric neutrino fluxes: perturbative QCD models and nuclear effects]{},” [*JHEP*]{} [**11**]{} (2016) 167, [[1607.00193]{}](http://arxiv.org/abs/1607.00193). M. Benzke, M. V. Garzelli, B. Kniehl, G. Kramer, S. Moch, and G. Sigl, “[Prompt neutrinos from atmospheric charm in the general-mass variable-flavor-number scheme]{},” [*JHEP*]{} [**12**]{} (2017) 021, [[1705.10386]{}](http://arxiv.org/abs/1705.10386). A. V. Giannini, V. P. Gonçalves, and F. S. Navarra, “[Intrinsic charm contribution to the prompt atmospheric neutrino flux]{},” [*Phys. Rev. D*]{} [**98**]{} (2018) 014012, [[ 1803.01728]{}](http://arxiv.org/abs/1803.01728). L. A. Anchordoqui, A. M. Cooper-Sarkar, D. Hooper, and S. Sarkar, “[Probing low-x QCD with cosmic neutrinos at the Pierre Auger Observatory]{},” [ *Phys. Rev. D*]{} [**74**]{} (2006) 043008, [[hep-ph/0605086]{}](http://arxiv.org/abs/hep-ph/0605086). V. Bertone, R. Gauld, and J. Rojo, “[Neutrino Telescopes as QCD Microscopes]{},” [*JHEP*]{} [**01**]{} (2019) 217, [[1808.02034]{}](http://arxiv.org/abs/1808.02034). M. Bustamante, J. F. Beacom, and W. Winter, “[Theoretically palatable flavor combinations of astrophysical neutrinos]{},” [*Phys. Rev. Lett.*]{} [ **115**]{} (2015) 161302, [[1506.02645]{}](http://arxiv.org/abs/1506.02645). J. G. Learned and S. Pakvasa, “[Detecting tau-neutrino oscillations at PeV energies]{},” [*Astropart. Phys.*]{} [**3**]{} (1995) 267, [[hep-ph/9405296]{}](http://arxiv.org/abs/hep-ph/9405296). J. F. Beacom, N. F. Bell, D. Hooper, S. Pakvasa, and T. J. Weiler, “[Measuring flavor ratios of high-energy astrophysical neutrinos]{},” [*Phys. Rev. D*]{} [**68**]{} (2003) 093005, [[ hep-ph/0307025]{}](http://arxiv.org/abs/hep-ph/0307025). \[Erratum: Phys. Rev. D 72, 019901 (2005)\]. S. Pakvasa, W. Rodejohann, and T. J. Weiler, “[Flavor Ratios of Astrophysical Neutrinos: Implications for Precision Measurements]{},” [*JHEP*]{} [**02**]{} (2008) 005, [[0711.4517]{}](http://arxiv.org/abs/0711.4517). M. Bustamante, A. M. Gago, and J. Jones Perez, “[SUSY Renormalization Group Effects in Ultra High Energy Neutrinos]{},” [*JHEP*]{} [**05**]{} (2011) 133, [[1012.2728]{}](http://arxiv.org/abs/1012.2728). M. Bustamante, A. M. Gago, and C. Pena-Garay, “[Energy-independent new physics in the flavour ratios of high-energy astrophysical neutrinos]{},” [*JHEP*]{} [**04**]{} (2010) 066, [[1001.4878]{}](http://arxiv.org/abs/1001.4878). P. Mehta and W. Winter, “[Interplay of energy dependent astrophysical neutrino flavor ratios and new physics effects]{},” [*JCAP*]{} [**1103**]{} (2011) 041, [[1101.2673]{}](http://arxiv.org/abs/1101.2673). C. A. Arg[ü]{}elles, T. Katori, and J. Salvado, “[New Physics in Astrophysical Neutrino Flavor]{},” [*Phys. Rev. Lett.*]{} [**115**]{} (2015) 161303, [[1506.02043]{}](http://arxiv.org/abs/1506.02043). I. M. Shoemaker and K. Murase, “[Probing BSM Neutrino Physics with Flavor and Spectral Distortions: Prospects for Future High-Energy Neutrino Telescopes]{},” [*Phys. Rev. D*]{} [**93**]{} (2016) 085004, [[1512.07228]{}](http://arxiv.org/abs/1512.07228). M. C. Gonzalez-Garcia, M. Maltoni, I. Martinez-Soler, and N. Song, “[Non-standard neutrino interactions in the Earth and the flavor of astrophysical neutrinos]{},” [*Astropart. Phys.*]{} [**84**]{} (2016) 15, [[1605.08055]{}](http://arxiv.org/abs/1605.08055). R. W. Rasmussen, L. Lechner, M. Ackermann, M. Kowalski, and W. Winter, “[Astrophysical neutrinos flavored with Beyond the Standard Model physics]{},” [*Phys. Rev. D*]{} [**96**]{} (2017) 083018, [[1707.07684]{}](http://arxiv.org/abs/1707.07684). M. Ahlers, M. Bustamante, and S. Mu, “[Unitarity Bounds of Astrophysical Neutrinos]{},” [*Phys. Rev. D*]{} [**98**]{} (2018) 123023, [[1810.00893]{}](http://arxiv.org/abs/1810.00893). D. Colladay and V. A. Kostelecky, “[Lorentz violating extension of the Standard Model]{},” [*Phys. Rev. D*]{} [**58**]{} (1998) 116002, [[hep-ph/9809521]{}](http://arxiv.org/abs/hep-ph/9809521). V. A. Kostelecky and M. Mewes, “[Lorentz and CPT violation in neutrinos]{},” [*Phys. Rev. D*]{} [**69**]{} (2004) 016005, [[hep-ph/0309025]{}](http://arxiv.org/abs/hep-ph/0309025). D. Hooper, D. Morgan, and E. Winstanley, “[Lorentz and CPT invariance violation in high-energy neutrinos]{},” [*Phys. Rev. D*]{} [**72**]{} (2005) 065009, [[hep-ph/0506091]{}](http://arxiv.org/abs/hep-ph/0506091). V. A. Kostelecky and N. Russell, “[Data Tables for Lorentz and CPT Violation]{},” [*Rev. Mod. Phys.*]{} [**83**]{} (2011) 11, [[0801.0287]{}](http://arxiv.org/abs/0801.0287). A. Kostelecky and M. Mewes, “[Neutrinos with Lorentz-violating operators of arbitrary dimension]{},” [*Phys. Rev. D*]{} [**85**]{} (2012) 096005, [[1112.6395]{}](http://arxiv.org/abs/1112.6395). P. W. Gorham [*et. al.*]{}, “[Implications of ultra-high energy neutrino flux constraints for Lorentz-invariance violating cosmogenic neutrinos]{},” [ *Phys. Rev. D*]{} [**86**]{} (2012) 103006, [[1207.6425]{}](http://arxiv.org/abs/1207.6425). E. Borriello, S. Chakraborty, A. Mirizzi, and P. D. Serpico, “[Stringent constraint on neutrino Lorentz-invariance violation from the two IceCube PeV neutrinos]{},” [*Phys. Rev. D*]{} [**87**]{} (2013) 116009, [[1303.5843]{}](http://arxiv.org/abs/1303.5843). F. W. Stecker and S. T. Scully, “[Propagation of Superluminal PeV IceCube Neutrinos: A High Energy Spectral Cutoff or New Constraints on Lorentz Invariance Violation]{},” [*Phys. Rev. D*]{} [**90**]{} (2014) 043012, [[1404.7025]{}](http://arxiv.org/abs/1404.7025). L. A. Anchordoqui, V. Barger, H. Goldberg, J. G. Learned, D. Marfatia, S. Pakvasa, T. C. Paul, and T. J. Weiler, “[End of the cosmic neutrino energy spectrum]{},” [*Phys. Lett. B*]{} [**739**]{} (2014) 99, [[1404.0622]{}](http://arxiv.org/abs/1404.0622). G. Tomar, S. Mohanty, and S. Pakvasa, “[Lorentz Invariance Violation and IceCube Neutrino Events]{},” [*JHEP*]{} [**11**]{} (2015) 022, [[1507.03193]{}](http://arxiv.org/abs/1507.03193). G. Amelino-Camelia, D. Guetta, and T. Piran, “[Icecube Neutrinos and Lorentz Invariance Violation]{},” [*Astrophys. J.*]{} [**806**]{} (2015) 269. J. Liao and D. Marfatia, “[IceCube’s astrophysical neutrino energy spectrum from CPT violation]{},” [*Phys. Rev. D*]{} [**97**]{} (2018) 041302, [[1711.09266]{}](http://arxiv.org/abs/1711.09266). L. A. Anchordoqui, C. A. Garcia Canal, H. Goldberg, D. Gomez Dumm, and F. Halzen, “[Probing leptoquark production at IceCube]{},” [*Phys. Rev. D*]{} [**74**]{} (2006) 125021, [[ hep-ph/0609214]{}](http://arxiv.org/abs/hep-ph/0609214). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Neutrino Interferometry for High-Precision Tests of Lorentz Symmetry with IceCube]{},” [*Nature Phys.*]{} [**14**]{} (2018) 961, [[1709.03434]{}](http://arxiv.org/abs/1709.03434). G. Amelino-Camelia, J. R. Ellis, N. E. Mavromatos, D. V. Nanopoulos, and S. Sarkar, “[Tests of quantum gravity from observations of gamma-ray bursts]{},” [*Nature*]{} [**393**]{} (1998) 763, [[astro-ph/9712103]{}](http://arxiv.org/abs/astro-ph/9712103). M. C. Gonzalez-Garcia, F. Halzen, and M. Maltoni, “[Physics reach of high-energy and high-statistics icecube atmospheric neutrino data]{},” [ *Phys. Rev. D*]{} [**71**]{} (2005) 093010, [[hep-ph/0502223]{}](http://arxiv.org/abs/hep-ph/0502223). L. A. Anchordoqui, H. Goldberg, M. C. Gonzalez-Garcia, F. Halzen, D. Hooper, S. Sarkar, and T. J. Weiler, “[Probing Planck scale physics with IceCube]{},” [*Phys. Rev. D*]{} [**72**]{} (2005) 065019, [[hep-ph/0506168]{}](http://arxiv.org/abs/hep-ph/0506168). J. L. Bazo, M. Bustamante, A. M. Gago, and O. G. Miranda, “[High energy astrophysical neutrino flux and modified dispersion relations]{},” [*Int. J. Mod. Phys. A*]{} [**24**]{} (2009) 5819, [[0907.1979]{}](http://arxiv.org/abs/0907.1979). J. S. Diaz, A. Kostelecky, and M. Mewes, “[Testing Relativity with High-Energy Astrophysical Neutrinos]{},” [*Phys. Rev. D*]{} [**89**]{} (2014) 043005, [[1308.6344]{}](http://arxiv.org/abs/1308.6344). F. W. Stecker, S. T. Scully, S. Liberati, and D. Mattingly, “[Searching for Traces of Planck-Scale Physics with High Energy Neutrinos]{},” [*Phys. Rev. D*]{} [**91**]{} (2015) 045009, [[ 1411.5889]{}](http://arxiv.org/abs/1411.5889). J. Ellis, N. E. Mavromatos, A. S. Sakharov, and E. K. Sarkisyan-Grinbaum, “[Limits on Neutrino Lorentz Violation from Multimessenger Observations of TXS 0506+056]{},” [*Phys. Lett. B*]{} [**789**]{} (2019) 352, [[1807.05155]{}](http://arxiv.org/abs/1807.05155). R. Laha, “[Constraints on neutrino speed, weak equivalence principle violation, Lorentz invariance violation, and dual lensing from the first high-energy astrophysical neutrino source TXS 0506+056]{},” [[1807.05621]{}](http://arxiv.org/abs/1807.05621). G. Sigl, S. Lee, D. N. Schramm, and P. Coppi, “[Cosmological neutrino signatures for grand unification scale physics]{},” [*Phys. Lett. B*]{} [ **392**]{} (1997) 129, [[ astro-ph/9610221]{}](http://arxiv.org/abs/astro-ph/9610221). V. Berezinsky, K. D. Olum, E. Sabancilar, and A. Vilenkin, “[UHE neutrinos from superconducting cosmic strings]{},” [*Phys. Rev. D*]{} [**80**]{} (2009) 023014, [[0901.0527]{}](http://arxiv.org/abs/0901.0527). V. Berezinsky, E. Sabancilar, and A. Vilenkin, “[Extremely High Energy Neutrinos from Cosmic Strings]{},” [*Phys. Rev. D*]{} [**84**]{} (2011) 085006, [[1108.2509]{}](http://arxiv.org/abs/1108.2509). C. Lunardini and E. Sabancilar, “[Cosmic Strings as Emitters of Extremely High Energy Neutrinos]{},” [*Phys. Rev. D*]{} [**86**]{} (2012) 085008, [[1206.2924]{}](http://arxiv.org/abs/1206.2924). P. B. Pal and L. Wolfenstein, “[Radiative Decays of Massive Neutrinos]{},” [ *Phys. Rev. D*]{} [**25**]{} (1982) 766. Y. Hosotani, “[Majorana Masses, Photon Gas Heating and Cosmological Constraints on Neutrinos]{},” [*Nucl. Phys. B*]{} [**191**]{} (1981) 411. \[Erratum: Nucl. Phys. B 197, 546 (1982)\]. J. F. Nieves, P. B. Pal, and D. G. Unger, “[Photon Mass in a Background of Thermal Particles]{},” [*Phys. Rev. D*]{} [**28**]{} (1983) 908. Y. Chikashige, R. N. Mohapatra, and R. D. Peccei, “[Spontaneously Broken Lepton Number and Cosmological Constraints on the Neutrino Mass Spectrum]{},” [*Phys. Rev. Lett.*]{} [**45**]{} (1980) 1926. \[, 921 (1980)\]. G. B. Gelmini, S. Nussinov, and M. Roncadelli, “[Bounds and Prospects for the Majoron Model of Left-handed Neutrino Masses]{},” [*Nucl. Phys. B*]{} [ **209**]{} (1982) 157. R. Tomas, H. Pas, and J. W. F. Valle, “[Generalized bounds on Majoron - neutrino couplings]{},” [*Phys. Rev. D*]{} [**64**]{} (2001) 095005, [[hep-ph/0103017]{}](http://arxiv.org/abs/hep-ph/0103017). J. F. Beacom, N. F. Bell, D. Hooper, S. Pakvasa, and T. J. Weiler, “[Decay of High-Energy Astrophysical Neutrinos]{},” [*Phys. Rev. Lett.*]{} [**90**]{} (2003) 181301, [[ hep-ph/0211305]{}](http://arxiv.org/abs/hep-ph/0211305). P. Baerwald, M. Bustamante, and W. Winter, “[Neutrino Decays over Cosmological Distances and the Implications for Neutrino Telescopes]{},” [*JCAP*]{} [ **1210**]{} (2012) 020, [[1208.4600]{}](http://arxiv.org/abs/1208.4600). M. Bustamante, J. F. Beacom, and K. Murase, “[Testing decay of astrophysical neutrinos with incomplete information]{},” [*Phys. Rev. D*]{} [**95**]{} (2017) 063013, [[1610.02096]{}](http://arxiv.org/abs/1610.02096). P. B. Denton and I. Tamborra, “[Invisible Neutrino Decay Could Resolve IceCube’s Track and Cascade Tension]{},” [*Phys. Rev. Lett.*]{} [**121**]{} (2018) 121802, [[1805.05950]{}](http://arxiv.org/abs/1805.05950). J. F. Beacom, N. F. Bell, D. Hooper, J. G. Learned, S. Pakvasa, and T. J. Weiler, “[PseudoDirac neutrinos: A Challenge for neutrino telescopes]{},” [*Phys. Rev. Lett.*]{} [**92**]{} (2004) 011101, [[hep-ph/0307151]{}](http://arxiv.org/abs/hep-ph/0307151). A. S. Joshipura, S. Mohanty, and S. Pakvasa, “[Pseudo-Dirac neutrinos via a mirror world and depletion of ultrahigh energy neutrinos]{},” [*Phys. Rev. D*]{} [**89**]{} (2014) 033003, [[ 1307.5712]{}](http://arxiv.org/abs/1307.5712). J. L. Feng, “[Dark Matter Candidates from Particle Physics and Methods of Detection]{},” [*Ann. Rev. Astron. Astrophys.*]{} [**48**]{} (2010) 495, [[1003.0904]{}](http://arxiv.org/abs/1003.0904). J. F. Beacom, N. F. Bell, and G. D. Mack, “[General Upper Bound on the Dark Matter Total Annihilation Cross Section]{},” [*Phys. Rev. Lett.*]{} [**99**]{} (2007) 231301, [[ astro-ph/0608090]{}](http://arxiv.org/abs/astro-ph/0608090). H. Yuksel, S. Horiuchi, J. F. Beacom, and S. Ando, “[Neutrino Constraints on the Dark Matter Total Annihilation Cross Section]{},” [*Phys. Rev. D*]{} [**76**]{} (2007) 123506, [[ 0707.0196]{}](http://arxiv.org/abs/0707.0196). K. Murase and J. F. Beacom, “[Constraining Very Heavy Dark Matter Using Diffuse Backgrounds of Neutrinos and Cascaded Gamma Rays]{},” [*JCAP*]{} [ **1210**]{} (2012) 043, [[1206.2595]{}](http://arxiv.org/abs/1206.2595). Collaboration, S. Adrian-Martinez [*et. al.*]{}, “[Search of Dark Matter Annihilation in the Galactic Centre using the ANTARES Neutrino Telescope]{},” [*JCAP*]{} [**1510**]{} (2015) 068, [[1505.04866]{}](http://arxiv.org/abs/1505.04866). Collaboration, M. G. Aartsen [*et. al.*]{}, “[All-flavour Search for Neutrinos from Dark Matter Annihilations in the Milky Way with IceCube/DeepCore]{},” [*Eur. Phys. J. C*]{} [**76**]{} (2016) 531, [[1606.00209]{}](http://arxiv.org/abs/1606.00209). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Search for Neutrinos from Dark Matter Self-Annihilations in the center of the Milky Way with 3 years of IceCube/DeepCore]{},” [*Eur. Phys. J. C*]{} [**77**]{} (2017) 627, [[1705.08103]{}](http://arxiv.org/abs/1705.08103). Collaboration, M. G. Aartsen [*et. al.*]{}, “[IceCube Search for Dark Matter Annihilation in nearby Galaxies and Galaxy Clusters]{},” [ *Phys. Rev. D*]{} [**88**]{} (2013) 122001, [[1307.3473]{}](http://arxiv.org/abs/1307.3473). B. Feldstein, A. Kusenko, S. Matsumoto, and T. T. Yanagida, “[Neutrinos at IceCube from Heavy Decaying Dark Matter]{},” [*Phys. Rev. D*]{} [**88**]{} (2013) 015004, [[1303.7320]{}](http://arxiv.org/abs/1303.7320). A. Esmaili and P. D. Serpico, “[Are IceCube neutrinos unveiling PeV-scale decaying dark matter?]{},” [*JCAP*]{} [**1311**]{} (2013) 054, [[1308.1105]{}](http://arxiv.org/abs/1308.1105). T. Higaki, R. Kitano, and R. Sato, “[Neutrinoful Universe]{},” [*JHEP*]{} [ **07**]{} (2014) 044, [[1405.0013]{}](http://arxiv.org/abs/1405.0013). C. Rott, K. Kohri, and S. C. Park, “[Superheavy dark matter and IceCube neutrino signals: Bounds on decaying dark matter]{},” [*Phys. Rev. D*]{} [**92**]{} (2015) 023529, [[ 1408.4575]{}](http://arxiv.org/abs/1408.4575). E. Dudas, Y. Mambrini, and K. A. Olive, “[Monochromatic neutrinos generated by dark matter and the seesaw mechanism]{},” [*Phys. Rev. D*]{} [**91**]{} (2015) 075001, [[1412.3459]{}](http://arxiv.org/abs/1412.3459). Y. Ema, R. Jinno, and T. Moroi, “[Cosmic-Ray Neutrinos from the Decay of Long-Lived Particle and the Recent IceCube Result]{},” [*Phys. Lett. B*]{} [**733**]{} (2014) 120, [[1312.3501]{}](http://arxiv.org/abs/1312.3501). J. Zavala, “[Galactic PeV neutrinos from dark matter annihilation]{},” [ *Phys. Rev. D*]{} [**89**]{} (2014) 123516, [[1404.2932]{}](http://arxiv.org/abs/1404.2932). K. Murase, R. Laha, S. Ando, and M. Ahlers, “[Testing the Dark Matter Scenario for PeV Neutrinos Observed in IceCube]{},” [*Phys. Rev. Lett.*]{} [**115**]{} (2015) 071301, [[1503.04663]{}](http://arxiv.org/abs/1503.04663). L. A. Anchordoqui, V. Barger, H. Goldberg, X. Huang, D. Marfatia, L. H. M. da Silva, and T. J. Weiler, “[IceCube neutrinos, decaying dark matter, and the Hubble constant]{},” [*Phys. Rev. D*]{} [**92**]{} (2015) 061301, [[1506.08788]{}](http://arxiv.org/abs/1506.08788). \[Erratum: Phys. Rev. D 94, 069901 (2016)\]. S. M. Boucenna, M. Chianese, G. Mangano, G. Miele, S. Morisi, O. Pisanti, and E. Vitagliano, “[Decaying Leptophilic Dark Matter at IceCube]{},” [*JCAP*]{} [**1512**]{} (2015) 055, [[ 1507.01000]{}](http://arxiv.org/abs/1507.01000). P. S. B. Dev, D. Kazanas, R. N. Mohapatra, V. L. Teplitz, and Y. Zhang, “[Heavy right-handed neutrino dark matter and PeV neutrinos at IceCube]{},” [*JCAP*]{} [**1608**]{} (2016), no. 08 034, [[1606.04517]{}](http://arxiv.org/abs/1606.04517). N. Hiroshima, R. Kitano, K. Kohri, and K. Murase, “[High-energy neutrinos from multibody decaying dark matter]{},” [*Phys. Rev. D*]{} [**97**]{} (2018) 023006, [[1705.04419]{}](http://arxiv.org/abs/1705.04419). M. Chianese, G. Miele, and S. Morisi, “[Interpreting IceCube 6-year HESE data as an evidence for hundred TeV decaying Dark Matter]{},” [*Phys. Lett. B*]{} [**773**]{} (2017) 591, [[ 1707.05241]{}](http://arxiv.org/abs/1707.05241). C. S. Fong, H. Minakata, B. Panes, and R. Zukanovich Funchal, “[Possible Interpretations of IceCube High-Energy Neutrino Events]{},” [*JHEP*]{} [ **02**]{} (2015) 189, [[1411.5318]{}](http://arxiv.org/abs/1411.5318). T. Cohen, K. Murase, N. L. Rodd, B. R. Safdi, and Y. Soreq, “[Gamma-ray Constraints on Decaying Dark Matter and Implications for IceCube]{},” [ *Phys. Rev. Lett.*]{} [**119**]{} (2017) 021102, [[1612.05638]{}](http://arxiv.org/abs/1612.05638). Y. Sui and P. S. Bhupal Dev, “[A Combined Astrophysical and Dark Matter Interpretation of the IceCube HESE and Throughgoing Muon Events]{},” [ *JCAP*]{} [**1807**]{} (2018) 020, [[ 1804.04919]{}](http://arxiv.org/abs/1804.04919). A. Bhattacharya, M. H. Reno, and I. Sarcevic, “[Reconciling neutrino flux from heavy dark matter decay and recent events at IceCube]{},” [*JHEP*]{} [**06**]{} (2014) 110, [[1403.1862]{}](http://arxiv.org/abs/1403.1862). A. Bhattacharya, A. Esmaili, S. Palomares-Ruiz, and I. Sarcevic, “[Probing decaying heavy dark matter with the 4-year IceCube HESE data]{},” [*JCAP*]{} [**1707**]{} (2017) 027, [[ 1706.05746]{}](http://arxiv.org/abs/1706.05746). Y. Bai, R. Lu, and J. Salvado, “[Geometric Compatibility of IceCube TeV-PeV Neutrino Excess and its Galactic Dark Matter Origin]{},” [*JHEP*]{} [**01**]{} (2016) 161, [[1311.5864]{}](http://arxiv.org/abs/1311.5864). C. A. Arg[ü]{}elles, A. Kheirandish, and A. C. Vincent, “[Imaging Galactic Dark Matter with High-Energy Cosmic Neutrinos]{},” [*Phys. Rev. Lett.*]{} [**119**]{} (2017) 201801, [[ 1703.00451]{}](http://arxiv.org/abs/1703.00451). J. Lykken, O. Mena, and S. Razzaque, “[Ultrahigh-energy neutrino flux as a probe of large extra-dimensions]{},” [*JCAP*]{} [**0712**]{} (2007) 015, [[0705.2029]{}](http://arxiv.org/abs/0705.2029). K. Ioka and K. Murase, “[IceCube PeV–EeV neutrinos and secret interactions of neutrinos]{},” [*PTEP*]{} [**2014**]{} (2014) 061E01, [[1404.2279]{}](http://arxiv.org/abs/1404.2279). K. C. Y. Ng and J. F. Beacom, “[Cosmic neutrino cascades from secret neutrino interactions]{},” [*Phys. Rev. D*]{} [**90**]{} (2014) 065035, [[1404.2288]{}](http://arxiv.org/abs/1404.2288). \[Erratum: Phys. Rev. D 90, 089904 (2014)\]. K. Blum, A. Hook, and K. Murase, “[High energy neutrino telescopes as a probe of the neutrino mass mechanism]{},” [[ 1408.3799]{}](http://arxiv.org/abs/1408.3799). W. Altmannshofer, C.-Y. Chen, P. S. Bhupal Dev, and A. Soni, “[Lepton flavor violating Z’ explanation of the muon anomalous magnetic moment]{},” [ *Phys. Lett. B*]{} [**762**]{} (2016) 389, [[1607.06832]{}](http://arxiv.org/abs/1607.06832). G. Barenboim, P. B. Denton, and I. M. Oldengott, “[Constraints on inflation with an extended neutrino sector]{},” [*Phys. Rev. D*]{} [**99**]{} (2019) 083515, [[1903.02036]{}](http://arxiv.org/abs/1903.02036). M. Bustamante and S. K. Agarwalla, “[Universe’s Worth of Electrons to Probe Long-Range Interactions of High-Energy Astrophysical Neutrinos]{},” [ *Phys. Rev. Lett.*]{} [**122**]{} (2019) 061103, [[1808.02042]{}](http://arxiv.org/abs/1808.02042). F. Capozzi, I. M. Shoemaker, and L. Vecchi, “[Neutrino Oscillations in Dark Backgrounds]{},” [*JCAP*]{} [**1807**]{} (2018) 004, [[1804.05117]{}](http://arxiv.org/abs/1804.05117). L. Anchordoqui, V. Barger, H. Goldberg, and D. Marfatia, “[Phase transition in the fine structure constant]{},” [*Phys. Lett. B*]{} [**660**]{} (2008) 529, [[0711.4055]{}](http://arxiv.org/abs/0711.4055). N. Klop and S. Ando, “[Effects of a neutrino-dark energy coupling on oscillations of high-energy neutrinos]{},” [*Phys. Rev. D*]{} [**97**]{} (2018) 063006, [[1712.05413]{}](http://arxiv.org/abs/1712.05413). A. Connolly, R. S. Thorne, and D. Waters, “[Calculation of High Energy Neutrino-Nucleon Cross Sections and Uncertainties Using the MSTW Parton Distribution Functions and Implications for Future Experiments]{},” [ *Phys. Rev. D*]{} [**83**]{} (2011) 113009, [[1102.0691]{}](http://arxiv.org/abs/1102.0691). R. Gandhi, C. Quigg, M. H. Reno, and I. Sarcevic, “[Neutrino interactions at ultrahigh energies]{},” [*Phys. Rev. D*]{} [**58**]{} (1998) 093009, [[hep-ph/9807264]{}](http://arxiv.org/abs/hep-ph/9807264). M. M. Block, L. Durand, and P. Ha, “[Connection of the virtual $\gamma^*p$ cross section of ep deep inelastic scattering to real $\gamma p$ scattering, and the implications for $\nu N$ and $ep$ total cross sections]{},” [ *Phys. Rev. D*]{} [**89**]{} (2014) 094027, [[1404.4530]{}](http://arxiv.org/abs/1404.4530). C. A. Arg[ü]{}elles, F. Halzen, L. Wille, M. Kroll, and M. H. Reno, “[High-energy behavior of photon, neutrino, and proton cross sections]{},” [*Phys. Rev. D*]{} [**92**]{} (2015) 074040, [[1504.06639]{}](http://arxiv.org/abs/1504.06639). C. Arg[ü]{}elles, M. Bustamante, J. Conrad, A. Kheirandish, S. Palomares-Ruiz, J. Salvad[ó]{}, and A. Vincent, “[A Roadmap for New Physics Searches Using km$^3$-scale Neutrino Detectors]{},” [*In preparation*]{} (2019). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Energy Reconstruction Methods in the IceCube Neutrino Telescope]{},” [*JINST*]{} [ **9**]{} (2014) P03009, [[1311.4767]{}](http://arxiv.org/abs/1311.4767). K. Murase, M. Ahlers, and B. C. Lacki, “[Testing the Hadronuclear Origin of PeV Neutrinos Observed with IceCube]{},” [*Phys. Rev. D*]{} [**88**]{} (2013) 121301, [[1306.3417]{}](http://arxiv.org/abs/1306.3417). C.-Y. Chen, P. S. Bhupal Dev, and A. Soni, “[Standard model explanation of the ultrahigh energy neutrino events at IceCube]{},” [*Phys. Rev. D*]{} [**89**]{} (2014) 033012, [[1309.1764]{}](http://arxiv.org/abs/1309.1764). C.-Y. Chen, P. S. Bhupal Dev, and A. Soni, “[Two-component flux explanation for the high energy neutrino events at IceCube]{},” [*Phys. Rev. D*]{} [ **92**]{} (2015) 073001, [[1411.5658]{}](http://arxiv.org/abs/1411.5658). Collaboration, M. G. Aartsen [*et. al.*]{}, “[A combined maximum-likelihood analysis of the high-energy astrophysical neutrino flux measured with IceCube]{},” [*Astrophys. J.*]{} [**809**]{} (2015) 98, [[1507.03991]{}](http://arxiv.org/abs/1507.03991). L. A. Anchordoqui, M. M. Block, L. Durand, P. Ha, J. F. Soriano, and T. J. Weiler, “[Evidence for a break in the spectrum of astrophysical neutrinos]{},” [*Phys. Rev. D*]{} [**95**]{} (2017) 083009, [[1611.07905]{}](http://arxiv.org/abs/1611.07905). A. C. Vincent, S. Palomares-Ruiz, and O. Mena, “[Analysis of the 4-year IceCube high-energy starting events]{},” [*Phys. Rev. D*]{} [**94**]{} (2016) 023009, [[1605.01556]{}](http://arxiv.org/abs/1605.01556). V. Berezinsky and G. Zatsepin, “[Cosmic rays at ultrahigh-energies (neutrino?)]{},” [*Phys. Lett. B*]{} [**28**]{} (1969) 423. V. S. Berezinsky and A. [Yu]{}. Smirnov, “[Cosmic neutrinos of ultra-high energies and detection possibility]{},” [*Astrophys. Space Sci.*]{} [**32**]{} (1975) 461. F. W. Stecker, “[Diffuse Fluxes of Cosmic High-Energy Neutrinos]{},” [ *Astrophys. J.*]{} [**228**]{} (1979) 919. C. T. Hill and D. N. Schramm, “[Ultrahigh-Energy Cosmic Ray Neutrinos]{},” [ *Phys. Lett. B*]{} [**131**]{} (1983) 247. \[, 495 (1983)\]. S. Yoshida and M. Teshima, “[Energy spectrum of ultrahigh-energy cosmic rays with extragalactic origin]{},” [*Prog. Theor. Phys.*]{} [**89**]{} (1993) 833. R. Engel, D. Seckel, and T. Stanev, “[Neutrinos from propagation of ultrahigh-energy protons]{},” [*Phys. Rev. D*]{} [**64**]{} (2001) 093010, [[astro-ph/0101216]{}](http://arxiv.org/abs/astro-ph/0101216). L. A. Anchordoqui, H. Goldberg, D. Hooper, S. Sarkar, and A. M. Taylor, “[Predictions for the Cosmogenic Neutrino Flux in Light of New Data from the Pierre Auger Observatory]{},” [*Phys. Rev. D*]{} [**76**]{} (2007) 123008, [[0709.0734]{}](http://arxiv.org/abs/0709.0734). H. Takami, K. Murase, S. Nagataki, and K. Sato, “[Cosmogenic neutrinos as a probe of the transition from Galactic to extragalactic cosmic rays]{},” [ *Astropart. Phys.*]{} [**31**]{} (2009) 201, [[0704.0979]{}](http://arxiv.org/abs/0704.0979). M. Ahlers, L. A. Anchordoqui, and S. Sarkar, “[Neutrino diagnostics of ultra-high energy cosmic ray protons]{},” [*Phys. Rev. D*]{} [**79**]{} (2009) 083009, [[0902.3993]{}](http://arxiv.org/abs/0902.3993). M. Ahlers, L. A. Anchordoqui, M. C. Gonzalez-Garcia, F. Halzen, and S. Sarkar, “[GZK Neutrinos after the Fermi-LAT Diffuse Photon Flux Measurement]{},” [ *Astropart. Phys.*]{} [**34**]{} (2010) 106, [[1005.2620]{}](http://arxiv.org/abs/1005.2620). K. Kotera, D. Allard, and A. V. Olinto, “[Cosmogenic Neutrinos: parameter space and detectabilty from PeV to ZeV]{},” [*JCAP*]{} [**1010**]{} (2010) 013, [[1009.1382]{}](http://arxiv.org/abs/1009.1382). S. Yoshida and A. Ishihara, “[Constraints on the origin of the ultra-high energy cosmic-rays using cosmic diffuse neutrino flux limits: An analytical approach]{},” [*Phys. Rev. D*]{} [**85**]{} (2012) 063002, [[1202.3522]{}](http://arxiv.org/abs/1202.3522). M. Ahlers and F. Halzen, “[Minimal Cosmogenic Neutrinos]{},” [*Phys. Rev. D*]{} [**86**]{} (2012) 083010, [[ 1208.4181]{}](http://arxiv.org/abs/1208.4181). R. Aloisio, D. Boncioli, A. di Matteo, A. F. Grillo, S. Petrera, and F. Salamida, “[Cosmogenic neutrinos and ultra-high energy cosmic ray models]{},” [*JCAP*]{} [**1510**]{} (2015) 006, [[1505.04020]{}](http://arxiv.org/abs/1505.04020). J. Heinze, D. Boncioli, M. Bustamante, and W. Winter, “[Cosmogenic Neutrinos Challenge the Cosmic Ray Proton Dip Model]{},” [*Astrophys. J.*]{} [**825**]{} (2016) 122, [[1512.05988]{}](http://arxiv.org/abs/1512.05988). A. Romero-Wolf and M. Ave, “[Bayesian Inference Constraints on Astrophysical Production of Ultra-high Energy Cosmic Rays and Cosmogenic Neutrino Flux Predictions]{},” [*JCAP*]{} [**1807**]{} (2018) 025, [[1712.07290]{}](http://arxiv.org/abs/1712.07290). R. Alves Batista, R. M. de Almeida, B. Lago, and K. Kotera, “[Cosmogenic photon and neutrino fluxes in the Auger era]{},” [*JCAP*]{} [**1901**]{} (2019) 002, [[1806.10879]{}](http://arxiv.org/abs/1806.10879). K. Møller, P. B. Denton, and I. Tamborra, “[Cosmogenic Neutrinos Through the GRAND Lens Unveil the Nature of Cosmic Accelerators]{},” [*JCAP*]{} [**1905**]{} (2019) 047, [[1809.04866]{}](http://arxiv.org/abs/1809.04866). A. van Vliet, R. Alves Batista, and J. R. H[ö]{}randel, “[Determining the fraction of cosmic-ray protons at ultra-high energies with cosmogenic neutrinos]{},” [[1901.01899]{}](http://arxiv.org/abs/1901.01899). J. Heinze, A. Fedynitch, D. Boncioli, and W. Winter, “[A new view on Auger data and cosmogenic neutrinos in light of different nuclear disintegration and air-shower models]{},” [*Astrophys. J.*]{} [**873**]{} (2019) 88, [[1901.03338]{}](http://arxiv.org/abs/1901.03338). J. F. Cherry, A. Friedland, and I. M. Shoemaker, “[Neutrino Portal Dark Matter: From Dwarf Galaxies to IceCube]{},” [[1411.1071]{}](http://arxiv.org/abs/1411.1071). S. Pakvasa, “[Neutrino Flavor Goniometry by High Energy Astrophysical Beams]{},” [*Mod. Phys. Lett. A*]{} [**23**]{} (2008) 1313, [[0803.1701]{}](http://arxiv.org/abs/0803.1701). A. Palladino, C. Mascaretti, and F. Vissani, “[The importance of observing astrophysical tau neutrinos]{},” [*JCAP*]{} [**1808**]{} (2018) 004, [[1804.04965]{}](http://arxiv.org/abs/1804.04965). S. Parke and M. Ross-Lonergan, “[Unitarity and the three flavor neutrino mixing matrix]{},” [*Phys. Rev. D*]{} [**93**]{} (2016) 113009, [[1508.05095]{}](http://arxiv.org/abs/1508.05095). G. Barenboim and C. Quigg, “Neutrino observatories can characterize cosmic sources and neutrino properties,” [*Phys. Rev. D*]{} [**67**]{} (2003) 073024, [[hep-ph/0301220]{}](http://arxiv.org/abs/hep-ph/0301220). T. Kashti and E. Waxman, “[Flavoring astrophysical neutrinos: Flavor ratios depend on energy]{},” [*Phys. Rev. Lett.*]{} [**95**]{} (2005) 181101, [[astro-ph/0507599]{}](http://arxiv.org/abs/astro-ph/0507599). P. Lipari, M. Lusignoli, and D. Meloni, “[Flavor Composition and Energy Spectrum of Astrophysical Neutrinos]{},” [*Phys. Rev. D*]{} [**75**]{} (2007) 123005, [[0704.0718]{}](http://arxiv.org/abs/0704.0718). Collaboration, M. G. Aartsen [*et. al.*]{}, “[Flavor Ratio of Astrophysical Neutrinos above 35 TeV in IceCube]{},” [*Phys. Rev. Lett.*]{} [**114**]{} (2015) 171102, [[ 1502.03376]{}](http://arxiv.org/abs/1502.03376). G. Amelino-Camelia, J. Kowalski-Glikman, G. Mandanici, and A. Procaccini, “[Phenomenology of doubly special relativity]{},” [*Int. J. Mod. Phys. A*]{} [**20**]{} (2005) 6007, [[ gr-qc/0312124]{}](http://arxiv.org/abs/gr-qc/0312124). J. Christian, “[Testing quantum gravity via cosmogenic neutrino oscillations]{},” [*Phys. Rev. D*]{} [**71**]{} (2005) 024012, [[gr-qc/0409077]{}](http://arxiv.org/abs/gr-qc/0409077). J. S. Diaz, “[Neutrinos as probes of Lorentz invariance]{},” [*Adv. High Energy Phys.*]{} [**2014**]{} (2014) 962410, [[1406.6838]{}](http://arxiv.org/abs/1406.6838). J. S. Diaz, “[Testing Lorentz and CPT invariance with neutrinos]{},” [ *Symmetry*]{} [**8**]{} (2016) 105, [[ 1609.09474]{}](http://arxiv.org/abs/1609.09474). K. Murase and I. M. Shoemaker, “[Neutrino Echoes from Multimessenger Transient Sources]{},” [[1903.08607]{}](http://arxiv.org/abs/1903.08607). M. J. Longo, “[Tests of Relativity From SN1987A]{},” [*Phys. Rev. D*]{} [ **36**]{} (1987) 3276. Z.-Y. Wang, R.-Y. Liu, and X.-Y. Wang, “[Testing the equivalence principle and Lorentz invariance with PeV neutrinos from blazar flares]{},” [*Phys. Rev. Lett.*]{} [**116**]{} (2016) 151101, [[1602.06805]{}](http://arxiv.org/abs/1602.06805). J.-J. Wei, X.-F. Wu, H. Gao, and P. Mészáros, “[Limits on the Neutrino Velocity, Lorentz Invariance, and the Weak Equivalence Principle with TeV Neutrinos from Gamma-Ray Bursts]{},” [*JCAP*]{} [**1608**]{} (2016) 031, [[1603.07568]{}](http://arxiv.org/abs/1603.07568). S. Boran, S. Desai, and E. O. Kahya, “[Constraints on differential Shapiro delay between neutrinos and photons from IceCube-170922A]{},” [*Eur. Phys. J. C*]{} [**79**]{} (2019) 185, [[1807.05201]{}](http://arxiv.org/abs/1807.05201). B. Baret [*et. al.*]{}, “[Bounding the Time Delay between High-energy Neutrinos and Gravitational-wave Transients from Gamma-ray Bursts]{},” [ *Astropart. Phys.*]{} [**35**]{} (2011) 1, [[1101.4669]{}](http://arxiv.org/abs/1101.4669). Collaboration, M. G. Aartsen [*et. al.*]{}, “[The IceCube Neutrino Observatory: Instrumentation and Online Systems]{},” [*JINST*]{} [ **12**]{} (2017) P03012, [[1612.05093]{}](http://arxiv.org/abs/1612.05093). Collaboration, M. G. Aartsen [*et. al.*]{}, “[IceCube-Gen2: A Vision for the Future of Neutrino Astronomy in Antarctica]{},” [[1412.5106]{}](http://arxiv.org/abs/1412.5106). L. A. Anchordoqui, C. Garcia Canal, and J. F. Soriano, “[Probing strong dynamics with cosmic neutrinos]{},” [[1902.10134]{}](http://arxiv.org/abs/1902.10134). S. L. Glashow, “[Resonant Scattering of Antineutrinos]{},” [*Phys. Rev.*]{} [**118**]{} (1960) 316. S. W. Li, M. Bustamante, and J. F. Beacom, “[Echo Technique to Distinguish Flavors of Astrophysical Neutrinos]{},” [*Phys. Rev. Lett.*]{} [**122**]{} (2019) 151101, [[1606.06290]{}](http://arxiv.org/abs/1606.06290). S.-H. Wang, P. Chen, M. Huang, and J. Nam, “[Feasibility of Determining Diffuse Ultra-High Energy Cosmic Neutrino Flavor Ratio through ARA Neutrino Observatory]{},” [*JCAP*]{} [**1311**]{} (2013) 062, [[1302.1586]{}](http://arxiv.org/abs/1302.1586). D. Fargion, A. Aiello, and R. Conversano, “[Horizontal tau air showers from mountains in deep valley: Traces of UHECR neutrino tau]{},” in [ *[Proceedings, 26th International Cosmic Ray Conference (ICRC), August 17-25, 1999, Salt Lake City: Invited, Rapporteur, and Highlight Papers]{}*]{}, p. 396, 1999. [[astro-ph/9906450]{}](http://arxiv.org/abs/astro-ph/9906450). A. V. Olinto [*et. al.*]{}, “[POEMMA: Probe Of Extreme Multi-Messenger Astrophysics]{},” [*PoS*]{} [**ICRC2017**]{} (2018) 542, [[1708.07599]{}](http://arxiv.org/abs/1708.07599). \[35, 542 (2017)\]. M. Sasaki and T. Kifune, “[Ashra Neutrino Telescope Array (NTA): Combined Imaging Observation of Astroparticles – For Clear Identification of Cosmic Accelerators and Fundamental Physics Using Cosmic Beams – ]{},” [*JPS Conf. Proc.*]{} [**15**]{} (2017) 011013. S. Wissel [*et. al.*]{}, “[A New Concept for High-Elevation Radio Detection of Tau Neutrinos]{},” [*accepted EPJ Web Conf.*]{} (2018). Presented at the Acoustic and Radio EeV Neutrino Detection Activities Conference 2018 (ARENA 2018), Catania, Italy, June 12–15, 2018 [ https://indico.cern.ch/event/667036/contributions/3005761/]{}. Collaboration, T. Liu, “[The status of the second station of Taiwan Astroparticle Radiowave Observatory for Geo-synchrotron Emissions (TAROGE-II)]{},” [*PoS*]{} [**ICRC2017**]{} (2018) 234. Collaboration, J. Alvarez-Muñiz [*et. al.*]{}, “[The Giant Radio Array for Neutrino Detection (GRAND): Science and Design]{},” [[1810.09994]{}](http://arxiv.org/abs/1810.09994). A. N. Otte, “[Studies of an air-shower imaging system for the detection of ultrahigh-energy neutrinos]{},” [*Phys. Rev. D*]{} [**99**]{} (2019) 083012, [[1811.09287]{}](http://arxiv.org/abs/1811.09287). P. Allison [*et. al.*]{}, “[Design and Initial Performance of the Askaryan Radio Array Prototype EeV Neutrino Detector at the South Pole]{},” [ *Astropart. Phys.*]{} [**35**]{} (2012) 457, [[1105.2854]{}](http://arxiv.org/abs/1105.2854). S. W. Barwick [*et. al.*]{}, “[Design and Performance of the ARIANNA HRA-3 Neutrino Detector Systems]{},” [*IEEE Trans. Nucl. Sci.*]{} [**62**]{} (2015) 2202, [[1410.7369]{}](http://arxiv.org/abs/1410.7369). Collaboration, P. W. Gorham [*et. al.*]{}, “[The Antarctic Impulsive Transient Antenna Ultra-high Energy Neutrino Detector Design, Performance, and Sensitivity for 2006-2007 Balloon Flight]{},” [ *Astropart. Phys.*]{} [**32**]{} (2009) 10, [[0812.1920]{}](http://arxiv.org/abs/0812.1920). Collaboration, S. Adrian-Martinez [*et. al.*]{}, “[Letter of intent for KM3NeT 2.0]{},” [*J. Phys. G*]{} [**43**]{} (2016) 084001, [[1601.07459]{}](http://arxiv.org/abs/1601.07459). C. W. James [*et. al.*]{}, “[Overview of lunar detection of ultra-high energy particles and new plans for the SKA]{},” [*EPJ Web Conf.*]{} [**135**]{} (2017) 04001, [[1704.05336]{}](http://arxiv.org/abs/1704.05336). Collaboration, R. Abbasi [*et. al.*]{}, “[Search for a Lorentz-violating sidereal signal with atmospheric neutrinos in IceCube]{},” [*Phys. Rev. D*]{} [**82**]{} (2010) 112003, [[1010.4096]{}](http://arxiv.org/abs/1010.4096).

ArXiv

--- abstract: | We investigate the formation and evolution of dwarf galaxies in a high resolution, hydrodynamical cosmological simulation of a Milky Way sized halo and its environment. Our simulation includes gas cooling, star formation, supernova feedback, metal enrichment and UV heating. In total, 90 satellites and more than 400 isolated dwarf galaxies are formed in the simulation, allowing a systematic study of the internal and environmental processes that determine their evolution. We find that $95\%$ of satellite galaxies are gas-free at $z=0$, and identify three mechanisms for gas loss: supernova feedback, tidal stripping, and photo-evaporation due to re-ionization. Gas-rich satellite galaxies are only found with total masses above $\sim5\times 10^9{\mathrm{M}_\odot}$. In contrast, for isolated dwarf galaxies, a total mass of $\sim10^9{\mathrm{M}_\odot}$ constitutes a sharp transition; less massive galaxies are predominantly gas-free at $z=0$, more massive, isolated dwarf galaxies are often able to retain their gas. In general, we find that the total mass of a dwarf galaxy is the main factor which determines its star formation, metal enrichment, and its gas content, but that stripping may explain the observed difference in gas content between field dwarf galaxies and satellites with total masses close to $10^9{\mathrm{M}_\odot}$. We also find that a morphological transformation via tidal stripping of infalling, luminous dwarf galaxies whose dark matter is less concentrated than their stars, cannot explain the high total mass-light ratios of the faint dwarf spheroidal galaxies. author: - | Till Sawala$^{1,3}$[^1], Cecilia Scannapieco$^{2}$, and Simon White$^{3}$\ $^{1}$Institute for Computational Cosmology, Department of Physics, University of Durham, South Road, Durham DH1 3LE, UK\ $^{2}$Astrophysikalisches Institut Potsdam, An der Sternwarte 16, 14482 Potsdam, Germany\ $^{3}$Max-Planck Institute for Astrophysics, Karl-Schwarzschild-Strasse 1, 85748 Garching, Germany bibliography: - 'aquila.bib' date: 'Accepted \*\*\*. Received \*\*\*; in original form 09 March 2011' title: 'Local Group Dwarf Galaxies: Nature [*And*]{} Nurture' --- \[firstpage\] cosmology: theory – galaxies: dwarf – galaxies: Local Group – galaxies: formation – galaxies: evolution – methods: N-body simulations Introduction ============ The Milky Way galaxy is believed to be surrounded by hundreds of dwarf galaxies [@Koposov-2008; @Tollerud-2008]. Most of the known satellites are faint, early type dwarf spheroidal galaxies, with stellar masses ranging from less than $10^3$ up to $10^7{\mathrm{M}_\odot}$, predominantly old and metal-poor stellar populations, no detectable interstellar gas, and eponymous, spheroidal morphologies. Stellar kinematics suggest that dwarf spheroidal galaxies are dominated by dark matter [@Faber-Lin-1983], with estimates for the total dynamical masses in the range of $\sim 2\times10^8$  – $10^9{\mathrm{M}_\odot}$ [@Walker-2007]. Dwarf galaxies of such low total mass are believed to be highly susceptible to both external and internal effects: Supernova feedback can establish a self-regulation of star formation, and also lead to an ejection of gas, while the interaction with the environment, and with neighbouring galaxies, can potentially strip their interstellar gas, and may also change their morphology. The effect of the cosmological UV-background, whilst external, is assumed to be independent of environment, and therefore identical for satellites and for isolated dwarf galaxies. Since the correlation of morphological type with proximity to larger galaxies was first demonstrated by [@Einasto-1974], the role of [*Nature*]{} versus [*Nurture*]{} in the formation of early-type dwarf galaxies has been the cause of debate [e.g. @Mateo-1998; @Grebel-2003; @Tolstoy-2009; @Annibali-2010]. Evidence for environmental effects on galaxy formation in general is well known, and expressed in the morphology-density relation [e.g. @Davis-1976]. Two studies comparing early type dwarf galaxies in the field to early type dwarf galaxies in clusters ([@Michielsen-2008] for the Virgo cluster and [@Koleva-2009] for the Fornax cluster), found no clear difference between the environments. However, comparing Coma cluster dwarf galaxies to dwarf galaxies in poor groups, [@Annibali-2010] reported that dwarf galaxies in low-density environments may experience more prolonged star formation; evidence that the morphology-density relation also extends to dwarf galaxies. It should also be emphasized that even if the environmental correlation is weak, external effects could still play an important role if, for dwarfs, the responsible mechanisms are also efficient in low-density environments. In a sample of SDSS galaxies with stellar masses in the range of $2\times 10^7-3\times10^8{\mathrm{M}_\odot}$, [@Geha-2006] report that the gas fraction of dwarf galaxies increases with distance to their nearest luminous neighbour. Within the Local Group, [@Einasto-1974] show that the HI mass in dwarf galaxies follow a similar trend with distance to M31 or the Milky Way. [@Grebel-2003] also report a morphology-distance relation: late-type, dwarf irregular galaxies are found predominantly at the outskirts of the Local Group; early type dwarf galaxies are mostly satellites; so-called transition type galaxies, which share some of the characteristics of early type dwarf galaxies whilst also containing some gas, are found at intermediate distances. [@Grebel-1997] also find a correlation between galactocentric distance and duration of star formation in Milky Way satellites. On the other hand, there is also evidence for supernova-driven outflows in starburst dwarf galaxies [e.g. @Meurer-1992; @Schwartz-2004]. [@Sanchez-2010] conclude that low-mass galaxies tend to be less strongly affected by tidal forces, and that the morphological trend of increasing thickness found in fainter galaxies, is related to the increasing importance of feedback mechanisms, rather than environmental effects. Future studies of isolated dwarf galaxies in the Local Group may answer the question whether internal processes suffice to expel gas from dwarf galaxies, or whether interactions are a necessary factor. Two isolated dwarf spheroidal galaxies, Cetus [@Whiting-1999] and Tucana [@Lavery-1992], have been discovered in the Local Group, with distances of $780 \pm 40$ kpc and $890\pm50$ kpc, respectively [@Bernard-2009]. Deep observations and modeling of the stellar populations by [@Monelli-2010] suggests that these presently isolated objects are similar to the oldest Milky Way dSph satellites; this runs counter to the morphology density relation and suggests a formation mechanism independent of environment. However, it is not known with certainty whether present isolation also implies isolated evolution. Numerical simulations of isolated dwarf galaxies [e.g. @Stinson-2007; @Valcke-2008; @Revaz-2009; @Sawala-2010] have shown that dark matter-dominated, early-type dwarf galaxies comparable to the Local Group dwarf spheroidals [*can*]{} form in isolation through the processes of supernova feedback and UV radiation, which regulate and eventually quench star formation, and eject most of the interstellar gas. In particular, this work showed that the dependence of the efficiency of the two effects on the depth of the gravitational potential can explain the observed scaling relations between stellar mass and mass-to-light ratio [e.g. @Mateo-1998; @Gilmore-2007; @Strigari-2008], and between stellar mass and metallicity [@Grebel-2003; @Tolstoy-2009]. Simulations of interacting dwarfs [@Mayer-2001; @Mayer-2006; @Klimentowski-2009] have indicated that evolved, disk-like galaxies can also be transformed into fainter, early-type dwarf galaxies in the environment of the Milky-Way. Two mechanisms that can lead to such a transformation are ram pressure and tidal interactions. Ram pressure refers to the pressure exerted on the interstellar gas of a satellite galaxy as it passes through the hot intergalactic medium in a dense environment [[@Einasto-1974]]{}. Ram-pressure stripping affects the interstellar gas, but not the stars or dark matter. While this mechanism is observable in galaxy clusters, estimates of the gas density of the Milky-Way halo [e.g. @Murali-2000] show it to be several orders of magnitude too low for ram pressure stripping to be efficient [@Mayer-2001]. Tidal perturbations are caused by the differential gravitational acceleration across the diameter of a satellite as it orbits the main halo. They affect gas, stars and dark matter alike, and can alter the internal kinematics and morphology of the satellite (“tidal distortion”), as well as remove mass from the objects (“tidal stripping”). Tidal stripping can still change the composition of objects, if the spatial distribution of the components differ. Clear observational evidence for tidal stripping within the Milky-Way halo is provided by the presence of tidal streams [e.g. @Johnston-1999 now also observed around other galaxies]. These elongated substructures consist of stars that were tidally unbound from satellite galaxies, and deposited along their orbits. Most present day dwarf spheroidals, however, show no sign of tidal distortion in their stellar kinematics [@Walker-2007]. Models in which similar-mass late type progenitors are transformed into different-mass dwarf spheroidals based on orbital parameters alone, may have also difficulty in explaining the strong-mass metallicity scaling relation [@Grebel-2003]. Most dwarf spheroidals also show metallicity gradients, which are reproduced in simulations where star formation is regulated by internal processes [@Revaz-2009], but are not typically found in dwarf irregulars, and may not be preserved during strong tidal stirring. So far, simulations have largely separated internal and external mechanisms: by investigating either the formation of isolated dwarf galaxies [@Stinson-2007; @Valcke-2008; @Revaz-2009; @Sawala-2010], or the transformation of evolved objects in an external gravitational potential [@Mayer-2001; @Mayer-2006; @Klimentowski-2009]. In reality, all effects are simultaneously present in the Local Group. In this work, we present results from the [*Aquila*]{} Simulation [@Scannapieco-2009], which follows the formation of a Milky Way sized galaxy and its environment in the fully cosmological context of a $\Lambda$CDM universe. The satellites that grow and evolve are thus subject to tidal forces, but also all the astrophysical processes associated with cooling, star formation, supernova feedback and UV heating. This simulation allows us to study all these effects in a consistent manner, and to compare their relative importance for the evolution of each object individually, as well as for the ensemble of the Milky Way satellites. This paper is organized as follows: Section \[aquila\_methods\] contains a description of the initial conditions for the simulation, the numerical methods, and the method for the identification of substructure. Section \[aquila\_evolution\] describes the formation and time-evolution of the halo along with its substructures. In Section \[aquila:gas\_loss\] we take a closer look at the different mechanisms for gas removal in individual satellites. Section \[aquila\_relations\] summarizes the statistical properties of the present-day satellite population and the derived scaling relations. In Section \[aquila\_outside\], we also compare the population of satellites to the population of isolated dwarf galaxies formed in the same simulation. We conclude with a summary in Section \[aquila\_summary\]. The Aquila Simulation {#aquila_methods} ===================== The simulation was set up with the principal aim of studying the formation of disk galaxies in a $\Lambda$CDM universe. The results presented here are based on simulation “AQ-C-5” of [@Scannapieco-2009], which resulted in the formation of a Milky-Way sized disk galaxy. In this paper, the focus is on the formation and evolution of the $\sim$90 satellites of the central galaxy, and the isolated dwarf galaxies that form in its environment. Initial Conditions ------------------ The initial conditions used here are based on one of several haloes (labeled halo “C”), which were extracted from the Millennium II Simulation, and resimulated with pure dark matter in the [ *Aquarius*]{} project [@Springel-2008]. The cosmological parameters are identical to those of the Millennium Simulations, $\Omega_m=0.25, \Omega_\Lambda = 0.75, h=0.73$ and $\sigma_8=0.9$, consistent with WMAP-1 cosmology. The simulation is performed with periodic boundary conditions in a box of side length 137 Mpc ($100 h^{-1}$ Mpc in comoving coordinates). The central Lagrangian region is filled with an equal number of high resolution dark matter and gas particles, at a mass ratio of $\Omega_{DM} = 0.21$ to $\Omega_b=0.04$. The particle masses in the level 5 [*Aquila*]{} simulation are $2.6\times10^6{\mathrm{M}_\odot}$ for dark matter and 2–4$\times10^5{\mathrm{M}_\odot}$ for gas particles. Star particles which form have a mass of 1–2$\times10^5{\mathrm{M}_\odot}$. The gravitational softening parameter is fixed to 1.37 (1$ h^{-1}$ kpc) in comoving coordinates, and equal for all particle types. Throughout this paper, masses and distances will be stated in physical units of ${\mathrm{M}_\odot}$ and kpc. Note that despite the limited accuracy of gravitational forces at the softening scale, tidal accelerations due to an external potential are almost unaffected by the softening. Computational Methods --------------------- The simulations are performed with the smoothed particle hydrodynamics code <span style="font-variant:small-caps;">GADGET-3</span> [@Springel-2005; @Springel-2008], together with the star formation, feedback and multiphase ISM model of [@Scannapieco-2005; @Scannapieco-2006]. The same model is also used in the higher resolution simulations of isolated dwarf galaxies presented in [@Sawala-2010; @Sawala-2011]. Specific to this simulation, a star formation parameter of $c_\star=0.1$ is used, whereas most simulations presented in [@Sawala-2010], and all simulations of [@Sawala-2011] use a value of $c_\star=0.05$. [@Sawala-2010] also contains a discussion of the effect of varying $c_\star$ for dwarf galaxies. The energy per supernova is set to $0.7 \times 10^{51}$ ergs, identical to the value of [@Sawala-2011], but slightly lower than the value of $1.0\times10^{51}$ ergs used in [@Sawala-2010]. The energy and metals are distributed equally between the hot and cold gas phases. The cooling model is based on [@Sutherland-1993], with no cooling below $10^4K$, and the UV heating mechanism is implemented following [@Haardt-1996], assuming a cosmic UV background active from $z=6$. The mass resolution of the [*Aquila*]{} simulation is several hundred times lower than can be achieved in simulations of individual dwarf galaxies. This reflects primarily the much larger Lagrangian volume, but the presence of a large galaxy with its high cold gas fraction and specific star formation rate also increases the computational cost per particle. As a result, the properties of individual dwarf galaxies are not resolved with the same level of detail as in simulations of isolated objects; dwarf galaxies with stellar masses of $\sim10^6 {\mathrm{M}_\odot}$ are clearly stretching the resolution limit. Still, the large number of satellites that form allow a number of statistical comparisons. In addition, the fact that, apart from the differences stated above, the same astrophysical and numerical models are used in the high resolution simulations of [@Sawala-2010; @Sawala-2011], allows a direct comparison and an estimation of the effects of resolution. We include the results of our high-resolution simulations in Figure \[fig:relation\_mdm-ms\]. The same initial conditions have also been used in other studies [@Okamoto-2009; @Okamoto-2010; @Wadepuhl-2010], with different physical astrophysical models and numerical codes. The star formation and feedback model of @Wadepuhl-2010 is based on [@Springel-2003], combined with a black hole wind model and a model for cosmic rays. They conclude that cosmic rays, or some other mechanism in addition to thermal supernova feedback and UV heating is necessary in order to bring the satellite luminosity function in agreement with observations. Our results agree qualitatively with their simulations, despite the fact that we do not include cosmic rays, indicating that supernova feedback produces similar effects in our ISM model to cosmic ray pressure in theirs. @Okamoto-2010 investigate how varying the treatment of feedback affects the properties of dwarfs. They show that stellar mass and metallicity scale strongly with circular velocity (i.e. subhalo mass), and conclude that a threshold established by re-ionization results in the small fraction of visible satellite galaxies populating a much larger number of satellite subhaloes. Neither study investigates the [*isolated*]{} dwarfs and compares them to the satellites; this is a major part of our work. Identification of Substructure {#aquila:substructure} ------------------------------ In each snapshot, haloes are identified in a two step process using a Friend-of-Friend (FoF) algorithm. The FoF groups are defined first by linking only the dark matter particles. In a second step, star and gas particles are linked to the particles already belonging to these groups in the same way. In the Aquila simulation, the larger haloes, and in particular, the main halo that will become the host of the “Milky Way” galaxy, also contain a number of gravitationally bound, over-dense substructures called [*subhaloes*]{}, which are identified using the [subfind]{} algorithm of [@Springel-2001]. For each FoF halo, [subfind]{} begins by computing a smoothed local density by an SPH-interpolation over all particles of the halo. A potential substructure is first identified as a local overdensity with respect to this smooth background density. It is then subjected to gravitational unbinding, whereby all unbound particles are iteratively removed, until the substructure either vanishes, falling below the threshold of 20 particles, or is identified as a genuine self-bound subhalo. In order to trace subhaloes over time between different snapshots, the 20 most bound dark matter particles of each subhalo in a given snapshot were compared to the list of particles in all subhaloes of the previous snapshot, identifying as the progenitor the subhalo that contained at least 11 of the 20 particles amongst its 20% of most bound particles. The process is repeated until a progenitor is no longer found, and we define this as the time of [*formation*]{}. If the subhalo can be linked to a subhalo in a previous snapshot, but not to its most bound particles, we consider the progenitor to have fragmented, and define this as the time of [ *fragmentation*]{}. Fragmentation can be due to the breaking-up of larger haloes, or to the amplification of substructures above the particle threshold, without the newly identified subhaloes hosting galaxies. In Table \[table:aquila-satellites\], fragmentation is indicated with an asterisk next to the formation redshift. We note that a large majority of subhaloes of satellite galaxies can be traced back as independent objects well beyond $z = 6$. Most of the subhaloes belonging to the main halo at $z=0$ belong to different groups at earlier times. We call the time when a subhalo is [*first*]{} identified as a subhalo of the main halo the time of [ *infall*]{}, noting that some subhaloes subsequently become transitorily isolated, and fall in again at a later time. -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- Time Evolution of the Aquila Simulation {#aquila_evolution} ======================================= Figures \[fig:aquila\_dm\], \[fig:aquila\_gas\] and \[fig:aquila\_stars\] respectively show the time-evolution of the projected dark matter, gas and stellar mass distributions in the central region of the simulation. Each panel is centred on the position of the main subhalo (which is the host of the “Milky Way”), and shows all particles within a cubic volume of side length 1 Mpc At each redshift, the $X,Y$ and $Z$-coordinates are defined parallel to the principal moments of the inertia tensor of the halo, with eigenvalues $I_x > I_y > I_z$. Keeping the volume fixed in comoving coordinates corresponds to a zoom-out in physical coordinates by a factor of 7.5 as the universe grows with time from $z=6.5$ to $z=0$. The squares indicate the position of all satellite subhaloes belonging to the main FoF halo identified at the time of the snapshot, with the size of the squares in all figures indicative of (but not strictly proportional to) the dark matter mass of the subhalo. In Figure \[fig:aquila\_gas\], where the blue colour indicates gas density, red boxes denote the subhaloes that contain gas, while white boxes denote subhaloes that are gas-free. It can be seen that even at high redshifts, the majority of subhaloes are gas-free, and only four relatively massive satellites contain gas at $z=0$. The smallest of these four satellites, which are the subject of Section \[aquila:massive\], has a total mass of $\sim 5\times10^9 {\mathrm{M}_\odot}$. All lower mass satellites, many of which formed stars, have lost their gas during their evolution. The different mechanisms of gas loss, internal and environmental, are discussed in Section \[aquila:gas\_loss\]. In Figure \[fig:aquila\_stars\], the green colour shows the stellar density, which is clearly dominated by the central object and its halo. Here, red boxes show subhaloes that contain stars, while white boxes show haloes that are essentially dark. At $z=0$, there are 90 subhaloes containing stars, including the four which contain gas, as shown in Figure \[fig:aquila\_gas\]. The highest mass subhalo that does not contain any stars at $z=0$ has a mass of $4.3\times10^8{\mathrm{M}_\odot}$. The lower mass limit for star formation becomes difficult to assess, because of the limited resolution of our simulation. The total number of subhaloes with stars is comparable, however, with observational estimates of the number of dwarf satellites around the Milky Way. The halo reaches a final virial mass of $1.6 \times 10^{12} {\mathrm{M}_\odot}$, comparable to recent observational estimates of the Milky Way halo, for example $10^{12}{\mathrm{M}_\odot}$ [@Xue-2008], $1.4\times10^{12}{\mathrm{M}_\odot}$ [@Smith-2007], $1.6\times10^{12}{\mathrm{M}_\odot}$ [@Gnedin-2010] and $2.4\times10^{12}{\mathrm{M}_\odot}$ [@Li-2008]. The corresponding spherical virial radius is $\sim 250$ kpc, but we include as satellites all subhaloes within the FoF group. 40% of the satellites are presently located outside of $r_{vir}$, with the furthest satellite at a galactocentric distance of $490$ kpc. The central galaxy reaches a stellar mass of $10.8 \times10^{10}{\mathrm{M}_\odot}$, higher than current observational estimates for the Milky Way, for example $5.5 \times10^{10}{\mathrm{M}_\odot}$ from [@Flynn-2006]. Distributions of the different mass components, and the positions of all satellites at $z=0$ are shown in three orthogonal projections along the principal axes of the inertia tensor of the halo in Figure \[fig:aquila\_projections\]. By comparison with the flattened distribution seen at high redshift, the final shape of the halo appears round and largely featureless. This transformation from a triaxial mass distribution, expected from purely gravitational assembly, to an oblate halo similar to that observed, is also studied in [@Tissera-2010], who attribute the difference to baryonic effects. Formation and Evolution of Satellites {#aquila:gas_loss} ===================================== When the environment is included, different mechanisms of gas loss can play a role, and all are observed in the simulation. Internal and external mechanisms often act simultaneously, and are not always easy to disentangle. Just as supernova heating aids in the gas removal by UV radiation [@Sawala-2010], the thermal expansion caused by the energy input can also make it easier for tidal interactions to remove gas. In section \[aquila:examples\], we discuss four exemplary cases of gas loss which are representative of the total subhalo population in terms of their final properties, but where the different mechanisms are relatively easily identified. Section \[aquila:extremes\] considers two extreme cases of subhaloes very heavily affected by stripping. Section \[aquila:massive\] describes four satellites that still contain gas at $z=0$, and contrasts them to the many gas-free satellites. Gas Loss by Example {#aquila:examples} ------------------- -- -- -- -- -- -- -- -- -- -- In the top panel of Figure \[fig:selected\_evolution\], we show the evolution of the three mass components, dark matter (black), gas (blue) and stellar mass (green) of 4 selected subhaloes, each representing a different evolutionary path. Also shown in the bottom panel of Figure \[fig:selected\_evolution\] is the distance from the satellite centre of mass to the main halo centre as a function of redshift. - First from the left in Figure \[fig:selected\_evolution\] is subhalo 14. Here, the gas is blown out by the combined effect of supernova feedback and UV heating while the dwarf halo is still in isolation. As it approaches the central halo, some gas is re-acquired at $z\sim3.5$, but this does not lead to renewed star formation. This residual gas is lost when the dwarf halo finally falls in to the central halo at $z=2.1$. As it spirals inwards on multiple orbits, the dark matter mass decreases from its peak of $2.4 \times 10^9{\mathrm{M}_\odot}$ to its $z=0$ value of $1.3\times10^9{\mathrm{M}_\odot}$. On each of the two final pericentre passages, the stellar mass also decreases slightly. - Second from the left, subhalo 23 loses its gas already at high redshift while still in isolation, in the same way as subhalo 14. It evolves passively and free of gas from $z=3$ onwards, first falling into the main halo at $z=1.53$. The dark matter mass and stellar mass are not significantly affected by tides between infall and $z=0$. - Subhalo 25, in the third column, keeps a significant amount of gas and continues to form stars up to its first infall at $z=0.96$. The ISM is stripped before it reaches pericentre for the first time and star formation ceases. The dark matter mass is subsequently reduced from its peak value of $4 \times 10^9 {\mathrm{M}_\odot}$ to $6.6 \times 10^8 {\mathrm{M}_\odot}$ on two close pericentre passages. The stellar mass also decreases, particularly during the second passage. - On the right, subhalo 56 is a peculiar case. Having lost some gas due to supernova feedback, the remainder is lost almost instantaneously at $z=6$, when the effect of the UV background sets in. As for subhalo 14, a small amount of gas gets re-accreted and is lost again as the subhalo approaches the central halo. The galaxy first becomes a satellite at $z=2.0$, but does not get close to the centre on its first orbit, resulting in no stripping of stars, and only a small reduction of the dark matter mass. The final two passages are much closer, and as a result, the dark matter and stellar mass are both significantly reduced. Three of the four objects have peak dark matter masses between $7 \times 10^8$ and $2.5 \times 10^9 {\mathrm{M}_\odot}$, with subhalo 56, which loses its gas due to UV heating, has a peak mass of $5 \times 10^8$. The final dark matter masses lie between $5.8 \times 10^8$ and $2 \times 10^8 {\mathrm{M}_\odot}$, with final stellar masses in the range of $2 \times 10^6$ to 4 $\times 10^7 {\mathrm{M}_\odot}$, and corresponding stellar mass - total mass ratios of 50-100. The four objects follow the overall scaling relations of Figures \[fig:relation\_mdm-ms\] to \[fig:ratio-infall-approach\], and are also similar to the isolated dwarf galaxies of [@Sawala-2010], except for higher final stellar mass - total mass ratios for those objects most strongly effected by stripping. Figure \[fig:aquila\_orbits\] shows the orbits of the four subhaloes described above in three different projections after z=1.9. While the four satellites are on similar orbits (in contrast to the two cases shown in Figure \[fig:extreme\]), it can be seen that only subhalo 25 (light blue curve), still has gas when it enters the main halo. The other three fall in gas-free. All four subhaloes are free of gas during most of their evolution as satellites. -- -- -- -- -- -- Two extreme cases {#aquila:extremes} ----------------- Two satellites, subhaloes 3 and 5, have final stellar masses that exceed their dark matter mass. They can be identified as outliers on the $M_\star$–$M_{DM}$ relation shown in Figure \[fig:relation\_mdm-ms\], and are off the chart in the $M_\star$/$M_{DM,z=0}$–$M_\star$/$M_{DM,infall}$ relations of Figure \[fig:ratio-infall-approach\]. This testifies to the fact that their dark matter masses decreased by a factor of 35 in the case of subhalo 5, and even $\sim 200$ for subhalo 3. Not surprisingly, the orbits of the two haloes, plotted in Figure \[fig:extreme\], both include a number of recent close pericentre passages. A difference in orbital shape is also apparent in comparison with Figure \[fig:aquila\_orbits\]. The dashed and solid line segments in Figure \[fig:extreme\] indicate the presence or absence of gas, respectively. It can be seen that both objects still have gas when they first fall into the main halo. Interestingly, they both go through a gas-free phase on their first pericentre passage, but both recollect gas twice during their subsequent evolution, only to lose it again at each pericentre passage. The two objects enter the halo with stellar mass – total mass ratios of $30$ to $40$ and properties similar to dwarf irregular galaxies, including gas to stellar mass ratios of order unity. They end up gas-free, with significantly reduced stellar masses, but also with final stellar mass – total mass ratios close to unity. Quantitatively, this result should be taken with caution, as the close pericentre passages make the evolution of the subhaloes also dependent on the detailed properties of the simulated central galaxy. In addition, some particles belonging to the outer parts of a satellite subhalo may be misattributed to the main halo instead, even though they remain gravitationally bound, and continue to move with the subhalo. However, qualitatively, this result suggests that strong tidal stripping [*decreases*]{}, rather than increases, the total mass-to-light ratio of satellite galaxies, and is therefore not a viable way to transform gas-rich, bright dwarf irregulars into the gas-free, faint dwarf spheroidals with [*high*]{} mass-light ratios observed around the Milky Way. Satellite Galaxies with Gas {#aquila:massive} --------------------------- Something can also be learned about gas loss by considering the four satellites (subhaloes 1, 2, 4 and 7) that still contain gas at $z=0$. With halo masses between $5\times 10^9$ and $8 \times 10^{10} {\mathrm{M}_\odot}$, these are among the most massive subhaloes; no subhalo less massive than $5 \times 10^9 {\mathrm{M}_\odot}$ still contains gas at $z=0$. None of these four satellite galaxies had particularly close encounters with the central galaxy, the minimum distance range from $\sim80$ to 330 kpc. However, this does not clearly separate them from the gas-free satellites, many of which are on even less bound orbits, or already fell in gas-free. The present galactocentric distances are also not significantly different among the two sub-populations. Interestingly, two of the only three satellite galaxies which are more massive than $5 \times 10^9 {\mathrm{M}_\odot}$ [*and*]{} gas-free, are hosted in subhaloes 3, 5, which underwent particularly strong tidal interactions and were discussed in more detail in Section \[aquila:extremes\]. Thus, it appears that maximal total masses of a few $\times10^9{\mathrm{M}_\odot}$ and orbits which avoid the inner halo are both required to retain any gas at $z=0$. Equivalently, masses below a few $\times10^9{\mathrm{M}_\odot}$ or very close orbits, are both sufficient to produce gas-free satellite galaxies. If the majority of dwarf spheroidals reside in subhaloes with masses below $10^9{\mathrm{M}_\odot}$ at present as well as at infall, cases of orbital metamorphosis are rare. It is also worth noting that the most massive surviving satellite (subhalo 1), fell in as late as $z=0.13$, and did so together with subhalo 7, another gas-rich companion. The two can easily be identified in Figures \[fig:aquila\_dm\] to \[fig:aquila\_stars\], where both are visible only in the last panels. It has been noted previously that the presence of two satellites as bright as the Magellanic Clouds near the Milky Way is rather unusual in $\Lambda$CDM [e.g @Boylan-Kolchin-2010]. [@Tremaine-1976] showed that dynamical friction in the halo of the Milky Way would lead to a rapid decay of the orbits of such large satellites, which would therefore be short-lived, surviving only a few Gyrs. Proper motions [e.g. @Besla-2007; @Piatek-2008] suggest that the Magellanic Clouds are indeed near their first pericentre after infall. -- -- -- -- -- -- -- -- -- -- -- -- -- -- Scaling Relations {#aquila_relations} ================= At redshift $z = 0$, the halo contains 199 satellite subhaloes, with masses between $10^{11}$ and $10^{8}{\mathrm{M}_\odot}$. Of these, 90 have stars, and 4 also contain gas. The properties of all satellites are listed in Table \[table:aquila-satellites\] of the appendix. They include the present stellar, gas and dark matter mass, the present galactocentric distance and the distance of closest approach, the infall redshift, the mass at infall, the median and maximum stellar metallicity. In this section, we explore the scaling relations among these properties. We focus in particular on the way in which the formation and evolution of the satellite population is linked to the subhalo mass, and to the influence of the environment. Stellar Mass - Halo Mass ------------------------ Figure \[fig:relation\_mdm-ms\] shows the relationship between present stellar mass and dark matter mass for all satellites that contain stars at $z=0$. In the left panel, the dark matter mass is the current mass of each subhalo, while in the right panel, the dark matter mass is the mass of the satellite at infall, i.e. when it first became a subhalo of the host halo (see Section \[aquila:substructure\]). In both cases, there is a clear correlation of stellar mass and halo mass, indicating that the processes that determine the amount of star formation per subhalo are regulated primarily by its mass. For a halo with an infall mass of $\sim10^9 {\mathrm{M}_\odot}$, the corresponding stellar mass is between a few times $10^5$ to a a few times $10^7 {\mathrm{M}_\odot}$. It should be noted that the minimum stellar mass resolved in the simulations is $2\times10^5 {\mathrm{M}_\odot}$. The two subhaloes discussed in Section \[aquila:extremes\], which underwent particularly strong tidal stripping, can be identified as outliers in the relation of stellar mass to present halo mass. Overall, the scatter is noticeably smaller when the mass at infall, rather than the present day mass is considered, suggesting that the evolution of the satellite after infall also plays a role in some cases. However, it is worth noting that environmental effects primarily reduce the [*halo mass*]{}, rather than the [*stellar mass*]{}, contrary to the scenario described in Section \[introduction\], whereby faint dwarf galaxies are formed through stripping of baryons. Figure \[fig:relation\_mdm-ms\] also includes a comparison with results from our earlier simulations of isolated dwarf galaxies with much higher resolution. In both panels, the black stars denote results from simulations labeled 12–20, with total masses of $2.3\times10^8$–$10^9{\mathrm{M}_\odot}$, presented in [@Sawala-2010], with stellar particle masses of $5.4\times 10^2$–$2.7 \times 10^3 {\mathrm{M}_\odot}$. Blue stars are adopted from [@Sawala-2011], where six haloes with representative merger histories and a common mass scale of $\sim10^{10}{\mathrm{M}_\odot}$ were re-simulated, with a stellar particle mass resolution of $9\times10^3{\mathrm{M}_\odot}$. We find that, despite the difference in resolution of up to two orders of magnitude, the results are in good agreement between the different sets of simulations, particularly when the dark matter masses are corrected for the effect of stripping, as shown in the right panel. Because the same code has been used in all three sets of simulations, it follows that the results are not strongly affected by resolution. The two panels in Figure \[fig:ratio-infall-approach\] both show the change in stellar mass – halo mass ratio of each object from infall to the present. The ratio at infall is shown on the x-axis, while the present ratio is shown on the y-axis. Most points lie close to the black line, which indicates a constant ratio. Notably however, the majority of haloes are above the line, meaning that their stellar mass fraction has increased since infall. This can be understood as a consequence of preferential stripping of dark matter compared to stellar matter, which is more centrally concentrated and therefore more strongly bound to the satellite. In the left panel, the colour-coding is by infall redshift; black and blue symbols indicate recent accretion, yellow and red symbols indicate infall at high redshift. In general, satellites that fell in earlier are more likely to have changed their ratio since infall, as expected if the change is due to continuous tidal stripping. In the right panel, the colour-coding is done by distance of closest approach between the subhalo and the halo of the central galaxy. As expected, haloes that had closer encounters are also the ones that underwent a slightly stronger change in the stellar mass to halo mass ratio since infall. It appears that the haloes with the greatest distance (D$_{min} > 300$kpc) have seen no change in the ratio, but these are commonly also subhaloes that have fallen in only recently (z$_{inf} < 0.2$). In both panels, the sizes of the symbols indicate total mass; larger satellites are typically found with higher stellar mass – total mass ratios, independent of infall time or orbit. Stellar Populations ------------------- Due to the small numbers of stellar particles per subhalo in the simulation, a detailed analysis of stellar populations is not possible. As a proxy for star formation history, we consider the maximum iron abundance \[Fe/H\] of the stars in each satellite galaxy. Because iron is formed only in the late stages of stellar evolution and injected into the interstellar medium via supernovae, the amount of iron observed in stars corresponds the specific degree of reprocessing of material within each galaxy, and the intensity and duration of star formation. Figure \[fig:fe-relations\] shows the maximum stellar iron abundance of the satellites, as a function of present distance (left), distance of closest approach (centre), and present-day stellar mass (right). Note that satellites with only a single generation of stars have primordial abundances, i.e. $[\mathrm{Fe/H}]\equiv- \infty$, and therefore do not appear on the plotted relations. The lack of a correlation on both the left and central panels indicate that the iron abundance does not depend strongly on either present distance, or distance of closest approach in the past. By contrast, there is a strong correlation with stellar mass, as observed in the Local Group, and also reproduced in our earlier simulations of isolated dwarf galaxies. At lower stellar mass, the scatter increases, similar to the trend in the the relation of stellar mass and halo mass seen in Figure \[fig:relation\_mdm-ms\]. Isolated Dwarf galaxies {#aquila_outside} ======================= ![Stellar mass per subhalo as a function of dark matter mass. In both panels, [*isolated*]{} galaxies are shown as blue dots, [*satellite*]{} galaxies are overplotted as red, open circles. The two blue, open circles represent dwarf galaxies presently isolated, but which had past interactions, as described in the text. In the upper panel, the dark matter mass of both the isolated and the satellite galaxies is taken at $z=0$, while the infall mass of the satellites and the interacting isolated galaxies is used in the lower panel, as in Figure \[fig:relation\_mdm-ms\]. The fraction of subhaloes without stars (not shown) is $55\%$ for the satellite subhaloes of the Milky Way, and $77\%$ for the isolated subhaloes.\[fig:relation\_mdm-ms\_outside\]](relation_mdm-ms_all.ps "fig:")\ ![Stellar mass per subhalo as a function of dark matter mass. In both panels, [*isolated*]{} galaxies are shown as blue dots, [*satellite*]{} galaxies are overplotted as red, open circles. The two blue, open circles represent dwarf galaxies presently isolated, but which had past interactions, as described in the text. In the upper panel, the dark matter mass of both the isolated and the satellite galaxies is taken at $z=0$, while the infall mass of the satellites and the interacting isolated galaxies is used in the lower panel, as in Figure \[fig:relation\_mdm-ms\]. The fraction of subhaloes without stars (not shown) is $55\%$ for the satellite subhaloes of the Milky Way, and $77\%$ for the isolated subhaloes.\[fig:relation\_mdm-ms\_outside\]](relation_mdm-ms_infall_all.ps "fig:") As is evident from Figures \[fig:aquila\_dm\] through \[fig:aquila\_projections\], the high resolution volume also contains plenty of structures outside the Milky Way halo. In this section, we compare the population of Milky Way satellites discussed in the previous section to the isolated dwarf galaxies that form in the remaining volume. At $z=0$, there are a total of 2097 subhaloes in the simulation, only $\sim 10\%$ of which are part of the most massive halo. Not all objects outside of the main group are truly isolated, as some of the other FoF groups also contain multiple subhaloes (see Section \[aquila:substructure\]). The second largest FoF group in the simulation hosts a galaxy with a stellar mass of $1.1 \times 10^{10} {\mathrm{M}_\odot}$ and contains 68 additional subhaloes. The next most massive galaxy in the simulation has a stellar mass of $1.7 \times 10^9{\mathrm{M}_\odot}$. Throughout this section, we consider as [*isolated*]{} all subhaloes which are not part of the two most massive haloes. Of these $1810$ subhaloes, 420 host stars, and 144 also contain gas. We compare these to the [*satellites*]{} of the Milky Way halo, that were discussed in the previous sections. Figure \[fig:relation\_mdm-ms\_outside\] shows the relation of stellar mass to halo mass for satellite galaxies (in red) and for these isolated dwarfs (in blue). When the mass ratios are compared at $z=0$ (upper panel), the satellites contain a systematically higher stellar mass for a given halo mass compared to the isolated galaxies. The difference is reduced when the infall masses are considered for the satellites (lower panel). The trend that low-mass, isolated subhaloes still show a higher mean dark matter mass in this second relation may be partly due to the fact that they typically grow in mass until $z=0$, while satellites peak at $z_{inf} > 0$. However, it is also partly attributable to the identification of substructures, which requires a higher density if the mean background density is higher. Interestingly, in the upper panel, the population of isolated dwarf galaxies also shows two outliers in the $M_\star-M_{DM}$ relation, which are denoted by blue open circles in both panels. While these two deviate not as significantly as the satellites which were discussed in Section \[aquila:extremes\], it points to the fact that even some dwarf galaxies which are isolated at $z=0$, may have undergone interactions in the past. The two objects with high stellar to total mass ratios were never satellites of the main halo, but have interacted with smaller groups of galaxies, which also lead to tidal stripping, mostly of dark matter. In the lower panel, we plot their dark matter masses before their last interactions, which brings them closer to the relation defined by both populations. It is also worth noting that the fraction of subhaloes without stars are different among the two populations: Whereas $45\%$ of the satellite subhaloes contain stars at $z=0$, only 23$\%$ of the isolated subhaloes contain stars. This indicates that the lowest mass subhaloes, which are unable to form stars even in isolation, often do not survive to $z=0$ if they become satellites. ![Gas mass as a function of present subhalo mass for [ *isolated*]{} dwarf galaxies (blue) and [*satellite*]{} dwarf galaxies (red). Filled (open) circles indicate isolated (satellite) galaxies with $M_{\mathrm{gas}} > 0$; arrows show the dark matter masses of galaxies that contain stars, but no gas. The red and blue numbers indicate the total number of gas-free subhaloes with stars for the two populations. The fraction of gas-free galaxies over the mass range shown, is $\sim95\%$ for satellites, and $\sim75\%$ for isolated dwarf galaxies. \[fig:mdm-mgas\]](relation_mdm-mgas_all.ps) In Figure \[fig:mdm-mgas\], we compare the gas content of satellites and isolated dwarf galaxies as a function of halo mass. As described in Section \[aquila:massive\], there are only four satellite galaxies with gas at $z=0$, which are found in some of the most massive subhaloes. Most of the isolated dwarf galaxies at $z=0$ are also gas-free, but about 1/3rd of the isolated galaxies still contain some gas at $z=0$. Among the isolated galaxies, there is a sharp drop in gas mass at subhalo masses of $\sim 2 \times 10^9 {\mathrm{M}_\odot}$. More massive isolated galaxies predominantly contain gas, less massive galaxies are most often gas-free, and resemble the satellite population. This points to a mass-threshold below which gas removal is efficient, and mostly independent of environment, while more massive galaxies can keep their gas in isolation. Close to this mass threshold, the populations differ; while some isolated dwarf galaxies still contain gas, no gas is found in the satellites. It should be noted, however, that while there are only four satellite galaxies with gas at $z=0$, the total number of satellite galaxies above the mass threshold is also very low. The most massive, gas-free galaxies are found in subhaloes of $\sim 5\times 10^9 {\mathrm{M}_\odot}$ in both the satellite and the isolated populations. [l]{} Figure \[fig:fe-relations\_outside\] compares the maximum iron enrichment \[Fe/H\] of stars in satellites and isolated galaxies, and is similar to the right panel in Figure \[fig:fe-relations\]. In both cases, there is a clear trend of increasing metallicity with stellar mass, with a large scatter at the low mass end, probably attributable in part to discreteness effects. At a given mass, the iron enrichment and the scatter are similar in the two populations, indicating that mass, rather than environment, is the primary determinant for the star formation history of dwarf galaxies. Summary {#aquila_summary} ======= Before summarizing, it should be emphasized again that some of the results presented in this paper are scraping the resolution limit of the [*Aquila*]{} simulation, which was designed to study much larger galaxies. We have, however, been able to demonstrate that the results are consistent with our much higher resolution simulations of individual dwarf galaxies. The simulation is also not enough to solve all the questions pertaining to the dwarf galaxy satellites of the Milky Way; the so-called [*ultra-faint*]{} dwarfs are clearly outside the scope of the simulation, and even some of the classical faint satellites are only resolved with a handful of particles. Tidal effects, believed to be the main stripping mechanism, should not depend very strongly on force resolution. However, if ram-pressure stripping were effective in the Local Group, it would likely be severely underestimated, because the gas pressure is only poorly determined, and pressure gradients are artificially smoothed in the SPH formalism. While the resolution is limited, it is reassuring that the properties of these galaxies tend to agree with the higher resolution simulations of isolated dwarf galaxies, and that the number of $\sim20-30$ dwarf-spheroidal-like galaxies with stellar masses in the range of $\sim10^6{\mathrm{M}_\odot}$ qualitatively agrees with the observational counts. At the same time, the fact that the simulation was not [*designed*]{} to study dwarf galaxies may actually be an advantage: Considering that dwarf galaxies and Milky Way sized galaxies form in the same universe, it is only natural that they are simulated with the same realization of the physical laws and parameters. We have described the formation of the satellite population in a cosmological simulation of a Milky Way sized halo and its environment that includes hydrodynamics, cooling, star formation, supernova feedback, enrichment and heating by the UV background. Of the 199 subhaloes with masses above $6\times10^7{\mathrm{M}_\odot}$, 90 contain stars and four also contain gas at $z=0$. We identified different mechanisms of gas-removal, both independent of environment (supernova feedback and UV heating), and caused by interactions with the host halo (tidal stripping). It was found that with a few notable exceptions, the properties of the satellites as a whole depend only weakly on environment, but very strongly on the mass of the subhalo. The satellite galaxies that contain stars follow a steep stellar mass – total mass relation, and a stellar mass–metallicity relation which are similar to those observed, and indistinguishable from those for isolated dwarf galaxies in the same simulation. The relations are also similar to those obtained by high resolution simulations of isolated dwarf galaxies [e.g. @Revaz-2009; @Sawala-2010]. In our simulation, tidal interactions after infall affect the dark matter haloes of satellites more strongly than their stellar components. The result is an average [*increase*]{} in total stellar mass – halo mass ratio, or a corresponding [*decrease*]{} in total mass – light ratio after infall. This is difficult to reconcile with the transformation of luminous, late-type galaxies with moderate mass-to-light ratios into dwarf-spheroidals with the high mass-to-light ratios inferred from observations, and required from abundance-matching arguments [e.g. @Guo-2010]. Furthermore, the trend of decreasing mass – light ratio is more pronounced for satellites on closer orbits, and with earlier infall times. This suggests that, if dwarf spheroidal galaxies of different luminosities originated from common, gas rich and bright progenitors subject to different levels of interaction, their final stellar mass after stripping would scale [*proportional*]{} to their total mass – light ratio; the opposite of which is commonly inferred from observations. [@Penarrubia-2010] have shown in purely gravitational simulations that preferential stripping of stars may be possible in a Milky Way potential if the infalling satellite haloes are cusped, so that the dark matter is more concentrated than the stars. Unfortunately, our fully cosmological, hydrodynamical simulations fall short of resolving the inner parts of satellite subhaloes to distinguish between cusped and cored profiles by several orders of magnitude, but the softened potential acts as an [*effective*]{} core. While the direct observational evidence for cusps or cores in dwarf spheroidals is still not clear, it should be noted that cores are not necessarily unique to “warm” dark matter: as shown by [@Navarro-1996], and recently confirmed in simulations by [@Governato-2010], baryonic feedback processes may result in cores in low-mass, cold dark matter haloes. If cores are a universal feature of dwarf galaxies, the transformation of dwarf irregulars to dwarf spheroidals purely via tidal effects would be difficult to reconcile with our results. Instead, we find that satellite galaxies that end up with high total mass-to-light ratios at $z=0$ are already faint on infall, and many of them have already lost their gas as a result of supernova feedback and UV radiation. While tidal interactions can remove remaining interstellar gas, as we observe in several cases, all these results suggest that the star formation of dwarf spheroidal galaxies is mostly determined independent of environment, and very strongly dependent on mass. At a total mass of $\sim 1-3\times10^9 {\mathrm{M}_\odot}$, the populations of isolated and satellite dwarf galaxies differ in the fraction of galaxies with gas. Whereas all satellite galaxies in this mass range are gas-free, isolated galaxies show a sharp decline in gas fraction, but many of them still contain gas at $z=0$. Qualitatively, this is in agreement with the HI mass–distance relation reported by [@Grebel-2003] and [@Geha-2006]. In the mass regime of dwarf spheroidals, however, with stellar masses below $10^7{\mathrm{M}_\odot}$ and inferred dynamical masses below $10^9{\mathrm{M}_\odot}$, both the satellites and most of the isolated dwarf galaxies are gas-free. Consequently, the dwarf spheroidal galaxies formed in the simulation do not follow a clear morphology-distance dichotomy. If such a sharp relation exists, the galaxies that constitute this relationship would not only have to form in a different way compared to the simulation, but also be on different orbits, as we find that present day position is not a good proxy for past interaction. Acknowledgements {#acknowledgements .unnumbered} ================ All simulations were carried out at the computing centre of the Max-Planck Society in Garching. \[lastpage\] [lccccccccc]{} \ \ Label & M$_\star$ & M$_\mathrm{gas}$ & M$_\mathrm{DM}$ & M$_\mathrm{inf}$ & Z$_\mathrm{inf}$ & Z$_\mathrm{form}$ & D & D$_\mathrm{min}$ & $[$Fe$/$H$]$\ & \[$10^6 {\mathrm{M}_\odot}$\] & \[$10^6 {\mathrm{M}_\odot}$\] & \[$10^8 {\mathrm{M}_\odot}$\] & \[$10^8 {\mathrm{M}_\odot}$\] & & & \[kpc\] & \[kpc\] &\ \ \ \ Label & M$_\star$ & M$_\mathrm{gas}$ & M$_\mathrm{DM}$ & M$_\mathrm{inf}$ & Z$_\mathrm{inf}$ & Z$_\mathrm{form}$ & D & D$_\mathrm{min}$ & $[$Fe$/$H$]$\ & \[$10^6 {\mathrm{M}_\odot}$\] & \[$10^6 {\mathrm{M}_\odot}$\] & \[$10^8 {\mathrm{M}_\odot}$\] & \[$10^8 {\mathrm{M}_\odot}$\] & & & \[kpc\] & \[kpc\] &\ \ \ \ \ 1 & 4745.17 & 1259.41 & 833.43 & 1142.22 & 0.13 & 9.37\* & 266.9 & 266.9 & -1.06\ 2 & 1960.70 & 2.78 & 133.21 & 310.85 & 0.53 & 10.91 & 383.1 & 83.5 &-0.94\ 3 & 1322.04 & – & 2.54 & 429.65 & 2.32 & 12.07 & 13.2 & 13.2 & -0.63\ 4 & 188.92 & 168.56 & 108.16 & 128.96 & 0.07 & 5.23\* & 233.5 & 233.5 & -1.82\ 5 & 623.97 & – & 4.18 & 247.15 & 2.10 & 15.47 & 133.3 & 13.2 & -0.72\ 6 & 276.84 & – & 37.15 & 121.12 & 0.50 & 14.73 & 321.9 & 62.8 & -1.49\ 7 & 40.32 & 0.84 & 48.05 & 50.31 & 0.13 & 13.34\* & 328.3 & 328.3 & -2.21\ 8 & 69.45 & – & 43.37 & 45.73 & 0.70 & 4.67\* & 205.4 & 205.4 & -1.89\ 9 & 122.04 & – & 18.19 & 51.53 & 1.07 & 12.07 & 246.1 & 121.5 & -1.84\ 10 & 42.12 & – & 23.53 & 46.69 & 0.15 & 9.86 & 121.0 & 121.0 & -2.05\ 11 & 19.61 & – & 25.15 & 27.30 & 0.13 & 9.37 & 490.2 & 490.2 & -2.36\ 12 & 7.68 & – & 25.90 & 31.83 & 0.38 & 10.38 & 390.0 & 334.5 & -2.74\ 13 & 36.82 & – & 20.05 & 29.71 & 0.09 & 8.45\* & 129.6 & 129.6 & -1.95\ 14 & 42.11 & – & 13.00 & 26.31 & 2.20 & 13.34 & 149.3 & 82.6& -2.09\ 15 & 73.51 & – & 6.02 & 81.83 & 2.90 & 15.47 & 96.6 & 52.5 & -1.43\ 16 & 13.64 & – & 12.06 & 20.23 & 0.13 & 16.25 & 289.9 & 289.9 & -1.93\ 17 & 22.03 & – & 9.76 & 17.05 & 1.36 & 5.83\* & 346.9 & 62.0 & -2.21\ 18 & 4.80 & – & 11.69 & 13.24 & 0.05 & 10.38 & 303.4 & 303.4 & \*\*\ 19 & 15.60 & – & 9.65 & 20.01 & 0.24 & 14.73 & 228.9 & 118.4 & -2.16\ 20 & 7.49 & – & 9.89 & 10.16 & 0.13 & 12.07 & 203.0 & 203.0 & -2.90\ 21 & 3.08 & – & 9.52 & 12.70 & 0.26 & 9.86 & 306.6 & 152.9 & -2.86\ 22 & 8.26 & – & 8.63 & 17.18 & 1.41 & 7.22 & 104.9 & 104.9 & -2.23\ 23 & 30.98 & – & 5.76 & 8.52 & 1.60 & 12.69 & 137.0 & 137.0 & -1.40\ 24 & 18.51 & – & 6.96 & 33.16 & 0.18 & 12.07 & 232.6 & 21.6 & -1.98\ 25 & 19.63 & – & 6.61 & 33.58 & 0.99 & 12.69 & 297.6 & 62.6 & -1.88\ 26 & 7.10 & – & 6.46 & 8.58 & 0.13 & 8.90 & 303.6 & 303.6 & -3.05\ 27 & 0.38 & – & 6.39 & 7.38 & 0.07 & 7.62 & 304.1 & 304.1 & \*\*\ 28 & 0.58 & – & 6.24 & 4.90 & 0.13 & 11.48 & 312.0 & 312.0 & \*\*\ 29 & 10.77 & – & 4.36 & 19.70 & 0.92 & 11.48 & 248.7 & 64.7 & -2.32\ 30 & 5.19 & – & 4.73 & 11.78 & 0.85 & 11.48 & 215.0 & 86.0 & -2.23\ 31 & 2.31 & – & 4.90 & 15.81 & 2.10 & 10.91 & 163.3 & 90.8 & -2.99\ 32 & 6.54 & – & 4.40 & 14.73 & 0.55 & 8.02 & 190.6 & 102.5 & -2.85\ 33 & 1.15 & – & 4.84 & 5.91 & 0.10 & 7.62 & 141.7 & 141.7 & -3.05\ 34 & 0.77 & – & 4.73 & 6.74 & 0.24 & 9.37 & 340.2 & 258.5 & -4.45\ 35 & 13.93 & – & 3.17 & 39.60 & 3.09 & 9.37 & 93.1 & 79.6 & -1.09\ 36 & 2.31 & – & 4.20 & 8.87 & 0.21 & 8.02 & 235.4 & 54.7 & -2.57\ 37 & 4.23 & – & 3.87 & 5.41 & 0.15 & 0.15 & 173.0 & 54.0 & -2.19\ 38 & 4.42 & – & 3.83 & 9.57 & 0.92 & 8.45 & 135.6 & 88.1 & -1.77\ 39 & 1.92 & – & 4.09 & 9.63 & 0.33 & 8.90 & 328.7 & 71.7 & -3.05\ 41 & 12.32 & – & 2.80 & 4.12 & 0.13 & 10.38 & 252.6 & 252.6 & -1.74\ 42 & 3.46 & – & 3.70 & 3.70 & 0.13 & 8.02 & 390.2 & 390.2 & -3.30\ 43 & 0.96 & – & 3.85 & 6.22 & 0.17 & 10.91 & 100.3 & 98.4 & -4.82\ 45 & 0.38 & – & 3.68 & 4.73 & 0.31 & 7.62 & 447.4 & 127.2 & \*\*\ 46 & 0.58 & – & 3.52 & 5.91 & 0.24 & 8.90 & 257.5 & 106.7 & -3.36\ 47 & 0.77 & – & 3.33 & 3.70 & 0.18 & 8.02 & 307.9 & 307.9 & -2.96\ 48 & 0.38 & – & 3.13 & 3.63 & 0.21 & 8.45 & 212.5 & 164.0 & -4.08\ 49 & 0.96 & – & 3.02 & 3.68 & 0.92 & 12.07 & 193.4 & 191.6 & \*\*\ 51 & 0.77 & – & 2.87 & 3.26 & 0.65 & 2.74 & 207.7 & 207.7 & -4.07\ 52 & 0.19 & – & 2.71 & 5.49 & 0.28 & 6.50 & 357.1 & 66.6 & \*\*\ 53 & 0.58 & – & 2.63 & 3.48 & 0.73 & 7.62 & 222.0 & 222.0 & -2.86\ 56 & 2.11 & – & 2.08 & 3.00 & 2.10 & 12.07 & 143.7 & 60.8 & -6.99\ 57 & 0.19 & – & 2.25 & 4.38 & 0.50 & 7.62 & 259.8 & 120.2 & \*\*\ 58 & 0.96 & – & 2.15 & 5.36 & 1.11 & 7.62 & 176.3 & 106.0 & -2.62\ 59 & 1.92 & – & 1.88 & 4.66 & 0.53 & 10.38 & 218.5 & 39.5 & -2.36\ 60 & 0.19 & – & 2.04 & 2.15 & 0.14 & 1.36 & 129.8 & 129.8 & \*\*\ 63 & 0.96 & – & 1.86 & 4.79 & 1.47 & 11.48 & 164.4 & 70.6 & -4.77\ 64 & 0.77 & – & 1.88 & 2.82 & 3.29 & 10.38 & 140.2 & 140.2 & -4.07\ 65 & 0.58 & – & 1.84 & 1.99 & 0.13 & 7.22 & 196.6 & 196.6 & -2.89\ 66 & 1.34 & – & 1.73 & 9.02 & 1.20 & 8.45\* & 327.3 & 36.7 & -2.98\ 67 & 0.96 & – & 1.75 & 4.44 & 1.67 & 10.38 & 223.5 & 17.5 & \*\*\ 68 & 0.38 & – & 1.80 & 2.47 & 1.03 & 10.91 & 267.2 & 97.9 & \*\*\ 70 & 0.38 & – & 1.80 & 2.80 & 1.07 & 8.02 & 235.1 & 172.1 & \*\*\ 71 & 0.19 & – & 1.71 & 1.69 & 0.13 & 4.42 & 367.7 & 367.7 & \*\*\ 73 & 0.96 & – & 1.55 & 11.76 & 0.79 & 10.91 & 158.5 & 42.6 & -2.38\ 74 & 0.38 & – & 1.55 & 2.23 & 1.11 & 1.11\* & 265.7 & 104.1 & -2.16\ 78 & 0.96 & – & 1.31 & 7.07 & 3.71 & 15.47 & 86.0 & 85.1 & -2.58\ 79 & 0.19 & – & 1.40 & 2.71 & 1.41 & 10.38 & 136.0 & 136.0 & \*\*\ 83 & 0.19 & – & 1.34 & 1.58 & 0.44 & 0.44\* & 207.3 & 207.3 & \*\*\ 84 & 0.19 & – & 1.34 & 2.01 & 1.47 & 8.02 & 251.0 & 214.5 & \*\*\ 86 & 0.19 & – & 1.31 & 1.69 & 1.53 & 7.22 & 307.0 & 170.2 & \*\*\ 87 & 0.58 & – & 1.23 & 6.76 & 2.45 & 7.22 & 127.3 & 105.6 & -2.13\ 94 & 0.38 & – & 1.16 & 6.41 & 1.30 & 10.38 & 265.5 & 49.5 & -2.28\ 95 & 0.19 & – & 1.18 & 1.16 & 1.36 & 1.36\* & 261.9 & 261.9 & \*\*\ 97 & 0.19 & – & 1.12 & 1.14 & 0.35 & 0.60 & 222.6 & 213.3 & -2.03\ 98 & 0.38 & – & 1.07 & 1.66 & 0.18 & 4.17 & 115.7 & 63.7 & -2.90\ 101 & 0.77 & – & 0.99 & 5.91 & 1.67 & 8.45 & 335.5 & 50.1 & -2.06\ 105 & 1.36 & – & 0.81 & 1.62 & 0.13 & 11.48 & 271.6 & 271.6 & -1.68\ 110 & 0.19 & – & 0.92 & 4.14 & 2.58 & 6.85 & 188.1 & 145.8 & \*\*\ 113 & 0.38 & – & 0.85 & 0.85 & 0.28 & 0.28\* & 115.6 & 115.6 & -4.92\ 119 & 0.19 & – & 0.77 & 4.27 & 0.79 & 10.91 & 66.7 & 43.3 & -2.19\ 120 & 0.19 & – & 0.77 & 2.08 & 1.67 & 1.82 & 78.8 & 78.8 & \*\*\ 123 & 0.19 & – & 0.74 & 0.81 & 0.13 & 10.91 & 331.6 & 331.6 & -2.14\ 128 & 0.19 & – & 0.70 & 0.77 & 0.42 & 0.42\* & 191.7 & 160.4 & -3.24\ 133 & 0.19 & – & 0.61 & 0.61 & 0.13 & 0.28 & 290.0 & 290.0 & -2.87\ 147 & 0.38 & – & 0.55 & 0.61 & 0.14 & 0.42 & 215.3 & 215.3 & -2.39\ 152 & 0.38 & – & 0.53 & 2.61 & 1.25 & 1.25\* & 204.6 & 86.8 & -2.83\ 159 & 0.19 & – & 0.55 & 1.07 & 0.29 & 0.29\* & 271.5 & 74.8 & \*\*\ 161 & 0.19 & – & 0.53 & 0.50 & 0.04 & 0.04\* & 144.7 & 144.7 & \*\*\ 167 & 0.19 & – & 0.50 & 0.48 & 0.05 & 0.05\* & 165.1 & 165.1 & \*\*\ 177 & 0.19 & – & 0.48 & 0.46 & 0.13 & 0.33\* & 379.8 & 379.8 & \*\*\ [^1]: E-Mail: [email protected]

ArXiv

--- abstract: 'A first-order percolation transition, called explosive percolation, was recently discovered in evolution networks with random edge selection under a certain restriction. However, the network percolation with more realistic evolution mechanisms such as preferential attachment has not yet been concerned. We propose a tunable network percolation model by introducing a hybrid mechanism of edge selection into the Bohman-Frieze-Wormald (BFW) model, in which a parameter adjusts the relative weights between random and preferential selections. A large number of simulations indicate that there exist crossover phenomena of percolation transition by adjusting the parameter in the evolution processes. When the strategy of selecting a candidate edge is dominated by random selection, a single discontinuous percolation transition occurs. When a candidate edge is selected more preferentially based on node‘s degree, the size of the largest component undergoes multiple discontinuous jumps, which exhibits a peculiar difference from the network percolation of random selection with a certain restriction. Besides, the percolation transition becomes continuous when the candidate edge is selected completely preferentially.' author: - Xiaolong Chen - Chun Yang - Linfeng Zhong - Ming Tang title: Crossover phenomena of percolation transition in evolution networks with hybrid attachment --- **Percolation is a pervasive concept in graph theory and statistics physics, which has a broad application in network science, epidemiology, and so on. Now it’s a useful theoretical tool to study the large-scale connectivity of complex networks. It was known as a continuous percolation process of random networks before the product rule was introduced into the network evolution process by Achlioptas in 2009. In the percolation model proposed by Achlioptas, the size of the largest component jumps abruptly at a transition point, which is called explosive percolation, and it is considered as a first-order transition. Stimulated by the work of Achlioptas, a great amount of attention has been given to the research of the explosive percolation in evolution networks with random edge selection under a certain restriction. However, the percolation transition in a more realistic evolution process of networks is still lack of studying. Based on the classical Bohman-Frieze-Wormald model, we propose a percolation model in evolution process of networks with hybrid attachment. In our model, a tunable parameter is taken into account to adjust the way of selecting a candidate edge. Through numeric simulations, we find that there are crossover phenomena of percolation transition in the evolution process, which exhibit a peculiar difference from the classical Bohman-Frieze-Wormald network percolation. Specifically, when the strategy of selecting a candidate edge is dominated by random selection, there is only a single discontinuous jump of the largest component size. When a candidate edge is selected more preferentially based on node‘s degree, the largest component grows with multiple discontinuous jumps. Besides, nearly completely preferential selection makes the size of the largest component increase continuously. Our work reveals the effect of hybrid attachment mechanism on the network percolation transition, and offer some insights into understanding the network percolation in a more realistic evolution process.** Introduction {#sec:intro} ============ Percolation is used to describe the movement and filtering of fluids through porous materials in materials science [@grimmett1999percolation]. During the last five decades, percolation theory has brought new understanding to a broad range of topics in networks science [@boccaletti2006complex; @castellano2009statistical; @newman2010networks] and epidemiology [@moore2000epidemics; @fu2008epidemic]. During 1959 to 1961, Erd[ö]{}s and R[é]{}nyi proposed the evolution model of random graphs, i.e., Erd[ö]{}s-R[é]{}nyi (ER) model [@erdds1959random; @erd6s1960evolution]. In this model, a randomly selected link is added to the network at each time step. Such random edge selection rule generates a second-order percolation transition that the largest connected component grows continuously. In 2009, Achlioptas introduced a product rule of edge selection [@achlioptas2009explosive]. In the model, two candidate edges are randomly selected, and the one merging the two components with smaller product of their sizes is added to the network. It postpones the emergence of the giant connected component and hence leads to a abrupt jump of the largest component size, which is called as explosive percolation. It is in striking contrast to the second-order percolation transition in the classical ER model, and a great deal of researches have been triggered [@ziff2009explosive; @radicchi2009explosive; @cho2009percolation; @friedman2009construction; @riordan2011explosive; @lee2011continuity; @araujo2010explosive; @cho2013avoiding; @chen2011explosive; @d2015anomalous]. In particular, refs. [@cho2009percolation; @radicchi2009explosive] explored the explosive percolation of the configuration scale-free networks in a cooperative Achlioptas growth process. Ref. [@cho2009percolation] found that there exists a critical degree-exponent parameter $\lambda_c\in(2.3 , 2.4)$. If $\lambda<\lambda_c$, the transition threshold tends to zero in the thermodynamic limit, and the transition is continuous; otherwise, a discontinuous percolation transition occurs near a finite threshold. Almost at the same time, ref. [@radicchi2009explosive] got similar results through numeric simulations. However, it is proven rigorously that the explosive percolations for all the Achlioptas processes are actually continuous by subsequent researches [@da2010explosive; @grassberger2011explosive; @lee2011continuity]. Besides the Achlioptas process, another percolation model was introduced by Bohman, Frieze and Wormald (BFW) [@bohman2004avoidance]. In this model, a single edge is randomly selected at each step. If the size of the resulting component is less than a specified value, this edge would be added to the network, or a critical function is used to decide whether to receive it or not. Later on, Chen *et al*. [@chen2011explosive] demonstrated that the transition of the BFW model is actually discontinuous. Although percolation transition in random networks has been studied extensively, its application to real-world networks is still scarcely investigated [@rozenfeld2010explosive; @pan2011using]. Actually, in the evolution processes of most real-world networks, the strategy of selecting a newly added edge follows the mechanism of preferential selection instead of random selection. The preferential attachment was introduced by Barab[á]{}si and Albert [@barabasi1999emergence], and it is considered as a vital mechanism to describe the evolution of the real-world networks with heterogeneous architecture [@newman2001clustering; @de2007preferential; @holme2002growing; @small2015growing]. In the model, the probability that a new node leads an edge to an exist node $i$ is in proportion to its degree $k_i$, that is $\prod_{i}\sim k_i$. However, the architectures of most real-world networks can not be described by strictly scale-free or random model [@newman2001random; @fararo1964study]. Thus a more realistic network evolution model with hybrid attachment was proposed [@liu2002connectivity; @liu2003propagation]. In the model, the probability that a node $i$ is attached by a new edge is in proportion to $\prod_{i}\sim(1-q)k_i+q$, where $q$ controls the relative weights between the heterogeneous and homogeneous components in the evolution process. In this paper, we propose a hybrid BFW model (HBFW) with mixture of preferential and random edge selection to study the network percolation in a more realistic case. When $q=0$, the model degenerates to the original BFW model, where each candidate edge is selected randomly, and the degrees thus satisfy exponential distribution. When $q=1$, the network is strictly scale free as the preferential attachment dominates network evolution process. Simulation results demonstrate that three separate parameter regions $\left[0,q^{\uppercase\expandafter{\romannumeral1}}_c \right]$, $\left(q^{\uppercase\expandafter{\romannumeral1}}_c,q^{\uppercase\expandafter{\romannumeral2}}_c \right]$, $\left(q^{\uppercase\expandafter{\romannumeral2}}_c, 1\right]$ exist in the model, where $q^{\uppercase\expandafter{\romannumeral1}}_{c}\in (0.84, 0.86)$ and $q^{\uppercase\expandafter{\romannumeral2}}_c\in (0.97, 0.98)$ respectively stand for two critical values. And crossover phenomena of percolation transition occur between these regions. Specifically, there is only a single discontinuous transition when $q\in\left[0, q^{\uppercase\expandafter{\romannumeral1}}_c \right]$. And when $q\in\left(q^{\uppercase\expandafter{\romannumeral1}}_c,q^{\uppercase\expandafter{\romannumeral2}}_c \right]$, multiple discontinuous jumps of the largest component size occur in the evolution process. Moreover, when $q\in\left(q^{\uppercase\expandafter{\romannumeral2}}_c, 1\right]$, the first transition of the HBFW model becomes continuous. Most extremely, there is single giant component all through the process, when $q$ approaches to 1. It shows totally different properties of percolation transitions in a network with more realistic evolution mechanism, compared with those in random networks or configuration networks. PERCOLATION MODEL {#sec:model} ================= In our model, we use the mechanism of the hybrid attachment to modify the rule of selecting a candidate edge in the BFW model. By introducing a tunable parameter *q*, we get a hybrid rule of preferential selection and random selection. Initially, the system is a network *G* with *N* isolated nodes. Denote *K* as the cap size and set $\emph{K=2}$. $u\left(l\right)$ and $L\left(l\right)$ respectively represent the number of candidate edges and accepted edges at the current step *l*. And $t=L(l)/N$ is the density of accepted edges at the current step. Then we consider the evolution process of the HBFW model. Namely, at each step *l*, the two nodes attaching to a candidate edge $e_u$ is selected according to the following two steps: $1)$ The first node is selected randomly from the system. $2)$ The second node is selected by the rule of hybrid attachment. The node $i$ is selected by the preferential (random) selection strategy with probability $q$ ($1-q$). According to the selection rule described above, a candidate edge $e_u$ is obtained, and then it is judged by the BFW algorithm. $3)$ Edge addition in the BFW algorithm. We can decide whether the candidate $e_u$ should be accepted or not according to the following criteria: (***i***) Denote $C_1$ as the size of the largest component if $e_u$ is added to the system; (***ii***) If $C_1\leq K$, we accept $e_u$ , and set $L(l+1)=L(l)+1$, $u(l+1)=u(l)+1$, $l=l+1$; (***iii***) Otherwise, if $L(l)/u(l)<g(K)=1/2+(2K)^{-1/2}$, set $K=K+1$, and go to (*ii*), until $C_1\leq K$ or $L(l)/u(l)>g(K)$. If $C_1\leq K$, $e_u$ is added to the network, and set $L(l+1)=L(l)+1$, $u(l+1)=u(l)+1$, $l=l+1$; or if $L(l)/u(l)>g(K)$, $e_u$ is rejected, and set $u(l+1)=u(l)+1$, $l=l+1$; Repeat steps $1) \sim 3)$. Note that initially all the nodes in the system are isolate, so the selection rule of preferential attachment can not be applied to any node. Thus without loss of generality, at the first step, the two nodes are randomly selected. SIMULATION RESULTS {#sec:simulation} ================== First of all, we explore the effect of the hybrid attachment on the architecture of the generated networks. In order to demonstrate the result clearly and without loss of generality, we set the generated networks of $\langle k \rangle =5$. We exhibit the degree distribution of the networks with five typical values of parameter $q$, as shown in Figs. \[digDist(mix)\] (a) and (b). When *q* is relatively small (e.g., $q=0.0$ and $q=0.6$), a pair of nodes connected by a candidate edge are selected almost by the strategy of random selection, thus degrees of all nodes satisfy an exponential distribution [@newman2001random] \[see fig.  \[digDist(mix)\] (a)\]. With the increase of $q$, it is more likely to select a large-degree node, then the heterogeneity of the networks grows gradually. And after about $q=0.86$, the degree distribution of the networks evolves asymptotically to power-law distribution [@de2007preferential], as shown in Fig. \[digDist(mix)\](b). Next, we discuss the properties of percolation transition in the HBFW model. For clarity, we denote $NC_i \thicksim o(N)$ if $\lim_{N\rightarrow\infty}C_{i}\rightarrow 0$, where $C_i$ is the fraction of nodes in the *ith* largest component, and $C_1$ is defined as the order parameter of network percolation. If $\lim_{N\rightarrow\infty}C_{i}\rightarrow c, c\in \left(0,1 \right]$, we denote $NC_i\thicksim O \left( N \right)$. In order to investigate the impact of hybrid attachment on percolation transition, we perform simulations with varieties of $q$. In the evolution process, edges are selected in strict accordance with the rules mentioned above. We first focus on the time evolutions of $C_1$ and $C_2$. Fig. \[orderparam\] illustrates the effect of the hybrid attachment on the percolation transition of the HBFW model by exhibiting the results for $q=0.0, q=0.6, q=0.96$ and $q=0.99$. Fig. \[orderparam\] shows that when $q=0.0$, it degenerates to the BFW model and there is only a unique discontinuous jump of the largest component at the percolation threshold. But for $q>0$, as shown in Figs. \[orderparam\](b)$\sim$(d), we can see that the transition point decreases gradually with the probability $q$ of preferential attachment. Most importantly, for a large value of $q$ multiple jumps of $C_1$ and $C_2$ occur after the first percolation transition \[see Figs. \[orderparam\](c) and (d)\]. When $q\rightarrow1$, as shown in Fig. \[orderparam\](d), the first transition of $C_1$ and $C_2$ is prone to be continuous. ![(Color online) Typical degree distribution of generated networks with $\langle k\rangle =5$. (a) the degree distribution for $q=0.0$ and $q=0.6$, (b) the degree distribution for $q=0.86$, $q=0.96$, $q=0.99$. System size is set sa $N=10^6$.[]{data-label="digDist(mix)"}](degDist.eps){height="80mm" width="58mm"} ![(Color online) Time evolutions of $C_1$ and $C_2$ for four typical values of $q$. (a) and (b) respectively depict the evolution processes for $q=0.0$ and $q=0.6$, which are both smaller than $q^{\uppercase\expandafter{\romannumeral1}}_c$; (c) and (d) respectively depict the processes for $q=0.96$ and $q=0.99$ , which are both larger than $q^{\uppercase\expandafter{\romannumeral1}}_c$. System size is set as $N=10^6$. []{data-label="orderparam"}](orderparam.eps){width="1\linewidth"} From large numbers of simulations, we find that there exists two critical values $q^{\uppercase\expandafter{\romannumeral1}}_c\in (0.84, 0.86)$ and $q^{\uppercase\expandafter{\romannumeral2}}_c\in (0.97, 0.0.98)$. For $q\in[0,q^{\uppercase\expandafter{\romannumeral1}}_c]$, there is only one jump of $C_1$ and $C_2$ at the transition point. And for $q\in(q^{\uppercase\expandafter{\romannumeral1}}_c, q^{\uppercase\expandafter{\romannumeral2}}_c]$, multiple jumps of the $C_1$ and $C_2$ emerge. However, the first transition of $C_1$ and $C_2$ is prone to be continuous for $q\in(q^{\uppercase\expandafter{\romannumeral2}}_c, 1]$. In the following section, we focus on the time evolution of the largest component, and demonstrate the above conclusions from two aspects. Firstly, by analyzing the jump size of the largest component $\Delta C_1$ [@nagler2011impact], and the time interval $\Delta t_i, i=1...3$ between two adjacent jumps of the largest component, we demonstrate the existence of multiple discontinuous jumps for $q\in(q^{\uppercase\expandafter{\romannumeral1}}_c, q^{\uppercase\expandafter{\romannumeral2}}_c]$. And we strengthen the above results by investigating the susceptibility $\chi$ of $C_1$. Secondly, by using the technology of “critical window” [@achlioptas2009explosive], we examine whether the first transition is discontinuous or not for different value of $q$. ![(Color online) Finite size effects of multiple jumps of the largest component $\Delta{C_1}$ and the time intervals between each two adjacent jumps $\Delta{t_1}\sim\Delta{t_3}$. The four largest jumps of order parameter $\Delta{C^{1}_{1}}\sim \Delta{C^{4}_{1}}$ for $q=0.6$ (a) and $q=0.96$ (b). Time delays $\Delta{t_1}\sim\Delta{t_3}$ for $q=0.6$ (c) and $q=0.96$ (d). Data are averaged over 100 network realizations.[]{data-label="jump"}](jump.eps){width="1\linewidth"} Firstly, we demonstrate that there is only one discontinuous jump of $C_1$ for $q\in[0, q^{\uppercase\expandafter{\romannumeral1}}_c]$, and multiple discontinuous jumps for $q\in(q^{\uppercase\expandafter{\romannumeral1}}_c, q^{\uppercase\expandafter{\romannumeral2}}_c]$. In fact, Ref. [@nagler2011impact] pointed out that if the maximum increase of the order parameter $\Delta C^{max}_{1}$ caused by adding a single edge satisfies the condition $$\label{maxJump} \lim_{N\rightarrow \infty}\Delta C^{max}_{1}/N>0,$$ the transition is strongly discontinuous. We consider the finite size effects of top 4 largest jumps of the largest component for different values of $q$. Figs. \[jump\] (a) and (b) show the cases of $q=0.6$ and $q=0.96$, where $\Delta C^{i}_1$ stands for the $ith$ largest jump size of the largest component. Simulations results illustrate that the size of the jumps are independent of the system size $N$. Therefore, the top four largest jumps of the largest component is discontinuous for $q\in[0,q^{\uppercase\expandafter{\romannumeral1}}_c]$ in the thermodynamic limit. Next, we investigate the time delays between two adjacent jumps of the largest component to judge how many discontinuous jumps occur in the evolution process. The $\Delta t_i, i=1,2,3$ stands for the time delay between the $ith$ jump and the$(i+1)th$ jump. Fig. \[jump\](c) shows that when $q=0.6$, $\Delta t_i(i=1,2,3)$ converge to zero in the thermodynamic limit. This means that these discontinuous jumps in the HBFW process with $q=0.6$ occur at almost the same time when *N* is large enough. Thus there is only one discontinuous jump of $C_1$ when $N\rightarrow\infty$. For $q=0.96$, the two time delays converge asymptotically to two positive constant values in the thermodynamic limit, i.e. $\lim_{N\rightarrow \infty}\Delta t_{1}=0.206\pm 0.002$ and $\lim_{N\rightarrow \infty}\Delta t_{2}=0.057\pm 0.004$ \[see Fig. \[jump\](d)\]. This means that three separate discontinuous jumps of $C_1$ occur during the percolation process. ![(Color online) Time evolution and finite size effects of the susceptibility $\chi$. Time evolution of the $\chi$ for $q=0.6$ (a) and $q=0.96$ (b). (c) The maximum value of susceptibility $\chi_{max}$ versus network size for $q=0.6$ and $q=0.96$. (d) The other two peak values of susceptibility $\chi_{2}$ and $\chi_{3}$ versus network size for $q=0.6$ and $q=0.96$. Data are averaged over 100 network realizations.[]{data-label="deviation"}](deviation.eps){width="1\linewidth"} To strengthen the above conclusions, we consider the susceptibility $$\label{chi} \chi=\sqrt{\langle C^2_1\rangle-\langle C_1\rangle^2},$$ which quantifies the amplitude of the fluctuations of the largest component size [@shu2015numerical]. When a transition occur in the evolution process, $\chi$ reaches to a peak value $\chi_{max}$. Further, a none zero value of $\chi_{max}$ in the thermodynamic limit indicates a discontinuous transition [@binder1981finite; @voss1984fractal]. In Fig. \[deviation\], we exhibit the susceptibility of the largest component sizes for $q=0.6$ and $q=0.96$ with corresponding to Fig. \[jump\]. Fig. \[deviation\](a) shows that there is only one peak of $\chi$, and Fig. \[deviation\](b) tells us that it has three obvious peaks of $\chi$ as time *t* increases. For the sake of clarity, we denote the second and third largest value of $\chi$ for $q=0.96$ as $\chi_{2}$ and $\chi_{3}$. Results of Figs. \[deviation\](a) and (b) show that for $q=0.96$, there are three jumps of the largest component size, but for $q=0.6$ only one jump of the largest component size occurs. Besides, Figs. \[deviation\](c) and (d) demonstrate that the three peak values of the susceptibility for $q=0.96$, and the maximum value of the susceptibility for $q=0.6$ tend to be nonzero values in the thermodynamic limit. Thus, it leads to the conclusion that the jump of the largest component for $q=0.6$ and the three jumps for $q=0.96$ are essentially discontinuous. ![(Color online) Relationship between the rescaled size of critical window $\Delta(\alpha,\beta)$ and network size *N*. (a) $\alpha=1/2,\beta=0.25$ and (b) $\alpha=1/2,\beta=0.1$ are chosen in $\Delta(\alpha,\beta)$ for $q=0.86,q=0.96$, and $q=0.99$. Data are averaged over 100 realizations.[]{data-label="cw"}](cw.eps){height="86mm" width="55mm"} At last, we demonstrate the continuity of the first transition for $q\in(q^{\uppercase\expandafter{\romannumeral2}}_c, 1]$ in the HBFW model by applying the theory of “critical window” proposed in Ref. [@achlioptas2009explosive]. Let $L_0$ denote the number of added edges at the last moment for $$\label{L0} NC_1\leq N^\alpha,$$ and $L_1$ as the number of added edges at the first moment for $$\label{L1} NC_1\geq \beta N$$ where $\alpha,\beta \in(0,1)$. $\Delta (\alpha,\beta)=L_1-L_0$ denotes the number of added edges during the evolution process, which is called as critical window. Fig. \[cw\] shows the rescaled value of $\Delta(\alpha, \beta)/N$ as a function of system size *N* for $q=0.86$, $q=0.96$, and $q=0.99$. We first set $\alpha=1/2, \beta=0.25$. As shown in Fig. \[cw\](a), $\Delta(\alpha, \beta)/N$ is sublinear to $N$ and satisfies $\lim_{N\rightarrow \infty}\Delta(\alpha, \beta)/N=0$ for $q=0.86$ and $q=0.96$, while $\lim_{N\rightarrow \infty}\Delta(\alpha, \beta)/N=0.732\pm0.002$ for $q=0.99$ \[see Fig. \[cw\](a)\]. For more accurate, we lower the value of $\beta$. In Fig. \[cw\](b), we set $\beta=0.1$ and get the accordant result as Fig. \[cw\](a). Figs. \[cw\](a) and (b) reveal the fact that the first jump of the HBFW process is discontinuous for $q=0.86$ and $q=0.96$, but it is continuous for $q=0.99$. More simulations demonstrate the existence of the critical value $q^{\uppercase\expandafter{\romannumeral2}}_c\in(0.97,0.98)$, below which the first jump of the largest component in HBFW process is discontinuous, but above which it is continuous. ![(Color online) The component distributions at the vicinity of transition threshold for $q=0.0$ (a), $q=0.96$ (b) and $q=0.99$ (c). System size is $N=10^6$.[]{data-label="cluster"}](clusterDist.eps){height="120mm" width="65mm"} Finally, we give a qualitative analysis for the two crossover phenomena of percolation transitions in the HBFW model. In the BFW model, the evolution of the largest component is restricted by the cap size $K$. Thus at each time step, the small components are merged together by a newly added edge with high probability. Consequently, a mass of components, whose sizes are approximately equal to the largest component size, emerge in the subcritical regime. In other words, a “powder key” is formed, which is a necessary condition for the discontinuous percolation transition [@friedman2009construction]. As shown in Fig. \[cluster\] (a), there are a mass of components with size of about two hundred nodes at $t=0.96$, and with size of about seven hundred nodes at $t=0.97$. Subsequently, a vanishingly small number of added edges can merge these small components, and thus lead to an explosive percolation. And at $t=0.98$, two giant components formed. The hybrid attachment mechanism in the HBFW model makes evolution procedure much more complex than the classical BFW process, which leads to peculiar properties of the percolation transitions. Generally, there are two underlying competed mechanisms behind the percolation procedures in the HBFW model, which are the suppression and promotion of growth of the largest component, respectively. On the one hand, the cap size $K$ restricts the formation of large components. On the other hand, with the increase of $q$, the probability of selecting large-degree nodes rises, which accelerates the growth of large components. When $q<q^{\uppercase\expandafter{\romannumeral1}}_c$, the strategy of selecting a candidate node is predominated by the random selection \[see Fig. \[digDist(mix)\] (a)\], the suppression plays a main role, rendering the percolation type similar to BFW model \[see Figs. \[orderparam\] (a) and (b)\]. With the increase of $q$, edges are more likely to attach to large-degree nodes \[see Fig. \[digDist(mix)\] (b)\]. Owing to the suppression of cap size *K*, the large-degree nodes absorb more and more small components or isolate nodes into the components, to which these hub nodes belong. Consequently, a mass of hub-centric components with sizes being sublinear to the system size $N$ appear in the subcritical regime. As shown in Fig. \[cluster\] (b), there are many components with size of about one thousand nodes at $t=0.6$. In the subsequent procedure, each added edge would combine two of these hub-centric components with a very small probability, and thus leads to a discontinuous jump of the largest component. As shown in Fig. \[cluster\], three discontinuous jumps of the largest component size occur in the time duration $[0.6,0.65]$, $(0.65,0.7]$ and $(0.7,0.85]$. When $q$ is large enough, i.e., $q>q^{\uppercase\expandafter{\romannumeral2}}_c$, the cap value $K$ that constrains the growth of the largest component won’t work anymore. Affected by the mechanism of preferential attachment, the large-degree nodes absorb small components or isolate nodes continuously, and giant components form at the early stage of the evolution process. Thus, it leads to the continuous growth of the largest component. As shown in Fig. \[cluster\] (c), several components with size of about ten thousand nodes form at $t=0.5$. We see that the large components absorb more small components, thus several giant components form at $t=0.6$. In the following time steps, these giant components are merged by edges, and at last the two stable giant component exist in the supercritical regime, e.g., $t=0.7$ and $t=0.95$. If $q$ approaches to 1, at each time step, the probability of selecting the second node approximates to $\prod_{i}\sim k_i$. The isolate nodes or small components will be absorbed into the large components continuously by the large degree nodes. And when $q=1$, we can imagine that there will be a single giant component to which those large-degree nodes belong, and it grows continuously all through the process. Thus, there will be no jump of the largest component. Discussion {#sec:dis} ========== In this paper, we propose a network percolation model with the hybrid attachment based on the BFW process. In the model, the tune parameter $q$ controls the evolution strategy. When $q$ is relatively small, the strategy of random edge selection dominates the evolution process. When $q$ becomes larger, the strategy of preferential edge selection dominates the evolution process. A large number of numeric simulations indicate that there exist crossover phenomena of percolation transition between three separate regions $\left[0,q^{\uppercase\expandafter{\romannumeral1}}_c \right]$, $\left(q^{\uppercase\expandafter{\romannumeral1}}_c,q^{\uppercase\expandafter{\romannumeral2}}_c \right]$, $\left(q^{\uppercase\expandafter{\romannumeral2}}_c, 1\right]$. In the region $q\in[0,q^{\uppercase\expandafter{\romannumeral1}}_c]$, the strategy of random edge selection dominates the evolution process, and with the constraint of forming the giant component, there is only one discontinuous jump of the largest component at the percolation threshold as the origin BFW model. In the region $q\in(q^{\uppercase\expandafter{\romannumeral1}}_c,q^{\uppercase\expandafter{\romannumeral2}}_c]$, the strategy of more preferential edge selection dominates the evolution process, and multiple discontinuous jumps occur as a result of competition between the mechanisms of constraining and accelerating the forming of the giant component, which are caused by the restricted of the cap size $K$, and the preferential selection of a candidate edge respectively. Further, in the region $q\in(q^{\uppercase\expandafter{\romannumeral2}}_c,1]$, with nearly complete preferential selection, the mechanism of constraining the giant connected component is no longer function, the largest component grows continuously at the first transition point. And if $q$ approaches to 1, a single giant component grows continuously throughout the evolution process. Percolation of networks has been one of the hottest topics in recent years, and it is considered as an important tool to study the robustness of networks [@parshani2010interdependent], spreading of epidemics [@newman2002spread] and so on. Although many important and interesting percolation phenomena are observed on both random networks and lattices, it is the first time to study the percolation in network evolution process with hybrid attachment, which will help us to understand the percolation transition in a more realistic process. It is unusual to find the crossover phenomena of percolation transition in the network evolution process, which have recently stimulated broad attention in other fields, such as the processes of social contagions [@wang2016dynamics]. Specially, a recent work studied the percolation transition in scale-free networks and the multiple discontinuous percolation transitions are observed [@chen2015multiple]. We study the percolation of networks in a more realistic evolution process from the aspect of numeric simulations, but its theoretical analysis need to be carried out. Besides, the percolation transition in real-world networks needs further works. This work was partially supported by the National Natural Science Foundation of China under Grants Nos. 11105025, 11575041 and 61433014, and Project No. 9140A06030614DZ02083. [100]{} G. Grimmett. *What is Percolation* (Springer, 1999). S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, D.-U. Hwang, Phys. Rep. **4**, 424 (2006). C. Castellano, S. Fortunato, V. Loreto, Rev. Mod. Phys. **81**, 591 (2009). M. Newman. *Networks: an introduction* (OUP Oxford, 2010). C. Moore, M. E. Newman, Phys. Rev. E **61**, 5678 (2000). X. Fu, M. Small, D.M. Walker, H. Zhang, Phys. Rev. E **77**, 036113 (2008). P. Erd[ö]{}s, A. R[é]{}nyi, Publ. Math. Debrecen **6**, 290 (1959). P. Erd[ö]{}s, A. R[é]{}nyi, Publ. Math. Inst. Hungar. Acad. Sci **5**, 17 (1960). D. Achlioptas, R. M. D’Souza, J. Spencer, Science **323**, 1453 (2009). R. M. Ziff, Phys. Rev. Lett. **103**, 045701 (2009). Y. S. Cho, J. S. Kim, J. Park, B. Kahng, D. Kim, Phys. Rev. Lett. **103**, 135702 (2009). F. Radicchi, S. Fortunato, Phys. Rev. Lett. **103**, 168701 (2009). E. J. Friedman, A. S. Landsberg, Phys. Rev. Lett. **103**, 255701 (2009). O. Riordan, L. Warnke, Science **333**, 322 (2011). H. K. Lee, B. J. Kim, H. Park, Phys. Rev. E **84**, 020101 (2011). N. A. Araujo, H. J. Herrmann, Phys. Rev. . **105**, 035701 (2010) Y. S. Cho, S. Hwang, H. J. Herrmann, B. Kahng, Science **339**, 1185 (2013). W. Chen, R. M. D’Souza, Phys. Rev. Lett. **106**, 115701 (2011). R. M. D’Souza, J. Nagler, Nature Phys. **11**, 531 (2015). J. Nagler, A. Levina, M. Timme, Nature Phys. **7**, 265 (2011) R. A. da Costa, S. N. Dorogovtsev, A. V. Goltsev, J. F. F. Mendes, Phys. Rev. Lett. **105**, 255701 (2010). P. Grassberger, C. Christensen, G. Bizhani, S.-W. Son, M. Paczuski, Phys. Rev. Lett. **106**, 225701 (2011). T. Bohman, A. Frieze, N. C. Wormald, Random Struct. Algor. **25**, 432 (2004). H. D. Rozenfeld, L. K. Gallos, H. A. Makse, Eur. Phys. J. B **75**, 305 (2010). R. K. Pan, M. Kivel[ä]{}, J. Saram[ä]{}ki, K. Kaski, J. Kert[é]{}sz, Phys. Rev. E **83**, 046112 (2011). A.-L. Barab[á]{}si, R. Albert, Science **286**, 509 (1999). M. E. Newman, Phys. Rev. E **64**, 025102 (2001). B. F. de Blasio, [Å]{}. Svensson, F. Liljeros, Proc. Natl. Acad. Sci. USA **104**, 10762 (2007). P. Holme, B. J. Kim, Phys. Rev. E **65**, 026107 (2002). M. Small, Y. Li, T. Stemler, K. Judd, Phys. Rev. E **91**, 042801 (2015). M. E. Newman, S. H. Strogatz, D. J. Watts, Phys. Rev. E **64**, 026118 (2001). T. J. Fararo, M. H. Sunshine. *A study of a biased friendship net* (Youth Development Center, Syracuse University, Syracuse, 1964). Z. Liu, Y.-C. Lai, N. Ye, P. Dasgupta, Phys. Lett. A **303(5)**, 337 (2002). Z. Liu, Y.-C. Lai, N. Ye, Phys. Rev. E **67(3)**, 031911 (2003). P. Shu, W. Wang, M. Tang, Chaos **25**, 063104 (2015). K. Binder, Z. Phys. B: Conden. Matter **43**, 119 (1981). R. F. Voss, J. Phys. A: Math. Gen. **17**, L373 (1984). R. Parshani, S. V. Buldyrev, S. Havlin, Phys. Rev. Lett. **105**, 048701 (2010). M. E. J. Newman, Phys. Rev. E **66**, 016128 (2002). W. Wang, M. Tang, P. Shu, Z. Wang, New J. Phys. **18**, 013029 (2016). W. Chen, Z. M. Zheng, R. M. D’Souza, J. Stat. Mech.: Theory E, **2015**, P04011 (2015).

ArXiv

--- abstract: 'We theoretically investigate spin decoherence of a single nitrogen-vacancy (NV) center in diamond. Using the spin coherent state P-representation method, coherence evolution of the NV center surrounded by nitrogen electron spins (N) is simulated. We find that spin decoherence time as well as free-induction decay of the NV center depend on the spatial configuration of N spins. Both the spin decoherence rate (1/$T_2$) and dephasing rate (1/$T_2^*)$ of the NV center increase linearly with the concentration of the N spins. Using the P-representation method, we also demonstrate extracting noise spectrum of the N spin bath, which will provide promising pathways for designing an optimum pulse sequence to suppress the decoherence in diamond.' author: - 'Zhi-Hui Wang' - Susumu Takahashi title: 'Spin decoherence and electron spin bath noise of a nitrogen-vacancy center in diamond' --- Nitrogen-vacancy (NV) impurity centers in diamond have been investigated extensively for fundamental research [@gruber97; @Jelezko02] and potential applications of quantum information processing devices [@JelezkoGate04; @Jiang09; @Neumann10], and a high-precision room-temperature magnetic sensor [@Maze08; @Balasubramanian08]. Coherent properties of NV centers in diamond play a crucial rule in the applications. Spin decoherence is due to couplings to noisy environments. For a NV center in diamond, major noise sources are paramagnetic impurities and nuclear spins. Diamond crystals containing $>$ 10 ppm of substitutional single-nitrogen (N) impurities are called type-Ib diamonds and the spin decoherence is caused by fluctuating N spin baths [@kennedy03; @Takahashi08; @hanson08]. On the other hand, in diamond containing much less N spins, the spin decoherence time $T_2$ is much longer and the decoherence is often limited by couplings to $^{13}$C nuclear spins ($\sim$1.1 $\%$ natural abundance) [@childress06; @Maze08b; @Balasubramanian08]. Understanding dynamics of an electron spin system and surrounding electron and nuclear spin baths has been a long-standing problem for theoretical investigation on electron spin resonance (ESR) in solids. The lineshape and width of continuous ESR spectrum as well as pulsed ESR signals including free induction decay (FID) and spin echo (SE) decay were successfully described by approximating spin baths by stochastic noise fields [@KlauderAnderson; @Chiba72; @Zhidomirov69; @deSousa03]. In order to realize spin-based quantum bits (qubits) in solid-state systems [@Kane98; @Loss98], investigation of spin baths is becoming more critical to understand decoherence. Methods like dynamical decoupling [@viola98; @khodjasteh05; @uhrig07] has been studied to suppress the decoherence in knowledge of noise spectrum of surrounding spin baths [@medford12; @Cywinski08PRB]. In semiconductor quantum dots [@hanson07], silicon [@MortonReview11], and diamond with very low concentration of N spins [@childress06; @Maze08b; @RBLiu12], nuclear spins are the major source of the decoherence in the system. The dipolar coupling between nuclear bath spins is much weaker than the hyperfine (HF) coupling between an electron (central) spin qubit and bath nuclear spins (the system-bath coupling). In such cases, microscopic treatments of nuclear spin baths have been thoroughly studied using the cluster expansion method [@Witzel05; @Witzel06; @Saikin07; @RBLiu06; @RBLiu07] in which the intra-bath coupling is treated as a perturbation to the system-bath coupling. On the other hand, for electron spin baths where the strength of the intra-bath coupling is comparable to that of the system-bath coupling (the strong intra-bath coupling regime), it is often challenging to implement the cluster expansion method [@Witzel10; @Witzel12]. In type-Ib diamond, the dipolar coupling strength between the NV center and the N spins is of the same order of that between N spins. Although it has been shown that experimentally observed FID and SE signals of a single NV center have been well described by a classical stochastic noise, the Ornstein-Uhlenbeck (O-U) process [@hanson08; @deLange10], spin dynamics of a NV center and surrounding N spins has not been fully understood yet. Variations of the P-representations [@glauber63; @sudarshan63; @drummond80] have been successfully used to formulate many-body problems in quantum optics, [*e.g.*]{} squeezing in optical solitons [@carter87]. The spin coherent state P-representation, as a variation of the time-dependent mean-field method, has been proposed to apply to spin-based qubit systems targeting at decoherence of the central spin [@PrepPRL06; @Zhang07JOP]. The wavefunction of the whole system is represented in the basis of direct product of wavefunctions for individual spins. Equations of motion for all spins are specially tailored to achieve a close approximation to quantum dynamics of the central spin. Simulation of FID and SE for a nuclear spin system has been demonstrated using the spin coherent P-representation method [@Zhang07JOP]. In this paper, we theoretically study spin decoherence of a single NV center in diamond with the N electron spin bath. We employ the P-representation method to simulate spin dynamics of the NV center and surrounding N spins. Simulated SE decays agree with the decays due to a classical noise field described by the O-U process. Our simulation shows that the decay rate of both FID ($1/T^*_2$) and SE ($1/T_2$) depend linearly on the concentration of N spins in the range from 1 to 100 ppm. Noise spectrum of N spins is also extracted using the P-representation method. The simulated noise spectrum is in good agreement with the noise spectrum of the O-U process in the range of high frequencies with which the spectrum of the SE sequence overlaps significantly. We consider a NV center in diamond ($S$ = 1) under a static magnetic field $B_0$ applied along the N-V axis (denoted as the $z-$axis). The Hamiltonian of the NV center system is $$\begin{aligned} H_S=D(S_0^z)^2+\gamma_0 B_0 S^z_0\end{aligned}$$ where $D$ = 2.87 GHz is the zero-field splitting due to the axial crystal field, $\gamma_0$ is the gyromagnetic ratio for the NV center, and $S_0$ is the electron spin of the NV center. The second term corresponds to the Zeeman energy of $S_0$. The HF coupling between $S_0$ and the nitrogen nuclear spin in the NV center is not considered. The degeneracy between $m_z=+1$ and $m_z=-1$ states is lifted by the external magnetic field $B_0$. We consider only the $m_z=-1$ and $m_z=0$ transition to treat the NV center as a two-level system, $s_0$. The Hamiltonian for an individual N spin ($S$ = 1/2) is $$\begin{aligned} \label{eq:Hami} H_k=\gamma B_0 S^z_k+ A_1S^z_kI^z_k\end{aligned}$$ where $\gamma$ is the gyromagnetic ratio for the N spin and the HF coupling constant is $A_1 = 114$ MHz for N spins delocalized along the $[111]$ axis and 86 MHz along the $[11\bar{1}]$, $[1\bar{1}1]$ and $[\bar{1}11]$ axes [@Loubser78]. We here consider cases where $B_0$ is not equal to 514 G, therefore there is a large mismatch in the transition energy of NV and N spins which suppresses the flip-flop process between NV and N spins greatly. In the rotating frame with the precession frequency of the N spins, $\gamma B_0 S_k$, and with the NV center, $(D-\gamma_0 B_0) s_0$, the Hamiltonian for the dipolar interaction between the NV and the N spins (the system-bath coupling) is given by [@hanson08], $$\begin{aligned} H_{SB}=(s_0^z-1/2)\sum_k A_k S^z_k\end{aligned}$$ where $A_k=[1-3(n^z_k)^2]a_k$ and $a_k=\hbar\gamma_0\gamma/r_k^3$ is the coupling constant. The Hamiltonian for the dipolar coupling between N bath spins $S_j$ and $S_k$ is given by $$\begin{aligned} H_{B}=\sum_{j,k} c_{j,k}[1-3(n^z_{j,k})^2] [S^z_jS^z_k-\frac{1}{4}(S^+_jS^-_k+S^-_jS^+_k)]\end{aligned}$$ where $c_{j,k}=\hbar\gamma^2/r_{j,k}^3$ and only secular terms are considered. ![ \[Fig1\] (a) A Schematic of a spatial configuration for a NV and N spins. The cube consists of the diamond lattice. A NV center (red large sphere) is located at $(0,0,0)$ and N spins (blue small spheres) are located randomly on diamond lattice sites. (b) Simulated FID signals of a NV center with three different N spin bath configurations. The concentration of N spins is $f = 10~$ppm for all cases. Blue crosses, green squares and red circles are simulated results, and lines are fit to $\exp[-(t/T^*_2)^2]$. Obtained $T^*_2=0.97$, $0.69$, and $0.61~\mu s$ from the fit are in good agreement with calculated values directly from the N spin bath configuration where $T^*_2=\sqrt 2/b$ are $0.95$, $0.67$, and $0.62$ $\mu s$ respectively. The inset shows a histogram of $b$ for the 90 instances with $f=10$ ppm. Black solid line represents the expected probability distribution of $b$ (see the supplemental material for details). ](Fig1.ps){width="100"} Fig. \[Fig1\] shows a spatial configuration of a NV center ($S_0$) and N spin ($S_k$) bath used in our simulation. The NV center spin is located at the center of a cube where the cube consists of a tetrahedral diamond unit cell with $a = 3.567\AA$ of the lattice constant. The number of the unit cells in the cube is $N_a^3$ ($N_a$ is the number of unit cell along the $x$,$y$ and $z$-axes). N spins are randomly distributed on the lattice sites. $N_a$ and the number of N spins (typically 80 $\sim$ 100 N spins are used) are adjusted according to the concentration of N spins ($f$). In the simulation, M initial vectors are sampled to represent unpolarized N spin bath corresponding to a high temperature limit of the bath, $T\gg \hbar\gamma B_0/k_B$, $\big({\bf R}_0^{(m)},{\bf R}_1^{(m)},\cdots,{\bf R}_J^{(m)}$) and $m=1,\ldots M$, are prepared where $J$ is the number of N spins and ${\bf R}_k=(\theta_k,\phi_k)$ is a classical vector on the Bloch sphere of spin $k$ (the $k$=0 spin is a NV center). Time evolution of the vector, ${\bf R}_k(t)$, is calculated according to a set of equations of motion $\dot {\bf R}_k={\bf B}_{k}\times {\bf R}_k$ where ${\bf B}_{k}$ is a local magnetic field for spin $k$ induced by other spins in the system. Coherence at time $t$ is given by calculating $\langle s^x_0(t) \rangle = \frac{1}{M}\sum_{m=1}^M R^{x,(m)}(t)$ (see the supplemental material for details). Simulated coherence are shown in Fig. \[Fig1\](b). The coherence decays as a function of time, and the decay corresponds to FID signals of a NV center. As shown in Fig. \[Fig1\](b), simulated FIDs are fit well by the fundamental Gaussian function. The Gaussian shape of the FID agrees with that of FID signals experimentally observed from a single NV center in type-Ib diamond [@hanson08]. The root-mean-square of the spin-bath coupling $b=\frac{1}{2}\sqrt{\sum_j A^2_j}$ quantifies the FID for the NV spin, [*i.e.*]{} $T^*_2=\sqrt 2/b$. Thus the FID time depends on the configuration of the local bath spins around the NV center. Fig. \[Fig1\](b) shows the simulated FID signals for the single NV center with different N bath configurations for $f=10~$ppm of the N spin concentration. We found that the values of $b$ obtained from $T^*_2$ agree with $b$ directly calculated from the configuration of the N bath spins. Inset of Fig. \[Fig1\](b) shows a histogram of $b$. The distribution of $b$ agrees with the theoretically expected distribution [@Dobrovitski08]. To represent a typical configuration of the N spin bath, simulated results with any pairs of spins coupled much stronger (50 times) than typical coupling strength are excluded. A deviation at high $b$ values is due to this exclusion (see the inset of Fig. \[Fig1\](b)). We obtained $T^*_2$ = $0.02\sim 1.73~\mu s$ with 0.53 $\mu s$ of the the mean value for $f=10~$ppm of the N spin concentration. ![ \[Fig2\] (a) Simulated SE and FID signals with $f=10~$ppm. Circles and crosses are simulation results, and a red solid line shows a fit to Eq. (\[eq:T2KA\]). With $b=1.44~\mu s^{-1}$ from the FID data, we obtained $\tau_C$ = $2.78~\mu s$. (b) Simulated SE signals for 90 instances of the N spin bath configuration with $f=10~$ppm. Blue circles are simulation results, and green lines are fits to Eq. (\[eq:T2KA\]). The inset shows a histogram of $\alpha$ for the 90 instances when SE signals are fitted by $\exp[-(t/T_2)^\alpha]$. ](Fig2.ps){width="120"} As shown in Fig. \[Fig2\](a), we simulated the time evolution of SE. The rephasing $\pi-$pulse is assumed to be perfect and instantaneous in the simulation. SE decay in electron spin baths has been described by treating the bath to be a classical noise field where the noise field $B(t)$ was modeled by the O-U process with the correlation function $C(t) = \langle B(0)B(t)\rangle = b^2 \exp(-|t|/\tau_C)$. $\tau_C$ is the correlation time of the bath, which measures the rate of the flip-flop process between the bath spins. The corresponding noise spectrum is Lorentzian with power $b^2$ and the half-width at half-maximum (HWHM) $1/\tau_C$. SE decay subject to the noise due to the O-U process is given by [@KlauderAnderson], $$\begin{aligned} \label{eq:T2KA} E(t)&=&\exp[-(b\tau_C)^2[t/\tau_C-3-e^{-t/\tau_C}+4e^{-t/(2\tau_C)}]].\end{aligned}$$ In the quasi-static limit ($b\tau_c \gg 1$) indicating slow bath dynamics, $E(t)=e^{-b^2t^3/(12\tau_C)}\sim\exp[-(t/T_2)^3]$. On the other hand, in the motional-narrowing limit ($b\tau_c \ll 1$), $E(t)=e^{-t/\tau_C}\sim\exp(-t/T_2)$. As shown in Fig. \[Fig2\](a), we found good agreement between our simulation results and Eq. (\[eq:T2KA\]). Using the value of $b$ determined by the bath configuration and confirmed by the FID, we determined $\tau_C$. Fig. \[Fig2\](b) shows 90 simulated SE decays with $f$=10 ppm. For many cases, we found $b\tau_C>1$, and SE decays are well described by $\exp[-(t/T_2)^\alpha]$ where the exponent $\alpha$ is typically between $2\sim 3$ as shown in inset of Fig. \[Fig2\](b). This is in consistence with previous experimental results ($\alpha \sim 3$) [@deLange10; @ZHNVDD]. We also found a few cases in the motional-narrowing regime ($b\tau_C<1$), in which the SE decay shapes are close to a single exponential function. ![ \[Fig3\] (a) $1/T^*_2$ as a function of the N spin concentration. Each cross is the mean value and the error bar is the standard deviation of all instances. Black solid line is a fit to a linear function. (b) $1/T_2$ as a function of the N spin bath concentration. Each blue circle is the mean value and the error bar is the standard deviation of all instances. Blue solid line is a fit to a linear function. The red cross and its error bar are the results with the HF interaction between N electron and $^{14}$N nuclear spins taken into account. Difference in simulated $1/T_2$ between with and without the HF coupling is smaller than the deviation (see the supplemental material for details). ](Fig3.ps){width="120"} We examined the concentration dependence of the FID and SE decay in the range from $1$ to $100$ ppm. In Fig. \[Fig3\](a) and (b), $1/T^*_2$ and $1/T_2$ are shown for the concentrations of $f = 1,~5,~10,~50,~100$ ppm. For each value of $f$, $80\sim 100$ random configurations of the bath are simulated and the mean of $1/T^*_2$ and $1/T_2$ with the standard deviation as an error bar are shown. We found a linear dependence on $f$ for both $1/T^*_2$ and $1/T_2$, namely $1/T^*_2 = 0.19f$ and $1/T_2 = 5\times10^{-2} f$ corresponding to $T_2 = 2.03$ and $0.2~\mu s$ for $f$ = 10 and 100 ppm respectively. The linear dependence of the concentration has been experimentally seen in $1/T_2$ of ensemble N spins in diamond [@Wyk97]. In addition, a linear concentration dependence of $1/T_2$ has been reported in phosphorous donors in silicon [@Witzel10]. The linear dependence can be understood as stemming from the dipolar nature of both the spin-bath and the intra-bath couplings. Consider the case of $b\tau_C\gg 1$. The average spin-bath coupling $\overline A_j \propto 1/\bar r^3\propto f$, hence $\overline {b^2}\propto f^2$, where $\bar r$ is the average distance between neighboring spins. Similarly for N bath spins, $\overline {\tau_C}$ is roughly proportional to average coupling, which is proportional to $1/f$. For $b\tau_C>1$, the expansion of Eq. \[eq:T2KA\] to the leading order in $t/\tau_C$ yields $\overline {1/T_2}\propto(\overline{b^2}/\overline{\tau_C})^{1/3}$, therefore $1/T_2$ is proportional to $f$. Finally we simulated noise spectrum of the N spin bath, $S(\omega)$. At time $t$, the local magnetic field due to N spins at the NV spin is expressed by $B_{NV}(t) \sim \sum_{j} A_{0,j}\langle R^z_j(t)\rangle$ where the sum is taken over all N bath spins. Direct Fourier transform of $B_{NV}(t)$ renders the noise spectrum, $S(\omega)$. Fig. \[fig:noiseSpec\] shows the simulated noise spectrum of N spins. Coherence of the NV center decays due to couplings to environmental noise of N spins. The amount of the decay is determined by overlaps between the noise spectrum and the spectrum of a pulse sequence used in ESR measurements. In SE measurement, the power spectrum of the pulse sequence is given by $|\sin^2(\omega t/4)/(\omega/4)|^2$ where $t$ is the total evolution time [@Cywinski08PRB]. As shown in Fig. \[fig:noiseSpec\], with $t=T_2$, significant overlap between the noise and the pulse-sequence spectrum happens at high frequencies where the noise spectrum agrees well with the spectrum of the O-U process. This supports the observation of good agreement between our simulated SE signals and the analytical solution with the O-U process (see Eq. \[eq:T2KA\]). On the other hand, in the present case, the dephasing time $T_2^*$ of the FID signals is determined by $e^{-t^2\int^{+\infty}_{-\infty}S(\omega)d\omega}$ where $T_2^*$ does not depend on details of the noise spectrum $S(\omega)$. We found excellent agreement in between $T_2^*$ obtained from the noise spectrum and $T_2^*$ extracted from simulated FID. ![ \[fig:noiseSpec\] Noise spectrum of N spins with $f=10~$ppm. Blue circles are the simulated noise spectrum. Red solid line shows the noise spectrum of the O-U process expressed by the Lorentzian function with $1/\tau_C=0.134~\mu s^{-1}$ where $\tau_C$ is extracted from the SE signal. Black dotted line shows the power spectrum of SE pulse sequence $|\sin^2(\omega t/4)/(\omega/4)|^2$ with $t = T_2 = 3.8\mu s$. ](Fig4.ps){width="120"} In summary, we investigated spin decoherence of a single NV center coupling to N electron spins using the P-representation method. The P-representation is a modified time-dependent mean-field technique that is suitable to simulate coherence evolution of a central spin surrounded by spin baths. Simulated results for SE signals are in good agreement with analytical expressions based on the O-U process and with previous experimental results. We found that the FID and SE decay rates, $1/T_2^*$ and $1/T_2$, depend linearly on the concentration of N spins. We also demonstrated simulating the noise spectrum of the electron spin bath using the P-representation method. The P-representation method is suitable to study decoherence in other spin-based qubit systems even in the strong intra-bath coupling regime, [*e.g.*]{} phosphorus donors in isotropically pure silicon ($^{29}$Si $< 50 $ppm). This capability will facilitate us to find optimum DD sequences to effectively suppress decoherence in spin-based qubit systems. We thank to V. V. Dobrovitski and D. A. Lidar for useful discussions. This work was supported by the Searle scholars program (S.T.). [44]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , , , , ****, (). , , , , , , , ****, (). , , , , , , ****, (). , , , , , , , , , , , ****, (). , , , , , , , , , , , ****, (). , , , , , , , , , , , ****, (). , , , , , , , , , , , ****, (). , , , , , ****, (). , , , , , ****, (). , , , , , ****, (). , , , , , , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , , , , ****, (). , , , , ****, (). , , , , , ****, (). , ****, (). , , , ****, (). , , , **** (). , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , , , , , ****, (). , , , , ****, (). , , , , , ****, (). , ****, (). , ****, (). , ****, (). , , , , ****, (). , , , , ****, (). , , , , ****, (). , ****, (). , , , , ****, (). , , , , , ****, (). , , , , ****, ().

ArXiv

--- abstract: '[ Facial attributes, emerging soft biometrics, must be automatically and reliably extracted from images in order to be usable in stand-alone systems. While recent methods extract facial attributes using deep neural networks (DNNs) trained on labeled facial attribute data, the robustness of deep attribute representations has not been evaluated. In this paper, we examine the representational stability of several approaches that recently advanced the state of the art on the CelebA benchmark by generating adversarial examples formed by adding small, non-random perturbations to inputs yielding altered classifications. We show that our fast flipping attribute (FFA) technique generates more adversarial examples than traditional algorithms, and that the adversarial robustness of DNNs varies highly between facial attributes. We also test the correlation of facial attributes and find that only for related attributes do the formed adversarial perturbations change the classification of others. Finally, we introduce the concept of natural adversarial samples, i.e., misclassified images where predictions can be corrected via small perturbations. We demonstrate that natural adversarial samples commonly occur and show that many of these images remain misclassified even with additional training epochs, even though their correct classification may require only a small adjustment to network parameters. ]{}' address: 'University of Colorado Colorado Springs, Vision and Security Technology (VAST) Lab, 1420 Austin Bluffs Parkway, Colorado Springs, CO 80918, USA' author: - Andras - Manuel - 'Ethan M.' - 'Terrance E.' bibliography: - 'refs.bib' title: 'Facial Attributes: Accuracy and Adversarial Robustness' --- Facial attributes; Adversarial Images; Deep learning [ ]{} Acknowledgment {#acknowledgment .unnumbered} ============== This research is based upon work funded in part by NSF IIS-1320956 and in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 2014-14071600012. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon.

ArXiv

--- abstract: 'The Galactic Ridge X-ray Emission (GRXE) spectrum has strong iron emission lines at 6.4, 6.7, and 7.0 keV, each corresponding to the neutral (or low-ionized), He-like, and H-like iron ions. The 6.4 keV fluorescence line is due to irradiation of neutral (or low ionized) material (iron) by hard X-ray sources, indicating uniform presence of the cold matter in the Galactic plane. In order to resolve origin of the cold fluorescent matter, we examined the contribution of the 6.4 keV line emission from white dwarf surfaces in the hard X-ray emitting symbiotic stars (hSSs) and magnetic cataclysmic variables (mCVs) to the GRXE. In our spectral analysis of 4 hSSs and 19 mCVs observed with Suzaku, we were able to resolve the three iron emission lines. We found that the equivalent-widths (EWs) of the 6.4 keV lines of hSSs are systematically higher than those of mCVs, such that the average EWs of hSSs and mCVs are $179_{-11}^{+46}$ eV and $93_{-3}^{+20}$ eV, respectively. The EW of hSSs compares favorably with the typical EWs of the 6.4 keV line in the GRXE of 90–300 eV depending on Galactic positions. Average 6.4 keV line luminosities of the hSSs and mCVs are $9.2\times 10^{39}$ and $1.6\times 10^{39}$ photons s$^{-1}$, respectively, indicating that hSSs are intrinsically more efficient 6.4 keV line emitters than mCVs. We compare expected contribution of the 6.4 keV lines from mCVs with the observed GRXE 6.4 keV line flux in the direction of $(l,b) \approx (28.5\arcdeg, 0\arcdeg$). We conclude that almost all the 6.4 keV line flux in GRXE may be explained by mCVs within current undertainties of the stellar number densities, while contribution from hSSs may not be negligible.' author: - 'Romanus <span style="font-variant:small-caps;">Eze</span>, Kei <span style="font-variant:small-caps;">Saitou</span>, and Ken <span style="font-variant:small-caps;">Ebisawa</span>' title: ' Iron Fluorescent Line Emission from the mCvs and Hard X-ray Emitting Symbiotic Stars as a Source of the Iron Fluorescent Line Emission from the Galactic Ridge ' --- Introduction ============ Presence of the seemingly extended hard X-ray emission from the Galactic Ridge has been known since early 1980’s (Galactic Ridge X-ray Emission; GRXE: [@worrall1982; @warwick1985; @koyama1986]). Strong iron K-line emission at $\sim$6.7 keV in the GRXE indicates its thermal plasma origin (e.g., [@koyama1986; @yamauchi1993]). More precise iron line diagnostics of the GRXE has been made possible with X-ray CCD cameras on-board ASCA [@kaneda1997] and Chandra [@ebisawa2005]. These instruments revealed that the line centroid energies in the GRXE are systematically lower than 6.7 keV (the energy expected from He-like ion in thermal equilibrium plasma), which imply that either the line emission is from non-ionization equilibrium plasma or there is an additional 6.4 keV fluorescent line emission. Suzaku, for the first time, resolved the GRXE iron line emission into three narrow lines, ones from neutral or low ionized (6.4 keV), He-like (6.7 keV), and H-like (7.0 keV) ions [@ebisawa2008], concluding that the GRXE iron line emission is both from hot thermal plasmas and fluorescence by cold materials. Regarding the origin of the GRXE, there is a strong argument in favor of collection of faint point sources as opposed to the diffused emission (e.g., [@revnivtsev2006; @krivonos2007; @revnivtsev2009; @revnivtsev2010] and references therein), although the question remains “what are these Galactic point sources?” Candidate point sources for the GRXE are cataclysmic variables (CVs) and active binaries (ABs). CVs are known to have such strong emission lines and hard spectra, but most of them are brighter than $\sim$$10^{31}$ erg s$^{-1}$, and their population may not be sufficient to account for all the GRXE. @revnivtsev2009 proposed that ABs dimmer than $\sim$$10^{31}$ erg s$^{-1}$ are likely candidates to account for the majority of the GRXE. However, ABs are well known to have thermal but much softer continuum spectra than CVs. @yuasa2010b proposed that intermediate polars (IPs), which are a subclass of magnetic CVs (mCVs), are main sources of the hard X-ray emission of the GRXE, giving support to the point source scenario for the origin of the GRXE. @yuasa2010b concluded that combination of IPs and ABs will explain most of the 6.7 keV and 7.0 keV emission lines in the GRXE as well as the continuum emission above 20 keV, whereas an *ad hoc* additional 6.4 keV line component is needed to explain the entire GRXE by the point source model. This X-ray fluorescence is, on the other hand, believed by some authors to be due to irradiation of the molecular clouds by X-ray photons or it may be as result of cosmic-ray particle bombardment [@koyama1986; @dogiel1998; @murakami2000; @valinia2000; @koyama2007; @yusefzadeh2007; @dogiel2009; @capelli2011]. This is also likely, since Galactic $\gamma$-ray *diffuse* emission above $\sim$100 keV is successfully explained by the cosmic-rays and interstellar matter interaction model (e.g., [@strong2005]). In this paper, we study origin of the 6.4 keV emission line in the GRXE, examining if this emission could be fully resolved by collection of point sources. We focused on hard X-ray emitting symbiotic stars (hSSs) and magnetic CVs (mCVs) observed with the Suzaku satellite [@mitsuda2007], since they are known to be significant 6.4 keV line emitters (e.g., [@ezuka1999; @yuasa2010a; @luna2007; @smith2008; @kennea2009; @eze2011]). We studied 4 hSSs and 19 mCVs (one polar and 18 IPs), all observed with the Suzaku satellite, and estimated their contributions to the 6.4 keV line emission flux of the GRXE. Our goal is to determine if they are the main sources of the GRXE 6.4 keV line emission flux, or if some additional sources are required. Data Selection ============== Our target sources, hSSs and mCVs, were selected based on the fact that they have been observed with Suzaku and have strong Fe K$\alpha$ emission lines with hard X-ray emission above 20 keV. All the four hard X-ray emitting symbiotic stars, SS73 17, RT Cru, T CrB, and CH Cyg observed with Suzaku were selected. In selecting the mCVs (polars and IPs), we used a CV catalog [@ritter2008] and the IP catalog[^1]. Five sources in the catalog, AE Aqr, AM Her, GK Per, 1RXS J070407.9+26250, and 1RXS J180340.0+40121 were dropped, even though observed with Suzaku, because they appear to had been too faint during their observations or have particular emission mechanism (AE Aqr: e.g., [@wynn1997]). A total of 23 sources were thus selected (table \[t1\]). Data Analysis and Results ========================= Analysis of our data were done using version 2 of the standard Suzaku pipeline products, and the HEASoft[^2] version 6.10. In majority of the sources we used $250\arcsec$ radius to extract all events for the XIS detector for the production of the source spectra but in some cases where the $250\arcsec$ radius overlaps with the calibration sources at the corners, we adjusted the radius accordingly. The XIS background spectra were extracted with $250\arcsec$ radius with no apparent sources and were offset from both the source and corner calibrations. The radius was also adjusted accordingly in some cases where it overlaps with the calibration sources at the corners. Response Matrices File and Ancillary Response File were generated for the XIS detector using the FTOOLS `xisrmfgen` and `xissimarfgen` [@ishisaki2007], respectively. Suzaku XIS 0, 2, and 3 have front-illuminated (FI) chips with similar features, so we merged the spectra of XIS 0 and 3, which we hereafter refer to as XIS FI (XIS 2 has been out of service since November 9, 2006 due to an anomaly). XIS 1 is back-illuminated (BI), and we hereafter refer it to as XIS BI. In the HXD PIN detector analysis, we used the non-X-ray background files and response matrix files appropriate for each observation provided by the Suzaku team. We used the `mgtime` FTOOLs to merge the good time intervals to get a common value for the PIN background and source event files. The source and background spectra extraction were done for each observation using the `xselect` filter time file routine. We corrected for the dead time of the observed spectra using the `hxddtcor` in the Suzaku FTOOLS. According to the standard analysis procedure, exposure time for all observations for the derived background spectra were increased by a factor of 10 to take care of the event rate in the PIN background event file which is made 10 times higher than the real background for suppression of the Poisson errors. We assumed a cosmic background model obtained with the HEAO-1 satellite [@boldt1987]. Spectral analysis of all observations were performed using XSPEC version 12.7. We modeled the spectrum using absorbed bremsstralung model with three Gaussian lines for the three Fe K$\alpha$ emission lines to measure the iron line fluxes. We assumed two types of absorption by full-covering and partial covering matter. Since we were primarily interested in the ion lines, our fitting covers 3–10 keV for the XIS BI, 3–12 keV for the XIS FI and 15–40 keV for the HXD PIN. We ignored energy range below 3 keV in the XIS FI and BI detector to avoid intrinsic absorption which is known to affect data at this energy range, and energies above 10 keV were ignored for XIS BI because the instrument background is higher compared to the XIS FI detectors. We also ignored energy range above 40 keV in the HXD PIN detector in order to obtain high signal-to-noise ratio signals. The three Fe lines, neutral or low ionized (6.4 keV), He-like (6.7 keV), and H-like (7.0 keV) ions, were clearly resolved in all the sources except IGR J17303–0601 where we were unable to detected the H-like (7.0 keV) significantly but the other two lines were detected. Spectra of all the sources can be found in an earlier work by @eze2015 and spectral parameters were shown in table \[t2\]. Furthermore, in order to determine average spectra of the hSSs and the mCVs, we used `addascaspec` to average the spectra of hSSs and mCVs (as well as responses). We used the same model to produce the spectra for the average hSSs and average mCVs. The spectra and spectral parameters for the average hSSs and average mCVs were presented in figure 1 and table \[t2\], respectively. We detected strong 6.4 keV iron line emission in the average hSSs spectrum with an equivalent width (EW) of $179_{-11}^{+46}$ eV and in the average mCVs spectrum with $93_{-3}^{+20}$ eV. We have found that the 6.4 keV line EW is much stronger in hSSs than in mCVs, which suggests that hSSs can be strong candidates of the GRXE 6.4 keV line emission. For comparison, 6.4 keV iron line EWs of the GRXE are of 90–390 eV, depending on the Galactic locations [@yamauchi2009]. Discussion ========== The 6.4 keV Line Emission ------------------------- The 6.4 keV fluorescence line emission is usually due to irradiation of the neutral (or low ionized) material (iron) by a hard X-ray source. @eze2015 used these sources used in this work to study the origin of the 6.4 keV fluorescence line and found that the emission could be partly from the reflection of hard X-rays from the accretion disks/white dwarf surfaces and from the absorption column. Generally, for all compact objects that accretes matter such as white dwarfs, neutron stars, and black holes, there are often emission of the Fe K$\alpha$ fluorescence line as well as Compton reflection which signifies presence of the surrounding cold matter. hSSs and mCVs are both binary systems in which white dwarfs accrete matter from their companions. The CVs are semi-detached systems in which the secondary star fills its Roche lobe and starts transferring mass into the lobe of the compact white dwarf primary. The transferred material has too much angular momentum to fall directly onto the surface of the white dwarf, but instead builds an accretion disk, which spirals round the white dwarf (e.g., [@warner1995]). On the other hand, SSs are interacting binaries formed from a red giant star and a hot degenerate companion which accretes mass from the stellar wind of the red giant. Such a nebulae often formed surrounding the system is typically detected via various optical emission lines, whereas accretion disks may not be formed [@kenyon1986]. Therefore, at least for hSSs, the 6.4 keV fluorescent emission lines are considered to be mainly from white dwarf surfaces. Hard X-rays emitted in the vicinity of the white dwarfs irradiate the white dwarf surfaces (e.g., [@luna2007; @eze2010]), leading to the emission of the Fe K$\alpha$ fluorescence line. We found that the EWs of the 6.4 keV line of hSSs are systematically higher than those of mCVs, such that the average EW of hSSs and mCVs are $179_{-11}^{+46}$ eV and $93_{-3}^{+20}$ eV, respectively (table \[t2\]). In order to see if hSSs are truly more efficient 6.4 keV line emitters than mCVs, we estimated 6.4 keV photon luminosities for those hSSs and mCVs whose distances are known (table \[t3\]). As a result, we found that the 6.4 keV line luminosities of hSSs are systematically higher than those of mCVs; that of the average hSSs is $9.2\times 10^{39}$ photons s$^{-1}$, and that of the average mCVs is $1.6\times 10^{39}$ photons s$^{-1}$. We note that one of the hSSs, RT Cru, has a high luminosity of $30\times 10^{39}$ photons s$^{-1}$, however, if we isolate this source, the average luminosity of hSSs will become $2.2\times 10^{39}$ photons s$^{-1}$, which will still be higher than that of the mCVs. We suppose there is a reason why the 6.4 keV line luminosities of hSSs are significantly higher than those of mCVs, even though hSSs may lack accretion disks. Many hSSs are believed to have a cocoon of gas coming from the red giant companion and surrounding the white dwarf (e.g., [@luna2007; @eze2010; @eze2011]). Hence, there is a possibility that the additional 6.4 keV line emission is from irradiation of the thick cold absorbing circumstellar gas partially covering the hard X-ray source. In fact, our spectral analysis indicate that the average hSSs spectrum has significantly higher hydrogen circumstallar partial covering fraction of $0.70\pm0.01$, compared to that of the average mCVs, $0.44\pm0.01$ with similar hydrogen column density of partial covering matter (table \[t2\]). The Contribution of the 6.4 keV Line Emission of hSSs and mCVs to that of the GRXE ---------------------------------------------------------------------------------- If the Galactic 6.4 keV line source has the line luminosity ($L_{\mathrm{6.4}}$ photons s$^{-1}$) and number the density ($n$ cm$^{-3}$), the 6.4 keV line emissivity may be expressed as $$\begin{aligned} \frac{nL_{\mathrm{6.4}}}{4\pi} \ \textrm{photons~s$^{-1}$~cm$^{-3}$~str$^{-1}$} \label{eq1}\end{aligned}$$ and the observed 6.4 keV line flux in photons s$^{-1}$ cm$^{-2}$ str$^{-1}$ will be $$\begin{aligned} F_{\mathrm{6.4}}= \int \frac{n\left( x\right) L_{\mathrm{6.4}}}{4\pi} dx \approx \frac{{\langle L_{\mathrm{6.4}}\rangle}}{4\pi} \int n(x) dx, \label{eq2}\end{aligned}$$ where the integration is made along the line of sight, and $\langle L_{\mathrm{6.4}}\rangle$ is the average 6.4 keV line luninosity either from hSSs or mCVs (see also [@yamauchi2009]). In order to be precise, we will have to take into account luminosity functions of the hSSs and mCV to estimate $\langle L_{\mathrm{6.4}}\rangle$. However, it is known that contribution of CVs to the GRXE is limited to a narrow range of the luminosities (e.g., @sazonov2006 [@warwick2014]). Thus, we approximate $\langle L_{\mathrm{6.4}}\rangle$ with the values estimated in the previous section from the current Suzaku samples, $\langle L_{\mathrm{6.4}}\rangle \sim9.2 \times 10^{39}$ photons s$^{-1}$ for hSSs and $\sim1.6 \times 10^{39}$ photons s$^{-1}$ for mCVs. Let’s consider the measurement of the 6.4 keV photon flux at the position $(l,b) \approx (28.5\arcdeg, 0\arcdeg)$, $(8\pm 2)\times 10^{-5}$ photons s$^{-1}$ cm$^{-2}$ deg$^{-2} \approx 0.26$ photons s$^{-1}$ cm$^{-2}$ str$^{-1}$ [@ebisawa2008]. We are going to estimate the contributions from mCVs using (\[eq2\]). Following @sazonov2006 [@revnivtsev2007] and @warwick2014, we take the model of the stellar mass distribution on the Galactic plane, $$\rho = \rho_{0, disk} \; \exp \left[ -\left(\frac{R_m}{R}\right)^3 -\frac{R}{R_{disk}} \right],\label{massdensity}$$ where $R_m$ and $R_{disk}$ represents the inner cut-off radius and e-fold disk radius, respectively. These authors claim that almost all the total GRXE flux is explained assuming the stellar density in the form of (\[massdensity\]). However, we found that the three authors adopt slightly different disk size and normalization, while they use the same outer disk radius, $R_{max}$ = 10 kpc: @sazonov2006 takes $\rho_{0, disk}=0.61 M_\odot$ pc$^{-3}$, $R= 3$ kpc, $R_{disk}=3$ kpc. @revnivtsev2007 takes $\rho_{0, disk}=5.5 M_\odot$ pc$^{-3}$, $R_m= 2.5$ kpc, $R_{disk}~2.2$ kpc. @warwick2014 takes the same $R_m$ and $R_{disk}$ as @revnivtsev2007, but $\rho_{0, disk}=1.6 M_\odot$ pc$^{-3}$. Because of these differences, the total surface mass density integrated in the direction of $l=28.5\arcdeg$ is significantly different: 1200 [@sazonov2006], 1900 [@warwick2014] or 6700 [@revnivtsev2007] $M_{\odot}$ pc$^{-2}$ (see Fig. 2). This may not be very surprising, since only less than half of the GRXE is resolved into point sources in this direction [@ebisawa2005], and estimate of the number of remaining extremely dim stars has a large uncertainty. If we assume the expected spatial number density of CVs normalized by stellar mass, $1.2 \times 10^{-5} M_\odot^{-1}$ [@sazonov2006] we may calculate the expected 6.4 keV line photon flux from CVs; 0.19, 0.31 or 1.1 photons s$^{-1}$ cm$^{-2}$ str$^{-1}$, depending on the three different stellar mass density estimates. Obviously, the latter two are over-estimates, exceeding the observed value. In any case, we conclude that the observed 6.4 keV line photon flux in the direction of $l=28.5\arcdeg$ may be explained totally by CVs, within the current uncertainty of the stellar densities. What about the contribution from hSSs? We have very few number of known symbiotic stars 200 [@belczynski200], of which only five are hSSs (see @kennea2009). This makes it difficult to estimate the space density of the hSSs and we therefore require more discoveries of hSSs in the Galaxy in other to make a proper estimate of the space density. However, we believe that there are many more hSSs that are yet to be discovered, hence hSSs could contribute a significant percentage of the GRXE 6.4 keV line flux. In summary, we conclude that the GRXE 6.4 keV line flux is primarily explained by mCVs. Contribution from the hSSs may not be neglected, since hSSs are intrinsically strong 6.4 keV line emitters. Taking account of contributions from other types of white dwarf X-ray binaries, the GRXE 6.4 keV line emission may be fully explained as being from a large number of accreting white dwarfs in the Galactic plane, most of which has yet to be identified. In order to confirm our model, further work should be done to search for more, still dimmer 6.4 keV line emitting sources in the Galactic plane, which may be hSSs, mCVs, or other types of white dwarf binaries. We hope that next generation Galactic surveys by the incoming hard X-ray satellite missions such as Spectrum-Roentgen-Gamma" (SRG) satellite will detect a large number of such sources that would account for most of the GRXE 6.4 keV line emission. We acknowledge the Suzaku team for providing data and some relevant files used in the analysis of this work. R.E. is very grateful to the Japan Society for the Promotion of Science (JSPS) for financial support under the JSPS Invitation Fellowship (Long Term), the Nigerian TETFund for National Research grant support and ISAS/JAXA, Sagamihara Campus for hosting him. This research made use of data obtained from Data ARchives and Transmission System (DARTS), provided by Center for Science-satellite Operation and Data Archives (C-SODA) at ISAS/JAXA. Some part of this work were[^3] Barlow, E. J., Knigge, C., Bird, A. J., Dean, A. J., Clark, D. J., Hill, A. B., Molina, M., & Sguera, V.  2006, , 372, 224 Belczyński, K., Mikolajewska, J., Munari, U., Ivison, R. J., & Friedjung, M.  2000, , 146, 407 Beuermann, K., Harrison, Th. E., McArthur, B. E., Benedict, G. F., & Ga̋nsicke, B. T.  2003, , 412, 821 Boldt, E.  1987, , 146, 215 Bonnet-Bidaud, J. -M., de Martino, D., Falanga, M., Mouchet, M., & Masetti, N.  2007, , 473, 185 Capelli, R., Warwick, R. S., Porquet, D., Gillessen, S., & Predehl, P.  2011, , 530, A38 de Martino, D., Bonnet-Bidaud, J. -M., Mouchet, M., Ga̋nsicke, B. T., Haberl, F., & Motch, C.  2006, , 449, 115 Dogiel, V. A., Ichimura, A., Inoue, H., & Masai, K.  1998, , 50, 567 Dogiel, V. A.,   2009, , 61, 901 Ebisawa, K.,   2005, , 635, 214 Ebisawa, K.,   2008, , 60, S223 Eze, R. N. C., Luna, G, J, M., & Smith, R. K. 2010, , 709, 816 Eze, R. N. C. 2011, Adv. in Space Research, 47, 1999 Eze, R. N. C. 2015, New Astronomy, 37, 35 Ezuka, H., & Ishida, M. 1999, , 120, 277 Ishisaki, Y., 2007, , 59, S113 Kaneda, H., Makishima, K., Yamauchi, S., Koyama, K., Matsuzaki, K., & Yamasaki, N. Y.  1997, , 491, 638 Kennea, J. A., Mukai, K., Sokoloski, J. L., Luna, G. J. M., Tueller, J., Markwardt, C. B., & Burrows, D. N.  2009, , 701, 1992 Kenyon, S. J.  1986, The symbiotic stars (Cambridge: Cambridge University Press) Koyama, K., Makishima, K., Tanaka, Y., & Tsunemi, H.  1986, , 38, 121 Koyama, K.,   2007, , 59, S221 Krivonos, R., Revnivtsev, M., Churazov, E., Sazonov, S., Grebenev, S., & Sunyaev, R.  2007, , 463, 957 Littlefair, S. P., Dhillon, V. S., & Marsh, T. R.  2001, , 327, 669 Luna, G. J. M., Sokoloski, J., Mukai, K., & Nelson, T.  2010, Astron. Telegram, 3053 Luna, G. J. M., & Sokoloski, J. L. 2007, , 671, 741 Masetti, N. et al. 2006, , 459, 21 McArthur, B. E., Benedict, G. F., Lee, J., et al.  2001, , 560, 907 Mitsuda, K.   2007, , 59, S1 Murakami, H., Koyoma, K., Sakano, M., & Tsujimoto, M.  2000, , 534, 283 Revnivtsev, M., Sazonov, S., Gilfanov, M., Churazov, E., Sunyaev, R., 2006. , 452, 169. Revnivtsev, M. & Sazonov, S. 2007, , 471, 159 Revnivtsev, M., Sazonov, S., Churazov, E., Froman, W., Vikhlinin, A., & Sunyaev, R.  2009, , 458, 1142 Revnivtsev, M., van den Berg, M., Burenin, R., Grindlay, J. E., Karasev, D., & Forman, W.  2010, , 515, A49 Ritter, H., & Kolb, U.   2003, , 404, 301 Sazonov, S, Revnivtsev, M., Gilfanov, M., Churazov, E. & Sunyaev, R., 2006, , 450, 117 Smith, R. K., Mushotzky, R., Mukai, K., Kallman, T., Markwardt, C. B., & Tueller, J.  2008, , 60, 43 Sokoloski, J. L., & Kenyon, S. J. 2003, , 584, 1021 Strong, A. W., Diehl, R., Halloin, H., Schnfelder, V., Bouchet, L., Mandrou, P., Lebrun, F., & Terrier, R.  2005, , 444, 495 Suleimanov, V., Revnivtsev, M., & Ritter, H.  2005, , 435, 191 Valinia, A., Tatischeff, V., Arnaud, K., Ebisawa, K., & Ramaty, R.  2000, , 543, 733 Warner, B. 1995, Cataclysmic variable stars (Cambridge: Cambridge University Press) Warwick, R. S., Turner, M. J. L., Watson, M. G., & Willingale, R.  1985, , 317, 218 Warwick, R. S. 2014, , 445, 66 Worrall, D. M., Marshall, F. E., Boldt, E. A., & Swank, J. H.  1982, , 255, 111 Yamauchi, S., & Koyama, K.  1993, , 404, 620 Yamauchi, S.,   2009, , 61, S225 Yuasa, T., Nakazawa, K., Makishima, K., Saitou, K., Ishida, M., Ebisawa, K., Mori, H., & Yamada, S.  2010, , 520, A25 Yuasa, T.  2010, Ph.D Thesis, University of Tokyo, Japan Yusef-Zadeh, F., Wardle, M., & Roy, S.  2007, , 665, L123 Wynn, G. A., King, A. R., & Horne, K.  1997, , 286, 436 (80mm,80mm)[f4-1.eps]{} (80mm,80mm)[f4-2.eps]{}  \[f1\] (80mm,80mm,angle=270)[gal.eps]{} [lccc]{} Source Name & ObsID & Obs. start (UT) & Exp.\ & & Date / Time & (ks)\ \ CH Cyg & 400016020 & 2006-05-28 / 07:28 & 35.2\ T CrB & 401043010 & 2006-09-06 / 22:44 & 46.3\ RT Cru & 402040010 & 2007-07-02 / 12:38 & 50.9\ SS73 17 & 403043010 & 2008-11-05 / 16:30 & 24.9\ \ V1432 Aql & 403027010 & 2008-04-16 / 21:33 & 24.9\ \ NY Lup & 401037010 & 2007-02-01 / 15:17 & 86.8\ RX J2133.7$+$5107 & 401038010 & 2006-04-29 / 06:50 & 62.8\ EX Hya & 402001010 & 2007-07-18 / 21:23 & 91.0\ V1223 Sgr & 402002010 & 2007-04-13 / 11:31 & 46.2\ MU Cam & 403004010 & 2008-04-14 / 00:55 & 50.1\ V2400 Oph & 403021010 & 2009-02-27 / 11:42 & 110.0\ YY Dra & 403022010 & 2008-06-15 / 18:37 & 27.4\ TV Col & 403023010 & 2008-04-17 / 18:00 & 30.1\ V709 Cas & 403025010 & 2008-06-20 / 10:24 & 33.3\ IGR J17303$-$0601 & 403026010 & 2009-02-16 / 10:09 & 27.7\ IGR J17195$-$4100 & 403028010 & 2009-02-18 / 11:03 & 26.9\ BG CMi & 404029010 & 2009-04-11 / 12:11 & 45.0\ PQ Gem & 404030010 & 2009-04-12 / 13:46 & 43.2\ TX Col & 404031010 & 2009-05-12 / 19:19 & 51.1\ FO Aqr & 404032010 & 2009-06-05 / 08:14 & 33.4\ AO Psc & 404033010 & 2009-06-22 / 11:50 & 35.6\ IGR J00234$+$6141 & 405022010 & 2010-06-25 / 00:06 & 77.4\ XY Ari & 500015010 & 2006-02-03 / 23:02 & 93.6\ [lcccccccccccccccccc]{} Source Name & Luminosity & Distance & Ref.\ & ($10^{39}$ photons s$^{-1}$) & (pc)\ \ CH Cyg & 1.52 & 250 & \[1\]\ T CrB & 2.90 & 500 & \[2\]\ RT Cru & 30.0 & 1500 & \[3\]\ SS73 17 & 2.29 & 500 & \[4\]\ \ V1432 Aql & 0.42 & 230 & \[5\]\ \ NY Lup & 6.05 & 840 & \[5, 6\]\ RX J2133.7$+$5107 & — & — & —\ EX Hya & 0.02 & 65 & \[7\]\ V1223 Sgr & 5.27 & 527 & \[2, 7\]\ MU Cam & — & — & —\ V2400 Oph & 2.93 & 500 & \[8\]\ YY Dra & 0.06 & 155 & \[9\]\ TV Col & 1.44 & 370 & \[10\]\ V709 Cas & 0.48 & 250 & \[6\]\ IGR J17303$-$0601 & — & — & —\ IGR J17195$-$4100 & 0.09 & 110 & \[11\]\ BG CMi & 1.10 & 500 & \[9\]\ PQ Gem & 0.97 & 400 & \[9\]\ TX Col & 0.48 & 500 & \[9\]\ FO Aqr & 5.50 & 400 & \[8\]\ AO Psc & 0.67 & 250 & \[9\]\ IGR J00234$+$6141 & 0.41 & 530 & \[12\]\ XY Ari & 0.22 & 270 & \[13\]\ \ hSSs Average & 9.20 & — & —\ mCVs Average& 1.63 & — & —\ [^1]: http://asd.gsfc.nasa.gov/Koji.Mukai/iphome/catalog/alpha.html [^2]: See http://heasarc.gsfc.nasa.gov/lheasoft/ for details. [^3]: Reprinted from Publication of New Astronomy, Vol. 36(2015), Author: R.N.C., Eze, Title: On the origin of the iron fluorescent line emission from the Galactic Ridge/ Page No. 64 - 69, Copyright (2015), with permission from Elsevier

ArXiv

--- abstract: 'There are several efficient methods to solve linear interval polynomial systems in the context of interval computations, however, the general case of interval polynomial systems is not yet covered as well. In this paper we introduce a new elimination method to solve and analyse interval polynomial systems, in general case. This method is based on computational algebraic geometry concepts such as polynomial ideals and Gröbner basis computation. Specially, we use the comprehensive Göbner system concept to keep the dependencies between interval coefficients. At the end of paper, we will state some applications of our method to evaluate its performance.' author: - | Sajjad Rahmany, Abdolali Basiri and Benyamin M.-Alizadeh\ School of Mathematics and Computer Sciences,\ Damghan University, Damghan, Iran. title: | An Elimination Method to Solve\ Interval Polynomial Systems --- Introduction ============ Many computational problems arising from applied sciences deal with floating-point computation and so require to import polynomial equations containing error terms in computers. This redounds polynomial equations to appear with perturbed coefficients i.e the coefficients range in specific intervals and so these are called [*interval polynomial equations*]{}. Interval polynomial equations come naturally from several problems in engineering sciences such as control theory \[35,36\], dynamical systems \[30\] and so on. One of the most important problems in the context of interval polynomial equations is to analyse and study the stability and solutions of an (or a system of) interval polynomial(s). More generally, the problem is to carry as much as possible information out from an interval polynomial system. Many scientific works are done in this direction using interval arithmetic \[2\], for instance computation of the roots in certain cases \[6,7\], however they do not enable us to obtain the desired roots, at least approximately \[6\]. Another example consists those works which contain (the most popular) method to solve an interval polynomial equation by computing the roots of some exact algebraic polynomials, while it is hard to solve an algebraic equation of high degree which has its own complexity challenging problems \[9,11,12,18,21\]. There is also a new method described in \[41\] which counts the zeros of a univariate interval polynomial. In addition to numerical methods, there are some attempts to combine numeric and symbolic methods to solve an interval polynomial system. In \[8\], Falai et al. state a modification of Wu’s characteristic set method for interval polynomial systems, and use numerical approximation to find an interval containing the roots. The essential trick in this work is to omit all the terms with interval coefficients containing zero, which permits the division of interval coefficients simply. This consideration may fail some important polynomials. In this paper, we try to use exact symbolic methods to facilitate analysing interval polynomial systems thanks to algebraic elimination methods like parametric computation techniques to analyse a parametric polynomial system which allows us to consider all exact polynomials arising from an interval polynomial. As we will state later, it is very important to keep the trace of interval coefficients during the computations. Roughly speaking we associate an auxiliary parameter to each interval coefficient provided that each parameter ranges over its own related interval only. Nowadays there are important results \[38,39\], efficient algorithms \[17,19,20,22,24,25\] and powerful implementations \[33,34\] in the context of parametric computations and analysing parametric polynomial systems. We introduce the new concept [*interval Gröbner system*]{} for a system of interval polynomials using the concept of [*comprehensive Gröbner systems*]{} \[39\] which is used to describe all different behaviours of a parametric polynomial system. Interval Gröbner system contains a finite number of systems where each one is a Gröbner basis for (non-interval) polynomial systems obtained from the main system. It is worth noting that against to \[8\], we don’t omit any interval coefficient and cover all possible cases for the exact coefficients arising from the intervals. We also design an algorithm to compute interval Gröbner system of an interval polynomial system. This paper is organized as follows. In Section 2 we state introductory definitions and recall interval arithmetic. In Section 3 we start to explain interval polynomials and their related concepts. We next receive to Section 4 which states the main idea behind this paper. To recall the concepts of computational algebraic tools we state Section 5 containing a brief introduction to Gröbner basis and comprehensive Gröbner system, together with their related algorithms. After, we describe our elimination method for interval polynomial systems in Section 6. Finally, as some applications and example, we state the Section 7 which contains two applied examples. Preliminaries {#pre} ============= In this section we recall the interval arithmetic and related concepts which are needed for the rest of this text. The main references of this section are \[16\] and \[26\]. Let $\R$ denote the set of real numbers while $\R^{*}$ is used to show the [*extended real numbers set*]{} i.e. $\R \cup \{-\infty,\infty\}$. Let $a,b\in \R^*$. We define $4$ kinds of [*real intervals*]{} defined by $a$ and $b$ as follows: $$\begin{aligned} \label{interval} \begin{array}{rcl} {\rm{Closed\ interval}} &:& [a,b] = \{x \mid a\leq x \leq b\} \ (a,b \ne \pm\infty) \\ {\rm{Left\ half\ open\ interval}} &:& (a,b] = \{x \mid a< x \leq b\}\ (b\ne \pm\infty)\\ {\rm{Right\ half\ open\ interval}} &:& [a,b) = \{x \mid a \leq x < b\}\ (a\ne \pm\infty)\\ {\rm{Open\ interval}} &:& (a,b) = \{x \mid a< x < b\} \\ \end{array}\end{aligned}$$ It is worth noting that approximately all of existing texts in the subject of interval computation deal with closed intervals. Most of times we denote the intervals by capitals, and their lower (resp. upper) bounds by underbar (resp. overbar), as $$X= [\underbar{X},\overline{X}]$$ to denote closed intervals. However, as there are some applied problems including non-closed intervals we preferred to consider all different types of intervals. Having all different kinds of intervals at once, we use the notion $[a,b,i,j]$ where $i,j \in \{0,1\}$ to denote the intervals in (1), as follows: $$\begin{aligned} [a,b,i,j] = \left\{ \begin{array}{lcl} [a,b] & {\rm if} & i=j=1\\ (a,b] & {\rm if} & i=0, j=1\\ [a,b) & {\rm if} & i=1, j=0\\ (a,b)& {\rm if} & i=j=0\\ \end{array} \right.\end{aligned}$$ However when all of intervals come from one sort of presentation, we prefer to use the (1) form. Now we recall the interval arithmetic and discuss on the interval dependencies what will occur in the evaluation of interval expressions. There exist two equivalent ways to state interval arithmetic. The first is based on the endpoints of intervals while the second considers each interval as a subset of real numbers. Let $[a_1, b_1, i_1, j_1]$ and $[a_2, b_2, i_2, j_2]$ be two real intervals. Note that each real number $a$ is considered as $[a,a,1,1]$ which is called a degenerate interval. Four essential arithmetic operations are defined as follows: $$\begin{aligned} [a_1, b_1, i_1, j_1] + [a_2, b_2, i_2, j_2] & = & [a_1 + a_2, b_1 + b_2, \min(i_1, i_2), \min(j_1, j_2)]\\ ~ [a_1, b_1, i_1, j_1] - [a_2, b_2, i_2, j_2] & = & [a_1 -b_2, b_1 -a_2, \min(i_1, j_2), \min(j_1, i_2)]\\ ~ [a_1, b_1, i_1, j_1] \times [a_2, b_2, i_2, j_2] & = & [a_kb_\ell, a_{k'}b_{\ell'}, \min(i_k,j_\ell), \min(i_{k'},j_{\ell'})]\\ ~\end{aligned}$$ where $a_kb_\ell$ and $a_{k'}b_{\ell'}$ are the minimum and maximum of the set $\{a_1a_2, a_1b_2, b_1a_2, b_1b_2\}$ respectively, and finally $$\begin{aligned} [a_1, b_1, i_1, j_1] / [a_2, b_2, i_2, j_2] & = & [a_1, b_1, i_1, j_1] \times [1/b_2, 1/a_2, j_2, i_2]\end{aligned}$$ provided that $a_2> 0$ or $b_2< 0$ or $a_2 = i_2 = 0$ or $b_2 = j_2 = 0$. Note in the above relations that all ambiguous cases $\infty-\infty$, $\pm\infty\times 0$, $\frac{\pm\infty}{\pm\infty}$ and $\frac{0}{0}$ will induce the biggest possible interval i.e. $\R$. As an easy observation, when an interval $X=[a,b,i,j]$ with $a\ne b$ contains zero, we can compute $1/X$ as follows: - If $a=0$ then $$\frac{1}{X} = \frac{1}{[a,b,0,j]}=[\frac{1}{b},+\infty,j,0],$$ - If $b=0$ then $$\frac{1}{X} = \frac{1}{[a,b,i,0]}=[-\infty, \frac{1}{a},0,i],$$ - If $ab<0$ then by seperating $X$ as $X=[a,0,i,0] \cup [0,b,0,j]$ we have $$\frac{1}{X} = \frac{1}{[a,0,i,0] \cup [0,b,0,j]} = [-\infty, \frac{1}{a},0,i] \cup [\frac{1}{b}, +\infty, j , 0].$$ Now consider $A$ and $B$ are two intervals as two sets of real numbers. We can state the above definitions of four essential arithmetic operations as $$A\ op \ B = \{x\ | \ \exists\ a\in A, \ b\in B: \ x=a \ op \ b\}$$ where $op \in \{+,-,\times, /\}$. Although interval arithmetic seems to be compatible with real numbers arithmetic, but this affects distributivity of multiplication over addition and the existance of inverse elements. More preciesly for each intervals $X,Y$ and $Z$, - $X\times (Y+Z)\subseteq X\times Y+X\times Z$, and if $X$ is non-degenerated then - $X\times \frac{1}{X} \ne 1$, but $1\in X\times \frac{1}{X}$ and - $X+(-X) \ne 0$, but $0\in X+(-X)$. Furthermore, if $X$ contains some negative real numbers, then $$X^n\ne \underbrace{X\cdots X}_{n \ times}.$$ To see this, let for instance $X=[a,b,i,j]$ where $a<0$ and $|a|<b$. Then, $X^2 = [0,b^2,1,j]$ while $X\times X = [ab,b^2, \min(i,j),j]$. To solve this inconsistency, we define the $n^,$th power of an interval for non-negative integer $n$ as follows: $$\begin{aligned} [a,b,i,j]^n=\left\{ \begin{array}{lcl} 1 & & n=0\\ [a^n, b^n, i, j] & & 0\leq a\\ [b^n, a^n, j, i] & & b\leq 0 \\ [0, \max(a^n,b^n), 1, c] & & a< 0 < b \\ \end{array} \right.\end{aligned}$$ where $ c=\left\{ \begin{array}{crl} i& & |b|<|a|\\ j& & otherwise \end{array} \right.. $ Let us now evaluate some expressions to illustrate more challenging problems dealing with interval arithmetic. Let $f(x,y)=\frac{x}{x+y}$, $X=[1,2]$ and $Y=[1,3]$. We compute $f(X,Y)$ in two ways. The first way is to compute $f(X,Y)$ as an usual evaluation using interval arithmetic: $$\begin{aligned} \label{eq:2} \frac{X}{X+Y} = [1,2]/[2,5] = [1/5, 1].\end{aligned}$$ However, one can manipulate the expression to see $$\begin{aligned} \label{eq:3} \frac{X}{X+Y} = \frac{1}{1+\frac{Y}{X}} = 1 /[3/2, 4] = [1/4 , 2/3]\end{aligned}$$ Let us separate $x$ and $y$ first depending on $f(x,y)$: we call $y$ (resp. $x$), a [*first*]{} (resp. [*second*]{}) [*class*]{} variable of $f$ as it appears one (resp. more than one) time in the structure of $f$. As it can be easily seen, the answer of (3) is a narrower interval and in fact the exact value. The reason is that $X$ is a second class variable for (2) and so it brings [*dependency*]{} between two parts of the expression. This is while there is no dependency in (3) given that $\frac{Y}{X}$ appears only one time, and so it is a first class variable. Dependency is one of the crucial points of this paper to find the solution set of interval polynomial systems. Dependencies are the main reason that causes to appear an amount of error by introducing larger intervals than the exact solution as all publications in this area try to avoid dependencies. Nevertheless it is possible to cancel dependencies by considering $X-X=0$ (see \[16\]) as well as $X/X=1$ easily, while sometimes this becomes a difficult work. Interval Polynomials {#IntPoly} ==================== Let $\R$ be the field of real numbers, considered as the ground field of computations all over the current text and consider the set $\{x_1,\ldots,x_n\}$ to be the set of variables. Each polynomial of the form $$\begin{aligned} \label{Def:1} [f]=\sum_{i=1}^m [a_i,b_i,\ell_i, k_i] x_1^{\alpha_{i1}}\cdots x_n^{\alpha_{in}}\end{aligned}$$ is called an interval polynomial, where $[a_i,b_i,\ell_i,k_i]$ is a real interval for each $i=1,\ldots,m$, and each power product $x_1^{\alpha_{i1}}\cdots x_n^{\alpha_{in}}$ is called a monomial where the powers are non negative integers. We denote the set of all interval polynomials by $[\R][x_1,\ldots,x_n]$. Let $[f]$ be an interval polynomial as defined in (4). The set of all polynomials arising from $[f]$ for different values of intervals in coefficients is called the family of $[f]$ and is denoted by $\mathcal{F}([f])$. More preciesly: $$\mathcal{F}([f]) = \{\sum_{i=1}^m c_i x_1^{\alpha_{i1}}\cdots x_n^{\alpha_{in}} \ | \ c_i \in [a_i,b_i,\ell_i, k_i], i=1,\ldots,m\}$$ Similar to the family of an interval polynomial, we can define the family of a set of interval polynomials as follows: \[Def:2\] Let $S:=\{[f]_1,\ldots,[f]_\ell\}$ be a set of interval polynomials with $\mathcal{F}([f]_j) = \mathcal{F}_j$ for each $j=1,\ldots, \ell$. We define the family of $S$ to be the set $\mathcal{F}_1\times\cdots \times \mathcal{F}_\ell$, denoted by $\mathcal{F}(S)$. We now define the concept of solution set for an interval polynomial. \[Def:3\] For an interval polynomial $[f] \in {[\R][x_1,\ldots,x_n]}$, we say that $a=(a_1,\ldots,a_n) \in \R^n$ is a real solution or a real root of $[f]$, if there exists a polynomial $p \in \mathcal{F}([f])$ such that $p(a) = 0$. Similarily, we say that a system of interval polynomials has a solution, if the contained interval polynomials have a common solution. Let us find the solution set of $$[-2,-1] x^2 + [1,2] x + [1,3]=0$$ where all intervals are closed. When $x\geq 0$, we have $$[-2,-1] x^2 + [1,2] x + [1,3] = [-2x^2+x+1, -x^2+2x+3] =[0,0]$$ So we must have $$0\leq -2x^2+x+1\ {\rm{and}}\ -x^2+2x+3\leq 0$$ which implies $$x\in[1,3].$$ Similarly when $x\leq 0$, we have $$[-2,-1] x^2 + [1,2] x + [1,3] = [-2x^2+2x+1, -x^2+x+3] =[0,0]$$ or equivalently $$0\leq -2x^2+2x+1\ {\rm{and}}\ -x^2+x+3\leq 0$$ which concludes $$x\in [-1.31, -0.36].$$ Thus the solution set of this interval polynomial is $$[-1.31, -0.36] \cup [1,3].$$ It is notable that using interval arithmetic in the well-known solution way of a quadratic polynomial equation due to discriminant, we receive to $$[-2.15, -0.06]\cup [0.382 , 8.4]$$ as the solution set that contains an amount of error. The idea {#Idea} ======== In this section we are going to describe the problems which may occur using the usual elimination method on a system of interval polynomials. To facilitate the description, let us give an example. Consider the system containing $$f_1=[1,2]x_1+x_2+2x_3,\ f_2=[1,4]x_1+x_2+1,\ f_3=[3,4]x_1+x_2+4x_3.$$ Going to eliminate the variable $x_2$, we have $$f_1-f_2=[-3,1]x_1+2x_3-1,\ f_3-f_2=[-1,3]x_1+4x_3-1.$$ One may now conclude that if we choose $0$ from both intervals $[-3,1]$ and $[-1,3]$ then the system has no solution because $2x_3-1=4x_3-1=0.$ However this case is impossible since both of intervals $[-3,1]$ and $[-1,3]$ can not be zero at once! To see this, notice that $[-3,1]$ comes from $[1,2]-[1,4]$ and so for $[-3,1]$ being zero, $[1,4]$ must give some values in $[1,2]$. On the other hand, $[-1,3]$ comes from $[3,4]-[1,4]$ and so this interval can be zero only when $[1,4]$ gives some values in $[3,4]$ and this is a contradiction. This simple linear system shows that the usual elimination method can cause a wrong decision or appearing some extra values in the solution set. The main reason is that we forgot the dependencies between $[-3,1]$ and $[-1,3]$ during the computation while they are both dependent on $[1,4]$ and so they are dependent. To solve this problem, we must keep the trace of each interval coefficient. In doing so, our idea is to use a parameter instead of each interval, to see how new coefficients are built. For instance let us substitute $[1,2]$, $[1,4]$ and $[3,4]$ by $a$, $b$ and $c$ as parameters in the above example. So we have $$\tilde{f_1}=ax_1+x_2+2x_3,\ \tilde{f_2}=bx_1+x_2+1,\ \tilde{f_3}=cx_1+x_2+4x_3$$ and doing elimination steps we have $$\tilde{f_1}-\tilde{f_2}=(a-b)x_1+2x_3-1,\ \tilde{f_3}-\tilde{f_2}=(c-b)x_1+4x_3-1.$$ Now one can conclude that under the assumption that $a-b=0$, the coefficient $c-b$ can not be zero. The reason is that if $c-b=0$ then $a=c$ while $1\leq a \leq 2$ and $3\leq c \leq 4$. Therefore using parameters prevent us taking wrong decisions about the solution set. The main question here is that how we can use elimination method when the coefficients contain some parameters. In fact as we will state in the next sections, we do the elimination steps thanks to Gröbner basis and for the parametric case, we use the concept of comprehensive Gröbner system, to see the simplest possible polynomials to solve. So our idea is to convert the interval polynomial systems to a parametric polynomial system and use the parametric algorithmic aspects, of course with some modifications, to solve the parametric system by dividing the solution set into finitely many components. At the end, we convert the result to see the solution set of the interval polynomial system. Gröbner Bases and Comprehensive Gröbner Systems {#GB} =============================================== In this section we recall the concepts and notations of ordinary and parametric polynomial rings. Let $K$ be a field and $ x_1,\ldots,x_n$ be $n$ (algebraically independent) variables. Each power product $x_1^{\alpha_1}\cdots x_n^{\alpha_n}$ is called a monomial where $\alpha_1,\ldots,\alpha_n\in \mathbf{Z}_{\geq 0}$. Because of simplicity, we abbreviate such monomials by ${\bf x}^\alpha$ where ${\bf x}$ is used for the sequence $x_1,\ldots,x_n$ and $\alpha=(\alpha_1,\ldots,\alpha_n)$. We can sort the set of all monomials over $K$ by special types of total orderings so called monomial orderings, recalled in the following definition. The total ordering $\prec$ on the set of monomials is called a monomial ordering whenever for each monomials ${\bf x}^\alpha, {\bf x}^\beta$ and ${\bf x}^\gamma$ we have: - ${\bf x}^\alpha\prec {\bf x}^\beta \Rightarrow {\bf x}^\gamma {\bf x}^\alpha\prec {\bf x}^\gamma {\bf x}^\beta$, and - $\prec$ is well-ordering. There are infinitely many monomial orderings, each one is convenient for a special type of problems. Among them, we point to pure and graded reverse lexicographic orderings denoted by $\prec_{lex}$ and $\prec_{grevlex}$ as follows. assume that $x_n\prec \cdots \prec x_1$. We say that - ${\bf x}^\alpha \prec_{lex} {\bf x}^\beta$ whenever $$\alpha_1=\beta_1,\ldots,\alpha_i=\beta_i \ {{\rm and}} \ \alpha_{i+1}<\beta_{i+1}$$ for an integer $1\leq i < n$. - ${\bf x}^\alpha \prec_{grevlex} {\bf x}^\beta$ if $$\sum_{i=1}^n \alpha_i < \sum_{i=1}^n \beta_i$$ breaking ties when there exists an integer $1\leq i < n$ such that $$\alpha_n=\beta_n,\ldots,\alpha_{n-i}=\beta_{n-i} \ { {\rm and}} \ \alpha_{n-i-1}>\beta_{n-i-1}.$$ It is worth noting that the former has many theoretical importance while the latter speeds up the computations and carries fewer information out. Each $K-$linear combination of monomials is called a polynomial on $x_1,\ldots,x_n$ over $K$. The set of all polynomials has the ring structure with usual polynomial addition and multiplication, and is called the polynomial ring on $x_1,\ldots,x_n$ over $K$ and denoted by $K[x_1,\ldots,x_n]$ or just by $K[{\bf x}]$. Let $f$ be a polynomial and $\prec$ be a monomial ordering. The greatest monomial w.r.t. $\prec$ contained in $f$ is called the leading monomial of $f$, denoted by $\LM(f)$ and the coefficient of $\LM(f)$ is called the leading coefficient of $f$ which is pointed by $\LC(f)$. Further, if $F$ is a set of polynomials, $\LM(F)$ is defined to be $\{\LM(f) | f\in F\}$ and if $I$ is an ideal, $in(I)$ is the ideal generated by $\LM(I)$ and is called the initial ideal of $I$. We are now going to remind the concept of Gröbner basis of a polynomial ideal which carries lots of useful information out about the ideal. Let $I$ be a polynomial ideal of $K[{\bf x}]$ and $\prec$ be a monomial ordering. The finite set $G\subset I$ is called a Gröbner basis of $I$ if for each non zero polynomial $f\in I$, $\LM(f)$ is divisible by $\LM(g)$ for some $g\in G$. Using the well-known Hilbert basis theorem (See \[4\] for example), it is proved that each polynomial ideal possesses a Gröbner basis with respect to each monomial ordering. There are efficient algorithms also to compute Gröbner basis. The first and the most simplest one is the Buchberger algorithm which is devoted in the same time of introduction of Gröbner basis concept while he most efficient known algorithm is the Faugère’s F$_5$ algorithm \[10\] and another signature-based algorithms such as G$^2$V \[13\] and GVW \[14\]. It is worth noting that Gröbner basis of an ideal is not unique necessarily. To have unicity, we define the reduced Gröbner basis concept. As an important fact the reduced Gröbner basis of an ideal is unique up to the monomial ordering. Let $G$ be a Gröbner basis for the ideal $I$ w.r.t. $\prec$. Then $G$ is called a reduced Gröbner basis of $I$ whenever each $g\in G$ is monic, i.e. $\LC(g)=1$ and none of the monomials appearing in $g$ is divisible by $\LM(h)$ for each $h\in G\setminus \{g\}$. One of the most important applications of Gröbner basis is its help to solve a polynomial system. Let $$\left\{ \begin{array}{ccc} f_1&=&0\\ &\vdots&\\ f_k&=&0\\ \end{array} \right.$$ be a polynomial system and $I=\langle f_1,\ldots,f_k\rangle$ be the ideal generated by $f_1,\ldots,f_k$. We define the affine variety associated to the above system or equivalently to the ideal $I$ to be $${\bf V}(I) = {\bf V}(f_1,\ldots,f_k) = \{\alpha \in \overline{K}^n | f_1(\alpha)=\cdots = f_k(\alpha) = 0\}$$ where $\overline{K}$ is used to denote the algebraic closure of $K$. Now let $G$ be a Gröbner basis for $I$ with respect to an arbitrary monomial ordering. As an interesting fact, $I=\langle G\rangle$ which implies that ${\bf V}(I) = {\bf V}(G)$. This is the key computational trick to solve a polynomial system. Let us continue by an example. \[Ex:1\] We are going to solve the following polynomial system: $$\left\{ \begin{array}{ccc} x^2-xyz+1&=&0\\ y^3+z^2-1&=&\\ xy^2+z^2&=&0\\ \end{array} \right.$$ By the nice properties of pure lexicographical ordering, the reduced Gröbner basis of the ideal $I=\langle x^2-xyz+1, y^3+z^2-1, xy^2+z^2 \rangle \subset \mathbf{Q}[x,y,z]$ has the form $$G=\{g_1(z), x-g_2(z), y-g_3(z)\}$$ w.r.t. $z\prec_{lex}y\prec_{lex}x$, where [ $$\left[ \begin{array}{ccl} g_1(z)&=& z^{15}-3z^{14}+5z^{12}-3z^{10}-z^9-z^8+4z^6-6z^4+4z^2-1\\ g_2(z)&=& 2z^{14}-9z^{13}+11z^{12}+2z^{11}-7z^{10}-3z^9+2z^8-z^7+4z^6+\\ &&+7z^5-10z^4-6z^3+11z^2+2z-4\\ g_3(z)&=&z^{13}-3z^{12}+z^{11}+2z^{10}+z^9-z^8-2z^6+2z^4-z^3-3z^2+1.\\ \end{array} \right.$$ ]{} This special form of Gröbner basis for this system allows us to find ${\bf V}(G)$ by solving only one univariate polynomial $g_1(z)$ and putting the roots into the two last polynomials in $G$. Suppose now that the same system of Example 5.4 is given as follows with parametric coefficients on parameters are $a,b$ and $c$: $$\left\{ \begin{array}{ccc} {a}x^2-({a}^2-{b}+1)xyz+1&=&0\\ y^3+{c}^2z^2-1&=&\\ ({a}+{b}+{c})xy^2+z^2&=&0\\ \end{array} \right.$$ The solutions of this system depend on the values of parameters apparently as we can see that the system has no solutions whenever ${a}=0$ and ${b}=1$ while it converts to the system of Example 5.4 for ${a}=1, {b}=1$ and ${c}=-1$ and so has some solutions. To manage all of different behaviours of parameters which cause to different behaviour of the main system, we recall the concept of comprehensive Gröbner system in the sequel. By this we can divide the space of parameters, i.e. $\overline{K}^t$ into a finite number of partitions, for which the general form of polynomials contained in assigned Gröbner basis is known. Let $K$ be a field and ${\bf a} := a_1,\ldots,a_t$ and ${\bf x}:=x_1,\ldots,x_n$ be the sequences of parameters and variables respectively. We call $K[{\bf a}][{\bf x}]$, the parametric polynomial ring over $K$, with parameters ${\bf a}$ and variables ${\bf x}$. This ring is in fact the set of all parametric polynomials as $$\sum_{i=1}^{m} p_{i} {\bf x}^{\alpha_i}$$ where $p_{i}\in K[{\bf a}]$ is a polynomial on ${\bf a}$ with coefficients in $K$, for each $i$. \[Def:CGS\] Let $I\subset K[{\bf a}][{\bf x}]$ be a parametric ideal and $\prec$ be a monomial ordering on ${\bf x}$. Then the set $$\mathcal{G}(I)=\{(E_i,N_i,G_i) \mid i=1,\ldots,\ell\} \subset K[{\bf a}]\times K[{\bf a}]\times K[{\bf a}][{\bf x}]$$ is said a comprehensive Gröbner system for $I$ if for each $(\lambda_1,\ldots,\lambda_t)\in \overline{K}^t$ and each specialization $$\begin{aligned} \sigma_{(\lambda_1,\ldots,\lambda_t)} :& K[{\bf a}][{\bf x}] &\rightarrow \ \ \ \ \ \ \ \overline{K}[{\bf x}]\\ &\sum_{i=1}^{m} p_{i} {\bf x}^{\alpha_i} &\mapsto \sum_{i=1}^{m} p_{i}(\lambda_1,\ldots,\lambda_t) {\bf x}^{\alpha_i}\end{aligned}$$ there exists an $1\leq i \leq \ell$ such that $(\lambda_1,\ldots,\lambda_t) \in {\bf V}(E_i) \setminus {\bf V}(N_i)$ and $\sigma_{(\lambda_1,\ldots,\lambda_t)}(G_i)$ is a Gröbner basis for $\sigma_{(\lambda_1,\ldots,\lambda_t)}(I)$ with respect to $\prec$. Because of simplicity, we call $E_i$ and $N_i$ the null and non-null conditions respectively. Remark that, by \[39, Theorem 2.7\], every parametric ideal has a comprehensive Gröbner system. Now we give an example from \[20\] to illustrate the definition of comprehensive Gröbner system. Consider the following parametric polynomial system in $\mathbf{Q}[a,b,c][x,y]$: $$\begin{aligned} \Sigma:\left\{ \begin{array}{lll} ax-b&=&0\\ by-a&=&0\\ cx^2-y&=&0\\ cy^2-x&=&0 \end{array} \right.\end{aligned}$$ Choosing the graded reverse lexicographical ordering $y\prec x$, we have the following comprehensive Gröbner system: For instance, for the specialization $\sigma_{(1,1,1)}$ for which $a\mapsto 1, b\mapsto 1$ and $c\mapsto 1$, $$\sigma_{(1,1,1)}(\{bx-acy, by-a\})=\{x-y,y-1\}$$ is a Gröbner basis of $\sigma_{(1,1,1)}(\langle \Sigma \rangle)$. It is worth noting that if ${\bf V}(E_i) \setminus {\bf V}(N_i) = \emptyset$ for some $i$, then the triple $(E_i,N_i,G_i)$ is useless, and so it must be omitted from the comprehensive Gröbner system. In this case we say that the pair $(E_i,N_i)$ is [*inconsistent*]{}. It is easy to see that inconsistency occurs if and only if $N_i \subset \sqrt{\langle E_i \rangle}$ and so we need to an efficient radical membership test to determine inconsistencies. In \[19, 20\] there is a new and efficient algorithm to compute comprehensive Gröbner system of a parametric polynomial ideal which uses a new and powerful radical membership criterion. Therefore we prefer to employ this algorithm so called [PGB]{} algorithm in our computations. Another essential trick which is used in \[20\] is the usage of [*minimal Dickson basis*]{} which reduces the content of computations in [PGB]{}. Before explain it, let us recall some notations which are used in the structure of [PGB]{}. Let $\prec_{\bf x}$ and $\prec_{\bf a}$ be two monomial orderings on $K[\bf x]$ and $K[\bf a]$ respectively. Let also $\prec_{{\bf x,a}}$ be the block ordering of $\prec_{\bf x}$ and $\prec_{\bf a}$, comparing two parametric monomials by $\prec_{\bf x}$, breaking tie by $\prec_{\bf a}$. For a parametric polynomial $f \in K[\bf a][\bf x]$, we denote by $\LM_{\bf x}(f)$ (resp. by $\LC_{\bf x}(f)$) the leading monomial (resp. the leading coefficient) of $f$ when it is considered as a polynomial in $K({\bf a})[{\bf x}]$, and so $\LC_{\bf x}(f) \in K[\bf a]$. \[Def:MDB\] By the above notations, let $P\subset K[{\bf a}][{\bf x}]$ be a set of parametric polynomials and $G\subset P$. Then, $G$ is called a minimal Dickson basis of $P$ denoted by ${\sc MDBasis}(P)$, if: - For each $p\in P$, there exist some $g\in G$ such that $\LM_{\bf x}(g) \mid \LM_{\bf x}(p)$ and - For each two distinct polynomials in $G$ as $g_1$ and $g_2$, none of $\LM_{\bf x}(g_1)$ and $\LM_{\bf x}(g_2)$ divides another. The case which occurs in [PGB]{} to compute a minimal Dickson basis for $P$ is only when $P$ is a Gröbner basis for $\langle P \rangle$ itself w.r.t. $\prec_{{\bf x}, {\bf a}}$ and $P \cap K[{\bf a}]=\{0\}$. In this situation, it suffices by Definition 5.7 to omitt all polynomials $p$ from $P$ for which there exists a $p'\in P$ such that $\LM_{\bf x}(p') \mid \LM_{\bf x}(p)$. The [PGB]{} algorithm as is shown below, uses [PGB-main]{} algorithm to introduce new branches in computations. ; The main work of [PGB-main]{} is to create all necessary branches and import them in comprehensive Gröbner system at output. In this algorithm $A*B$ is defined to be the set $\{ab \mid a\in A, b\in B \}$. $(E,N,\{1\})$; ; ; $(PGB)$; ; As it is shown in the algorithm, it computes first a Gröbner basis of the ideal $\langle P \rangle$ over $K[\bf{a,x}]$ i.e. $G$, before performing any branches based on parametric constraints, according to \[20, Lemma 32\] as follows: By the notations used in the algorithm, for each specialization $\sigma_{(\lambda_1,\ldots,\lambda_t)}$ if $$(\lambda_1,\ldots,\lambda_t) \in {\bf V}(G_r) \setminus {\bf V}(\prod_{g\in G\setminus G_r}\LC_{\bf x}(g))$$ then $\sigma_{(\lambda_1,\ldots,\lambda_t)}(G)$ is a Gröbner basis for $\sigma_{(\lambda_1,\ldots,\lambda_t)}(\langle P \rangle)$. After this, the algorithm computes a minimal Dickson basis i.e. $G_m$ and continues by taking a decision for each situation that one of the leading coeffiecients of $G_m$ is zero. By this, [PGB-main]{} constructs all necessary branches to import in comprehensive Gröbner system. All over the algorithm, when it needs to add a new branch $(E_i,N_i,G_i)$ into the system, the algorithm [IsConsistent]{} is used as follow to test the consistency of parametric conditions $(E_i,N_i)$. ; The main part of this algorithm is radical membership test. The powerful trick which is used in \[19,20\] to radical membership check is based on linear algebra methods tackling with a probabilistic check. We refer the reader to \[20, Section 5\] for more details. Elimination Method to Solve Interval Polynomial Systems {#Elim} ======================================================= In this section we introduce the new concept of [*interval Gröbner system*]{} and its related definitions and statements. Now we state the following proposition as an immediate consequence of Definition 3.4. Recall that for a polynomial system $S \subset \R[x_1,\ldots,x_n]$ the variety of $S$ is the set of all complex solutions of $S$, denoted by $\V(S)$. A system $ [S]$ of interval polynomials has a solution if and only if there exists a polynomial system $S$ in $\F( [S])$ with $\V(S)\neq \emptyset$. There is an efficient criterion due to the well-known Hilbert Nullestelensatz theorem which determines if $\V(S) \neq \emptyset$ by Gröbner basis: $\V(S) \neq \emptyset$ if and only if the Gröbner basis of $\langle S \rangle$ does not contain any constant. Note that there are infinitely many polynomial systems in $\F ( [S])$ for an interval polynomial system $ [S]$ and so it is practical impossible to check all of them by Nullestelensatz theorem. Nevertheless, we give a finite partition on the set of all polynomial systems arising from $ [S]$ using the concept of comprehensive Gröbner system. Let $ [S]=\{[f]_1,\ldots,[f]_\ell\}$ be a system of interval polynomials. We define the ideal family of $ [S]$, denoted by $\idf( [S])$ to be the set $$\idf( [S]) = \{\langle p_1,\ldots,p_\ell \rangle \ | \ (p_1,\ldots,p_\ell) \in \F( [S])\}$$ \[Thm:main\] Let $ [S]$ be a system of interval polynomials and $\prec$ be a monomial ordering on $\R[x_1,\ldots,x_n]$. Then - The set of initial ideals $\{in(I) \ | \ I \in \idf( [S])\}$ is a finite set, and - For each set of ideals of $\idf( [S])$ with the same initial ideal, there exists a set of parametric polynomials which induces the ideals by different specializations. To prove this theorem, we use the concept of comprehensive Gröbner system. Suppose that $S^* $ is obtained by replacing each interval coefficient by a parameter. Note that if an interval appears in $t\geq 1$ coefficients, then we assign $t$ distinct parameters to it. It is easy to check that each element of $\idf( [S])$ is the image of $S^*$ under a suitable specialization. On the other hand by \[39, Theorem 2.7\] $S^*$ has a finite comprehensive Gröbner system as $\mathcal{G} = \{(E_1,N_1,G_1), \ldots, (E_k,N_k,G_k)\}$, where for each specialization $\sigma$ there exists a $1\leq j \leq k$ such that $\LM(\sigma(S^*)) = \LM(G_i)$. It is worth noting that although there is a finite number of branches in $\mathcal{G}$, we can also remove those specializations with complex values, and also those with values out of the assigned interval. Thus for each $I \in \idf( [S])$ there exists an $1\leq i \leq k$ with $in(I) = \langle\LM(G_i)\rangle$ and this finishes the proof. What is explained in the proof of Theorem 6.3 yields to extend the concept of comprehensive Gröbner system for interval polynomials. \[Def:IGS\] Let $ [S]\subset [\R][x_1,\ldots,x_n]$ be a system of interval polynomials with $t$ interval coefficients, and $\prec$ be a monomial ordering on $\R[x_1,\ldots,x_n]$. Let also that $\mathcal{G} = \{(E_1,N_1,G_1), \ldots, (E_k,N_k,G_k)\}$ be a set of triples $$(E_i,N_i,G_i) \in \R[h_1,\ldots,h_t] \times \R[h_1,\ldots,h_t]\times \R[h_1,\ldots,h_t][x_1,\ldots,x_n]$$ where $\{h_1,\ldots,h_t\}$ is the set of parameters assigned to each interval coefficient. Then we call $\mathcal{G}$ an interval Gröbner system for $ [S]$ denoted by $\mathcal{G}_{\prec}( [S])$ if for each $t-$tuple $(a_1,\ldots,a_t)$ of the inner values of interval coefficients there exists an $1\leq i \leq k$ such that: - For each $p \in E_i$, $p(a_1,\ldots,a_t) = 0$, - There exist some $q \in N_i$ such that $q(a_1,\ldots,a_t) \neq0,$ and - $\sigma(G_i)$ is a Gröbner basis for $\langle \sigma( [S]) \rangle$ with respect to $\prec$, where $\sigma$ is the specialization $h_j \mapsto a_j$ for $j=1,\ldots,t$. Each interval polynomial system possesses an interval Gröbner system. Let $S^*$ be the parametric polynomial system obtained by assigning each interval coefficient to a parameter. As mentioned in the proof of Theorem 6.3, $\mathcal{G}_{\prec}( [S])$ is the same comprehensive Gröbner system of $S^*$ where each parameter is bounded to give values from its assigned ideal. On the other hand it is proved that each system of parametric polynomials has a comprehensive Gröbner system, which terminates the proof. We give now an easy example to illustrate what described above. \[IGSexample\] Consider the interval polynomial system $$\begin{aligned} \label{IGSexamplesystem} [S] = \left\{ \begin{array}{ccc} [-1, 2)xy+[0, 1)y+[3, 5)&=&0\\ ~ [-3, 1)xy^2+[1, 3)y&=&0 \end{array} \right. \end{aligned}$$ To obtain a parametric polynomial system, we assign the intervals $[-1,2), [0,1), [3,5), [-3,1)$ and $[1,3)$ by $h_1,\ldots,h_5$ respectively. Then we observe the parametric polynomial system $$S^* = \{h_1 x y+h_2 y+h_3, h_4 x y^2 + h_5y \} \subset \R[h_1,\ldots,h_5][x,y]$$ Using the lexicographic monomial ordering $y\prec x$ we can compute a comprehensive Gröbner system for $\langle S^* \rangle$ which contains about $19$ triples. However some of them are admissible only for some values of parameters [*out*]{} of their assigned interval. For instance the triple $(\{1\}, \{h_1,h_2,h_4,h_5\}, \{h_3\})$ is not acceptable in this example, since $h_5 \in [1,3]$ and so it can not be zero. By removing such triples, there remains only $8$ one shown in the following table. Therefore the following table shows $\mathcal{G}_\prec( [S])$. Computing Interval Gröbner Systems {#Sec3:IGS} ---------------------------------- In tis section we state our algorithm so called [IGS]{} to compute interval Gröbner system for an interval polynomial system. This algorithm is based on the [PGB]{} algorithm with some extra conditions for the definition of consistency. To begin let $ [S]=\{[f]_1,\ldots,[f]_\ell\}\subset [\mathbf{R}][x_1,\ldots,x_n]$ be a system of interval polynomials, where for each $1\leq j\leq \ell$, $$[f]_j=\sum_{i=1}^{m_j} [a_{ij},b_{ij}] x_1^{\alpha_{ij,1}}\cdots x_n^{\alpha_{ij,n}}$$ and $(\alpha_{ij,1},\ldots,\alpha_{ij,n})\in \mathbf{Z}^n_{\geq 0}$, for each $i$. As it is mentioned in Theorem 6.3, we assign to each interval coefficient $[a_{ij}, b_{ij}]$ a parameter $h_{ij}$ to convert $ [S]$ to a parametric polynomial system $S^*$. The following proposition describes the relations between comprehensive Gröbner systems of $S^*$ and interval Gröbner bases of $ [S]$. \[Prop:1\] Using the above notations, let $[\mathcal{G}]$ and $\mathcal{G}$ be an interval Gröbner basis for $ [S]$ and a comprehensive Gröbner basis for $S^*$ respectively w.r.t. the same monomial ordering. Then for each $(E,N,G) \in [\mathcal{G}]$, there exists $(E',N',G') \in \mathcal{G}$ such that ${\bf V}(E)\setminus {\bf V}(N) \subset {\bf V}(E')\setminus {\bf V}(N')$ and $G,G'$ have the same initial ideal. This comes from Definitions 6.4 and 5.5. According to the above proposition, to compute an interval Gröbner basis for $ [S]$, it is enough to compute a comprehensive Gröbner basis for $S^*$, and use a criterion to omit those triples $(E,N,G)$ lying in $\mathcal{G} \setminus [\mathcal{G}]$, we call them [*redundant*]{} triples. \[Rem:1\] Note that for a triple $(E,N,G)$ in $\mathcal{G} \setminus [\mathcal{G}]$, the intersection of ${\bf V}(E)\setminus {\bf V}(N) $ with the cartesian product of interval coefficients is empty. We are now going to present a criterion to determine the elements of $\mathcal{G} \setminus [\mathcal{G}]$. This criterion of course is based on the answer of this question:   [*How can we sure that a system of polynomials $E\subset \mathbf{R}[a_1,\ldots,a_t]$ has a real root in the interval $[\alpha_1,\beta_1)\times \cdots\times [\alpha_t,\beta_t)$?*]{}   In the case for which $\langle E \rangle$ is zero dimensional, this question is answered totally thanks to some efficient computational tools like Sturm’s chain by isolating the real roots. However in the case of positive dimensional, there exist only some algorithm to isolate the real roots. Among them there exists an algorithm which determines whether a multivariate polynomial system has real root or not. Because of the reasons declared above, we convert the above key question to the problem of determining whether a polynomial system has a real root or not. \[Thm:Crit\] Let $E\subset \mathbf{R}[a_1,\ldots,a_t]$ be a finite polynomial set. Let also $$F=E \cup \{a_i+(a_i-\beta_i)b_i^2-\alpha_i\ | \ i=1,\ldots,t\} \subset \mathbf{R}[a_1,\ldots,a_t,b_1,\ldots,b_t]$$ where $b_j$’s are algebraic independent by $a_i$’s and $[\alpha_i,\beta_i)$ be a real interval for each $i=1,\ldots,t$. Then the system $E=0$ has a solution in $[\alpha_1,\beta_1)\times \cdots\times [\alpha_t,\beta_t)$ if and only if the system $F=0$ has a real solution. Let $E=0$ has a solution $(\gamma_1,\ldots,\gamma_t)\in [\alpha_1,\beta_1)\times \cdots\times [\alpha_t,\beta_t)$. Let also $$\eta_i=\sqrt{\frac{\alpha_i-\gamma_i}{\gamma_i-\beta_i}}$$ for each $i=1,\ldots,t$. It is easy to see that $$\gamma_i+(\gamma_i-\beta_i)\eta_i^2-\alpha_i = 0$$ which implies that $(\gamma_1,\ldots,\gamma_t,\eta_1,\ldots,\eta_t)$ is a solution of $F=0$. Conversely, suppose that there exists $(\gamma_1,\ldots,\gamma_t,\eta_1,\ldots,\eta_t)\in \mathbf{R}^{2t}$ which is a solution of $F=0$, i. e. $f(\gamma_1,\ldots,\gamma_t)=0$ for each $f\in F$ and $\gamma_i+(\gamma_i-\beta_i)\eta_i^2-\alpha_i = 0$, for each $i=1\ldots,t$. It is enough to show that $\gamma_i \in [\alpha_i,\beta_i)$. In doing so, we see that $$\gamma_i = \frac{\alpha_i+\beta_i \eta_i^2}{1+\eta_i^2} = (\beta_i-\alpha_i)\frac{\eta_i^2}{1+\eta_i^2}+\alpha_i.$$ Indeed, $0\leq\frac{\eta_i^2}{1+\eta_i^2}<1$ and this shows that $\alpha_i\leq \underbrace{(\beta_i-\alpha_i)\frac{\eta_i^2}{1+\eta_i^2}+\alpha_i}_{\gamma_{_i}} <\beta_i$ which finishes the proof. \[aux\] Note that for the intervals $[\alpha, \infty)$ and $(-\infty, \beta]$ we can use the auxiliary polynomials $a-\alpha-b^2$ and $a-\beta+b^2$ respectively. Using Theorem 6.9 and the Remarks 6.8, 6.10 we can determine the elements of $\mathcal{G} \setminus [\mathcal{G}]$ exactly (see the notations of Proposition 6.7). Let $(E,N,G) \in \mathcal{G}$ and $[\alpha_1,\beta_1),\ldots,[\alpha_t,\beta_t)$ be $t$ real intervals. Then $(E,N,G)$ is redundant if and only if the system $F=0$ has no real roots, where $$F=E \cup \{a_i+(a_i-\beta_i)b_i^2-\alpha_i\ | \ i=1,\ldots,t\} \cup \{\prod_{g\in N}(c_g g-1)\}$$ $ \subset \mathbf{R}[a_1,\ldots,a_t,b_1,\ldots,b_t,c_g: g\in N].$ The proof comes from Theorem 6.9 and this fact that if $\prod_{g\in N}(c_g g-1)=0$ then there exists a $g\in N$ for which $c_g g-1 = 0$ which implies that $g\ne 0$. The above corollary is the criterion which determines all redundant triples, and so tackling this criterion with [PGB]{} algorithm we can design our new algorithm to compute interval Gröbner systems. We design now the [IGS]{} algorithm by its main procedure. ; The [PGB-main]{} algorithm is the same which which is used in [PGB]{} algorithm. We only change the definition of consistency as below. Let $[\alpha_1,\beta_1),\ldots,[\alpha_t,\beta_t)$ be $t$ real intervals and $E,N\subset \mathbf{R}[a_1,\ldots,a_t]$. The pair $(E,N)$ is called consistent if it is not redundant, or equivalently, $$[{\bf V}(E) \setminus {\bf V}(N)] \cap [\alpha_1,\beta_1)\times \cdots \times [\alpha_t,\beta_t)\neq \emptyset.$$ According to the above definition, we change the [IsConsistent]{} algorithm to [Interval-IsConsistent]{}, which checks the consistency for radical membership and redundancy determination. \[Alg:IsCo\] ; It is worth noting that redundant triples will be omitted before the algorithm goes to continue with them. This property causes that [IGS]{} returns less triples than [PGB]{}. Examples {#Examples} ======== Solving Fuzzy Polynomial Systems {#Sec4:App} -------------------------------- In this section we state the ability of interval Gröbner system to solve parametric polynomial systems for which the parameters range over specific intervals, which may appear, for instance, when analysing fuzzy polynomial systems. In doing so, we can use the same [IGS]{} algorithm, after converting an interval polynomial system to a parametric one. To give an example, we point at resolution of fuzzy polynomial systems which converts to solve a parametric polynomial system with parameters range over $[0,1]$ (see \[1\]). (\[1, Example 4.3\] resolved) Consider the following system of fuzzy polynomials: $$\left\{ \begin{array}{lll} (0.25, 0.75, 0.625)&=& (0, 0.5, 0,5)x^5 + (0.5, 0.5, 0.25)y\\ (1.25, 1.25, 0.75)&=&(0.5,1,0.5)x^4+(1.5, 0.5, 0.5)xy\\ \end{array} \right.$$ To solve this system, the general way is to decompose it into four parametric polynomial systems. Here we solve only the third and fourth systems. Consider the third system as follows: $$(3): \left\{ \begin{array}{lll} f_1&=& (\frac{1}{2}-\frac{1}{2}h)x^5+(\frac{1}{2}h)y+\frac{1}{2}-\frac{3}{4}h\\ f_2&=& (-\frac{1}{2}+\frac{1}{2}h)x^5+(\frac{3}{4}-\frac{1}{4}h)y-\frac{7}{8}+\frac{5}{8}h\\ f_3&=& (-\frac{1}{2}+h)x^4+(2-\frac{1}{2}h)xy-\frac{5}{4}h\\ f_4&=& (1-\frac{1}{2}h)x^4+(1+\frac{1}{2}h)xy-2+\frac{3}{4}h\\ \end{array} \right.%\ \ \ {\rm where}\ h\in [0,1]$$ Computing an interval Gröbner basis w.r.t. $y\prec_{lex} x$, we have only a triple $(\{\ \}, \{h-1\}, \{1\})$ which means that we have no solutions whenever $h\ne 1$. As our algorithm deals with the values of $h\in [0,1)$, we must consider the case $h=1$ seperately. In this case, we find a solution $(h=1, x= -1.473157368, y=0.5)$ regarding the sign of variables. The fourth system is: $$(4): \left\{ \begin{array}{lll} f_1&=& (\frac{1}{2}-\frac{1}{2}h)x^5+(\frac{3}{4}-\frac{1}{4}h)y+\frac{1}{2}-\frac{3}{4}h\\ f_2&=& (-\frac{1}{2}+\frac{1}{2}h)x^5+(\frac{1}{2}h)y-\frac{7}{8}+\frac{5}{8}h\\ f_3&=& (-\frac{1}{2}+h)x^4+(1+\frac{1}{2}h)xy-\frac{5}{4}h\\ f_4&=& (1-\frac{1}{2}h)x^4+(2-\frac{1}{2}h)xy-2+\frac{3}{4}h\\ \end{array} \right.%\ \ \ {\rm where}\ h\in [0,1]$$ where $h\in [0,1]$, $x\leq 0$ and $y\leq 0$. Computing an interval Gröbner basis, we see only one triple $(\{\ \}, \{h^2-2h+3\}, \{1\})$ for which its Gröbner basis equals to $\{1\}$. Therefore the above system has no solution for $h\in [0,1)$. Note that our method deals with $[0,1)$ here and so we must check the system for $h=1$ separately. By this, the system has no solution again by considering the sign of $x$ and $y$. and so this system has no solution. Note that another way to inform that these system have or have not any real solutions (and not the exact form of solutions), is to check their consistency by what we have stated in Algorithm 6.1, since all of variables have interval form in these systems. In doing so we must add three auxiliary polynomials $h+(h-1)\tilde{h}^2, x+\tilde{x}^2$ and $y-\tilde{y}^2$ to the System (3) and $h+(h-1)\tilde{h}^2, x+\tilde{x}^2$ and $y+\tilde{y}^2$ to the System (4), where $\tilde{h}, \tilde{x}$ and $\tilde{y}$ are some extra real variables. The result of course is that (3) is consistent however (4) is not. Real Factors ------------ One of the interesting problems in the context of interval polynomials is the [*Divisibility Problem*]{} stated as follows (See \[32\]): [*For an interval polynomial ${[f]}\in [\mathbb{R}]\{x_1,\ldots,x_n\}$ and a real polynomial $g\in \mathbb{R}[x_1.\ldots,x_n]$, determine whether there is a polynomial $p \in \mathcal{F}({[f]})$ such that $g$ is a factor of $p$.* ]{} In \[32\] there is stated an efficient method based on linear programming techniques and nice properties of polytopes. In the sequel we exalin our method to solve this problem by interval Gröbner basis. To continue, let us say that $g$ i-divides $[f]$ whenever the answer of the above problem is [*yes*]{} (Note that the letter “i” stands for interval). By this conception we can now explain the following criterion based on interval Gröbner basis. \[idiv\] Using the above notations, let $[S]=\{ [f], g\}\subset [\mathbb{R}]\{x_1,\ldots,x_n\}$ and $[\mathcal{G}]$ be a reduced interval Gröbner system for $[S]$ with respect to a monomial ordering $\prec$. Then $g$ i-divides $[f]$ if and only if there exists a triple $(E,N,G)\in [\mathcal{G}]$ where $G=\{g\}$ and ${\bf V}(E) \setminus {\bf V}(N) \ne \{0\}$. It is easy to see that $g$ i-divides $[f]$ if and only if there exists an specialization $\sigma$ for which $g | \sigma([f])$ and of course $\sigma([f])\ne 0$, by the statement of divisibility problem. This implies that $\sigma([S])$ can be expressed only by $ \{ g \}$ and therefore there exists a pair of parametric sets $(E,N)$ such that $(E,N,\{g\}) \in [\mathcal{G}]$. We are going to solve the divisibility problem for $$[f]=[-1,1]x^2+[-3,1]y^2+[1/2,2]xy\in [\mathbb{R}]\{x,y\}$$ and $g=x-\sqrt{2}y\in \mathbb{R}[x,y]$. By computing a reduced interval Gröbner basis for $[S]=\{ [f], g \}$, we find the triple $$(\{c\sqrt{2}+2a+b\}, \{1\}, \{x-\sqrt{2}y\})$$ where $a,b$ and $c$ denote inner values of the intervals $[-1,1]$, $[-3,1]$ and $[1/2,2]$. This means that if the values of $a,b$ and $c$ satisfy the equation $c\sqrt{2}+2a+b = 0$, then there exists some $p \in \mathcal{F}([f])$ such that $g | p$. For instance, by evaluating $a=1/3, b=-2$ and so $c=2 \sqrt{2}/3$ we find $p=1/3x^2-2y^2+2 \sqrt{2}/3 xy \in \mathcal{F}([f])$ which is divided by $g$ (Note that $p=1/3(x+3\sqrt{2}y) g$). Therefore $g$ i-divides $[f]$. We can use Theorem 7.2 for further aims. Let $f,g \in \mathbb{R}[x_1,\ldots,x_n]$ where $g$ does not divide $f$. Then one can use Theorem 7.2 to find a polynomial $\tilde{f}$ with the same coefficients of $f$ which contain a few perturbation and $g\mid \tilde{f}$. In doing so, one can convert $f$ to an interval polynomial $[f]$ by putting the interval $[c-\epsilon, c+\epsilon]$ instead of the coefficient $c$, for each coefficient $c$ appearing in $f$. Then using Theorem 7.2 one can increase the $\epsilon$ up enough until $g$ i-divides $[f]$ with the desired precision. Conclusion ========== In the current paper we have introduced the concept of interval Gröbner system as a novel computational tool to analyse interval polynomial systems. We have further designed a complete algorithm to compute it using the existing methods to analyse parametric polynomial systems. This concept can solve some important problems from the applied and engineering research fields such as solving fuzzy polynomial system, as devoted in this paper. Acknowledgement {#acknowledgement .unnumbered} =============== The authors would like to express their sincere thanks to Professor Deepak Kapur, who devoted his time and knowledge during the preparation of this paper. The third author would also like to give an special thanks to professor Prungchan Wongwises for her kindness on posting her Ph. D. thesis for him. [2]{} A. Abbasi Molai, A. Basiri and S. Rahmany. ., , [**220**]{}, pages 541–558, 2013. G. Alefeld and D. M. Claudio. ., , [**67**]{}, pages 3–8, 1998. ., ., ., [**152**]{}(1–2)., pages 1–15., 2003. . . Springer, 1993. . . PhD thesis, Universität Innsbruck, 1965. D. M. Claudio and P. W. Oliveira. ., , [**76**]{}(S1), pages T375–T376, 1996. I. V. Dugarova, Y. M. Smagina. ., , pages 25–29, 1991. C. Falai and Y. Wu. Applications of interval arithmetic in solving polynomial equations by Wu’s elimination method. , [**48**]{}(9), pages 1260–1273, 2005. X. Fan, J. Deng and F. Chen. ., , [**216**]{}(2), pages 563–573, 2008. J.-C. Faug[è]{}re. . , pages 75–83, 2002. J. A. Ferreira, F. Patrício and F. Oliveira. , ., [**136**]{}, pages 271–281, 2001. J. A. Ferreira, F. Patrício and F. Oliveira. ., , [**173**]{}, pages 295–302, 2005. S. Gao, Y. Guan and F. Volny IV. . , pages 13–19, 2010. S. Gao, F. Volny and M. Wang. . [http://www.icar.org.641]{}, 2010. G.-M. Greuel, G. Pfister, and H. Schönemann. 3.0. , Centre for computer algebra, University of Kaiserslautern, 2011. . 2nd Edition, 2004. . , , [**38**]{}(3), pages 715–724, 2012. D. Hertz, C. S. Adjiman and C. A. Floudas. , ., [**23**]{}, pages 1333–1339, 1999. D. Kapur, Y. Sun and D. Wang. . , pages 29–36, 2010. D. Kapur, Y. Sun and D. Wang. . , [**49**]{}, pages 27–44, 2013. A. Levkovich, N. Cohen and E. Zezeb. . ., [**13**]{}, pages 321–333, 1996. . . [*J. Symbolic Comput*]{}., [**41**]{}(11), pages 1245–1263, 2006. . . ., [**44**]{}(5), pages 463–478, 2009. . . ., [**33**]{}(1-2), pages 183–208, 2002. . . ., [**45**]{}(12), pages 1391–1425, 2010. 2009. . ., ., (1–4)., pages 345–356., 2004. ., ., ., [**52**]{}(1)., pages 179–200., 2012. H. H. Rosembrock. ., , London, 1990. . , pages 326–331, 2006. H. Sekigava. , , [**44**]{}, pages 908–922, 2009. (to download a beta version of the Singular grobcov.lib library). An efficient algorithm for computing a comprehensive Gröbner system of a parametric polynomial system., [http://www.mmrc.iss.ac.cn/dwang/]{}. Y. M. Smagina. ., , 1990. C. B. Sohn. ., , [**37**]{}(2), 1992. ., [**4029**]{}, pages 150–159., 2006. . . ., [**36**]{}(3-4), pages 669–683, 2003. . . ., [**14**]{}(1), pages 1–29, 1992. ., ., report 75/1, University of Karlsruhe., 1975. M. Zhang and J. Deng. . , [**273**]{}, pages 102–110, 2013.

ArXiv

--- abstract: 'The Barzilai-Borwein (BB) gradient method is efficient for solving large-scale unconstrained problems to the modest accuracy and has a great advantage of being easily extended to solve a wide class of constrained optimization problems. In this paper, we propose a new stepsize to accelerate the BB method by requiring finite termination for minimizing two-dimensional strongly convex quadratic function. Combing with this new stepsize, we develop gradient methods which adaptively take the nonmonotone BB stepsizes and certain monotone stepsizes for minimizing general strongly convex quadratic function. Furthermore, by incorporating nonmonotone line searches and gradient projection techniques, we extend these new gradient methods to solve general smooth unconstrained and bound constrained optimization. Extensive numerical experiments show that our strategies of properly inserting monotone gradient steps into the nonmonotone BB method could significantly improve its performance and the new resulted methods can outperform the most successful gradient decent methods developed in the recent literature.' author: - Yakui Huang - 'Yu-Hong Dai' - 'Xin-Wei Liu' - Hongchao Zhang date: 'Received: date / Accepted: date' title: 'On the acceleration of the Barzilai-Borwein method' --- Introduction {#intro} ============ Gradient descent methods have been widely used for solving smooth unconstrained nonlinear optimization $$\label{eqpro} \min_{x\in\mathbb{R}^n}~f(x)$$ by generating a sequence of iterates $$\label{eqitr} x_{k+1}=x_k-\alpha_kg_k,$$ where $f: \mathbb{R}^n \to \mathbb{R}$ is continuously differentiable, $g_k=\nabla f(x_k)$ and $\alpha_k>0$ is the stepsize along the negative gradient. Different gradient descent methods would have different rules for determining the stepsize $\alpha_k$. The classic steepest descent (SD) method proposed by Cauchy [@cauchy1847methode] determines its stepsize by the so-called exact line search $$\label{sd} \alpha_k^{SD}=\arg\min_{\alpha \in \mathbb{R}}~f(x_k-\alpha g_k).$$ Although the SD method locally has the most function value reduction along the negative gradient direction, it often performs poorly in practice. Theoretically, when $f$ is a strongly convex quadratic function, i.e., $$\label{qudpro} f(x)=\frac{1}{2}x ^T Ax-b ^T x,$$ where $b\in\mathbb{R}^n$ and $A\in\mathbb{R}^{n\times n}$ is symmetric and positive definite, SD method converges $Q$-linearly [@akaike1959successive] and will asymptotically perform zigzag between two orthogonal directions [@forsythe1968asymptotic; @huang2019asymptotic]. In 1988, Barzilai and Borwein [@barzilai1988two] proposed the following two novel stepsizes that significantly improve the performance of gradient descent methods: $$\label{sBB} \alpha_k^{BB1}=\frac{s_{k-1}^Ts_{k-1}}{s_{k-1}^Ty_{k-1}} \qquad \mbox{and} \qquad \alpha_k^{BB2}=\frac{s_{k-1}^Ty_{k-1}}{y_{k-1}^Ty_{k-1}},$$ where $s_{k-1}=x_k-x_{k-1}$ and $y_{k-1}=g_k-g_{k-1}$. Clearly, when $s_{k-1}^Ty_{k-1}>0$, one has $\alpha_{k}^{BB1}\geq\alpha_{k}^{BB2}$. Hence, $\alpha_{k}^{BB1}$ is often called the *long* BB stepsize while $\alpha_{k}^{BB2}$ is called the *short* BB stepsize. When the objective function is quadratic , the BB stepsize $\alpha_{k}^{BB1}$ will be exactly the steepest descent stepsize, but with one step retard, while $\alpha_{k}^{BB2}$ will be just the stepsize of minimal gradient (MG) method [@dai2003altermin], that is $$\alpha_k^{BB1}=\frac{g_{k-1}^Tg_{k-1}}{g_{k-1}^TAg_{k-1}}=\alpha_{k-1}^{SD} \quad \mbox{and} \quad \alpha_k^{BB2}=\frac{g_{k-1}^TAg_{k-1}}{g_{k-1}^TA^2g_{k-1}}=\alpha_{k-1}^{MG}.$$ It is proved that the Barzilai-Borwein (BB) method converges $R$-superlinearly for minimizing two-dimensional strongly convex quadratic function [@barzilai1988two] and $R$-linearly for the general $n$-dimensional case [@dai2002r]. Although the BB method does not decrease the objective function value monotonically, extensive numerical experiments show that it performs much better than the SD method [@fletcher2005barzilai; @raydan1997barzilai; @yuan2008step]. And it is commonly accepted that when a not high accuracy is required, BB-type methods could be even competitive with nonlinear conjugate gradient (CG) methods for solving smooth unconstrained optimization [@fletcher2005barzilai; @raydan1997barzilai]. Furthermore, by combing with gradient projection techniques, the BB-type methods have a great advantage of easy extension to solve a wide class of constrained optimization, for example the bound or simplex constrained optimization [@dai2005projected]. Hence, BB-type methods enjoy many important applications, such as image restoration [@wang2007projected], signal processing [@liu2011coordinated], eigenvalue problems [@jiang2013feasible], nonnegative matrix factorization [@huang2015quadratic], sparse reconstruction [@wright2009sparse], machine learning [@tan2016barzilai], etc. Recently, Yuan [@yuan2006new; @yuan2008step] propose a gradient descent method which combines a new stepsize $$\label{syv} % \alpha_k^{Y}=\frac{2}{\sqrt{(1/\alpha_{k-1}^{SD}-1/\alpha_k^{SD})^2+ % 4\|g_k\|^2/(\alpha_{k-1}^{SD}\|g_{k-1}\|)^2}+(1/\alpha_{k-1}^{SD}+1/\alpha_k^{SD})}, \alpha_k^{Y}=\frac{2}{\frac{1}{\alpha_{k-1}^{SD}}+\frac{1}{\alpha_{k}^{SD}}+ \sqrt{\left(\frac{1}{\alpha_{k-1}^{SD}}-\frac{1}{\alpha_{k}^{SD}}\right)^2+ \frac{4\|g_k\|^2}{(\alpha_{k-1}^{SD}\|g_{k-1}\|)^2}}},$$ in the SD method so that the new method enjoys finite termination for minimizing a two-dimensional strongly convex quadratic function. Based on this new stepsize $\alpha_k^Y$, Dai and Yuan [@dai2005analysis] further develop the DY method, which alternately employs $\alpha_k^{SD}$ and $\alpha_k^{Y}$ stepsizes as follows $$\label{sdy} \alpha_k^{DY}=\left\{ \begin{array}{ll} \alpha_k^{SD}, & \hbox{if mod($k$,4)$<2$;} \\ \alpha_k^{Y}, & \hbox{otherwise.} \end{array} \right.$$ It is easy to see that $\alpha_k^{Y}\leq\alpha_k^{SD}$. Hence, DY method is a monotone method. Moreover, it is shown that DY method performs better than the nonmonotone BB method [@dai2005analysis]. The property of nonmonotonically reducing objective function values is an intrinsic feature that causes the efficiency of BB method. However, it is also pointed out by Fletcher [@fletcher2012limited] that retaining monotonicity is important for a gradient method, especially for minimizing general objective functions. On the other hand, although the monotone DY method performs well, using $\alpha_k^{SD}$ and $\alpha_k^{Y}$ in a nonmonotone fashion may yield better performance, see [@de2014efficient] for example. Moreover, it is usually difficult to compute the exact monotone stepsize $\alpha_k^{SD}$ in general optimization. Hence, in this paper, motivated by the great success of the BB method and the previous considerations, we want to further improve and accelerate the *nonmonotone* BB method by incorporating some *monotone* steps. For a more general and uniform analysis, we first consider to accelerate the class of gradient descent methods for quadratic optimization using the following stepsize $$\label{glstep1} \alpha_k(\Psi(A))=\frac{g_{k-1}^T\Psi(A) g_{k-1}}{g_{k-1}^T\Psi(A)Ag_{k-1}},$$ where $\Psi(\cdot)$ is a real analytic function on $[\lambda_1,\lambda_n]$ that can be expressed by a Laurent series $$\Psi(z)=\sum_{k=-\infty}^\infty c_kz^k,~~c_k\in\mathbb{R},$$ such that $0<\sum_{k=-\infty}^\infty c_kz^k<+\infty$ for all $z\in[\lambda_1,\lambda_n]$. Here, $\lambda_1$ and $\lambda_n$ are the smallest and largest eigenvalues of $A$, respectively. Clearly, the method is generally nonmonotone and the two BB stepsizes $\alpha_k^{BB1}$ and $\alpha_k^{BB2}$ can be obtained by setting $\Psi(A)=I$ and $\Psi(A)=A$ in , respectively. More particularly, we will derive a new stepsize, say $\tilde{\alpha}_k(\Psi(A))$, which together with the stepsize $\alpha_k(\Psi(A))$ in can minimize the two-dimensional convex quadratic function in no more than $5$ iterations. To the best of our knowledge, this is the first nonmonotone gradient method with finite termination property. We will see that $\tilde{\alpha}_k(I)\leq\alpha_k^{SD}$ and $\tilde{\alpha}_k(A)\leq\alpha_k^{MG}$. Hence, this finite termination property is essentially obtained by inserting monotone stepsizes into the generally nonmonotone stepsizes . In fact, we show that this finite termination property can be maintained even the algorithm uses different function $\Psi$’s during its iteration. Based on this observation, to achieve good numerical performance, we propose an adaptive nonmonotone gradient method (ANGM), which adaptively takes some nonmonotone steps involving the long and short BB stepsizes , and some monotone steps using $\tilde{\alpha}_k(A)$. Moreover, to efficiently minimize more general nonlinear objective function, we propose two variants of ANGM, called ANGR1 and ANGR2, using certain retard stepsize. By combing nonmonotone line search and gradient projection techniques, these two variants of gradient methods are further extended to solve bound constrained optimization. Our numerical experiments show that the new proposed methods significantly accelerate the BB method and perform much better on minimizing quadratic function than the most successful gradient decent methods developed in the recent literature, such as DY method [@dai2005analysis], ABBmin2 method [@frassoldati2008new] and SDC method [@de2014efficient]. In addition, we also compare ANGR1 and ANGR2 with the spectral projected gradient (SPG) method [@birgin2000nonmonotone; @birgin2014spectral] and the BB method using Dai-Zhang nonmonotone line search (BB1-DZ) [@dai2001adaptive] for solving the general unconstrained problems from [@andrei2008unconstrained] and the bound constrained problems from the CUTEst collection [@gould2015cutest]. The numerical results highly suggest the potential benefits of our new proposed methods for solving more general unconstrained and bound constrained optimization. The paper is organized as follows. In Section \[secabb\], we derive the new stepsize $\tilde{\alpha}_k(\Psi(A))$ by requiring finite termination on minimizing two-dimensional strongly convex quadratic function. In Section \[secarbbmd\], we first derive the ANGM, ANGR1 and ANGR2 methods for minimizing general strongly convex quadratic function and then generalize ANGR1 and ANGR2 to solve bound constrained optimization. Our extensive numerical experiments on minimizing strongly convex quadratic function and solving general unconstrained and bound constrained optimization are presented in Section \[secnum\]. We finally draw some conclusion remarks in Section \[secclu\]. Derivation of new stepsize {#secabb} ========================== In this section, we derive a new monotone stepsize based on the nonmonotone gradient method to minimize quadratic function . This new stepsize is motivated by requiring finite termination for minimizing two-dimensional strongly convex quadratic function. Such an idea was originally proposed by Yuan [@yuan2006new] to accelerate SD method. However, new techniques need to be developed for accelerating the class of nonmonotone gradient descent methods since the key orthogonal property of successive two gradients generated by SD method no longer holds for methods . Observe that the method is invariant under translations and rotations when minimizing quadratics. Hence, for theoretical analysis to minimize , without loss of generality, we may simply assume that the matrix $A$ is diagonal, i.e., $$\label{formA} A=\textrm{diag}\{\lambda_1,\ldots,\lambda_n\},$$ where $0 < \lambda_1\leq\ldots\leq\lambda_n$. Motivation {#submot} ---------- First, let us investigate the behavior of gradient method with $\Psi(A)=I$ (i.e., the BB1 method). Particularly, we apply it to the non-random quadratic minimization problem proposed in [@de2014efficient], which has the form with a diagonal matrix $A$ given by $$\label{pro2} A_{jj}=10^{\frac{ncond}{n-1}(n-j)}, ~~~j=1,\ldots,n,$$ and $b$ being a null vector. Here, $ncond=\log_{10} \kappa$ and $\kappa>0$ is the condition number of $A$. We set $n=10$, $\kappa = 10^3$ and use $(10,10,\ldots,10)^T$ as the initial point. The iteration was stopped once the gradient norm is reduced by a factor of $10^{-6}$. Denote the $i$-th component of $g_k$ by $g_k^{(i)}$ and the indices of the components of $g_k$ with two largest magnitudes by $i_1$ and $i_2$, respectively. Then, the percentage of the magnitudes of the first two largest components to that of the whole gradient can be computed by $$\Upsilon(g_k)=\frac{|g_k^{(i_1)}|+|g_k^{(i_2)}|}{\sum_{i=1}^n|g_k^{(i)}|}.$$ This $\Upsilon(g_k)$ is plotted in Fig. \[bbdom2\] (left), where we can see that $\Upsilon(g_k)\geq0.8$ holds for more than half of the iterations (145 out of 224 total iterations). Hence, roughly speaking, the searches of the BB1 method are often dominated in some two-dimensional subspaces. The history of index $i_1$ against the iteration number is also plotted in Fig. \[bbdom2\] (right), where we can see that $|g_k^{(i_1)}|$ corresponds more frequently to the largest eigenvalues $\lambda_{10}$ or $\lambda_{9}$. Since $$\label{updgki} g_{k+1}^{(j)}=(1-\alpha_k\lambda_j)g_k^{(j)},~j=1,\ldots,n.$$ and $1/\lambda_{n}\leq\alpha_k\leq1/\lambda_1$, the history of $i_1$ in Fig. \[bbdom2\] (right) in fact indicates that, most stepsizes generated by the BB1 method are often much larger than $1/\lambda_{10}$ or $1/\lambda_{9}$. As a result, the BB1 method may need many iterations to reduce the corresponding components of the gradients $g_k^{(9)}$ or $g_k^{(10)}$. \ In [@huang2019asymptotic], we showed that a family of gradient methods including SD and MG will asymptotically reduce their searches in a two-dimensional subspace and could be accelerated by exploiting certain orthogonal properties in this two-dimensional subspace. In a similar spirit, we could also accelerate the convergence of the class of gradient methods in a lower dimensional subspace if certain orthogonal properties hold. Suppose that, for a given $k>0$, there exists a $q_k$ satisfying $$\label{defqk} (I-\alpha_{k-1}A)q_k=g_{k-1}.$$ Since this $q_k$ is also invariant under translations and rotations, for later analysis we may still assume $A$ in is diagonal as in . The following lemma shows a generalized orthogonal property for $q_k$ and $g_{k+1}$, which is a key property for deriving our new stepsize in the next subsection. \[orpty\] Suppose that the sequence $\{g_k\}$ is obtained by applying gradient method with stepsizes to minimize a quadratic function and $q_k$ satisfies . Then, we have $$\label{orthg} q_k^T\Psi(A)g_{k+1}=0.$$ By , and we get $$\begin{aligned} q_k^T\Psi(A)g_{k+1} &=q_k^T\Psi(A)(I-\alpha_kA)(I-\alpha_{k-1}A)g_{k-1} \\ &=g_{k-1}^T\Psi(A)(I-\alpha_kA)g_{k-1} \\ &=g_{k-1}^T\Psi(A)g_{k-1}-\alpha_kg_{k-1}^T\Psi(A)Ag_{k-1}=0.\end{aligned}$$ This completes the proof. A new stepsize {#sectwod} -------------- In this subsection, we derive a new stepsize based on the iterations of gradient method . We show that combining the new stepsize with gradient method , we can achieve finite termination for minimizing two-dimensional strongly convex quadratic functions. By Lemma \[orpty\], we have that $g_{k}^T\Psi(A)q_{k-1}=0$ for $k>0$. Now, suppose both $\Psi^r(A)q_{k-1}$ and $\Psi^{1-r}(A)g_{k}$ are nonzero vectors, where $r \in \mathbb{R}$. Let us consider to minimize the function $f$ in a two-dimensional subspace spanned by $\frac{\Psi^r(A)q_{k-1}}{\|\Psi^r(A)q_{k-1}\|}$ and $\frac{\Psi^{1-r}(A)g_{k}}{\|\Psi^{1-r}(A)g_{k}\|}$, and let $$\begin{aligned} \label{ftu1} \varphi(t,l) &:= f\left(x_{k}+t\frac{\Psi^r(A)q_{k-1}}{\|\Psi^r(A)q_{k-1}\|} +l\frac{\Psi^{1-r}(A)g_{k}}{\|\Psi^{1-r}(A)g_{k}\|}\right) \nonumber\\ & = f(x_{k}) +\vartheta_k^T \begin{pmatrix} t \\ l \\ \end{pmatrix} + \frac{1}{2} \begin{pmatrix} t \\ l \\ \end{pmatrix} ^T H_k \begin{pmatrix} t \\ l \\ \end{pmatrix},\end{aligned}$$ where $$\label{eqgrad} \vartheta_k= B_k g_{k} = \begin{pmatrix} \frac{g_{k}^T\Psi^r(A)q_{k-1}}{\|\Psi^r(A)q_{k-1}\|}\\ \frac{g_{k}^T\Psi^{1-r}(A)g_{k}}{\|\Psi^{1-r}(A)g_{k}\|} \\ \end{pmatrix} \mbox{ with } B_k=\begin{pmatrix} \frac{\Psi^r(A)q_{k-1}}{\|\Psi^r(A)q_{k-1}\|}, \frac{\Psi^{1-r}(A)g_{k}}{\|\Psi^{1-r}(A)g_{k}\|} \\ \end{pmatrix} ^T$$ and $$\label{eqhes} H_k= B_k A B_k^T = \begin{pmatrix} \frac{q_{k-1}^T\Psi^{2r}(A)Aq_{k-1}}{\|\Psi^r(A)q_{k-1}\|^2} & \frac{q_{k-1}^T\Psi(A) Ag_{k}}{\|\Psi^r(A)q_{k-1}\|\|\Psi^{1-r}(A)g_{k}\|} \\ \frac{q_{k-1}^T\Psi(A)Ag_{k}}{\|\Psi^r(A)q_{k-1}\|\|\Psi^{1-r}(A)g_{k}\|} & \frac{g_{k}^T\Psi^{2(1-r)}(A)Ag_{k}}{\|\Psi^{1-r}(A)g_{k}\|^2}\\ \end{pmatrix}.$$ Denote the components of $H_k$ by $H_k^{(ij)}$, $i,j=1,2$ and notice that $B_k B_k^T = I$ by $g_{k}^T\Psi(A)q_{k-1}=0$. Then, we have the following finite termination theorem. \[thftm\] Suppose that a gradient method is applied to minimize a two-dimensional quadratic function with $\alpha_k$ given by for all $k \neq k_0$ and uses the stepsize $$\begin{aligned} \label{new2} \tilde{\alpha}_{k_0} &= \frac{2}{(H_{k_0}^{(11)}+H_{k_0}^{(22)})+\sqrt{(H_{k_0}^{(11)}-H_{k_0}^{(22)})^2+4(H_{k_0}^{(12)})^2}}\end{aligned}$$ at the $k_0$-th iteration where $k_0 \geq2$. Then, the method will find the minimizer in at most $k_0+3$ iterations. Let us suppose $x_k$ is not a minimizer for all $k=1, \ldots, k_0+2$. We then show $k_0+3$ must be the minimizer, i.e., $g_{k_0+3} = 0$. For notation convenience, in the following proof of this theorem, let’s simply use $k$ to denote $k_0$. First, we show that using stepsize at the $k$-th iteration implies $$\label{parallel-g2} g_{k+1}\quad \mbox{is parallel to} \quad - B_k^T H_k^{-1} \vartheta_k + \tilde{\alpha}_{k} g_{k},$$ where $\vartheta_k, B_k$ and $H_k$ is given by and . In fact, $\tilde{\alpha}_{k}$ given by satisfies the following quadratic equation $$\label{eqalp2} \tilde{\alpha}_{k}^2\Delta -\tilde{\alpha}_{k}(H_k^{(11)}+H_k^{(22)})+1=0,$$ where $\Delta=\textrm{det}(H_k) = \textrm{det}(A) >0$. Let $$\Theta=(H_k^{(12)}\vartheta_k^{(1)}+H_k^{(22)}\vartheta_k^{(2)})\vartheta_k^{(1)} -(H_k^{(11)}\vartheta_k^{(1)}+H_k^{(12)}\vartheta_k^{(2)})\vartheta_k^{(2)},$$ where $\vartheta_k^{(i)}$ are components of $\vartheta_k$, $i=1,2$. Then, multiplying $\Theta$ to , we have $$\label{eqalp1} \tilde{\alpha}_{k}^2\Delta \Theta- \tilde{\alpha}_{k}(H_k^{(11)}+H_k^{(22)})\Theta+\Theta=0,$$ which is exactly $$\begin{aligned} \label{parallel-explicit} &(H_k^{(22)}\vartheta_k^{(1)}-H_k^{(12)}\vartheta_k^{(2)}-\tilde{\alpha}_k\Delta \vartheta_k^{(1)}) [\vartheta_k^{(2)}- \tilde{\alpha}_k(H_k^{(12)}\vartheta_k^{(1)}+H_k^{(22)}\vartheta_k^{(2)})] \nonumber\\ =&(H_k^{(11)}\vartheta_k^{(2)}-H_k^{(12)}\vartheta_k^{(1)}-\tilde{\alpha}_k\Delta \vartheta_k^{(2)}) [\vartheta_k^{(1)}-\tilde{\alpha}_k(H_k^{(11)}\vartheta_k^{(1)}+H_k^{(12)}\vartheta_k^{(2)})].\end{aligned}$$ The above identity implies the vector $$\begin{pmatrix} \vartheta_k^{(1)}- \tilde{\alpha}_k(H_k^{(11)}\vartheta_k^{(1)}+H_k^{(12)}\vartheta_k^{(2)}) \\ \vartheta_k^{(2)}- \tilde{\alpha}_k(H_k^{(12)}\vartheta_k^{(1)}+H_k^{(22)}\vartheta_k^{(2)})\\ \end{pmatrix}$$ is parallel to $$\begin{pmatrix} H_k^{(22)}\vartheta_k^{(1)}-H_k^{(12)}\vartheta_k^{(2)}- \tilde{\alpha}_k\Delta \vartheta_k^{(1)} \\ H_k^{(11)}\vartheta_k^{(2)}-H_k^{(12)}\vartheta_k^{(1)}-\tilde{\alpha}_k\Delta \vartheta_k^{(2)}\\ \end{pmatrix},$$ which written in a matrix format just means $$\label{g2pal2} \vartheta_k+H_k(- \tilde{\alpha}_{k}\vartheta_k) \quad \mbox{is parallel to} \quad H_k^{-1}\vartheta_k - \tilde{\alpha}_{k} \vartheta_k.$$ Since $n=2$, we have $B_k B_k^T =B_k^T B_k=I$. Then, it follows from $g_k = B_k^T \vartheta_k$, $\vartheta_k = B_k g_k$, $H_k = B_k A B_k^T$ and $g_{k+1} = g_k - \tilde{\alpha}_k A g_k$ that $g_{k+1}= B_k^T \vartheta_k+ B_k^T H_k(- \tilde{\alpha}_{k}\vartheta_k)$. So, we have from that holds. Therefore, implies holds. Now, it follows from and $H_{k}^{-1} = B_k A^{-1} B_k^T$ that $$- B_k^T H_{k}^{-1} \vartheta_k + \tilde{\alpha}_{k} g_{k} = - A^{-1} g_{k} + \tilde{\alpha}_{k} g_{k} = - A^{-1}(g_{k}- \tilde{\alpha}_{k} A g_{k}) = - A^{-1} g_{k+1}.$$ Hence, implies $g_{k+1}$ is parallel to $A^{-1} g_{k+1}$. So, if $\tilde{\alpha}_{k}$ given by is used at the $k$-th iteration, then $g_{k+1}$ is parallel to $A^{-1} g_{k+1}$. Since $x_{k+1}$ is not the minimizer, we have $g_{k+1} \ne 0$. So, $g_{k+1}$ is an eigenvector of $A$, i.e. $Ag_{k+1} = \lambda g_{k+1}$ for some $\lambda > 0$. Since $x_{k+2}$ is not the minimizer, we have $g_{k+2} \ne 0 $ and the algorithm will not stop at the $k+2$-th iteration. So, by , we have $\alpha_{k+2}=\frac{g_{k+1}^T \Psi(A) g_{k+1}}{g_{k+1}^T \Psi(A)A g_{k+1}} = 1/\lambda$. Hence, we have $g_{k+3}=(1-\alpha_{k+2}\lambda)g_{k+2} =0$, which implies $x_{k+3}$ must be the minimizer. We complete the proof. Notice that by setting $k_0 =2$ in the above Theorem \[thftm\], the new gradient method in Theorem \[thftm\] will find the exact minimizer in at most $5$ iterations when minimizing a two-dimensional strongly convex quadratic function. In fact, since $\Delta=\lambda_1\lambda_2$ and $H_k^{(11)}+H_k^{(22)}=\lambda_1+\lambda_2$, the equation has two positive roots $1/\lambda_1$ and $1/\lambda_2$. This observation allows us to use the stepsize $\tilde{\alpha}_{k_0}$ with some retards as stated in the following corollary, which would lead us a more convenient way for choosing stepsizes when the objective function is not quadratic. Suppose that a gradient method is applied to a two-dimensional quadratic function with $\alpha_{k_0+m}=\tilde{\alpha}_{k_0}$ for $k_0\geq2$ and some positive integer $m$, and $\alpha_k$ given by for all $k \neq k_0+m$. Then, the method stops in at most $k_0+m+3$ iterations. By setting $\Psi(A)=I$, $\Psi(A)=A$ and $r=1/2$ in , and setting $k_0=k$ in , we can derive the following two stepsizes: $$\label{newsbb1} \tilde{\alpha}_k^{BB1}= \frac{2}{\frac{q_{k-1}^TAq_{k-1}}{\|q_{k-1}\|^2} + \frac{1}{\alpha_k^{SD}}+ \sqrt{\left(\frac{q_{k-1}^TAq_{k-1}}{\|q_{k-1}\|^2} - \frac{1}{\alpha_k^{SD}}\right)^2+\frac{4(q_{k-1}^TAg_k)^2}{\|q_{k-1}\|^2\|g_k\|^2}}}$$ and $$\label{newsbb2} \tilde{\alpha}_{k}^{BB2}= \frac{2}{\frac{1}{\hat{\alpha}_{k-1}} +\frac{1}{\alpha_k^{MG}}+ \sqrt{\left(\frac{1}{\hat{\alpha}_{k-1}}-\frac{1}{\alpha_k^{MG}}\right)^2 +\Gamma_k}},$$ respectively, where $$\label{snew21} \hat{\alpha}_{k}=\frac{q_{k}^TAq_{k}}{q_{k}^TA^2q_{k}}~~\textrm{and}~~ \Gamma_k=\frac{4(q_{k-1}^TA^2g_k)^2}{q_{k-1}^TAq_{k-1}\cdot g_k^TA g_k}. %=\frac{\alpha_{k-1}(\|q_{k}\|^2-q_{k}^Tg_{k-1})} %{\|q_{k}\|^2-2q_{k}^Tg_{k-1}+\|g_{k-1}\|^2}.$$ By and , we have $$\label{upbdtalp3} \tilde{\alpha}_k^{BB1}\leq\min\left\{\alpha_k^{SD},\frac{\|q_{k-1}\|^2}{q_{k-1}^TAq_{k-1}}\right\} ~~\textrm{and}~~ \tilde{\alpha}_{k}^{BB2}\leq\min\{\alpha_k^{MG},\hat{\alpha}_{k-1}\}.$$ Hence, both $\tilde{\alpha}_k^{BB1}$ and $\tilde{\alpha}_k^{BB2}$ are short monotone steps for reducing the value and gradient norm of the objective function, respectively. And it follows from Theorem \[thftm\] that by inserting the monotone steps $\tilde{\alpha}_k^{BB1}$ and $\tilde{\alpha}_k^{BB2}$ into the BB1 and BB2 methods, respectively, the gradient method will have finite termination for minimizing two-dimensional strongly convex quadratic functions. To numerically verify this finite termination property, we apply the method with $\Psi(A)=I$ (i.e., the BB1 method) and $\tilde{\alpha}_2^{BB1}$ given by to minimize a two-dimensional quadratic function with $$\label{twoquad} A=\textrm{diag}\{1,\lambda\} \quad \mbox{and} \quad b=0.$$ We run the algorithm for five iterations using ten random starting points. The averaged values of $\|g_5\|$ and $f(x_5)$ are presented in Table \[tb2ft\]. Moreover, we also run BB1 method for a comparison purpose. We can observe that for different values of $\lambda$, the values of $\|g_5\|$ and $f(x_5)$ obtained by BB1 method with $\tilde{\alpha}_2^{BB1}$ given by are numerically very close to zero. However, even for the case $\lambda=10$, $\|g_5\|$ and $f(x_5)$ obtained by pure BB1 method are far away from zero. These numerical results coincide with our analysis and show that the nonmonotone method can be significantly accelerated by incorporating proper monotone steps. New methods {#secarbbmd} =========== In this section, based on the above analysis, we propose an adaptive nonmonotone gradient method (ANGM) and its two variants, ANGR1 and ANGR2, for solving both unconstrained and box constrained optimization. These new gradient methods adaptively take some nonmonotone steps involving the long and short BB stepsizes , and some monotone steps using the new stepsize developed in the previous section. Quadratic case -------------- As mentioned in Section \[submot\], the stepsizes $\alpha_{k}^{BB1}$ generated by the BB1 method may be far away from the reciprocals of the largest eigenvalues of the Hessian matrix $A$ of the quadratic function . In other words, the stepsize $\alpha_{k}^{BB1}$ may be too large to effectively decrease the components of gradient $g_k$ corresponding to the first several largest eigenvalues, which, by , can be greatly reduced when small stepsizes are employed. In addition, it has been observed by many works in the recent literature that gradient methods using long and short stepsizes adaptively generally perform much better than those using only one type of stepsizes, for example see [@dai2003altermin; @de2014efficient; @di2018steplength; @gonzaga2016steepest; @huang2019gradient; @huang2019asymptotic]. So, we would like to develop gradient methods that combines the two nonmonotone BB stepsizes with the short monotone stepsize given by . We first extend the orthogonal property developed in Lemma \[orpty\] and the finite termination result given in Theorem \[thftm\]. \[gorpty\] Suppose that a gradient method with stepsizes in the form of is applied to minimize a quadratic function . In particular, at the $k-1$-th and $k$-th iteration, two stepsizes $\alpha_{k-1}(\Psi(A))$ and $\alpha_k(\Psi_1(A))$ are used, respectively, where $\Psi$ and $\Psi_1$ may be two different analytic functions used in . If $q_k \in \mathbb{R}^n$ satisfies $$\label{gdefqk} (I-\alpha_{k-1}(\Psi(A))A)q_k=g_{k-1},$$ then we have $$\label{gorthg} q_k^T\Psi_1(A)g_{k+1}=0.$$ Notice that by , we have $$g_k=g_{k-1}-\alpha_{k-1}(\Psi(A))Ag_{k-1} \quad \mbox{and} \quad g_{k+1}=g_k-\alpha_k(\Psi_1(A))Ag_k.$$ Then, the proof is essential the same as those in the proof of Lemma \[orpty\]. Based on Lemma \[gorpty\] and using the same arguments as those in the proof of Theorem \[thftm\], we can obtain the following finite termination result even different function $\Psi$’s are used in to obtain the stepsizes. \[gthftm\] Suppose that a gradient method is applied to minimize a two-dimensional quadratic function with $\alpha_k$ given by for all $k \neq k_0$ and $k \ne k_0-1$, and uses the stepsizes $\alpha_{k-1}(\Psi_1(A))$ and $\alpha_k(\Psi_1(A))$ at the $k-1$-th and $k$-th iteration, respectively, where $k_0 \geq2$. Then, the method will find the minimizer in at most $k_0+3$ iterations. Theorem \[gthftm\] allows us to incorporate the nonmonotone BB stepsizes $\alpha_{k}^{BB1}$ and $\alpha_{k}^{BB2}$, and the short monotone stepsize $\tilde{\alpha}_{k}^{BB2}$ in one gradient method. Alternate or adaptive scheme has been employed for choosing long and short stepsizes in BB-type methods [@dai2005projected; @zhou2006gradient]. And recent studies show that adaptive strategies are more preferred than the alternate scheme [@dhl2018; @zhou2006gradient]. Hence, we would like develop adaptive strategies to choose proper stepsizes for our new gradient methods. In particular, our adaptive nonmonotone gradient method (ANGM) takes the long BB stepsize $\alpha_{k}^{BB1}$ when $\alpha_{k}^{BB2}/\alpha_{k}^{BB1}\geq\tau_1$ for some $\tau_1\in(0,1)$. Otherwise, a short stepsize $\alpha_{k}^{BB2}$ or $\tilde{\alpha}_{k}^{BB2}$ will be taken depending on the ratio $\|g_{k-1}\|/\|g_{k}\|$. Notice that $\alpha_{k}^{BB2}$ minimizes the gradient in the sense that $$\alpha_{k}^{BB2}=\alpha_{k-1}^{MG}=\arg\min_{\alpha \in \mathbb{R}} \|g_{k-1}-\alpha Ag_{k-1}\|.$$ So, when $\|g_{k-1}\|/\|g_{k}\| > \tau_2 $ for some $\tau_2 >1$, i.e. the gradient norm decreases, the previous stepsize $\alpha_{k-1}$ is often a reasonable approximation of $\alpha_{k}^{BB2}$. By our numerical experiments, when BB method is applied the searches are often dominated in some two-dimensional subspaces. And the new gradient method in Theorem \[gthftm\] would have finite convergence for minimizing two-dimensional quadratic function when the new stepsize $\tilde{\alpha}_{k}^{BB2}$ is applied after some BB2 steps. Hence, our ANGM would employ the new monotone stepsize $\tilde{\alpha}_{k}^{BB2}$ when $\|g_{k-1}\|\geq\tau_2\|g_{k}\|$; otherwise, certain BB2 steps should be taken. In practice, we find that when $\|g_{k-1}\| \le \tau_2\|g_{k}\|$, ANGM often has good performance by taking the stepsize $\min\{\alpha_{k}^{BB2},\alpha_{k-1}^{BB2}\}$. To summarize, our ANGM applies the following adaptive strategies for choosing stepsizes: $$\label{newmd0} \alpha_{k}= \begin{cases} \min\{\alpha_{k}^{BB2},\alpha_{k-1}^{BB2}\},& \text{if $\alpha_{k}^{BB2}<\tau_1\alpha_{k}^{BB1}$ and $\|g_{k-1}\|<\tau_2\|g_{k}\|$}; \\ \tilde{\alpha}_{k}^{BB2},& \text{if $\alpha_{k}^{BB2}<\tau_1\alpha_{k}^{BB1}$ and $\|g_{k-1}\|\geq\tau_2\|g_{k}\|$}; \\ \alpha_{k}^{BB1},& \text{otherwise}, \end{cases}$$ Notice that the calculation of $\tilde{\alpha}_{k}^{BB2}$ needs to compute $\alpha_k^{MG}$ which is not easy to obtain when the objective function is not quadratic. In stead, the calculation of $\tilde{\alpha}_{k-1}^{BB2}$ will just require $\alpha_k^{BB2}$, which is readily available even for general objective function. Moreover, it is found in recent research that gradient methods using retard stepsizes can often lead better performances [@friedlander1998gradient]. Hence, in the first variant of ANGM, we simply replace $\tilde{\alpha}_{k}^{BB2}$ in by $\tilde{\alpha}_{k-1}^{BB2}$, i.e. the stepsizes are chosen as $$\label{newmd1} \alpha_{k}= \begin{cases} \min\{\alpha_{k}^{BB2},\alpha_{k-1}^{BB2}\},& \text{if $\alpha_{k}^{BB2}<\tau_1\alpha_{k}^{BB1}$ and $\|g_{k-1}\|<\tau_2\|g_{k}\|$}; \\ \tilde{\alpha}_{k-1}^{BB2},& \text{if $\alpha_{k}^{BB2}<\tau_1\alpha_{k}^{BB1}$ and $\|g_{k-1}\|\geq\tau_2\|g_{k}\|$}; \\ \alpha_{k}^{BB1},& \text{otherwise}. \end{cases}$$ We call the gradient method using stepsize ANGR1. On the other hand, since the calculation of $\tilde{\alpha}_{k-1}^{BB2}$ also needs $\hat{\alpha}_{k-2}$ and $\Gamma_{k-1}$ and by , $$\label{upbdtalp2} \tilde{\alpha}_{k-1}^{BB2}\leq\min\{\alpha_k^{BB2},\hat{\alpha}_{k-2}\},$$ to simplify ANGR1, we may further replace $\tilde{\alpha}_{k-1}^{BB2}$ in by its upper bound in . As a result, we have the second variant of ANGM, which chooses stepsizes as $$\label{newmd2} \alpha_{k}= \begin{cases} \min\{\alpha_{k}^{BB2},\alpha_{k-1}^{BB2}\},& \text{if $\alpha_{k}^{BB2}<\tau_1\alpha_{k}^{BB1}$ and $\|g_{k-1}\|<\tau_2\|g_{k}\|$}; \\ \min\{\alpha_k^{BB2},\hat{\alpha}_{k-2}\},& \text{if $\alpha_{k}^{BB2}<\tau_1\alpha_{k}^{BB1}$ and $\|g_{k-1}\|\geq\tau_2\|g_{k}\|$}; \\ \alpha_{k}^{BB1},& \text{otherwise}. \end{cases}$$ We call the gradient method using stepsize ANGR2. In terms of global convergence for minimizing quadratic function , by , we can easily show the $R$-linear global convergence of ANGM since it satisfies the property in [@dai2003alternate]. Similarly, $R$-linear convergence of ANGR1 and ANGR2 can be also established. See the proof of Theorem 3 in [@dhl2018] for example. Compared with other gradient methods, ANGM, ANGR1 and ANGR2 do not need additional matrix-vector products. In fact, it follows from that $Aq_k=\frac{1}{\alpha_{k-1}}(q_k-g_{k-1})$, which gives $$\label{snew2} \hat{\alpha}_{k}=\frac{q_{k}^TAq_{k}}{q_{k}^TA^2q_{k}} =\frac{\alpha_{k-1}q_{k}^T(q_{k}-g_{k-1})} {(q_{k}-g_{k-1})^T(q_{k}-g_{k-1})}.$$ Hence, no additional matrix-vector products are needed for calculation of $\hat{\alpha}_{k-1}$ in $\tilde{\alpha}_{k}^{BB2}$, $\hat{\alpha}_{k-2}$ in $\tilde{\alpha}_{k-1}^{BB2}$ and the stepsize used in ANGR2. Since the calculation of $Ag_k$ is necessary for the calculation of $g_{k+1}$, $\Gamma_k$ in $\tilde{\alpha}_{k}^{BB2}$ requires no additional matrix-vector products either. As for $\tilde{\alpha}_{k-1}^{BB2}$, it follows from $g_{k-1}^TA^2q_{k-2}=\frac{1}{\alpha_{k-3}}(q_{k-2}-g_{k-3})^T A g_{k-1}$ and $Ag_{k-1}=\frac{1}{\alpha_{k-1}}(g_{k-1}-g_{k})$ that $$\begin{aligned} \Gamma_{k-1}& %=\frac{4(q_{k-1}^TA^2g_k)^2}{q_{k-1}^TAq_{k-1}\cdot g_k^TA g_k} =\frac{4((q_{k-2}-g_{k-3})^T Ag_{k-1})^2}{\alpha_{k-3}((q_{k-2}-g_{k-3})^Tq_{k-2})\cdot g_{k-1}^TA g_{k-1}}\nonumber\\ &=\frac{4((q_{k-2}-g_{k-3})^T(g_{k-1}-g_{k}))^2} {\alpha_{k-3}\alpha_{k-1}((q_{k-2}-g_{k-3})^Tq_{k-2})\cdot g_{k-1}^T(g_{k-1}-g_{k})}.\end{aligned}$$ Thus, no additional matrix-vector products are required for calculation of $\Gamma_{k-1}$ in $\tilde{\alpha}_{k-1}^{BB2}$. Notice that all the new methods, ANGM, ANGR1 and ANGR2, require the vector $q_{k}$ for calculation of their stepsizes. However, computing $q_k$ exactly from maybe as difficult as minimizing the quadratic function. Notice that the $q_k$ satisfying also satisfies the secant equation $$\label{secant-cond} q_k^Tg_{k}=\|g_{k-1}\|^2.$$ Hence, we may find an approximation of $q_k$ by requiring the above secant condition holds. One efficient way to find such a $q_k$ satisfying the secant equation is to simply treat the Hessian matrix $A$ as the diagonal matrix and derive $q_k$ from , that is when $g_k^{(i)} \ne 0$, $$\label{defq} q_k^{(i)}=\frac{g_{k-1}^{(i)}}{1-\alpha_{k-1}\lambda_i}= \frac{(g_{k-1}^{(i)})^2}{g_k^{(i)}},~~i=1,\ldots,n,$$ And we can just let $q_k^{(i)} = 0$, if $g_k^{(i)}=0$. To summarize, the approximated $q_k$ can be computed by $$\label{defq2} q_k^{(i)}= \left\{ \begin{array}{ll} \frac{(g_{k-1}^{(i)})^2}{g_k^{(i)}}, & \hbox{if $g_k^{(i)}\neq0$;} \\ 0, & \hbox{if $g_k^{(i)}=0$.} \end{array} \right.$$ As we will see in Section \[secnum\], this simple way of calculating $q_k$ leads very efficient algorithm. For a simple illustration of numerical behavior of ANGR1, we again applied ANGR1 with $\tau_1=0.85$ and $\tau_2=1.3$ to solve problem with $n=10$. Fig. \[bbabbrg1\] shows the largest component $|g_k^{(i_1)}|$ of the gradient generated by BB1 and ANGR1 methods against the iteration number, where circle means the ANGR1 method takes the new stepsize $\tilde{\alpha}_{k}^{BB2}$ at that iteration. It can be seen that $|g_k^{(i_1)}|$ generated by BB1 method often increases significantly with a much larger value at the iteration where the new stepsize $\tilde{\alpha}_{k}^{BB2}$ is applied. On the other hand, $|g_k^{(i_1)}|$ generated by the ANGR1 method is often reduced and kept small after the new stepsize $\tilde{\alpha}_{k}^{BB2}$ is applied. A detail correspondence of $i_1$ and $\lambda_j$ is presented in Table \[bb1angmnj\], where $n_j$ is the total number of $i_1's$ for which $i_1=j$, $j=1,\ldots,10$. We can see from the last three columns in Table \[bb1angmnj\] that ANGR1 is much efficient than BB1 for decreasing those components of $g_k$ corresponding to large eigenvalues. Hence, the undesired behavior of BB1 discussed in the motivation Section \[submot\] is greatly eliminated by ANGR1. ![Problem with $n=10$: history of $|g_k^{(i_1)}|$ generated by the BB1 and ANGR1 methods[]{data-label="bbabbrg1"}](BB1ANGR1g1t085.eps){width="65.00000%" height="45.00000%"} $n_j$ 1 2 3 4 5 6 7 8 9 10 ------- ---- --- ---- ---- ---- --- ---- ---- ---- ---- BB1 40 5 6 4 6 2 10 21 42 87 ANGR1 51 4 22 18 16 4 1 8 12 17 : The correspondence of $i_1$ and $\lambda_j$ by the BB1 and ANGR1 methods[]{data-label="bb1angmnj"} Bound constrained case ---------------------- In this subsection, we would like to extend ANGR1 and ANGR2 methods for solving the bound constrained optimization $$\label{conprob} \min_{x\in\Omega} ~~f(x),$$ where $f$ is Lipschitz continuously differentiable on the feasible set $\Omega=\{x\in\mathbb{R}^n|~l\leq x\leq u\}$. Here, $l \le x \le u$ means componentwise $l_i \le x_i \le u_i$ for all $i=1, \ldots, n$. Clearly, when $l_i=-\infty$ and $u_i=+\infty$ for all $i$, problem reduces to the smooth unconstrained optimization. Our methods will incorporate the gradient projection strategy and update the iterates as $$x_{k+1}=x_{k}+\lambda_{k}d_{k},$$ with $\lambda_{k}$ being a step length determined by some line searches and $d_k$ being the search direction given by $$\label{dirc} d_k=P_{\Omega}(x_k-\alpha_kg_k)-x_k,$$ where $P_{\Omega}(\cdot)$ is the Euclidean projection onto $\Omega$ and $\alpha_k$ is our proposed stepsize. It is well-known that the components of iterates generated by gradient descent methods corresponding to optimal solutions at the boundary will be finally unchanged when the problem is nondegenerate. Hence, in [@huang2019gradient], the authors suggest to use the following modified BB stepsizes for bound constrained optimization $$\label{bsbb-1-2} \bar{\alpha}_k^{BB1} =\frac{s_{k-1}^Ts_{k-1}}{s_{k-1}^T\bar{y}_{k-1}} \quad \mbox{and} \quad \bar{\alpha}_k^{BB2}=\frac{s_{k-1}^T\bar{y}_{k-1}}{\bar{y}_{k-1}^T\bar{y}_{k-1}},$$ where $\bar{y}_{k-1}$ is given by $$\label{bary} \bar{y}_{k-1}^{(i)}=\left\{ \begin{array}{ll} 0, & \hbox{ if $s_{k-1}^{(i)}=0$;} \\ g_k^{(i)}-g_{k-1}^{(i)}, & \hbox{ otherwise.} \end{array} \right.$$ Notice that $\alpha_k^{BB1} = \bar{\alpha}_k^{BB1}$. We will also do this modifications for solving bound constrained optimization and replace the two BB stepsizes in our new methods by $\bar{\alpha}_k^{BB1}$ and $\bar{\alpha}_k^{BB2}$. As mentioned before, we expect to get short steps using our new stepsizes. Since may not hold for general functions, we would impose $\bar{\alpha}_{k}^{BB2}$ as a safeguard. As a result, our ANGR1 and ANGR2 methods for solving bound constrained optimization employ the following stepsizes: $$\label{newgmd1} \bar{\alpha}_{k}= \begin{cases} \min\{\bar{\alpha}_{k-1}^{BB2},\bar{\alpha}_{k}^{BB2}\},& \text{if $\bar{\alpha}_{k}^{BB2}<\tau_1\bar{\alpha}_{k}^{BB1}$ and $\|\bar{g}_{k-1}\|<\tau_2\|\bar{g}_{k}\|$}; \\ \min\{\bar{\alpha}_{k}^{BB2},\tilde{\alpha}_{k-1}^{BB2}\},& \text{if $\bar{\alpha}_{k}^{BB2}<\tau_1\bar{\alpha}_{k}^{BB1}$ and $\|\bar{g}_{k-1}\|\geq\tau_2\|\bar{g}_{k}\|$}; \\ \bar{\alpha}_{k}^{BB1},& \text{otherwise}, \end{cases}$$ and $$\label{newgmd2} \bar{\alpha}_{k}= \begin{cases} \min\{\bar{\alpha}_{k-1}^{BB2},\bar{\alpha}_{k}^{BB2}\},& \text{if $\bar{\alpha}_{k}^{BB2}<\tau_1\bar{\alpha}_{k}^{BB1}$ and $\|\bar{g}_{k-1}\|<\tau_2\|\bar{g}_{k}\|$}; \\ \min\{\bar{\alpha}_{k}^{BB2},\hat{\alpha}_{k-2}\},& \text{if $\bar{\alpha}_{k}^{BB2}<\tau_1\bar{\alpha}_{k}^{BB1}$ and $\|\bar{g}_{k-1}\|\geq\tau_2\|\bar{g}_{k}\|$}; \\ \bar{\alpha}_{k}^{BB1},& \text{otherwise}, \end{cases}$$ respectively, where $\tau_1\in(0,1)$, $\tau_2\geq1$, and $\bar{g}_{k}=P_{\Omega}(x_{k}-g_{k})-x_{k}$. The overall algorithm of ANGR1 and ANGR2 for solving bound constrained optimization are given in Algorithm \[al1\], where the adaptive nonmonotone line search by Dai and Zhang [@dai2001adaptive] is employed to ensure global convergence and achieve better numerical performance. In particular, the step length $\lambda_{k} = 1$ is accepted if $$\label{nonmls} f(x_{k}+d_{k})\leq f_{r} +\sigma g_{k}^Td_{k},$$ where $f_{r}$ is the so-called reference function value and is adaptively updated by the rules given in [@dai2001adaptive] and $\sigma\in(0,1)$ is a line search parameter. Once is not accepted, it will perform an Armijo-type back tracking line search to find the step length $\lambda_{k}$ such that $$\label{nonmls-2} f(x_{k}+\lambda_{k}d_{k})\leq \min\{f_{\max},f_{r}\}+\sigma\lambda_{k}g_{k}^Td_{k},$$ where $f_{\max}$ is the maximal function value in recent $M$ iterations, i.e., $$f_{\max}=\max_{0\leq i\leq \min\{k,M-1\}} f(x_{k-i}).$$ This nonmonotone line search is observed specially suitable for BB-type methods [@dai2001adaptive]. Moreover, under standard assumptions, Algorithm \[al1\] ensures convergence in the sense that $\lim\inf_{k\rightarrow\infty}\|\bar{g}_{k}\|=0$, see [@hager2006new]. Numerical results {#secnum} ================= In this section, we present numerical comparisons of ANGM, ANGR1, ANGR2 with some recent very successful gradient descent methods on solving quadratic, general unconstrained and bound constrained problems. All the comparison methods were implemented in Matlab (v.9.0-R2016a) and run on a laptop with an Intel Core i7, 2.9 GHz processor and 8 GB of RAM running Windows 10 system. Quadratic problems ------------------ In this subsection, we compare ANGM, ANGR1 and ANGR2 with the BB1 [@barzilai1988two], DY [@dai2005analysis], ABBmin2 [@frassoldati2008new], and SDC [@de2014efficient] methods on solving quadratic optimization problems. We first solve some randomly generated quadratic problems from [@yuan2006new]. Particularly, we solve $$\label{quad-test1} \min_{x \in \mathbb{R}^n} \: f(x)=(x-x^*)^TV(x-x^*),$$ where $x^*$ is randomly generated with components between $-10$ and $10$, $V=\textrm{diag}\{v_1,\ldots,v_n\}$ is a diagonal matrix with $v_1=1$ and $v_n=\kappa$, and $v_j$, $j=2,\ldots,n-1$, are generated by the *rand* function between 1 and $\kappa$. -- ----------------------------------------------------------------- $\{v_2,\ldots,v_{n-1}\}\subset(1,\kappa)$ $\{v_2,\ldots,v_{n/5}\}\subset(1,100)$ $\{v_{n/5+1},\ldots,v_{n-1}\}\subset(\frac{\kappa}{2},\kappa)$ $\{v_2,\ldots,v_{n/2}\}\subset(1,100)$ $\{v_{n/2+1},\ldots,v_{n-1}\}\subset(\frac{\kappa}{2},\kappa)$ $\{v_2,\ldots,v_{4n/5}\}\subset(1,100)$ $\{v_{4n/5+1},\ldots,v_{n-1}\}\subset(\frac{\kappa}{2},\kappa)$ $\{v_2,\ldots,v_{n/5}\}\subset(1,100)$ $\{v_{n/5+1},\ldots,v_{4n/5}\}\subset(100,\frac{\kappa}{2})$ $\{v_{4n/5+1},\ldots,v_{n-1}\}\subset(\frac{\kappa}{2},\kappa)$ -- ----------------------------------------------------------------- : Distributions of $v_j$[]{data-label="tbspe"} We have tested five sets of problems with $n=1,000$ using different spectral distributions of the Hessian listed in Table \[tbspe\]. The algorithm is stopped once the number of iteration exceeds 20,000 or the gradient norm is reduced by a factor of $\epsilon$, which is set to $10^{-6}, 10^{-9}$ and $10^{-12}$, respectively. Three different condition numbers $\kappa=10^4, 10^5$ and $10^6$ are tested. For each value of $\kappa$ or $\epsilon$, 10 instances of the problem are randomly generated and the averaged results obtained by the starting point $x_0=(0,\ldots,0)^T$ are presented. For the ABBmin2 method, $\tau$ is set to 0.9 as suggested in [@frassoldati2008new]. The parameter pair $(h,s)$ of the SDC method is set to $(8,6)$ which is more efficient than other choices for this test. The $q_k$ is calculated by for our methods. ![Performance profiles of compared methods on solving random quadratic problems with spectral distributions in Table \[tbspe\], iteration metric[]{data-label="frandp"}](randp_ang_ng_all_ar.eps){width="60.00000%" height="45.00000%"} -- ------- ------------ ----------- -------- -------- --------- --------- --------- --------- --------- --------- --------- ----------- --------- 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1e-06 200.6 198.6 213.3 193.6 204.1 196.5 211.1 219.0 230.3 221.0 197.1 **177.1** 183.2 1e-09 665.0 678.1 677.5 937.3 718.7 768.5 1259.4 1198.1 1195.3 2590.0 2672.7 **428.8** 2029.6 1e-12 939.7 941.6 901.5 1159.2 951.2 1043.7 1464.3 1339.0 1411.9 6032.9 6353.9 **560.3** 4087.2 1e-06 140.3 **140.2** 143.1 161.2 172.8 171.1 217.3 265.8 310.3 311.2 261.5 302.5 160.6 1e-09 **546.1** 590.5 641.1 779.5 850.5 891.9 1055.7 1161.3 1299.0 1665.4 1340.0 1321.1 735.1 1e-12 **895.1** 1025.2 1098.1 1328.6 1446.6 1598.2 1739.6 1908.7 2107.0 2820.8 2434.6 2267.5 1346.8 1e-06 **163.2** 170.1 177.3 193.3 216.7 231.1 283.3 318.4 363.9 388.4 329.4 356.3 235.5 1e-09 **566.5** 640.0 680.7 811.0 970.8 986.0 1188.3 1218.7 1364.4 1783.1 1511.8 1470.8 818.1 1e-12 **928.4** 1030.0 1163.7 1412.0 1575.2 1733.2 1973.6 2075.2 2191.1 2977.7 2780.2 2288.4 1310.6 1e-06 **212.3** 213.7 237.1 259.0 254.7 291.8 365.3 431.4 475.1 500.5 431.3 519.0 262.8 1e-09 **616.1** 655.7 759.7 885.5 956.4 1107.9 1232.4 1405.3 1533.3 1859.4 1659.9 1489.5 805.5 1e-12 **996.0** 1078.4 1250.9 1452.1 1629.5 1786.8 2041.6 2179.0 2427.6 3051.5 2785.4 2383.9 1469.3 1e-06 **623.1** 654.8 663.4 671.0 683.4 697.2 761.3 813.7 931.0 832.5 650.8 816.0 668.3 1e-09 **2603.0** 2654.7 2847.6 2901.4 2936.9 3161.7 3228.6 3306.4 3807.2 4497.2 3185.5 2929.7 3274.5 1e-12 **4622.7** 4675.2 4905.9 4634.2 4818.7 4944.7 5224.0 5480.7 5972.1 7446.7 7024.1 4808.7 5816.6 1e-06 **1339.5** 1377.5 1434.2 1478.1 1531.7 1587.7 1838.3 2048.4 2310.6 2253.7 1870.1 2170.9 1510.4 1e-09 **4996.7** 5219.0 5606.6 6314.7 6433.3 6916.0 7964.4 8289.7 9199.2 12395.0 10370.0 7640.0 7662.7 1e-12 **8382.0** 8750.3 9320.1 9986.1 10421.3 11106.5 12443.0 12982.6 14109.7 22329.5 21378.3 12308.9 14030.4 -- ------- ------------ ----------- -------- -------- --------- --------- --------- --------- --------- --------- --------- ----------- --------- : The numbers of averaged iterations of the ANGM, BB1, DY, ABBmin2 and SDC methods on solving quadratic problems with spectral distributions in Table \[tbspe\][]{data-label="tbANGM"} -- ------- ------------ ------------ ----------- ----------- ----------- -------- -------- -------- -------- --------- --------- ----------- --------- 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1e-06 198.7 185.8 188.1 182.5 189.2 188.2 184.5 200.1 214.1 221.0 197.1 **177.1** 183.2 1e-09 655.5 641.6 625.4 667.9 621.7 680.9 762.7 811.5 1012.2 2590.0 2672.7 **428.8** 2029.6 1e-12 897.3 854.7 844.5 838.4 775.2 813.1 917.4 984.7 1172.7 6032.9 6353.9 **560.3** 4087.2 1e-06 137.9 119.0 119.2 118.7 **115.3** 124.5 152.4 157.5 178.9 311.2 261.5 302.5 160.6 1e-09 **466.5** 470.9 487.2 512.0 540.2 589.9 672.6 691.0 722.3 1665.4 1340.0 1321.1 735.1 1e-12 788.1 **783.3** 802.1 835.3 908.1 1004.5 1074.2 1110.3 1215.2 2820.8 2434.6 2267.5 1346.8 1e-06 164.5 149.6 **146.3** 153.0 157.8 166.8 183.3 208.9 229.1 388.4 329.4 356.3 235.5 1e-09 507.5 **473.0** 516.6 541.2 563.3 609.0 682.5 747.2 850.5 1783.1 1511.8 1470.8 818.1 1e-12 818.3 **801.5** 865.4 894.4 965.4 1046.0 1108.8 1200.2 1419.8 2977.7 2780.2 2288.4 1310.6 1e-06 189.7 **172.2** 168.8 179.6 195.2 212.7 229.4 262.9 282.1 500.5 431.3 519.0 262.8 1e-09 539.0 **519.2** 558.7 570.2 617.1 689.6 786.3 853.1 901.8 1859.4 1659.9 1489.5 805.5 1e-12 **849.0** 851.1 894.2 934.8 1011.8 1065.5 1197.3 1291.8 1408.0 3051.5 2785.4 2383.9 1469.3 1e-06 625.7 587.6 600.1 **583.1** 614.2 632.2 675.8 726.9 769.0 832.5 650.8 816.0 668.3 1e-09 2539.9 **2518.1** 2612.7 2530.4 2636.0 2581.2 2773.6 2875.2 3026.2 4497.2 3185.5 2929.7 3274.5 1e-12 **4207.3** 4238.2 4292.4 4256.2 4240.7 4285.1 4397.2 4534.8 4729.0 7446.7 7024.1 4808.7 5816.6 1e-06 1316.5 **1214.1** 1222.6 1217.0 1271.8 1324.5 1425.3 1556.4 1673.1 2253.7 1870.1 2170.9 1510.4 1e-09 4708.5 **4622.7** 4800.7 4821.7 4978.4 5150.6 5677.7 5977.9 6512.9 12395.0 10370.0 7640.0 7662.7 1e-12 7560.1 **7528.8** 7698.7 7759.1 7901.3 8214.1 8694.9 9121.9 9944.7 22329.5 21378.3 12308.9 14030.4 -- ------- ------------ ------------ ----------- ----------- ----------- -------- -------- -------- -------- --------- --------- ----------- --------- : The numbers of averaged iterations of the ANGR1, BB1, DY, ABBmin2 and SDC methods on problems in Table \[tbspe\][]{data-label="tbANGR1"} We compared the algorithms by using the performance profiles of Dolan and Moré [@dolan2002] on iteration metric. In these performance profiles, the vertical axis shows the percentage of the problems the method solves within the factor $\rho$ of the metric used by the most effective method in this comparison. Fig. \[frandp\] shows the performance profiles of ANGM, ANGR1 and ANGR2 obtained by setting $\tau_1=0.1$, $\tau_2=1$ and other four compared methods. Clearly, ANGR2 outperforms all other methods. In general, we can see ANGM, ANGR1 and ANGR2 are much better than the BB1, DY and SDC methods. To further analyze the performance of our methods, we present results of them with different values of $\tau_1$ from 0.1 to 0.9 in Tables \[tbANGM\], \[tbANGR1\] and \[tbANGR2\]. From Table \[tbANGM\], we can see that, for the first problem set, ANGM is much faster than the BB1, DY and SDC methods, and competitive with the ABBmin2 method. As for the other four problem sets, ANGM method with a small value of $\tau_1$ outperforms the other compared methods though its performance seems to become worse as $\tau_1$ increases. The results shown in Tables \[tbANGR1\] and \[tbANGR2\] are slightly different from those in Table \[tbANGM\]. In particular, ANGR1 and ANGR2 outperform other four compared methods for most of the problem sets and values of $\tau_1$. For each given $\tau_1$ and tolerance level, ANGR1 and ANGR2 always perform better than other methods in terms of total number of iterations. -- ------- ----------- ------------ ------------ ----------- -------- -------- ----------- -------- -------- --------- --------- ----------- --------- 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1e-06 196.9 191.8 184.8 186.8 183.3 189.0 **175.6** 182.2 189.5 221.0 197.1 177.1 183.2 1e-09 591.0 607.8 599.4 613.3 669.5 699.6 704.0 763.6 1005.1 2590.0 2672.7 **428.8** 2029.6 1e-12 870.4 811.3 763.1 790.7 800.0 881.4 844.8 971.0 1146.5 6032.9 6353.9 **560.3** 4087.2 1e-06 129.7 117.6 115.0 **111.9** 114.6 116.3 140.6 146.9 164.4 311.2 261.5 302.5 160.6 1e-09 **455.4** 474.8 470.8 491.4 530.6 566.0 595.9 601.8 691.5 1665.4 1340.0 1321.1 735.1 1e-12 786.0 **750.5** 811.7 819.3 938.1 908.4 1000.3 1008.5 1089.0 2820.8 2434.6 2267.5 1346.8 1e-06 156.2 141.2 **136.3** 149.7 146.7 167.9 179.9 188.0 206.8 388.4 329.4 356.3 235.5 1e-09 495.8 **465.6** 490.4 529.0 530.8 606.4 667.6 718.5 726.6 1783.1 1511.8 1470.8 818.1 1e-12 809.0 **775.1** 796.4 876.6 918.8 997.0 1080.8 1130.1 1157.8 2977.7 2780.2 2288.4 1310.6 1e-06 183.2 168.0 **164.9** 166.9 180.9 195.7 219.8 234.2 225.3 500.5 431.3 519.0 262.8 1e-09 **516.2** 521.9 523.2 541.6 603.9 637.3 700.0 709.8 739.2 1859.4 1659.9 1489.5 805.5 1e-12 842.6 **830.1** 845.8 899.5 959.4 1039.2 1099.3 1133.9 1192.2 3051.5 2785.4 2383.9 1469.3 1e-06 611.7 **580.5** 605.4 594.8 586.2 605.1 659.3 644.1 653.1 832.5 650.8 816.0 668.3 1e-09 2472.7 **2394.4** 2510.8 2504.4 2447.2 2502.1 2551.9 2648.6 2763.9 4497.2 3185.5 2929.7 3274.5 1e-12 4122.4 4108.7 **4103.2** 4107.2 4161.8 4227.9 4363.6 4324.0 4483.7 7446.7 7024.1 4808.7 5816.6 1e-06 1277.8 **1199.1** 1206.3 1210.1 1211.6 1274.0 1375.2 1395.4 1439.1 2253.7 1870.1 2170.9 1510.4 1e-09 4531.1 **4464.5** 4594.7 4679.7 4781.9 5011.5 5219.4 5442.3 5926.3 12395.0 10370.0 7640.0 7662.7 1e-12 7430.3 **7275.7** 7320.2 7493.4 7778.1 8053.8 8388.8 8567.6 9069.2 22329.5 21378.3 12308.9 14030.4 -- ------- ----------- ------------ ------------ ----------- -------- -------- ----------- -------- -------- --------- --------- ----------- --------- : The numbers of averaged iterations of the ANGR2, BB1, DY, ABBmin2 and SDC methods on problems in Table \[tbspe\][]{data-label="tbANGR2"} Secondly, we compared the methods on solving the non-rand quadratic problem with $n=10,000$. For ANGM, ANGR1 and ANGR2, $\tau_1$ and $\tau_2$ were set to 0.4 and 1, respectively. The parameter pair $(h,s)$ used for the SDC method was set to $(30,2)$. Other settings are the same as above. Table \[tbqp\] presents averaged number of iterations over 10 different starting points with entries randomly generated in $[-10,10]$. It can be seen that ANGM, ANGR1 and ANGR2 are significantly better than the BB1 and DY methods. In addition, ANGR1 and ANGR2 often outperform the ABBmin2 and SDC methods while ANGM is very competitive with them. -- ------- --------- --------- --------- --------- --------- ------------ ------------- 1e-06 626.8 527.7 513.0 597.1 539.1 **500.6** 512.1 1e-09 1267.0 972.1 894.5 1000.6 976.7 893.7 **890.2** 1e-12 1741.9 1396.8 1277.8 1409.2 1399.1 1298.0 **1257.4** 1e-06 1597.8 1326.7 1266.3 1254.3 1209.7 **1046.0** 1127.9 1e-09 3687.5 3168.3 2559.8 2647.4 2605.2 2424.3 **2399.8** 1e-12 5564.8 4892.4 3895.0 4156.4 4139.0 3858.5 **3663.3** 1e-06 4060.9 2159.4 3130.2 1986.2 2112.9 1992.0 **1936.0** 1e-09 10720.4 10134.3 7560.8 7178.5 7381.1 **6495.1** 6550.1 1e-12 17805.5 18015.6 12193.6 11646.7 11922.9 10364.9 **10280.2** 1e-06 6285.5 4013.8 4909.5 3837.6 3861.7 **3538.6** 3576.0 1e-09 15674.9 14274.7 11015.1 10826.5 10963.0 **9813.1** 9840.1 1e-12 25112.2 24304.8 17366.4 17212.3 17461.0 15521.4 **15200.9** -- ------- --------- --------- --------- --------- --------- ------------ ------------- : The numbers of iterations of the compared methods on problem with $n=10,000$[]{data-label="tbqp"} Finally, we compared the methods on solving two large-scale real problems Laplace1(a) and Laplace1(b) described in [@fletcher2005barzilai]. The two problems require the solution of a system of linear equations derived from a 3D Laplacian on a box, discretized using a standard 7-point finite difference stencil. Each problem has $n=N^3$ variables with $N$ being the number of interior nodes taken in each coordinate direction. The solution is fixed by a Gaussian function centered at $(\alpha,\beta,\gamma)$ and multiplied by $x(x-1)y(y-1)z(z-1)$. The parameter $\sigma$ is used to control the rate of decay of the Gaussian. See [@fletcher2005barzilai] for more details on these problems. Here, we set the parameters as follows: $$\begin{aligned} &(a)~~ \sigma=20,~ \alpha=\beta=\gamma=0.5;\\ &(b)~~ \sigma=50,~ \alpha=0.4,~ \beta=0.7,~ \gamma=0.5.\end{aligned}$$ The null vector was used as the starting point. We again stop the iteration when $\|g_k\|\leq\epsilon\|g_0\|$ with different values of $\epsilon$. For ANGM, ANGR1 and ANGR2, $\tau_1$ and $\tau_2$ were set to 0.7 and 1.2, respectively. The parameter pair $(h,s)$ used for the SDC method was chosen for the best performance in our test, i.e., $(2,6)$ and $(8,6)$ for Laplace1(a) and Laplace1(b), respectively. Other settings are the same as above. The number of iterations required by the compared methods for solving the two problems are listed in Table \[tbnLap\]. It can be seen that our methods are significantly better than the BB1, DY and SDC methods and is often faster than ABBmin2 especially when a tight tolerance is used. [|c|c|c|c|c|c|c|c|c|]{} & & & & & & & &\ \ & 1e-06 & 259 & 249 & **192** & 213 & 245 & 195 & 233\ & 1e-09 & 441 & 373 & 329 & 393 & 313 & 322 & **308**\ & 1e-12 & 680 & 546 & 401 & 529 & 367 & 373 & **364**\ & 1e-06 & 359 & 383 & 289 & 297 & 291 & 332 & **288**\ & 1e-09 & 591 & 570 & 430 & 553 & 408 & 446 & **396**\ & 1e-12 & 882 & 789 & 608 & 705 & 620 & **516** & 591\ & 1e-06 & 950 & 427 & 351 & 513 & 450 & **303** & 416\ & 1e-09 & 1088 & 651 & 485 & 609 & 584 & 519 & **503**\ & 1e-12 & 1241 & 918 & 687 & 825 & 694 & 604 & **597**\ & 1e-06 & 1568 & 1059 & 832 & 1023 & 986 & **830** & 937\ & 1e-09 & 2120 & 1594 & 1244 & 1555 & 1305 & 1287 & **1207**\ & 1e-12 & 2803 & 2253 & 1696 & 2059 & 1681 & **1493** & 1552\ \ & 1e-06 & 246 & 236 & 217 & **213** & 242 & 217 & 214\ & 1e-09 & 473 & 399 & 365 & 437 & **333** & 338 & 409\ & 1e-12 & 651 & 532 & 502 & 555 & **451** & 478 & 573\ & 1e-06 & 288 & 454 & 294 & 309 & 296 & **290** & 324\ & 1e-09 & 607 & 567 & **433** & 485 & 517 & 499 & 495\ & 1e-12 & 739 & 794 & 634 & 766 & 686 & **590** & 645\ & 1e-06 & 544 & 371 & 369 & 379 & 381 & 406 & **358**\ & 1e-09 & 646 & 700 & 585 & 653 & 638 & **558** & 648\ & 1e-12 & 937 & 1038 & 880 & 965 & 854 & **785** & 810\ & 1e-06 & 1078 & 1061 & **880** & 901 & 919 & 913 & 896\ & 1e-09 & 1726 & 1666 & **1383** & 1575 & 1488 & 1395 & 1552\ & 1e-12 & 2327 & 2364 & 2016 & 2286 & 1991 & **1853** & 2028\ Unconstrained problems ---------------------- Now we present comparisons of ANGR1 and ANGR2 with SPG method[^1] in [@birgin2000nonmonotone; @birgin2014spectral], and the BB1 method using Dai-Zhang nonmonotone line search [@dai2001adaptive] (BB1-DZ) on solving general unconstrained problems. For our methods, the parameter values are set as the following: $$\alpha_{\min}=10^{-30},~\alpha_{\max}=10^{30},~M=8,~\sigma=10^{-4},~\alpha_0=1/\|g_0\|_\infty.$$ In addition, the parameter $\tau_1$ and $\tau_2$ for ANGR1 and ANGR2 are set to 0.8 and 1.2, respectively. Default parameters were used for SPG. Each method was stopped if the number of iteration exceeds 200,000 or $\|g_k\|_\infty\leq10^{-6}$. Our test problems were taken from [@andrei2008unconstrained]. We have tested 59 problems listed in Table \[tbunpro\] with $n=1,000$ and the performance profiles are shown in Fig. \[Andrei\], which shows that ANGR1 and ANGR2 outperform SPG and BB1-DZ in terms of the iteration number, and ANGR2 is faster than ANGR1. Moreover, BB1-DZ is slightly better than SPG. Detail numerical results are also presented in Table \[tbunc\]. Since the only difference between BB1-DZ with ANGR1 and ANGR2 lies in the choice of stepsizes, these numerical results show our adaptive choices of stepsizes in ANGR1 and ANGR2 are very effective and can indeed greatly accelerate the convergence of BB-type methods. Problem name Problem name --------- --------------------------------------- --------- --------------------------------- 1 Extended Freudenstein & Roth function 31 NONDIA 2 Extended Trigonometric 32 DQDRTIC 3 Extended White & Holst 33 Partial Perturbed Quadratic 4 Extended Beale 34 Broyden Tridiagonal 5 Extended Penalty 35 Almost Perturbed Quadratic 6 Perturbed Quadratic 36 Perturbed Tridiagonal Quadratic 7 Raydan 1 37 Staircase 1 8 Raydan 2 38 Staircase 2 9 Diagonal 1 39 LIARWHD 10 Diagonal 2 40 POWER 11 Diagonal 3 41 ENGVAL1 12 Hager 42 EDENSCH 13 Generalized Tridiagonal 1 43 BDEXP 14 Extended Tridiagonal 1 44 GENHUMPS 15 Extended TET 45 NONSCOMP 16 Generalized Tridiagonal 2 46 VARDIM 17 Diagonal 4 47 QUARTC 18 Diagonal 5 48 Extended DENSCHNB 19 Extended Himmelblau 49 Extended DENSCHNF 20 Extended PSC1 50 LIARWHD 21 Generalized PSC1 51 BIGGSB1 22 Extended Powell 52 Generalized Quartic 23 Extended Cliff 53 Diagonal 7 24 Perturbed quadratic diagonal 54 Diagonal 8 25 Quadratic QF1 55 Full Hessian FH3 26 Extended quadratic exponential EP1 56 SINCOS 27 Extended Tridiagonal 2 57 Diagonal 9 28 BDQRTIC 58 HIMMELBG 29 TRIDIA 59 HIMMELH 30 ARWHEAD : Test problems from [@andrei2008unconstrained][]{data-label="tbunpro"} ![Performance profiles of compared methods on 59 unconstrained problems in Table \[tbunpro\], iteration metric[]{data-label="Andrei"}](Andrei_ng_ar.eps){width="55.00000%" height="40.00000%"} ---- ------- ----------- ------------------ ------- ----------- ------------------ ------- ----------- ---------- ------- ----------- ---------- -- -- -- -- iter $f_k$ $\|g_k\|_\infty$ iter $f_k$ $\|g_k\|_\infty$ iter $f_k$ iter $f_k$ 1 87 2.45e+04 8.95e-07 41 2.45e+04 4.06e-08 26 2.45e+04 1.46e-07 25 2.45e+04 1.46e-07 2 45 5.72e-13 3.83e-07 87 5.26e-07 3.97e-07 100 4.89e-07 9.32e-07 100 4.73e-07 2.20e-07 3 110 5.80e-14 3.05e-08 120 6.30e-20 4.95e-11 91 9.95e-12 2.93e-07 88 1.05e-11 2.90e-07 4 46 3.92e-10 6.57e-07 65 8.28e-18 3.74e-10 31 9.52e-15 4.15e-08 31 1.94e-12 6.12e-08 5 41 8.83e+02 9.02e-07 107 8.83e+02 1.24e-08 107 8.83e+02 1.24e-08 107 8.83e+02 1.24e-08 6 597 2.39e-13 9.83e-07 457 2.16e-13 9.24e-07 300 7.98e-16 5.66e-08 287 1.69e-13 8.03e-07 7 465 5.01e+04 9.64e-07 339 5.01e+04 9.04e-07 330 5.01e+04 5.56e-08 301 5.01e+04 8.39e-07 8 1 1.00e+03 0.00e+00 1 1.00e+03 0.00e+00 1 1.00e+03 0.00e+00 1 1.00e+03 0.00e+00 9 415 -2.71e+06 8.50e-07 361 -2.71e+06 5.43e-07 294 -2.71e+06 9.05e-07 276 -2.71e+06 7.04e-07 10 168 3.13e+01 6.05e-07 173 3.13e+01 5.68e-07 135 3.13e+01 5.93e-07 131 3.13e+01 8.61e-07 11 668 -4.96e+05 8.82e-07 404 -4.96e+05 8.63e-07 336 -4.96e+05 8.85e-07 286 -4.96e+05 9.93e-07 12 63 -4.47e+04 2.91e-07 58 -4.47e+04 5.70e-07 52 -4.47e+04 5.14e-07 52 -4.47e+04 5.59e-07 13 29 9.97e+02 6.53e-07 27 9.97e+02 9.67e-07 26 9.97e+02 4.21e-07 25 9.97e+02 9.79e-07 14 30 2.13e-07 3.75e-07 18 2.68e-07 4.46e-07 20 5.20e-07 7.33e-07 20 5.20e-07 7.33e-07 15 10 1.28e+03 4.34e-10 5 1.28e+03 1.33e-08 5 1.28e+03 1.33e-08 5 1.28e+03 1.33e-08 16 69 2.38e+00 8.58e-07 58 9.58e-01 7.14e-07 56 9.58e-01 8.40e-07 57 9.58e-01 6.13e-07 17 3 0.00e+00 0.00e+00 3 0.00e+00 0.00e+00 3 0.00e+00 0.00e+00 3 0.00e+00 0.00e+00 18 4 6.93e+02 4.81e-08 1 6.93e+02 0.00e+00 1 6.93e+02 0.00e+00 1 6.93e+02 0.00e+00 19 15 4.21e-19 1.92e-10 15 4.21e-19 1.92e-10 15 4.21e-19 1.92e-10 15 4.21e-19 1.92e-10 20 23 9.99e+02 8.27e-07 22 9.99e+02 8.65e-07 22 9.99e+02 8.65e-07 22 9.99e+02 8.65e-07 21 15 3.87e+02 5.87e-07 13 3.87e+02 3.78e-07 13 3.87e+02 3.78e-07 13 3.87e+02 3.78e-07 22 246 3.19e-07 6.80e-07 266 3.67e-07 5.97e-07 164 7.08e-07 9.62e-07 205 5.55e-07 9.82e-07 23 154 9.99e+01 7.79e-07 126 9.99e+01 1.57e-07 541 9.99e+01 2.61e-07 539 9.99e+01 6.10e-08 24 373 2.41e-11 8.95e-07 246 1.41e-11 6.83e-07 249 2.05e-12 2.98e-07 314 4.05e-12 4.14e-07 25 612 -5.00e-04 9.45e-07 575 -5.00e-04 9.95e-07 319 -5.00e-04 8.92e-08 282 -5.00e-04 7.45e-07 26 4 7.93e+03 4.19e-08 4 7.93e+03 3.71e-09 4 7.93e+03 3.71e-09 4 7.93e+03 3.71e-09 27 34 3.89e+02 7.50e-07 39 3.89e+02 9.74e-07 36 3.89e+02 8.49e-07 36 3.89e+02 5.15e-07 28 77 3.98e+03 7.64e-07 98 3.98e+03 8.91e-07 69 3.98e+03 1.49e-07 64 3.98e+03 9.57e-07 29 3910 2.47e-13 6.97e-07 2403 9.93e-15 6.71e-07 816 9.62e-16 4.67e-07 611 6.71e-14 5.86e-07 30 4 0.00e+00 1.50e-09 4 0.00e+00 1.50e-09 4 0.00e+00 1.50e-09 4 0.00e+00 1.50e-09 31 15 1.35e-14 9.86e-09 15 1.35e-14 9.86e-09 16 1.10e-12 8.20e-08 16 1.10e-12 8.20e-08 32 41 4.20e-16 2.86e-07 26 1.80e-13 8.47e-07 17 2.99e-16 2.14e-07 17 2.99e-16 2.14e-07 33 312 2.51e-14 5.96e-07 229 1.42e-13 9.63e-07 257 6.00e-15 2.16e-07 294 7.03e-14 8.24e-07 34 47 5.31e-15 3.20e-07 51 7.94e-14 8.28e-07 41 7.67e-14 7.35e-07 47 2.62e-14 7.66e-07 35 621 2.36e-13 9.76e-07 363 2.47e-13 9.97e-07 358 1.53e-13 7.77e-07 272 4.53e-15 6.22e-07 36 595 2.49e-13 9.97e-07 606 2.48e-13 9.95e-07 368 3.46e-13 1.00e-06 306 1.82e-13 8.61e-07 37 15936 1.81e-12 7.80e-07 10239 1.54e-12 7.68e-07 7234 1.39e-13 1.89e-07 6543 1.07e-12 9.13e-07 38 28528 1.97e-14 2.43e-07 25640 2.74e-12 9.97e-07 60283 2.86e-12 9.93e-07 27783 2.98e-12 9.74e-07 39 61 2.64e-14 2.06e-08 59 2.62e-13 4.93e-07 47 9.15e-14 6.15e-08 47 1.30e-12 2.36e-07 40 5435 2.84e-13 9.98e-07 5472 2.94e-13 9.98e-07 327 1.89e-13 6.99e-07 366 5.63e-13 9.98e-07 41 33 1.11e+03 8.58e-07 31 1.11e+03 8.09e-07 29 1.11e+03 1.70e-07 29 1.11e+03 1.70e-07 42 32 6.00e+03 7.33e-07 32 6.00e+03 7.33e-07 31 6.00e+03 4.46e-07 31 6.00e+03 9.35e-07 43 15 6.52e-05 9.93e-07 15 6.52e-05 9.93e-07 15 6.52e-05 9.93e-07 15 6.52e-05 9.93e-07 44 747 1.39e-17 1.50e-09 3 4.49e-23 3.00e-12 3 4.49e-23 3.00e-12 3 4.49e-23 3.00e-12 45 48 9.98e-12 8.75e-07 89 1.97e-14 2.57e-07 59 2.30e-13 5.85e-07 63 2.18e-13 7.89e-07 46 1 4.62e-30 4.44e-16 30 2.17e-27 5.77e-15 30 2.17e-27 5.77e-15 30 2.17e-27 5.77e-15 47 1 0.00e+00 0.00e+00 1 0.00e+00 0.00e+00 1 0.00e+00 0.00e+00 1 0.00e+00 0.00e+00 48 8 1.37e-11 4.68e-07 8 1.37e-11 4.68e-07 8 1.37e-11 4.68e-07 8 1.37e-11 4.68e-07 49 12 1.78e-19 4.76e-10 17 6.46e-21 5.91e-11 17 6.46e-21 5.91e-11 17 6.46e-21 5.91e-11 50 61 2.64e-14 2.06e-08 59 2.62e-13 4.93e-07 47 9.15e-14 6.15e-08 47 1.30e-12 2.36e-07 51 30459 1.24e-05 9.95e-07 14877 3.73e-07 6.46e-07 7558 6.51e-07 8.62e-07 8969 6.08e-12 6.45e-07 52 9 4.98e-15 1.99e-07 9 4.98e-15 1.99e-07 9 4.98e-15 1.99e-07 9 4.98e-15 1.99e-07 53 7 -8.17e+02 1.97e-09 7 -8.17e+02 1.97e-09 7 -8.17e+02 1.97e-09 7 -8.17e+02 1.97e-09 54 6 -4.80e+02 5.85e-08 6 -4.80e+02 4.36e-10 6 -4.80e+02 4.36e-10 6 -4.80e+02 4.36e-10 55 3 -2.50e-01 4.56e-10 3 -2.50e-01 4.56e-10 3 -2.50e-01 4.56e-10 3 -2.50e-01 4.56e-10 56 15 3.87e+02 5.87e-07 13 3.87e+02 3.78e-07 13 3.87e+02 3.78e-07 13 3.87e+02 3.78e-07 57 553 -2.70e+06 1.91e-07 762 -2.70e+06 6.06e-07 305 -2.70e+06 9.96e-07 346 -2.70e+06 5.60e-07 58 16 4.79e-04 8.77e-07 17 3.78e-04 6.91e-07 17 3.78e-04 6.91e-07 17 3.78e-04 6.91e-07 59 12 -5.00e+02 1.09e-07 12 -5.00e+02 7.41e-11 12 -5.00e+02 7.41e-11 12 -5.00e+02 7.41e-11 ---- ------- ----------- ------------------ ------- ----------- ------------------ ------- ----------- ---------- ------- ----------- ---------- -- -- -- -- : Results of compared methods on unconstrained problems in Table \[tbunpro\][]{data-label="tbunc"} Bound constrained problems -------------------------- We further compare ANGR1 and ANGR2, with SPG [@birgin2000nonmonotone; @birgin2014spectral] and the BB1-DZ method [@dai2001adaptive] combined with gradient projection techniques on solving bound constrained problems from the CUTEst collection [@gould2015cutest] with dimension more than $50$. We deleted $3$ problems from this list since none of these comparison algorithms can solve them. Hence, in total there are $47$ problems left in our test. The iteration was stopped if the number of iteration exceeds 200,000 or $\|P_{\Omega}(x_k-g_k)-x_k\|_{\infty}\leq10^{-6}$. The parameters $\tau_1$ and $\tau_2$ for ANGR1 and ANGR2 are set to 0.4 and 1.5, respectively. Other settings are the same as before. Fig. \[cutest\] shows the performance profiles of all the compared methods on iteration metric. Similar as the unconstrained case, from Fig. \[cutest\], we again see that both ANGR1 and ANGR2 perform significantly better than SPG and BB1-DZ. Hence, our new gradient methods also have potential great benefits for solving constrained optimization. ![Performance profiles of compared methods on solving 47 bound constrained problems from CUTEst, iteration metric[]{data-label="cutest"}](cutest_angr_ng_ar.eps "fig:"){width="55.00000%" height="40.00000%"}\ Conclusions {#secclu} =========== We have developed techniques to accelerate the Barzilai-Borwein (BB) method motivated from finite termination for minimizing two-dimensional strongly convex quadratic functions. More particularly, by exploiting certain orthogonal properties of the gradients, we derive a new monotone stepsize that can be combined with BB stepsizes to significantly improve their performance for minimizing general strongly convex quadratic functions. By adaptively using this new stepsize and the two BB stepsizes, we develop a new gradient method called ANGM and its two variants ANGR1 and ANGR2, which are further extended for solving unconstrained and bound constrained optimization. Our extensive numerical experiments show that all the new developed methods are significantly better than the BB method and are faster than some very successful gradient methods developed in the recent literature for solving quadratic, general smooth unconstrained and bound constrained optimization. This work was supported by the National Natural Science Foundation of China (Grant Nos. 11631013, 11701137, 11671116), the National 973 Program of China (Grant No. 2015CB856002), China Scholarship Council (No. 201806705007), and USA National Science Foundation (1522654, 1819161). [10]{} \[1\][[\#1]{}]{} urlstyle \[1\][DOI \#1]{} Akaike, H.: On a successive transformation of probability distribution and its application to the analysis of the optimum gradient method. Ann. Inst. Stat. Math. **11**(1), 1–16 (1959) Andrei, N.: An unconstrained optimization test functions collection. Adv. Model. Optim. **10**(1), 147–161 (2008) Barzilai, J., Borwein, J.M.: Two-point step size gradient methods. IMA J. Numer. Anal. **8**(1), 141–148 (1988) Birgin, E.G., Mart[í]{}nez, J.M., Raydan, M.: Nonmonotone spectral projected gradient methods on convex sets. SIAM J. Optim. **10**(4), 1196–1211 (2000) Birgin, E.G., Mart[í]{}nez, J.M., Raydan, M.: Spectral projected gradient methods: review and perspectives. J. Stat. Softw. **60**(3), 539–559 (2014) Cauchy, A.: M[é]{}thode g[é]{}n[é]{}rale pour la r[é]{}solution des systemes d¡¯[é]{}quations simultan[é]{}es. Comp. Rend. Sci. Paris **25**, 536–538 (1847) Dai, Y.H.: Alternate step gradient method. Optimization **52**(4-5), 395–415 (2003) Dai, Y.H., Fletcher, R.: Projected Barzilai–Borwein methods for large-scale box-constrained quadratic programming. Numer. Math. **100**(1), 21–47 (2005) Dai, Y.H., Huang, Y., Liu, X.W.: A family of spectral gradient methods for optimization. Comp. Optim. Appl. **74**(1), 43–65 (2019) Dai, Y.H., Liao, L.Z.: $R$-linear convergence of the Barzilai and Borwein gradient method. IMA J. Numer. Anal. **22**(1), 1–10 (2002) Dai, Y.H., Yuan, Y.X.: Alternate minimization gradient method. IMA J. Numer. Anal. **23**(3), 377–393 (2003) Dai, Y.H., Yuan, Y.X.: Analysis of monotone gradient methods. J. Ind. Manag. Optim. **1**(2), 181 (2005) Dai, Y.H., Zhang, H.: Adaptive two-point stepsize gradient algorithm. Numer. Algorithms **27**(4), 377–385 (2001) De Asmundis, R., Di Serafino, D., Hager, W.W., Toraldo, G., Zhang, H.: An efficient gradient method using the yuan steplength. Comp. Optim. Appl. **59**(3), 541–563 (2014) Di Serafino, D., Ruggiero, V., Toraldo, G., Zanni, L.: On the steplength selection in gradient methods for unconstrained optimization. Appl. Math. Comput. **318**, 176–195 (2018) Dolan, E.D., Mor[é]{}, J.J.: Benchmarking optimization software with performance profiles. Math. Program. **91**(2), 201–213 (2002) Fletcher, R.: On the Barzilai–Borwein method. In: Qi, L., Teo, K., Yang, X. (eds.) Optimization and Control with Applications. pp. 235–256. Springer, Boston, (2005) Fletcher, R.: A limited memory steepest descent method. Math. Program. **135**(1-2), 413–436 (2012) Forsythe, G.E.: On the asymptotic directions of the s-dimensional optimum gradient method. Numer. Math. **11**(1), 57–76 (1968) Frassoldati, G., Zanni, L., Zanghirati, G.: New adaptive stepsize selections in gradient methods. J. Ind. Manag. Optim. **4**(2), 299 (2008) Friedlander, A., Mart[í]{}nez, J.M., Molina, B., Raydan, M.: Gradient method with retards and generalizations. SIAM J. Numer. Anal. **36**(1), 275–289 (1998) Gonzaga, C.C., Schneider, R.M.: On the steepest descent algorithm for quadratic functions. Comp. Optim. Appl. **63**(2), 523–542 (2016) Gould, N.I., Orban, D., Toint, P.L.: CUTEst: a constrained and unconstrained testing environment with safe threads for mathematical optimization. Comp. Optim. Appl. **60**(3), 545–557 (2015) Hager, W.W., Zhang, H.: A new active set algorithm for box constrained optimization. SIAM J. Optim. **17**(2), 526–557 (2006) Huang, Y., Dai, Y.H., Liu, X.W., Zhang, H.: Gradient methods exploiting spectral properties. arXiv preprint arXiv:1905.03870 (2019) Huang, Y., Dai, Y.H., Liu, X.W., Zhang, H.: On the asymptotic convergence and acceleration of gradient methods. arXiv preprint arXiv:1908.07111 (2019) Huang, Y., Liu, H., Zhou, S.: Quadratic regularization projected Barzilai–Borwein method for nonnegative matrix factorization. Data Min. Knowl. Disc. **29**(6), 1665–1684 (2015) Jiang, B., Dai, Y.H.: Feasible Barzilai–Borwein-like methods for extreme symmetric eigenvalue problems. Optim. Method Softw. **28**(4), 756–784 (2013) Liu, Y.F., Dai, Y.H., Luo, Z.Q.: Coordinated beamforming for miso interference channel: Complexity analysis and efficient algorithms. IEEE Trans. Signal Process. **59**(3), 1142–1157 (2011) Raydan, M.: The Barzilai and Borwein gradient method for the large scale unconstrained minimization problem. SIAM J. Optim. **7**(1), 26–33 (1997) Tan, C., Ma, S., Dai, Y.H., Qian, Y.: Barzilai–Borwein step size for stochastic gradient descent. In: Advances in Neural Information Processing Systems, pp. 685–693 (2016) Wang, Y., Ma, S.: Projected Barzilai–Borwein method for large-scale nonnegative image restoration. Inverse Probl. Sci. En. **15**(6), 559–583 (2007) Wright, S.J., Nowak, R.D., Figueiredo, M.A.: Sparse reconstruction by separable approximation. IEEE Trans. Signal Process. **57**(7), 2479–2493 (2009) Yuan, Y.X.: A new stepsize for the steepest descent method. J. Comput. Math. **24**(2), 149–156 (2006) Yuan, Y.X.: Step-sizes for the gradient method. AMS IP Studies in Advanced Mathematics. **42**(2), 785–796 (2008) Zhou, B., Gao, L., Dai, Y.H.: Gradient methods with adaptive step-sizes. Comp. Optim. Appl. **35**(1), 69–86 (2006) [^1]: codes available at <https://www.ime.usp.br/~egbirgin/tango/codes.php>

ArXiv

--- abstract: 'We study the negative gradient flow of the spinorial energy functional (introduced by Ammann, Weiß, and Witt) on 3-dimensional Berger spheres. For a certain class of spinors we show that the Berger spheres collapse to a 2-dimensional sphere. Moreover, for special cases, we prove that the volume-normalized standard 3-sphere together with a Killing spinor is a stable critical point of the volume-normalized version of the flow. Our results also include an example of a critical point of the volume-normalized flow on the 3-sphere, which is not a Killing spinor.' author: - Johannes Wittmann bibliography: - 'Literatur.bib' title: 'The Spinorial Energy Functional: Solutions of the Gradient Flow on Berger Spheres' --- Introduction ============ Let $M$ be a compact spin manifold and $\mathcal{N}$ the union of all pairs $(g,\varphi)$ where $g$ is a Riemannian metric on $M$ and $\varphi\in\Gamma(\Sigma(M,g))$ is a spinor of the spin manifold $(M,g)$ whose pointwise norm is constant and equal to $1$. The *spinorial energy functional* $\mathcal{E}$, introduced in [@AWWsflow], is defined by $$\mathcal{E}\colon \mathcal{N}\rightarrow [0,\infty), \hspace{3em}(g,\varphi)\mapsto \frac{1}{2}\int_M|\nabla^{\Sigma(M,g)}\varphi|^2dv^g,$$ where $dv^g$ is the Riemannian volume form of $(M,g)$ and $|.|$ is the pointwise norm on $T^*M\otimes \Sigma(M,g)$. If $\textup{dim}M\ge 3$, then the critical points of $\mathcal{E}$ are precisely the pairs $(g,\varphi)$ consisting of a Ricci-flat Riemannian metric $g$ and a parallel spinor $\varphi$. In the surface case, the spinorial energy functional is related to the Willmore energy of immersions and treated in detail in [@AWWsflowsurf]. On the Fréchet-bundle $\mathcal{N}\rightarrow\mathcal M$, $\mathcal{M}:=\{\text{Riemannian metrics on } M\}$, there exists a natural connection, which is defined in [@AWWsflow] with the aid of results in [@BourGaud]. This connection defines a splitting of $T\mathcal{N}$ in horizontal and vertical subbundles, which allows us to define a Riemannian metric on $\mathcal{N}$. The negative gradient flow of $\mathcal{E}$ with respect to this Riemannian metric is called the *spinor flow*. Short time existence and uniqueness of the spinor flow was shown in [@AWWsflow] with a variant of DeTurck’s trick. In this paper the spinor flow on 3-dimensional Berger spheres is treated. We view the 3-sphere $S^3$ as a $S^1$-principal bundle over $S^2$ via the Hopf fibration $\pi\colon S^3\rightarrow S^2$. Rescaling the standard metric $g_{S^3}$ along the fibers of the Hopf fibration by $\varepsilon >0$ yields the Berger metrics $g^\varepsilon$ on $S^3$. We call $(S^3,g^\varepsilon)$ a Berger sphere. There is a certain class of spinors on $S^3$, the so-called $S^1$-invariant spinors [@ABCollaps], [@Mor], which are in one-to-one correspondence to the spinors on the base manifold $S^2$. Our first theorem concerns these spinors. Let $M=S^3$ and as initial value $(g_0,\varphi_0)$ choose $g_0=g^\varepsilon$ and $\varphi_0$ a spinor of unit length that corresponds to an arbitrary Killing spinor on the base $S^2$. Then, if the fibers are sufficiently short (i.e. $\varepsilon$ is small enough), the spinor flow converges to a 2-dimensional sphere in infinite time. This theorem can be seen as a special case of the conjecture that $S^1$-principal bundles with suitable Riemannian metrics and sufficiently short fibers together with $S^1$-invariant spinors collapse to the base manifold under the spinor flow. In [@AWWsflow] it was observed that the volume-normalized standard metric on $S^3$ together with a Killing spinor is a critical point of the volume-normalized spinor flow. It is not clear whether this critical point is stable. However, there are such stability results for other geometric flows, see e.g. [@Huisken 1.1 Theorem] in the case of the mean curvature flow. Our second theorem is a first positive result concerning this stability question. Let $M=S^3$ and as initial value $(g_0,\varphi_0)$ choose $g_0=c(\varepsilon)g^\varepsilon$ the volume-normalized Berger metric and $\varphi_0$ a spinor that is obtained via parallel transport of an arbitrary Killing spinor of unit length from $(S^3,g_{S^3})$ to $(S^3,g^\varepsilon)$ as described in Remark \[rmk main res 2\]. Then, if we are not too far away from $c(1)g_{S^3}$ (i.e. $\varepsilon$ is sufficiently close to $1$), the volume-normalized spinor flow converges in infinite time to the volume-normalized standard metric on $S^3$ together with a Killing spinor. Overview of the proof --------------------- First of all, in [@AWWsflow] it was shown that under the splitting of $T\mathcal{N}$ the negative gradient of $\mathcal{E}$ has an expression $$-\textup{grad}\mathcal{E}_{(g,\varphi)}=(Q_1(g,\varphi),Q_2(g,\varphi))\in\Gamma(\odot^2T^*M)\oplus\Gamma(\Sigma(M,g)),$$ where $Q_1(g,\varphi)$ and $Q_2(g,\varphi)$ depend mainly on $\nabla^{(M,g)}\varphi$. This fact is important for us, because it means, essentially, that we can understand $-\textup{grad}\mathcal{E}_{(g,\varphi)}$ by understanding $\nabla^{(M,g)}\varphi$. Furthermore, one of the main tools for us to prove the above theorems are generalized cylinders [@BaerGaudMoroi], which provide a way to identify spinors for different metrics. To be more concrete, given a smooth $1$-parameter family $(g_t)_{t\in I}$ of Riemannian metrics on a manifold $M$, $I$ an interval, the generalized cylinder is the manifold $\mathcal{Z}:=I\times M$ together with the Riemannian metric $g_\mathcal{Z}:=dt^2+g_t$. If the dimension of $M$ is odd, as in our case, we get an identification $\Sigma^+(\mathcal Z,g_\mathcal{Z})|_{\{t\}\times M}\cong\Sigma (M,g_t)$. In particular, we can think of sections $\varphi\in\Gamma(\Sigma^+(\mathcal Z,g_\mathcal{Z}))$ as families of sections $(\varphi_t)_{t\in I}$ with $\varphi_t\in\Gamma(\Sigma(M,g_t))$ where $\varphi_t(.):=\varphi(t,.)$. Denote by $\pi\colon S^3\rightarrow S^2$ the Hopf fibration as above. We write $$\begin{aligned} \label{eq ansatz for g_t} g_t(X_1+Y_1,X_2+Y_2):=g_{S^3}(\alpha(t)X_1+\beta(t)Y_1, \alpha(t)X_2+\beta(t)Y_2)\end{aligned}$$ for $t\in I=[0,b)$, $b\in(0,\infty]$, $X_i\in \textup{ker}(d\pi)$, $Y_i\in\textup{ker}(d\pi)^\bot$, $i=1,2$, and smooth functions $\alpha,$ $\beta\colon I\rightarrow (0,\infty)$. We will choose as ansatz for the metric part of the solution where we require that $\alpha(0)$ and $\beta(0)$ are chosen so that $g_0$ is the metric part of our initial value. Using the generalized cylinder with respect to we write $$\begin{aligned} \label{eq ansatz for varphi_t} \varphi_t(p):=\mathcal{P}_{0,t}(p)(\varphi_0(p))\end{aligned}$$ where $\mathcal{P}_{0,t}(p)$ is the parallel transport in $\Sigma^+(\mathcal{Z},g_\mathcal{Z})$ with respect to $\nabla^{\Sigma^+(\mathcal{Z},g_{\mathcal{Z}})}$ along the curve $\gamma_p(s):=(s,p)$ from $\gamma_p(0)$ to $\gamma_p(t)$. Then we choose as ansatz for the spinor part of the solution. In the next step, we derive an expression for $\nabla^{\Sigma(S^3,g_t)}\varphi_t$ that depends in particular on $\alpha$ and $\beta$. To achieve this, we use curvature terms to construct suitable differential equations in $\Sigma^+(\mathcal{Z},g_\mathcal{Z})$. We use these expressions for $\nabla^{\Sigma(S^3,g_t)}\varphi_t$ to calculate $Q_1(g_t,\varphi_t)$ and $Q_2(g_t,\varphi_t)$. After that we show $\frac{\partial}{\partial t}\varphi_t=0=Q_2(g_t,\varphi_t)$ independent of the choice of $\alpha$ and $\beta$. Finally, we will see that $\frac{\partial}{\partial t}g_t=Q_1(g_t,\varphi_t)$ is equivalent to a system of two non-linear ordinary differential equations for $\alpha$ and $\beta$. We solve these systems to get the desired properties of the solutions. Acknowledgments {#acknowledgments .unnumbered} --------------- I would like to thank Bernd Ammann for his ongoing support and many fruitful discussions. I am also grateful to Nicolas Ginoux for his insightful comments at the early stages of my research. Preliminaries ============= Spin geometry ------------- In this section we fix notation and review basics of spin geometry which will be relevant in the following. For more details we refer to e.g. [@LaMi], [@Hij], [@Fri] and [@Roe]. Let $M$ be an oriented $n$-dimensional manifold and denote by $\textup{GL}^+M$ the $GL^+(n,\mathbb{R})$-principal bundle of oriented frames for $M$. Moreover, we denote by $\theta\colon \widetilde{GL}^+(n,\mathbb{R})\rightarrow GL^+(n,\mathbb{R})$ the universal covering for $n\ge 3$ and the connected twofold covering for $n=2$. A *topological spin structure on $M$* is a $\theta$-reduction of $\textup{GL}^+M$, i.e. a topological spin structure on $M$ is a $\widetilde{GL}^+(n,\mathbb{R})$-principal bundle $\widetilde{\textup{GL}}^+M$ over $M$ together with a twofold covering $\Theta\colon\widetilde{\textup{GL}}^+M\rightarrow\textup{GL}^+M$ such that the following diagram commutes $$\begin{xy} \xymatrix{ \widetilde{\textup{GL}}^+M\times \widetilde{GL}^+(n,\mathbb{R}) \ar[r] \ar[dd]^{\Theta\times\theta} & \widetilde{\textup{GL}}^+M \ar[dd]^\Theta \ar[rd] & \\ & & M\\ \textup{GL}^+M\times GL^+(n,\mathbb{R}) \ar[r] & \textup{GL}^+M \ar[ru] & } \end{xy}$$ where the horizontal arrows denote the group actions of the principal bundles. Now let $(M,g)$ be an oriented Riemannian manifold and $\textup{SO}(M,g)$ the $SO(n,\mathbb{R})$-principal bundle of oriented orthonormal frames for $M$. Restricting $\theta$ to the *spin group* given by $\textup{Spin}(n):=\theta^{-1}(SO(n,\mathbb{R}))$, we define a *metric spin structure on $M$* to be a $\theta|_{\textup{Spin}(n)}$-reduction of $\textup{SO}(M,g)$. Again, this means that a metric spin structure on $M$ is a $\textup{Spin}(n)$-principal bundle $\textup{Spin}(M,g)$ over $M$ together with a twofold covering $\Theta\colon \textup{Spin}(M,g)\rightarrow \textup{SO}(M,g)$ such that the following diagram commutes $$\begin{xy} \xymatrix{ \textup{Spin}(M,g)\times \textup{Spin}(n) \ar[r] \ar[dd]^{\Theta\times\theta} & \textup{Spin}(M,g) \ar[dd]^\Theta \ar[rd] & \\ & & M\\ \textup{SO}(M,g)\times SO(n,\mathbb{R}) \ar[r] & \textup{SO}(M,g) \ar[ru] & } \end{xy}$$ Given a topological spin structure $\widetilde{\textup{GL}}^+M$ on an oriented manifold $M$, every Riemannian metric $g$ on $M$ defines a metric spin structure on $(M,g)$ by $\textup{Spin}(M,g):=\widetilde{\textup{GL}}^+M|_{\textup{SO}(M,g)}$. In the following, the term *spin structure* refers to a topological or metric spin structure and it should always be clear from the context which one we mean. In order to introduce (complex) spinors, we consider representations of $\textup{Spin}(n)$. We first note that the spin group can be realized as a subgroup of the group of invertible elements in $\mathbb{C}l_n$ where $\mathbb{C}l_n$ is the Clifford algebra of $\mathbb{C}^n$ with inner product given by the complex bilinear extension of the standard inner product of $\mathbb{R}^n$, namely $\textup{Spin}(n)\cong\{x_1\cdot\ldots\cdot x_{2k}\text{ }|\text{ } x_i\in S^{n-1}\subset\mathbb{R}^n\subset\mathbb{C}l_n,\text{ }k\in\mathbb{N}\}$. If $n$ is even, then there exists exactly one equivalence class of irreducible complex representations of $\mathbb{C}l_n$ and every such representation is of dimension $2^\frac{n}{2}$. If $n$ is odd, then there exist exactly two equivalence classes of irreducible complex representations of $\mathbb{C}l_n$, each of dimension $2^{\frac{n-1}{2}}$. Introducing the *complex volume element* $\omega_n:=i^{\lfloor\frac{n+1}{2}\rfloor} e_1\cdot\ldots\cdot e_n\in\mathbb{C}l_n$ where $a=\lfloor b\rfloor$ is the largest integer $a\le b$ and $(e_1,\ldots,e_n)$ is the standard basis of $\mathbb{C}^n$, we can distinguish the two different equivalence classes for $n$ odd by the action of $\omega_n$, i.e. $\omega_n$ acts as the identity $id$ on one equivalence class and as $-id$ on the other. The *complex spinor representation* $\rho\colon \textup{Spin}(n)\rightarrow \textup{Aut}(\Sigma_n)$ is the restriction of an irreducible complex representation $\rho\colon \mathbb{C}l_n\rightarrow \textup{End}(\Sigma_n)$ of $\mathbb{C}l_n$ to $\textup{Spin}(n)$ where for $n$ odd we require $\rho(\omega_n)=id_{\Sigma_n}$. For $n$ odd, the complex spinor representation is irreducible. For $n$ even, it splits into two irreducible representations $\rho=\rho^+\oplus\rho^-$ where $\rho^\pm\colon \textup{Spin}(n)\rightarrow \textup{Aut}(\Sigma_n^\pm)$ have dimension $2^{\frac{n}{2}-1}$ and $\Sigma_n^\pm$ are the $\pm1$-eigenspaces of $\rho(\omega_n)$. Let $\textup{Spin}(M,g)$ be a spin structure on $(M,g)$. The *(complex) spinor bundle* $\Sigma (M,g)$ is the complex vector bundle associated to the spin structure and the complex spinor representation, i.e. $\Sigma(M,g):=\textup{Spin}(M,g)\times_\rho \Sigma_n$. For $n$ even, we have an isomorphism $\Sigma (M,g)\cong \Sigma^+ (M,g) \oplus\Sigma^-(M,g)$ where $\Sigma^\pm(M,g):=\textup{Spin}(M,g)\times_{\rho^\pm}\Sigma_n^\pm$. Next we introduce the so-called Clifford multiplication, which allows to multiply spinors and tangent vectors. To that end, notice that $TM\cong \textup{Spin}(M,g)\times_{\tau\circ\theta|_{\textup{Spin(n)}}}\mathbb{R}^n$, where $\tau$ is the standard representation of $SO(n,\mathbb{R})$ on $\mathbb{R}^n$. Given $\varphi=[p,\sigma]\in\Sigma_x(M,g)$ and $X=[p,v]\in T_xM$ we define the *Clifford multiplication (on $\Sigma (M,g)$)* by $X\cdot\varphi:=[p,\rho(v)(\sigma)]$. From the relations of the Clifford algebra $\mathbb{C}l_n$, it follows that $$\begin{aligned} \label{eq spingeo cliffordm} X\cdot (Y\cdot \varphi) + Y\cdot (X\cdot \varphi)=-2g(X,Y)\varphi,\end{aligned}$$ for all $X,Y\in T_xM$ and $\varphi\in\Sigma_x(M,g)$. For $n$ even, Clifford multiplication interchanges the factors $\Sigma^\pm(M,g)$. Moreover, given an oriented orthonormal basis $(e_1,\ldots,e_n)$ of $T_xM$ and $\varphi\in\Sigma_x(M,g)$ for $n$ odd, respectively $\varphi\in\Sigma^+_x(M,g)$ for $n$ even, we have $$\begin{aligned} \label{eq spingeo cliffordm onb} i^{\lfloor\frac{n+1}{2}\rfloor} e_1\cdot\ldots\cdot e_n\cdot\varphi=\varphi.\end{aligned}$$ To measure the length of spinors, we introduce a natural bundle metric on $\Sigma(M,g)$. First, given an irreducible representation $\rho\colon\mathbb{C}l_n\rightarrow \textup{End}(\Sigma_n)$ of $\mathbb{C}l_n$, there exists a hermitian inner product $\langle.,.\rangle_{\Sigma_n}$ on $\Sigma_n$ such that $\langle\rho(x)(\psi),\varphi\rangle_{\Sigma_n}=\langle\psi,\rho(x)(\varphi)\rangle_{\Sigma_n}$ for all $x\in\mathbb{R}^n$, $\varphi$, $\psi\in\Sigma_n$. In particular, the inner product $\langle.,.\rangle_{\Sigma_n}$ is $\textup{Spin}(n)$-invariant and therefore induces a bundle metric on $\Sigma(M,g)$, which we denote by $\langle.,.\rangle$. It holds that $$\begin{aligned} \label{eq spingeo cliffordm scp} \langle X\cdot\psi,\varphi\rangle=-\langle\psi,X\cdot\varphi\rangle,\end{aligned}$$ for all $X\in T_xM$, $\varphi$, $\psi\in\Sigma_x(M,g)$. For $\varphi\in\Gamma(\Sigma(M,g))$ we set $|\varphi|:=\sqrt{\langle\varphi,\varphi\rangle}$. In order to differentiate spinors we note that the Levi-Civita connection $\nabla$ on $(M,g)$ can be lifted to a metric connection $\nabla^{\Sigma(M,g)}$ on $\Sigma(M,g)$, the *spinorial Levi-Civita connection*. For all $X$, $Y\in\Gamma(TM)$ and $\varphi\in\Gamma(\Sigma(M,g))$ we have $$\begin{aligned} \nabla^{\Sigma(M,g)}_X(Y\cdot\varphi)=(\nabla_XY)\cdot\varphi + Y\cdot \nabla^{\Sigma(M,g)}_X\varphi.\end{aligned}$$ For $n$ even, the factors $\Sigma^\pm(M,g)$ are invariant under $\nabla^{\Sigma(M,g)}$. In particular, we get connections $\nabla^{\Sigma^\pm(M,g)}$ on $\Sigma^\pm(M,g)$. Denote by $R^{\Sigma(M,g)}$ the curvature of $\Sigma(M,g)$. Let $(e_1,\ldots,e_n)$ be a local orthonormal frame for $(M,g)$. It holds that $$\begin{aligned} \label{eq spingeo curvature} R^{\Sigma(M,g)}(X,Y)\varphi=\frac12\sum_{1\le i < j\le n} g(R^M(X,Y)e_i,e_j)e_i\cdot(e_j\cdot \varphi),\end{aligned}$$ for all $X$, $Y\in\Gamma(TM)$, $\varphi\in\Gamma(\Sigma(M,g))$ where $R^M$ is the curvature of $(M,g)$. Generalized cylinders --------------------- Details concerning this section can be found in [@BaerGaudMoroi]. Let $M$ be a manifold, $I\subset\mathbb{R}$ an interval and $(g_t)_{t\in I}$ a smooth 1-parameter family of Riemannian metrics on $M$. The *generalized cylinder* is the Riemannian manifold $(\mathcal{Z},g_\mathcal{Z})$, where $\mathcal{Z}:=I\times M$ and $g_\mathcal{Z}:=dt^2+g_t$. The Riemannian hypersurface $\{t\}\times M$ is isometric to $(M,g_t)$ and we denote both by $M_t$. Moreover, the vector field $\nu:=\frac{\partial}{\partial t}\in\Gamma(T\mathcal{Z})$ is of unit length and $\nu|{M_t}$ is orthogonal to $M_t$. We write $W=W_t$ for the Weingarten map of $M_t$ with respect to $\nu|_{M_t}$. The following lemma will be used later. For all $U$, $X$, $Y\in T_pM$, $p\in M$, and $t\in I$, it holds that $$\begin{aligned} \nabla^{\mathcal{Z}}_\nu \nu &=0,\notag \\ g_t(W_t(X),Y)&=-\frac{1}{2}\dot{g}_t(X,Y),\notag \\ g_\mathcal{Z}(R^\mathcal{Z}(X,Y)U,\nu)&=\frac{1}{2}((\nabla^{M_t}_Y\dot{g}_t)(X,U)-(\nabla^{M_t}_X\dot{g}_t)(Y,U)),\label{eq cy 3}\\ g_\mathcal{Z}(R^\mathcal{Z}(X,\nu)\nu,Y)&=-\frac{1}{2}(\ddot{g}_t(X,Y)+\dot{g}_t(W_t(X),Y))\label{eq cy 4}. \end{aligned}$$ If $\tilde{Z}\in\Gamma(T\mathcal{Z})$ with $\tilde{Z}(t,p)=(0,Z_t(p))\in T_tI\times T_pM$ for all $t\in I$, $p\in M$, then $$\begin{aligned} [\nu,\tilde{Z}](t,p)&=(0,\frac{d}{ds}\bigg|_{s=t}Z_s(p))\in T_tI\times T_pM,\label{eq cy 5}\\ (\nabla^\mathcal{Z}_\nu \tilde{Z})(t,p)&=(0,\frac{d}{ds}\bigg|_{s=t}Z_s(p)-W_t(Z_t(p)))\label{eq cy 6} \end{aligned}$$ for all $t\in I$, $p\in M$ where $\nabla^{\mathcal{Z}}$ is the Levi-Civita connection of $(\mathcal{Z},g_\mathcal{Z})$. The next lemma describes how we can identify spinors of different spinor bundles with the help of generalized cylinders. \[lemma cyl identi\] Let $M$ be an oriented manifold together with a topological spin structure $\widetilde{\textup{GL}}^+M$. The topological spin structure on $M$ induces a metric spin structure on $(\mathcal{Z}, g_\mathcal{Z})$ and metric spin structures $\textup{Spin}(M,g_t):=\widetilde{\textup{GL}}^+M|_{\textup{SO}(M,g_t)}$ on $(M,g_t)$. For the respective spinor bundles we have the following isomorphisms of vector bundles: If $n$ is even, then $\Sigma(\mathcal Z,g_\mathcal{Z})|_{M_t}\cong\Sigma M_t$. If $n$ is odd, then $\Sigma^+(\mathcal Z,g_\mathcal{Z})|_{M_t}\cong\Sigma M_t$. The bundle metrics $\langle.,.\rangle$ are preserved by these isomorphisms. Moreover, if “$\cdot$” and “$\cdot_t$” denote the Clifford multiplications in $\Sigma^{(+)}(\mathcal{Z},g_\mathcal{Z})$ and $\Sigma M_t$, then it holds that $$\nu\cdot(X\cdot\varphi)=X\cdot_t\varphi$$ for all $X\in TM$, $\varphi\in\Sigma M_t$. If we write $\nabla^t=\nabla^{\Sigma M_t}$, then we have $$\nabla^{\Sigma(\mathcal{Z},g_\mathcal{Z})}_X\varphi=\nabla^t_X\varphi -\frac{1}{2}W_t(X)\cdot_t\varphi$$ for all $\varphi\in\Gamma(\Sigma M_t)$. The spinorial energy functional and its gradient flow {#sect sflow} ----------------------------------------------------- In the following we work with the real part of $\langle.,.\rangle$ and we write $(.,.):=\textup{Re}\langle.,.\rangle$. It will be useful that $(X\cdot\varphi,Y\cdot\varphi)=g(X,Y)(\varphi,\varphi)$ and $(X\cdot\varphi,\varphi)=0$ hold for all $X$, $Y\in T_xM$ and $\varphi\in\Sigma_x(M,g)$. These identities follow directly from and . Let $M$ be a connected, compact, oriented manifold with a fixed topological spin structure $\widetilde{\textup{GL}}^+M$ and $\textup{dim}M\ge 2$. As stated before, every choice of Riemannian metric $g$ on $M$ defines a metric spin structure $\textup{Spin}(M,g):=\widetilde{\textup{GL}}^+M|_{\textup{SO}(M,g)}$ and so we have the corresponding spinor bundles $\Sigma(M,g)$. We set $\mathcal{N}_g:=\{\varphi\in\Gamma(\Sigma(M,g))\text{ }|\text{ }|\varphi|=1\}$, and $\mathcal{N}:=\bigsqcup_{g\in\mathcal{M}}\mathcal{N}_g$. The *spinorial energy functional* $\mathcal{E}$ is defined by $$\mathcal{E}\colon \mathcal{N}\rightarrow [0,\infty), \hspace{3em}(g,\varphi)\mapsto \frac{1}{2}\int_M|\nabla^{\Sigma(M,g)}\varphi|^2dv^g,$$ As mentioned in the introduction there exists a natural connection on the Fréchet-bundle $\mathcal{N}\rightarrow\mathcal M$. For details we refer to [@AWWsflow]. From that connection we get horizontal tangent spaces $\mathcal{H}_{(g,\varphi)}\cong\Gamma(\odot^2T^*M)$ and a splitting $$\begin{aligned} \label{eq z2} T_{(g,\varphi)}\mathcal{N}\cong\Gamma(\odot^2T^*M)\oplus V_{(g,\varphi)}\end{aligned}$$ where $$V_{(g,\varphi)}=\{\psi\in\Gamma(\Sigma(M,g))\text{ }|\text{ }(\varphi,\psi)=0 \}.$$ On the first factor, we choose the inner product which we get by integrating the natural inner product on $(2,0)$-tensors. On the second factor, we choose the $L^2$-inner product defined by $$(\psi_1,\psi_2)_{L^2}:=\int_M(\psi_1,\psi_2)dv^g,$$ for $\psi_1$, $\psi_2\in V_{(g,\varphi)}$. The negative gradient flow of $\mathcal{E}$, $$\frac{\partial}{\partial t}(g_t,\varphi_t)=-\textup{grad}\mathcal{E}_{(g_t,\varphi_t)},$$ is called the *spinor flow*. Under the splitting we have $$-\textup{grad}\mathcal{E}_{(g,\varphi)}=(Q_1(g,\varphi),Q_2(g,\varphi)),$$ with $$\begin{aligned} Q_1(g,\varphi)&=-\frac{1}{4}|\nabla^{\Sigma(M,g)}\varphi|^2g-\frac{1}{4}\textup{div}_g T_{g,\varphi}+\frac{1}{2}\left(\nabla^{\Sigma(M,g)}\varphi\otimes\nabla^{\Sigma(M,g)}\varphi\right),\\ Q_2(g,\varphi)&=-\left(\left({\nabla^{\Sigma (M,g)}}\right)^*\nabla^{\Sigma (M,g)}\right)\varphi+|\nabla^{\Sigma(M,g)}\varphi|^2\varphi,\end{aligned}$$ for all $(g,\varphi)\in\mathcal{N}$. Here, $\left(\nabla^{\Sigma(M,g)}\varphi\otimes\nabla^{\Sigma(M,g)}\varphi\right)(X,Y):=\left(\nabla^{\Sigma(M,g)}_X\varphi,\nabla^{\Sigma(M,g)}_Y\varphi\right)$ for all $X,Y\in\Gamma(TM)$, and $T_{g,\varphi}$ is the symmetrization of $\left(X\cdot Y\cdot\varphi + g(X,Y)\varphi,\nabla^{\Sigma(M,g)}_Z\varphi\right)$ in the second and third component where $X,Y,Z\in\Gamma(TM)$. The *(volume) normalized spinor flow* is the negative gradient flow of $\mathcal{E}|_{\mathcal{N}_1}$ for $\mathcal{N}_1:=\{(g,\varphi)\in\mathcal{N}\text{ }|\text{ } \textup{vol}(M,g)=1\}$. We have $-\textup{grad}(\mathcal{E}|_{\mathcal{N}_1})=(\tilde{Q}_1,Q_2|_{\mathcal{N}_1})$ with $$\tilde{Q}_1(g,\varphi)=Q_1(g,\varphi)+\frac{n-2}{2n}\frac{1}{\textup{vol}(M,g)}\mathcal{E}(g,\varphi)g$$ for all $(g,\varphi)\in\mathcal{N}_1$. Note that in the above identity for $\tilde{Q}_1$ the “$\frac{1}{\textup{vol}(M,g)}$” is equal to $1$. However, it is important in our strategy of solving the normalized spinor flow. More concretely, we will construct $(g_t,\varphi_t)$ with $|\varphi_t|=1$ for all $t$ and $\frac{\partial}{\partial t}(g_t,\varphi_t)=(\tilde{Q}_1(g_t,\varphi_t),Q_2(g_t,\varphi_t))$ where for the initial value we have $\textup{vol}(M,g_0)=1$. To make sure that $(g_t,\varphi_t)$ is a solution of the normalized spinor flow, we then need to verify that $\textup{vol}(M,g_t)=1$ for all $t$. To that end, we calculate $$\begin{aligned} \frac{d}{dt}\bigg|_{t=s}\textup{vol}(M,g_t)&=\frac{d}{dt}\bigg|_{t=s}\int_M dv^{g_t}=\int_M\frac{d}{dt}\bigg|_{t=s}dv^{g_t}\\ &=\int_M (\frac{1}{2}\textup{tr}_{g_{s}}\dot{g}_{s})dv^{g_{s}}\\ &=\int_M (\frac{1}{2}\textup{tr}_{g_{s}}\tilde{Q}_1(g_s,\varphi_s))dv^{g_{s}}\\ &=\frac{1}{2}\int_M\textup{tr}_{g_{s}}Q_1(g_s,\varphi_s)dv^{g_{s}}+\frac{1}{2}\frac{n-2}{2n}\frac{1}{\textup{vol}(M,g_s)}\mathcal{E}(g_s,\varphi_s)\int_M\textup{tr}_{g_s}g_sdv^{g_s}\\ &=\frac{1}{2}\int_M-\frac{n-2}{4}|\nabla^{\Sigma(M,g_s)}\varphi_s|^2dv^{g_s}+\frac{1}{2}\frac{n-2}{2}\frac{1}{\textup{vol}(M,g_s)}\mathcal{E}(g_s,\varphi_s)\textup{vol}(M,g_s)\\ &=-\frac{1}{2}\frac{n-2}{2}\mathcal{E}(g_s,\varphi_s)+\frac{1}{2}\frac{n-2}{2}\mathcal{E}(g_s,\varphi_s)\\ &=0,\end{aligned}$$ so $\textup{vol}(M,g_t)=1$ for every $t$ and $(g_t,\varphi_t)$ is in fact a solution of the normalized spinor flow. Solutions of the spinor flow on Berger spheres ============================================== In this section we state and prove our main results, Theorem \[main result 1\] and Theorem \[main result 2\]. First we collect necessary technical ingredients. Then we define $S^1$-invariant spinors, which are part of the initial values of theorem \[main result 1\]. After that we prove our main results with the strategy explained in the introduction. As stated in the introduction we view $S^3$ as a $S^1$-principal bundle over $S^2$ via the Hopf fibration $\pi\colon S^3\rightarrow S^2$. If we equip $S^2=\mathbb{C}P^1$ with the Fubini-Study metric $g_{FS}$, then the Hopf fibration $\pi\colon (S^3,g_{S^3})\rightarrow (S^2,g_{FS})$ turns into a Riemannian submersion. The action of $S^1$ on $S^3$ induces a global flow whose infinitesimal generator we denote by $K\in\Gamma(TS^3)$. We have that $K(p)\in\textup{ker}(d\pi_p)$ for all $p\in S^3$ and $K$ is of unit length with respect to $g_{S^3}$. On $S^3$ we choose the connection which is given by $$p\mapsto \textup{ker}(d\pi_p)^\bot.$$ This connection induces a connection form $\tilde{\omega}\colon TS^3\rightarrow i\mathbb{R}$. It holds that $\tilde{\omega}(K)= i$. We write $\omega:=\frac{1}{i}\tilde{\omega}$ and denote by $d\omega$ the differential of $\omega$, i.e. $$d\omega(X,Y)=L_X(\omega(Y))-L_Y(\omega(X))-\omega([X,Y]).$$ With $X^*$ we denote the horizontal lift of $X$ (with respect to the above connection). For the rest of this paper we fix an orientation on $S^2$. On $S^3$ we fix the orientation that satisfies the following: If $(f_1,f_2)$ is any oriented local orthonormal frame for $(S^2,g_{FS})$, then $(K,f_1^*,f_2^*)$ is an oriented local orthonormal frame for $(S^3,g_{S^3})$. In Remark \[rem change or conv\] we explain how our results change if we choose the other orientation on $S^3$ (i.e. $(-K,f_1^*,f_2^*)$ is oriented). For the rest of this paper we use the following notation: $(f_1,f_2)$ denotes an arbitrary oriented local orthonormal frame for $(S^2,g_{FS})$. Moreover, we set $$\begin{aligned} (f_1(t),f_2(t))&:=(\frac{1}{\beta(t)}f_1,\frac{1}{\beta(t)}f_2),\\ (e_0(t), e_1(t), e_2(t))&:=(\frac{1}{\alpha(t)}K, f_1(t)^*,f_2(t)^*).\end{aligned}$$ Then $(e_0(t), e_1(t), e_2(t))$ is an oriented local orthonormal frame for $(S^3,g_t)$ where $g_t$, $\alpha(t)$, and $\beta(t)$ are defined by . We set $$a:=d\omega(f_1^*,f_2^*)=\pm 2.$$ Note that $a$ is a constant that does not depend on the choice of the oriented local orthonormal frame $(f_1,f_2)$. \[lemma z1\]If $\nabla^t$ denotes the Levi-Civita connection on $(S^3,g_t)$ and $(S^2,\beta(t)^2g_{FS})$ respectively, we have $$\begin{aligned} \nabla^t_{e_0(t)}e_0(t)&=0,\hspace{5em} & \nabla^t_{e_1(t)}e_2(t)&=-\frac{1}{2}\frac{\alpha(t)}{\beta(t)^2}ae_0(t)+\left(\nabla^t_{f_1(t)}f_2(t)\right)^*,\\ \nabla^t_{e_0(t)}e_1(t)&=\frac{1}{2}\frac{\alpha(t)}{\beta(t)^2}ae_2(t),\hspace{3em} & \nabla^t_{e_2(t)}e_1(t)&=\frac{1}{2}\frac{\alpha(t)}{\beta(t)^2}ae_0(t)+\left(\nabla^t_{f_2(t)}f_1(t)\right)^*,\\ \nabla^t_{e_0(t)}e_2(t)&=-\frac{1}{2}\frac{\alpha(t)}{\beta(t)^2}ae_1(t),\hspace{3em} &\nabla^t_{e_2(t)}e_0(t)&=-\frac{1}{2}\frac{\alpha(t)}{\beta(t)^2}ae_1(t),\\ \nabla^t_{e_1(t)}e_0(t)&=\frac{1}{2}\frac{\alpha(t)}{\beta(t)^2}ae_2(t), \hspace{3em} &\nabla^t_{e_2(t)}e_2(t)&=\left(\nabla^t_{f_2(t)}f_2(t)\right)^*,\\ \nabla^t_{e_1(t)}e_1(t)&=\left(\nabla^t_{f_1(t)}f_1(t)\right)^*.\end{aligned}$$ Since horizontal lifts are right invariant, it follows that $$[e_0(t),e_j(t)]=0$$ on $S^3$ for $j=1,2$. Using the Koszul formula we then compute the Christoffel symbols $\tilde{\Gamma}_{ij}^k$ of $(e_0(t), e_1(t), e_2(t))$ with respect to $\nabla^t$: $$\begin{aligned} -\tilde{\Gamma}_{ij}^0&=\tilde{\Gamma}_{i0}^j=\tilde{\Gamma}_{0i}^j=\frac{1}{2}\frac{\alpha(t)}{\beta(t)^2}a,\\ \tilde{\Gamma}_{00}^i&=\tilde{\Gamma}_{i0}^0=\tilde{\Gamma}_{0i}^0=\tilde{\Gamma}_{00}^0=0,\\ \tilde{\Gamma}_{ij}^k&=\Gamma_{ij}^k\circ \pi,\end{aligned}$$ for $i,j,k\in\{1,2\}$ where $\Gamma_{ij}^k$ are the Christoffel symbols of $(f_1(t), f_2(t))$ with respect to $\nabla^t$. The lemma now follows from an easy computation. \[lemma curv in cyl\]For all $\varphi\in\Gamma(\Sigma^+(\mathcal Z,g_\mathcal{Z}))$ and all horizontal vector fields $Y\in\Gamma(TS^3)$ it holds that $$\begin{aligned} R^{\Sigma^+(\mathcal Z,g_\mathcal{Z})}(\nu,K)\varphi&=\frac{1}{2}\left(\frac{\alpha''(t)}{\alpha(t)}+\frac{\alpha'(t)}{\beta(t)^2}a-\frac{\alpha(t)\beta'(t)}{\beta(t)^3}a\right)\nu\cdot K\cdot\varphi,\\ R^{\Sigma^+(\mathcal Z,g_\mathcal{Z})}(\nu,Y)\varphi&=\frac{1}{2}\left(\frac{\beta''(t)}{\beta(t)}-\frac{1}{2}\frac{\alpha'(t)}{\beta(t)^2}a+\frac{1}{2}\frac{\alpha(t)\beta'(t)}{\beta(t)^3}a\right)\nu\cdot Y\cdot\varphi\end{aligned}$$ where $(\mathcal{Z},g_\mathcal{Z})$ is with respect to . We use . As local orthonormal frame for $\mathcal{Z}$ we choose $(\nu,e_0(.), e_1(.), e_2(.))$. Furthermore, we use the notation $$R_{X,Y,Z,W}:=g_{\mathcal{Z}}(R^\mathcal Z(X,Y)Z,W)$$ With the aid of Lemma \[lemma z1\] and - it follows from straight forward calculations (for details we refer to [@JW Lemma 6.9]), that $$\begin{aligned} R_{\nu,e_0,e_0,e_1}&=R_{\nu,e_0,e_0,e_2}=R_{\nu,e_0,\nu,e_1}=R_{\nu,e_0,\nu,e_2}=0,\\ R_{\nu,e_0,\nu,e_0}&=\frac{\alpha''(t)}{\alpha(t)},\\ R_{\nu,e_0,e_1,e_2}&=\frac{\alpha'(t)}{\beta(t)^2}a-\frac{\alpha(t)\beta'(t)}{\beta(t)^3}a.\end{aligned}$$ Plugging this into and using , we get $$\begin{aligned} R^{\Sigma^+\mathcal{Z}}(\nu,e_0)\varphi&=\frac{1}{2}\left(R_{\nu,e_0,\nu,e_0}\nu\cdot e_0\cdot\varphi + R_{\nu,e_0,e_1,e_2}e_1\cdot e_2\cdot\varphi\right)\\ &=\frac{1}{2}\left(\frac{\alpha''(t)}{\alpha(t)}\nu\cdot e_0\cdot\varphi + \left(\frac{\alpha'(t)}{\beta(t)^2}a-\frac{\alpha(t)\beta'(t)}{\beta(t)^3}a\right)\underbrace{e_1\cdot e_2\cdot\varphi}_{=\nu\cdot e_0\cdot \varphi}\right)\\ &=\frac{1}{2}\left(\frac{\alpha''(t)}{\alpha(t)}+\frac{\alpha'(t)}{\beta(t)^2}a-\frac{\alpha(t)\beta'(t)}{\beta(t)^3}a\right)\nu\cdot e_0\cdot\varphi.\end{aligned}$$ From this, the first equation in Lemma \[lemma curv in cyl\] directly follows. The second equation follows from $$R^{\Sigma^+\mathcal{Z}}(\nu,e_i)\varphi=\frac{1}{2}\left(\frac{\beta''(t)}{\beta(t)}-\frac{1}{2}\frac{\alpha'(t)}{\beta(t)^2}a+\frac{1}{2}\frac{\alpha(t)\beta'(t)}{\beta(t)^3}a\right)\nu\cdot e_i\cdot\varphi,$$ $i=1,2$, which is shown with the same method as above. $S^1$-invariant spinors {#s1invspinors} ----------------------- For details concerning this section we refer to [@ABCollaps] and also [@Mor]. Define $(g_t)_{t\in I}$ by . The $S^1$-action on $S^3$ induces an $S^1$-action on $\textup{SO}(S^3,g_t)$ which lifts uniquely to an $S^1$-action on $\textup{Spin}(S^3,g_t)$ as follows: We use the fact that for every Riemannian metric on $S^3$ (respectively, $S^2$) there exists, up to equivalence of reductions, exactly one metric spin structure. Pulling back $\textup{Spin}(S^2,\beta(t)^2g_{FS})$ along the Hopf fibration and enlarging the structure group to $\textup{Spin}(3)$ we get the spin structure on $(S^3,g_t)$, $$\textup{Spin}(S^3,g_t)=\pi^*(\textup{Spin}(S^2,\beta(t)^2g_{FS}))\times_{\textup{Spin}(2)}\textup{Spin}(3).$$ Now we define the $S^1$-action by $$[(x,\sigma),g]\cdot e^{is}:=[(x\cdot e^{is},\sigma),g]$$ for $(x,\sigma)\in\pi^*(\textup{Spin}(S^2,\beta(t)^2g_{FS}))\subset S^3\times \textup{Spin}(S^2,\beta(t)^2g_{FS})$, $g\in\textup{Spin}(3)$, and $e^{is}\in S^1\subset\mathbb{C}$. This action is the desired lift of the $S^1$-action on $\textup{SO}(S^3,g_t)$. Uniqueness follows from the fact that $\textup{Spin}(S^3,g_t)\cong S^3\times\textup{Spin}(3)$ and $\textup{SO}(S^3,g_t)\cong S^3\times SO(3,\mathbb{R})$ are connected. This yields a $S^1$-action on $\Sigma(S^3,g_t)$. Spinors which are invariant under this action are called *$S^1$-invariant*. Denote by $V(t)\subset\Gamma(\Sigma(S^3,g_t))$ the vector space of $S^1$-invariant spinors. For every $\varphi\in V(t)$ we have $$\begin{aligned} \label{eq s1 inv 1} \nabla_K^t\varphi=\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}aK\cdot_t\varphi,\end{aligned}$$ see [@ABCollaps Lemma 4.3.]. The $S^1$-invariant spinors are in one-to-one correspondence to the spinors on the base manifold. To be more precise, by [@ABCollaps Lemma 4.4.] there is an isomorphism of vector spaces $$Q=Q(t)\colon \Gamma(\Sigma(S^2,\beta(t)^2g_{FS}))\rightarrow V(t).$$ The following identities will be used later: For every vector field $X\in\Gamma(TS^2)$ and every spinor $\sigma\in\Gamma(\Sigma(S^2,\beta(t)^2g_{FS}))$ $$\begin{aligned} \label{eq s1 inv 2} \nabla_{X^*}^tQ(\sigma)=Q(\nabla_X^t\sigma)-\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}aX^*\cdot_t Q(\sigma),\end{aligned}$$ $$\begin{aligned} \label{eq s1 inv 3} Q(X\cdot_t \sigma)=X^*\cdot_t Q(\sigma).\end{aligned}$$ In - we denote by $\nabla^t$ the spinorial Levi-Civita connection on $\Sigma(S^3,g_t)$ and $\Sigma(S^2,\beta(t)^2g_{FS})$ respectively and “$\cdot_t$” is the clifford multiplication in the respective spinor bundles. A collapsing theorem -------------------- Our first main result is the following theorem. \[main result 1\] Let $\varepsilon>0$ and $\lambda\in\{\pm1\}$. Write $(g_0,\varphi_0):=(g^\varepsilon,Q(\sigma))$ for $\sigma$ a $\lambda$-Killing spinor on $(S^2,g_{FS}=\frac{1}{4}g_{S^2})$ such that $|\varphi_0|=1$. Then the solution of the spinor flow on $M=S^3$ with initial value $(g_0,\varphi_0)$ is given by - where $b=t_{max}$, $t_{max}\in(0,\infty]$ maximum time of existence (to the right), such that: - If $a\lambda=2$, then: - For $0<\varepsilon<\frac{2}{3}$ we have $t_{max}=\infty$, $\lim\limits_{t\rightarrow \infty}\alpha(t)=0$, and $\lim\limits_{t\rightarrow \infty}\beta(t)=:\beta_\infty>0$. - For $\varepsilon\ge \frac{2}{3}$ we have $\lim\limits_{t\rightarrow t_{max}}\alpha(t)=\lim\limits_{t\rightarrow t_{max}}\beta(t)=0$. - For $\varepsilon =\frac{2}{3}$ we have $t_{max}=12$, $\alpha(t)=\frac{2}{3}\beta(t)$, and $\beta(t)=\frac{1}{6}\sqrt{36-3t}$. - For $\varepsilon =1$ we have $t_{\max}=16$ and $\alpha(t)=\beta(t)=\frac{1}{4}\sqrt{16-t}$. - If $a\lambda=-2$, then: - For every $\varepsilon >0$ we have $t_{max}=\infty$, $\lim\limits_{t\rightarrow \infty}\alpha(t)=0$, and $\lim\limits_{t\rightarrow \infty}\beta(t)=:\beta_\infty>0$. Moreover, in any of the above cases the spinor flow preserves the class of $S^1$-invariant spinors which correspond to Killing spinors on $S^2$. More precisely, $$\varphi_t=Q(\sigma_t)$$ for every $t\in I$ where $\sigma_t$ is a $\frac{\lambda}{\beta(t)}$-Killing spinor on $(S^2,\beta(t)^2g_{FS})$. In the case $a\lambda=2$ the result can be interpreted as follows: If we start with fibers that are sufficiently short ($\varepsilon <\frac{2}{3}$), then the $S^1$-fiber converges to a point under the spinor flow ($\alpha\to 0$), but the complement does not ($\beta\to\beta_\infty>0$). In that sense the $S^1$-principal bundle $S^3$ collapses against its base $S^2$. If we start with fibers that are too long ($\varepsilon >\frac{2}{3}$), then $S^3$ converges to a point under the spinor flow. In the case $a\lambda=-2$ the collapse is independent of the length of the fibers. Now we carry out the steps mentioned in the introduction to prove Theorem \[main result 1\]. \[lemma eq for varphi\_t\]Choose $(g_0,\varphi_0)$ as in Theorem \[main result 1\] and define $(g_t,\varphi_t)_{t\in I}$ by -. Then, for every $t\in I$ and every horizontal vector field $Y\in \Gamma(TS^3)$, it holds that $$\begin{aligned} \nabla^t_K\varphi_t&=\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}aK\cdot_t\varphi_t,\\ \nabla^t_Y\varphi_t&=(\frac{1}{\beta(t)}\lambda-\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}a)Y\cdot_t\varphi_t\end{aligned}$$ where $\nabla^t=\nabla^{\Sigma(S^3,g_t)}$ and “$\cdot_t$” is the Clifford multiplication in $\Sigma(S^3,g_t)$. First of all, from - we get $$\begin{aligned} \nabla^0_K\varphi_0&=\frac{1}{4}\varepsilon aK\cdot_0\varphi_0,\\ \nabla^0_Y\varphi_0&=(\lambda-\frac{1}{4}\varepsilon a)Y\cdot_0\varphi_0,\end{aligned}$$ for every horizontal vector field $Y\in\Gamma(TS^3)$. From we get $\nabla^\mathcal{Z}_\nu e_0=0$ and yields $[\nu,e_0](t,p)=-\frac{\alpha'(t)}{\alpha(t)}e_0(t,x)$. In the following we write $\nabla^{\Sigma^+\mathcal Z}=\nabla^{\Sigma^+{(\mathcal{Z},g_\mathcal{Z})}}$. Using Lemma \[lemma cyl identi\] and Lemma \[lemma curv in cyl\] it follows that $$\begin{aligned} \nabla&^{\Sigma^+\mathcal{Z}}_\nu\left(\nabla^{t}_{e_0(t)}\varphi_t -\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}ae_0(t)\cdot_t\varphi_t\right)\\ &=\nabla^{\Sigma^+\mathcal{Z}}_\nu\left(\nabla^{\Sigma^+\mathcal{Z}}_{e_0}\varphi+\frac{1}{2}\nu\cdot \underbrace{W_t(e_0(t))}_{=-\frac{\alpha'(t)}{\alpha(t)}e_0(t)}\cdot\varphi\right)- \frac{1}{4}a\nabla^{\Sigma^+\mathcal{Z}}_\nu\left(\frac{\alpha(t)}{\beta(t)^2}\nu\cdot e_0\cdot\varphi\right)\\ &=\nabla^{\Sigma^+\mathcal{Z}}_\nu\nabla^{\Sigma^+\mathcal{Z}}_{e_0}\varphi -\frac{1}{2}L_\nu(\frac{\alpha'(t)}{\alpha(t)})\nu\cdot e_0\cdot\varphi-\frac{1}{4}aL_\nu(\frac{\alpha(t)}{\beta(t)^2})\nu\cdot e_0\cdot \varphi\\ &=R^{\Sigma^+\mathcal{Z}}(\nu,e_0)\varphi + \nabla^{\Sigma^+\mathcal{Z}}_{e_0}\underbrace{\nabla^{\Sigma^+\mathcal{Z}}_\nu\varphi}_{=0} + \nabla^{\Sigma^+\mathcal{Z}}_{[\nu,e_0]}\varphi -\left(\frac{1}{2}L_\nu(\frac{\alpha'(t)}{\alpha(t)})+\frac{1}{4}aL_\nu(\frac{\alpha(t)}{\beta(t)^2})\right)\nu\cdot e_0\cdot \varphi\\ &=-\frac{\alpha'(t)}{\alpha(t)}\nabla^{\Sigma^+\mathcal{Z}}_{e_0}\varphi + R^{\Sigma^+\mathcal{Z}}(\nu,e_0)\varphi - \left(\frac{1}{2}(\frac{\alpha'(t)}{\alpha(t)})'+\frac{1}{4}a(\frac{\alpha(t)}{\beta(t)^2})'\right)\nu\cdot e_0\cdot\varphi\\ &=-\frac{\alpha'(t)}{\alpha(t)}\nabla^{\Sigma^+\mathcal{Z}}_{e_0}\varphi + \frac{1}{2}\left(\frac{\alpha''(t)}{\alpha(t)}+\frac{\alpha'(t)}{\beta(t)^2}a-\frac{\alpha(t)\beta'(t)}{\beta(t)^3}a\right)\nu\cdot e_0\cdot\varphi\\ &\hspace{5em}-\left(\frac{1}{2}\frac{\alpha''(t)}{\alpha(t)}-\frac{1}{2}\frac{\alpha'(t)^2}{\alpha(t)^2}+\frac{1}{4}a\frac{\alpha'(t)}{\beta(t)^2}-\frac{1}{2}a\frac{\alpha(t)\beta'(t)}{\beta(t)^3}\right) \nu\cdot e_0\cdot\varphi\\ &=-\frac{\alpha'(t)}{\alpha(t)}\nabla^{\Sigma^+\mathcal{Z}}_{e_0}\varphi + \left(\frac{1}{4}\frac{\alpha'(t)}{\beta(t)^2}a+\frac{1}{2}\frac{\alpha'(t)^2}{\alpha(t)^2}\right)\nu\cdot e_0\cdot\varphi\\ &=-\frac{\alpha'(t)}{\alpha(t)}\left(\nabla^t_{e_0(t)}\varphi_t+\frac{1}{2}\frac{\alpha'(t)}{\alpha(t)}\nu\cdot e_0\cdot\varphi\right)+ \left(\frac{1}{4}\frac{\alpha'(t)}{\beta(t)^2}a+\frac{1}{2}\frac{\alpha'(t)^2}{\alpha(t)^2}\right)\nu\cdot e_0\cdot\varphi\\ &=-\frac{\alpha'(t)}{\alpha(t)}\left(\nabla^{t}_{e_0(t)}\varphi_t -\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}ae_0(t)\cdot_t\varphi_t\right).\end{aligned}$$ We have shown $ \left\{ \begin{array}{lll} \nabla^{\Sigma^+\mathcal{Z}}_\nu&\left(\nabla^{t}_{e_0(t)}\varphi_t -\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}ae_0(t)\cdot_t\varphi_t\right)&=-\frac{\alpha'(t)}{\alpha(t)}\left(\nabla^{t}_{e_0(t)}\varphi_t -\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}ae_0(t)\cdot_t\varphi_t\right),\\ &\nabla^{0}_{e_0(0)}\varphi_0-\frac{1}{4}\frac{\alpha(0)}{\beta(0)^2} ae_0(0)\cdot_0\varphi_0&=0. \end{array} \right.$ This differential equation with initial value has zero as unique solution, so $$\nabla^{t}_{e_0(t)}\varphi_t -\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}ae_0(t)\cdot_t\varphi_t=0$$ on $\mathcal{Z}$. To prove the second equation, we define $\tilde{Y}_t\in \Gamma(TS^3)$ by $\tilde{Y}_t(x):=\frac{1}{\beta(t)}Y(x)$. Using the same ideas as above, we get $$\nabla^t_{\widetilde{Y}_t}\varphi_t-\left(\frac{1}{\beta(t)}\lambda-\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}a\right)\widetilde{Y}_t\cdot_t\varphi_t=0.$$ Using the previous lemma, a straightforward calculation yields the following lemma. (Details can be found in [@JW Lemma 6.14].) \[lemma q-terms 1\]Let $(g_t,\varphi_t)_{t\in I}$ as in Lemma \[lemma eq for varphi\_t\]. For every $t\in I$ it holds that $$\begin{aligned} Q_1(g_t,\varphi_t)(e_0(t),e_0(t))&=-\frac{9}{64}\frac{\alpha(t)^2}{\beta(t)^4}a^2+\frac{1}{2}\frac{\alpha(t)}{\beta(t)^3}a\lambda -\frac{1}{2}\frac{1}{\beta(t)^2}\lambda^2,\\ Q_1(g_t,\varphi_t)(e_1(t),e_1(t))&=\frac{3}{64}\frac{\alpha(t)^2}{\beta(t)^4}a^2-\frac{1}{8}\frac{\alpha(t)}{\beta(t)^3}a\lambda,\\ Q_1(g_t,\varphi_t)(e_2(t),e_2(t))&=Q_1(g_t,\varphi_t)(e_1(t),e_1(t)),\\ Q_1(g_t,\varphi_t)(e_i(t),e_j(t))&=0 \text{ for }i\neq j,\\ Q_2(g_t,\varphi_t)&=0. \end{aligned}$$ First of all, we have $$\frac{\partial}{\partial t}\varphi_t=\nabla^{\Sigma^+(\mathcal{Z},g_\mathcal{Z})}_\frac{\partial}{\partial t}\varphi =0=Q_2(g_t,\varphi_t).$$ Moreover, $|\varphi_t|=1$ for all $t\in I$ follows from the fact that $\nabla^{\Sigma^+(\mathcal{Z},g_\mathcal{Z})}$ is a metric connection. From Lemma \[lemma q-terms 1\] we deduce that $\frac{\partial}{\partial t}g_t=Q_1(g_t,\varphi_t)$ with $g_0=g^\varepsilon$ holds iff $(\alpha,\beta)$ is the solution of the following system of two non-linear ordinary differential equations: $$\begin{aligned} \alpha'(t)&=-\frac{9}{128}\frac{\alpha(t)^3}{\beta(t)^4}a^2+\frac{1}{4}\frac{\alpha(t)^2}{\beta(t)^3}a\lambda -\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}\lambda^2,\\ \beta'(t)&=\frac{3}{128}\frac{\alpha(t)^2}{\beta(t)^3}a^2-\frac{1}{16}\frac{\alpha(t)}{\beta(t)^2}a\lambda,\\ \alpha(0)&=\varepsilon,\\ \beta(0)&=1.\end{aligned}$$ Let $F\colon U:=\mathbb{R}^2\setminus\{(x,y)\in\mathbb{R}^2|\text{ }y=0\}\rightarrow \mathbb{R}^2$ be the vector field associated to that system, i.e. $$F\begin{pmatrix} x \\ y \end{pmatrix}=\begin{pmatrix} -\frac{9}{128}\frac{x^3}{y^4}a^2+\frac{1}{4}\frac{x^2}{y^3}a\lambda -\frac{1}{4}\frac{x}{y^2}\lambda^2 \\ \frac{3}{128}\frac{x^2}{y^3}a^2-\frac{1}{16}\frac{x}{y^2}a\lambda \end{pmatrix},$$ see figure \[fig plot of F from main result 1\]. \ Let $c=(x,y)\colon J\rightarrow U$, $J\subset\mathbb{R}$ interval, be an integral curve of $F$. If there exists $t\in J$ such that $x(t)\neq 0$, then $c(J)$ lies in one quadrant of $\mathbb{R}^2$. Using $$x'(t)=\frac{x(t)}{y(t)^2}\left(-\frac{9}{128}\left(\frac{x(t)}{y(t)}a\lambda\right)^2+\frac{1}{4}\left(\frac{x(t)}{y(t)}a\lambda\right) -\frac{1}{4}\right)$$ and $-\frac{9}{128}z^2+\frac{1}{4}z -\frac{1}{4}<0$ for all $z\in\mathbb{R}$, we get that $x(t)$ is either strictly decreasing or strictly increasing (depending on in which quadrant $c(J)$ lies). It follows that the critical points of $F$ are precisely the points $(0,k)$ for $k\neq0$. Let us now prove the case $a\lambda=-2$. First we show that the integral curves of $F$ remain in certain compact subsets of $U$. To that end, let $$K(v,w):=\{(x,y)\in\mathbb{R}^2\text{ }|\text{ }0\le x\le v, \text{ } w\le y \le w+v-x\}.$$ If $c$ is an integral curve of $F$ as above with $x(l),y(l)>0$ for some $l\in J$, then $c(t)\in K(c(l))$ for all $t\in J$ with $t\ge l$. The idea to prove that is as follows: For every boundary point $(v,w)\in \partial K(c(l))$ with $v>0$ the vector $F(v,w)$ points inside $K(c(l))$. Then the integral curve $c$ can’t leave $K(c(l))$ since its movement is prescribed by $F$. Let $c\colon J\rightarrow U$ be a maximal integral curve of $F$ with $c(0)=(\varepsilon,1)$, $\varepsilon>0$, and $J=[0,t_{max})$. Then $c(J)\subset K(c(0))\subset U$. Therefore, we have $t_{max}=\infty$. Using the Poincaré-Bendixson theorem (we use the version from [@CodLevODE]) and observing that $F$ has no periodic orbits, we get a sequence $(t_n)_{n\in \mathbb{N}}$ such that $t_n\ge 0$, $t_n\to \infty$, and $c(t_n)\to p$ for $n\to \infty$ where $p$ is a critical point of $F$. So we have $p=(0,k)$ for some $k>0$. It follows that $$\lim\limits_{t\to \infty}c(t)=(0,k),$$ since for every $n\in\mathbb{N}$ we have $c([t_n,\infty))\subset K(c(t_n))$. This proves Theorem \[main result 1\] in the case $a\lambda =-2$. The case $a\lambda=2$ can be treated with the same methods, i.e. by showing that integral curves remain in certain compact sets and using the Poincaré-Bendixson theorem. This time, however, we have to consider three different cases depending on the value of $\varepsilon >0$. We briefly outline the proof in this case. Define $$\begin{aligned} K_1(v,w)&:=\{(x,y)\in\mathbb{R}^2\text{ }|\text{ }0\le x\le v,\text{ }\frac32x+w-\frac32v\le y\le w\} \text{ for } 0<v<\frac23 w,\\ K_2(v)&:=\{(x,y)\in\mathbb{R}^2\text{ }|\text{ }0\le x\le v,\text{ } x\le y\le\frac32 x\}\text{ for } v>0,\\ K_3(v,w)&:=\{(x,y)\in\mathbb{R}^2\text{ }|\text{ }0\le x\le v, \frac{w}{v}x\le y\le x\}\text{ for } 0<w<v,\\\end{aligned}$$ see figure \[K123\]. (-0.25,1) .. controls (1.75,4).. (2.75,5.5) .. controls (1.75,5.5).. (-0.25,5.5) .. controls (-0.25,3).. (-0.25,1); (-0.5,0)– (3.75,0) node\[below\] [$x$]{}; (-0.25,-0.25)–(-0.25,6)node\[right\][$y$]{}; (-0.25,0) – (3.75,6); (-0.25,0) – (3.75,4); (-0.25,1) – (2.75,5.5); (-0.25,5.5) – (2.75,5.5); (2.75,5.5) circle (0.03)node\[above\] [$(v,w)$]{}; (-0.05,4.3) node\[right\] [$K_1(v,w)$]{}; (4.25,0)– (8.5,0) node\[below\] [$x$]{}; (4.5,-0.5)–(4.5,6)node\[right\][$y$]{}; (4.5,0) .. controls (5.5,1).. (8,3.5) .. controls (8,4).. (8,10.5/2) .. controls (5.5,1.5).. (4.5,0); (4.5,0) – (8.5,6); (4.5,0) – (8.5,4); (8,3.5) – (8,10.5/2); (6.85,3.62) node\[right\] [$K_2(v)$]{}; (8,0) circle (0.03)node\[below\] [$v$]{}; (9,0)– (13.25,0) node\[below\] [$x$]{}; (9.25,-0.5)–(9.25,6)node\[right\][$y$]{}; (9.25,0) .. controls (11.25,6/8).. (13.25,1.5) .. controls (13.25,2).. (13.25,4) .. controls (12.25,3).. (9.25,0); (9.25,0) – (13.25,6); (9.25,0) – (13.25,4); (9.25,0)– (13.25,1.5); (13.25,1.5) – (13.25,4); (11.45,1.9) node\[right\] [$K_3(v,w)$]{}; (13.25,1.5) circle (0.03); (13.25,1.3) circle (0.00)node\[below\] [$(v,w)$]{}; **Proof for $a\lambda=2$ and $0<\varepsilon<\frac{2}{3}$:** We show as before: If $c=(x,y)\colon J \rightarrow U$ is an integral curve of $F$ with $0<x(l)<\frac{2}{3}y(l)$ and $y(l)>0$ for some $l\in J$, then $c(t)\in K_1(c(l))$ for all $t\in J$ with $t\ge l$. Using the Poincaré-Bendixson theorem as above proves the theorem in this case. **Proof for $a\lambda=2$ and $\frac{2}{3}<\varepsilon<1$:** Again we have: If $c=(x,y)\colon J \rightarrow U$ is an integral curve of $F$ with $0<x(l)<y(l)<\frac{3}{2}x(l)$ for some $l\in J$, then $c(t)\in K_2(x(l))$ for all $t\in J$ with $t\ge l$. Let $c=(x,y)\colon J\rightarrow U$ be a maximal integral curve of $F$ with $c(0)=(\varepsilon,1)$, $\frac{2}{3}<\varepsilon<1$, and $J=[0,t_{max})$. We show $c(t)\xrightarrow{t\to t_{max}}0$. It holds that $c(J)\subset K_2(x(0))$. Because of the Poincaré-Bendixson theorem, there exists no $\delta>0$ such that $c(J)\subset K_2(x(0))\cap \{(x,y)\in\mathbb{R}^2\text{ }|\text{ }x\ge\delta\}$. Together with the fact that $x(t)$ is strictly decreasing we get $x(t)\xrightarrow{t\to t_{max}}0$ and therefore $c(t)\xrightarrow{t\to t_{max}}0$. This completes the proof in that case. **Proof for $a\lambda=2$ and $\varepsilon>1$:** If $c=(x,y)\colon J \rightarrow U$ is an integral curve of $F$ with $0<y(l)<x(l)$ for some $l\in J$, then $c(t)\in K_3(c(l))$ for all $t\in J$ with $t\ge l$. Now proceed as in the case $\frac{2}{3}<\varepsilon<1$. It remains to prove the last statement of the theorem. Let $\widetilde{\textup{GL}}^+S^3$ be the topological spin structure on $S^3$. For fixed $e^{is}\in S^1$ we construct a spin-diffeomorphism $F\colon \widetilde{\textup{GL}}^+S^3\rightarrow \widetilde{\textup{GL}}^+S^3$ which restricts to the action of $e^{is}$ on $\textup{Spin}(S^3,g_t)$ defined in Section \[s1invspinors\], for every $t\in I$. (We use the definition of “spin-diffeomorphism” which is given in [@AWWsflow Section 4.1].) Then we use the diffeomorphism invariance of the spinor flow (see [@AWWsflow Corollary 4.5. (ii)]) together with the uniqueness of the solution finish the proof. The $S^1$-action on $S^3$ induces an $S^1$-action on $\textup{GL}^+S^3$ which lifts uniquely to a $S^1$-action on $\widetilde{\textup{GL}}^+S^3$. This can be shown as in the case of metric spin structures (see Section \[s1invspinors\]) using topological spin structures instead. From the action of $e^{is}$ on $S^3$, $\widetilde{\textup{GL}}^+S^3$ and on $\textup{Spin}(S^3,g_t)$ (the latter is defined in Section \[s1invspinors\]) we then get maps $$F\colon \widetilde{\textup{GL}}^+S^3\rightarrow \widetilde{\textup{GL}}^+S^3,$$ $$F_t\colon \textup{Spin}(S^3,g_t)\rightarrow \textup{Spin}(S^3,g_t),$$ $$f\colon S^3\rightarrow S^3$$ where $F$ is a spin-diffeomorphism. The $S^1$-actions on $\textup{GL}^+S^3$ and $\textup{SO}(S^3,g_t)$ coincide on $\textup{SO}(S^3,g_t)\subset\textup{GL}^+S^3$. Combining that with the uniqueness of the $S^1$-action on $\textup{Spin}(S^3,g_t)$ we get $$F|_{\textup{Spin}(S^3,g_t)}=F_t$$ for every $t\in I$. Using [@AWWsflow Section 4.1] we get a map $$F_*\colon \mathcal{N}\rightarrow\mathcal{N},\hspace{3em} \mathcal{N}_g\ni\varphi\mapsto F_*\varphi\in\mathcal{N}_{(f^{-1})^*g},$$ defined by: If locally $\varphi=[\tilde{s},\tilde{\varphi}]$, then $F_*\varphi=[F\circ\tilde{s}\circ f^{-1},\tilde{\varphi}\circ f^{-1}]$. From the definitions it follows that for every $t\in I$ and every spinor $\varphi\in\mathcal{N}_{g_t}$ $$\begin{aligned} \label{eq z1} \left(F_*\varphi\right)(x)=\varphi(x\cdot e^{-is})\cdot e^{is}.\end{aligned}$$ Now let $(g_t,\varphi_t)_{t\in I}$ be the solution of the spinor flow with initial value as in Theorem \[main result 1\]. By the diffeomorphism invariance of the spinor flow, $((f^{-1})^*g_t,F_*\varphi_t)_{t\in I}$ is also a solution. We have $(f^{-1})^*g_0=(f^{-1})^*g^\varepsilon=g^\varepsilon$ and from we get $F_*\varphi_0=\varphi_0$. Because of the uniqueness of the solution of the spinor flow it follows that $$F_*\varphi_t=\varphi_t$$ for every $t\in I$. Using again and noting that $e^{is}\in S^1$ was arbitrary, we see that $\varphi_t$ is $S^1$-invariant, i.e. $\varphi_t\in V(t)$. Define $\sigma_t\in\Gamma(\Sigma(S^2,\beta(t)^2g_{FS}))$ by $Q(\sigma_t)=\varphi_t$. Combining the second equation in Lemma \[lemma eq for varphi\_t\] with and yields that $\sigma_t$ is a $\frac{\lambda}{\beta(t)}$-Killing spinor. This finishes the proof of the theorem. \[rem change or conv\] If we choose the other orientation on $S^3$ (i.e. the orientation that satisfies: If $(f_1,f_2)$ is any oriented local orthonormal frame on $(S^2,g_{FS})$, then $(-K,f_1^*,f_2^*)$ is an oriented local orthonormal frame on $(S^3,g_{S^3})$), then Theorem \[main result 1\] still holds if we switch the results for the cases “$a\lambda=2$” and “$a\lambda=-2$”. This can be seen as follows: In Lemma \[lemma curv in cyl\] one has to replace “$a$” by “$-a$”. The additional sign enters because in the proof we used . For the same reason we have to replace “$a$” by “$-a$” in -. Then we also need to replace “$a$” by “$-a$” in Lemma \[lemma eq for varphi\_t\] and Lemma \[lemma q-terms 1\] and therefore also in Theorem \[main result 1\]. A stability theorem ------------------- For $\varepsilon>0$ we define $c(\varepsilon)>0$ by $\textup{vol}(S^3,c(\varepsilon)g^\varepsilon)=1$, i.e. $c(\varepsilon)=(\frac{1}{2\pi^2\varepsilon})^{\frac{2}{3}}$. Our second main result is the following theorem. (Stability)\[main result 2\] Let $\varepsilon>0$ and $\mu\in\{\pm\frac{1}{2}\}$. Moreover, let $g_0:=c(\varepsilon)g^\varepsilon$ and $\varphi_0$ a spinor which is obtained via parallel transport of a $\mu$-Killing spinor from $(S^3,g_{S^3})$ to $(S^3,g^\varepsilon)$ such that $|\varphi_0|= 1$. (In Remark \[rmk main res 2\] we will explain how $\varphi_0$ is defined in a more formal way.) Then the solution of the normalized spinor flow on $M=S^3$ with initial value $(g_0,\varphi_0)$ is given by - where $b=t_{max}$, $t_{max}\in(0,\infty]$ maximum time of existence (to the right), such that: - If $a\mu=1$, then: - For $0<\varepsilon<\frac{2}{3}$ we have $\lim\limits_{t\rightarrow t_{max}}\alpha(t)=0$ and $\lim\limits_{t\rightarrow t_{max}}\beta(t)=\infty$. - For $\varepsilon =\frac{2}{3}$ we have $t_{max}=\infty$, $\alpha(t)\equiv\frac{2}{3}\sqrt{c(\frac{2}{3})}$, and $\beta(t)\equiv\sqrt{c(\frac{2}{3})}$. In particular, $(c(\frac23)g^\frac23,\varphi_0)$ is a critical point of the normalized spinor flow. - For $\varepsilon >\frac{2}{3}$ we have $t_{max}=\infty$ and $\lim\limits_{t\rightarrow \infty}\alpha(t)=\lim\limits_{t\rightarrow \infty}\beta(t)=\sqrt{c(1)}$. Moreover, there exist smooth functions $f,g\colon I\rightarrow (0,\infty)$ with $\lim\limits_{t\rightarrow \infty}f(t)=\lim\limits_{t\rightarrow \infty}g(t)=\frac{\mu}{\sqrt{c(1)}}$ and $$\nabla^t_K\varphi_t=f(t)K\cdot_t\varphi_t,$$ $$\nabla^t_Y\varphi_t=g(t)Y\cdot_t\varphi_t,$$ for every horizontal vector field $Y\in\Gamma(TS^3)$ and every $t\in I$. - For $\varepsilon =1$ we have $t_{max}=\infty$ and $\alpha(t)\equiv\beta(t)\equiv\sqrt{c(1)}$. In particular, $(c(1)g_{S^3},\varphi_0)$ is a critical point of the normalized spinor flow. - If $a\mu=-1$, then: - For every $\varepsilon >0$ we have $t_{max}=\infty$ and $\lim\limits_{t\rightarrow \infty}\alpha(t)=\lim\limits_{t\rightarrow \infty}\beta(t)=\sqrt{c(1)}$. Moreover, there exist smooth functions $f,g\colon I\rightarrow (0,\infty)$ with $\lim\limits_{t\rightarrow \infty}f(t)=\lim\limits_{t\rightarrow \infty}g(t)=\frac{\mu}{\sqrt{c(1)}}$ and $$\nabla^t_K\varphi_t=f(t)K\cdot_t\varphi_t,$$ $$\nabla^t_Y\varphi_t=g(t)Y\cdot_t\varphi_t,$$ for every horizontal vector field $Y\in\Gamma(TS^3)$ and every $t\in I$. - For $\varepsilon =1$ we have $t_{max}=\infty$ and $\alpha(t)\equiv\beta(t)\equiv\sqrt{c(1)}$. In particular, $(c(1)g_{S^3},\varphi_0)$ is a critical point of the normalized spinor flow. \[rmk main res 2\] 1. In the case $a\mu=1$ we can interpret the result as follows: If we are not too far away ($\varepsilon >\frac23$) from the normalized standard metric $c(1)g_{S^3}$ together with a Killing spinor, then the metric part of the solution flows back to the normalized standard metric ($\alpha,\beta\to\sqrt{c(1)}$) and the spinor part of the solution flows back to a Killing spinor ($f,g\to\frac{\mu}{\sqrt{c(1)}}$). However, if we are too far away ($\varepsilon \le \frac23$), then the solution no longer flows back ($\alpha\to 0$, $\beta\to \infty$). 2. If we choose $g_0=g^\varepsilon$ and $\varphi_0$ as in Theorem \[main result 2\], then $(g_0,\varphi_0)$ converges to a point under the unnormalized spinor flow (for $a\mu=-1$ or $a\mu=1$ and $\varepsilon\ge\frac{2}{3}$), see [@JW Theorem 6.17]. In that sense the interesting behavior is only captured in the normalized flow. 3. In the following we make precise what we mean by $\varphi_0$ in Theorem \[main result 2\]. To that end, let $\sigma_0\in\Gamma(\Sigma(S^3,g_{S^3}))$ be a $\mu$-Killing spinor with $|\sigma_0|=1$. **Case $\varepsilon >1$:** Define $\sigma_t$ by $\eqref{eq ansatz for varphi_t}$ where $\alpha(t)=1+t$, $\beta(t)\equiv 1$ and $b=\varepsilon -1$. Set $$\varphi_0:=\sigma_{\varepsilon-1}\in\Gamma(\Sigma(S^3,g^\varepsilon)).$$ **Case $\varepsilon <1$:** Define $\sigma_t$ by $\eqref{eq ansatz for varphi_t}$ where $\alpha(t)=1-t$, $\beta(t)\equiv 1$, and $b=1-\varepsilon$. Set $$\varphi_0:=\sigma_{1-\varepsilon}\in\Gamma(\Sigma(S^3,g^\varepsilon)).$$ **Case $\varepsilon =1$:** Simply set $\varphi_0:=\sigma_0$. Note that $\Sigma(S^3,g^\varepsilon)=\Sigma(S^3,c(\varepsilon)g^\varepsilon)$, so $\varphi_0\in\Gamma(\Sigma(S^3,c(\varepsilon)g^\varepsilon))$. \[lemma eq for varphi\_t 2\] Choose $(g_0,\varphi_0)$ as in Theorem \[main result 2\] and define $(g_t,\varphi_t)_{t\in I}$ by -. Then, for every $t\in I$ and every horizontal vector field $Y\in \Gamma(TS^3)$, it holds that $$\begin{aligned} \nabla^t_K\varphi_t&=\left(\frac{\mu-\frac{1}{4}a}{\alpha(t)}+\frac{1}{4}\frac{\alpha(t)}{\beta(t)^2}a\right)K\cdot_t\varphi_t,\\ \nabla^t_Y\varphi_t&=\frac{1}{\beta(t)}\left(-\frac{1}{4}a\left(\frac{\alpha(t)}{\beta(t)}-1\right)+\mu\right)Y\cdot_t\varphi_t. \end{aligned}$$ First, these equations hold for $t=0$. (This can be shown as follows: Let $\sigma_t$ be as in Remark \[rmk main res 2\]. We can deduce equations for $\nabla^t\sigma_t$ similar to the proof of Lemma \[lemma eq for varphi\_t\]. Evaluating the equations for $\nabla^t\sigma_t$ at $t=\varepsilon -1$ and $t=1-\varepsilon$, respectively, yields the equations of Lemma \[lemma eq for varphi\_t 2\] for $t=0$, see also [@JW Lemma 6.21].) Using the same method as in the proof of Lemma \[lemma eq for varphi\_t\] yields the equations for all $t\in I$. Using Lemma \[lemma eq for varphi\_t 2\] one easily proves the following lemma. \[lemma q-terms 2\]Let $(g_t,\varphi_t)_{t\in I}$ as in Lemma \[lemma eq for varphi\_t 2\]. For every $t\in I$ it holds, that $$\begin{aligned} Q_1(g_t,\varphi_t)(e_0(t),e_0(t))&=\frac{1}{4}\frac{1}{\alpha(t)^2}\left(\mu-\frac{1}{4}a\right)^2+\frac{1}{\beta(t)^2}\left(-\frac{1}{2}\mu^2-\frac{3}{8}a\mu\right)\\ &\hspace{2em}+\frac{\alpha(t)}{\beta(t)^3}\left(\frac{1}{8}a^2+\frac{1}{2}a\mu\right)-\frac{9}{64}\frac{\alpha(t)^2}{\beta(t)^4}a^2,\\ Q_1(g_t,\varphi_t)(e_1(t),e_1(t))&=-\frac{1}{4}\frac{1}{\alpha(t)^2}\left(\mu-\frac{1}{4}a\right)^2+\frac{\alpha(t)}{\beta(t)^3}\left(-\frac{1}{32}a^2-\frac{1}{8}a\mu\right)+\frac{3}{64}\frac{\alpha(t)^2}{\beta(t)^4}a^2,\\ Q_1(g_t,\varphi_t)(e_2(t),e_2(t))&=Q_1(g_t,\varphi_t)(e_1(t),e_1(t)),\\ Q_1(g_t,\varphi_t)(e_i(t),e_j(t))&=0 \text{ for }i\neq j,\\ Q_2(g_t,\varphi_t)&=0,\\ \frac16 \frac{1}{\textup{vol}(S^3,g_t)}\mathcal{E}(g_t,\varphi_t)&=\frac{1}{12}(\mu-\frac{1}{4}a)^2\frac{1}{\alpha(t)^2}+ \frac{1}{64}a^2\frac{\alpha(t)^2}{\beta(t)^4}+\frac{1}{24}a(\mu-\frac{1}{4}a)\frac{1}{\beta(t)^2}\\ &\hphantom{=}+\frac{1}{6}(\frac{1}{4}a+\mu)^2\frac{1}{\beta(t)^2}-\frac{1}{12}a(\frac{1}{4}a+\mu)\frac{\alpha(t)}{\beta(t)^3}. \end{aligned}$$ From Lemma \[lemma q-terms 2\] we get that $\frac{\partial}{\partial t}g_t=\tilde{Q}_1(g_t,\varphi_t)$ with $g_0=c(\varepsilon)g^\varepsilon$ is equivalent to the following systems of two non-linear ordinary differential equations for $(\alpha,\beta)$ with initial values $\alpha(0)=\sqrt{c(\varepsilon)}\varepsilon$, $\beta(0)=\sqrt{c(\varepsilon)}$: For $a \mu=1$: $$\begin{aligned} \alpha'(t)&=-\frac{1}{4}\frac{\alpha(t)^3}{\beta(t)^4}+\frac{5}{12}\frac{\alpha(t)^2}{\beta(t)^3}-\frac{1}{6}\frac{\alpha(t)}{\beta(t)^2},\\ \beta'(t)&=\frac{1}{8}\frac{\alpha(t)^2}{\beta(t)^3}-\frac{5}{24}\frac{\alpha(t)}{\beta(t)^2}+\frac{1}{12}\frac{1}{\beta(t)},\\ \end{aligned}$$ and for $a\mu=-1$: $$\begin{aligned} \alpha'(t)&=-\frac{1}{4}\frac{\alpha(t)^3}{\beta(t)^4}+\frac{1}{12}\frac{\alpha(t)}{\beta(t)^2}+\frac{1}{6}\frac{1}{\alpha(t)},\\ \beta'(t)&=\frac{1}{8}\frac{\alpha(t)^2}{\beta(t)^3}-\frac{1}{12}\frac{\beta(t)}{\alpha(t)^2}-\frac{1}{24}\frac{1}{\beta(t)}. \end{aligned}$$ Denote by $F=F(x,y)$ the vector field associated to these differential equations as in the proof of Theorem \[main result 1\]. First, we show that it suffices to restrict $F$ to a certain 1-dimensional submanifold of $\mathbb{R}^2$. To that end, define $$u\colon (0,\infty)\rightarrow \mathbb{R}^2,\hspace{3em} t\mapsto \left( \sqrt{c(t)}t,\sqrt{c(t)} \right)=\left((2\pi^2)^{-\frac{1}{3}}t^{\frac{2}{3}},(2\pi^2)^{-\frac{1}{3}}t^{-\frac{1}{3}}\right).$$ The image of $u$ is an embedded submanifold of $\mathbb{R}^2$ and precisely the set of the initial values we are interested in. Noting that $$F(u(t))=k(t)u'(t),$$ with $$k(t)=\left\{\begin{array}{ll} \frac{1}{8}(2\pi^2)^{\frac{2}{3}}t^{\frac{5}{3}}(-3t^2+5t-2), & \text{for }a\mu=1, \\ \frac{1}{8}(2\pi^2)^{\frac{2}{3}}t^{-\frac{1}{3}}(-3t^4+t^2+2), & \text{for }a\mu=-1,\end{array}\right.$$ we see that $F$ is tangent to the image of $u$. So we only have to understand the integral curves of $F|_{\textup{Im}(u)}$. One way to do this is to look at the corresponding vector field on $(0,\infty)$ which is given by $d(u^{-1})(V\circ u)$. Since $d(u^{-1})(V\circ u)=k(t)$ the claims about $\alpha$, $\beta$ and $t_{max}$ of Theorem \[main result 2\] follow easily. The claimed convergence of the functions $f$ and $g$ follows from Lemma \[lemma eq for varphi\_t 2\] and the shown convergence of $\alpha$ and $\beta$. This finishes the proof of the theorem.

ArXiv

--- abstract: | Accretion disks in binary systems can experience hydrodynamic impact at inner as well as outer edges. The first case is typical for protoplanetary disks around young T Tau stars. The second one is typical for circumstellar disks in close binaries. As a result of such an impact, perturbations with different scales and amplitudes are excited in the disk. We investigated the nonlinear evolution of perturbations of a finite, but small amplitude, at the background of sub-Keplerian flow. Nonlinear effects at the front of perturbations lead to the formation of a shock wave, namely the discontinuity of the density and radial velocity. At this, the tangential flow in the neighborhood of the shock becomes equivalent to the flow in in the boundary layer. Instability of the tangential flow further leads to turbulization of the disk. Characteristics of the turbulence depend on perturbation parameters, but $\alpha$-parameter of Shakura-Sunyaev does not exceed $\sim 0.1$. PACS numbers: 47.27.-i, 97.10.Gz, 97.80.-d author: - 'E. P. Kurbatov[^1]' - 'D. V. Bisikalo[^2]' bibliography: - 'paper.bib' title: 'Excitation of turbulence in accretion disks of binary stars by non-linear perturbations' --- Introduction ============ Turbulent viscosity is the most effective mechanism of the angular momentum transfer in astrophysical disks . The turbulence is caused by development of instability of small perturbations. In accretion disks of nonmagnetic stars, as well as in the protoplanetary disks (in cold non-ionized gas), the turbulence can develop from instabilities of the hydrodynamic type only. The stability of linear hydrodynamic perturbations in Keplerian disks is studied in numerous papers (see, for instance, references in [@Godon1999ApJ...521..319G]). Both radial and non-axisymmetric perturbations of the modal type turn out to be stable according to the Rayleigh criterion [@Rayleigh1917RSPSA..93..148R; @Landau1987theorphys-6]. The reason for this is spectral stability of the Keplerian flow. There are non-modal type perturbations or transient ones, which display a growth in the linear stage (see a review in [@Razdoburdin2015PhyU...58.1031R]). However, the problem of the mechanism of their continuous generation in the Keplerian flow, as well as the problem of efficiency of removal of angular momentum, on large time scales needs further study. There are at least two classes of astrophysical disks in which accretion flow experiences an external action: the disks around components of close binary stars and disks around young stellar systems (protoplanetary ones). In the objects of the first kind, the stream from the donor-star acts upon the outer part of the accretion disk, generating a tangential discontinuity and associated with the latter shock wave, the so-called  [@Bisikalo1998MNRAS.300...39B; @Boyarchuk2002mtcb.book.....B]. In the objects of the second kind act detached shocks from the components of the binary system [@Kaygorodov2010] or shockless perturbations, if the star moves with a subsonic velocity [@Sytov2016AstRep]. The last affect the protoplanetary disk at the inner side. As a result, nonlinear waves of different amplitude and scale may be exited in the accretion disk. Although in a certain conditions the nonlinear waves may provide a conducive background for linear instability and the turbulence formation [@Kurbatov2014], it is not the case in the protoplanetary disks, since the nonlinear perturbations are too weak. In hydrodynamics, the mechanism of formation of discontinuities in nonlinear waves is known [@Zeldovich1966shockwaves.book; @Landau1987theorphys-6]. This phenomenon is due to the “breaking” of the wave profile. If the discontinuity is a shock wave, dissipative effects will lead to the damping of the perturbation with time. On the other hand, in the accretion flows the tangential velocity of a gas can many times larger than the sound velocity. Therefore, it can not be ruled out that, for certain perturbation parameters the rollover of the profile may cause a tangential discontinuity. In this case, the difference of energy at different sides of the discontinuity will transform into kinetic energy of the turbulence. In the present paper, we investigate whether evolution of nonlinear waves in the accretion disk may lead to the appearance of unstable configurations and, as a result, to turbulence. we We perform an analysis of nonlinear perturbations in sub-Keplerian disks with inhomogeneous distribution of density and pressure. The paper is organized as follows: in the Section 2 we describe the structure of accretion flow in semi-detached and young binary systems. In Section 3 we present the model of an equilibrium sub-Keplerian disk and a briefly analyze linear perturbations in the latter. In the fourth Section we explore evolution of nonlinear radial perturbations in accretion disks. In Section 5 we apply our results to two simple models of accretion disks. Conclusions and deductions are presented in the last Section. The structure of accretion flows in young binaries with circumstellar envelope and in close binaries ==================================================================================================== Accretion disks in binary systems have the distribution of angular momentum close to the Keplerian one. Velocity of tangential motion in them can be hundreds times larger than the sound velocity (in the central regions). Often, accretion disks experience hydrodynamic action from accretion flows, encounter with the streams of incoming gas or detached shock waves. Such effects, along to large tangential velocities, can be potential sources of turbulence. Below, we will consider the flow patterns for two classes of typical binary stars: young stars surrounded by a protoplanetary disk, and close binary systems. The envelope of a T Tauri binary star is a remnant of the protostellar cloud from which the matter falls and forms a Keplerian protoplanetary disk. In the central part of the disk around the binary forms a gas-free region, a “gap”. The reason for the existence of the gap can be both tidal action (Lindblad resonances) of the binary [@Artymowicz1994ApJ...421..651A] and the effects of the detached shock waves produced by the components of the system [@Kaygorodov2010]. Typical gap width is of the order of several orbital separations (see, e.g., [@Jensen1997AJ....114..301J; @Najita2003ApJ...589..931N]). Inside the gap, around each of the stars forms its own circumstellar accretion disk and in the case of the supersonic orbital motion of components (which is typical of such systems), detached shock waves appear in the front of the stars . In the region of the inner disk boundary, cold and dense gas of the Keplerian flow is contiguous to the hot and rarefied gas of the gap. Keplerian velocity in this region can exceed the sound velocity by many dozen times. At the same time, the impact of the detached shock wave upon the disk is concentrated on a small part of the boundary. Since the impact has a shock nature, and it is quite intense, it is obvious that it will result in destruction of the Keplerian flow in the vicinity of the contact. Blurring of this region by differential rotation of the gas can lead to the formation of a turbulent layer over a time of the order of one revolution of the disk at the radius of the inner boundary. In close binary systems, one of the components (the donor) fills its Roche lobe. This results in the flow of the matter onto the second component (accretor). Under the influence of the dissipative processes, a disk forms around the accretor. The former has a Keplerian angular momentum distribution. Typical is the case when the accretor is more massive and is a $\sim 1~M_\odot$ white dwarf, while the mass of the donor is lower. The radius of the accretion disk comprises the tenth of the orbital separation, while velocity of Keplerian motion even at the outer edge of the disk can exceed the sound speed be several tens of times. The region of the outer edge of the disk where the stream from the donor interacts with accretion disk, is a tangential discontinuity [@Bisikalo1998MNRAS.300...39B; @Boyarchuk2002mtcb.book.....B; @Bisikalo2003AstRep-hotline]. The step of tangential velocity at the discontinuity can be of the order of the Keplerian velocity at the edge of the disk. It is known that such discontinuities are unstable for many types of perturbations [@Fridman2008PhyU...51..213F]. Because of the high flow velocities, the instability must have a drift character and manifest itself as a turbulent wake from the tangential discontinuity in the direction of rotation of the gas. In both examples considered above, the turbulence is excited in the Keplerian flow: in the inner region of the protoplanetary disks and in the outer region of the circumstellar disk in semi-detached binary systems. The turbulence, however, engulfs only a small part of the disk flow. The problem of possible further propagation of turbulence into entire volume of the accretion disk is of particular interest, since if this is a case, existing nonlinear perturbations can be considered as the source of turbulence of nonmagnetic accretion disks. The effect of detached shock waves upon internal boundary of the protoplanetary disk should not only lead to the destruction of the boundary Keplerian flow, but also to the propagation of the perturbations with different amplitudes and scales inside the disk. A similar situation should take place also in the disks of close binaries — slight perturbations of the rate of mass loss by the donor star or characteristics of the flow at the edge of the accretion disk (for example, due to the intrinsic precession of the accretion disk) can cause changes in the position and intensity of the tangential discontinuity, which, in turn, will disturb the accretion flow. In the Introduction, we mentioned that small amplitude perturbations are either stable (for instant, radial ones are stable by the Rayleigh criterion [@Balbus1996ApJ...467...76B]), or display only a weak instability (some nonaxisymmetric perturbations are unstable [@Razdoburdin2015PhyU...58.1031R], but they have rather large time scale of growth). On the contrary, non-linear waves can participate in the universal scenario of generation of shock waves and tangential discontinuities which is realized in any environment with ordinary equation of state [@Zeldovich1966shockwaves.book; @Landau1987theorphys-6]. At the conditions of fast shear flow these discontinuities will be unstable and will decay directly, generating turbulence. A nonlinear wave is a propagating perturbation of density, pressure, and velocity, with a finite amplitude. Two factors are responsible for the formation of a discontinuity in a nonlinear wave: (i) in the areas of enhanced pressure, propagation velocity of perturbations is higher than in the low-pressure regions; (ii) velocity of perturbations is added with the velocity of the background flow. So, if consider the leading edge of a nonlinear wave, where the pressure and velocity vary from the maximum to the minimum values, the region of the maximum will catch up with the minimum region. This will lead to the steepening of the wave profile and to its “rollover” and formation of a discontinuous solution in a finite time. This pattern becomes somewhat more complicated if one takes into account the differential rotation of the disk matter. For instance, the wave solution for small radial perturbations exists only in the region of the Keplerian disk, where the frequency of the wave is not less than the angular velocity of the matter of the disk [@Goldreich1979ApJ...233..857G]. However, we will show below that the scenario of formation of discontinuities can take place in the rotating flows too. Next, we will perform a brief analysis of the linear perturbations in the accretion disks and will consider the mechanism of formation of discontinuities in the nonlinear perturbations in such disks. Linear perturbations in the accretion disk ========================================== Equilibrium disk {#sec:equillibrium} ---------------- Let consider a stationary disk flow, symmetric respective to the $r = 0$ axis and $z = 0$ plane. The equation of radial balance of forces in the mid-plane of the disk: $$\frac{1}{\rho_0}\,{\frac{\partialp_0}{\partialr}} = - r \Omega_\mathrm{K}^2 + \frac{\lambda_0^2}{r} \;,$$ where $\rho_0$, $p_0$, $\lambda_0$ are, respectively, density, pressure, and angular momentum in the $z = 0$ plane; $\Omega_\mathrm{K}$ is Keplerian angular velocity $$\Omega_\mathrm{K} = \Omega_{\mathrm{K}\ast} \left( \frac{r}{r_\ast} \right)^{-3/2} \;.$$ By  we will mark the values of variables in the point $r_\ast$ at $z = 0$. In particular, $\Omega_{\mathrm{K}\ast} = (G M / r_\ast^3)^{1/2}$. The gas pressure usually decreases with the radius. Therefore the centrifugal acceleration is not exactly equal to the gravitational one, but is somewhat lower . Let the disk to be sub-Keplerian, if the balance of gravitational and centrifugal force is violated by a small amount [@Armitage2007astro.ph..1485A]: $$\label{eq:subkeplerian_disk} r \Omega_\mathrm{K}^2 - \frac{\lambda_0^2}{r^3} = \chi\,\frac{H^2}{r^2}\,r \Omega_\mathrm{K}^2 \;,$$ where $H$ is semi-thickness of the disk; $\chi$ — a dimansionless parameter. Let distribution of the matter to be isentropic with the exponent $\gamma$, distribution of density $\rho_0$ on radius is a power law with exponent $d$. Then, for the pressure, density and sound velocity one may write down $$\begin{gathered} \rho_0(r) = \rho_\ast \left( \frac{r}{r_\ast} \right)^{-d} \;, \\ p_0 = p_\ast \left( \frac{\rho_0}{\rho_\ast} \right)^\gamma = p_\ast \left( \frac{r}{r_\ast} \right)^{-\gamma d} \;, \\ c_0^2(r) = c_\ast^2 \left( \frac{r}{r_\ast} \right)^{-(\gamma-1)d} \;,\end{gathered}$$ where $c_\ast^2 = \gamma p_\ast / \rho_\ast$. Then the equation of the radial balance of forces becomes $$\label{eq:radial_balance} d c_\ast^2 \left( \frac{r}{r_\ast} \right)^{-(\gamma-1)d+3} = \chi \Omega_{\mathrm{K}\ast}^2 H^2(r) \;.$$ Let assume that the vertical structure of the disk is isothermal. In a thin disk, $H \ll r$, the density is distributed as $$\rho(r, z) = \rho_0(r)\,\operatorname{exp}\left[ - \frac{z^2}{2 H^2(r)} \right] \;,$$ where semi-thickness of the disk is determined from the condition of the vertical balance of forces $$\label{eq:vertical_balance} H^2(r) = \frac{c_0^2(r)}{\Omega_\mathrm{K}^2(r)} = \frac{c_\ast^2}{\Omega_{\mathrm{K}\ast}^2} \left( \frac{r}{r_\ast} \right)^{-(\gamma-1)d+3} \;.$$ From this expression, using Eq. (\[eq:radial\_balance\]) we may derive the factor $\chi$ in the Eq. (\[eq:subkeplerian\_disk\]): $$\chi = d \;.$$ For the accepted distribution $\rho(r, z)$ the dependence of the surface density on the radius will be a power law: $$\Sigma(r) = 2 \int_0^\infty dz\,\rho(r, z) = \sqrt{2 \pi}\,\frac{\rho_\ast c_\ast}{\Omega_{\mathrm{K}\ast}} \left( \frac{r}{r_\ast} \right)^{[-(\gamma+1)d+3]/2} \;.$$ In the realistic models of the accretion disks, surface density declines with radius, approximately, as $1/r^{0 \dots 1}$ (see [@Armitage2007astro.ph..1485A] and references therein). For instance, in the Shakura-Sunyaev disk model $\Sigma \propto r^{-3/5}$ . In the paper [@Armitage2015arXiv150906382A] as a typical dependence is assumed $\sigma \propto r^{-1}$. In our model, we shall assume $\Sigma \propto r^{-1/2}$. Below, this model will allow to obtain more simple expressions for some of dependences, while remaining qualitatively consistent with more complete models. The slope of the surface density distribution in combination with the value of the adiabatic exponent $\gamma$, determines the values of exponents in power-laws for remaining thermodynamical quantities: volumetric density, pressure, and sound velocity. Further in the present study, we will consider the gas as monatomic. Thus, we obtain $$\gamma = 5/3 \;,\qquad d = 3/2 \;,\qquad (\gamma-1)d = 1 \;.$$ Let write down final expressions for equilibrium distributions: $$\begin{gathered} \rho_0 = \rho_\ast \left( \frac{r}{r_\ast} \right)^{-3/2} \;, \\ c_0 = c_\ast \left( \frac{r}{r_\ast} \right)^{-1/2} \;, \\ \label{eq:subkeplerian_momentum} \lambda_0 = \left( 1 - \frac{3}{2}\,\frac{c_\ast^2}{r_\ast^2 \Omega_{\mathrm{K}\ast}^2} \right)^{1/2} \left( \frac{r}{r_\ast} \right)^{1/2} r_\ast^2 \Omega_{\mathrm{K}\ast} \;, \\ H = \frac{c_\ast}{\Omega_{\mathrm{K}\ast}}\,\frac{r}{r_\ast} \;.\end{gathered}$$ The typical values of the gas velocity in the disks, according to numerical models, are equal to tens of sound velocities [@Bisikalo1998MNRAS.300...39B]. Thus, correction to the Keplerian momentum in expression (\[eq:subkeplerian\_momentum\]) does not exceed one per cent. Linear perturbations {#sec:linear_pert} -------------------- Let consider the dynamics of gas flow in the mid-plane of the disk and confine ourselves to perturbations of the equilibrium configurations that (i) are symmetric about the axis $r = 0$; (ii) do not depend on the vertical coordinate z; (iii) have a zero velocity in the direction of the z-axis. The system of equations that describes such a class of flows has the form: $$\begin{gathered} \label{eq:continuity_general} {\frac{\partial\rho}{\partialt}} + \frac{1}{r}\,{\frac{\partial(r \rho v)}{\partialr}} = 0 \;, \\ \label{eq:radial_momentum_general} {\frac{\partialv}{\partialt}} + v\,{\frac{\partialv}{\partialr}} = - \frac{1}{\rho}\,{\frac{\partialp}{\partialr}} - r \Omega_\mathrm{K}^2 + \frac{\lambda^2}{r^3} \;, \\ \label{eq:angular_momentum_general} {\frac{\partial\lambda}{\partialt}} + v\,{\frac{\partial\lambda}{\partialr}} = 0 \;,\end{gathered}$$ where $v$ is radial velocity; $\lambda$ — angular momentum per unit mass. Let assume that the system of equations (\[eq:continuity\_general\])–(\[eq:angular\_momentum\_general\]) describes a small perturbation of the equilibrium configuration, which was determined in the previous section. The perturbation itself we describe as a radial displacement of gas elements. Let denote by $\xi(t, r)$ the value of radial displacement of the particle whose position at the time $t$ is $r$, i.e., the radial coordinate of the nonmoved particle is $r - \xi(t, r)$. From the conservation of the angular momentum law in form (\[eq:angular\_momentum\_general\]) it follows that the moment of each element of the gas is conserved, i. e., we may write down $$\label{eq:angular_momentum_expansion} \lambda^2(t, r) = \lambda_0^2\bigl(r - \xi(t, r)\bigr) = \lambda_0^2(r) - {\frac{d\lambda_0^2(r)}{dr}}\,\xi(t, r) + \mathcal{O}\!\left( r \Omega_\mathrm{K}^2 \frac{\xi^2}{r^2} \right) \;.$$ Here we took into account that $|\xi| \ll r$. In our model the derivative $d\lambda_0^2/dr$ does not depend on the coordinate and is equal to $${\frac{d\lambda_0^2}{dr}} = \left( 1 - \frac{3}{2}\,\frac{c_\ast^2}{r_\ast^2 \Omega_{\mathrm{K}\ast}^2} \right) r_\ast^3 \Omega_{\mathrm{K}\ast}^2 \;.$$ Let require that perturbations of interest have a small spatial scale of variability. Then we can neglect the  term $\rho v/r$ in the equation of continuity, which is determined by the cylindrical geometry of the problem, and the advective term in (\[eq:continuity\_general\]) will take a Cartesian form. This approximation was used by many authors in the analysis of perturbations in the accretion disks (see, for instance, [@Lubow1993ApJ...409..360L; @Balbus1996ApJ...467...76B]). Let denote perturbed values by an index . Then, after linearization of equation of continuity (\[eq:continuity\_general\]) over small perturbations, we get $$\label{eq:continuity_pert} {\frac{\partial\rho_1}{\partialt}} + {\frac{\partial(\rho_0 v_1)}{\partialr}} = 0 \;.$$ By means of (\[eq:angular\_momentum\_expansion\]) linearization of the Euler equation (\[eq:radial\_momentum\_general\]) will be $$\label{eq:radial_momentum_pert} {\frac{\partialv_1}{\partialt}} + \frac{1}{\rho_0}\,{\frac{\partialp_1}{\partialr}} - {\frac{dp_0}{dr}}\,\frac{\rho_1}{\rho_0^2} + \frac{1}{r^3 }\,{\frac{d\lambda_0^2}{dr}}\,\xi = 0 \;.$$ Angular momentum conservation law was already taken into account explicitly when we employed expansion (\[eq:angular\_momentum\_expansion\]). The gradients of thermodynamical values for adiabatic equation of state have the form: $$\begin{gathered} {\frac{dp_0}{dr}} = c_0^2\,{\frac{d\rho_0}{dr}} \;, \\ p_1 = c_0^2 \rho_1 \;, \\ {\frac{dc_0^2}{dr}} = (\gamma - 1)\,\frac{c_0^2}{\rho_0}\,{\frac{d\rho_0}{dr}} \;.\end{gathered}$$ To close the system (\[eq:continuity\_pert\]), (\[eq:radial\_momentum\_pert\]), let use the relation between dislocation, density and velocity [@Lynden-Bell1974MNRAS.168..603L]: $$\begin{gathered} \label{eq:rho_xi} \rho_1 = - {\frac{\partial(\rho_0 \xi)}{\partialr}} \;, \\ \label{eq:v_xi} v_1 = {\frac{\partial\xi}{\partialt}} \;.\end{gathered}$$ It easy to show that, if these definitions are used, continuity equation (\[eq:continuity\_pert\]) is satisfied automatically. Inserting (\[eq:rho\_xi\]) and (\[eq:v\_xi\]) into the radial component of equations of motion equation (\[eq:radial\_momentum\_pert\]) we get a closed equation for the displacement: $$\label{eq:xi} {\frac{\partial^2 \xi}{\partialt^2}} = \frac{c_0^2}{\rho_0}\,{\frac{\partial^2 (\rho_0 \xi)}{\partialr^2}} + (\gamma - 2)\,\frac{c_0^2}{\rho_0^2}\,{\frac{\partial\rho_0}{\partialr}}\,{\frac{\partial(\rho_0 \xi)}{\partialr}} - \frac{1}{r^3}\,{\frac{d\lambda_0^2}{dr}}\,\xi \;.$$ The algorithm of solution of this equation is presented in the Appendix A. Solution has the form of a progressing wave: $$\label{eq:xi_solution} \xi = r_\ast \Re\!\left\{ C e^{\mp i \omega t} \left( \frac{r}{r_\ast} \right)^{7/4} H_\nu\!\left[ \frac{2 r_\ast \omega}{3 c_\ast} \left( \frac{r}{r_\ast} \right)^{3/2} \right] \right\} \;,$$ where $H_\nu$ is the Hankel function of the first kind of the order $\nu$; $C$ is a complex constant; $\Re\{\cdot\}$ — real part of its argument. The order is defined as $$\nu = \frac{1}{6} \left( 16\,\frac{r_\ast^2 \Omega_{\mathrm{K}\ast}^2}{c_\ast^2} - 23 \right)^{1/2} \;.$$ It is necessary to note that the solution of Eq. (\[eq:xi\_solution\]) has a wave form only in the case, when the following inequality keeps (see Appendix A): $$\label{eq:wave_constraint} \omega^2 > \left( 1 - \frac{3}{2}\,\frac{c_\ast^2}{r_\ast^2 \Omega_{\mathrm{K}\ast}^2} \right) \Omega_\mathrm{K}^2 \approx \Omega_\mathrm{K}^2 \;.$$ A similar result was obtained in [@Goldreich1979ApJ...233..857G] by WKB-method. The last inequality may be rewritten as $$\frac{r}{r_\ast} \gtrsim \left( \frac{\Omega_{\mathrm{K}\ast}}{\omega} \right)^{2/3} \;.$$ Thus, perturbation wave with a given frequency $\omega$ may progress in the outer part of the disk only, where $\omega > \Omega_\mathrm{K}$. At large distance from the center dependence of radial displacement on the radius gets a simpler form (see Appendix A): $$\label{eq:xi_approx_solution} \xi \approx |C|\,r_\ast \left( \frac{3 c_\ast}{\pi r_\ast \omega} \right)^{1/2} \frac{r}{r_\ast} \,\cos\!\left[ \frac{2 r_\ast \omega}{3 c_\ast} \left( \frac{r}{r_\ast} \right)^{3/2} \mp \omega t + (\dots) \right] \;,$$ where $(\dots)$ is an insignificant phase addition. Using (\[eq:rho\_xi\]) and (\[eq:v\_xi\]) we can obtain expressions for perturbations of density and velocity: $$\begin{gathered} \label{eq:rho_approx} \rho_1 \approx |C|\,\rho_\ast \left( \frac{3 r_\ast \omega}{\pi c_\ast} \right)^{1/2} \,\sin\!\left[ \frac{2 r_\ast \omega}{3 c_\ast} \left( \frac{r}{r_\ast} \right)^{3/2} \mp \omega t + (\dots) \right] \;, \\ \label{eq:v_approx} v_1 \approx \pm |C|\,c_\ast \left( \frac{3 r_\ast \omega}{\pi c_\ast} \right)^{1/2} \frac{r}{r_\ast} \,\sin\!\left[ \frac{2 r_\ast \omega}{3 c_\ast} \left( \frac{r}{r_\ast} \right)^{3/2} \mp \omega t + (\dots) \right] \;.\end{gathered}$$ Let get an expression for perturbation of the angular momentum. Using expansion (\[eq:angular\_momentum\_expansion\]) one may write down $$\lambda_1 \approx - {\frac{d\lambda_0^2}{dr}}\,\frac{\xi}{2 \lambda_0} \;.$$ Applying approximate relation (\[eq:xi\_approx\_solution\]), we get $$\label{eq:lambda_approx} \lambda_1 \approx - \frac{|C|}{2}\,r_\ast^2 \Omega_{\mathrm{K}\ast} \left( 1 - \frac{3}{2}\,\frac{c_\ast^2}{r_\ast^2 \Omega_{\mathrm{K}\ast}^2} \right)^{1/2} \left( \frac{3 c_\ast}{\pi r_\ast \omega} \right)^{1/2} \left( \frac{r}{r_\ast} \right)^{1/2} \,\cos\!\left[ \frac{2 r_\ast \omega}{3 c_\ast} \left( \frac{r}{r_\ast} \right)^{3/2} \mp \omega t + (\dots) \right] \;.$$ It is easy to estimate the typical space scale of the wave $\ell$ from oscillating part of solution (\[eq:xi\_approx\_solution\]). As the definition of $\ell$ let write down the equation for the phase: $$\label{eq:phase} 0 = \left( {\frac{\partial}{\partialt}} \pm c_0\,{\frac{\partial}{\partialr}} \right) \cos[ \dots ] = \pm \left( \omega - \frac{c_0}{\ell} \right) \sin[ \dots ] \;,$$ where $[\dots]$ stands for the term in square brackets in (\[eq:xi\_approx\_solution\]). Then $$\label{eq:wavelength} \ell = \frac{c_0}{\omega} = \frac{\Omega_\mathrm{K}}{\omega}\,H \;.$$ Comparing this expression with condition (\[eq:wave\_constraint\]) one may note that the perturbations have a wave form, if their  $\ell$ does not exceed semi-thickness of the disk. In other words, the waves can not propagate into the region $\ell > H$. Linear perturbations in the rotating flows were considered by many authors. For instance, in the classical paper [@Goldreich1979ApJ...233..857G] by WKB-method, was obtained a dispersion relation for the waves: $$\omega^2 = \Omega_\mathrm{K}^2 + c_0^2 k^2 \;,$$ where $k$ is the parameter of the WKB-method, which is interpreted as the wave number. Such an interpretation implies that the sense of the value $k^{-1}$ is the length of the perturbation wave. This value is related to $\ell$ from (\[eq:wavelength\]) as $$\label{eq:wavelength_wkb} k^{-1} = \frac{\ell}{(1 - \Omega_\mathrm{K}^2/\omega^2)^{1/2}} \;.$$ In the depth of the wave zone ($\omega \gg \Omega_\mathrm{K}$) the values of both expressions comply with precision up to value of the order $\Omega_\mathrm{K}^2/\omega^2$. However, in the region $\omega \approx \Omega_\mathrm{K}$ definition of wave length by expression (\[eq:wavelength\_wkb\]) provides a physically unacceptable result, since $k^{-1} \to \infty$. Below, as the length of perturbation wave we will bear in mind expression (\[eq:wavelength\]), which, in essence, is an estimate of the distance between neighboring maxima of the wave. Description of the dynamics of small perturbations in the terms of radial displacement $\xi$ is completely equivalent to the usual approach to the analysis of linear waves in the accretion disks, examples of which can be found in [@Lubow1993ApJ...409..360L; @Zhuravlev2007AstL...33..740Z] and many other papers. This method was used repeatedly earlier, see, e.g., description of the general approach to the analysis of Lagrangian perturbations in [@Lynden-Bell1974MNRAS.168..603L]. In the present study, the use of the Lagrangian approach is preferable, since it allows to express more easily the initial conditions for the analysis of nonlinear perturbations, which is carried out in the next Section. Formation of discontinuities ============================ It is known that nonlinear perturbations propagating in the gas can evolve with time into shock waves [@Landau1987theorphys-6; @Zeldovich1966shockwaves.book]. Usually, the proof of this statement is based on the solution in the form of a simple Riemann wave for one-dimensional gas dynamics [@Landau1987theorphys-6]. In the polytropic flow, described by a simple wave, all gasdynamic quantities are expressed through one variable. As it can be seen from the expressions (\[eq:rho\_approx\]) and (\[eq:v\_approx\]), in the case of a perturbation of inhomogeneous medium, density and velocity distributions depend on the coordinate differently and, therefore, they do not form a simple wave. However, it is possible to show that in this case the perturbation can also evolve into a shock wave. If there is a nonzero tangential velocity, on both sides of the shock wave forms a flow, equivalent to the flow in the boundary layer. To prove this assertion, we use the formulation of equations for one-dimensional gas flow presented in the book [@Landau1987theorphys-6]. First, we note that we are interested in perturbations from the wave part of the accretion disk, with wavelength which does not exceed the semi-thickness of the disk (see the previous Section). Let us consider, for definitiveness, a perturbation in the form of a wave (\[eq:rho\_approx\]), (\[eq:v\_approx\]). If nonlinear effects would be taken into account, the form of this perturbation will be distorted, but such characteristics as the amplitude and typical scale of variability ( $\ell$) will be about the same as in the linear case, at least up to the beginning of the formation of discontinuities. Let estimate the contribution of various terms to the linearized Euler equation (\[eq:radial\_momentum\_pert\]). The amplitude of velocity perturbations we normalize by the Mach number $\mathcal{M}$: $|v_1| \sim \mathcal{M} c_0$. Then the amplitude of density perturbations can be estimated as $|\rho_1| \sim \mathcal{M} \rho_0$. During one wave period, every gas element oscillates along the direction of propagation of the perturbation with an amplitude of the order of $|v_1| \ell/c$. This value can be used as an estimate of the displacement amplitude $|\xi|$. It follows from these estimates that in the progressing wave, the ratio of pressure force and the sum of the gravitational and centrifugal forces is $$\frac{c_0^2}{\rho_0}\,\frac{|\rho_1|}{\ell} : \Omega_\mathrm{K}^2 |\xi| \;\;\sim\;\; \frac{\mathcal{M} c_0^2}{\ell} : \Omega_\mathrm{K}^2 \mathcal{M} \ell \;\;\sim\;\; 1 : \frac{\ell^2}{H^2} \;.$$ In the wave part of the disk $\ell < H$, therefore, in order to avoid complicating of further presentation, we neglect the influence of gravitational and centrifugal acceleration upon dynamics of the wave. It may seem that, at the background of our approximations, all effects caused by differential rotation disappear. Indeed, we demanded that perturbations have sufficiently small scale of variability in comparison with the typical disk radius and the perturbation amplitude is small enough that it would be possible to interpret the sum of the gravitational and centrifugal accelerations in terms of the displacement of the gas element $\xi$. In additive, as it can be seen from the solution of the linearized problem (subsection 3.2), for sufficiently large $r$ the differential rotation law does not determine directly the dependence $\xi(t, r)$, but enters only the phase of the oscillating term. In reality, the role of differential rotation will be manifested in the fact that the distribution of the unperturbed density and the speed of sound are inhomogeneous in space and the law of this distribution is dictated, among other factors, by the law of differential rotation (see \[sec:equillibrium\]). We reduced the gas flow equations to the one-dimensional Euler equation and continuity equation. It is shown in [@Landau1987theorphys-6] that an arbitrary one-dimensional isentropic gas flow with the adiabatic index $\gamma$ may be described by a linear equation of the form $$\label{eq:chi} (\gamma-1)\,w\,{\frac{\partial^2\chi}{\partialw^2}} - {\frac{\partial^2\chi}{\partialv^2}} + {\frac{\partial\chi}{\partialw}} = 0 \;,$$ where $v$ is the velocity; $w = \int dp/\rho = c^2/(\gamma-1)$ — enthalpy. Solution of this equation, the function $\chi(v, w)$, is defined as $$\chi = \varphi - v r + \left( \frac{v^2}{2} + w \right) t \;,$$ where $\varphi$ is velocity potential, i.e. $v = \partial \varphi/\partial r$. Time $t$ and coordinate $r$ are expressed as: $$\begin{gathered} \label{eq:chi_t_general} t = {\frac{\partial\chi}{\partialw}} \;, \\ \label{eq:chi_r_general} r - v t = - {\frac{\partial\chi}{\partialv}} \;.\end{gathered}$$ We are not interested in the complete solution of the problem of nonlinear evolution of a perturbations with a given initial form, but only in the analysis of the possibility of formation of discontinuities in it. The location of the discontinuity can be defined as the point at which the plots of the velocity and/or enthalpy dependence on the coordinate become multivalued. In terms of the function $r(v, w)$ the position of discontinuity corresponds to the inflection point: $$\label{eq:shock_conditions_v} {\left(\frac{\partialr}{\partialv}\right)_{\!\!t}} = 0 \;,\qquad {\left(\frac{\partial^2r}{\partialv^2}\right)_{\!\!t}} = 0 \;,$$ while the derivatives are computed at the fixed instant of the time. It is easy to see that the r.h.s. of (\[eq:chi\_r\_general\]) describes dependence of $r(v, w)$ at $t = 0$; we will denote it as $R(v, w)$. The first of expressions (\[eq:shock\_conditions\_v\]) together with (\[eq:chi\_r\_general\]) provides the time of formation of discontinuity in solution: $$t_\mathrm{sh} = - {\left(\frac{\partialR}{\partialv}\right)_{\!\!t}} \;.$$ The calculation of this derivative in the general case may turn out to be a non-trivial task. However, we can express the position of the element we can express position of a gas element $R$ кas a function of velocity, by inversion of the dependence of the velocity on coordinates at $t = 0$. Let take as an initial perturbation configuration a section of the profile of a linear wave (\[eq:v\_approx\]): $$v|_{t=0} = - \mathcal{M}_\ast c_\ast\,\frac{r}{r_\ast} \,\sin\!\left[ \frac{2 r_\ast \omega}{3 c_\ast} \left( \frac{r}{r_\ast} \right)^{3/2} \right] \;,$$ where $\mathcal{M}_\ast$ is Mach number at $r_\ast$. Without limiting generality, let select $r_\ast$ as a point, where velocity becomes zero and declines in the direction of progression of the wave. It is evident that this point is also inflection point for dependence $r(v)$. Let consider the wave in a small vicinity of this point. Let introduce $x = r - r_\ast$; it will not be an additiveal restriction to require satisfaction of condition $|x| \ll r_\ast$: $$v|_{t=0} = - \mathcal{M}_\ast \omega x + \mathcal{O}\!\left( \frac{x^2}{r_\ast^2} \right) \;.$$ Now, with precision down to small quantities of the second order, we may write down: $$R = r_\ast - \frac{v}{\mathcal{M}_\ast \omega} \;.$$ Inserting this expression into (\[eq:chi\_r\_general\]) and applying the first of conditions (\[eq:shock\_conditions\_v\]) (the second one is fulfilled automatically), we obtain an expression for the time of discontinuity formation: $$\label{eq:shock_time} t_\mathrm{sh} = \frac{1}{\mathcal{M}_\ast \omega} \;,$$ where $\mathcal{M}_\ast$ and $\omega$ should be interpreted as parameters of perturbation in the point $r_\ast$. It is easy to find the point of formation of discontinuity $r_\mathrm{sh}$, if we recall that the local phase velocity of small-scale perturbations is equal to the sound velocity. In the general case of an arbitrary direction of progression of the wave we obtain $$\label{eq:shock_position_general} t_\mathrm{sh} = \pm \int_{r_\ast}^{r_\mathrm{sh}} \frac{dr}{c_0} \;.$$ The sign in the front of r.h.s. depends on direction of the wave propagation — the outer edge of the disk () or to the inner one (). Inserting (\[eq:shock\_time\]) and expression for the sound velocity $c_0 = c_\ast\,(r/r_\ast)^{-1/2}$ into (\[eq:shock\_position\_general\]), we obtain $$\label{eq:shock_position} r_\mathrm{sh} = r_\ast \left( 1 \pm \frac{3 c_\ast}{2 \mathcal{M}_\ast r_\ast \omega} \right)^{2/3} \;.$$ Note, for the waves progressing to the center of the accretion disk, the formula (\[eq:shock\_position\]) is valid in the wave region, i.e., as long as condition $\omega > \Omega_\mathrm{K}(r_\mathrm{sh})$ is valid (see subsection \[sec:linear\_pert\]). The time it takes for perturbation to reach the border of the wave zone may be estimated from the latter condition and expression (\[eq:shock\_position\_general\]): $$t_\mathrm{sh,max} = \frac{2 r_\ast}{3 c_\ast} \left( 1 - \frac{\Omega_{\mathrm{K}\ast}}{\omega} \right) \;.$$ If the discontinuity did not form during this time, then, following energy conservation law, the wave should reflect from the wave region boundary and head to the outer part of the disk. In this case, the wave has one more chance to form the discontinuity. However, the very reflection of the wave from the boundary $\ell = H$ can not be described in terms of the present model. It is evident that all above mentioned is valid also for dependence of perturbation of enthalpy on radius. At this, from the form of functions (\[eq:rho\_approx\]) and (\[eq:v\_approx\]) it is clear that inflection points for $r(v)$ and $r(w)$ coincide. It is important to note that the angular momentum distribution in the perturb flow will behave differently. Indeed, expressions (\[eq:rho\_approx\]), (\[eq:v\_approx\]) and (\[eq:lambda\_approx\]) show that, if the dependence of density and radial velocity passes through zero (the front of the perturbation wave), the angular momentum passes through the extremum. This means that in the plot of the dependence of the angular momentum on the radius does not appear discontinuity. However, since the angular momentum is transferred by the gas as a passive scalar, see (\[eq:angular\_momentum\_general\]), at the instant of formation of the discontinuity of density and radial velocity, the distribution of angular momentum will get the shape of a caustic with a peak located at the point $r_\mathrm{sh}$. Schematically this is shown in Fig. \[fig:shock\_form\]. As it is seen, the amplitude of the density and radial velocity step is determined by the change of the corresponding value over the scale of half wavelength, $\ell/2$, of the initial perturbation. The change of tangential velocity on the caustic is determined mainly by the difference of the values of the tangential velocity in the original distribution over the scale $\ell/4$.  Application to accretion disks ============================== In the previous Section we obtained the estimates of the formation time of discontinuities in a rotating flow. As was shown in [@Landau1987theorphys-6], discontinuities of density and velocity, formed as a result of rollover of the wave profile are asymptotically damped because of microscopic viscosity. However, in the case of a rotating flow, the picture of the evolution of a perturbation after formation of a discontinuity may change qualitatively. We noted above that at the time of the formation of the discontinuous solution the tangential velocity profile takes the form of a caustic. In the neighborhood of the caustic tangential velocity will experience a strong change. Let $\delta \lambda$ be the difference of the angular moment in the initial tangential velocity distribution over the scale $\ell/4$: $$\delta \lambda \approx \left| {\frac{d\lambda_0^2}{dr}}\,\frac{\ell/4}{2 \lambda_0} \right|_{r_\mathrm{sh}} \approx \frac{\Omega_\mathrm{K}(r_\mathrm{sh})}{8 \omega}\,r_\mathrm{sh} c_0(r_\mathrm{sh}) \;.$$ Here, we neglected a small deviation from Keplerian distribution of angular momentum in the expression (\[eq:subkeplerian\_momentum\]). The change of tangential velocity at the caustic is equal to $\delta \lambda/r_\mathrm{sh}$ or, in the units of sound velocity, $$\label{eq:alpha} \alpha \equiv \frac{\Omega_\mathrm{K}(r_\mathrm{sh})}{8 \omega} = \frac{\Omega_{\mathrm{K}\ast}}{8 \omega} \left( 1 \pm \frac{3 c_\ast}{2 \mathcal{M}_\ast r_\ast \omega} \right)^{-1} \;.$$ In the reference frame connected with the surface of discontinuity, configuration of the flow resembles a laminar boundary layer of the . Indeed, the typical width of the caustic is much smaller than the wavelength $\ell$, which in the wave zone can not exceed semi-thickness of the disk. At the same time, in the tangential direction the flow is fairly homogeneous. It is well known that the flow of this type becomes unstable if the Reynolds number $\mathsf{Re} \approx 3.7$ (see, for example, [@Monin1986SvPhU..29..843M]). With increase of the Reynolds number, the range of the wave numbers of perturbations which cause the instability of the boundary layer expands without limits. For a typical accretion disk, one can estimate that $\mathsf{Re} \sim 10^{10}$ [@Razdoburdin2015PhyU...58.1031R]. Even if we determine the Reynolds number from the value of the tangential velocity difference (\[eq:alpha\]), its estimate decreases by two or three orders of magnitude only. This allows to claim that the region of the caustic in the distribution of the tangential velocity is unstable with respect to perturbations of practically any scale. We note that the region of the caustics is located at both sides of the density and radial velocity jump. Since the jump is a shock wave, it propagates in the gas with supersonic velocity. By virtue of this, small-scale perturbations can propagate behind the shock only [@Landau1987theorphys-6], and, therefore, the instability can develop only there. Given that accretion disks are typically have huge Reynolds numbers, instability of the distribution of tangential velocity in the caustic may well become a source of turbulence in the disk. Then the quantity $\alpha$ (\[eq:alpha\]) should be interpreted as an estimate of the Shakura-Sunyaev coefficient . Expression (\[eq:alpha\]) is written down for the waves that move from the center (sign ) and to the center (sign ). If the perturbation propagates toward the outer region of the disk, the value of $\alpha$ increases with increasing amplitude of the initial perturbation $\mathcal{M}_\ast$, but it can not exceed $1/8$ in principle due to condition (\[eq:wave\_constraint\]). Let us consider as an example a protoplanetary disk around binary system with the total mass $M_\mathrm{tot} = 2 M_\odot$, orbital separation $A = 10 R_\odot$ and the inner disk radius $r_\mathrm{in} = 2 A$. We set the temperature of the disk equal to $10^4$ К. Let assume that the wave moves from the inner boundary of the disk to the outer one (in this case it is necessary to assume $r_\ast = r_\mathrm{in}$). Figure \[fig:shock\_out\] shows the plots of the dependence on time of the radius of formation of the discontinuity and the amplitude of the jump of tangential velocity on the spatial scale and Mach number of the initial perturbation. It is seen that for $\mathcal{M}_\ast \lesssim 10^{-2}$ the value of $\alpha$ weakly depends on the initial wavelength of perturbation, but increases with the amplitude as $\alpha \sim \mathcal{M}_\ast$. The radius, at which the discontinuity is formed and the time of its formation rapidly grow with increasing wavelength, i.e., long-wave perturbations of small amplitude can reach the outer boundary of the disk, without having had time to form a discontinuity. Perturbations of larger amplitudes, naturally, quite quickly form high intensity jumps, where $\alpha \lesssim 1/8$.  Scenario of formation of discontinuities in the case when the wave comes from the outer side of the disk, is somewhat different. Let consider as an example a circumstellar disk around a white dwarf with the mass $M = 1 M_\odot$. Let set the inner radius of the disk equal to $r_\mathrm{in} = 0.03 R_\odot$ and the outer one to $r_\mathrm{out} = 0.3 R_\odot$, the temperature of the disk to $10^4$ К; $r_\ast = r_\mathrm{out}$. The parameters of the discontinuities that form in such a disk are shown in Fig. \[fig:shock\_in\]. Unlike the first case, here the time of propagation of the wave is limited by the time it intersects the wave zone (a situations when the discontinuity forms strictly at the boundary of the wave zone are marked in Fig. \[fig:shock\_in\] by asterisks). As the wave propagates into the disc, sound velocity increases. This explains a large gradient of the dependences $r_\mathrm{sh}(\mathcal{M}_\ast)$ and $\alpha(\mathcal{M}_\ast)$ in the neighborhood of the boundary of the wave zone. The value of $\alpha$ at this point attains its maximum value $1/8$. Perturbations of a small initial amplitude, as a rule, lead to higher values of the coefficient $\alpha$ than in the first case, up to two orders of magnitude. Wave region of small-scale perturbations can cover entire disk. If the amplitude of such perturbations is small, they can reach the inner boundary of the disk, before they form a discontinuity.  Conclusion ========== In this paper we studied the evolution of nonlinear perturbations in a single class of sub- Keplerian accretion disks. It was assumed that disks can be exposed to external action at inner and outer borders. The first case corresponds to protoplanetary disks around binary systems of young stars of T Tau type; the source of impact can be detached shock waves. The second case can be realized in accretion disks of close binary stars, where perturbations can be generated by the region of interaction of the accretion flow and the disk. It is found that at the front of a radial longitudinal perturbation wave a discontinuity of the density and radial velocity forms with time, which is a shock wave. Distribution of the tangential velocity in this case takes the form of a caustic. Position of the peak ofthe latter corresponds to the position of the shock wave. Formation of the shock wave occurs due to the well-known mechanism of the  of the profile of the nonlinear perturbation. The value of the jump of the tangential velocity at the caustic depends on the amplitude and length of the perturbation wave. In some cases it can reach about $0.1$ of the sound velocity. Such a configuration is unstable and disintegrates. The natural consequence of the decay is turbulization of the accretion disk. The estimates of the Shakura-Syunyaev parameter for the accretion disks — $\alpha < 0.1$ — in both considered cases indicate a substantially subsonic turbulence. Acknowledgments =============== E.P. Kurbatov was supported by the Russian Foundation for Basic research (contracts No. 14-29-06059 and 15-02-06365). Appendix A ========== After substitution $\eta = \rho_0 \xi$, Eq. (\[eq:xi\]) for displacement of the gas elements becomes: $$\label{app:eq:eta} {\frac{\partial^2 \eta}{\partialt^2}} = c_0^2\,{\frac{\partial^2\eta}{\partialr^2}} + (\gamma - 2)\,\frac{c_0^2}{\rho_0}\,{\frac{\partial\rho_0}{\partialr}}\,{\frac{\partial\eta}{\partialr}} - \frac{1}{r^3}\,{\frac{d\lambda_0^2}{dr}}\,\eta \;.$$ We do not provide boundary conditions for this equation, but seek solution in the form of propagating perturbance. Therefore, we may apply the standard method of separation of variables: $$\eta(t, r) = T(t)\,R(r) \;.$$ Then, denoting by $-\omega^2$ integration constant, we get (dot and prime mark denote differentiation over time and spatial coordinate, respectively) $$\frac{\ddot{T}}{T} = \frac{c_0^2}{R} \left[ R'' + (\gamma - 2)\,(\ln \rho_0)' R' \right] - \frac{1}{r^3}\,{\frac{d\lambda_0^2}{dr}} = - \omega^2 \;.$$ Let write down solution for the temporal part as $$T \propto e^{\mp i \omega t} \;.$$ For the spatial part we get an equation $$R'' + (\gamma - 2)\,(\ln \rho_0)' R' + \left( \frac{\omega^2}{c_0^2} - \frac{1}{c_0^2 r^3}\,{\frac{d\lambda_0^2}{dr}} \right) R = 0 \;.$$ Let insert to the last Eq. radial profiles of the density and sound velocity obtained in subsection \[sec:equillibrium\]. Then, after substitution of the variables, $x = r/r_\ast$, $R(r) = X(x)$, and introduction of denotations: $$\begin{gathered} \mu = \frac{r_\ast \omega}{c_\ast} \;, \\ \varkappa = \frac{r_\ast \Omega_{\mathrm{K}\ast}}{c_\ast} \;,\end{gathered}$$ we obtain: $$x^2 X'' + \frac{x}{2}\,X' + \left( \mu^2 x^3 - \varkappa^2 + \frac{3}{2} \right) X = 0 \;.$$ Solution of this equation will be expressed via Hankel function of the first type:[^3] [@Zaitsev2003ode.book с. 169]: $$\label{app:eq:x} X = C r_\ast \rho_\ast x^{1/4} H_\nu\!\left( \frac{2 \mu}{3}\,x^{3/2} \right) \;,$$ where $C$ is a complex constant; the index of Hankel function is equal to $$\nu = \frac{1}{6}\,(16 \varkappa^2 - 23)^{1/2} \;.$$ For larger arguments of Hankel function, $\mu x^{3/2} \gg 1$, one has $H_\nu(z) \approx (\pi z/2)^{-1/2} \operatorname{exp}[i z + i (\dots)]$, and then $$\label{app:eq:x_aprx} X \approx C r_\ast \rho_\ast \left( \frac{3}{\pi \mu} \right)^{1/2} x^{-1/2} \operatorname{exp}\left[ i \frac{2 \mu}{3}\,x^{3/2} + i (\dots) \right] \;,$$ where $i (\dots)$ — insignificant phase addition. Note, if solution has to have wave character, it is necesarry to satisfy the following inequality: $$\label{app:eq:wave_constraint} \mu^2 x^3 - \varkappa^2 + \frac{3}{2} > 0 \;.$$ or $$x > \left( \frac{\varkappa^2}{\mu^2} - \frac{3}{2\mu^2} \right)^{1/3} \;.$$ For low $x$ solution (\[app:eq:x\]) very rapidly aperiodically approaches zero. Figure \[fig:wave\_cart\] shows approximately the dependence $X(x)$.  Exact solution for linear perturbation may be obtained also for full set up of the problem, without neglect of the  term in the continuity equation (\[eq:continuity\_general\]). In this case, the value of displacement and density will be related as $$\rho_1 = - \frac{1}{r}\,{\frac{\partial(r \rho_0 \xi)}{\partialr}} \;.$$ After substitution $\eta = r \rho_0 \xi$ we obtain an equation analogous to (\[app:eq:eta\]): $${\frac{\partial^2 \eta}{\partialt^2}} = c_0^2 r\,{\frac{\partial}{\partialr}} \left( \frac{1}{r}\,{\frac{\partial\eta}{\partialr}} \right) + (\gamma - 2)\,\frac{c_0^2}{\rho_0}\,{\frac{\partial\rho_0}{\partialr}}\,{\frac{\partial\eta}{\partialr}} - \frac{1}{r^3}\,{\frac{d\lambda_0^2}{dr}}\,\eta \;.$$ From this equation, by means of separation of variables and substitutions listed above, it is possible to get an equation for $X$, which differs from the last Eq. by the sign in the front of the second term only: $$x^2 X'' - \frac{x}{2}\,X' + \left( \mu^2 x^3 - \varkappa^2 + \frac{3}{2} \right) X = 0 \;.$$ Its solution [@Zaitsev2003ode.book]: $$X = C r_\ast^2 \rho_\ast x^{3/4} H_\nu\!\left( \frac{2 \mu}{3}\,x^{3/2} \right) \;,$$ where the index of hankel function is $$\nu = \frac{1}{6}\,(16 \varkappa^2 - 15)^{1/2} \;.$$ This solution does not differ essentially from Eq. (\[app:eq:x\]), obtained in quasi-Cartesian approximation. [^1]: [email protected] [^2]: [email protected] [^3]: Hankel functions are related to the Bessel functions of the first and second type: H $H_\nu = J_\nu + i Y_\nu$.

ArXiv

--- abstract: 'We report the *Fermi* Large Area Telescope discovery of [$\gamma$]{}-ray pulsations from the 22.7 ms pulsar A in the double pulsar system J0737$-$3039A/B. This is the first mildly recycled millisecond pulsar (MSP) detected in the GeV domain. The 2.7 s companion object PSR J0737$-$3039B is not detected in [$\gamma$]{} rays. PSR J0737$-$3039A is a faint [$\gamma$]{}-ray emitter, so that its spectral properties are only weakly constrained; however, its measured efficiency is typical of other MSPs. The two peaks of the [$\gamma$]{}-ray light curve are separated by roughly half a rotation and are well offset from the radio and X-ray emission, suggesting that the GeV radiation originates in a distinct part of the magnetosphere from the other types of emission. From the modeling of the radio and the [$\gamma$]{}-ray emission profiles and the analysis of radio polarization data, we constrain the magnetic inclination $\alpha$ and the viewing angle $\zeta$ to be close to 90$^\circ$, which is consistent with independent studies of the radio emission from [PSR J0737$-$3039A]{}. A small misalignment angle between the pulsar’s spin axis and the system’s orbital axis is therefore favored, supporting the hypothesis that pulsar B was formed in a nearly symmetric supernova explosion as has been discussed in the literature already.' author: - 'L. Guillemot, M. Kramer, T. J. Johnson, H. A. Craig, R. W. Romani, C. Venter, A. K. Harding, R. D. Ferdman, I. H. Stairs, and M. Kerr' bibliography: - 'J0737-3039A.bib' title: '*Fermi* LAT pulsed detection of [PSR J0737$-$3039A]{} in the double pulsar system' --- Introduction {#sec:introduction} ============ Pulsed GeV [$\gamma$]{}-ray emission from more than 100 pulsars[^1] has been observed by the Large Area Telescope [LAT; @FermiLAT] aboard the *Fermi* satellite, launched in 2008 June [see The Second Fermi Large Area Catalog of Gamma-Ray Pulsars; @Fermi2PC hereafter 2PC]. The current population of [$\gamma$]{}-ray pulsars includes objects known from independent radio or X-ray observations and detected in [$\gamma$]{} rays by folding the *Fermi* LAT data at the known rotational periods [e.g., @Fermi8MSPs; @FermiJ2043; @Espinoza2013], and [$\gamma$]{}-ray pulsars found through direct blind searches of the LAT data [@FermiBlindSearch; @Pletsch9PSRs] or radio searches for pulsars in unassociated [$\gamma$]{}-ray sources [see @Ray2012 and references therein]. These pulsars are energetic (spin-down luminosities $\dot E = 4 \pi^2 I \dot P / P^3 > 10^{33}$ erg s$^{-1}$, where $P$ is the spin period, $\dot P$ is its first time derivative, and $I$ denotes the moment of inertia, assumed to be $10^{45}$ g cm$^2$ in this work) and are typically nearby. They therefore have large values of the “spin-down flux” $\dot E / d^2$ and of the heuristic detectability metric $\sqrt{\dot E} / d^2$ [@FermiPSRCatalog]. The double pulsar system PSR J0737$-$3039A/B [@Burgay2003; @Lyne2004] consists of two radio-emitting neutron stars in a tight 2.4 hr orbit. Radio timing observations of the two pulsars provide high-precision tests of strong field gravity [@Kramer2006]. The 2.7 s pulsar J0737$-$3039B has a low spin-down luminosity $\dot E \sim 1.7 \times 10^{30}$ erg s$^{-1}$, three orders of magnitude smaller than that of the least energetic [$\gamma$]{}-ray pulsar currently known, making this pulsar unlikely to be detectable by the LAT. On the other hand, the higher $\dot E$ of $\sim 5.94 \times 10^{33}$ erg s$^{-1}$—for which the Shklovskii effect [@Shklovskii1970] is negligible due to the system’s low transverse velocity—and the modest parallax distance of $1150_{-160}^{+220}$ pc measured with very long baseline interferometry observations [@Deller2009] make the 22.7 ms pulsar J0737$-$3039A a credible candidate for a detection in [$\gamma$]{} rays. Searching for high-energy pulsations from PSRs J0737$-$3039A and B allows exploration of regions of the $P-\dot P$ diagram that are devoid of [$\gamma$]{}-ray pulsars [@FermiPSRCatalog]: mildly recycled millisecond pulsars (MSPs) for the former (mildly recycled pulsars being pulsars that were only partially spun-up by accretion of matter from a companion star), and slowly-rotating normal pulsars for the latter. Increasing the variety of [$\gamma$]{}-ray pulsar types helps understand the phenomenology of high-energy emission from these stars, and how emission properties evolve with, e.g., $\dot E$ or age (with the caveat that the evolution could be masked by effects arising from the geometry). Additionally, the modeling of pulse profiles at different wavelengths can yield constraints on the geometry of emission, in particular, the inclination angle of the magnetic axis ($\alpha$) and the observer viewing angle ($\zeta$) with respect to the spin axis. Constraints on $\alpha$ and $\zeta$ can also be determined by studying the radio polarization. Since the inclination of the binary system with respect to the line-of-sight is accurately known, $\zeta$ directly measures the misalignment between the pulsar’s spin axis and the orbital angular momentum, a quantity which gives important clues about the formation of the binary system [for a discussion, see, e.g., @Ferdman2008]. We here report the discovery of [$\gamma$]{}-ray pulsations from the 22.7 ms pulsar J0737$-$3039A, in 43 months of data recorded by the *Fermi* LAT. In Section \[sec:gammaanalysis\] we describe the analysis of the LAT data and the resulting light curve and spectrum for [PSR J0737$-$3039A]{}. In Sections \[sec:lcmodeling\] and \[sec:pol\], we model the offset radio and [$\gamma$]{}-ray light curves, as well as the radio polarization, under various emission models, allowing us to constrain the pulsar’s geometrical characteristics. We conclude with a brief discussion on the implication of these results. LAT analysis {#sec:gammaanalysis} ============ We analyzed *Fermi* LAT events recorded between 2008 August 4 and 2012 March 7, belonging to the “Source” class under the P7\_V6 instrument response functions, and with zenith angles smaller than 100$^\circ$. We rejected events recorded when the rocking angle of the telescope exceeded 52$^\circ$, when the instrument was not operating in the science observations mode or when the data quality flag was not set as good. Analyses of the LAT data were carried out using the *Fermi* Science Tools[^2] (STs) v9r28p0. The data were phase-folded with the *Fermi* plug-in [@Ray2011] distributed with the <span style="font-variant:small-caps;">Tempo2</span> pulsar timing package [@tempo2], and the ephemerides for PSRs J0737$-$3039A and B published in @Kramer2006. Since 2008 March PSR J0737$-$3039B has been invisible in the radio domain [@Perera2010] because of the precession of its spin axis [@Breton2008] causing its radio beam to miss the Earth, precluding timing measurements contemporaneous with the *Fermi* mission. Additionally, changes in the radio profile of PSR J0737$-$3039B prevented @Kramer2006 from constructing a coherent timing model for this pulsar. Our ability to accurately fold the *Fermi* LAT data using this ephemeris is therefore limited. Nevertheless, the very low spin-down luminosity of this pulsar makes it unlikely to be detected in [$\gamma$]{} rays. From initial pulsation searches using standard data selection cuts we found no evidence of [$\gamma$]{}-ray pulsations from PSR J0737$-$3039B. However, selecting events found within $0\fdg5$ from [PSR J0737$-$3039A]{} and with energies above 0.2 GeV we obtained a value for the bin-independent *H*-test parameter [@deJager2010] of 21.3, corresponding to a pulsation significance of $\sim 3.7\sigma$. The [$\gamma$]{}-ray pulse profile of PSR J0737$-$3039A comprises two peaks, which is reminiscent of known [$\gamma$]{}-ray pulsar light curves [e.g., @FermiPSRCatalog], with no evidence for emission in the $[0 ; 0.2] \cup [0.5 ; 0.8]$ phase interval. [lc]{} Rotational period, $P$ (ms) & $22.7$\ Period derivative, $\dot P$ ($10^{-18}$) & $1.76$\ Spin-down luminosity, $\dot E$ ($10^{33}$ erg s$^{-1}$) & $5.94$\ Magnetic field at the light cylinder, $B_{\rm LC}$ ($10^3$ G) & $4.97$\ Distance, $d$ (pc) & $1150^{+220}_{-150}$\ First peak position, $\Phi_1$ & $0.43 \pm 0.01$\ First peak full width at half-maximum, FWHM$_1$ & $0.03 \pm 0.02$\ Distance from the closest radio peak maximum, $\delta_1$ & $0.20 \pm 0.01$\ Second peak position, $\Phi_2$ & $0.93 \pm 0.01$\ Second peak full width at half-maximum, FWHM$_2$ & $0.02_{-0.02}^{+0.03}$\ Distance from the closest radio peak maximum, $\delta_2$ & $0.13 \pm 0.01$\ [$\gamma$]{}-ray peak separation, $\Delta = \Phi_2 - \Phi_1$ & $0.49 \pm 0.01$\ Photon index, $\Gamma$ & $< 1.3$\ Cutoff energy, $E_c$ (GeV) & $ 0.4 \pm 0.4$ $(<0.9)$\ Photon flux, $F$ ($\geq 0.1$ GeV) ($10^{-9}$ photons cm$^{-2}$ s$^{-1}$)& $ 6 \pm 3$ $(<9)$\ Energy flux, $G$ ($\geq 0.1$ GeV) ($10^{-12}$ erg cm$^{-2}$ s$^{-1}$) & $ 4 \pm 1$ $(<5)$\ Luminosity, $L_\gamma / f_\Omega = 4 \pi G d^2$ ($10^{32}$ erg s$^{-1}$) & $ 6_{-2}^{+3}$ $(<8)$\ Efficiency, $\eta / f_\Omega = 4 \pi G d^2 / \dot E$ & $0.10_{-0.04}^{+0.05}$ $(<0.13)$\ As illustrated in @Kerr2011weights and @Guillemot2012a, weighting the [$\gamma$]{}-ray events by the probability that they originate from the putative pulsar increases the sensitivity to faint pulsations. To measure these probabilities we performed a binned likelihood analysis of the spectra of the sources in the region of interest, using the `PyLikelihood` module of the STs. Our spectral model included the 50 2FGL catalog sources [@Fermi2FGL] within 20$^\circ$ of [PSR J0737$-$3039A]{}. The contribution from the pulsar, which is not associated with any 2FGL source, was modeled as an exponentially cutoff power law (ECPL) of the form $dN/dE \propto E^{-\Gamma} \exp \left( - E / E_c \right)$, where $\Gamma$ is the photon index and $E_c$ is the cutoff energy of the spectrum. The source model also included contributions from the extragalactic diffuse emission and the residual instrumental background, jointly modeled using the *iso\_p7v6source* template, and from the Galactic diffuse emission, modeled with the *gal\_2yearp7v6\_v0* map cube[^3]. The parameters of [PSR J0737$-$3039A]{}, of the seven sources within 10$^\circ$ of the pulsar, and the normalization factors of diffuse components were left free in the fit. A Test Statistic [TS; for a definition see @Fermi2FGL] value of 20.3 was found for [PSR J0737$-$3039A]{} at this stage. To improve the quality of the spectral results we analyzed the $[0 ; 0.2] \cup [0.5 ; 0.8]$ phase interval (OFF pulse), removing the contribution from the pulsar. This allowed us to obtain an improved fit of the spectra of neighboring sources. Sources with TS values below 2 were excluded from the best-fit model obtained at this point, and the spectral parameters of sources beyond 3$^\circ$ from the double pulsar system were frozen at their best-fit values. We finally analyzed events in the complementary $[0.2 ; 0.5] \cup [0.8 ; 1]$ phase interval (ON pulse) with the resulting source model, refitting the contributions of the pulsar and of the three sources within 3$^\circ$, to determine the spectral parameters for [PSR J0737$-$3039A]{} listed in Table \[tab:params\]. A TS value of 31.2 is obtained for [PSR J0737$-$3039A]{}. Refitting the data with a power law shape for the pulsar, we found that the ECPL model is preferred by the likelihood at the [4]{}$\sigma$ level. A cross-check of this analysis performed with the independent <span style="font-variant:small-caps;">Pointlike</span> analysis tool [@KerrThesis] yielded results consistent with those listed in Table \[tab:params\]. Also listed in the table are the [$\gamma$]{}-ray luminosity $L_\gamma = 4 \pi f_\Omega G d^2$ where $G$ is the phase-averaged energy flux above 0.1 GeV and assuming a beaming correction factor $f_\Omega = 1$, and the derived efficiency $\eta = L_\gamma / \dot E$. The photon flux $F$ and energy flux $G$ of [PSR J0737$-$3039A]{} are among the lowest values of any [$\gamma$]{}-ray pulsars detected to date, probably a consequence of its low $\dot E$ value. We also note that [PSR J0737$-$3039A]{} has the lowest value of the magnetic field at the light cylinder of any recycled pulsars detected in [$\gamma$]{} rays, $B_{\rm LC} \sim 4.97 \times 10^3$ G [see Equation 3.22 of @Handbook]. None of the different attempts to model the spectrum of [PSR J0737$-$3039A]{} presented above yielded satisfactory measurements of its photon index, $\Gamma$. In all cases, a small value consistent with 0 was favored. This is likely a consequence of the cutoff energy $E_c$ being too low to reliably determine the photon index $\Gamma$. Limits on the other spectral parameters can thus be placed by repeating the analysis using a value for the photon index which is representative of the $\Gamma$ values observed for other MSPs. Table \[tab:params\] lists the results obtained by fixing $\Gamma$ at 1.3, this value being the average photon index measured for the strongest MSPs in 2PC. The spectral parameters quoted in parentheses represent canonical values, arising from the lack of constraints on $\Gamma$. The tool `gtsrcprob` was finally used to assign each event a probability that it originated from [PSR J0737$-$3039A]{} based on the fluxes and spectra obtained from the likelihood analysis. The integrated pulse profile of [PSR J0737$-$3039A]{} over 0.1 GeV and for probabilities larger than 0.05 is shown in Figure \[fig:lc\]. The zero of phase is defined by the maximum of the first Fourier harmonic of the signal transferred back to the time domain. We find a weighted $H$-test parameter of [41]{} [@Kerr2011weights; @deJager2010], corresponding to a significance of $\sim$[5.4]{}$\sigma$. As can be seen from Figure \[fig:lc\], the [$\gamma$]{}-ray profile comprises two peaks, with indications for additional complexity. We fitted the profile above 0.1 GeV with Lorentzian functions and found the peak positions and full widths at half-maxima for the first and second [$\gamma$]{}-ray peaks as listed in Table \[tab:params\]. Defining $\delta_1$ and $\delta_2$ as the separation between the [$\gamma$]{}-ray peaks and the maxima of the closest 1.4 GHz radio peaks, respectively at $\Phi_{r,1} \sim 0.24$ and $\Phi_{r,2} \sim 0.79$, we find $\delta_1 = 0.20 \pm 0.01$ and $\delta_2 = 0.13 \pm 0.01$. The uncertainty on $\delta_1$ and $\delta_2$ caused by the error on the dispersion delay is $\Delta (\mathrm{DM}) / (k f^2)$, where $\Delta (\mathrm{DM}) = 5 \times 10^{-3}$ pc cm$^{-3}$ is the uncertainty on the dispersion measure (DM) reported in @Kramer2006, $k$ is the dispersion constant [see @Handbook] and $f = 1.4$ GHz, is found to be $\sim 5 \times 10^{-4}$ in phase, and can thus be neglected. The [$\gamma$]{}-ray emission from [PSR J0737$-$3039A]{} is thus offset from the radio emission, suggesting distinct origins in the magnetosphere of the pulsar. This also holds for the X-ray peaks detected by @Chatterjee2007 in *Chandra* HRC data, which are likewise aligned with the radio peaks. Finally, we have searched the OFF-pulse fraction of the data for modulation at the orbital period, caused by the collision of the particle winds from the pulsars, and we have also searched the ON-pulse signal for attenuation of the emission from pulsar A around conjunction, caused by, e.g., photon–light-pseudoscalar-boson oscillation in the magnetosphere of B, as proposed by @Dupays2005. In both cases, we observed only steady emission as a function of orbital phase. Constraining the Viewing\ Geometry of [PSR J0737$-$3039A]{} {#sec:geom} ================================= We can place constraints on the magnetic inclination $\alpha$ and viewing angle $\zeta$ by modeling the radio and [$\gamma$]{}-ray light curves and the radio polarization. For all models we assume the vacuum, retarded-dipole magnetic field geometry [@Deutsch1955]. The emission is assumed to originate in the open zone, determined by field lines that do not close within the light cylinder (at $R_{\rm LC} = c / \Omega = c P / 2 \pi$), traced to foot points on the star surface at $R_{\rm NS}$ (radius of the neutron star), which define the polar cap. The computation and fitting procedure are described in @Watters2009, @Venter2009, @RomaniWatters2010, and references therein. In two of the prevalent models of [$\gamma$]{}-ray emission, uniformly emissive zones of the magnetosphere stand in for more physical models. The first is the outer gap picture [OG; @Cheng1986; @Yadigaroglu1995] in which radiation is emitted between the null charge surface where $\mathbf{\Omega} \cdot \mathbf{\mathit{B}} = 0$ and $R_{\rm LC}$. The second is the two-pole caustic model [TPC; @Dyks2003], which we take to be a geometric realization of the slot gap model [@Muslimov2004]. In the original TPC model emission extends from the surface to 0.75 $R_{\rm LC}$. In both geometries the [$\gamma$]{}-ray emission is confined toward the edge of the open zone coinciding with the last open field lines, in a “gap” idealized as a region of width $w_{\rm em}$, interior to an accelerating layer of width $w_{\rm acc}$. In the OG model the [$\gamma$]{}-ray emission originates from a thin layer on the inner edge of the gap, closest to the magnetic pole, whereas in the TPC model emission is produced throughout the gap. The radio emission can occupy a large fraction of the open field lines, albeit generally at lower altitude. Joint Radio and [$\gamma$]{}-Ray Light Curve Fits {#sec:lcmodeling} ------------------------------------------------- One approach is to fit the [$\gamma$]{}-ray and radio light curves directly, following the geometrical light curve modeling of @Venter2009, with the fitting procedure described in more detail in, e.g., @Guillemot2012a. For these computations we assume $P = 25$ ms and $\dot{P}=1\times10^{-18}$ s s$^{-1}$, and treat the radio pulse as a classical single-altitude, hollow-cone beam centered on the magnetic axis as described by @Story2007. For this analysis we start from a grid of models with 1$^\circ$ steps in both $\alpha$ and $\zeta$, and fit the emission zone accelerating and emitting gap widths, $w_{\rm acc}$ and $w_{\rm em}$, with steps of 2.5% of the polar cap opening angle ($\Theta_{\rm PC} \approx (\Omega R_{\rm NS}/c)^{1/2}$). For this computation we extend the emission to 0.95 $R_{\rm LC}$. The models are fit to an unweighted-counts [$\gamma$]{}-ray light curve for [PSR J0737$-$3039A]{}, constructed by selecting events from the data described in Section \[sec:gammaanalysis\] found within 2$^\circ$ of the pulsar, with energies above 0.1 GeV and with probabilities of being associated with the pulsar larger than 0.05. In the case of low-statistics [$\gamma$]{}-ray light curves with sharp peaks such as those observed for [PSR J0737$-$3039A]{}, using a $\chi^2$ statistic with binned light curves may not be optimal as this statistic is often insensitive to the peaks. We have thus found that it is preferable to use the unweighted counts, which allow the use of Poisson likelihood, as opposed to the weighted counts which require using a $\chi^2$ statistic for binned light curves. The radio profile is fit with a $\chi^{2}$ statistic, assuming the same relative uncertainty for each phase bin, which is combined with the likelihood from the [$\gamma$]{}-ray fit to determine the best-fit model parameters from a scan over the model phase space. We used 30 bins for both the [$\gamma$]{}-ray and radio light curves in these fits. One difficulty in joint fitting of the radio and [$\gamma$]{}-ray profiles is that the low statistical uncertainty of the radio light curve compared to the uncertainty of the [$\gamma$]{}-ray bins causes the fit to be dominated by the radio data and to ignore the constraints provided by the [$\gamma$]{} rays. Because our radio model is simplistic, this can lead to unrealistic solutions. We have therefore investigated different prescriptions for the “uncertainty” on the radio profile bins, in order to balance the radio and [$\gamma$]{}-ray contributions to the joint likelihood. The following prescription gives the most satisfactory results. First, we select an on-peak interval for the [$\gamma$]{}-ray light curve, which should dominate the likelihood, and define $\sigma_{\rm g\,ave}$ as the average, fractional uncertainty of the [$\gamma$]{}-ray bins in this interval. We then define a radio uncertainty $\sigma_{\rm r} = r_{\rm max}\times\sigma_{\rm g\,ave}$, where $r_{\rm max}$ denotes the maximum radio light curve bin. Although this prescription gives satisfactory results, it is arbitrary. We tested its sensitivity by varying $\sigma_{\rm r}$ by a factor of two [see @Guillemot2012b]; the resulting best-fit geometries changed by $\alpha \lesssim 4^{\circ}\ (12^{\circ})$ and $\zeta \lesssim 4^{\circ}\ (9^{\circ})$ for the TPC (OG) model. The latter values can be considered systematic errors on our fit results. The best-fitting model light curves are shown in Figure \[fig:lcmodeling\], and the parameters are given in Table \[tab:fitpars\]. The $1\sigma$ uncertainties are estimated from two-dimensional likelihood profiles for $\alpha$ and $\zeta$ and one-dimensional profiles for the other parameters. The best-fit gap widths for both models have size 0, indicating that they are unresolved by our simulations. The simple geometric models we use neglect important physics, e.g., the effects of magnetospheric charges and currents on the gaps, and cannot perfectly fit the data. We use this discrepancy to set the scale of our parameter constraints by rescaling the likelihood by its best fit value, reduced by one-half the degrees of freedom (dof). In the Gaussian approximation, this is equivalent to setting the best-fit reduced $\chi^2=1$. With this normalization, the $\chi^{2}$ differences between the TPC and OG models are not significant. Also shown in Figure \[fig:lcmodeling\] are the closest approaches to the magnetic axis under the best-fit TPC and OG light curves. We note that our modeling finds different locations for the magnetic axis under the two geometries, due to the ambiguity in defining the first and second [$\gamma$]{}-ray peaks when the phase separation is close to 0.5. Using these geometric models, the light curves obtained when reflecting across $\alpha = 90^\circ$ and $\zeta = 90^\circ$ are the same, but shifted by 0.5 in phase. We report confidence regions for $\alpha < 90^\circ$ and $\zeta < 90^\circ$, but note that identical, reflected regions exist in the other three quadrants. From the simulations, we can estimate the beaming correction factor $f_{\Omega}$ [@Watters2009] which relates emission along a given line-of-sight to the total flux. The values for the models are given in Table \[tab:fitpars\] with estimated uncertainties based on the confidence regions in the $\alpha$–$\zeta$ plane. Both models suggest that this factor should be $\sim$1, implying corrected [$\gamma$]{}-ray efficiencies close to 10%. [l c c c c c c c c]{} TPC & $80^{+9}_{-3}$ & $86^{+2}_{-14}$ & . . . & 0.0$\pm$2.5 & 121.0 & 54 & $0.89^{+0.15}_{-0.20}$\ OG & $88^{+1}_{-17}$ & $74^{+14}_{-4}$ & 0.0$\pm$10.0 & 0.0$\pm$2.5 & 123.1 & 53 & $0.95^{+0.10}_{-0.37}$ In addition to standard TPC and OG models, we have explored two alternative geometries. In the first approach, a very low-altitude radio cone (low-altitude slot gap radio geometry) was invoked in conjunction with the usual TPC model for the [$\gamma$]{} rays. This was done in the context of a radio cone producing peaks leading the caustic [$\gamma$]{}-ray peaks. The second alternative, motivated by the idea that the radio may indeed have a dominating leading peak with the radio profile lagging the [$\gamma$]{}-ray light curve, assumed conal radio *and* [$\gamma$]{}-ray geometries, with the [$\gamma$]{}-ray cone being *lower* than that of the radio. The fits from these alternative approaches were not satisfactory, as they were unable to reproduce the observed radio and [$\gamma$]{}-ray peak shapes and separations, leading us to abandon these scenarios. However, they point to interesting avenues of model refinement, e.g., investigation of cones with non-uniform emissivities (such as patchy or one-sided radio cones), or non-aligned [$\gamma$]{}-ray and radio cones. Such refinements, which are expected to lead to improved light curve fits, are beyond the scope of the current paper and will be developed in future work [@Seyffert2013]. We remark that the fitting results presented in this section were obtained by assuming that the observed radio emission from [PSR J0737$-$3039A]{} originates from both magnetic poles. It has been proposed that the non-$180^\circ$ separation of the two radio peaks and the overall symmetry of the profile imply that the radio emission originates from magnetic field lines associated with a single pole [see, e.g., @Manchester2005]. However, @Ferdman2013 showed that this scenario would imply a high altitude origin for the radio emission in the pulsar’s magnetosphere. If both the radio and the [$\gamma$]{} rays originate from the outer magnetosphere, we would expect the pulses to be aligned in phase [@Venter2012]. The two-pole origin for the radio emission therefore seems to be a more natural solution. Polarization Fitting {#sec:pol} -------------------- When attempting to understand pulsar viewing geometry, radio polarization can provide very powerful constraints. Although recycled pulsars are notoriously difficult to model, [PSR J0737$-$3039A]{} shows substantial linear polarization structure, and thus offers good prospects of constraining the orientation. In our polarization modeling we follow the conventional assumption that the electric vector position angle (P.A.) follows the projection onto the plane of the sky of the magnetic field line at the emission point. The P.A. sweep as one moves past the radio pole(s) thus probes the viewing geometry, with particularly strong constraints on $\beta=\zeta-\alpha$. Our modeling extends beyond the point dipole rotating vector model [RVM; @Radhakrishnan1969], to follow the distortion of the magnetic field as the emission zone moves to a non-negligible fraction of $R_{\rm LC}$ [@Craig2012]. In addition, we follow @Karastergiou2009 in treating “orthogonal mode jumps” for which the P.A. shifts by $\pm 90^\circ$, and optionally account for the effects of interstellar scattering (negligible for [PSR J0737$-$3039A]{}). For details see @Craig2013. [c c c c c c c c c]{} ${98.8}^{+8}_{-1.5}$ & ${95.8}^{+13.2}_{-4.3}$ & $0.01^{+0.22}_{-0.01}$ & $0.11^{+0.49}_{-0.05}$ & $48$ & $35$ & $0.443^{+0.008}_{-0.055}$ Figure \[fig:polarfit\] shows a polarimetric profile for [PSR J0737$-$3039A]{} measured at 1.4 GHz with the Parkes radio telescope (R. N. Manchester, private communication). The Stokes parameters and Gaussian decomposition of the linear intensity components are displayed in the top panel, while the lower panel shows the P.A. data. In this section we refer to the brightest radio peak as P2. Both it and P1 are multi-component. In polarization fitting, these components are interpreted as different emission zones with different linear and circular polarization fractions and, possibly, different orthogonal mode states and emission altitudes. Progressively across the P1 pulse we see a component with strong polarization at nearly constant P.A., a largely unpolarized peak, a component with rapid linear polarization sweep after the maximum and a separate, weakly linear polarized component on the flat pulse tail. For P2 the pattern appears reversed, with a weak linear tail leading the pulse, a linear component with rapid P.A. sweep, an unpolarized peak and then a strongly linearly polarized component with little P.A. sweep. The flat P.A. components at the front of P1 and the back of P2 are not naturally produced by any component locked to a local magnetic field. We suspect that these components are controlled by magnetospheric plasma, either in fixing the P.A. directly or in strongly distorting the field at very high altitude. Here we fit the rapid P.A. sweep in the central linearly polarized component of each pulse. The parameters are $\alpha$, $\zeta$, $\phi_0$ (the phase of the total intensity peak with respect to the surface dipole magnetic axis), $\Psi_0$, the P.A. offset and the two emission heights $r_1$ and $r_2$ in units of $R_{\rm LC}$. The fit values, the $\chi^{2}$ and the full projected (multi-parameter) uncertainty ranges on the fit parameters are shown in Table \[tab:polarfitpars\]. Figure \[fig:polarchisq\] shows a projection of the $\chi^2$ surface of the fit onto the $\alpha$–$\zeta$ plane. The best-fit P.A. models for the two poles are displayed in the lower panel of Figure \[fig:polarfit\]. Note that the weak trailing linear component in P1 is separated from the central component by a zero in the linear intensity L, and so is very likely an orthogonal mode jump. The jumped P.A. points cluster around the model prediction. We do not attempt to model the total intensity light curve or the P.A. behavior of the flat P.A. components leading P1 and trailing P2, since these do not follow simple magnetospheric models. However, we can make some comments about the origin of this emission. First, the circles mark the phase ranges of the open zone at the altitude fit for the two central components. These components occupy the second half of the open zone at this altitude, terminating at the closed zone boundary, an excellent consistency check. However, the wings of the pulses, especially at large $\phi$, lie beyond the circles and hence must either come from higher altitude or else arise from the classical closed zone. Our picture of each pole is thus a cut through a hollow cone with the central component arising from low altitude and higher altitude components from the pulse wings. Since P1 has the smallest $|\beta|=3^\circ$, it is viewed at lower altitude than P2 with $\beta=14\fdg6$. These fits also determine the phase of closest approach of the surface dipole axis. This can be compared with predictions of the outer magnetosphere [$\gamma$]{}-ray models at the fit $\alpha$ and $\zeta$. We follow @RomaniWatters2010 in examining the $\alpha$ and $\zeta$ for the best fits to the [$\gamma$]{}-ray light curve and then checking against polarization constraints. For the OG model the best light curve matches are found very close to the angles fit to the polarization data. For the TPC model (here limited to $0.75 R_{\rm LC}$), the best fits also occur near $(90^\circ,\, 90^\circ)$. These are somewhat worse than the best OG values, since they include too much unpulsed emission at all phases, but they are quite acceptable. In this regard the polarization fitting agrees well with the fits of Section \[sec:lcmodeling\], when transformed to the $\alpha>90^{\circ}$ and $\zeta>90^{\circ}$ region. However, the very accurate phase measurements from the P.A. sweep highlight a tension between the two bands: the radio-determined pole is always $\sim 0.1$ in phase before the best-fit axis for the [$\gamma$]{} pulse. This tension is also evident in the fits to the cruder pulse profiles of Section \[sec:lcmodeling\]. However, non-vacuum effects provide a plausible explanation for this offset. In @Kalapotharakos2012 it was noted that the primary effect on the high altitude pulse of increasing the magnetosphere conductivity was a growing lag in phase with respect to the surface dipole. Lags of $\delta \phi \sim 0.1$ were seen for high conductivity magnetospheres. If we disconnect the [$\gamma$]{}-ray and radio zero-phase definitions in the joint light curve fits, letting $\delta$ be larger by as much as 0.1 in phase, the best-fit $\alpha$ and $\zeta$ values changed by $\leq4^{\circ}$, generally making the agreement between the two methods better. In Figure \[fig:lcmodeling2\] we show the predicted OG and TPC [$\gamma$]{}-ray light curves at the pulsar geometry determined from the polarization study, superimposed on the radio and the [$\gamma$]{}-ray data, and using the phase of closest approach to the surface dipole axis as defined by the polarization fits. Both modeled light curves are acceptable, with the OG one being somewhat better (reduced $\chi^2$ of 1.45 compared to 1.56 for the TPC model). For these $\alpha$ and $\zeta$ angles we find beaming correction factors $f_\Omega$ of 0.92 for the OG model and 1.15 for the TPC model, similar to the ones given in Table \[tab:fitpars\] and therefore also implying [$\gamma$]{}-ray efficiencies of $\sim 10\%$. Discussion {#sec:discussion} ========== The Low Spectral Cutoff Energy and Magnetic Field at the Light Cylinder of [PSR J0737$-$3039A]{} ------------------------------------------------------------------------------------------------ In all outer-magnetospheric models, we expect the accelerating electric field $E_{||}$ to scale as [e.g., @FermiVela2]: $$\begin{aligned} E_{||} \propto C(r) B_{\rm LC} w_{\rm acc}^2,\end{aligned}$$ with $C(r)$ some function of radius, and $w_{\rm acc}$ the accelerating gap width. On the other hand, if leptons are accelerated in the radiation-reaction-limited regime, i.e., when the gain in energy due to acceleration by the electric field is balanced by curvature radiation losses, we expect that the spectral cutoff energy should scale as: $$\begin{aligned} E_{c} \propto E_{||}^{3/4} \rho_c^{1/2} \propto C(r)^{3/4} w_{\rm acc}^{3/2} B_{\rm LC}^{3/4} \rho_c^{1/2},\end{aligned}$$ with $\rho_c$ the radius of curvature (at the position where the photons of maximal energy are produced). If we take $C(r)$, $\rho_c$ and $w_{\rm acc}$ constant for all MSPs (e.g., independent of $P$ and $\dot{P}$), this leads us to expect: $$\begin{aligned} E_{c} \propto B_{\rm LC}^{3/4}.\end{aligned}$$ One does see a trend when plotting published values of $E_{c}$ versus $B_{\rm LC}$ for all known gamma-ray MSPs, but it is much weaker than expected, and the index is much smaller than 0.75, implying that our assumption of constant $C(r)$, $\rho_c$ and $w_{\rm acc}$ across the MSP population is too simplistic. [PSR J0737$-$3039A]{} is a mildly recycled MSP, having a relatively long $P$ and hence the lowest $B_{\rm LC} \propto P^{-5/2}\dot{P}^{1/2}$ of all currently known gamma-ray MSPs. Assuming that the above empirical trend may be extrapolated down to low $B_{\rm LC}$ values, one would expect a relatively low cutoff energy. This is indeed seen, although [PSR J0737$-$3039A]{}’s cutoff energy falls significantly below even the extrapolated trend. Apart from invoking the weak empirical trend discussed above, one may more generally observe that the relatively long period of [PSR J0737$-$3039A]{} implies a lower acceleration potential and hence a lower spectral cutoff than for typical MSPs, as is observed from the spectral analysis. Constraints on the System’s Formation History --------------------------------------------- From the modeling of the radio and [$\gamma$]{}-ray light curves of [PSR J0737$-$3039A]{} under the OG and the TPC geometries, and the modeling of radio polarization data (see Section \[sec:geom\]), we constrained the magnetic inclination angle $\alpha$ and the viewing angle $\zeta$ to be close to 90$^\circ$. The best-fit $\alpha$ and $\zeta$ angles obtained from the polarization study (see Table \[tab:polarfitpars\]) agree well with those obtained from the light curve modeling (Table \[tab:fitpars\]). The conclusion that $\alpha$ and $\zeta$ are close to $90^\circ$ is in line with the results of @Ferdman2013, based on the analysis of six years of radio observations of [PSR J0737$-$3039A]{}, revealing no significant variations of its radio profile with time. @Ferdman2013 find $\alpha$ and $\zeta$ values of $90\fdg2^{+16.3}_{-16.2}$ and $90\fdg8^{+0.27}_{-0.46}$, respectively. These values were calculated assuming that radio emission is seen from both magnetic poles, that the orbital inclination is $88\fdg7$, and averaging the results from all radio pulse heights used (30% to 50%) and all observation epochs. By virtue of the relationship between $\zeta$ and the misalignment angle between the spin axis and the orbital angular momentum, $\delta_{SO}$ [cf. Equation (3.36a) of @Damour1992], our values for $\zeta$ close to $90^\circ$ imply small $\delta_{SO}$ values, unless the pulsar is at a special precession phase. Nevertheless, the fact that @Ferdman2013 find a 95% confidence upper limit on $\delta_{SO}$ of $\sim 3\fdg2$ from their observations taken between 2005 June and 2011 June indicates that the pulsar was very unlikely to be at a special precession phase during the *Fermi* observation. Our constraints on $\zeta$ therefore provide an independent confirmation that $\delta_{SO}$ is close to 0$^\circ$. Since the spin axis of pulsar A and the angular momentum of the binary are likely to have become aligned through the accretion of matter from pulsar B’s progenitor star by pulsar A, a value close to 0$^\circ$ for the misalignment angle $\delta_{SO}$ supports the scenario under which a small kick was imparted to the system by the supernova explosion of pulsar B’s progenitor star. An electron-capture supernova, resulting from a collapsed O-Ne-Mg core as proposed by @Podsiadlowski2005, could provide such a formation scenario for pulsar B [see @Ferdman2013 and references therein for a detailed discussion]. Because [PSR J0737$-$3039A]{} is a faint [$\gamma$]{}-ray emitter, we were not able to investigate any evolution of the [$\gamma$]{}-ray profile as a function of time caused by a variation of the line-of-sight angle, $\zeta$, in the four years of LAT data considered here. Since the misalignment angle $\delta_{SO}$ has been shown to be close to 0$^\circ$, changes in $\zeta$ are likely small, but are still expected. An exciting prospect, when significantly more *Fermi* LAT data are available, will be to monitor the evolution of the [$\gamma$]{}-ray profile of [PSR J0737$-$3039A]{}. To date, no [$\gamma$]{}-ray pulsar light curves have even been observed to vary with time. We are very grateful to Richard N. Manchester (ATNF) for providing valuable comments which helped to improve the manuscript, for providing us with the Parkes polarization data analyzed in this paper, and for carrying out independent cross-checks of the relative alignment of the radio and [$\gamma$]{}-ray profiles using Parkes timing data taken concurrently with the *Fermi* dataset. The Parkes radio telescope is part of the Australia Telescope, which is funded by the Commonwealth of Australia for operation as a National Facility managed by the Commonwealth Scientific and Industrial Research Organisation (CSIRO). We also wish to express our gratitude to Ismaël Cognard (LPC2E, Orléans) for providing Nançay timing solutions for PSR J0737$-$3039A, that were used for searching for pulsations before this manuscript was written. The Nançay Radio Observatory is operated by the Paris Observatory, associated with the French Centre National de la Recherche Scientifique. The *Fermi* LAT Collaboration acknowledges support from a number of agencies and institutes for both development and the operation of the LAT as well as scientific data analysis. These include NASA and DOE in the United States, CEA/Irfu and IN2P3/CNRS in France, ASI and INFN in Italy, MEXT, KEK, and JAXA in Japan, and the K. A. Wallenberg Foundation, the Swedish Research Council and the National Space Board in Sweden. Additional support from INAF in Italy and CNES in France for science analysis during the operations phase is also gratefully acknowledged. Portions of this research performed at NRL are sponsored by NASA DPR S-15633-Y. [^1]: A list of [$\gamma$]{}-ray pulsar detections is available at\ https://confluence.slac.stanford.edu/display/GLAMCOG/Public+List+of+LAT-Detected+Gamma-Ray+Pulsars. [^2]: http://fermi.gsfc.nasa.gov/ssc/data/analysis/scitools/overview.html [^3]: See http://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html

ArXiv

--- abstract: 'In [@BCEN], the authors consider a random walk $(Z_{n,1},\ldots,Z_{n,K+1})\in \ZZ^{K+1}$ with the constraint that each coordinate of the walk is at distance one from the following one. A functional central limit theorem for the first coordinate is proved and the limit variance is explicited. In this paper, we study an extended version of this model by conditioning the extremal coordinates to be at some fixed distance at every time. We prove a functional central limit theorem for this random walk. Using combinatorial tools, we give a precise formula of the variance and compare it with the one obtained in [@BCEN].' address: 'Institut Camille Jordan, CNRS UMR 5208, Université de Lyon, Université Lyon 1, 43, Boulevard du 11 novembre 1918, 69622 Villeurbanne, France.' author: - Thibault Espinasse - 'Nadine Guillotin-Plantard' - Philippe Nadeau title: A combinatorial approach to a model of constrained random walkers --- Introduction and results ======================== Central limit theorems for additive functionals of Markov chains have attracted continuing interest for over half a century. One approach, due to Gordin [@Gor], rests on martingale approximation. Roughly speaking in good cases the asymptotic normality for functionals of Markov chain is derived from the central limit theorem for martingales. Various conditions for the central limit theorem to hold are now known and several expressions for the asymptotic variance can be found in the literature (see for instance [@HR]), however it is usually quite difficult to compute it theoretically. In this note, we are faced with this problem in the study of a generalized version of the so-called “prisoners model” introduced in [@BCEN]. We are able to give an explicit formula for the asymptotic variance of the Markov chain we are interested in. However our approach is not robust enough to give a full description of the asymptotic variance for more general models (see Remark (iii) in Section \[sec:conclusion\]). In what follows we will use the notation $[\![1;n]\!] :=\{1,\ldots,n\}$. Let $K$ be a positive integer. For any $h\in[\![0;K]\!]$, we define $$\mathcal{C}_{K, h}=\{ z\in \ZZ^{K+1}; \forall i\in [\![1;K]\!], |z_{i+1}-z_i| =1 \mbox{ and }z_{K+1}- z_1= h\}.$$ The set $ \mathcal{C}_{K, h}$ is empty unless $K-h$ is even; so let us set $K-h=2{g}$. The set $\mathcal{C}_{K,0}$ corresponds to the Bernoulli bridges with length $K$. We can define a neighbourhood structure on $\mathcal{C}_{K,h}$ through the following relation $$\forall z,z' \in \mathcal{C}_{K,h}, z \sim z' \Leftrightarrow \forall i \in [\![1;K+1]\!], |z_i-z'_i| = 1.$$ The set of neighbors of $z\in\mathcal{C}_{K,h}$ will be denoted by $\Gamma(z)$ and its cardinality by $\deg_{\mathcal{C}_{K,h}}(z)$. Let us denote by $(Z_n)_{n\geq 0}$ the Markov chain defined on $\mathcal{C}_{K,h}$ corresponding to $K+1$ simple random walks on $\ZZ$ under the shape constraint. In other words, $(Z_n)_{n\geq 0}$ is the Markov chain with state space $\mathcal{C}_{K,h}$ and transition probabilities given by $$\PP[Z_{n+1} = z' | Z_{n} = z] = \left\{ \begin{array}{ll} \frac{1}{\deg_{\mathcal{C}_{K,h}}(z)} & \text{if} \ z'\sim z,\\ \ \ 0 & \text{otherwise.} \end{array} \right.$$ We will denote by $Z_{n,i}$ the $i^{\text{th}}$ coordinate of $Z_n$ and by $\lfloor x \rfloor$ the integer part of a real number $x$. We assume that $Z_{0,1}=0$ almost surely. In this paper we are interested in the distributional limit of the first coordinate $(Z_{n,1})$ as $n$ tends to infinity. Actually, the convergence to the Brownian motion is not surprising, and we are mainly interested here in the exact value of the variance. \[Th1\] The sequence of random processes $\left(\frac{Z_{\lfloor nt\rfloor,1}}{\sqrt{n}}\right)_{t\geq 0, n\geq 1}$ weakly converges to the Brownian motion with variance $$\sigma_{K,h}^2=\frac{1}{K} \frac{A_{K,h}}{B_{K,h}}$$ where $$A_{K,h}= \sum_{k=0}^{{\left\lfloor {K/2} \right\rfloor}} \binom{K}{2k+1} \binom{K-2k}{{g}-k}$$ and $$B_{K,h}=\sum_{k=0}^{{\left\lfloor {K/2} \right\rfloor}} \binom{K}{2k}\ \binom {K-2k}{{g}-k}.$$ The variance $\sigma_{K,h}^2$ has the following properties: - For any $h\in [\![0;K]\!]$, as $K$ tends to infinity, $$\sigma_{K,h}^2 \sim \frac{2}{K}.$$ - For any $K$, $$\sigma_{K,0}^2 < 2 / K .$$ - For $K$ large enough, $$\sigma_{K,0}^2 > 2 / (K+2).$$ With our notations, the variance obtained in [@BCEN] is equal to $2/ (K+2)$, so item iii) means that for large $K$, our variance is larger than the one obtained in the unconstrained case. We conjecture that this should hold for any $K$. The paper is organized as follows: in Section \[sec:comb\], various enumerative results about walks are proved. They will be used in Section \[sec:proba\]. Theorem \[Th1\] is proved in Section \[sec:proba\]. Our approach is based on a decomposition of $Z_{n,1}$ as a sum of a martingale and a bounded function. We will compute explicitly the two parts, and provide a geometrical interpretation for the decomposition. The asymptotic properties of the variance will be proved using local limit theorems for random walks on $\ZZ$. In Section \[sec:unconstrained\] we explain how to adapt our method to derive the main result of [@BCEN]. In Section \[sec:conclusion\] some variations of the model are proposed and open problems are discussed. Combinatorics {#sec:comb} ============= We prove various enumerative results that will be used in Section \[sec:proba\]. If $z\in\mathcal{C}_{K,h}$, its shape is defined as $\F(z)=(F_1,\ldots,F_K)$ by $F_i=z_{i+1}-z_{i}$. We refer naturally to the elements of $\F(z)$ as the [*steps*]{} of $z$, which can be either up or down. The set of possible shapes is given by $$Sh_{K,h}=\{(F_i)_i\in\{\pm 1\}^{K}; \sum_{i=1}^KF_i=h\}.$$ Clearly $z\in\mathcal{C}_{K,h}$ if and only if $(z_1,\F(z))\in\ZZ\times Sh_{K,h}$. Let us denote, for every $z\in\mathcal{C}_{K,h}$, $$\Gamma^{\pm }(z)=\{ z'\sim z; z_1'-z_1= z_{K+1}'-z_{K+1} = \pm 1\}.$$ ![Two paths $z\sim z'$ with $z'\in\Gamma^+(z)$ \[Figure\_Chemins\]](Figure_Chemins) Now let $z,z'\in\mathcal{C}_{K,h}$ be such that $z'\in\Gamma^+(z)$, and let $F=\F(z),F'=\F(z')$ be the shapes of $z,z'$ respectively. For any $i\in \{1,\ldots,K\}$, define $M_i=\pm 1$ if $F_i=F'_i=\pm 1$ and $M_i=0$ otherwise. So $M=(M_1,\ldots,M_{K})$ has values in $\{1,0,-1\}$, and can thus be represented as a Grand Motzkin path; we define $\Phi^+(z,z')=M$. Note that $\sum_i M_i=h$, so this path $M$ goes from $(0,0)$ to $(K,h)$; we note $\mathcal{M}_{K,h}$ the set of such sequences, which has cardinality $$\label{GMP} |\mathcal{M}_{K,h}|=\sum_{k=0}^{{\left\lfloor {K/2} \right\rfloor}}\binom{K}{2k}\binom{K-2k}{{g}-k}.$$ Now if we fix the starting point of $z$, then from any element of $\mathcal{M}_{K,h}$ we can clearly reconstruct a pair $z,z'$, so we have the following result \[prop:bijection\_motzkin\] For any $z_1$, $\Phi^+$ is a bijection between $\{(z,z')\in\mathcal{C}_{K,h}^2; z'\in\Gamma^+(z)\}$ and $\mathcal{M}_{K,h}$. Denote by $A(z)$ the algebraic area between the path $z$ and the $X$-axis. Note that $$A(z) = \frac{1}{2}z_{1}+ z_2 + \cdots + z_{K} + \frac{1}{2}z_{K+1}$$ ![\[Figure\_Area\] The area $A(z')-A(z)$ is equal to $-1+4-3+3=3$](Figure_Area) We now prove that for any fixed path $z\in \mathcal{C}_{K,h} $ the sum over $z'\in\Gamma(z)$ of the algebraic areas boils down to zero. Fix a path $z\in \mathcal{C}_{K,h} $. Let $(z',p)$ be a [*marked path*]{} (with respect to $z$), which we define as a path $z'\in\Gamma(z)$ together with a step $p\in z'$ which does not cross $z$. Define $sign(z;z',p)=1$ (*resp. $-1$*) if the step $p$ occurs above (*resp.* below) $z$. With these notations it is clear that $A(z')-A(z)=\sum_p sign(z;z',p)$, where the sum is over all marking steps in $z'$; see Figure \[Figure\_Area\]. We now define an involution $\I$ of these marked paths. Note first that $z'$ crosses $z$ an even number of times, and let $2k$ be this quantity. The case $k=0$ occurs only when $z'\in\{z^+,z^-\}$ which are defined as the two paths obtained from $z$ by shifting it one unit up or down. In this case define simply $\I(z;z^\pm,p)=(z;z^\mp,p)$. Suppose now $k\neq 0$, and denote by $L$ (*resp.* $R$) the nearest crossing left (*resp.* right) of $p$. There are two special cases: if $p$ occurs before the first crossing of $z'$, then $L$ is defined as the last crossing of $z'$, while if $p$ occurs after the last crossing of $z'$, then $R$ is defined as the first crossing of $z'$. Examples of the definition of $L,R$ are illustrated in Figure \[Figure\_Involution\]. Now exactly one step among $\{L,R\}$ is a step of $z'$ of a different up/down type as $p$. Let $q$ be this crossing step, and exchange $p$ with $q$ in $z'$, keeping all other steps unchanged. This defines a new path $z''\in \Gamma(z)$, in which $p$ becomes crossing while $q$ is noncrossing. So $(z'',q)$ is a marked path, and we define $\I(z;z',p):=(z;z'',q)$. We have the following lemma, whose proof is immediate. For any $z$, $\I$ is an involution on marked paths which is sign reversing. ![Two cases of the application of the involution $\I$. \[Figure\_Involution\]](Figure_Involution){width="90.00000%"} \[prop:zerosum\] For any $z\in\mathcal{C}_{K,h}$, $\displaystyle{ \sum_{z'\in \Gamma(z)} \left(A(z') - A(z)\right) = 0.}$ The sum to compute is equal to the sum of $sign(z;z',p)$ for all marked paths $(z',p)$. By the previous lemma, the involution $\I$ pairs such marked paths two by two with opposite signs, so the sum boils down to zero. \[prop:total\_area\] For any $K,h$, $\displaystyle{\sum_{\stackrel{z,z'\in\mathcal{C}_{K,h}}{z'\in\Gamma^+(z)}}[ A(z') - A(z)]=\sum_{k=0}^{{\left\lfloor {K/2} \right\rfloor}} \binom{K}{2k+1} \binom{K-2k}{{g}-k}}.$ The l.h.s. can be rewritten as the sum of $sign(z;z',p)$ over all $z,z',p$ such that $z'\in\Gamma^+(z)$ and $(z',p)$ is marked. Define $(z,z'',q):=\I(z,z',p)$. If $z''$ also belongs to $\Gamma^+(z)$, then the corresponding terms in the sum will cancel each other. The surviving terms correspond to $z''\in\Gamma^-(z)$, which by inspection occurs precisely in the following cases: 1. \[it1\] $p$ occurs before the first crossing, and $p$ is a descent; 2. \[it2\] $p$ occurs after the last crossing, and $p$ is an ascent; 3. \[it3\] $z'=z^+$ and $p$ is any step. In all three cases, $sign(z,z',p)$ is always equal to $1$. Thus we need to enumerate the triplets $(z;z',p)$ verifying , or . There are clearly $K \binom{K}{{g}}$ triplets satisfying . We now compute the number of triplets $(z,z',p)$ verifying case or , and such that there are $2k>0$ crossings between $z$ and $z'$. The position of $p$ together with the position of these crossings form a subset $\{j_1<j_2<\cdots<j_{2k+1}\}$ of $\{1,\ldots,K\}$ ; here the position of $p$ is $j_1$ in case and $j_{2k+1}$ in case . There are $\binom{K}{2k+1}$ such subsets, and, conversely, the knowledge of such a subset determines the positions of $p$ and the crossings of $z$ and $z'$ in each case. The remaining $K-2k-1$ steps are the same in $z$ and $z'$, so we only need to count possible steps for $z$. We want to ensure that $z$ belongs to $\mathcal{C}_{K,h}$. Notice that among the crossings there are $k$ up steps and $k$ down steps. Now in case , $p$ is a down step, so there must be exactly $g-k-1$ down steps among the remaining $K-2k-1$ steps; in case , $p$ is an up step, so there must be exactly $g-k$ down steps among the remaining $K-2k-1$ steps. In total, this represents $\binom{K-2k-1}{g-k-1}+\binom{K-2k-1}{g-k}=\binom{K-2k}{g-k}$ possibilities. Adding everything up we obtain the desired expression (note that the case $k=0$ corresponds to ). Proof of Theorem \[Th1\] {#sec:proba} ======================== Martingale ---------- Denote by $(\mathcal{F}_n)_{n\geq 0}$ the natural filtration of the Markov chain $Z_n$. \[Le1\] The sequence $(A (Z_n))_{n\geq 0}$ is a martingale with respect to the filtration $(\mathcal{F}_n)_{n\geq 0}$. Since $(Z_n)_n$ is a Markov chain, it is enough to prove that for any $z\in\mathcal{C}_{K,h}$, $$\mathbb{E}[ A(Z_{n+1}) - A(Z_n) | Z_n=z]=0$$ or equivalently (since the Markov chain moves uniformly on its neighbours) for any $z\in\mathcal{C}_{K,h}$, $$\sum_{z'\sim z} [ A(z') - A(z)] = 0.$$ We recognize here the content of Proposition \[prop:zerosum\]. The coupling ------------ Denote by $f_K(z)$ the area between a path $z=(z_1,z_2,\ldots, z_{K},z_{K+1})\in \mathcal{C}_{K,h}$, and the segment $(z_1, z_1+1, ...., z_1+K)$. Then, for any $z\in \mathcal{C}_{K,h}$ (see Figure \[Figure\_Coupling\]), $$\label{aire} f_K(z) +A(z) = K z_1 + \frac{K^2}{2}.$$ ![Relation between the first coordinate $z_1$ and the area under the path. \[Figure\_Coupling\]](Figure_Coupling){width="20.00000%"} Since $f_{K}$ is bounded by $K^2$, the position of the first walker and the area under the path can be asymptotically related as $$\label{aire2} \frac{Z_{[nt],1}}{\sqrt{n}} = \frac{1}{K}\frac{ A (Z_{[nt]})}{\sqrt{n}}+o(1).$$ Theorem \[Th1\] will then be deduced from the functional central limit theorem for the rescaled martingale $(A(Z_{n})/\sqrt{n})_{n\geq 1}.$ If $z\in\mathcal{C}_{K,h}$, its shape is defined as $\F(z)=(F_1,\ldots,F_{K+1})$ by $F_i=z_{i}-z_{1}$. The set of possible shapes is given by $$\overline{Sh}_{K,h}=\{z\in\mathcal{C}_{K,h}; z_1=0\}.$$ This description of the possible shapes of $\mathcal{C}_{K,h}$ differs from the one used in Section \[sec:comb\] but will be more convenient in the next computations. Note that a natural neighbourhood structure is given on $\overline{Sh}_{K,h}$ by the relation $$F \sim_S F' \Leftrightarrow \forall i \in [\![1;K+1]\!], F_i - F'_i \in \{0,2\} \text{ or } \forall i \in [\![1;K+1]\!], F_i - F'_i \in \{0,-2\} .$$ If $F,F'$ are neighbours in $\overline{Sh}_{K,h}$, then one of following statements is true: - If $F_i - F'_i \in \{0,2\}$ then $F-(1, \cdots, 1)$ and $F'$ are neighbours in $\mathcal{C}_{K,h}$. - If $F_i - F'_i \in \{0,-2\}$ then $F+(1, \cdots, 1)$ and $F'$ are neighbours in $\mathcal{C}_{K,h}$. Note that the special case $F = F'$ gives two neighbours in $\mathcal{C}_{K,h}$ while the others cases provide a one-to-one transformation.\ Conversely, if $z,z'$ are neighbours in $\mathcal{C}_{K,h}$, then either $\F(z) = \F(z')$ and $z = z' \pm (1, \cdots, 1)$, or $\F(z)$ and $\F(z')$ are distinct neighbours in $\overline{Sh}_{K,h}$. We denote by $\deg_S(F)$ the number of neighbours of $F\in \overline{Sh}_{K,h}$, i.e. $$\deg_S(F) := \sharp \left\{G \in \overline{Sh}_{K,h}; G \sim_S F \right\}.$$ According to the previous remark, we have $$\deg_{\mathcal{C}_{K,h}}(z) = \deg_S(\F(z))+1.$$ Consider the Markov chain $(F_n)_{n}$ with values in $\overline{Sh}_{K,h}$, with transition probabilities given by $$\displaystyle p(F,G) = \left\{ \begin{array}{lll} \frac{1}{ \deg_S(F) +1} & \mbox{ if } & G \sim_S F \mbox{ and } G\neq F, \\ \frac{2}{\deg_S(F)+1} & \mbox{ if } & G=F. \end{array} \right.$$ This irreducible and ergodic Markov chain is reversible with stationary measure $$\pi_{S}(F) := \frac{\deg_S(F) +1} { \sum_{F\in\overline{Sh}_{K,h} } (\deg_S(F) +1)}.$$ The Markov chains $(\F(Z_n))_{n \in \mathbb{N}}$ and $(F_n)_{n \in \mathbb{N}}$ are then identically distributed. Indeed, $$\begin{aligned} \mathbb{P}[\F(Z_{n+1}) = G|Z_n = z]& = \frac{1}{\deg_{\mathcal{C}_{K,h}}(z)}\sum_{z'\sim z} 1_{\F(z') = G} \\& =\left\{ \begin{array}{ll} \frac{1}{\deg_{S}(\F(z))+1} & \text{if} ~ G \sim_S \F(z), G \neq \F(z), \\ \frac{2}{\deg_{S}(\F(z))+1} & \text{if} ~ G =\F(z) \end{array} \right. \\ & = \mathbb{P}[F_{n+1} = G|F_n = \F(z)] \end{aligned}$$ Hence, $$\begin{aligned} \mathbb{P}[\F(Z_{n+1}) = G|\F(Z_n) = F]& = \sum_{z \in \mathcal{C}_{K,h}} \mathbb{P}[\F(Z_{n+1}) = G|\F(Z_n) = F; Z_n = z]\mathbb{P}[Z_n = z |\F(Z_n) = F] \\ & = \mathbb{P}[F_{n+1} = G|F_n = F] \Big(\sum_{z \in \mathcal{C}_{K,h}} \mathbb{P}[Z_n = z |\F(Z_n) = F] \Big) \\ & = \mathbb{P}[F_{n+1} = G|F_n = F].\end{aligned}$$ Let us remark that for every $z,z' \in \mathcal{C}_{K,h}$ s.t. $z'\sim z$ with $z'\neq z$, the difference of area $|A(z')-A(z)|$ can be computed with respect to $\F(z')$ and $\F(z)$ only. Indeed, we have the following relations $$\begin{aligned} |A(z') - A(z)|&=& \left\{ \begin{array}{lll} |A(\F(z')) - A(\F(z)+(1,\cdots,1)) | & \text{ if } \exists i \text{ s.t. } \F(z')_{i} -\F(z)_{i} = 2,\\ |A( \F(z')) - A(\F(z)-(1,\cdots,1) ) | & \text{ if } \exists i \text{ s.t. } \F(z')_{i} -\F(z)_{i} = -2,\\ \hspace{2.5cm} K & \text{ if } \F(z')=\F(z). \end{array} \right.\end{aligned}$$ The quantity $|A(z')-A(z)|$ will then be denoted by $\delta A(\F(z),\F(z'))$. From the central limit theorem for martingales (see for instance [@Rick], Theorem 7.4, p.374), we have to compute the a.s. limit as $n$ tends to infinity of $$\begin{aligned} \Sigma_n &:=& \frac{1}{n} \sum_{k=1}^{n} \mathbb{E}[ (A(Z_k) -A(Z_{k-1}) )^2 |Z_{k-1}]\\ &=& \frac{1}{n} \sum_{k=1}^{n} \mathbb{E}[ \delta A(\F(Z_{k-1}),\F(Z_k))^2 |\F(Z_{k-1}); Z_{k-1,1} ]\\ &= & \sum_{F\in \overline{Sh}_{K,h}} \frac{1}{n} \sum_{k=1}^{n} 1_{ \{ \F(Z_{k-1}) = F\}} \mathbb{E}[ \delta A(\F(Z_{k-1}),\F(Z_k))^2|\F(Z_{k-1}) =F ]\\ &=& \sum_{F\in \overline{Sh}_{K,h}} \Big( \frac{1}{n} \sum_{k=1}^{n} 1_{ \{\F(Z_{k-1}) = F\}} \Big) \mathbb{E}[ \delta A(\F(Z_{0}),\F(Z_1))^2 |\F(Z_0) =F ]\\ &\stackrel{law}{=}& \sum_{F\in \overline{Sh}_{K,h}} \left( \frac{1}{n} \sum_{k=1}^{n} 1_{ \{F_{k-1} = F\}} \right) \mathbb{E}[ \delta A (F_0,F_1 )^2 |F_0 =F ]\end{aligned}$$ Since the Markov chain $(F_n)_n$ is ergodic with invariant measure $\pi_{S}$, the sequence of random variables $(\Sigma_n)_n$ converges almost surely to the constant $$\sum_{F\in \overline{Sh}_{K,h}} \mathbb{E} [ (\delta A (F_0,F_1))^2| F_0=F] \pi_S(F),$$ which can be rewritten as $\mathbb{E}_{\pi_S} [ (A( Z_1) -A (Z_0))^2]$ by remarking that $\F(Z_0)=Z_0$. Then, using (\[aire2\]), we obtain that $$\sigma_{K,h}^2 = \frac{1}{K^2} \mathbb{E}_{\pi_S} [ (A( Z_1) -A (Z_0) )^2].$$ From (\[aire\]), we deduce that $$f_K(Z_1) - f_K(Z_0) = (K-1) (Z_{1,1}- Z_{0,1} ) - [A(Z_1) - A(Z_0)].$$ From the reversibility of the Markov chain $(F_n)_n$, we have $$\mathbb{E}_{\pi_S} [ (A(Z_1) -A(Z_0) ) (f_K(Z_1) -f_K(Z_0) ) ] =0$$ Indeed, $\mathbb{E}_{\pi_S} [ (A(Z_1) -A(Z_0) ) f_K(Z_1) ]$ can be rewritten as $$\begin{aligned} & &\sum_{F\in \overline{Sh}_{K,h}}\pi_S(F) \sum_{F'\neq F} p(F,F') (A(z_1)-A(z_0)) f_K(z_1) {\bf 1}_{\{z_0\sim z_1;\F(z_0)=F;\F(z_1)=F'\}} \\ &=& \sum_{(F,F'); F\neq F'} \pi_S(F') p(F',F) (A(z_1)-A(z_0))f_K(z_1) {\bf 1}_{\{z_0\sim z_1;\F(z_0)=F;\F(z_1)=F'\}} \\ &=& - \sum_{F'} \pi_S (F') \sum_{F\neq F'} p(F',F) (A(z_0)-A(z_1))f_K(z_1) {\bf 1}_{\{z_0\sim z_1;\F(z_0)=F;\F(z_1)=F'\}} \\ &=& -\mathbb{E}_{\pi_S} [ (A(Z_1) -A(Z_0) ) f_K(Z_0) ] \end{aligned}$$ and the last expectation is zero since for any fixed shape $F\in \overline{Sh}_{K,h}$, we have $$\begin{aligned} \mathbb{E}[(A(Z_1)-A(Z_0)) f_K(Z_0)|\F(Z_0) =F] &=& \mathbb{E}[(A(Z_1)-A(Z_0)) f_K(Z_0)| Z_0 =F] \\ &= & f_K(F) \mathbb{E}[ A(Z_1)-A(z_0) | Z_0 = F] \\ &=& 0\end{aligned}$$ from Lemma \[Le1\]. Therefore, $$\sigma_{K,h}^2 = \frac{1}{K}\ \mathbb{E}_{\pi_S} [ (A(Z_1) -A(Z_0)) ( Z_{1,1}-Z_{0,1}) ].$$ From Proposition \[prop:zerosum\] and the expression of the measure $\pi_{S}$, we deduce $$\begin{aligned} \sigma_{K,h}^2 &=& \frac{1}{K}\ \frac{A_{K,h}}{B_{K,h}}\end{aligned}$$ where $$A_{K,h}= 2 \sum_{\stackrel{\small z,z'\in\mathcal{C}_{K,h}}{ z'\in\Gamma^+(z)}} (A(z')-A(z))\ \text{ and }\ B_{K,h}=\sum_{F\in\overline{Sh}_{K,h}} (\deg_S (F)+1).$$ Applying Propositions \[prop:total\_area\] and \[prop:bijection\_motzkin\] combined with (\[GMP\]), we get $$A_{K,h}= 2\sum_{k=0}^{{\left\lfloor {K/2} \right\rfloor}} \binom{K}{2k+1} \binom{K-2k}{{g}-k}\ \text{ and }\ B_{K,h}=2|\mathcal{M}_{K,h}|=2 \sum_{k=0}^{{\left\lfloor {K/2} \right\rfloor}}\binom{K}{2k}\binom{K-2k}{{g}-k}.$$ Properties of the variance -------------------------- We are interested in the properties of the variance $\sigma_{K,h}^2.$ Denote by $(S_n)_{n\geq 0}$ the random walk on $\mathbb{Z}$ starting from 0 and moving according to the following rule $$\PP[S_{n+1} = y | S_n = x] =\frac{1}{3} \mbox{ if } y \in\{x-1,x,x+1\}.$$ Then, for any positive integer $K$ and any $h\in [\![0;K]\!]$ with $K-h$ even, $$\PP[S_{K} = h]=\frac{ |\mathcal{M}_{K,h}| }{3^{K}}$$ where $|\mathcal{M}_{K,h}| $ denotes the number of Grand Motzkin paths from $(0,0)$ to $(K,h)$. The variance $\sigma_{K,h}^2$ can then be rewritten $$\label{tll} \sigma_{K,h}^2 = \frac{1}{K} \frac{\PP[S_{K}=h+1]+\PP[S_{K}=h-1]}{\PP[S_{K}=h]}.$$ Item i) directly follows from the local central limit theorem for the random walk $(S_n)_{n\geq0}$ ( see for instance Proposition P4 p.46 in [@Spi]). Remark that from the symmetry of the random walk $(S_n)_n$, we have $$\sigma_{K,0}^2 = \frac{2}{K} \frac{\PP[S_{K}=1]}{\PP[S_{K}=0]}.$$ From Fourier inversion formula, for any $x\in \ZZ$, we have $$\begin{aligned} \PP[S_n= x] &=&\frac{1}{2\pi} \int_{-\pi}^{\pi} e^{i\theta x} \EE\Big[ e^{i\theta S_n}\Big]\, d\theta\\ &=& \frac{1}{2\pi} \int_{-\pi}^{\pi} e^{i\theta x} \EE\left[ e^{i\theta S_1}\right]^n\, d\theta\\ &=& \frac{1}{2\pi} \int_{-\pi}^{\pi} e^{i\theta x} \left( \frac{1+2\cos \theta}{3} \right)^n\, d\theta\end{aligned}$$ since the increments of the random walk $(S_n)_n$ are independent and identically distributed according to the uniform distribution on the discrete set $\{-1,0,+1\}$. Therefore, we deduce $$\begin{aligned} 3^{K} \PP[S_{K}=1] &=& (2 \pi )^{-1} \left|\int_{-\pi}^{\pi} ( 1 + 2 \cos(\theta))^{K} e^{i\theta} d\theta\right|\\ &< & (2 \pi )^{-1} \int_{-\pi}^{\pi} ( 1 + 2 \cos(\theta))^{K} d\theta = 3^{K} \PP(S_{K}=0),\end{aligned}$$ which yields assertion ii). Let us now prove item iii). Since the increments of the random walk $(S_n)_n$ are symmetric and bounded, from Theorem 2.3.5 (Formula (2.23)) in [@LL], there exists a constant $c$ such that for any $n$ and $x$, $$\label{vlada} \Big| \PP[S_{n} = x] - \frac{\sqrt{3} e^{-3 x^2/(4n)}}{2\sqrt{\pi n}} \Big| \leq \frac{c}{n^{3/2}}.$$ Now concerning iii) it is enough to prove that the limit as $m$ tends to infinity of the sequence $$u_K:=\frac{2\ \PP[S_{K}=0]}{(K+2) (\PP[S_{K}=0] -\PP[S_{K}=1])}$$ is strictly greater than one. From (\[vlada\]), we have $$\begin{aligned} u_K&=&\frac{2}{K+2}\ \frac{ \frac{\sqrt{3}}{\sqrt{4\pi K}}+\mathcal{O}(K^{-3/2})}{ \frac{\sqrt{3}}{\sqrt{4\pi K}}(1- e^{-3/4K})+\mathcal{O}(K^{-3/2}) }\\ &=& \frac{2}{K+2}\ \frac{ \frac{\sqrt{3}}{\sqrt{4\pi K}}+\mathcal{O}(K^{-3/2})}{ \frac{3\sqrt{3}}{4\sqrt{4\pi} K^{3/2}}+\mathcal{O}(K^{-3/2}) }\\ &\sim & \frac{8}{3} \ \text{as}\ K\rightarrow +\infty.\end{aligned}$$ The unconstrained case {#sec:unconstrained} ====================== In this section we indicate how to adapt the ideas we developed in order to give an alternative proof of the main result of [@BCEN]. We now consider the Markov chain $(Z_n^{*})_{n\geq 0}$ with values in the state space $\mathcal{C}_{K}=\{ z\in \ZZ^{K+1}; \forall i\in [\![1;K]\!], |z_{i+1}-z_i| =1\}$, so that $\mathcal{C}_{K}$ is the union $\cup_h\mathcal{C}_{K, h}$. The difference is that both ends of the walks are allowed to move in different directions. The set of paths $\Gamma(z)=\Gamma^+(z)\sqcup\Gamma^-(z)$ is defined as before. By Proposition \[prop:bijection\_motzkin\], $\Phi^+$ is a bijection between $\{(z,z')\in\mathcal{C}_{K}^2; z'\in\Gamma^+(z)\}$ and $\mathcal{M}_{K}:=\cup_h\mathcal{M}_{K,h}$, which clearly has cardinality $3^K$ since each step can be chosen independently. Now modify the area of $z\in \mathcal{C}_{K}$ by setting $$A^{*}(z) := \frac{3}{2}z_{1}+ z_2 + \cdots + z_{K} + \frac{3}{2}z_{K+1}=z_1+A(z)+z_{K+1}$$ Extend any path $z\in\mathcal{C}_{K}$ by unit horizontal steps at both ends: then $A^{*}(z)$ is the algebraic area below this extended path. We have the following proposition which is the counterpart of Propositions \[prop:zerosum\] and \[prop:total\_area\] in this unconstrained setting. \[prop1\] For any $z\in\mathcal{C}_{K}$, $\displaystyle{ \sum_{z'\in \Gamma(z)} \left(A^{*}(z') - A^{*}(z)\right) = 0.}$ For any $K$, $\displaystyle{\sum_{\stackrel{z,z'\in\mathcal{C}_{K}}{z'\in\Gamma^+(z)}}[ A^{*}(z') - A^{*}(z)]=2\cdot 3^K}$. The difference $A^{*}(z')-A^{*}(z)$ between two paths can be written as the sum of $sign(z;z',p)$ over all marked paths $(z;z',p)$ where the mark $p$ can now also be one of the two extra horizontal steps, see figure below.  We need to define $\I^{*}$, a modification of the involution $\I$ from Section \[sec:comb\]. Given a marked path $(z;z',p)$, its image $(z;z'',q)$ by $\I^{*}$ is defined as follows. If $p$ is the initial horizontal step, then find the first crossing: $z''$ is obtained by changing $z''$ at this step, which becomes the new mark $q$ (If there is no crossing, so that $z'=z^\pm$, then change it to $z''=z^\mp$ and let $q$ be the final horizontal step). Define $\I^{*}$ symmetrically when $p$ is the final horizontal step. In all other cases, define $\I^*$ as in $\I$, except in the cases where one needs to use the special cases of $L$ and $R$: in this case the mark $q$ will be one of the horizontal steps. It is easily seen that $\I^{*}$ is bijective and reverses signs, so that the first formula of the proposition is proved. For the second one, we notice as in Section \[sec:comb\] that the l.h.s. can be written as the sum of $sign(z;z',p)$ over all marked paths satisfying $z'\in \Gamma^+(z)$ and $z''\in \Gamma^-(z)$. This happens when $p$ is the initial horizontal step for any $z,z'$ with $z'\in \Gamma^+(z)$, and these cases contribute $3^K$ to the sum. The other possibility is that $p$ is a down step which has no crossing to its left. Clearly $sign(z;z',p)=1$ here also since $z'\in \Gamma^+(z)$, and there are also $3^K$ such possibilities: indeed, their images by $\I^*$ in this case are exactly the marked paths $(z;z'',q)$ where $z''\in\Gamma^-(z)$ and $q$ is the initial horizontal step. This completes the proof. Denote by $f_K^{*}(z)$ the area between a path $z=(z_1,z_2,\ldots, z_{K},z_{K+1})\in \mathcal{C}_{K}$, and the segment $(z_1, z_1+1, ...., z_1+K+1,z_1+K+2)$. Then, for any $z\in \mathcal{C}_{K}$ (see Figure \[Figure\_Coupling\_2\]), $$\label{aire*} f_{K}^{*}(z) +A^{*}(z) = (K+2) z_1 + \frac{(K+2)^2}{2}.$$ ![Relation between the first coordinate $z_1$ and the area under the path. \[Figure\_Coupling\_2\]](Figure_Coupling_2){width="20.00000%"} Since $f_{K}^{*}$ is bounded by $(K+2)^2$, the position of the first walker and the area under the extended path can be asymptotically related as $$\label{aire2} \frac{Z_{[nt],1}^{*}}{\sqrt{n}} = \frac{1}{K+2}\frac{ A^{*} \big(Z_{[nt]}^{*}\big)}{\sqrt{n}}+o(1).$$ The first assertion in Proposition \[prop1\] implies that the sequence $(A^{*} (Z_n^{*}))_{n\geq 0}$ is a martingale with respect to the natural filtration of the Markov chain. The main result of [@BCEN] is then deduced from the functional central limit theorem for the rescaled martingale $(A^{*}(Z_{n}^{*})/\sqrt{n})_{n\geq 1}.$ The proof is similar to the one given in Section \[sec:proba\] in the constrained case. The asymptotic variance is then rewritten as $$\begin{aligned} \sigma_{K,*}^{2} &=& \frac{1}{K+2}\ \frac{A_{K}^{*}}{B_{K}^{*}}\end{aligned}$$ where $$A_{K}^{*}= 2 \sum_{\stackrel{\small z,z'\in\mathcal{C}_{K}}{ z'\in\Gamma^+(z)}} (A^{*}(z')-A^{*}(z))\ \text{ and }\ B_{K}^{*}=2 |\mathcal{M}_{K}|=2\cdot 3^K.$$ The second assertion in Proposition \[prop1\] leads to $\sigma_{K,*}^{2} = \frac{2}{K+2}.$ Discussion and open problems {#sec:conclusion} ============================ 1. The case when the distance $h$ between the extremal coordinates depends on $K$ can also be considered. It is not difficult to see that Theorem \[Th1\] still holds with $h$ replaced by $h(K)$ in the variance formula. The asymptotic behavior of the variance for $K$ large (item i) in Theorem \[Th1\]) is still true if $h(K)=o(K^{3/4})$. Indeed, relation (\[tll\]) is still valid and local limit theorem 2.3.11 in [@LL] (p.46) gives the result after elementary computations. The asymptotic behavior of the variance for $h \gg K^{3/4}$ is not known. 2. Instead of $ \mathcal{C}_{K,h}$ we can consider the following set of paths $$\mathcal{C}_{K}=\{ z\in \mathbb{Z}^{K+1}; \forall i\in [\![1;K]\!], |z_{i+1}-z_i| \in\{0,1\} \}$$ and $(Z_n)_{n\geq 0}$ the Markov chain defined on $\mathcal{C}_{K}$ corresponding to $K+1$ simple random walks on $\mathbb{Z}$ under the shape constraint. In [@BCEN], the set of paths $$\mathcal{D}_{K}=\{ z\in \mathbb{Z}^{K+1}; \forall i\in [\![1;K]\!], |z_{i+1}-z_i| =1 \}$$ and the corresponding Markov chain $(Z_n^{(K)})_{n\geq 0}$ were studied. Whatever the distribution of $Z_0^{(K)}$, the sequence of random variables $(Z_{n,1}^{(K)}/\sqrt{n})_n$ converges to a centered gaussian law with variance $2/(K+2)$. By remarking that given the number of zeroes $N_0$ of the initial random variable $Z_0$, the position of the first walkers $(Z_{n,1})_{n}$ and $(Z_{n,1}^{(K-N_0)})$ are identically distributed (with the following convention: if $N_0=K$, the first random walker evolves as the simple symmetric random walk on $\mathbb{Z}$), it follows that the sequence of random variables $(Z_{n,1}/\sqrt{n})_n$ converges in law to a mixture of gaussian distributions with variance equal to $$\sigma_K^2 = 2 \sum_{l=0}^{K} \frac{ \PP[ N_0 =l]}{K+2-l}.$$ Note that whatever the distribution of the random variable $N_0$ the variance is still greater than the one obtained in the unconstrained case considered in [@BCEN] and is rational when $Z_0$ is uniformly distributed on $ \mathcal{C}_{K}$. 3. The generalization of Theorem \[Th1\] to more general sets of paths does not seem to be obvious. Except in the case of [@BCEN] where our proof can be adapted the construction of a convenient martingale from which the computation of the variance can be deduced is far from clear. For instance consider the set of Bernoulli bridges with length $K$ with the additional constraint: $L$ of the $K+1$ random walkers stay at the same height at each step. A central limit theorem for the first random walker should also hold. We conjecture that the variance should be increasing in $L$. [**Acknowledgments.**]{} We are grateful to James Norris and Serge Cohen for stimulating discussions. [00]{} (2014) Diffusivity of a random walk on random walks. [*Random Structures and algorithms*]{}, to appear. arXiv:1210.4745. \(1991) [*Probability: Theory and examples.*]{} Wadsworth & Brooks/Cole Advanced Books & Software, Pacific Grove, CA. \(1969) The central limit theorem for stationary processes. [*Dokl. Akad.Nauk SSSR*]{} 188, 739 – 741. and [J. S. Rosenthal]{} (2007) On variance conditions for Markov chain CLTs. [*Electron. Comm. Probab.*]{} 12, 454 – 464. \(2010) [*Random walk: A modern introduction.*]{} Cambridge University Press. \(1976) [*Principles of random walks.*]{} Springer-Verlag, Second edition.

ArXiv

--- abstract: 'We analyze an optimal control problem for a fractional semilinear PDE; control constraints are also considered. We adopt the integral definition of the fractional Laplacian and establish the well-posedness of a fractional semilinear PDE; we also analyze suitable finite element discretizations. We thus derive the existence of optimal solutions and first and second order optimality conditions for our optimal control problem; regularity properties are also studied. We devise a fully discrete scheme that approximates the control variable with piecewise constant functions; the state and adjoint equations are discretized via piecewise linear finite elements. We analyze convergence properties of discretizations and obtain a priori error estimates.' author: - 'Enrique Otárola[^1]' bibliography: - 'bibliography.bib' title: ' Fractional semilinear optimal control: optimality conditions, convergence, and error analysis[^2]' --- optimal control problem, fractional diffusion, integral fractional Laplacian, regularity estimates, finite elements, convergence, a priori error estimates. 35R11, 49J20, 49M25, 65K10, 65N15, 65N30. Introduction ============ In this work we shall be interested in the analysis and discretization of a distributed optimal control problem for a semilinear, elliptic, and fractional partial differential equation (PDE). To make matters precise, we let $\Omega \subset \mathbb{R}^n$ be an open and bounded domain in $\mathbb{R}^n$ $(n \in \{2,3\})$ with Lipschitz boundary $\partial \Omega$; additional regularity requirements on $\partial \Omega$ will be imposed in the course of our regularity and convergence rate analyses ahead. Let us define the cost functional $$\label{eq:cost_functional} J(u,z) := \int_{\Omega} L (x,u(x)) \mathrm{d}x + \frac{\alpha}{2} \int_{\Omega} |z(x)|^2 \mathrm{d}x,$$ where $L: \Omega \times \mathbb{R} \rightarrow \mathbb{R}$ denotes a Carathéodory function of class $C^2$ with respect to the second variable and $\alpha > 0$ corresponds to the so-called regularization parameter. Further assumptions on $L$ will be deferred until section \[sec:assumption\]. In this work, we shall be concerned with the following PDE-constrained optimization problem: Find $ \min J(u,z) $ subject to the *semilinear, elliptic, and fractional PDE* $$\label{eq:state_equation} (-\Delta)^s u + a(\cdot,u) = z \textrm{ in } \Omega, \qquad u = 0 \textrm{ in } \Omega^c,$$ with ${{\Omega^c}}={{\mathbb{R}^n}}\setminus {{\Omega}}$, and the *control constraints* $ \mathfrak{a} \leq z(x) \leq \mathfrak{b} $ for a.e. $x \in \Omega$. The control bounds $\mathfrak{a},\mathfrak{b} \in \mathbb{R}$ are such that $\mathfrak{a} < \mathfrak{b}$. Assumptions on the nonlinear function $a$ will be deferred until section \[sec:assumption\]. We will refer to the previously defined PDE-constrained optimization problem as the *fractional semilinear optimal control problem*. For smooth functions $w: \mathbb{R}^n \rightarrow \mathbb{R}$, there are several equivalent definitions of the fractional Laplace operator $(-\Delta)^s$ in $\mathbb{R}^n$ [@MR3613319]. Indeed, $(-\Delta)^s$ can be naturally defined by means of the following pointwise formula: $$(-\Delta)^s w(x) = C(n,s) \, \mathrm{ p.v } \int_{\mathbb{R}^n} \frac{w(x) - w(y)}{|x-y|^{n+2s}} \mathrm{d}y, \qquad C(n,s) = \frac{2^{2s} s \Gamma(s+\frac{n}{2})}{\pi^{n/2}\Gamma(1-s)}, \label{eq:integral_definition}$$ where $\textrm{p.v}$ stands for the Cauchy principal value and $C(n,s)$ is a positive normalization constant that depends only on $n$ and $s$. Equivalently, $(-\Delta)^s$ can be defined via Fourier transform: $\mathcal{F}( (-\Delta)^s w) (\xi) = | \xi |^{2s} \mathcal{F}(w) (\xi)$. A proof of the equivalence of these two definitions can be found in [@Landkof section 1.1]. In addition to these two definitions, several other *equivalent definitions* of $(-\Delta)^s$ in $\mathbb{R}^n$ are available in the literature [@MR3613319]. Regarding *equivalence*, the scenario in bounded domains is *substantially different*. For functions supported in $\bar \Omega$, we may utilize the integral representation to define $(-\Delta)^s$. This gives rise to the so-called *restricted* or *integral* fractional laplacian. Notice that we have materialized a zero Dirichlet condition by restricting the operator to act only on functions that are zero outside $\Omega$. We must immediately mention that in bounded domains, and in addition to the *restricted* or *integral* fractional Laplacian there are, at least, two others *non-equivalent* definitions of nonlocal operators related to the fractional Laplacian: the *regional* fractional Laplacian and the *spectral* fractional Laplacian; see [@MR3393253 Section 2] and [@MR3503820 Section 6] for details. In this work, we adopt the restricted or integral definition of the fractional Laplace operator $(-\Delta)^s$, which, from now on, we shall simply refer to as the *integral fractional Laplacian*. During the very recent past, there has been considerable progress in the design and analysis of solution techniques for problems involving fractional diffusion. We refer the interested reader to [@MR3893441; @acta_marta] for a complete overview of the available results and limitations. In contrast to these advances, the numerical analysis of PDE-constrained optimization problems involving $(-\Delta)^s$ has been less explored. Restricting ourselves to problems considering the spectral definition, we mention [@MR3429730; @MR4015150; @MR3702421] within the linear–quadratic scenario, [@MR3850351] for optimization with respect to order, and [@MR3939497] for bilinear optimal control. We also mention [@semilinear], where the authors analyze, at the continuous level, a semilinear optimal control problem for the spectral fractional Laplacian. Concerning the integral fractional Laplacian, it seems that the results are even scarcer; the linear–quadratic case has been recently analyzed in [@MR3990191]. We conclude this paragraph by mentioning [@MR3596859; @MR3158780] for numerical approximations of optimal control problems involving suitable nonlocal operators. In addition to this exposition being the first one that studies numerical schemes for semilinear optimal control problems involving the *integral* fractional Laplacian, the analysis itself comes with its own set of difficulties. Overcoming them has required us to provide several results. Let us briefly detail some of them: 1. *Fractional PDEs:* Let $s \in (0,1)$, $n \geq 2$, $r>n/2s$, and $z \in L^r(\Omega)$. We show that is well-posed for $a = a(x,u)$ being a Carathéodory function, monotone increasing in $u$, satisfying and $a(\cdot,0) \in L^r(\Omega)$ (Theorem \[thm:stata\_equation\_well\_posedness\]). \[ii\] 2. *FEM discretizations:* We prove convergence of finite element discretizations on Lipschitz polytopes and obtain error estimates on smooth domains; the latter under additional assumptions on $a$ and the underlying forcing term that guarantee the regularity estimates of Theorem \[thm:regularity\_space\_state\_equation\]; see section \[sec:fem\]. \[iii\] 3. *Existence of an optimal control:* Assuming that, in addition, $L=L(x,u)$ is a Carathéodory function and $a$ and $L$ are locally Lipschitz in $u$, we show that our control problem admit at least a solution; see Theorem \[thm:existence\_control\]. \[iv\] 4. *Optimality conditions:* Let $n \in \{2,3\}$ and $s > n/4$. Under additional assumptions on $a$ and $L$, we derive second order necessary and sufficient optimality conditions with a minimal gap; see Section \[sec:2nd\_order\]. \[v\] 5. *Regularity estimates:* Let $n \geq 2$ and $s \in (0,1)$. We analyze regularity properties for the optimal variables. We prove that $\bar u,\bar p \in H^{s+1/2-\epsilon}(\Omega)$ and $\bar z \in H^{\gamma}(\Omega)$, where $\gamma = \min \{ 1, s + 1/2 - \epsilon \}$ and $\epsilon >0$ is arbitrarily small. \[vi\] 6. *Convergence of discretization and error estimates:* Let $n \geq 2$ and $s \in (0,1)$. We prove that the sequence $\{ \bar z_h \}_{h>0}$ of global solutions of suitable discrete control problems converge to a solution of the fractional semilinear optimal control problem. When $n \in \{2,3\}$ and $s>n/4$, we derive error estimates. Over the last 20 years, several contributions have delineated the numerical analysis of semilinear optimal control problems. Without a doubt, these studies have paved the way for the achievement of the aforementioned results. In particular, we have followed [@MR2583281], for the analysis of and the optimal control problem, [@MR3586845], for deriving second order optimality conditions, and [@MR2350349; @MR3023751; @MR3586845], for analyzing convergence properties and deriving error estimates. The rest of the paper is organized as follows. In section \[sec:state\_equation\], we analyze the fractional state equation . Section \[sec:optimal\_control\_problem\] is devoted to the study of the fractional semilinear optimal control problem. In section \[sec:fem\], we study finite element discretizations for . Section \[sec:fem\_control\] is advocated to the analysis of finite element discretizations for the fractional semilinear optimal control problem: convergence and error estimates. Notation and preliminaries {#sec:notation} ========================== Throughout this work $\Omega$ is an open and bounded domain of $\mathbb{R}^n$ $(n \geq 2)$ with Lipschitz boundary $\partial \Omega$; we will impose additional assumptions on $n$ and $\partial \Omega$ when needed. We will denote by $\Omega^c$ the complement of $\Omega$. If $\mathcal{X}$ and $\mathcal{Y}$ are normed spaces, we write $\mathcal{X} \hookrightarrow \mathcal{Y}$ to denote that $\mathcal{X}$ is continuously embedded in $\mathcal{Y}$. Let $\{ x_n \}_{n=1}^{\infty}$ be a sequence in $\mathcal{X}$. We will denote by $x_n \rightarrow x$ and $x_n \rightharpoonup x$ the strong and weak convergence, respectively, of $\{ x_n \}_{n=1}^{\infty}$ to $x$. The relation ${\sf a} \lesssim {\sf b}$ indicates that ${\sf a} \leq C {\sf b}$, with a positive constant $C$ that depends neither on ${\sf a}$, ${\sf b}$ nor on the discretization parameters but it might depend on $s$, $n$, and $\Omega$. The value of $C$ might change at each occurrence. Assumptions {#sec:assumption} ----------- We will operate under the following assumptions on $a$ and $L$. We must, however, immediately mention that some of the results obtained in this work are valid under less restrictive requirements; when possible we explicitly mention the assumptions on $a$ and $L$ that are needed to obtain a particular result. 1. \[A1\] $a:\Omega\times \mathbb{R}\rightarrow \mathbb{R}$ is a Carathéodory function of class $C^2$ with respect to the second variable such that $a(\cdot,0)\in L^{r}(\Omega)$ for $r>n/2s$. 2. \[A2\] $\frac{\partial a}{\partial u}(x,u)\geq 0$ for a.e. $x\in\Omega$ and for all $u \in \mathbb{R}$. 3. \[A3\] For all $\mathfrak{m}>0$, there exists a positive constant $C_{\mathfrak{m}}$ such that $$\sum_{i=1}^{2}\left|\frac{\partial^{i} a}{\partial u^{i} }(x,u)\right|\leq C_{\mathfrak{m}}, \qquad \left|\frac{\partial^{2} a}{\partial u^{2} }(x,u)- \frac{\partial^{2} a}{\partial u^{2} }(x,w)\right|\leq C_{\mathfrak{m}} |u-w|$$ for a.e. $x\in \Omega$ and $u,w \in [-\mathfrak{m},\mathfrak{m}]$. <!-- --> 1. \[B1\] $L: \Omega \times \mathbb{R} \rightarrow \mathbb{R}$ is a Carathéodory function of class $C^2$ with respect to the second variable such that $L(\cdot,0) \in L^1(\Omega)$. 2. \[B2\] For all $\mathfrak{m}>0$, there exist $\psi_{\mathfrak{m}}, \phi_{\mathfrak{m}} \in L^{r}(\Omega)$, with $r>n/2s$, such that $$\left| \frac{\partial L}{\partial u}(x,u) \right| \leq \psi_{\mathfrak{m}}(x) \quad \left| \frac{\partial^2 L}{\partial u^2}(x,u) \right| \leq \phi_{\mathfrak{m}}(x),$$ for a.e. $x\in \Omega$ and $u \in [-\mathfrak{m},\mathfrak{m}]$. The following assumptions are particularly needed to derive regularity estimates: 1. \[C1\] $a(\cdot,0) \in L^2(\Omega) \cap H^{\frac{1}{2} - s - \epsilon}(\Omega)$, 2. \[C2\] $\frac{\partial L}{\partial u}(\cdot,0) \in L^2(\Omega) \cap H^{\frac{1}{2} - s - \epsilon}(\Omega)$. In these assumptions $\epsilon$ denotes an arbitrarily small positive constant. Function spaces --------------- For any $s \geq 0$, we define $H^s(\mathbb{R}^n)$, the Sobolev space of order $s$ over $\mathbb{R}^n$, by [@Tartar Definition 15.7] $$H^s(\mathbb{R}^n) := \left \{ v \in L^2(\mathbb{R}^n): (1+|\xi|^2)^{s/2} \mathcal{F}(v) \in L^2(\mathbb{R}^n)\right \}.$$ With the space $H^s(\mathbb{R}^n)$ at hand, we define $\tilde H^s(\Omega)$ as the closure of $C_0^{\infty}(\Omega)$ in $H^s(\mathbb{R}^n)$. This space can be equivalently characterized by [@McLean Theorem 3.29] $$\tilde H^s(\Omega) = \{v|_{\Omega}: v \in H^s(\mathbb{R}^n), \textrm{ supp } v \subset \overline\Omega\}.$$ When $\partial \Omega$ is Lipschitz $\tilde H^s(\Omega)$ is equivalent to $\mathbb{H}^s(\Omega)=[L^2(\Omega),H_0^1(\Omega)]_s$, the real interpolation between $L^2(\Omega)$ and $H_0^1(\Omega)$ for $s \in (0,1)$ and to $H^s(\Omega) \cap H_0^1(\Omega)$ for $s \in (1,3/2)$ [@McLean Theorem 3.33]. We denote by $H^{-s}(\Omega)$ the dual space of $\tilde H^s(\Omega)$ and by $\langle \cdot, \cdot \rangle$ the duality pair between these two spaces. We define the bilinear form $$\label{eq:bilinear_form} \mathcal{A}(v,w) = \frac{C(n,s)}{2} \iint_{\mathbb{R}^n \times \mathbb{R}^n} \frac{( v(x) - v(y) ) (w(x)-w(y))}{|x-y|^{n+2s}} \mathrm{d}x \mathrm{d}y,$$ and denote by $\| \cdot \|_s$ the norm that $ \mathcal{A}(\cdot,\cdot)$ induces, which is just a multiple of the $H^s(\mathbb{R}^n)$-seminorm: $ \| v \|_s = \sqrt{ \mathcal{A}(v,v)} = \mathfrak{C}(n,s) |v|_{H^s(\mathbb{R}^n)}, $ where $\mathfrak{C}(n,s) = \sqrt{C(n,s)/2}$. We will repeatedly use the following continuous embedding: $H^s(\Omega) \hookrightarrow L^{\mathfrak{q}}(\Omega)$ for $1 \leq \mathfrak{q} \leq 2n/(n-2s)$ [@MR2424078 Theorem 7.34]; observe that $n>2s$. If $\mathfrak{q} < 2n/(n-2s)$ the embedding $H^s(\Omega) \hookrightarrow L^{\mathfrak{q}}(\Omega)$ is compact [@MR2424078 Theorem 6.3]. The state equation {#sec:state_equation} ================== Let $f \in H^{-s}(\Omega)$ be a forcing term. In this section, we analyze the following fractional semilinear PDE: Find $u \in \tilde H^s(\Omega)$ such that $$\label{eq:weak_semilinear_pde} \mathcal{A}(u,v) + \langle a(\cdot,u),v \rangle = \langle f , v \rangle \quad \forall v \in \tilde H^{s}(\Omega).$$ Here, $a = a(x,u) : \Omega \times \mathbb{R} \rightarrow \mathbb{R}$ denotes a Carathéodory function that is monotone increasing in $u$. In addition, we assume that, for $\mathfrak{m}>0$, there exits $$\varphi_{\mathfrak{m}} \in L^{\mathfrak{t}}(\Omega): \quad |a(x,u)| \leq | \varphi_{\mathfrak{m}}(x)| ~\textrm{a.e.}~x \in \Omega,~u \in [-\mathfrak{m},\mathfrak{m}], \quad \mathfrak{t} =2n/(n+2s). \label{eq:assumption_on_phi_state_equation}$$ We present the following existence and uniqueness result. Let $n \geq 2$, $s \in (0,1)$, and $r>n/2s$. Let $\Omega \subset \mathbb{R}^n$ be an open and bounded domain with Lipschitz boundary. If $f \in L^{r}(\Omega)$, $a$ satisfies , and $a(\cdot,0) \in L^r(\Omega)$, then problem has a unique solution $u \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ which satisfies the stability estimate $$| u |_{H^s(\mathbb{R}^n)} + \| u \|_{L^{\infty}(\Omega)} \lesssim \| f - a(\cdot,0) \|_{L^{r}(\Omega)}, \label{eq:stability}$$ with a hidden constant that is independent of $u$, $a$, and $f$. \[thm:stata\_equation\_well\_posedness\] We proceed in four steps: Let us assume, for the moment, that, in addition, $a = a(x,u)$ is globally bounded, i.e., $|a(x,u)| \leq \mathfrak{c}$ for a.e. $x \in \Omega$ and all $u \in \mathbb{R}$, and $a(\cdot,0) = 0$. Define the mapping $$\mathfrak{A}: \tilde H^s(\Omega) \rightarrow H^{-s}(\Omega): \quad \langle \mathfrak{A} u,v \rangle = \mathcal{A}(u,v) + \langle a(\cdot,u),v \rangle \quad \forall v \in \tilde H^{s}(\Omega).$$ Since $\mathcal{A}$ is bilinear, continuous, and coercive on $\tilde H^s(\Omega) \times \tilde H^s(\Omega)$ and $a = a(x,u)$ is globally bounded and monotone increasing in $u$, it is immediate that $\mathfrak{A}$ is *well-posed*, *monotone*, and *coercive*. In addition, since $a = a(x,u)$ is continuous in $u$ for a.e. $x \in \Omega$, dominated convergence yields the *hemicontinuity* of $\mathfrak{A}$. Existence and uniqueness of $u \in \tilde H^s(\Omega)$ follows from the main theorem on monotone operators [@MR1033498 Theorem 26.A], [@MR3014456 Theorem 2.18]. Set $v = u$ to obtain $ | u |_{H^s(\mathbb{R}^n)} \lesssim \| f \|_{H^{-s}(\Omega)}$. Define, for $k >0$, $v_k$ by $ v_k(x) = u(x) - k \textrm{ if } u(x) \geq k, $ $ v_k(x) = 0 \textrm{ if } |u(x)| < k, $ and $ v_k(x) = u(x) + k \textrm{ if } u(x) \leq- k. $ We also define the set $$\Omega(k):= \{ x \in \Omega: |u(x)| \geq k\}.$$ Since $a = a(x,u)$ is monotone increasing in $u$ and $a(\cdot,0) = 0$, we have $\langle a(\cdot,u), v_k \rangle = \int_{\Omega} a(x,u(x)) v_k(x) \mathrm{d}x \geq 0$. This yields $\mathcal{A}(u,v_k) \leq \langle f, v_k \rangle$. The relations and inequalities (2.22)–(2.30) in [@MR3745164] reveal that $\mathcal{A}(v_k,v_k) \leq \mathcal{A}(u,v_k)$. We can thus obtain $ \| v_k \|^2_{s} = \mathcal{A}(v_k,v_k) \leq \langle f, v_k \rangle$. Define $\mathfrak{q}:=2n/(n-2s)$. Thus, for $t<2n/(n-2s)$, $$\| v_k \|_{L^{\mathfrak{q}}(\Omega)}^2 \lesssim | v_k |^2_{H^s(\mathbb{R}^n)} \lesssim \| v_k \|_{L^{\mathfrak{q}}(\Omega)} | \Omega(k)|^{\frac{1}{t}} \|f \|_{L^{r}(\Omega)}, \quad \mathfrak{q}^{-1} + r^{-1} + t^{-1} = 1.$$ On the other hand, $\| v_k \|_{L^{\mathfrak{q}}(\Omega)}^{\mathfrak{q}} = \int_{\Omega(k)} |v_k(x)|^{\mathfrak{q}} \mathrm{d}x = \int_{\Omega(k)} | |u(x)|-k |^{\mathfrak{q}} \mathrm{d}x$. Let $h>k$, then $\Omega(h) \subset \Omega(k)$ and $ \int_{\Omega(k)} | |u(x)|-k |^{\mathfrak{q}} \mathrm{d}x \geq (h-k)^{\mathfrak{q}} |\Omega(h)|$. Consequently, $$(h-k) |\Omega(h)|^{\frac{1}{\mathfrak{q}}} \leq \| v_k \|_{L^{\mathfrak{q}}(\Omega)} \lesssim ( | \Omega(k)|^{\frac{1}{\mathfrak{q}}})^{\frac{\mathfrak{q}}{t}} \|f \|_{L^{r}(\Omega)}.$$ Since $r>n/2s$, $\mathfrak{q}/t > 1$. An application of [@MR1786735 Lemma B.1] yields the existence of $\mathfrak{h} > 0$ such that $|\Omega(\mathfrak{h}) | = 0$, which implies that $u \in L^{\infty}(\Omega)$ and $\| u \|_{L^{\infty}(\Omega)} \lesssim \| f\|_{L^r(\Omega)}$. We remove the boundedness assumption on $a$. Define, for $k >0$, $a_k$ by $ a_k(x,u) = a(x,k) \textrm{ if } u > k, $ $ a_k(x,u) = a(x,u) \textrm{ if } |u| \leq k, $ and $ a_k(x,u) = a(x,-k) \textrm{ if } u < - k. $ Since $a_k$ is bounded in the second variable and satisfies , there exists a unique solution $u$ to problem with $a$ replaced by $a_k$. In addition, $| u |_{H^s(\mathbb{R}^n)} + \| u \|_{L^{\infty}(\Omega)} \leq c_{\infty} \| f \|_{L^{r}(\Omega)}$ with $c_{\infty}$ being independent of $a_k$ and $k$. Choose $k>c_{\infty} \| f \|_{L^{r}(\Omega)}$ so that $a_k(x,u(x))= a(x,u(x))$ for a.e. $x \in \Omega$. Consequently, $u$ solves . Uniqueness of solutions follows from the monotonicity properties of $a$. We remove the condition $a(\cdot,0) = 0$ by replacing $a(\cdot,u)$ by $a(\cdot,u) - a(\cdot,0)$. The optimal control problem {#sec:optimal_control_problem} =========================== In this section, we analyze the following weak version of the fractional semilinear optimal control problem: Find $$\label{eq:min} \min \{ J(u,z): (u,z) \in \tilde H^s(\Omega) \times \mathbb{Z}_{ad}\}$$ subject to the *fractional, semilinear, and elliptic state equation* $$\label{eq:weak_st_eq} \mathcal{A}( u, v)+(a(\cdot,u),v)_{L^2(\Omega)}=(z,v)_{L^2(\Omega)} \quad \forall v \in \tilde H^s(\Omega).$$ Here, $ \mathbb{Z}_{ad}:=\{ v \in L^2(\Omega): \mathfrak{a} \leq v(x) \leq \mathfrak{b}~\text{a.e.}~x \in \Omega \}$. Let $r>n/2s$ and $a = a(x,u): \Omega \times \mathbb{R} \rightarrow \mathbb{R}$ be a monotone increasing in $u$ Carathéodory function satisfying and $a(\cdot,0) \in L^r(\Omega)$. In view of Theorem \[thm:stata\_equation\_well\_posedness\], the existence of a unique solution $u$ to problem is guaranteed. We thus introduce the control to state map $\mathcal{S}: L^{r}(\Omega) \rightarrow \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ which, given a control $z$, associates to it the unique state $u$ that solves . With $\mathcal{S}$ at hand, we also introduce the reduced cost functional $j:\mathbb{Z}_{ad} \rightarrow \mathbb{R}$ by the relation $j(z)=J(\mathcal{S}z,z)$. Existence of optimal controls ----------------------------- Since the PDE-constrained optimization problem – is not convex, we analyze existence results and optimality conditions in the context of local solutions. We begin our studies with an existence result. \[thm:existence\_control\] Let $n \geq 2$, $s \in (0,1)$ and $r>n/2s$. Let $a = a(x,u) : \Omega \times \mathbb{R} \rightarrow \mathbb{R}$ be a Carathéodory function that is monotone increasing in $u$ with $a(\cdot,0) \in L^r(\Omega)$. Let $L = L(x,u) : \Omega \times \mathbb{R} \rightarrow \mathbb{R}$ be a Carathéodory function with $L(\cdot,0) \in L^1(\Omega)$. Assume that $a$ and $L$ are locally Lipschitz with respect to the second variable. Thus, – admits at least one solution $(\bar{u},\bar{z}) \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega) \times \mathbb{Z}_{ad}$. Let $\{ (u_k,z_k) \}_{k=1}^{\infty}$ be a minimizing sequence, i.e., for $k \in \mathbb{N}$, $z_k \in \mathbb{Z}_{ad}$ and $u_k = \mathcal{S}z_k$ are such that $J(u_k,z_k) \rightarrow \mathfrak{j}:= \inf J(u,z)$ as $k \uparrow \infty$. Since $\mathbb{Z}_{ad}$ is bounded in $L^{\infty}(\Omega)$, there exits a nonrelabeled subsequence $\{ z_k \}_{k=1}^{\infty}$ such that $z_k \mathrel{\ensurestackMath{\stackon[1pt]{\rightharpoonup}{\scriptstyle\ast}}} \bar{z}$ in $L^{\infty}(\Omega)$ as $k \uparrow \infty$. Observe that, since $z_k \in \mathbb{Z}_{ad}$ for every $k \in \mathbb{N}$, there exists $\mathfrak{m}>0$ such that $|u_k(x)| \leq \mathfrak{m}$ for $k \in \mathbb{N}$ and a.e. $x \in \Omega$. This implies that $\{a(\cdot,u_k)\}_{k=1}^{\infty}$ is bounded in $L^r(\Omega)$. We can thus conclude the existence of a nonrelabeled subsequence $\{ u_k \}_{k=1}^{\infty}$ such that $u_k \rightharpoonup \bar{u}$ in $\tilde H^s(\Omega)$ and $u_k \rightarrow \bar{u}$ in $L^2(\Omega)$ as $k \uparrow \infty$. For $k \in \mathbb{N}$, $u_k \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ solves $$\mathcal{A}(u_k,v) + \langle a(\cdot,u_k),v \rangle = \langle z_k ,v \rangle \quad \forall v \in \tilde H^s(\Omega). \label{eq:u_n}$$ Since $\mathfrak{M}:=\{ v \in L^{\infty}(\Omega): |v(x)|\leq M ~\mathrm{a.e}.~x \in \Omega\}$ is weakly sequentially closed, we conclude that $\bar u \in \mathfrak{M}$. We can thus invoke the local Lipschitz property of $a$ in $u$ to obtain $\| a(\cdot,\bar u) - a(\cdot,u_k) \|_{L^2(\Omega)} \leq \mathfrak{l}_M \| \bar u - u_k \|_{L^2(\Omega)} \rightarrow 0$ as $k \uparrow \infty$. In view of the previous convergence results, passing to the limit in yields $\bar u = \mathcal{S} \bar z$. On the other hand, the map $L^2(\Omega) \ni v \mapsto \| v\|_{L^2(\Omega)}^2 \in \mathbb{R}$ is continuous and convex; it is thus weakly lower continuous. Consequently, $$\mathfrak{j} = \lim_{n \uparrow \infty} J(u_n,z_n) = \int_{\Omega} L(x,\bar u(x) ) \mathrm{d}x + \liminf_{n \uparrow \infty} \frac{\alpha}{2} \| z_n \|^2_{L^2(\Omega)} \geq J(\bar u, \bar z).$$ Notice that $\left| \int_{\Omega} [ L(x,\bar u(x)) - L(x,u_n(x)) ] \mathrm{d}x \right| \leq \mathfrak{l}_{\mathfrak{m}} \| \bar u - u_n \|_{L^1(\Omega)} \rightarrow 0$ as $n \uparrow 0$ because of the local Lipschitz property of $L$ in the second variable and $ u_n \rightarrow \bar u$ in $L^2(\Omega)$. To obtain the result of Theorem \[thm:existence\_control\] we have assumed that $a$ is locally Lipschitz in $u$ and $a(\cdot,0) \in L^r(\Omega)$ with $r>n/2s$. Observe that condition can thus be guaranteed because $n/2s>2n/(n+2s)$. First order necessary optimality conditions {#sec:1st_order} ------------------------------------------- In this section, we analyze differentiability properties for the control to state map $\mathcal{S}$ and derive first order necessary optimality conditions. We begin our studies by precisely introducing the concept of local minimum. Let $q \in [1,\infty)$ and $\epsilon >0$, we define the closed ball in $L^q(\Omega)$ of radius $\epsilon$ and centered at $\bar z$, $$B_{\epsilon}(\bar z) := \{ z \in L^q(\Omega): \| \bar z - z \|_{L^q(\Omega)} \leq \epsilon \}.$$ Let $q \in [1,\infty)$. We say that $\bar z \in \mathbb{Z}_{ad}$ is a local minimum, or locally optimal, in $L^{q}(\Omega)$ for – if there exists $\epsilon >0$ such that $j(\bar z) \leq j(z)$ for every $z \in B_{\epsilon}(\bar z) \cap \mathbb{Z}_{ad}$. Since $\mathbb{Z}_{ad}$ is bounded in $L^{\infty}(\Omega)$, it can be proved that $\bar z \in \mathbb{Z}_{ad}$ is locally optimal in $L^q(\Omega)$ if and only if $\bar z \in \mathbb{Z}_{ad}$ is a local minimum in $L^2(\Omega)$; see [@MR3586845 Section 5] for details. \[rk:local\_minimum\] In what follows we will operate in $L^2(\Omega)$ regarding local optimally. \[thm:properties\_C\_to\_S\] Let $n \geq 2$, $s \in (0,1)$, and $r > n/2s$. Assume that [\[A1\]]{.nodecor}, [\[A2\]]{.nodecor}, and [\[A3\]]{.nodecor} hold. Then, the control to state map $\mathcal{S}: L^{r}(\Omega) \rightarrow \tilde H^{s}(\Omega)\cap L^{\infty}(\Omega)$ is of class $C^2$. In addition, if $z,w \in L^{r}(\Omega)$, then $\phi = \mathcal{S}'(z) w \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ corresponds to the unique solution to the problem $$\label{eq:aux_adjoint} \mathcal{A}(\phi,v)+\left(\tfrac{\partial a}{\partial u}(\cdot,u)\phi, v \right)_{L^2(\Omega)}=(w,v)_{L^2(\Omega)} \quad \forall v \in \tilde H^s(\Omega),$$ where $u = \mathcal{S}z$. In addition, for every $w_1,w_2\in L^{r}(\Omega)$, $\psi=\mathcal{S}''(z)(w_1,w_2)\in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ corresponds to the unique solution to $$\label{eq:aux_adjoint_diff_2} \mathcal{A}(\psi, v)+\left(\tfrac{\partial a}{\partial u}(\cdot,u)\psi,v\right)_{L^2(\Omega)}=-\left(\tfrac{\partial^2 a}{\partial u^2}(\cdot,u)\phi_{w_1}\phi_{w_2}, v \right)_{L^2(\Omega)} \quad \forall v \in \tilde H^s(\Omega),$$ where $u = \mathcal{S}z$ and $\phi_{w_i}=\mathcal{S}'(z)w_i$, with $i=1,2$. The first order Fréchet differentiability of $\mathcal{S}$ from $L^{r}(\Omega)$ into $\tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ follows from a slight modification of the proof of [@MR2583281 Theorem 4.17] that basically entails to replace $H^1(\Omega)$ by $\tilde H^s(\Omega)$ and $C(\bar \Omega)$ by $L^{\infty}(\Omega)$. These arguments also show that $\phi=\mathcal{S}'(z)w\in \tilde H^{s}(\Omega)\cap L^{\infty}(\Omega)$ corresponds to the unique solution to ; since $w \in L^r(\Omega)$ and $\frac{\partial a}{ \partial u} (x,u) \geq 0$ for a.e. $x\in \Omega$ and all $u \in \mathbb{R}$, problem is well-posed. The second order Fréchet differentiability of $\mathcal{S}$ can be obtained by using the implicit function theorem. Let us introduce the linear mapping $\mathfrak{f} \mapsto \mathfrak{u}$ by $$\mathfrak{R}: L^r(\Omega) \rightarrow \tilde H^s(\Omega) \cap L^{\infty}(\Omega): \quad \mathcal{A}(\mathfrak{u},v) = \langle \mathfrak{f},v\rangle \quad \forall v \in \tilde H^s(\Omega).$$ Define $\mathfrak{F}: [\tilde H^s(\Omega) \cap L^{\infty}(\Omega)] \times L^r(\Omega) \rightarrow L^{\infty}(\Omega)$ by $\mathfrak{F}(u,z) := u - \mathfrak{R}(z - a(\cdot,u))$. We first observe that $\mathfrak{F}$ is of class $C^2$. Second, $\mathfrak{F}(\mathcal{S}z,z) = 0$. Third, since $\partial \mathfrak{F} / \partial u (u,z) v = v + \mathfrak{R} \partial a/ \partial u (\cdot,u)v$, it can be deduced that $\partial \mathfrak{F} / \partial u (u,z)$ is surjective from $L^{\infty}(\Omega)$ into $\tilde H^s(\Omega) \cap L^{\infty}(\Omega)$. The implicit function theorem thus implies that $\mathcal{S}$ is of class $C^2$. The fact that $\psi$ solves follows from differentiating $\mathfrak{F}(\mathcal{S}z,z) = 0$. The following result is standard: If $\bar{z}\in \mathbb{Z}_{ad}$ denotes a locally optimal control for problem –, then $ j'(\bar z) (z - \bar z) \geq 0 $ for all $z \in \mathbb{Z}_{ad}$ [@MR2583281 Lemma 4.18]. To explore this inequality, we define the adjoint state $p \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ as the solution to $$\label{eq:adj_eq} \mathcal{A}(v,p) + \left(\tfrac{\partial a}{\partial u}(\cdot,u)p,v\right)_{L^2(\Omega)} = \left(\tfrac{\partial L}{\partial u}(\cdot,u),v\right)_{L^2(\Omega)} \quad \forall v \in \tilde H^s(\Omega).$$ Assumption [\[A1\]]{.nodecor} guarantees that $\partial a/\partial u (x,u) \geq 0$ for a.e. $x \in \Omega$ and all $u \in \mathbb{R}$. Assumption [\[B2\]]{.nodecor} yield $\partial L/ \partial u(\cdot,u) \in L^r(\Omega)$ for $r>n/2s$. The existence and uniqueness of $p$ in $\tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ is thus immediate. We present first order necessary optimality conditions for –. \[thm:optimality\_cond\] Let $n \geq 2$, $s \in (0,1)$, and $r>n/2s$. Assume that [\[A1\]]{.nodecor}–[\[A3\]]{.nodecor} and [\[B1\]]{.nodecor}–[\[B2\]]{.nodecor} hold. Then every locally optimal control $\bar z \in \mathbb{Z}_{ad}$ satisfies the variational inequality $$\label{eq:var_ineq} \left( \bar p +\alpha \bar{z},z-\bar{z}\right)_{L^2(\Omega)}\geq 0 \quad \forall z\in \mathbb{Z}_{ad},$$ where $\bar p \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ denotes the solution to with $u$ replaced by $\bar{u} = \mathcal{S} \bar z$. Define $\ell: L^{\infty}(\Omega) \rightarrow \mathbb{R}$ by $\ell(u) = \int_{\Omega} L(x,u(x)) \mathrm{d}x$ and observe that [\[B1\]]{.nodecor}–[\[B2\]]{.nodecor} yield the Fréchet differentiability of $\ell$ on bounded sets of $L^{\infty}(\Omega)$. Since $\mathcal{S}: L^r(\Omega) \rightarrow \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ is differentiable, we thus deduce the Fréchet differentiability of $\ell$ as a map from $L^{\sigma}(\Omega)$ to $\mathbb{R}$, where $\sigma = \max \{ n/2s, 2 \}$, upon noticing that $L^2(\Omega) \ni z \mapsto \| z \|^2_{L^2(\Omega)} \in \mathbb{R}$ is also differentiable. Basic computations thus reveal $$j'(\bar z) h = \int_{\Omega} \left( \frac{\partial L}{\partial u} (x, \mathcal{S} \bar z(x) ) \mathcal{S}'(\bar z)h(x) + \alpha \bar z(x) h(x) \right) \mathrm{d}x, \quad h \in L^{\sigma}(\Omega). \label{eq:aux_variational_inequality}$$ Set $h = z - \bar z \in \mathbb{Z}_{ad}$ and define $\chi= \mathcal{S}'(\bar z)h$. Setting $v = \chi$ in problem and $v = \bar p$ in the problem that $\chi$ solves allow us to obtain $(z - \bar z, \bar p)_{L^2(\Omega)} = (\tfrac{ \partial L}{\partial u}(\cdot,\bar u),\chi)_{L^2(\Omega)}$. Replace this identity into to obtain . This concludes the proof. Define $\Pi_{[\mathfrak{a},\mathfrak{b}]} : L^1(\Omega) \rightarrow \mathbb{Z}_{ad}$ by $\Pi_{[\mathfrak{a},\mathfrak{b}]}(v) := \min\{ \mathfrak{b}, \max\{ v, \mathfrak{a} \} \}$ a.e in $\Omega$. The local optimal control $\bar{z}$ satisfies if and only if [@MR2583281 Section 4.6] $$\label{eq:projection_control} \bar{z}(x):=\Pi_{[\mathfrak{a},\mathfrak{b}]}(-\alpha^{-1}\bar{p}(x)) \textrm{ a.e.}~x \in \Omega.$$ Since $\bar p \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ and $s \in (0,1)$, it is immediate that $\bar z \in H^{s}(\Omega) \cap L^{\infty}(\Omega)$; further regularity properties for $\bar z$ are obtained in Theorem \[thm:regularity\_space\] below. Second order sufficient optimality condition {#sec:2nd_order} -------------------------------------------- In Theorem \[thm:optimality\_cond\] we obtained a first order necessary condition for local optimality. Since our optimal control problem is not convex, sufficiency requires the use of second order optimality conditions. The purpose of this section is thus to derive such conditions. To accomplish this task, we begin by introducing some preliminary concepts. Let $\bar{z} \in \mathbb{Z}_{ad}$ satisfies . Define $\bar{\mathfrak{p}} := \bar p + \alpha \bar z$. Observe that immediately yields $$\label{eq:derivative_j} \bar{\mathfrak{p}}(x) \begin{cases} = 0 & \text{ a.e.}~x \in \Omega \text{ if } \texttt{a}< \bar{z} < \texttt{b}, \\ \geq 0 & \text{ a.e.}~x \in \Omega \text{ if }\bar{z}=\texttt{a}, \\ \leq 0 & \text{ a.e.}~x \in \Omega \text{ if } \bar{z}=\texttt{b}. \end{cases}$$ Define the *cone of critial directions* $ C_{\bar{z}}:=\{v\in L^2(\Omega): \eqref{eq:sign_cond} \text{ holds and } \bar{\mathfrak{p}}(x) \neq 0 \implies v(x) = 0 \}, $ where condition reads as follows: $$\label{eq:sign_cond} v(x) \geq 0 \text{ a.e.}~x\in\Omega \text{ if } \bar{z}(x)=\mathfrak{a}, \qquad v(x) \leq 0 \text{ a.e.}~x\in\Omega \text{ if } \bar{z}(x)=\mathfrak{b}.$$ The following result is instrumental. \[pro:diff\_properties\_j\] Let $n \geq 2$, $s \in (0,1)$, $r>n/2s$, and $\sigma = \max \{ 2, n/2s \}$. Assume that [\[A1\]]{.nodecor}–[\[A3\]]{.nodecor} and [\[B1\]]{.nodecor}–[\[B2\]]{.nodecor} hold. Then the reduced cost $j: L^{\sigma}(\Omega) \rightarrow \mathbb{R}$ is of class $C^2$. In addition, for $z,w_1,w_2 \in L^{\sigma}(\Omega)$, we have $$\label{eq:charac_j2} j''(z)(w_1,w_2) = \int_{\Omega} \left( \frac{\partial^2 L}{\partial u^2}(x,u) \phi_{w_1}\phi_{w_2} + \alpha w_1 w_2 - p \frac{\partial^2 a}{\partial u^2}(x,u)\phi_{w_1}\phi_{w_2} \right) \mathrm{d}x,$$ where $u = \mathcal{S}z$, $p$ solves and $\phi_{w_i}=\mathcal{S}'(z)w_i$, with $i \in \{1,2 \}$. The first order Fréchet differentiability of $j$ has been obtained in Theorem \[thm:optimality\_cond\]. Theorem \[thm:properties\_C\_to\_S\] guarantees that $\mathcal{S}$ is second order Fréchet differentiable as a map from $L^{r}(\Omega)$ into $\tilde H^s(\Omega) \cap L^{\infty}(\Omega)$. In view of [\[B1\]]{.nodecor}–[\[B2\]]{.nodecor}, the map $u \mapsto \ell(u) := \int_{\Omega}L(x,u(x)) \mathrm{d}x$ is second order Fréchet differentiable as well as a map from the set $\{ u \in L^{\infty}(\Omega): |u(x)| \leq \mathfrak{m} \}$ to $\mathbb{R}$. The chain rule allows us to conclude that $j \in C^2$. The identity follows from the arguments elaborated in [@MR2583281 Section 4.10]. We now formulate second order necessary optimality conditions. \[thm:nec\_opt\_cond\] Let $n \in \{ 2,3 \}$ and $s > n/4$. Let $\bar{z} \in \mathbb{Z}_{ad}$ be a locally optimal control for problem –. Thus, $$\label{eq:second_order_nec} j''(\bar{z})v \geq 0 \quad \forall v\in C_{\bar{z}},$$ where $C_{\bar{z}}:=\{v\in L^2(\Omega): \eqref{eq:sign_cond} \text{ holds and } \bar{\mathfrak{p}}(x) \neq 0 \implies v(x) = 0 \}$. Let $v\in C_{\bar{z}}$. Define, for every $k \in \mathbb{N}$ and for a.e. $x\in \Omega$, the function $$v_{k}(x):= \begin{cases} \qquad \quad 0 \quad &\text{ if }\quad x: \mathfrak{a} < \bar{z}(x) < \mathfrak{a}+ \tfrac{1}{k}, \quad \mathfrak{b}-\tfrac{1}{k} < \bar{z}(x) < \mathfrak{b}, \\ \Pi_{[-k,k]}(v(x)) &\text{ otherwise}. \end{cases}$$ Since $v \in C_{\bar z}$, we have that $v_k \in C_{\bar z}$. In addition, $|v_{k}(x)|\leq |v(x)|$ and $v_{k}(x)\rightarrow v(x)$ as $k \uparrow \infty$ for a.e. $x \in \Omega$; therefore $v_k \rightarrow v$ in $L^2(\Omega)$. Now, since $\bar z + \rho v_k \in \mathbb{Z}_{ad}$, for $\rho \in (0,k^{-2}]$, and $\bar z$ is locally optimal for $j$ we deduce, for $\rho$ sufficiently small, $$0 \leq \tfrac{1}{\rho}[ j(\bar z + \rho v_k) - j(\bar z) ] = j'(\bar z) v_k + \tfrac{\rho}{2} j''(\bar z + \theta_k \rho v_k) v_k^2, \quad \theta_k \in (0,1). \label{eq:aux_second_order}$$ Observe that and $v_k \in C_{\bar z}$ yield $j'(\bar z) v_k = \int_{\Omega} \mathfrak{p}(x) v_k(x) \mathrm{d}x = 0$. Thus, diving by $\rho$ in , utilizing the characterization , and letting $\rho \downarrow 0$ yield $j''(\bar z) v_k^2 \geq 0$. Let $k \uparrow \infty$ and invoke , again, and $\| v_k - v\|_{L^2(\Omega)} \rightarrow 0$ to conclude. We now provide a sufficient second order optimality condition with a minimal gap with respect to the the necessary one proved in Theorem \[thm:nec\_opt\_cond\]. \[thm:suff\_opt\_cond\] Let $n \in \{ 2,3 \}$ and $s > n/4$. Let $\bar{z} \in \mathbb{Z}_{ad}$ be a locally optimal control for problem – satisfying $$\label{eq:second_order_suff} j''(\bar{z})v > 0 \quad \forall v\in C_{\bar{z}} \setminus \{ 0 \}.$$ Then, there exists $\kappa > 0$ and $\mu >0$ such that $$\label{eq:quadratic_growing} j( z) \geq j(\bar z) + \tfrac{\kappa}{2} \| z - \bar z\|_{L^2(\Omega)}^2$$ for all $z \in \mathbb{Z}_{ad}$ such that $\| \bar z - z \|_{L^2(\Omega)} \leq \mu$. We proceed by contradiction. Assume that for every $k \in \mathbb{N}$ there is an element $z_k \in \mathbb{Z}_{ad}$ such that $$\label{eq:contradiction_argument} \| \bar z - z _{k} \|_{L^2(\Omega)} < \tfrac{1}{k}, \quad j( z_k) < j(\bar z) + \tfrac{1}{2k} \| \bar z - z_k\|_{L^2(\Omega)}^2.$$ Define $\rho_k:= \| z _{k} - \bar z \|_{L^2(\Omega)}$ and $v_k:= \rho_k^{-1} ( z_k - \bar z)$. Notice that there exists a nonrelabeled subsequence $\{ v_k \}_{k=1}^{\infty} \subset L^2(\Omega)$ such that $v_k \rightharpoonup v$ in $L^2(\Omega)$ as $k \uparrow \infty$. We now proceed in three steps: We prove that $v \in C_{\bar z}$. Since the set of elements satisfying is closed and convex in $L^2(\Omega)$ and, for every $k \in \mathbb{N}$, $v_k$ belongs to this set, we deduce that $v$ satisfies . It suffices to prove that $\mathfrak{p}(x) \neq 0$ implies $v(x) = 0$. In view of , we deduce that $\int_{\Omega} \mathfrak{p}(x) v(x) \mathrm{d}x \geq 0$ because $ \int_{\Omega} \mathfrak{p}(x) v_k(x) \mathrm{d}x = \rho_k^{-1} \int_{\Omega} \mathfrak{p}(x)(z_k(x) - \bar z(x) )\mathrm{d}x \geq 0$. On the other hand, observe that and the mean value theorem reveal $$j(z_k) - j(\bar z) = j'(\bar z + \theta_k( z_k - \bar z )) (z_k - \bar z) < \tfrac{1}{2k} \| \bar z - z_k\|_{L^2(\Omega)}^2 = \tfrac{\rho_k^2}{2k}, \quad \theta_k \in (0,1).$$ Divide by $\rho_k$ and let $k \uparrow \infty$ to arrive at $ j'(\bar z + \theta_k( z_k - \bar z )) v_k < (2k)^{-1}\rho_k \rightarrow 0$ as $k \uparrow \infty$. Define $\hat z_k := \bar z + \theta_k( z_k - \bar z )$. Since $s>n/4$ and $\hat z_k \rightarrow \bar z$ in $L^2(\Omega)$, as $k \uparrow \infty$, we have $$\hat u_k:= \mathcal{S}(\hat z_k) \rightarrow \mathcal{S}(\bar z) = \bar u \textrm{ in } \tilde H^s(\Omega) \cap L^{\infty}(\Omega), \quad \tfrac{\partial L}{\partial u} (\cdot ,\hat u_k) \rightarrow \tfrac{\partial L}{\partial u} (\cdot, \bar u) \textrm{ in } L^{r}(\Omega),$$ upon invoking [\[B2\]]{.nodecor}. Consequently, $\hat p_k \rightarrow \bar p$ in $H^s(\Omega) \cap L^{\infty}(\Omega)$ as $k \uparrow \infty$. Here, $\hat p_k$ denotes the solution to with $z$ replaced by $\hat z_k$ and $u$ replaced by $\hat u_k$. Thus, $$\int_{\Omega} \bar{\mathfrak{p}}(x) v(x) \mathrm{d}x = \lim_{k \uparrow \infty} \int_{\Omega} \left[ \hat p_k (x) + \alpha \hat{z_k}(x) \right]v_k(x)\mathrm{d}x = \lim_{k \uparrow \infty} j'( \bar z + \theta_k( z_k - \bar z ) )v_k \leq 0.$$ We have thus deduced that $\int_{\Omega}\mathfrak{p}(x) v(x) \mathrm{d}x = \int_{\Omega} |\mathfrak{p}(x) v(x)| \mathrm{d}x = 0$. Consequently, $\mathfrak{p}(x) \neq 0$ implies $v(x) = 0$ for a.e. $x\in \Omega$. This proves that $v \in C_{\bar z}$. We prove that $v=0$. We begin with an application of Taylor’s theorem: $$j(z_k) = j(\bar z) + \rho_k j'(\bar z) v_k + \tfrac{\rho_k^2}{2}j''(\hat z_k) v_k^2, \quad \theta_k \in (0,1),$$ where $\hat z_k = \bar z + \theta_k(z_k - \bar z)$ and $\rho_kv_k = z_k - \bar z$. Now, $j'(\bar z) v_k \geq 0$ and yield $$\tfrac{\rho_k^2}{2}j''(\hat z_k) v_k^2 \leq j(z_k) - j(\bar z) < \tfrac{1}{2k} \| \bar z - z_k\|_{L^2(\Omega)}^2.$$ This implies that $j''(\hat z_k) v_k^2 < k^{-1}$. Consequently, $j''(\hat z_k) v_k^2 \rightarrow 0$ as $k \uparrow \infty$. We now prove that $j''(\bar z) v^2 \leq \liminf_{k} j''(\hat z_k) v_k^2$. We begin by noticing that $$j''(\hat z_k)v_k^2 = \int_{\Omega} \left( \frac{\partial^2 L}{\partial u^2}(x, \hat u_k) \phi_{v_k}^2 - \hat p_k \frac{\partial^2 a}{\partial u^2}(x,\hat u_k)\phi_{v_k}^2 + \alpha v_k^2 \right) \mathrm{d}x.$$ As $k \uparrow \infty$, $\hat z_k \rightarrow \bar z$ and $v_k \rightharpoonup v$ in $L^2(\Omega)$. We thus have $ \hat u_k \rightarrow \bar u $ and $ \hat p_k \rightarrow \bar p $ in $\tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ and $\phi_{v_k} \rightharpoonup \phi_{v}$ in $\tilde H^s(\Omega)$; the latter implies that $\phi_{v_k} \rightarrow \phi_{v}$ in $L^{\mathfrak{q}}(\Omega)$ as $k \uparrow \infty$ for $\mathfrak{q} < 2n/(n-2s)$. Invoke [\[B2\]]{.nodecor} to obtain $$\begin{gathered} \left| \int_{\Omega} \left( \tfrac{\partial^2 L}{\partial u^2}(x, \hat u_k) \phi_{v_k}^2 - \tfrac{\partial^2 L}{\partial u^2}(x, \bar u) \phi_{v}^2 \right) \mathrm{d}x\right| \leq \| \phi_{v_k} \|^2_{L^{\mathfrak{q}}(\Omega)} \left \| \tfrac{\partial^2 L}{\partial u^2}(\cdot, \hat u_k) - \tfrac{\partial^2 L}{\partial u^2}(\cdot, \bar u)\right\|_{L^r(\Omega)} \\ + \| \psi_{\mathfrak{m}} \|_{L^r(\Omega)} \| \phi_v + \phi_{v_k} \|_{L^{\mathfrak{q}}(\Omega)} \| \phi_v - \phi_{v_k} \|_{L^{\mathfrak{q}}(\Omega)} \rightarrow 0, \quad k \uparrow \infty.\end{gathered}$$ On the other hand, invoke [\[A2\]]{.nodecor} to derive $$\begin{gathered} \left| \int_{\Omega} \left( \bar p \frac{\partial^2 a}{\partial u^2}(x, \bar u )\phi_{v}^2 - \hat p_k \frac{\partial^2 a}{\partial u^2}(x,\hat u_k)\phi_{v_k}^2 \right) \mathrm{d}x\right| \leq C_{\mathfrak{m}} \| \phi_v \|^2_{L^{2}(\Omega)} \| \bar p - \hat p_k \|_{L^{\infty}(\Omega)} \\ + C_{\mathfrak{m}} \| \hat p_k \|_{L^{r}(\Omega)} \left( \| \bar u - \hat u_k \|_{L^{\infty}(\Omega)} \| \phi_v \|^2_{L^{\mathfrak{q}}(\Omega)} + \| \phi_v + \phi_{v_k} \|_{L^{\mathfrak{q}}(\Omega)} \| \phi_v - \phi_{v_k} \|_{L^{\mathfrak{q}}(\Omega)} \right) \rightarrow 0\end{gathered}$$ as $k \uparrow \infty$. Finally, observe that $\| v \|^2_{L^2(\Omega)} \leq \liminf_{k} \| v_k\|^2_{L^2(\Omega)}$ because $\| \cdot \|^2_{L^2(\Omega)}$ is weakly lower semicontinuous. We can thus conclude that $j''(\bar z) v^2 \leq \liminf_{k} j''(\hat z_k) v_k^2$. Finally, since $ \liminf_{k} j''(\hat z_k) v_k^2\leq 0$ and $v \in C_{\bar z}$, implies that $v = 0$. Since $v=0$, $\phi_{v_k} \rightarrow 0$ in $L^{\mathfrak{q}}(\Omega)$ as $k \uparrow \infty$ for $\mathfrak{q} < 2n/(n-2s)$. This implies $$\alpha = \alpha \| v_k\|_{L^2(\Omega)}^2 \leq \liminf_{k \uparrow \infty} j''(\hat z_k) v_k^2 \leq 0,$$ which contradicts the fact that $\alpha > 0$. This concludes the proof. Define, for $\tau >0$, $$C_{\bar{z}}^\tau:=\{v\in L^2(\Omega): \eqref{eq:sign_cond} \textnormal{ holds and } |\bar{\mathfrak{p}}(x)|>\tau \implies v(x)=0 \}. \label{def:critical_cone_tau}$$ \[thm:equivalent\_opt\_cond\] Let $n \in \{ 2,3 \}$ and $s > n/4$. Let $\bar z \in \mathbb{Z}_{ad}$ be locally optimal for problem –. Thus, is equivalent to $$\label{eq:second_order_equivalent} \exists \mu, \tau >0: \quad j''(\bar{z})v^2 \geq \mu \|v\|_{L^2(\Omega)}^2 \quad \forall v \in C_{\bar{z}}^\tau,$$ where $C_{\bar{z}}^\tau$ is defined in . Since $C_{\bar z} \subset C_{\bar z}^{\tau}$, we immediately conclude that implies . To prove that implies we proceed by contradiction. Assume that, for $\tau >0$, $$\exists v_{\tau} \in C_{\bar z}^{\tau}: \quad j''(\bar z) v_{\tau}^2 < \tau \| v_{\tau}\|_{L^2(\Omega)}^2.$$ Define $w_{\tau} := \|v_\tau \|_{L^2(\Omega)}^{-1}v_{\tau} $. Note that, up to a nonrelabeled subsequence if necessary, $$\label{eq:weak_seq} w_{\tau} \in C_{\bar z}^{\tau}, \qquad \|w_{\tau}\|_{L^2(\Omega)}=1, \qquad j''(\bar{z})w_{\tau}^2<\tau, \qquad w_{\tau} \rightharpoonup w \text{ in } L^2(\Omega).$$ We prove that $w \in C_{\bar z}$. Since the set of elements satisfying is weakly closed in $L^2(\Omega)$, we conclude that $w$ satisfies . On the other hand, $$\int_{\Omega} \mathfrak{p}(x) w(x) \mathrm{d}x = \lim_{\tau \downarrow 0} \int_{\Omega} \mathfrak{p}(x) w_{\tau}(x) \mathrm{d}x = \lim_{\tau \downarrow 0} \int_{|\mathfrak{p}(x)| \leq \tau} \mathfrak{p}(x) w_{\tau}(x) \mathrm{d}x \leq \lim_{\tau \downarrow 0} \tau \sqrt{|\Omega|} = 0,$$ where we have used that $\mathfrak{p} \in L^2(\Omega)$, $w_{\tau}\rightharpoonup w$ in $L^2(\Omega)$, $w_{\tau} \in C_{\bar z}^{\tau}$, and $\| w_{\tau}\|_{L^2(\Omega)} = 1$. As a result, $\int_{\Omega} |\mathfrak{p}(x) w(x)| \mathrm{d}x = \int_{\Omega}\mathfrak{p}(x) w(x) \mathrm{d}x = 0$. This implies that if $|\mathfrak{p}(x)| \neq 0$, then $w(x) = 0$ for a.e. $x \in \Omega$. We have thus obtained that $w \in C_{\bar z}$. We prove that $w = 0$. Since $w \in C_{\bar z}$, implies that $w = 0$ or $j''(\bar z) w^2 >0$. On the other hand, the arguments elaborated in the step 2 of the proof of Theorem \[thm:suff\_opt\_cond\] in conjunction with yield $ j''(\bar z) w^2 \leq \liminf_{\tau \downarrow 0} j''(\bar z) w_{\tau}^2 \leq \limsup_{\tau \downarrow 0} j''(\bar z) w_{\tau}^2 \leq 0. $ Consequently, $w = 0$ and $\lim_{\tau \downarrow 0} j''(\bar z) w_{\tau}^2 = 0$. Finally, since $w = 0$ and $w_{\tau} \rightharpoonup 0$ in $L^2(\Omega)$ as $\tau \downarrow 0$, we have that $\phi_{w_{\tau}} \rightarrow 0$ in $L^{\mathfrak{q}}(\Omega)$, as $\tau \downarrow 0$, for $\mathfrak{q} < 2n/(n-2s)$. Thus, $ \alpha = \alpha \| w_{\tau} \|_{L^2(\Omega)}^2 \leq \liminf_{\tau \downarrow 0} j''(\bar z) w_{\tau}^2 = 0, $ which is a contradiction. This concludes the proof. Regularity estimates {#sec:regularity} -------------------- In this section, we derive regularity estimates for the optimal control variables. In doing so, the following regularity result for the linear case $a \equiv 0$ will be of importance. Let $n \geq 1$, $s \in (0,1)$, and $\Omega$ be a domain such that $\partial \Omega \in C^{\infty}$. Let $\mathsf{u}$ be the solution to $(-\Delta)^s \mathsf{u} = \mathsf{f}$ in $\Omega$ and $\mathsf{u} = 0$ in $\Omega^c$. If $\mathsf{f} \in H^{t}(\Omega)$, for some $t \geq -s$, then $\mathsf{u} \in H^{s + \vartheta}(\Omega)$, where $\vartheta = \min \{s+t,1/2-\epsilon \}$ and $\epsilon > 0$ is arbitrarily small. In addition, we have $$\| \mathsf{u} \|_{H^{s+\vartheta}(\Omega)} \lesssim \| \mathsf{f} \|_{H^{t}(\Omega)}, \label{eq:regularity_state_smooth}$$ where the hidden constant depends on $\Omega$, $n$, $s$, and $\vartheta$. \[pro:state\_regularity\_smooth\] See [@MR3276603; @MR0185273]. Observe that smoothness of $\mathsf{f}$ does not ensure that the solution to $(-\Delta)^s \mathsf{u} = \mathsf{f}$ in $\Omega$ and $\mathsf{u} = 0$ in $\Omega^c$ is any smoother than $\cap \{ H^{s+1/2-\epsilon}(\Omega): \epsilon > 0 \}$. To present regularity estimates, we will assume that, in addition to \[A1\]–\[A3\] and \[B1\]–\[B2\], the nonlinear functions $a$ and $L$ satisfy \[C1\]–\[C2\]. Let $n \geq 2$ and $s \in (0,1)$. If $\Omega$ is such that $\partial \Omega \in C^{\infty}$, then $\bar u, \bar p \in H^{s + 1/2 - \epsilon}(\Omega)$ and $\bar z \in H^{\gamma}(\Omega)$, where $\gamma = \min \{ 1, s + 1/2 - \epsilon \}$ and $\epsilon$ denotes an arbitrarily small positive constant. \[thm:regularity\_space\] Since $\bar z \in \mathbb{Z}_{ad}$ and $a(\cdot,0) \in L^2(\Omega)$, we apply Proposition \[pro:state\_regularity\_smooth\] with $t = 0$ to obtain $\bar u \in H^{s + \nu}(\Omega)$, where $\nu = \min \{ s , 1/2 - \epsilon \}$ and $\epsilon >0$ is arbitrarily small, and $$\| \bar u \|_{H^{s + \nu}(\Omega)} \lesssim \| \bar z - a(\cdot, \bar u)\|_{L^2(\Omega)} \lesssim \| \bar z \|_{L^{2}(\Omega)} + \| a(\cdot,0) \|_{L^2(\Omega)}, \label{eq:first_estimate_u}$$ upon using that $a$ is locally Lipschitz in the second variable and $\| \bar u \|_s \lesssim \| \bar z \|_{H^{-s}(\Omega)}$. We now obtain a first regularity estimate for $\bar p$. To do this, we first observe that, since $\frac{\partial a}{\partial u} (\cdot, u)$ is locally Lipschitz in $u$ and $\bar u, \bar p \in H^s(\Omega)$, we have $\frac{\partial a}{\partial u}(\cdot, \bar u) \bar p \in H^s(\Omega)$. In fact, notice that, for $x,y \in \Omega$ and $u \in \mathbb{R}$, we have $$\left| \tfrac{\partial a }{\partial u} (x,u) \bar p(x) - \tfrac{\partial a }{\partial u} (y,u) \bar p(y) \right| \leq \left| \tfrac{\partial a }{\partial u} (x,u) \right | |\bar p(x) - \bar p(y)| + |\bar p(y)| \left| \tfrac{\partial a }{\partial u} (x,u) -\tfrac{\partial a }{\partial u} (y,u) \right |.$$ The definition of $| \cdot |_{H^s(\Omega)}$ thus imply $| \tfrac{\partial a }{\partial u} (\cdot,\bar u) \bar p |_{H^s(\Omega)} \lesssim |\bar p|_{H^s(\Omega)} + \| \bar p \|_{L^{\infty}(\Omega)} | \bar u |_{H^s(\Omega)}$, upon using \[A3\] Invoke \[C2\] and Proposition \[pro:state\_regularity\_smooth\] to obtain $\bar p \in H^{s + \iota}(\Omega)$, where $\iota = \min \{ s + \lambda , \frac{1}{2} - \epsilon \}$, $\lambda = \min \{ s, \frac{1}{2} - s - \epsilon \}$, and $\epsilon >0$ is arbitrarily small. In addition, we have the following estimate: $$\begin{gathered} \| \bar p \|_{H^{s + \iota}(\Omega)} \lesssim \| \tfrac{\partial a }{\partial u} (\cdot,\bar u) \bar p \|_{H^{s}(\Omega)} + \| \tfrac{\partial L}{\partial u}(\cdot,\bar u) \|_{H^{\frac{1}{2} - s - \epsilon}(\Omega)} \\ \lesssim |\bar p|_{H^s(\Omega)} + \| \bar p \|_{L^{\infty}(\Omega)} | \bar u |_{H^s(\Omega)} + \| \tfrac{\partial L}{\partial u} (\cdot,\bar u) \|_{H^{\frac{1}{2} - s - \epsilon}(\Omega)}. \label{eq:first_estimate_p}\end{gathered}$$ A nonlinear interpolation result based on [@MR1786735 Theorem A.1] and [@Tartar Lemma 28.1] allows to obtain that $\bar z \in H^{\upsilon}(\Omega)$, for $\upsilon = \min \{ 1, s + \iota \}$, with a similar estimate. We now consider four cases. $s \in ( \frac{1}{2}, 1 )$: Observe that $\nu = \iota = \tfrac{1}{2} - \epsilon$. Thus, $\bar u , \bar p \in H^{s + \frac{1}{2} - \epsilon}(\Omega)$, for $\epsilon >0$ being arbitrarily small, and $\bar z \in H^1(\Omega)$. In addition, and yield $$\begin{gathered} \| \bar u \|_{H^{s+\frac12 - \epsilon}(\Omega)} + \| \bar p \|_{H^{s+\frac12 - \epsilon}(\Omega)} + \| \bar z \|_{H^1(\Omega)} \lesssim |\bar p|_{H^s(\Omega)} + \| \bar p \|_{L^{\infty}(\Omega)} | \bar u |_{H^s(\Omega)} \\ + \| \bar z \|_{L^{2}(\Omega)} + \| \tfrac{\partial L}{\partial u} (\cdot,\bar u) \|_{H^{\frac{1}{2} - s - \epsilon}(\Omega)} + \| a(\cdot,0) \|_{L^2(\Omega)} =: \mathfrak{B}.\end{gathered}$$ $s = \tfrac{1}{2}$. The proof follows similar arguments. For brevity, we skip the details. $s \in [\tfrac{1}{4},\tfrac{1}{2})$. Here, $\nu = s$ and $\iota = \tfrac{1}{2} - \epsilon$. Thus, $\bar u \in H^{2s}(\Omega)$ and $\bar p \in H^{s+ \frac{1}{2} - \epsilon}(\Omega)$, where $\epsilon >0$ is arbitrarily small. In view of , a nonlinear interpolation argument yields $\bar z \in H^{s+ \frac{1}{2} - \epsilon}(\Omega)$. In addition, we have the estimate $$\| \bar u \|_{H^{2s}(\Omega)} + \| \bar p \|_{H^{s+\frac12 - \epsilon}(\Omega)} + \| \bar z \|_{H^{s+\frac12 - \epsilon}(\Omega)} \lesssim \mathfrak{B}.$$ Invoke Proposition \[pro:state\_regularity\_smooth\] with $t = \tfrac{1}{2} - s - \epsilon$ and \[C1\] to obtain that $\bar u \in H^{s + \frac{1}{2} - \epsilon}(\Omega)$. Observe that $ s + \tfrac{1}{2} - \epsilon > \tfrac{1}{2} - s - \epsilon$ and $2s>\tfrac{1}{2} - s - \epsilon$ for $\epsilon >0$ arbitrarily small. $s \in (0,\tfrac{1}{4})$. We proceed on the basis of a bootstrap argument as in [@MR3990191]. Since $s < \tfrac{1}{4}$, $\nu = s$ and $\iota = 2s$. Thus $\bar u \in H^{2s}(\Omega)$, $\bar p \in H^{3s}(\Omega)$, $\bar z \in H^{3s}(\Omega)$, and $$\| \bar u \|_{H^{2s}(\Omega)} + \| \bar p \|_{H^{3s}(\Omega)} + \| \bar z \|_{H^{3s}(\Omega)} \lesssim \mathfrak{B}.$$ $s \in [\tfrac{1}{6},\tfrac{1}{4})$. Observe that $2s > \tfrac{1}{2} - s - \epsilon$ for $\epsilon >0$ being arbitrarily small. Invoke Proposition \[pro:state\_regularity\_smooth\] with $t = 1/2 - s - \epsilon$ to obtain $$\begin{gathered} \| \bar u \|_{H^{s + \frac{1}{2} - \epsilon}(\Omega)} \lesssim \| z \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)} + \| a(\cdot,\bar u) - a(\cdot,0) \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)} + \| a(\cdot,0) \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)} \\ \lesssim \| \bar z \|_{H^{2s}(\Omega)} + \| \bar u \|_{H^{2s}(\Omega)} + \| a(\cdot,0) \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)} \lesssim \mathfrak{B} + \| a(\cdot,0) \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)}. $$ By assumption $\frac{\partial L}{\partial u}(\cdot,0) \in L^2(\Omega) \cap H^{\frac{1}{2} - s - \epsilon}(\Omega)$. On the other hand, $\tfrac{\partial a}{\partial u}(\cdot,\bar u) \bar p \in H^{2s}(\Omega)$, because $\bar p \in H^{3s}(\Omega)$ and $\bar u \in H^{2s}(\Omega)$. Thus, Proposition \[pro:state\_regularity\_smooth\] with $t = \frac{1}{2} - s - \epsilon$ yields $$\| \bar p \|_{H^{s + \frac{1}{2} - \epsilon}(\Omega)} \lesssim |\bar p|_{H^{2s}(\Omega)} + \| \bar p \|_{L^{\infty}(\Omega)} | \bar u |_{H^{ 2s}(\Omega)} + \| \tfrac{\partial L}{\partial u} (\cdot,\bar u) \|_{H^{ \frac{1}{2} - s - \epsilon}(\Omega)}. $$ A nonlinear interpolation argument yields $\bar z \in H^{s+\frac{1}{2} - \epsilon}(\Omega)$ with a similar estimate. $s \in (0,\tfrac{1}{6})$. Invoke Proposition \[pro:state\_regularity\_smooth\] with $t = 2s$ to obtain $\bar u \in H^{3s}(\Omega)$ and $$\| \bar u \|_{H^{3s}(\Omega)} \\ \lesssim \| \bar z \|_{H^{2s}(\Omega)} + \| \bar u \|_{H^{2s}(\Omega)} + \| a(\cdot,0) \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)} \lesssim \mathfrak{B} + \| a(\cdot,0) \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)}.$$ $s \in [\tfrac{1}{8},\tfrac{1}{6})$. Observe that $3s>\frac{1}{2} -s-\epsilon$ for $\epsilon >0$ being arbitrarily small. Invoke Proposition \[pro:state\_regularity\_smooth\] with with $t = \frac{1}{2} - s - \epsilon$ to obtain $\bar u \in H^{s+ \frac{1}{2} - \epsilon}(\Omega)$ with $$\| \bar u \|_{H^{s + \frac{1}{2} - \epsilon}(\Omega)} \lesssim \| \bar z\|_{H^{3s}(\Omega)} + \| \bar u \|_{H^{3s}(\Omega)} + \| a(\cdot,0) \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)} \lesssim \mathfrak{B} + \| a(\cdot,0) \|_{H^{\frac{1}{2} -s-\epsilon}(\Omega)}.$$ Invoke Proposition \[pro:state\_regularity\_smooth\] again to deduce that $\bar p, \bar z \in H^{s+ \frac{1}{2} - \epsilon}(\Omega)$. $s \in (0, \tfrac{1}{8})$. Since $\bar u, \bar p \in H^{3s}(\Omega)$, we have that $\frac{\partial a}{\partial u}(\cdot,\bar u) \bar p \in H^{3s}(\Omega)$. Apply Proposition \[pro:state\_regularity\_smooth\] with $t = 3s$ to obtain $$\| \bar p \|_{H^{4s}(\Omega)} \lesssim |\bar p|_{H^{3s}(\Omega)} + \| \bar p \|_{L^{\infty}(\Omega)} | \bar u |_{H^{ 3s}(\Omega)} + \| \tfrac{\partial L}{\partial u} (\cdot,\bar u) \|_{H^{\frac{1}{2} - s - \epsilon}(\Omega)}.$$ Observe that $3s \leq \tfrac{1}{2} - s - \epsilon$ for $\epsilon >0$ being arbitrarily small. A nonlinear interpolation argument yields $\bar z \in H^{4s}(\Omega)$ with a similar estimate. Similarly, we obtain $\bar u \in H^{4s}(\Omega)$. $s \in [\tfrac{1}{10}, \tfrac{1}{8})$. Observe that $4s>\frac{1}{2} -s-\epsilon$ for $\epsilon >0$ being arbitrarily small. We thus immediately obtain that $\bar u, \bar p, \bar z \in H^{s+ \frac{1}{2} - \epsilon}(\Omega)$ with a suitable estimate. $s \in (0,\tfrac{1}{10})$. Since $\bar u, \bar p \in H^{4s}(\Omega)$, we have that $\frac{\partial a}{\partial u}(\cdot,\bar u) \bar p \in H^{4s}(\Omega)$. Apply Proposition \[pro:state\_regularity\_smooth\] with $t = 4s$ to obtain $\bar p \in H^{5s}(\Omega)$. A nonlinear interpolation argument yields $\bar z \in H^{5s}(\Omega)$ with a suitable estimate. From this procedure we note that, at every step, there is a regularity gain. Consequently, after a finite number of steps, which is proportional to $s^{-1}$, we can conclude that the desired regularity results hold. This concludes the proof. Finite element approximation of fractional semilinear PDEs {#sec:fem} ========================================================== In this section, we analyze the convergence properties of suitable discretizations and derive a priori error estimates. For analyzing convergence properties, it will be sufficient to assume that $\Omega$ is an open and bounded Lipschitz polytope. However, additional assumptions on $\Omega$ will be imposed for deriving error estimates: $\Omega$ is smooth and convex; convexity being assumed for simplicity. Since in this case $\Omega$ cannot be meshed exactly, we consider curved simplices to discretize $\Omega \setminus \Omega_h$. Here, $\Omega_h$ denotes a suitable polytopal domain that *approximates* $\Omega$. For the sake of brevity, we restrict the presentation to open and bounded domains $\Omega \subset \mathbb{R}^n$ $(n \geq 2)$ such that $\partial \Omega \in C^2$; for Lipschitz polytopes the presentation is simpler (see Remark \[rem:polytopes\]). We follow [@MR773854 Section 5.2] and consider a family of open, bounded, and convex polytopal domains $\{ \Omega_h \}_{h>0}$, based on a family of quasi-uniform partitions made of closed simplices $\{ {\mathscr{T}}_h \}_{h>0}$, that approximate $\Omega$ in the following sense: $$\mathcal{N}_h \subset \bar \Omega_h, \quad \mathcal{N}_h \cap \partial \Omega_h \subset \partial \Omega, \quad |\Omega \setminus \Omega_h| \lesssim h^2. \label{eq:properties_of_Omegah}$$ Here, $h = \max_{T \in {\mathscr{T}}_h} h_T$ denotes the mesh-size of the quasi-uniform partition ${\mathscr{T}}_h = \{ T \}$, where $h_T = \mathrm{diam}(T)$, and $\mathcal{N}_h$ correspond to the set of all nodes of the mesh ${\mathscr{T}}_h$. We shall also assume that $\Omega$ is convex so that $\Omega_h \subset \Omega$ for every $h>0$. Given a mesh ${\mathscr{T}}_h$, we define the finite element space of continuous piecewise polynomials of degree one as $${\mathbb{V}}_h = \left\{ v_h \in C^0( \overline {\Omega} ): {v_h}_{|T} \in \mathbb{P}_1(T) \ \forall T \in {\mathscr{T}}_h, \ v_{h} = 0 \textrm{ on } \overline \Omega \setminus \Omega_h \right\}. \label{eq:defFESpace}$$ Note that discrete functions are trivially extended by zero to $\Omega^c$ and that we enforce a classical homogeneous Dirichlet boundary condition at the degrees of freedom that are located at the boundary of $\Omega_h$. If $\Omega$ is a Lipschitz polytope the previous construction is not necessary: $\Omega = \Omega_h$ and ${\mathbb{V}}_h = \{ v_h \in C^0( \overline{\Omega} ): {v_h}_{|T} \in \mathbb{P}_1(T) \ \forall T \in {\mathscr{T}}_h \}$. \[rem:polytopes\] The discrete problem -------------------- We introduce the following finite element approximation of problem : Find $u_h \in {\mathbb{V}}_h$ such that $$\label{eq:discrete_semilinear_pde} \mathcal{A} (u_h,v_h) + \int_{\Omega_h} a(x,u_h(x)) v_h(x) \mathrm{d}x = \int_{\Omega_h} f(x) v_h(x) \mathrm{d}x \quad \forall v_h \in {\mathbb{V}}_h.$$ Let $r>n/2s$ and $f \in L^r(\Omega)$. Let $a = a(x,u) : \Omega \times \mathbb{R} \rightarrow \mathbb{R}$ be a Carathéodory function that is monotone increasing in $u$. Assume, in addition, that $a$ satisfies and $a(\cdot,0) \in L^r(\Omega)$. Withing this setting, Theorem \[thm:stata\_equation\_well\_posedness\] guarantees that the continuous problem admits a unique solution $u \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ satisfying . Since $\mathcal{A}$ is coercive and $a$ is monotone increasing in $u$, an appllication of Brouwer’s fixed point theorem [@MR816732 Proposition 2.6] yields the existence of a unique solution for ; see also the proof of [@MR1033498 Theorem 26.A]. In addition, $\| u_h \|_s \lesssim \| f \|_{H^{-s}(\Omega)}$ for every $h>0$. Regularity estimates {#regularity-estimates} -------------------- Before deriving error estimates, it is of fundamental importance the understanding of regularity estimates for the solution of . \[thm:regularity\_space\_state\_equation\] Let $n \geq 2$, $s \in (0,1)$, and $\Omega$ be a domain such that $\partial \Omega \in C^{\infty}$. Assume, in addition, that $a$ is locally Lipschitz with respect to the second variable. If both $a(\cdot,0)$ and $f$ belong to $H^{1/2 - s - \epsilon}(\Omega)$, with $\epsilon$ arbitrarily small, then $u \in H^{s + 1/2 - \epsilon}(\Omega)$. Define $\lambda := \min \{ s, \frac{1}{2} - s - \epsilon \}$, where $\epsilon >0$ is arbitrarily small. Apply Proposition \[pro:state\_regularity\_smooth\] with $t = \lambda$ to obtain $u \in H^{s + \nu}(\Omega)$, where $\nu = \min \{ s + \lambda, \frac{1}{2} - \epsilon \}$, and $$\| u \|_{H^{s + \nu }(\Omega)} \lesssim \| f \|_{H^{\frac12 -s - \epsilon}(\Omega)} + \| a (\cdot,u ) - a (\cdot,0 ) \|_{H^{\lambda}(\Omega)} + \| a (\cdot,0 ) \|_{H^{1/2 - s - \epsilon}(\Omega)}. \label{eq:reg_u_aux1}$$ We now proceed in several steps on the basis of a bootstrap argument. If $s \geq \tfrac{1}{4}$, then $s > \tfrac{1}{2} -s - \epsilon$, with $\epsilon >0$ arbitrarily small. Thus, $\lambda = \frac{1}{2} - s - \epsilon$ and $$\| u \|_{H^{s+\frac12 - \epsilon}(\Omega)} \lesssim \| f \|_{H^{\frac12 -s - \epsilon}(\Omega)} + \| u \|_{H^s(\Omega)} + \| a (\cdot,0 ) \|_{H^{1/2 - s - \epsilon}(\Omega)},$$ where we have used the local Lipschitz property of $a$ in the second variable. Invoke the basic estimate $\| u \|_s \lesssim \| f \|_{H^{-s}(\Omega)}$ to bound $\| u \|_{H^s(\Omega)}$. If $s \in (0,\tfrac{1}{4})$, we have $\lambda = s$ and $\nu = 2s$. Consequently, $u \in H^{3s}(\Omega)$ with the estimate . Define $\iota := \min \{ 3s,\frac{1}{2} - s - \epsilon\}$, where $\epsilon >0$ is arbitrarily small. Apply Proposition \[pro:state\_regularity\_smooth\] with $t = \iota$ to obtain $u \in H^{s + \kappa}(\Omega)$, where $\kappa = \min \{ s + \iota, \tfrac{1}{2} - \epsilon \}$, and $$\| u \|_{H^{s + \kappa}(\Omega)} \lesssim \| f \|_{H^{\frac12 -s - \epsilon}(\Omega)} + \| a (\cdot,u ) - a (\cdot,0 ) \|_{H^{\iota}(\Omega)} + \| a (\cdot,0 ) \|_{H^{1/2 - s - \epsilon}(\Omega)}. \label{eq:reg_u_aux2}$$ If $s \in [\tfrac{1}{8},\tfrac{1}{4})$, then $\iota = \tfrac{1}{2} -s - \epsilon$, $\kappa = \tfrac{1}{2} - \epsilon$, and $u \in H^{s+\frac12 - \epsilon}(\Omega)$. If $s \in (0,\tfrac{1}{8})$, we have $\iota = 3s$ and $\kappa = 4s$. Consequently, $u \in H^{5s}(\Omega)$ with the estimate . Define $\mu:= \min \{ 5s, \tfrac{1}{2} -s - \epsilon\}$, where $\epsilon >0$ is arbitrarily small. Invoke Proposition \[pro:state\_regularity\_smooth\] with $t = \mu$ to obtain that $u \in H^{s + \upsilon}(\Omega)$, where $\upsilon = \min \{ s + \mu, \tfrac{1}{2} - \epsilon \}$. If $s \in [\tfrac{1}{12},\tfrac{1}{4})$, then $\mu = \tfrac{1}{2} -s - \epsilon$, $\upsilon = \tfrac{1}{2} - \epsilon$, and $u \in H^{s+\frac12 - \epsilon}(\Omega)$. From this procedure we note that, at every step, there is a regularity gain of $2s$. Consequently, after a finite number of steps we can conclude that $u \in H^{s+\frac{1}{2} - \epsilon}(\Omega)$, with $\epsilon >0$ arbitrarily small. Error estimates --------------- We now present error estimates. In doing so, we will assume, in addition, that there exists $\phi \in L^{r}(\Omega)$, with $r> n/2s$, such that $$|a(x,u) - a(x,v)| \leq |\phi(x)| |u-v| ~\textrm{a.e.}~x \in \Omega,~u,v \in \mathbb{R}. \label{eq:assumption_on_a_phi}$$ Let $n \geq 2$, $s \in (0,1)$, and $r>n/2s$. Let $\Omega$ be an open and bounded domain with Lipschitz boundary. Assume that $a$ is as in the statement of Theorem \[thm:stata\_equation\_well\_posedness\]. Assume, in addition, that $a$ satisfies . Let $u \in \tilde H^s(\Omega)$ be the solution to and let $u_h \in \mathbb{V}_h$ be its finite element approximation obtained as the solution to . Then, we have the quasi–best approximation result $$\label{eq:error_estimate_semilinear_s} \| u - u_h \|_{s} \lesssim \| u - v_h\|_{s} \quad \forall v_h \in \mathbb{V}_h.$$ If, in addition, $\Omega$ is smooth and convex and both $a(\cdot,0)$ and $f$ belong to $H^{1/2 - s - \epsilon}(\Omega)$, with $\epsilon>0$ arbitrarily small, then $$\label{eq:error_estimate_semilinear_s_final} \| u - u_h \|_{s} \lesssim h^{\frac{1}{2} - \epsilon } \| u \|_{H^{s + 1/2 - \epsilon}(\Omega)},$$ If, in addition, holds with $r \geq n/s$, then $$\label{eq:error_estimate_semilinear_s_final_L2} \| u - u_h \|_{L^2(\Omega)} \lesssim h^{\vartheta + \frac{1}{2} - \epsilon} \| u \|_{H^{s + 1/2 - \epsilon}(\Omega)}.$$ Here, $\vartheta = \min \{ s , \tfrac{1}{2} - \epsilon \}$ with $\epsilon >0$ being arbitrarily small. In all three estimates the hidden constant is independent of $u$, $u_h$, and $h$. \[thm:error\_estimates\_state\_equation\] Since $a$ is monotone increasing in the second variable, we obtain $$\begin{aligned} \| u - u_h \|_s^2 & = \mathcal{A}(u - u_h,u - u_h) \leq \mathcal{A}(u - u_h,u - u_h) + (a(\cdot,u) - a(\cdot,u_h), u - u_h)_{L^2(\Omega)} \\ & = \mathcal{A}(u - u_h,u - v_h) + (a(\cdot,u) - a(\cdot,u_h), u - v_h)_{L^2(\Omega)}, \quad v_h \in \mathbb{V}_h\end{aligned}$$ upon utilizing Galerkin orthogonality. Invoke and $H^s(\Omega) \hookrightarrow L^{\mathfrak{q}}(\Omega) $ with $\mathfrak{q} \leq 2n/(n-2s)$ to obtain . Assume now that $\Omega$ is smooth and convex so Theorem \[thm:regularity\_space\_state\_equation\] applies; convexity being assumed for simplicity. To bound $\| u - v_h \|_s$ we first invoke [@McLean Theorem 3.3.3]: $$\| u -v_h \|_s \lesssim \| u - v_h \|_{H^s(\Omega)} \quad \forall v_h \in \mathbb{V}_h, \qquad s \in (0,1) \setminus \{ \tfrac{1}{2} \}.$$ The second ingredient is the localization of fractional order Sobolev seminorms [@MR1752263; @MR1930387]: $$|v|^2_{H^s(\Omega)} \leq \sum_{T} \left[ \int_{T} \int_{S_T} \frac{|v(x) - v(y)|^2}{|x-y|^{n+2s}} \mathrm{d}y \mathrm{d}x + \frac{\mathfrak{c}}{ s h_T^{2s} } \| v \|^2_{L^2(T)} \right], \quad s \in (0,1), \quad \mathfrak{c}>0,$$ for $v \in H^s(\Omega)$; $S_T $ denotes a suitable patch associated to $T$. We stress that curved domains/simplices are also handled in [@MR1752263; @MR1930387]. It thus suffices to note that, if $\mathfrak{T}$ denotes a boundary curved simplex, the fact that $u \in \tilde H^s(\Omega) \cap H^{s+1/2 - \epsilon}(\Omega)$, with $\epsilon >0$ arbitrarily small, implies $$\| u - u_h \|_{L^2(\mathfrak{T})} = \| u \|_{L^2(\Omega \setminus \Omega_h)} \lesssim h^{2\upsilon} \| u \|_{H^{\upsilon}(\Omega)}, \quad \upsilon = \min \{1,s+1/2 - \epsilon \},$$ which follows from interpolating [@MR773854 estimate (5.2.18)] and $\| v \|_{L^2(\Omega \setminus \Omega_h)} \leq \| v \|_{L^2(\Omega)}$. On the other hand, if $v \in \tilde H^s(\Omega) \cap H^{s+1/2 - \epsilon}(\Omega)$, with $\epsilon>0$ arbitrarily small, then $$\int_{\mathfrak{T}} \int_{S_{\mathfrak{T}}} \frac{|v(x) - v(y)|^2}{|x-y|^{n+2s}} \mathrm{d}y \mathrm{d}x \leq |v|^2_{H^s(S_\mathfrak{T})} \lesssim h^{2(1/2 - \epsilon)} \| v \|^2_{H^{s+1/2 - \epsilon}(\Omega)}.$$ We thus utilize interpolation error estimates for the Scott–Zhang operator [@MR3893441 Proposition 3.6] and Theorem \[thm:regularity\_space\_state\_equation\] to arrive at the estimate ; see [@MR3893441 Section 3.2] for details and the particular treatment of the case $s = 1/2$. The error estimate in $L^2(\Omega)$ follows from duality. Define $ 0 \leq \chi \in L^{r}(\Omega)$ by $$\chi(x) = \frac{a(x,u(x)) - a(x,u_h(x))}{u(x) - u_h(x)} ~\mathrm{if}~u(x) \neq u_h(x), \quad \chi(x) = 0 ~\mathrm{if}~u(x) = u_h(x).$$ Let $\mathfrak{z} \in \tilde H^s(\Omega)$ be the solution to $ \mathcal{A}(v,\mathfrak{z}) + (\chi \mathfrak{z} ,v)_{L^2(\Omega)} = \langle \mathfrak{f} , v \rangle $ for all $v \in \tilde H^s(\Omega)$; $\mathfrak{f} \in H^{-s}(\Omega)$. Let $\mathfrak{z}_h$ be the finite element approximation of $\mathfrak{z}$ within $\mathbb{V}_h$. Thus, $$\begin{aligned} \langle \mathfrak{f} , u - u_h \rangle & = \mathcal{A}(u-u_h,\mathfrak{z}) + (\chi \mathfrak{z}, u - u_h)_{L^2(\Omega)} = \mathcal{A}(u-u_h,\mathfrak{z} - \mathfrak{z}_h) + \mathcal{A}(u-u_h,\mathfrak{z}_h) \\ & + (\chi \mathfrak{z}, u - u_h)_{L^2(\Omega)} = \mathcal{A}(u-u_h,\mathfrak{z} - \mathfrak{z}_h) + (a(\cdot,u) - a(\cdot,u_h), \mathfrak{z} - \mathfrak{z}_h)_{L^2(\Omega)} \\ & \leq \| u - u_h \|_s \| \mathfrak{z} - \mathfrak{z}_h \|_s + \| \phi \|_{L^{r}(\Omega)} \| u - u_h \|_{L^{q}(\Omega)} \| \mathfrak{z} - \mathfrak{z}_h \|_{L^{q}(\Omega)}.\end{aligned}$$ Here, $q$ satisfies $2q^{-1} + r^{-1} = 1$, i.e., $q < 2n/(n-2s)$. Set $\mathfrak{f} = u - u_h \in L^2(\Omega)$. Notice that, since $\phi \in L^r(\Omega)$, with $r \geq n/s$, $\chi \mathfrak{z}$ belong to $L^2(\Omega)$. We can thus invoke Proposition \[pro:state\_regularity\_smooth\] with $t=0$ to obtain $\| \mathfrak{z} \|_{H^{s+\theta}(\Omega)} \lesssim \| u - u_h \|_{L^2(\Omega)}$. Consequently, $$\| u - u_h \|^2_{L^2(\Omega)} \lesssim \| u - u_h \|_s \| \mathfrak{z} - \mathfrak{z}_h \|_s \lesssim h^{ \frac{1}{2} - \epsilon} \|u\|_{H^{s+\frac{1}{2} - \epsilon}(\Omega)} h^{\vartheta} \| u - u_h \|_{L^{2}(\Omega)},$$ where $\vartheta = \min \{s,1/2-\epsilon\}$ and $\epsilon >0$ is arbitrarily small. This concludes the proof. Convergence properties ---------------------- Let $1<p<\infty$ and let $\{ f_h \}_{h>0}$ be a sequence such that $f_h \in L^p(\Omega_h)$. We will say that $f_h \rightharpoonup f$ in $L^{p}(\Omega)$ as $h \downarrow0$ if $f \in L^p(\Omega)$ and $$\int_{\Omega_h} f_h(x) v(x) \mathrm{d}x \rightarrow \int_{\Omega} f(x) v(x) \mathrm{d}x \quad \forall v \in L^{q}(\Omega), \quad h \downarrow 0, \quad p^{-1} + q^{-1} = 1. \label{eq:weak_convergence_h}$$ If $p = \infty$, we will say that $f_h \mathrel{\ensurestackMath{\stackon[1pt]{\rightharpoonup}{\scriptstyle\ast}}} f$ in $L^{\infty}(\Omega)$ if $f \in L^{\infty}(\Omega)$ and holds for every $v \in L^1(\Omega)$. Observe that, upon considering a suitable extension of $f_h$ to $\Omega \setminus \Omega_h$, $f_h$ can be understood as an element of $L^{p}(\Omega)$. Since $| \Omega \setminus \Omega_h | \rightarrow 0$ as $h \downarrow 0$, is equivalent to $\int_{\Omega} \tilde{f}_h(x) v(x) \mathrm{d}x \rightarrow \int_{\Omega} f(x) v(x) \mathrm{d}x$ for $\{ \tilde v_h \}_{h>0}$ being a uniformly bounded extension of $\{ v_h\}_{h>0}$ to $\Omega$. If $\Omega$ is a Lipschitz polytope, then reduces to the standard concept of weak convergence in $L^p(\Omega)$ because $\Omega_ h = \Omega$ for every $h>0$. \[rem:polytopes\_weak\_convergence\] Let $n \geq 2$, $s \in (0,1)$, and $r>n/2s$. Let $\Omega$ be an open, bounded, and convex domain such that $\partial \Omega \in C^{2}$. Assume that $a$ is as in the statement of Theorem \[thm:stata\_equation\_well\_posedness\] and satisfies, in addition, . Let $u \in \tilde H^s(\Omega)$ solves . Let $\mathfrak{u}_h \in \mathbb{V}_h$ be the solution to with $f$ replaced by $f_h \in L^{r}(\Omega_h)$. Then, $$f_h \rightharpoonup f \textrm{ in }L^{r}(\Omega) \implies \mathfrak{u}_h \rightarrow u \textrm{ in } L^{\mathfrak{t}}(\Omega), \quad h \downarrow 0, \quad \mathfrak{t} \leq 2n/(n-2s).$$ Here, $f_h \rightharpoonup f$ in $L^r(\Omega)$ is understood in the sense of . \[eq:proposition\_convergence\] A simple application of the triangle inequality yields $$\| u - \mathfrak{u}_h \|_{L^{\mathfrak{t}}(\Omega)} \leq \| u - u_h \|_{L^{\mathfrak{t}}(\Omega)} + \| u_h - \mathfrak{u}_h \|_{L^{\mathfrak{t}}(\Omega)}, \quad \mathfrak{t} \leq 2n/(n-2s),$$ where $u_h$ denotes the solution to . Since $H^s(\Omega) \hookrightarrow L^{\mathfrak{q}}(\Omega)$ for $\mathfrak{q} \leq 2n/(n-2s)$, the quasi–best approximation estimate yields $\| u - u_h \|_{L^{\mathsf{q}}(\Omega)} \lesssim \| u - v_h \|_s$ for an arbitrary $v_h \in \mathbb{V}_h$. A density argument as in [@MR0520174 Theorem 3.2.3] reveals the convergence result $\| u - u_h \|_{L^{\mathsf{q}}(\Omega)} \rightarrow 0$ as $h \downarrow 0$. To control $\| u_h - \mathfrak{u}_h \|_{L^{\mathfrak{t}}(\Omega)}$ we invoke the problems that $u_h$ and $ \mathfrak{u}_h$ solve: $$\begin{gathered} \| u_h - \mathfrak{u}_h \|^2_s = \mathcal{A}(u_h - \mathfrak{u}_h , u_h - \mathfrak{u}_h ) = (f - f_h, u_h - \mathfrak{u}_h)_{L^2(\Omega)} \\ - \left( a(\cdot,u_h) - a(\cdot,\mathfrak{u}_h), u_h - \mathfrak{u}_h \right)_{L^2(\Omega)} \leq \| f - f_h \|_{H^{-s}(\Omega)} \| u_h - \mathfrak{u}_h \|_s.\end{gathered}$$ This immediately yields $\| u_h - \mathfrak{u}_h \|_{L^{\mathfrak{t}}(\Omega)} \lesssim \| f - f_h \|_{H^{-s}(\Omega)}$. Since $f_h \rightharpoonup f$ in $L^{r}(\Omega)$ we can thus obtain that $ \| u_h - \mathfrak{u}_h \|_{L^{\mathfrak{t}}(\Omega)} \rightarrow 0$ as $h \downarrow 0$. This concludes the proof. The result of Proposition \[eq:proposition\_convergence\] can also be obtained for Lipschitz polytopes; observe that the involved arguments do not utilize further regularity beyond what is natural for the problem: $u \in \tilde H^s(\Omega) \cap L^{\infty}(\Omega)$. \[rem:polytopes\_state\_convergence\] Finite element approximation for the optimal control problem {#sec:fem_control} ============================================================ In this section, we propose a finite element discretization for our control problem. We analyze convergence properties and derive, when possible, error estimates. To accomplish this task, we operate within the discrete setting introduced in section \[sec:fem\] and introduce, in addition, the finite element space of piecewise constant functions $$\mathbb{Z}_h = \left\{ v_{{\mathscr{T}}} \in L^{\infty}( \Omega_h ): {v_{{\mathscr{T}}}}_{|T} \in \mathbb{P}_0(T) \ \forall T \in {\mathscr{T}}_h \right\} \label{eq:piecewise_constant_functions}$$ and the space of discrete admissible controls $ \mathbb{Z}_{\mathrm{ad},h} = \mathbb{Z}_{\mathrm{ad}} \cap \mathbb{Z}_h. $ The discrete optimal control problem {#sec:discrete_optimal_control_problem} ------------------------------------ We consider the following discrete counterpart of the continuous optimal control problem –: Find $$\label{eq:min_discrete} \min \{ J_h(u_h,z_h): (u_h,z_h) \in \mathbb{V}_h \times \mathbb{Z}_{ad,h} \}$$ subject to the *discrete state equation* $$\label{eq:weak_st_eq_discrete} \mathcal{A}( u_h, v_h)+\int_{\Omega_h} a(x,u_h(x)) v_h(x) \mathrm{d}x = \int_{\Omega_h} z_h(x) v_h(x) \mathrm{d}x \quad \forall v \in \mathbb{V}_h.$$ Here, $J_h: \mathbb{V}_h \times \mathbb{Z}_{ad,h} \ni (u_h,z_h) \mapsto J_h(u_h,z_h):= \int_{\Omega_h} L(x,u_h(x))\mathrm{d}x + \frac{\alpha}{2} \| z_h \|^2_{L^2(\Omega_h)} \in \mathbb{R}$. We present the following result. \[thm:existence\_discrete\_control\] Let $n \geq 2$ and $s \in (0,1)$. Assume that [\[A1\]]{.nodecor}–[\[A3\]]{.nodecor} and [\[B1\]]{.nodecor}–[\[B2\]]{.nodecor} hold. Thus, the discrete optimal control problem – admits at least one solution $(\bar{u}_h,\bar{z}_h) \in \mathbb{V}_h \times \mathbb{Z}_{ad,h}$. In addition, if $(\bar{u}_h,\bar{z}_h)$ denotes an optimal solution for –, then the triple $(\bar u_h, \bar p_h, \bar z_h) \in \mathbb{V}_h \times \mathbb{V}_h \times \mathbb{Z}_{ad,h}$ satisfies $$\begin{aligned} \mathcal{A} ( \bar u_h,v_h) + ( a(\cdot, \bar u_h), v_h )_{L^2(\Omega_h)} & = (\bar z_h, v_h)_{L^2(\Omega_h)} \quad \forall v_h \in \mathbb{V}_h, \label{eq:optimal_state_discrete} \\ \mathcal{A}(v_h,\bar p_h) + \left( \tfrac{\partial a}{\partial u}(\cdot, \bar u_h) \bar p_h, v_h \right)_{L^2(\Omega_h)} & = \left( \tfrac{\partial L}{\partial u}(\cdot,\bar u_h), v_h \right)_{L^2(\Omega_h)} \quad \forall v_h \in \mathbb{V}_h, \label{eq:optimal_adjoint_state_discrete}\end{aligned}$$ and the variational inequality $$(\bar p_h + \alpha \bar z_h, z_h - \bar z_h)_{L^2(\Omega_h)} \geq 0 $$ for every $z_h \in \mathbb{Z}_{ad,h}$. The proof follows from the arguments developed, for the continuous counterpart, in Theorems \[thm:existence\_control\] and \[thm:optimality\_cond\]. For brevity, we skip details. Convergence of discretizations {#sec:convergence} ------------------------------ We begin with the following convergence result: *the sequence $\{ \bar z_h \}_{h>0}$ of global solutions of the discrete control problems – converge, as $h \downarrow 0$, to a solution of the continuous fractional semilinear optimal control problem –.* Let $n \geq 2$ and $s \in (0,1)$. Let $\Omega$ be a Lipschitz domain satisfying the exterior ball condition. Assume that \[A1\]–\[A3\] and \[B1\]–\[B2\] hold. Assume that $a = a(x,u)$ satisfies, in addition, . Assume that $\frac{\partial L}{\partial u}(\cdot,0) \in L^{\infty}(\Omega)$. Let $\bar z_h$, for every $h>0$, be a global solution of the discrete optimal control problem. Then, there exist nonrelabeled subsequences $\{ \bar z_h \}_{h>0}$ such that $\bar z_h \mathrel{\ensurestackMath{\stackon[1pt]{\rightharpoonup}{\scriptstyle\ast}}} \bar{z}$, in $L^{\infty}(\Omega)$, with $\bar z$ being a solution to –. In addition, we have $$\label{eq:convergence} \| \bar z - \bar z_h \|_{L^{2}(\Omega)} \rightarrow 0, \qquad j_{h}( \bar z_h) \rightarrow j(\bar z), $$ as $h \downarrow 0$. \[thm:convergence\] Since $\{ \bar z_h \}_{h>0}$ is uniformly bounded in $L^{\infty}(\Omega)$, we deduce the existence of a nonrelabeled subsequence $\{ \bar z_h \}_{h>0}$ such that $\bar z_h \mathrel{\ensurestackMath{\stackon[1pt]{\rightharpoonup}{\scriptstyle\ast}}} \bar{z}$ in $L^{\infty}(\Omega)$ as $h \downarrow 0$. In what follows we prove that $\bar z$ is locally optimal for the continuous optimal control problem and that $j_h(\bar z_h) \rightarrow j(\bar z)$ as $h \downarrow 0$. Let $\tilde z \in \mathbb{Z}_{ad}$ be locally optimal for – and define $\tilde p$ as the solution to with $u$ replaced by $\tilde u := \mathcal{S} \tilde z$. Define $\tilde z_h \in \mathbb{Z}_{ad,h}$ by $\tilde z_h|_{T} := \int_{T} \tilde z(x) \mathrm{d}x / |T|$ for $T \in {\mathscr{T}}_h$. Now, observe that, since \[A3\] holds and $\tilde p, \partial L/\partial u(\cdot,0) \in L^{\infty}(\Omega)$, we can deduce that $\partial L/\partial u(\cdot, \tilde u) - \partial a/\partial u(\cdot,\tilde u) \tilde p \in L^{\infty}(\Omega)$. In view of the fact that $\Omega$ is Lipschitz and satisfies the exterior ball condition, we can thus invoke [@MR3168912 Proposition 1.1] to obtain that $\tilde p \in C^s(\mathbb{R}^n)$. The projection formula thus yields $\tilde z \in C^s(\overline \Omega)$. Consequently, $\| \tilde z - \tilde z_h \|_{L^{\infty}(\Omega_h)} \rightarrow 0$ as $h \downarrow 0$. Invoke that $\tilde z$ is locally optimal for – and that $\bar z_h$ corresponds to the global solution of the discrete control problem to arrive at $$j(\tilde z) \leq j(\bar z) \leq \liminf_{h \downarrow 0} j_h( \bar z_h ) \leq \limsup_{h \downarrow 0} j_h( \bar z_h ) \leq \limsup_{h \downarrow 0} j_h( \tilde z_h ) = j( \tilde z).$$ To obtain the last equality, we used that $\| \tilde z - \tilde z_h \|_{L^{\infty}(\Omega)} \rightarrow 0$ implies $j_h(\tilde z_h) \rightarrow j(\tilde z)$ as $h \downarrow 0$. We have thus proved that $\bar z$ is locally optimal and $j_h(\bar z_h) \rightarrow j(\bar z)$ as $h \downarrow 0$. We now prove that $\| \bar z - \bar z_h \|_{L^2(\Omega)} \rightarrow 0$ as $h \downarrow 0$. In view of Proposition \[eq:proposition\_convergence\], we have that $\bar u_h \rightarrow \bar u$ in $L^{\mathfrak{q}}(\Omega)$, for $\mathfrak{q} \leq 2n/(n-2s)$, as $h \downarrow 0$. Invoke the local Lipschitz property of $L$ in the second variable to arrive at $$\left| \int_{\Omega_h} \left[ L(x,\bar u(x)) - L(x, \bar u_h(x)) \right] \mathrm{d}x \right| \leq \mathfrak{l}_M \| \bar u - \bar u_h \|_{L^1(\Omega)} \rightarrow 0, \quad h \downarrow 0.$$ Similar arguments yield $ \| L(\cdot, \bar u ) - L (\cdot,0) \|_{L^1(\Omega \setminus \Omega_h)} \leq \mathfrak{l}_M \| \bar u\|_{L^1(\Omega \setminus \Omega_h)} \rightarrow 0$ as $h \downarrow 0$. Consequently, in view of the convergence result $j_h(\bar z_h) \rightarrow j(\bar z)$, we obtain $$\tfrac{\alpha}{2} \| \bar z_h \|^2_{L^2(\Omega)} \rightarrow \tfrac{\alpha}{2} \| \bar z\|^2_{L^2(\Omega)}, \quad h \downarrow 0.$$ This and the weak converge $\bar z_h \rightharpoonup \bar{z}$ in $L^{2}(\Omega)$ imply that $ \bar z_h \rightarrow \bar z$ in $L^2(\Omega)$, as $h \downarrow 0$, and concludes the proof. We now prove a somehow reciprocal result: *every strict local minimum of the continuous problem – can be approximated by local minima of the discrete optimal control problems.* Let $n \geq 2$ and $s \in (0,1)$. Let $\Omega$ be a Lipschitz domain satisfying the exterior ball condition. Assume that \[A1\]–\[A3\] and \[B1\]–\[B2\] hold. Assume that $a = a(x,u)$ satisfies, in addition, . Assume that $\frac{\partial L}{\partial u}(\cdot,0) \in L^{\infty}(\Omega)$. Let $\bar z$ be a strict local minimum of problem –. Then, there exists a sequence $\{ \bar z_h \}_{h>0}$ of local minima of the discrete problems such that $$\label{eq:convergence_local_minima} \| \bar z - \bar z_h \|_{L^{2}(\Omega)} \rightarrow 0, \qquad j_{h}( \bar z_h) \rightarrow j(\bar z), $$ as $h \downarrow 0$. \[thm:convergence\_local\_minima\] Since $\bar z$ is a strict local minimum for problem –, we deduce the existence of $\epsilon >0$ such that the minimization problem $$\min \{ j(z): z \in \mathbb{Z}_{ad}~\mathrm{and}~\| \bar z - z \|_{L^2(\Omega)} \leq \epsilon \} \label{eq:local_continuous_problem}$$ admits a unique solution $\bar z \in \mathbb{Z}_{ad}$. On the other hand, let us consider, for $h>0$, the discrete problem $$\min \{ j_h(z_h): z_h \in \mathbb{Z}_{ad,h}~\mathrm{and}~\| \bar z - z_h \|_{L^2(\Omega)} \leq \epsilon \}. \label{eq:local_discrete_problem}$$ We extend discrete functions $z_h \in \mathbb{Z}_{ad,h}$, defined over $\Omega_h$, to $\Omega$ by setting $z_h(x) = \bar z(x)$ for $x \in \Omega\setminus \Omega_h$. To conclude that problem admits at least a solution, we need to verify that the set where the minimum is sought is nonempty; notice that such a set is compact. To accomplish this task, we define, as in the proof of Theorem \[thm:convergence\], $\hat z_h \in \mathbb{Z}_{ad,h}$ by $\hat z_h |_T := \int_{T} \bar z(x) \mathrm{d}x / |T|$ for $T \in {\mathscr{T}}_h$. Observe that, on $\Omega \setminus \Omega_h$, $\hat z_h = \bar z$. Since $\bar z \in C^s(\Omega)$, we have that $\| \bar z - \hat z_h \|_{L^{\infty}(\Omega)} \rightarrow 0$ as $h \downarrow 0$. As a result, if $h$ is sufficiently small, $\hat z_h \in \mathbb{Z}_{ad,h}$ is such that $\| \bar z - \hat z_h \|_{L^2(\Omega)} \leq \epsilon$. We can thus conclude the existence of $h_{\star} > 0$ such that problem admits at least a solution for $h \leq h_{\star}$. Let $h \leq h_{\star}$ and let $\bar z_h$ be a solution to problem . Since $\{ \bar z_h \}_{0 < h \leq h_{\star}}$ is bounded in $L^{\infty}(\Omega)$, there exist a subsequence $\{ \bar z_{h_k} \}_{k= 1}^{\infty}$ of $\{ \bar z_h \}_{0 < h \leq h_{\star}}$ such that $\bar z_{h_k} \mathrel{\ensurestackMath{\stackon[1pt]{\rightharpoonup}{\scriptstyle\ast}}} \tilde{z}$ in $L^{\infty}(\Omega)$ as $k \uparrow \infty$. Proceed as in the proof of Theorem \[thm:convergence\] to obtain that $\tilde z$ is a solution to the continuous problem and $\bar z_{h_k} \rightarrow \tilde z$ in $L^2(\Omega)$ as $k \uparrow \infty$. Since problem admits a unique solution, we must have $\tilde z = \bar z$ and $\bar z_h \rightarrow \bar z$ in $L^2(\Omega)$ as $h \downarrow 0$. Observe that, for $h$ sufficiently small, the constraint $\| \bar z - \bar z_h \|_{L^2(\Omega)} \leq \epsilon$ is not active in problem . Consequently, $\bar z_h$ solves the original discrete problem. We conclude by mentioning that the arguments elaborated in the proof of Theorem \[thm:convergence\] yield . Error estimates {#sec:error_estimates} --------------- Let $\{ \bar z_h \} \in \mathbb{Z}_{ad,h}$ be a sequence of local minima of the discrete optimal control problems such that $\| \bar z - \bar z_h \|_{L^{2}(\Omega_h)} \rightarrow 0$ as $h \downarrow 0$; $\bar z$ being a local solution of the continuous problem –; see Theorems \[thm:convergence\] and \[thm:convergence\_local\_minima\] . The main goal of this section is to provide an error estimate for $\bar z - \bar z_h$ in $L^2(\Omega)$, namely $$\label{eq:error_estimate_in L2_aux} \| \bar z - \bar z_{h} \|_{L^2(\Omega_h)}\lesssim h^{\gamma}, \quad \gamma = \min \left\{ 1, s + \tfrac{1}{2} - \epsilon \right\}, \quad \forall h \leq h_{\star}.$$ Here, $\epsilon >0$ is arbitrarily small. In what follows, if necessary, we extend discrete functions $z_h \in \mathbb{Z}_{ad,h}$, defined over $\Omega_h$, to $\Omega$ by setting $z_h(x) = \bar z(x)$ for $x \in \Omega\setminus \Omega_h$. We begin with the following instrumental result. Let $n \in \{ 2,3 \}$ and $s > n/4$. Let $\Omega$ be a convex domain such that $\partial \Omega \in C^{\infty}$. Assume that \[A1\]–\[A3\], \[B1\]–\[B2\], and \[C1\]–\[C2\] hold. Let $\bar z \in \mathbb{Z}_{ad}$ satisfies the second order optimality condition , or equivalently . Let as assume that is false. Then, there exists a positive constants $h_{\star}$ such that $$\label{eq:basic_estimate} \tfrac{\mathfrak{C}}{2}\| \bar z - \bar z_{h} \|^2_{L^2(\Omega_h)} \leq \left[ j'(\bar z_h) - j'(\bar z) \right](\bar z_h - \bar z)$$ for every $h \leq h_{\star}$, where $\mathfrak{C} = \min \{ \mu, \alpha \}$, $\mu$ is the constant appearing in , and $\alpha$ denotes the regularization parameter. \[thm:instrumental\_error\_estiamate\] Since is false, there exists a sequence $\{ h_{\ell} \}_{\ell=1}^{\infty}$ such that $h_{\ell} \downarrow 0$ as $\ell \uparrow \infty$ and $\{ \bar z_{h_{\ell}} \}_{\ell=1}^{\infty}$ satisfies $\| \bar z - \bar z_{h_{\ell}} \|_{L^2(\Omega_h)}/h_{\ell}^{\gamma} \rightarrow \infty$ as $h_{\ell} \downarrow 0$. In what follows, to simplify notation we omit the subindex $\ell$. Invoke the mean value theorem to obtain $$\left( j'(\bar z_h) - j'(\bar z) \right) (\bar z_h - \bar z) = j''( \bar z + \theta_h ( \bar z_h - \bar z ) ) (\bar z_h - \bar z)^2, \quad \theta_h \in (0,1). \label{eq:aux_second_order_2}$$ Define $v_h:= (\bar z_{h} - \bar z)/ \| \bar z_h - \bar z \|_{L^2(\Omega)}$. Observe that, for every $h >0$, we have $\| v_h \|_{L^2(\Omega)} = 1$. Upon considering a subsequence, if necessary, we can assume that $v_h \rightharpoonup v$ in $L^2(\Omega)$. Since the set of elements satisfying is weakly closed in $L^2(\Omega)$ and each $v_{h}$ satisfies , we conclude that $v$ satisfies as well. We now prove that $|\mathfrak{p}(x)| > 0$ implies $v(x) = 0$; recall that $\mathfrak{p} = \bar p + \alpha \bar z$. Define $\bar{\mathfrak{p}}_h:= \bar p_h + \alpha \bar z_h$. Observe that $\| \mathfrak{p} - \mathfrak{p}_h \|_{L^2(\Omega)} \rightarrow 0$ as $h \downarrow 0$. As a result, we obtain $$\begin{gathered} \int_{\Omega} \mathfrak{p}(x) v(x) \mathrm{d}x = \lim_{h \rightarrow 0} \int_{\Omega_{h}} \mathfrak{p}_{h}(x) v_{h}(x) \mathrm{d}x \\ = \lim_{h \rightarrow 0} \frac{1}{ \| \bar z_{h} - \bar z \|_{L^2(\Omega)} } \left[ \int_{\Omega_{h}} \mathfrak{p}_{h}(x) [ (\Pi_{h} \bar z(x) - \bar z(x)) + (\bar z_{h}(x) - \Pi_{h} \bar z(x)) ] \mathrm{d}x \right],\end{gathered}$$ where $\Pi_h: L^2(\Omega) \rightarrow \mathbb{Z}_h$ denotes the orthogonal projection operator onto piecewise constant functions over ${\mathscr{T}}_h$. Since $0 \leq j_{h}'( \bar z_{h}) ( \Pi_{h} \bar z - \bar z_{h} ) = \int_{\Omega_h} \mathfrak{p}_{h}(x) (\Pi_{h} \bar z(x) - \bar z_{h}(x)) \mathrm{d}x$, because $\Pi_{h} \bar z \in \mathbb{Z}_{ad,h}$, we have $$\int_{\Omega} \mathfrak{p}(x) v(x) \mathrm{d}x \leq \lim_{h \rightarrow 0} \frac{1}{ \| \bar z_{h} - \bar z \|_{L^2(\Omega)} } \left[ \int_{\Omega_{h}} \mathfrak{p}_{h}(x)(\Pi_{h} \bar z(x) - \bar z(x)) \mathrm{d}x \right].$$ Invoke the regularity results for $\bar z$ obtained in Theorem \[thm:regularity\_space\], namely $\bar z \in H^{\gamma}(\Omega)$, where $\gamma = \min \{s+1/2-\epsilon,1\}$ and $\epsilon >0$ is arbitrarily small, standard error estimates for $\Pi_h$, and $\lim_{h \rightarrow 0} \| \bar z - \bar z_{h} \|_{L^2(\Omega_h)}/ h^{\gamma} = \infty$ to obtain $ \int_{\Omega} \mathfrak{p}(x) v(x) \mathrm{d}x \leq 0. $ In view of , we can thus conclude that $\int_{\Omega} | \mathfrak{p}(x) v(x)| \mathrm{d}x = 0$ and thus that $|\mathfrak{p}(x)| > 0$ implies $v(x) = 0$ for a.e $x \in \Omega$. Consequently, $v \in C_{\bar z}$. Define $\hat{z}_h := \bar z + \theta_h ( \bar z_h - \bar z )$, $\hat{u}_h:= \mathcal{S} \hat{z}_h$, and $\hat{p}_h$ as the solution to with $u$ replaced by $\hat{u}_h$. Invoke to obtain $$\begin{aligned} \lim_{h \rightarrow 0} j''(\hat z_h) v_{h}^2 & = \lim_{h \rightarrow 0} \int_{\Omega} \left( \frac{\partial^2 L}{\partial u^2}(x, \hat u_{h}) \phi_{v_h}^2 - \hat p_h \frac{\partial^2 a}{\partial u^2}(x,\hat u_{h})\phi_{v_h}^2 + \alpha v_{h}^2 \right) \mathrm{d}x \\ & = \alpha + \int_{\Omega} \left( \frac{\partial^2 L}{\partial u^2}(x, \bar u) \phi_{v}^2 - \bar p \frac{\partial^2 a}{\partial u^2}(x,\bar u)\phi_{v}^2 \right) \mathrm{d}x, \end{aligned}$$ where we have used $ \hat u_{h} \rightarrow \bar u $ and $ \hat p_{h} \rightarrow \bar p $ in $\tilde H^s(\Omega) \cap L^{\infty}(\Omega)$ and $\phi_{v_h} \rightharpoonup \phi_{v}$ in $\tilde H^s(\Omega)$; the latter implies that $\phi_{v_h} \rightarrow \phi_{v}$ in $L^{\mathfrak{q}}(\Omega)$ as $k \uparrow \infty$ for $\mathfrak{q} < 2n/(n-2s)$; see the proof of Theorem \[thm:suff\_opt\_cond\] for details. Invoke that $v \in C_{\bar z}^{\tau}$ and $\bar z$ satisfies to obtain $$\lim_{h \rightarrow 0} j''(\hat z_h) v_{h}^2 = \alpha + j''(\bar z)v^2 - \alpha \| v \|^2_{L^2(\Omega)} \geq \alpha + (\mu - \alpha) \| v \|^2_{L^2(\Omega)}.$$ Since $\| v \|_{L^2(\Omega)} \leq 1$, we can thus conclude that $\lim_{h \rightarrow 0} j''(\hat z_h) v_{h}^2 \geq \mathfrak{C}$, where $\mathfrak{C} = \min \{ \mu, \alpha \}$. As a result, there exists $h_{\star} >0$ such that, for every $h \leq h_{\star}$, we have $ j''(\hat z_h) v_{h}^2 \geq \mathfrak{C}/2. $ In view of , we can finally derive and conclude the proof. We now provide an error estimate for the difference $\bar z - \bar z_h$ in $L^2(\Omega)$. Let $n \in \{ 2,3 \}$ and $s > n/4$. Let $\Omega$ be a convex domain such that $\partial \Omega \in C^{\infty}$. Assume that \[A1\]–\[A3\], \[B1\]–\[B2\], and \[C1\]–\[C2\] hold. Let $\bar z \in \mathbb{Z}_{ad}$ satisfies the second order optimality condition , or equivalently . Then, there exist $h_{\star} >0$ such that $$\label{eq:error_estimate_in L2} \| \bar z - \bar z_{h} \|_{L^2(\Omega_h)}\lesssim h^{\gamma}, \quad \gamma = \min \{ 1, s + 1/2 - \epsilon \}, \quad \forall h \leq h_{\star},$$ where $\epsilon >0$ is arbitrarily small. We proceed by contradiction. Let us assume that is false so that we have at hand the instrumental estimate of Theorem \[thm:instrumental\_error\_estiamate\]. We begin by observing that $j_h'(\bar z_h)(z_h - \bar z_h) \geq 0$ for every $z_h \in \mathbb{Z}_{ad,h}$ and $j'(\bar z)(\bar z_h - \bar z) \geq 0$. In view of these inequalities, we invoke to obtain $$\tfrac{\mathfrak{C}}{2}\| \bar z - \bar z_h \|_{L^2(\Omega_h)}^2 \leq [j_h'(\bar z_h) -j'(\bar z_h)](z_h - \bar z_h) + j'(\bar z_h)( z_h - \bar z) \label{eq:first_step}$$ for every $z_h \in \mathbb{Z}_{ad,h}$. Let $\Pi_h: L^2(\Omega) \rightarrow \mathbb{Z}_h$ be the orthogonal projection operator onto piecewise constant functions over ${\mathscr{T}}_h$. Set $z_h = \Pi_h \bar z \in \mathbb{Z}_{ad,h}$ in to obtain $$\tfrac{\mathfrak{C}}{2}\| \bar z - \bar z_h \|_{L^2(\Omega_h)}^2 \leq [j_h'(\bar z_h) -j'(\bar z_h)]( \Pi_h \bar z - \bar z_h) + j'(\bar z_h)( \Pi_h \bar z - \bar z) =: \mathrm{I} + \mathrm{II}.$$ We bound the term $\mathrm{II}$ as follows. First, standard properties of $\Pi_h$ reveal that $$\begin{gathered} \mathrm{II}_{\Omega_h} := (p(\bar z_h) + \alpha \bar z_h, \Pi_h \bar z - \bar z)_{L^2(\Omega_h)} = ( p(\bar z_h), \Pi_h \bar z - \bar z)_{L^2(\Omega_h)} \\ = (p(\bar z_h) - \Pi_h p(\bar z_h), \Pi_h \bar z - \bar z)_{L^2(\Omega_h)} \lesssim h^{\gamma + \vartheta + \frac{1}{2}-\epsilon} |\bar p(\bar z_h)|_{H^{s+\frac{1}{2}-\epsilon}(\Omega)} | \bar z |_{H^{\gamma}(\Omega)},\end{gathered}$$ where $\vartheta = \min \{s,1/2-\epsilon \}$, $\gamma = \min \{ 1, s + 1/2 - \epsilon \}$, $\epsilon >0$ is arbitrarily small, and $p(\bar z_h)$ denotes the solution to with $u$ replaced by $\mathcal{S} \bar z_h$. Theorem \[thm:regularity\_space\] guarantees that $\bar z \in H^{\gamma}(\Omega)$ so that the term $| \bar z |_{H^{\gamma}(\Omega)}$ is uniformly bounded. On the other, Proposition \[pro:state\_regularity\_smooth\] reveals that $p(\bar z_h) \in H^{s+1/2-\epsilon}(\Omega)$, where $\epsilon >0$ is arbitrarily small. The remaining term $\mathrm{II}_{\Omega\setminus \Omega_h}$ vanishes: $$| \mathrm{II}_{\Omega\setminus \Omega_h}| = |(p(\bar z_h) + \alpha \bar z_h, \Pi_h \bar z -\bar z )_{L^2(\Omega\setminus \Omega_h)}| = 0. $$ We now control $\mathrm{I}$. To accomplish this task, we first observe that $ \mathrm{I} = (\bar p_h - p(\bar z_h) ), \Pi_h \bar z - \bar z_h)_{L^2(\Omega)}. $ Second, we split $\mathrm{I} = \mathrm{I}_{\Omega_h} + \mathrm{I}_{\Omega \setminus \Omega_h}$ and control $\mathrm{I}_{\Omega_h}$ as follows: $$\begin{gathered} \mathrm{I}_{\Omega_h} := (\bar p_h - p(\bar z_h) , \Pi_h \bar z - \bar z_h)_{L^2(\Omega_h)} = (\bar p_h - p(\bar z_h), \Pi_h (\bar z - \bar z_h) )_{L^2(\Omega_h)} \\ \lesssim \| \bar p_h - p(\bar z_h) \|_{L^2(\Omega)} \| \bar z - \bar z_h\|_{L^2(\Omega)} \leq \frac{\mathfrak{C}}{4} \| \bar z - \bar z_h\|^2_{L^2(\Omega)} +C h^{\vartheta + 1/2 - \epsilon} | p(\bar z_h) |_{H^{s+\frac{1}{2}-\epsilon}(\Omega)},\end{gathered}$$ where $\vartheta = \min \{ s, 1/2 - \epsilon \}$, $\epsilon >0$ is arbitrarily small and $C>0$. Since discrete functions $z_h \in \mathbb{Z}_{ad,h}$ are extended to $\Omega$ upon setting $z_h(x) = \bar z(x)$ in $\Omega \setminus \Omega_h$, $\mathrm{I}_{\Omega \setminus \Omega_h} = 0$. A collection of the derived estimates yields the bound $\| \bar z - \bar z_h \|_{L^2(\Omega)} \lesssim h^{\gamma}$, which is a contradiction. This concludes the proof. [^1]: Departamento de Matemática, Universidad Técnica Federico Santa María, Valparaíso, Chile ([[email protected]]{}). [^2]: Submitted to the editors .

ArXiv

--- abstract: 'In this paper we present Gelisp, a new library to represent musical Constraint Satisfaction Problems and search strategies intuitively. Gelisp has two interfaces, a command-line one for Common Lisp and a graphical one for OpenMusic. Using Gelisp, we solved a problem of automatic music generation proposed by composer Michael Jarrell and we found solutions for the All-interval series.' author: - Mauricio Toro - Camilo Rueda - Carlos Agón - | \ Gérard Assayag bibliography: - 'gelisp.bib' title: 'Gelisp: A Library to Represent Musical CSPs and Search Strategies' --- Introduction ============ A *Constraint Satisfaction Problem* (CSP) is a formalism to represent combinatorial problems. To solve a CSP we need to find objects that satisfy a number of constraints (i.e., criteria over those variables). CSPs provide a declarative way to represent combinatorial problems, specifying cons-traints instead of a sequence of steps to find the solution (as used in imperative programming). Additionally, it is possible to specify strategies to choo-se between branches during search. CSPs in computer music can be used to solve harmonic, rhythmic or melodic problems. In addition, they can be used for automatic generation of musical structures satisfying a set of rules. For instance, we can find solutions for the *All-interval series* [@allinterval], where we need to find a sequence of 12 different pitches with 12 different intervals. In order to solve a CSP, we can use constraint programming languages such as Prolog or Mozart-Oz [@VanRoyHaridi:2004]. In order to solve a CSP, those languages use a *Constraint Solving Library (CSL)* such as Gecode [@fastprop]. CLSs are usually written in C++. The problem ----------- Using traditional CSL’s or programming languages to solve CSPs is time-demanding and it is intended for specialized users because they usually require deep knowledge on C++ or logic programming. This makes these tools often unpractical to specify musical CSPs. Furthermore, these tools do not provide a representation for musical data structures. Our solution ------------ *Gelisp*[^1] is a wrapper for Gecode to Common Lisp. *Gelisp* was originally developed by Rueda in 2006 and we modified it to work with current version of Gecode. Furthermore, we added support to model CSPs and search strategies graphically on *OpenMusic (OM)* [@OM98]. In addition, *Gelisp* can take advantage of the musical data structures and functions defined for OM. The novelty of *Gelisp* is to provide a graphical representation for search strategies (e.g., Depth\ First Search) and global constraints (e.g., “all the intervals of a sequence must be different”), based on an efficient CSL. Related Work ------------ Several graphical CSLs for OM have been developed in the last decade. Situation [@situation] generates music based on constraints, OmRc [@omrc] finds structures corresponding to rhythmical constraints, OmClouds [@omclouds] finds approximated solutions to a CSP, and OMBacktrack (<http://www.ircam.fr/>\ [equipes/repmus/](equipes/repmus/)) is a wrapper for the CSL Screamer [@screamer] (a CSL written on Lisp). A graphical CSL to solve musical CSPs should be able to setup search strategies in a graphical way, post multiple kinds of constraints graphically without declaring explicitly loops and recursion, and solve the problem using state-of-art algorithms. Unfortunately, OmRC and OmSituation are designed to solve specific problems. OmBacktrack is no longer available for current versions of OM. Finally, OmClouds does not guarantee a solution satisfying all the constraints (i.e., a complete solution). Gecode ====== Gecode is a Constraint Solving Library (CSL) written in C++. Gecode provides a *propagator* for each type of constraint. Propagators translate a constraint into basic constraints supplying the same information. Basic (finite domain) constraints have the form $x \in [a..b]$. For instance, in a store (i.e., a set with all the constraints asserted) containing [$pitch_1 \in [36..72]$ and $pitch_2 \in [60..80]$]{}, a propagator for the constraint $pitch_1 > pitch_2 + 2$ would add constraints $pitch_1 \in [63..72]$ and $pitch_2 \in [60..69]$. As described in the above example, the action of propagators ends up narrowing down the set of possible values for each variable. This, however, does not guarantee that it will eventually be inferred a single value for each variable. Gecode thus include *search engines*. The purpose of a search engine is to choose additional basic constraints to add into the store until all variables have reduced their domain to a single value. Using them we can find one, many, or all the solutions for a CSP. Gecode works on different operating systems and it will be used as the CSL for Mozart-Oz, therefore it is very likely to be maintained for a long time. Furthermore, it provides an extensible API, allowing the user to create new propagators and user-defined search engines. For instance, we can extend Gecode to reason about trees and graphs, which are useful in musical CSPs. Gelisp ====== *Gelisp* provides an interface for Common Lisp and another for OM. In *Gelisp*, sequences of variables are represented by lists, as opposed to Gecode, where they are represented by arrays. This makes the power of list processing (provided by Lisp and OM) available for *Gelisp* users. Interface for Common Lisp ------------------------- To solve a problem using this interface, we need to write a script. A script is a function to define the problem variables and their domains (the possible values that a variable can take), post constraints over the variables, and setup a search strategy. This interface allows the user to call most of Gecode propagators for both, Finite Domain (FD) and Finite Set (FS) constraints. Basic FD constraints deal with expressions of the form $x \in R$, where $R$ is a range or a set of ranges of integers. On the other hand, FS constraints deal with expressions among sets of FD variables. In what follows, we present some propagators that *Gelisp* provides for FD and FS. *Gelisp* provides FD propagators for defining domains (e.g., $Domain(X) = [2,5]$), equalities and inequalities (e.g., $ X + Y < Z$), cardinality (e.g., 1 occurs two times in $[X Y Z]$), boolean constraints, regular expression constraints and the all-distinct constraint. The all-distinct constraint makes the elements of a sequence pairwise different. On the other hand, for FS we provide constraints for defining domains (e.g., $V \subseteq{ \{1, 2, 3 \} }$) and set relations (e.g., $X \subset{ A \cup B }$). In addition, *Gelisp* includes two search engines, Depth Search First (DSF) and Branch-and-bound (BAB). The DFS engine works by choosing some variable, then a value for that variable, if this does not succeed (a constraint does not hold) then choo-ses another value. If the value succeed, then choo-ses another variable, then a value for it, etc. The BAB engine works in a similar way, but solutions are computed in such a way that each subsequent solution increases or decreases the value of some user specified FD variable. Both engines can be used for both FS and FD. In addition, we can define search heuristics for value (i.e., the order to assign a value to a variable) and variable order (i.e., the order to choose a variable). These heuristics are parameters for the search engines. Graphical Interface for OpenMusic --------------------------------- Instead of writing a script, in the graphical interface we represent a program with a special patch, called *CSP patch*. A patch is a visual algorithm, in which boxes represent functional calls, and connections are functional compositions. Inside a *CSP patch*, we can place special boxes to define a constraint in the CSP, variable and value heuristics, the variable to be optimized during the search, and a time limit in the search. For instance, we provide a variety of boxes to represent simple constraints (e.g., $a = 2$) and global constraints (e.g., “all the intervals from a sequence must be different”). Using the graphical interface we can express a variety of problems declaratively with global constraints. Global constraints have parameters. For instance, the graphical box to find the intervals of a list “$x \rightarrow dx$” has a parameter to choose among absolute, non-absolute, or modulo $n$ intervals (calculated as $(V_{i+1} - V_i) \% n$). Additionally, it has a parameter to post an *all-distinct* constraint over the intervals. Moreover, the output of a CSP patch can be connected to a box to find one solution or a box to find $n$ the solutions Applications ============ In this section, we describe both, an intuitive and formal definition of two CSPs and we explain how to solve them with *Gelisp*. Formally, a CSP is triple $<X,D,C>$, where $X$ is a set of variables, $D$ is the domain for each variable, and C is a set of constraints (read as conjunction) over the variables. All-interval series ------------------- In this problem, we need to find a sequence of 12 different pitches with 12 different intervals (fig. \[fig:allintervals-ex\]). This problem can be generalized to find $n$ different pitches with $n$ different intervals equivalent under inversion [^2]. For instance, a value of $n=24$ represents the *all-interval series* for microtones. Therefore, a solution to this CSP is a sequence of $n$ pairwise different variables with domain $[1..n]$, where all modulo $n$ intervals of the sequence are pairwise different. We give bellow a formalization of this problem\ \ **Variables**: $V_1$ ... $V_n$\ **Domains**: $[1..n]$ ... $[1..n]$\ **Constraints**: - $C_1$ alldiff$(V)$ - $C_2$ alldiff$( (V_{i+1} - V_i) \% n, i \leq n)$ There is not a constraint over the interval $(V_{n} - V_0$) because that interval is always six, according to the literature. Furthermore, it is enough to calculate the series where $V_0 = 0$ because the other ones can be obtained from that one using transposition. In addition, we know that if $V_1..V_n$ is an all-interval serie, $V_n...V_1$ is also one. For those reasons, we include these two constraints to avoid symmetrical solutions: - $C_3$ $V_0 = 0$ - $C_4$ $V_0 < V_n$ We represent graphically this CSP (fig. \[fig:allintervals\]) with a box to create $n$ all-different variables with domain $[1..n]$, an $x \rightarrow dx$ box for $C_2$ with an all-different parameter , an *equality* box for $C_3$, and an *inequality* box for $C_4$. Jarrell CSP ----------- Composer Michael Jarrell proposed an idea for automatic music generation [@jarrell]. The goal is to generate a sequence of $n$ notes. There is a fix number of occurrences $OM_{1}$...$OM_{A}$ for each sequences of intervals (called motives) $M_{1}$...$M_{A}$ over the sequence of non-absolute intervals of the output sequence. In addition, each note of the output sequence belongs to a Chord $Ch$. Moreover, the first $L_1$ and the last note $L_2$ of the output sequence are fixed. We give bellow a formalization of this problem\ \ **Inputs**: - Motives \[$M_{1}$...$M_{A}$\], Limits $L_{1}$ and $L_{2}$,\ Occurrences \[$OM_{1}$...$OM_{A}$\], Chord $Ch$ **Variables**: $V_1$ ... $V_n$\ **Domains**: $[0..127]$ ... $[0..127]$\ **Constraints**: - $C_1$ $\forall_{1 < i < A}$ $|\{j, M_{i}$ is a subsequence of the variables’ intervals that starts on $j\}| = OM_{i}$ - $C_2$ $\forall_{1 < i < n}$ $V_i \in Ch$ - $C_3$ $V_1 = L_1 \wedge V_n = L_2$ We represent graphically (fig.\[fig:jarrell\]) the constraint $C_1$. We use the *x$\rightarrow$dx* and *motives-occurs=* boxes to fix the number of occurrences of each motive over the intervals of the output sequence. Jarrell also proposes in [@jarrell] to consider absolute intervals and octaviation for the chords, the limits and the motives. For instance, using absolute intervals, an interval $V_{i+1} - V_i$ is equal to $V_{i} - V_{i+1}$ and using octaviation, a pitch G4 is equivalent to G1,G2,G5, etc. Finally, he also proposes to have specific motives and chords for each segment of the output sequence, according to a user-defined segmentation. For simplicity, we do not present those constraints in this paper. However, a complete model of this problem can be found at *Gelisp* website. Concluding Remarks and Future Work ================================== We presented a library for Common Lisp and OM providing a variety of constraints and search engines. *Gelisp* provides graphical boxes to represent some constraints and search strategies. *Gelisp* abstracts minor details that are not necessary for musicians and mathematicians. It would be pretentious to conclude that we can easily model any musical CSP using *Gelisp* graphical interface, or using the command-line interface. However, we can model a variety of problems using *Gelisp* in a simple way taking advantage of the state-of-the-art propagators and search engines provided by Gecode. An approach related to CSPs is concurrent constraint programming, a family of process calculi often used to model musical interactions problems. Process calculi has been applied to the modeling of interactive music systems [@is-chapter; @tdcr14; @ntccrt; @cc-chapter; @torophd; @torobsc; @Toro-Bermudez10; @Toro15; @ArandaAOPRTV09; @tdcc12; @toro-report09; @tdc10; @tdcb10; @tororeport] and ecological systems [@PT13; @TPSK14; @PTA13; @mean-field-techreport]. In future works, we will explore a bigger sample of musical CSPs and their representation using global constraints. In addition, the idea of representing CSPs and their search strategies with business rules from *Rules2Cp* [@rules4cp] can be extended to generate a musical CSP based on musical rules. ACKNOWLEDGMENTS =============== Thanks to Moreno Andreatta, Jean Bresson, Serge Lemouton, Killian Sprotte, and Guido Tack for their valuable comments when developing *Gelisp*. Thanks to Carlos Toro and Jorge Pérez for their remarks on this paper. [^1]: http://gelisp.sourceforge.net/ [^2]: For instance, an interval C-E is equivalent to E-C.

ArXiv

--- abstract: 'We investigate the neutron’s response to magnetic fields on a torus with the aid of chiral perturbation theory, and expose effects from non-vanishing holonomies. The determination of such effects necessitates non-perturbative treatment of the magnetic field; and, to this end, a strong-field power counting is employed. Using a novel coordinate-space method, we find the neutron propagates in a coordinate-dependent effective potential that we obtain by integrating out charged pions winding around the torus. Knowledge of these finite volume effects will aid in the extraction of neutron properties from lattice QCD computations in external magnetic fields. In particular, we obtain finite volume corrections to the neutron magnetic moment and magnetic polarizability. These quantities have not been computed correctly in the literature. In addition to effects from non-vanishing holonomies, finite volume corrections depend on the magnetic flux quantum through an Aharonov-Bohm effect. We make a number of observations that demonstrate the importance of non-perturbative effects from strong magnetic fields currently employed in lattice QCD calculations. These observations concern neutron physics in both finite and infinite volume.' author: - 'Brian C. Tiburzi' bibliography: - 'magtron.bib' title: 'Neutron in a Strong Magnetic Field: Finite Volume Effects' --- ł Ł ø Ø § \#1 \#1[[(\[\#1\])]{}]{} Introduction ============ Electromagnetic interactions of quarks are simple at the level of the action, however, the confining dynamics of quarks in QCD lead to electromagnetic properties of non-perturbative bound-state hadrons. While these properties can be parameterized in terms of a few low-energy hadronic parameters, determining their values from QCD provides a window to hadron structure in terms of quark degrees of freedom. Additionally the QCD response to external electromagnetic fields provides a lever-arm with which to study the modification of hadronic structure. For example, the energy levels of a neutron in a weak magnetic field depend on the neutron’s magnetic moment and magnetic polarizability. The former hadronic parameter provides a glimpse at the current distribution within the neutron, while the latter parameter encompasses the response of this current distribution to an external magnetic field. Lattice QCD computations in classical electromagnetic fields constitute a fruitful method to determine hadronic properties, especially polarizabilities, see [@Martinelli:1982cb; @Bernard:1982yu; @Fiebig:1988en; @Lee:2005ds; @Christensen:2004ca; @Lee:2005dq; @Shintani:2006xr; @Engelhardt:2007ub; @Shintani:2008nt; @Aubin:2008qp; @Detmold:2009dx; @Alexandru:2009id; @Detmold:2010ts; @Alexandru:2010dx; @Freeman:2012cy; @Primer:2013pva; @Lujan:2013qua; @Lujan:2014kia]. Lattice gauge theory techniques allow one to determine the neutron magnetic moment and magnetic polarizability by studying neutron correlation functions in external magnetic fields. Because the notion of a magnetic moment, or a dipole polarizability, is intimately linked with rotational invariance, lattice calculations of these quantities are subject to non-trivial finite volume effects. Indeed the imposition of a uniform magnetic field on a torus is already constrained by the finite volume [@'tHooft:1979uj; @Smit:1986fn; @Damgaard:1988hh]. The magnetic flux quantization for a torus leads to two notable effects. First is the restriction to magnetic fields that are not considerably small compared to hadronic scales. As a result, lattice computations are often probing hadronic response to strong magnetic fields. Second is the accompanying gauge holonomy, which is an artifact of the finite volume. Both of these effects complicate external field spectroscopy.[^1] To address rigorously the response of the neutron in a strong magnetic field, we use chiral perturbation theory and combine three crucial ingredients for the first time. We use the heavy-nucleon chiral expansion to address nucleon-pion dynamics. Secondly the magnetic field is treated non-perturbatively by summing charged pion couplings to the external field with the aid of Schwinger’s proper-time trick [@Schwinger:1951nm]. In this way, effects of charged pion Landau levels are included in our computation. Thirdly the computation is performed in finite volume, and properly includes topological effects from charged pions winding around the torus, and the flux quantization for magnetic fields.[^2] A few highlights of the various results obtained are as follows: - We obtain the infinite volume spin-independent energy shift, and Zeeman splitting of the neutron in strong magnetic fields, Eqs.  and . These results are required to address the breakdown of perturbative magnetic field expansions, which is necessitated by the size of current-day lattices. - We obtain expressions for finite volume corrections to the neutron magnetic moment and magnetic polarizability by considering what we deem to be extreme limits, Eqs.  and . These results differ from those in the literature, however, those calculations are known to be inconsistent by keeping the lattice size $L$ finite in one part of the calculation, and $L = \infty$ in another. We caution lattice practitioners from using such formulae due to the extremes required for their applicability. - We obtain finite volume corrections to the neutron magnetic moment and polarizability that can be used in practice for external field computations in lattice QCD, Eqs.  and . The simplicity of these results reflects that charged pion winding transverse to the magnetic field is suppressed due to Landau level confinement. At finite volume, we show non-perturbative magnetic field effects also remain relevant on current-day lattices. The organization of our presentation is as follows. We begin by reviewing the treatment of uniform magnetic fields on a torus in Sec. \[s:Gen\]. Included in this treatment is the solution to the magnetic periodic Green’s function for charged scalars, and that for a charged heavy fermion in the static approximation. The one-loop computation of the neutron effective action in a magnetic field is pursued in Sec. \[s:One\]. The computation uses heavy nucleon chiral perturbation theory with strong-field power counting in the so-called $p$-regime, and details pertinent to the computation are reviewed. We employ a direct coordinate-space approach that is first exemplified by computing finite volume corrections to the neutron mass. By extending the computation to the neutron two-point function in an external magnetic field, expressions are obtained for the spin-dependent and spin-independent terms in the neutron effective potential. The effective potential is coordinate dependent, which directly reflects the non-vanishing holonomies of the gauge field. Perturbative expansions in powers of the magnetic field are contrasted with non-perturbative results obtained in strong-field chiral perturbation theory. The final section, Sec. \[s:Disc\], concerns assessing the effect of finite volume. In this section, we evaluate corrections to the neutron magnetic moment and magnetic polarizability. While the effects are complicated by the coordinate dependence of the effective potential, we find that such corrections to the ground-state neutron energy are suppressed. At the end of this section, we conclude with a few directions for future work. Various technical details have been relegated to appendices. In Appendix \[s:D\], we detail the inclusion of delta-pion intermediate states in our calculation. The main text discusses only proton-pion intermediate states for simplicity. One-dimensional integral formulae for finite volume effects are collected in Appendix \[s:A\]. A final appendix, Appendix \[s:B\], demonstrates how finite volume corrections to the magnetic moment are related between two different extreme limits. Generalities {#s:Gen} ============ To begin, we discuss the inclusion of a uniform magnetic field on a torus, and the implications for charged particle two-point correlation functions. Such generalities will be required to pursue chiral perturbation theory computations at one-loop order for the neutron. We consider four-dimensional Euclidean spacetime, with three directions compactified to form a torus. The length of each compact direction is taken to be $L$. The fourth direction will correspond to Euclidean time and is non-compact. Throughout we use Greek indices for quantities having four components, $x_\mu = (x_1, x_2, x_3, x_4)$, and Latin indices for quantities having three spatial components, $x_j$. We additionally employ arrow notation, $\vec{x} = ( x_1, x_2, x_3, 0)$, for quantities with three spatial components, and hat notation, $\hat{x}_j$, for unit spatial vectors. Bold symbols are reserved for quantities having the particular form $\bm{x} = (x_1, 0, x_3, x_4)$, i.e. a quantity that appears similar to a four-vector but with the second component deleted. Lastly we will often append a subscript $\perp$ to denote the spatial directions perpendicular to the magnetic field. Magnetic Field -------------- In infinite volume, the vector potential $A_\mu(x) = ( - B x_2, 0, 0, 0)$ gives rise to a uniform magnetic field, $\vec{B} = B \hat{x}_3$. On a torus, however, one needs to worry about boundary conditions. A particularly clear exposition of these issues is given in [@AlHashimi:2008hr]. While a uniform magnetic field is trivially periodic, the gauge potential $A_\mu (x)$ is not periodic, but is periodic only up to a gauge transformation, namely $$A_\mu ( x + L \hat{x}_j ) = A_\mu (x) + \partial_\mu \Lambda_j (x) \label{eq:nonP} ,$$ where the transformation function is given by $\Lambda_j (x) = ( 0, - B L x_1, 0)$. On a torus, moreover, the electric and magnetic fields are not the only gauge invariant quantities. There are additionally holonomies of the gauge field which can be expressed in terms of Wilson lines spanning the compact directions. For our application, it is useful to consider a matter field of electric charge $Q$, for which the Wilson lines have the form $$W_j (x) = e^{i Q \int_0^L dx_j A_j(x)} e^{ - i Q \Lambda_j(x)} ,$$ where the Einstein summation convention has been suspended. For the linearly rising gauge potential,[^3] there are two non-vanishing holonomies which are coordinate dependent, namely $$\begin{aligned} W_1(x_2) &=& e^{ - i Q B L x_2} ,\notag \\ W_2(x_1) &=& e^{i Q B L x_1} .\end{aligned}$$ Addressing effects of these holonomies on neutron correlation functions is one of the central goals of this work. Let the hadron possessing charge $Q$ be described by a field $\phi (x)$, and further that the gauge covariant derivative acting on the field has the form $D_\mu \phi = \partial_\mu \phi + i Q A_\mu \phi$. Accordingly when the gauge field transforms as $A_\mu \to A_\mu + \partial_\mu \alpha$ under a gauge transformation, the matter field has the corresponding transformation $\phi \to e^{-i Q \alpha} \phi$. In this way, the gauge covariant derivative transforms in precisely the same way as the matter field, $D_\mu \phi \to e^{- i Q \alpha} D_\mu \phi$. Given these transformation properties, the non-periodicity of the gauge potential exhibited in Eq.  can be gauged away at the boundary. As a result, a periodic matter field will satisfy a modified boundary condition after such a gauge transformation is made. Carrying out the boundary gauge transformation leads to the boundary condition $$\phi(x + L \hat{x}_j) = e^{ - i Q \Lambda_j (x)} \phi(x), \label{eq:MPBC}$$ in accordance with gauge invariance of the Euclidean action density. We refer to Eq.  as a magnetic periodic boundary condition (MPBC). For the gauge potential at hand, the matter field satisfies the MPBC: $\phi( x + L \hat{x}_2 ) = W_2(x_1) \phi (x)$, with periodicity obeyed in the other two spatial directions. The interpretation of this MPBC is simple, each time the charged particle winds around the $x_2$-direction, it picks up a Wilson line in accordance with gauge covariance. Although the field $\phi(x)$ is periodic in the $x_1$-direction, the other non-vanishing holonomy, $W_1(x_2)$, remains relevant as we show through explicit computation. It enters through the violation of translational invariance. Consistency of the MPBC requires quantization of the magnetic field. Consider winding around once in each direction transverse to the magnetic field to arrive at the field $\phi (x + L \hat{x}_1 + L \hat{x}_2)$. As the field is periodic in the $\hat{x}_1$ direction, we have $\phi (x + L \hat{x}_1 + L \hat{x}_2) = \phi(x + L \hat{x}_2) = W_2(x_1) \phi(x)$. On the other hand, the field satisfies a MPBC in the $\hat{x}_2$ direction, which can be taken into account first, $\phi (x + L \hat{x}_1 + L \hat{x}_2) = W_2( x_1 + L) \phi(x + L \hat{x}_1) = W_2(x_1 + L) \phi(x)$. Consequently the Wilson loops must satisfy a consistency condition $W_2(x_1) W_2^\dagger (x_1 + L) = 1$, which translates into a restriction on the size of the magnetic field $$Q B L^2 = 2 \pi N_\Phi \label{eq:quant} ,$$ where the integer $N_\Phi$ is the magnetic flux quantum of the torus. The quantization of the magnetic field naturally leads to a discrete $\mathbb{Z}_{N_\Phi}$ translational invariance. For integers $n \in [ 0, 1, \cdots, N_\Phi-1]$, we have invariance of the Wilson loops under the translations $$\begin{aligned} W_1 \left( x_2 + \frac{n}{N_\phi} L \right) &=& W_1 (x_2), \notag \\ W_2 \left(x_1 + \frac{n}{N_\phi} L \right) &=& W_2 (x_1) .\end{aligned}$$ Such discrete translational invariance is a constraint that must be satisfied by the finite volume neutron effective action. Charged Scalar Propagator ------------------------- In considering finite volume corrections to the neutron two-point function in chiral perturbation theory, we will require the finite volume propagator for the charged pion. On a spatial torus, the Euclidean action for a charged scalar field $\phi$ has the familiar Klein-Gordon form $$S = \int_0^L d\vec{x} \int_{- \infty}^\infty dx_4 \, \phi^\dagger(x) \left[ - D_\mu D_\mu + m^2 \right] \phi(x) .$$ On account of Eq. , the time-ordered correlation function for the scalar field, $G_{FV} (x',x) = \langle 0 | T \left\{ \phi(x') \phi^\dagger(x) \right\} | 0 \rangle$, must satisfy the following MPBCs $$\begin{aligned} G_{FV} ( x' + L \hat{x}_2, x ) &=& W_2(x'_1) G_{FV} ( x', x), \notag \\ G_{FV} ( x', x + L \hat{x}_2) &=& G_{FV} (x', x) W_2^\dagger (x_1) \label{eq:GMPBC} .\end{aligned}$$ The scalar two-point correlation function obeys periodic boundary conditions in the remaining two spatial directions. In order to satisfy the MPBCs required on the charged scalar two-point function, we follow [@Tiburzi:2012ks] and construct the finite volume propagator from magnetic periodic images of the infinite volume propagator. To satisfy magnetic periodicity, we must take $$G_{FV} (x', x) = \sum_{\vec{\nu}} [ W^\dagger_2 (x'_1)]^{\nu_2} G_\infty ( x' + \vec{\nu} L, x) \label{eq:GFV} .$$ Here we employ a shorthand notation for the sum over all winding numbers, $\sum_{\vec{\nu}} \equiv \sum_{\nu_1 = - \infty}^\infty \sum_{\nu_2 = -\infty}^\infty \sum_{\nu_3 = -\infty}^\infty$, and the function $G_\infty (x',x)$ is the infinite volume two-point function, which satisfies the Green’s function equation $\left[ - D'_\mu D'_\mu + m^2 \right] G_\infty (x', x) = \delta^{(4)} (x' - x)$. As a consequence, the finite volume two-point function satisfies an analogous equation $$\left[ - D'_\mu D'_\mu + m^2 \right] G_{FV} (x',x) = \delta^{(3)}_L (\vec{x} \,' - \vec{x}\,) \delta( x'_4 - x_4) \label{eq:FVGF} ,$$ where $\delta_L(x - y)$ is the Dirac delta-function appropriate for variables having compact support, $x, y \in [0, L]$. Without knowing the explicit form of the infinite volume correlation function, the first MPBC in Eq.  can easily be demonstrated by reindexing the sum over magnetic periodic images. Notice that despite the coordinate dependence introduced by the Wilson loops $W_2^\dagger(x'_1)$ in Eq. , the finite volume two-point function remains periodic in the $x_1$-direction. Demonstrating that the second MPBC is satisfied by Eq.  requires the explicit form of the infinite volume correlation function, which appears as [@Tiburzi:2008ma] $$\begin{aligned} G_\infty (x', x) &=& \frac{1}{2} \int_0^\infty ds \int \frac{d \bm{k}}{(2 \pi)^3} e^{ i \bm{k} \cdot ( \bm{x}' - \bm{x} )} e^{ - \frac{1}{2} s \cM^2} \notag \\ && \phantom{sp} \times \Big\langle x'_2 - \frac{k_1}{Q B}, s \Big | x_2 - \frac{k_1}{QB}, 0 \Big\rangle \label{eq:SHOprop} .\end{aligned}$$ In this expression, we use the notation $\int d \bm{k}$ for the integration $\int_{- \infty}^\infty dk_1 \int_{- \infty}^\infty dk_3 \int_{-\infty}^\infty dk_4$, and the parameter $\cM$ is defined by $\cM^2 = m^2 + k_3^2 + k_4^2$. The bracketed quantity is the quantum mechanical propagator for the simple harmonic oscillator $$\begin{aligned} &&\langle x', t' | x, t \rangle = \langle x' | e^{- \Delta t \, H } | x \rangle \notag \\ && \phantom{space} = \sqrt{\frac{Q B}{2 \pi \sinh (Q B \Delta t) }} \exp \Bigg[ - \frac{Q B}{2 \sinh ( Q B \Delta t)} \notag \\ && \phantom{spacing} \times \Big\{ ( x'^2 + x^2 ) \cosh (Q B \Delta t) - 2 x' x \Big\} \Bigg] ,\end{aligned}$$ where $\Delta t = t' - t$ is the Euclidean time difference with $\Delta t > 0$ understood, and $H = \frac{1}{2} p_x^2 + \frac{1}{2} ( Q B x)^2$ is the simple harmonic oscillator Hamiltonian. The quantum mechanical propagator is manifestly an even function of $Q B$, and thus we avoid superfluous $| Q B |$ notation. The scalar propagator is invariant under $Q B \to - QB$ accompanied by a mirror reflection in the $\hat{x}_1$-direction. The propagator is also invariant under a simultaneous reflection of the $\hat{x}_1$ and $\hat{x}_2$ directions. With the explicit form of the infinite volume two-point function, we can verify that the second MPBC is satisfied provided the magnetic field is appropriately quantized, as in Eq. . To simplify computations below, it is useful to perform the $\bm{k}$ integrals appearing in Eq. , which are each Gaußian. The integration produces a one-dimensional, proper-time integral representation for the coordinate-space propagator $$\begin{aligned} G_\infty (x',x) &=& e^{i Q B \Delta x_1 \ol x_2} \int_0^\infty \frac{ds}{(4 \pi s)^2} \frac{Q B s}{\sinh Q B s} e^{ - m_\pi^2 s} \notag \\ && \times \exp \left[ - \frac{ QB \Delta \vec{x} \, {}^2_\perp }{4 \tanh Q B s} - \frac{\Delta x_3^2 + \Delta x_4^2}{4 s} \right] ,\notag \\ \label{eq:Bprop}\end{aligned}$$ with $\Delta x_\mu = x'_\mu - x_\mu$, and $\ol x_\mu = \frac{1}{2} ( x'_\mu + x_\mu)$. The overall coordinate-dependent phase violates translational invariance, and accordingly vanishes when the external field is turned off. This phase factor leads to the appearance of Wilson loops $W_1(x_2)$ in finite volume computations, as well as an Aharonov-Bohm effect. Static Fermion Propagator ------------------------- The propagator for a charged fermion in a magnetic field can similarly be derived. We shall use the static approximation throughout, for which a simple form for the fermion propagator emerges. We also carefully check that there are no effects from magnetic periodic images in the static limit. The infinite volume static fermion propagator has the simple form $$D_\infty (x', x) = \delta^{(3)} (\vec{x} \, ' - \vec{x} \, ) \theta(x'_4 - x_4) \cP_+ ,$$ where $\cP_+ = \frac{1}{2} ( 1 + \gamma_4 )$ is the positive parity projection matrix. As the fermion is charged, let us construct its finite volume propagator by taking a sum over magnetic periodic images $$D_{FV} (x', x) = \sum_{\vec{\nu}} [ W^\dagger_2 (x'_1) ]^{\nu_2} D_\infty (x' + \vec{\nu} L, x) \label{eq:pprop} ,$$ so that the propagator satisfies MPBCs. Given the form of the infinite volume propagator, however, the finite volume correlation function can be simplified dramatically. Before attempting to simplify the heavy fermion correlation function, it is efficacious to relate the Dirac delta-function with compact support to a sum over non-compact delta-functions. This can be achieved through the Poission summation formula $$\frac{1}{L} \sum_{n = - \infty}^\infty \delta ( k - 2 \pi n / L ) = \frac{1}{2 \pi} \sum_{\nu = - \infty}^\infty e^{ i k L \nu} .$$ The Dirac delta-function with compact support can be written as a Fourier decomposition over quantized momentum modes $$\delta_L ( x - y) = \frac{1}{L} \sum_{n = - \infty}^\infty e^{ 2 \pi i n (x - y) / L} .$$ This sum over modes can then be converted into a winding number expansion by utilizing Poisson’s formula. As a result, the compact Dirac delta-function is given by the alternate expression $$\delta_L (x - y) = \sum_{\nu = - \infty}^\infty \delta ( x + \nu L - y) ,$$ which is written in terms of periodic images of the non-compact delta-function Having recalled these useful properties of Dirac delta-functions, we are now ready to handle the case of the static fermion propagator with magnetic periodic images. From Eq. , we can separate out the image contributions by writing the heavy fermion propagator in the form $$D_{FV}(x',x) = \cD_{FV} (\vec{x} \, ', \vec{x} \, ) \, \theta(x'_4 - x_4) \cP_+ \label{eq:heavy} ,$$ where the winding number dependence has been relegated to the function $$\cD_{FV} (\vec{x} \, ', \vec{x} \, ) = \sum_{\vec{\nu}} [W_2^\dagger (x'_1)]^{\nu_2} \delta^{(3)} ( \vec{x} \, ' + \vec{\nu} L - \vec{x} \, ) \label{eq:pimage} .$$ If the Wilson loops were absent from Eq. , we would obviously have the simple relation $\cD_{FV} (\vec{x} \, ', \vec{x} \, ) = \delta^{(3)}_L ( \vec{x} \, ' - \vec{x} \, )$. The presence of the Wilson loops, however, does not alter this relation. Physically this is quite sensible as the static fermion should remain at a fixed spatial location. To demonstrate that the relation holds in the presence of Wilson loops, we note that because the function $\cD_{FV} (\vec{x} \, ', \vec{x} \, )$ is written as a sum over non-compact Dirac delta-functions, we may utilize their continuous Fourier decomposition to arrive at the expression $$\cD_{FV} (\vec{x} \, ', \vec{x} \, ) = \sum_{\vec{\nu}} \int \frac{d \vec{k}}{( 2 \pi)^3} e^{ - i Q B L x'_1 \nu_2} e^{ i \vec{k} \cdot ( \vec{x} \, ' + \vec{\nu} L - \vec{x} \, )} .$$ The sum over winding numbers can be converted into a momentum mode sum by utilizing the Poisson summation formula. This produces the discrete Fourier mode decomposition $$\begin{aligned} \cD_{FV} (\vec{x} \, ', \vec{x} \, ) &=& e^{ i Q B x'_1 ( x'_2 - x_2) } \frac{1}{L^3} \sum_{\vec{n}} e^{ 2 \pi i \vec{n} \cdot ( \vec{x} \, ' - \vec{x} \, )} \notag \\ &=& \delta_L^{(3)} (\vec{x} \, ' - \vec{x} \, ) .\end{aligned}$$ The second equality holds because the overall phase factor becomes unity when multiplied by the delta-function. Crucial to this observation is that $x'_2$ and $x_2$ have compact support in $\cD_{FV} (\vec{x} \, ', \vec{x} \, )$. The diligent reader will realize that the derivation of this relation is also required above to arrive at Eq. . Nonetheless, the static charged fermion propagator maintains a simple form in an external magnetic field. One-Loop Computation {#s:One} ==================== Having spelled out the physics of non-interacting particles in uniform magnetic fields on a torus, we are now in a position to calculate the neutron two-point function in a magnetic field including leading-order effects of pion-nucleon interactions. We first give a brief review of heavy nucleon chiral perturbation theory with strong-field power counting in Sec. \[s:HNCPT\]. This review will explain the necessary ingredients of the finite volume computation. As the computation of the neutron two-point function will be carried out in coordinate space, we provide a simple example in zero magnetic field to exhibit features that may be unfamiliar. This zero-field example is presented in Sec. \[s:EX\]. Finally in Sec. \[s:CMF\], we derive the spin-dependent and spin-independent terms in the neutron effective potential by integrating out charged pions winding around the torus. Heavy Nucleon Chiral Perturbation Theory {#s:HNCPT} ---------------------------------------- The dynamics of the neutron at low-energies can be described in terms of an effective field theory, which is chiral perturbation theory. The main ingredient of chiral perturbation theory is the symmetry breaking pattern of QCD. For two massless quark flavors, the QCD action maintains an $SU(2)_L \otimes SU(2)_R \otimes U(1)_B$ symmetry that is spontaneously broken to $SU(2)_V \otimes U(1)_B$ by the formation of the chiral condensate. The resulting Goldstone manifold can be parameterized by a field $\Sigma$ that lives in the coset $SU(2)_L \otimes SU(2)_R / SU(2)_V$. The Goldstone modes are non-linearly realized in $\Sigma$: $$\Sigma = \exp \left( 2 i \varphi / f \right),$$ with the pions contained in $\varphi$ as $$\varphi = \begin{pmatrix} \frac{1}{\sqrt{2}} \pi^0 & \pi^+ \\ \pi^- & - \frac{1}{\sqrt{2}} \pi^0 \end{pmatrix} .$$ Our conventions are such that the pion decay constant $f$ has value $f \approx 130 \, \texttt{MeV}$. Under a chiral transformation $(L, R) \in SU(2)_L \otimes SU(2)_R$, we have the coset transformation $\Sigma \to L \Sigma R^\dagger$. The effective theory of pions can be constructed by writing down the most general chirally invariant Lagrangian. In constructing the chiral Lagrange density, one must consider additional sources of symmetry breaking. In an external magnetic field, for example, chiral symmetry is explicitly broken to the product of diagonal subgroups, $SU(2)_L \otimes SU(2)_R \to U(1)_L \otimes U(1)_R$. Non-vanishing quark masses explicitly break chiral symmetry down to the vector subgroup. With these two sources of explicit symmetry breaking taken into account, the resulting low-energy theory maintains only a $U(1)_V \otimes U(1)_B$ symmetry. Accounting for this pattern of spontaneous and explicit symmetry breaking, the chiral Lagrange density has the form $$\cL = \frac{f^2}{8} \operatorname{Tr}\left( D_\mu \Sigma^\dagger D_\mu \Sigma \right) - \frac{\lambda}{2} \operatorname{Tr}\left( m_Q [ \Sigma^\dagger + \Sigma ] \right) .$$ Appearing above is the quark mass matrix $m_Q = \operatorname{diag}( m_u, m_d )$, and we will work in the limit of strong isospin symmetry, $m_u = m_d = m$. The action of the covariant derivative on the coset field is specified by $$D_\mu \Sigma = \partial_\mu \Sigma + i A_\mu \left[ \cQ, \Sigma \right] ,$$ where the quark electric charges appear in the matrix $\cQ = e \operatorname{diag}\left( \frac{2}{3}, - \frac{1}{3} \right)$. The parameter $- \lambda$ is the chiral limit value of the chiral condensate. In writing the chiral Lagrange density, we have only included the lowest-order terms. We define these terms to scale as $\cO(p^2)$, where $p$ represents a small dimensionless number. The power counting we employ hence assumes that $$\frac{m_\pi^2}{\Lambda_\chi^2} \sim \frac{k^2}{\Lambda_\chi^2} \sim \frac{e F_{\mu \nu}}{\Lambda_\chi^2} \sim \frac{(e A_\mu)^2}{\Lambda_\chi^2} \sim p^2 \label{eq:PC} ,$$ where $\Lambda_\chi \sim 2 \sqrt{2} \pi f$ is the chiral symmetry breaking scale. In this counting, $m_\pi$ is the pion mass, which satisfies the Gell-Mann–Oakes–Renner relation $f^2 m_\pi^2 = 4 \lambda m$. The first condition of the power counting is implicitly a restriction on the size of the quark mass, which should be well satisfied for the up and down quarks. The second condition of the power counting concerns the typical momentum $k$. In finite volume, the available momentum modes of a free particle are quantized in the form $\vec{k} = \frac{2 \pi}{L} \vec{n}$, and thus the second condition implies the restriction to modes that satisfy $\vec{n} \, {}^2 / 2 ( f L)^2 \sim p^2$. Combining the first and second conditions, we have the $p$-regime constraint [@Gasser:1987zq], namely $m_\pi L \sim 2 \pi | \vec{n} | \gg 1$, for $| \vec{n} | \neq 0$. This constraint ensures that the pion Compton wavelength remains relatively large compared to the size of the torus. When this condition is not met, the pion zero modes become strongly coupled and the chiral condensate will be substantially depleted [@Gasser:1987ah]. The third condition of the power counting concerns the external magnetic field, whose strength must be weak compared to the chiral symmetry breaking scale. This condition treats the magnetic field as strong; because, combining with the first condition, we have $|e B| / m_\pi^2 \sim 1$. Determination of corrections to the chiral condensate in this regime has been carried out in [@Cohen:2007bt]. Combining the second and the third conditions, we see the power counting supports many values of the magnetic flux quantum, $N_\Phi \sim 2 \pi | \vec{n} |^2 \gg 1$. The final condition of the power counting allows for the holonomies to appear at leading order. The argument of the Wilson loops have the general form $e B L x_\perp$, where $x_\perp = ( x_1, x_2)$ refers to either of the coordinates transverse to the direction of the magnetic field. By combining the second and final conditions, we arrive at $(e B x_2)^2 \sim k^2$, which translates into the condition $| e B | L x_2 \sim 2 \pi | \vec{n} | \gg 1$. By the cubic symmetry of the torus, we must have $| e B | L x_\perp \gg 1$, too. In this counting, the Wilson loops are not amenable to perturbative expansion in the strength of the magnetic field. In the strong-field, $p$-regime power counting, the neutral pion propagator retains its Klein-Gordon form. The charged pion propagator requires the summation of charge couplings as mandated by Eq. , see Fig. \[f:prop\] for a graphical depiction of the charged pion propagator. The leading-order charged pion Lagrange density appears as $$\cL = D_\mu \pi^- D_\mu \pi^+ + m_\pi^2 \, \pi^- \pi^+ ,$$ and thus the charged pion propagator has the form determined above in Eq. . This propagator, $G_{FV}(x',x)$, is the crucial new ingredient for the finite volume computation of the neutron two-point function. The Lagrange density for a nucleon in an external field has the form $$\cL = \ol N \left( \gamma_\mu D_\mu + M \right) N ,$$ where $N$ is the nucleon isodoublet, $N = \begin{pmatrix} p \\ n \end{pmatrix}$. We treat the mass $M$ as the same order as the chiral symmetry breaking scale $\Lambda_\chi$. In order to have a low-energy expansion, we must make a field redefinition to eliminate the mass term [@Jenkins:1990jv]. To this end, we use the heavy nucleon field $N_v$ defined by $$N_v(x) = e^{M x_4} \cP_+ N(x) .$$ With this redefinition, derivatives acting on the heavy nucleon field produce a residual momentum $k$ that can be treated as small, $k / M \sim \cO(p)$ in the power counting. The relevant terms of the $\cO(p)$ nucleon chiral Lagrange density appear as $$\cL = \ol N_v \cD_4 N_v + g_A \ol N_v \vec{\sigma} \cdot \vec{\cA} \, N_v ,$$ where $\vec{\sigma}$ are the Pauli spin matrices, and $g_A \sim 1.25$ is the nucleon axial charge. In practice, the nearby delta-resonance makes sizable contributions to nucleon properties. Inclusion of these resonances is described in Appendix \[s:D\]. We omit the delta resonance in the main text only to keep the discussion as simple as possible. Electromagnetism is coupled into the pion-nucleon Lagrange density through the axial $\cA_\mu$ and vector $\mathcal{V}_\mu$ fields of pions. These are given by $$\begin{aligned} \mathcal{V}_\mu &=& i Q A_\mu + \frac{1}{2 f^2} \left( \phi D_\mu \phi - D_\mu \phi \, \phi \right) + \cdots, \notag \\ \cA_\mu &=& \frac{1}{f} D_\mu \phi + \cdots ,\end{aligned}$$ where terms of order $p^2$ and higher have been dropped. The action of the chirally covariant derivative $\cD_\mu$ on the nucleon field is specified by $$(\cD_\mu)_i = \partial_\mu N_i + (\mathcal{V}_\mu)_i {}^{i'} N_{i'} + \operatorname{Tr}\left( \mathcal{V}_\mu \right) N_i .$$ There are additional local interactions that are required to determine the neutron two-point function to $\cO(p^3)$. The magnetic field independent term at $\cO(p^2)$ is the nucleon mass operator, which gives rise to linear quark mass dependence of the nucleon mass away from the chiral limit. There are also magnetic field dependent operators, which are simply the isosinglet and isovector magnetic moment operators $$\cL = \frac{e}{2 M} \left[ \kappa_0 \ol N_v \vec{\sigma} \cdot \vec{B}\, N_v + \kappa_1 \ol N_v \vec{\sigma} \cdot \vec{B} \, \tau^3 N_v \right] \label{eq:magmom} ,$$ and also count as $\cO(p^2)$. Above we demonstrate that there is no effect from magnetic periodic images in the static limit. As a result, the propagators of both nucleons are given by $D_{FV}(x',x)$ in Eq. . Example in Zero Magnetic Field {#s:EX} ------------------------------ Before computing the neutron two-point function in an external magnetic field, it is efficacious to show how the finite volume computation of the nucleon mass proceeds in coordinate space. A salient feature of the coordinate space approach is that the propagators are written in a winding number expansion from the outset. The one-loop sunset diagram for the nucleon two-point function is shown in Fig. \[f:sunset\]. Evaluating this diagram gives rise to the leading volume dependence of the nucleon mass. In coordinate space, the correction to the two-point function has the form $$\begin{aligned} \delta D_{FV} (x',x) &=& C \sumint D_{FV}(x',y) \sigma_i D_{FV}(y,z) \sigma_j \notag \\ && \phantom{sp} \times \left[ \frac{\partial^2}{\partial y_i \partial z_j} G^{(0)}_{FV}(y,z) \right] D_{FV}(z,x) ,\notag \\ \label{eq:mess}\end{aligned}$$ where $C$ is a dimensionful constant given by $\frac{3}{2} g_A^2 / f^2$. The factor of $\frac{3}{2} = \frac{1}{2} + 1$ takes into account both the neutral pion loop ($\frac{1}{2}$) and the charged pion loop ($1$). The sum/integral notation reflects the $\mathbb{T}^3 \times \mathbb{R}$ spacetime $$\sumint = \int_0^L d \vec{y} \int_{- \infty}^\infty dy_4 \int_0^L d \vec{z} \int_{- \infty}^\infty dz_4 .$$ For this zero-field example, the pion propagator is that of a free scalar $$G_{FV}^{(0)} (x',x) = \sum_{\vec{\nu}} G_\infty^{(0)} (x' + \vec{\nu} L,x) ,$$ with the infinite volume propagator $$G^{(0)}_\infty (x',x) = \int_0^\infty \frac{ds}{ (4 \pi s)^2} e^{ - m_\pi^2 s} e^{ - (x'-x)^2 / 4 s} .$$ This familiar proper-time integral representation for the propagator emerges from the zero-field limit of Eq. . To compute the nucleon mass shift from Eq. , we first amputate the external legs. In coordinate space, this is easy to achieve because propagators for the external legs are Green’s functions satisfying the equation $\frac{\partial}{\partial x'_4} D_{FV} (x',x) = \delta_{L}^{(3)} (\vec{x} \, ' - \vec{x} \, ) \delta(x'_4 - x_4)$. Defining the amputated contribution to the two-point function as $[\delta D_{FV}(x',x)]_{\text{amp}} \equiv - \frac{\partial^2}{\partial x'_4 \partial x_4} \delta D_{FV} (x'x)$, we see that the amputated sunset diagram is $$[\delta D_{FV}(x',x)]_{\text{amp}} = C \, D_{FV}(x',x) \vec{\nabla}' \cdot \vec{\nabla} G^{(0)}_{FV}(x',x) ,$$ upon simplifying the spin structure. Corrections to the nucleon mass are identified by projecting onto zero residual nucleon energy, $P_4 =0$.[^4] In general, we can write this projection in terms of an effective potential $$\label{eq:recipe} \int_{- \infty}^\infty d(x'_4 - x_4) [\delta D_{FV}(x',x)]_{\text{amp}} = - V(\vec{x} \, ', \vec{x} \, ) .$$ The effective potential then appears as a correction to the nucleon Lagrangian $$L(x_4) = \int_0^L d\vec{x} \,' \int_0^L d\vec{x} \,\, \ol N_v(\vec{x} \, ', x_4) V(\vec{x} \, ', \vec{x} \, ) N_v( \vec{x} \, , x_4) .$$ We have not appealed to spatial symmetries to simplify the form of the effective potential. Indeed for the simple example at hand, the potential is local, and otherwise coordinate independent. In an external magnetic field, however, we will find that the effective potential remains local, but is coordinate dependent. Carrying out the projection in Eq. , we indeed find a local potential, $V(\vec{x} \, ' , \vec{x} \, ) = \Delta M \, \delta^{(3)}_L (\vec{x} \, ' - \vec{x} \, )$, with the constant $\Delta M$ given by $$\begin{aligned} \Delta M &=& C \frac{m_\pi^2 \sqrt{\pi}}{(4 \pi)^2} \sum_{\vec{\nu}} \int_0^\infty \frac{ds}{s^{3/2}} e^{- m_\pi^2 s} e^{ - \frac{\vec{\nu} \, {}^2 L^2}{ 4 s}} ,\end{aligned}$$ having utilized a proper-time integration by parts. In this simple example, the constant $\Delta M$ is just the correction to the nucleon mass. Separating out the infinite volume contribution, $\Delta M_\infty$, which arises form the sector of zero winding number $\vec{\nu} = \vec{0}$, we have $$\Delta M = \Delta M_\infty + \Delta M(L) ,$$ where the infinite volume contribution after dimensional regularization results in the well-known expression [@Jenkins:1991ts; @Bernard:1993nj] $$\Delta M_\infty = - \frac{3 g_A^2}{16 \pi f^2} m_\pi^3 \label{eq:infinite} ,$$ and the finite volume correction is $$\Delta M(L) = \frac{3 g^2_A m^2_\pi}{16 \pi f^2} \sum_{\vec{\nu} - \{ \vec{0} \} } \frac{ e^{ - |\vec{\nu}| m_\pi L}} { |\vec{\nu}| L } .$$ This expression for the finite volume correction to the nucleon mass is that derived in [@Beane:2004tw]. Having recovered familiar results from the coordinate-space approach, we are now ready to pursue the computation in an external magnetic field. Computation in Magnetic Field {#s:CMF} ----------------------------- To compute the $\cO(p^3)$ correction to the neutron effective action, we must evaluate the contributions depicted diagrammatically in Fig. \[f:sunsetB\]. The diagrams in the first row vanish, and the remaining four diagrams are related by gauge invariance. As a result, we can express the four sunset diagrams as one contribution by utilizing the gauge covariant derivative at each pion-nucleon vertex. Amputating the propagators of external legs produces $$[\delta D_{FV}(x',x)]_{\text{amp}} = C \, \sigma_i D_{FV}(x',x) \sigma_j D'_i D_j G_{FV}(x',x) \label{eq:amp} ,$$ where the parameter $C$ now arises solely from the charged pion loop, and is accordingly given by $C = g_A^2 / f^2$. This amputated correction must be projected onto vanishing neutron residual energy, $P_4 = 0$, to derive the neutron effective potential. There are both spin-independent and spin-dependent contributions to the effective potential. ### Spin-Independent Term The spin-independent part can be evaluated by considering how the action of $\vec{D}' \cdot \vec{D}$ on $G_{FV} (x',x)$ differs from that of $- \vec{D}' {}^2$. To this end, we write $$\vec{D}' \cdot \vec{D} = - \vec{D}' {}^{2} + \vec{D}' \cdot \vec{D}_+ ,$$ where $\vec{D}_+ = \vec{D}' + \vec{D}$. From the expression for the propagator in position space, Eqs.  and , we see the only terms that can contribute to $\vec{D}_+$ are the phase factors. This is not surprising because it is precisely these factors that break translational invariance down to the discrete $\mathbb{Z}_{N_\Phi}$ subgroup. As translational invariance is preserved in the $\hat{x}_3$-direction, we have $(D_+)_3 = 0$. Now consider the derivatives acting in the $\hat{x}_2$-direction. We have simply $$(D_+)_2 \, G_\infty (x',x) = i Q B \Delta x_1 \, G_\infty (x',x) .$$ A slightly more involved computation is required for the derivatives acting in the $\hat{x}_1$-direction. On the $\vec{\nu} \, {}^\text{th}$ image, we find that $$\begin{aligned} &&(D_+)_1 \, [W^\dagger_2(x'_1)]^{\nu_2} G_\infty (x' + \vec{\nu} L,x) \notag \\ &&\phantom{spa} = - i Q B ( \Delta x_2 + \nu_2 L ) \, [W^\dagger_2(x_1)]^{\nu_2} G_\infty (x' + \vec{\nu} L,x) . \notag \\\end{aligned}$$ Consequently on the $\vec{\nu} \, {}^\text{th}$ image, we are led, after some cancellations within the proper-time integral, to the equivalence $$\vec{D}' \cdot \vec{D}_+ \longrightarrow - \frac{1}{2} ( Q B L \vec{\nu}_\perp)^2 .$$ To evaluate the action of $\vec{D}' {}^2$, we write it in two terms, $\vec{D}' {}^2 = D'_\mu D'_\mu - (D'_4)^2$. For the former term, we can utilize the Green’s function equation; while for the latter term, it vanishes after projecting the amputated diagram onto definite $P_4 = 0$. Assembling results of these observations, we can deduce the spin-independent neutron effective potential. To exhibit various features, we write the result in the form $$V_0 ( \vec{x} \, ', \vec{x} \, ) = \delta_L^{(3)} ( \vec{x} \, ' - \vec{x} \, ) \left[ E_0 + V_0 (\vec{x}_\perp) \right] \label{eq:VeeZero} .$$ Locality emerges due to the static limit of the proton propagator. The coordinate independent piece $E_0$ appearing above is the spin-independent neutron energy shift in an external magnetic field. This term arises in infinite volume. Subtracting off the charged pion contribution to the nucleon mass appearing in Eq. , we have $$\begin{aligned} E_0 = \frac{g_A^2 m_\pi^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} e^{ - m_\pi^2 s} \left( \frac{e B s}{\sinh e B s} - 1 \right) \label{eq:strong0} .\end{aligned}$$ In weak magnetic fields, $|e B| / m_\pi^2 \ll 1$, we can expand the neutron energy shift $E_0$ in the form $$E_0 = - \frac{1}{2} 4 \pi \beta_M B^2 + \cdots,$$ where $\beta_M$ is the magnetic polarizability given by $\beta_M = \frac{e^2}{4 \pi} \frac{g_A^2}{96 \pi f^2 m_\pi}$, and agrees with the known value from chiral perturbation theory [@Bernard:1991rq]. The strong field result in Eq.  treats the magnetic field non-perturbatively, $|e B |/ m_\pi^2 \sim 1$, and was originally derived in [@Tiburzi:2008ma].[^5] As a matter of curiosity, the Laplace transform required for $E_0$ can be expressed in terms of well-known functions, namely $$E_0 = \frac{g_A^2 m_\pi^2 }{(4 \pi f)^2} \sqrt{\pi | e B|} \, \mathcal{J} \left( \frac{m_\pi^2}{| e B|} \right) \label{eq:E0nonP} ,$$ with $$\begin{aligned} \mathcal{J} (x) &=& \sqrt{2 \pi} \left[ \sqrt{2 x} + \zeta \left( \frac{1}{2}, \frac{x+1}{2} \right) \right] \label{eq:Jzeta} ,\end{aligned}$$ where $\zeta$ is the generalized zeta-function. In Fig. \[f:E0\], the spin-independent neutron energy in Eq.  is compared with its weak-field limit. In particular, we focus outside the perturbative regime, where the expansion parameter, $\xi = e B / m_\pi^2$, satisfies $|\xi| \gtrsim 1$. Magnetic fields of this size and larger are readily encountered on current-day lattices. Outside the perturbative regime, adding higher-order corrections worsens the agreement with the non-perturbative result. This is the expected behavior of asymptotic expansions. Notice, however, that the non-perturbative result can be reasonably well approximated as a quadratic plus a quartic correction. Hence a fit to lattice data that line up with the non-perturbative curve will not necessarily be able to distinguish between a perturbative model, and more complicated zeta-function behavior. Because the leading-order term of an asymptotic expansion results in the best approximation outside the perturbative regime, it may well be that the value of the quadratic coefficient determined from a perturbative model fit has a value close to the true leading-order coefficient. It should be noted that this procedure has a fundamental limitation that can only be addressed by comparing perturbative model fits to those using the expected non-perturbative behavior. The remaining term $V_0(\vec{x}_\perp)$ appearing in the spin-independent neutron energy, Eq. , is the effective potential for the neutron, and it arises in this computation from finite volume effects. This spin-independent potential has the form $$\begin{aligned} V_0(\vec{x}_\perp) &=& \frac{g_A^2 \sqrt{\pi}}{(4 \pi f)^2} \sum_{\vec{\nu} - \{\vec{0} \}} (-)^{N_\Phi \nu_1 \nu_2} [W^\dagger_1(x_2)]^{\nu_1} [W^\dagger_2(x_1)]^{\nu_2} \notag\\ && \times \int_0^\infty \frac{ds}{s^{3/2}} \frac{e B s}{\sinh e B s} \left[ m_\pi^2 + \frac{1}{2} ( e B \vec{\nu}_\perp L)^2 \right] \notag \\ && \times \exp \left[ - m_\pi^2 s - \frac{e B \vec{\nu}_\perp^{\, 2} L^2}{4 \tanh e B s} - \frac{\nu_3^2 L^2}{4 s} \right] . \label{eq:V0}\end{aligned}$$ Because this result depends on the Wilson loops, the effective potential obeys the discrete $\mathbb{Z}_{N_\Phi}$ magnetic translational invariance of the torus. It is amusing to note that the oscillating sign factor, $(-)^{N_\Phi \nu_1 \nu_2}$, naturally arises from the Aharonov-Bohm effect, $(-)^{N_\Phi \nu_1 \nu_2} = e^{i e B (\nu_1 L) (\nu_2 L) / 2}$, where, by periodicity, the flux penetrates a triangular region in the plane transverse to the magnetic field, see Fig. \[f:AB\]. In the limit of vanishing magnetic field, the coordinate dependence of the effective potential disappears, and we recover the charged pion contribution to the finite volume neutron mass, $V_0(\vec{x}_\perp) \overset{B\to0}{=} \Delta M(L)$. In a non-vanishing magnetic field, the winding number sums can be cast in terms of Jacobi elliptic-theta functions. As the resulting expression for $V_0(\vec{x}_\perp)$ is rather lengthy, we display the result in Appendix \[s:A\]. Various features of the result will be exhibited below in Sec. \[s:Disc\]. ### Spin-Dependent Term We now evaluate the spin-dependent term entering the one-loop expression for the neutron two-point function. Using the familiar properties of Pauli matrices, we see that the spin-dependent term in Eq.  is proportional to $i \vec{\sigma} \cdot (\vec{D}' \times \vec{D} )$. Hence whenever the action of $\vec{D}$ is the same as $- \vec{D}'$, there is a spin-dependent contribution of the form $- i \vec{\sigma} \cdot (\vec{D}' \times \vec{D}' ) = Q \, \vec{\sigma} \cdot \vec{B} = Q B \sigma_3$. To this end, we write $$i \vec{\sigma} \cdot (\vec{D}' \times \vec{D} ) = Q B \sigma_3 + i \vec{\sigma} \cdot ( \vec{D}' \times \vec{D}_+) ,$$ where the action of $\vec{D}_+$ on the $\vec{\nu} \, {}^\text{th}$ magnetic periodic image has been determined above. To simplify the cross product, we notice that $D'_3$ is proportional to $\Delta x_3$; and, in the $\vec{\nu} \, {}^\text{th}$ image, this displacement becomes $\nu_3 L$ which is odd under reindexing $\nu_3 \to - \nu_3$. All of the remaining factors in propagators of magnetic periodic images are even in $\nu_3$. We can therefore eliminate $D'_3$ from the cross product. Because $\vec{D}_+$ has only components transverse to the magnetic field, and the $D'_3$ contribution is zero, the finite volume effect only allows the neutron spin to point in the direction of the magnetic field, as in infinite volume. Thus we have $$i \vec{\sigma} \cdot ( \vec{D}' \times \vec{D}_+ ) \longrightarrow i \sigma_3 [ D'_1 \, (D_+)_2 - D'_2 \, (D_+)_1] .$$ In determining this contribution, we must not forget terms of the form $i \sigma_3 \Big( [ D'_1, (D_+)_2] - [D'_2, (D_+)_1] \Big)$, which evaluate to $ - 2 Q B \sigma_3$. Assembling the results of these observations produces the expression for the spin-dependent neutron effective potential. A convenient way to write the result is $$V_1 (\vec{x} \, ', \vec{x} ) = e B \sigma_3 \, \delta_L^{(3)} (\vec{x} \, ' - \vec{x} ) \left[ E_1 + V_1 ( \vec{x}_\perp ) \right] \label{eq:VeeOne} ,$$ where locality again emerges due to the static limit of the intermediate-state proton propagator. The coordinate independent piece, $E_1$ determines the spin-dependent neutron energy in the magnetic field, which leads to a Zeeman energy splitting: $2 e B E_1$. This contribution arises from the infinite volume limit, and can be written in the form $$E_1 = \frac{\kappa_n}{ 2 M} + \Delta E_1 \label{eq:strong1} .$$ Here $\kappa_n $ is the neutron magnetic moment, which includes the tree-level contribution, $\kappa_0 - \kappa_1$, from the magnetic moment operators in Eq. , as well as the charged pion loop contribution. The latter emerges from our computation in the sector of zero winding number, $\vec{\nu} = \vec{0}$, and is the contribution at linear order in the magnetic field. Using dimensional regularization, we see $$\kappa_n = \kappa_0 - \kappa_1 + \frac{g_A^2 M m_\pi}{4 \pi f^2} ,$$ which is the standard result from chiral perturbation theory [@Jenkins:1992pi; @Meissner:1997hn]. The contribution to the energy proportional to $\Delta E_1$ arises from treating the magnetic field as strong. It is given by the expression $$\Delta E_1 = - \frac{g_A^2}{( 4 \pi f)^2} \sqrt{ \pi | e B |} \mathcal{J} \left( \frac{m_\pi^2}{|e B|} \right) \label{eq:E1nonP} ,$$ where $\mathcal{J}(x)$ is defined in Eq. . In Fig. \[f:E1\], the neutron Zeeman splitting is plotted as a function of the magnetic field. The splitting is defined to be the energy difference between spin-polarized states normalized by the neutron mass. We use the physical neutron magnetic moment to generate this plot. As with the spin-independent case, our primary focus in the spin-dependent case is outside the perturbative regime, and our consideration is motivated by the size of uniform magnetic fields available on typical lattices. In the case of the Zeeman splitting, we consider values of the expansion parameter satisfying $|\xi| \sim 5$. For such magnetic fields, the perturbative expansion has broken down, as is evidenced by the figure. Despite this breakdown, the non-perturbative result follows a strikingly linear curve, with only slight curvature appearing as $|\xi|$ approaches $10$. Such field strengths, however, satisfy the relation $| e B| /M m_\pi \sim 1$, and our results are not reliable for these considerably large values of the magnetic field. It would be interesting to study whether this fortuitous behavior persists when recoil-order corrections are added. Nonetheless, fits to lattice data that follow the non-perturbative curve will not necessarily be able to distinguish between a perturbative model, and more complicated non-perturbative behavior. The figure suggests that the value of the magnetic moment deduced from fitting the non-perturbative curve to a straight line will not be far off from the actual value of the magnetic moment. While the leading-order term of an asymptotic expansion results in the best approximation outside the perturbative regime, this approximation is uncontrolled with no guarantee that the leading-order term is even close to the non-perturbative result. Chiral perturbation theory suggests that the magnetic moment might be a fortuitous case, but care is needed to evaluate the behavior of the expansion in powers of the magnetic field. The remaining term $V_1(\vec{x}_\perp)$ appearing in the spin-dependent neutron energy, Eq. , arises solely from finite volume effects. This spin-dependent neutron potential takes the form $$\begin{aligned} V_1(\vec{x}_\perp) &=& \frac{g_A^2 \sqrt{\pi}}{(4 \pi f)^2} \sum_{\vec{\nu} - \{\vec{0} \}} (-)^{N_\Phi \nu_1 \nu_2} [W^\dagger_1(x_2)]^{\nu_1} [W^\dagger_2(x_1)]^{\nu_2} \notag \\ && \times \int_0^\infty \frac{ds}{s^{3/2}} \frac{e B s}{\sinh e B s} \left[ - 1 + \frac{ e B \vec{\nu}^{\, 2}_\perp L^2}{2 \tanh e B s} \right] \notag \\ && \times \exp \left[ - m_\pi^2 s - \frac{e B \vec{\nu}_\perp^{\, 2} L^2}{4 \tanh e B s} - \frac{\nu_3^2 L^2}{4 s} \right] .\label{eq:V1} $$ Despite the explicit coordinate dependence, the potential maintains the discrete $\mathbb{Z}_{N_\Phi}$ magnetic translational invariance. As with the spin-independent potential, the oscillating sign factor present in this expression, $(-)^{N_\Phi \nu_1 \nu_2}$, reflects the Aharonov-Bohm effect. The winding number sums can be recast in terms of Jacobi elliptic-theta functions, and we display this lengthy result in Appendix \[s:A\]. Because the potential takes a rather complicated form, we discuss various features by first considering simplifying limits. Discussion and Conclusion {#s:Disc} ========================= Discussion of Results --------------------- Above we determine the effective potential for the neutron in a magnetic field by integrating out charged pions winding around the torus. The charged pion winding is accompanied by Wilson loops that reflect the non-trivial holonomy of the gauge field in accordance with gauge invariance. As the general result is rather complicated, we exhibit features by using various limits before explaining how to determine the finite volume effect more generally. ### Extreme Weak-Field Limit As the simplest limit to consider, let us imagine the limit of an extremely weak field, $| e B| / m_\pi^2 \ll 1$ and $| e B | L x_\perp \ll 1$. The latter constraint requires localization of the physics to the bulk of the lattice. In this limit, we could expand the Wilson loops perturbatively in the strength of the magnetic field; although, we argue against this below. As the spin-dependent potential starts out at linear order in the magnetic field, we can imagine taking the strict zero-field limit of the function $V_1(x_\perp)$, which results in a coordinate-independent finite volume effect $$\begin{aligned} V_1(x_\perp)\Big|_{B = 0} &=& \frac{g_A^2 \sqrt{\pi}}{(4 \pi f)^2} \sum_{\vec{\nu} - \{\vec{0} \}} \int_0^\infty \frac{ds}{s^{3/2}} \left[ - 1 + \frac{\vec{\nu}^{\, 2} L^2}{3 s} \right] \notag \\ && \times \exp \left[ - m_\pi^2 s - \frac{\vec{\nu}^{\, 2} L^2}{4 s} \right] .\end{aligned}$$ In an extremely weak magnetic field, this contribution could be identified as the finite volume effect on the neutron magnetic moment. Performing the proper-time integration results in $$\Delta \kappa_n(L) = \frac{g_A^2 M m_\pi}{6 \pi f^2} \sum_{\vec{\nu} - \{\vec{0} \}} \left[ 1 - \frac{1}{2 |\vec{\nu} \, | m_\pi L} \right] e^ {- |\vec{\nu}\, | m_\pi L} \label{eq:FVmoment} .$$ This result for the finite volume modification to the magnetic moment differs from the one obtained in [@Beane:2004tw].[^6] That result, however, was obtained by taking the derivative $\partial/\partial m_\pi^2$ of the finite volume correction to the nucleon mass. While infinite volume results are related by taking a quark mass derivative, the finite volume corrections are not. In a three-point function computation of the current matrix element, additional terms appear in finite volume due to the quantization of momentum transfer between initial- and final-state nucleons. These terms reflect the breaking of rotational invariance. A detailed critique of the computation of [@Beane:2004tw] appears in [@Tiburzi:2007ep]. On the other hand, isospin twisted boundary conditions provide a method to overcome the limitation to quantized momentum transfer. Finite volume corrections to the nucleon isovector magnetic moment with isospin twisted boundary conditions have been determined in [@Tiburzi:2006px]. When one takes the limit of vanishing twist parameter $\theta$, the finite volume correction to the neutron magnetic moment is precisely that appearing above in Eq. . This equivalence is demonstrated in Appendix \[s:B\]. The situation is different with respect to the extreme weak-field limit of the spin-independent neutron potential. In the unpolarized case, the strict zero-field limit produces the finite volume correction to the neutron mass (from the charged pion loop). We must then expand to second order in the magnetic field to arrive at new results. Potentially contained in such results would be the finite volume correction to the magnetic polarizability. At this order, however, terms arising from expanding the holonomies will be present and these lead to coordinate dependence. Analogous results have been considered for pion two-point functions in perturbatively small external fields, see [@Tiburzi:2008pa]. The oscillating signs arising from Aharonov-Bohm phases, moreover, cannot be sensibly expanded even in an extremely weak magnetic field. We will avoid the temptation to expand any phase factors in powers of the magnetic field. These topological effects are inherently non-perturbative in the field strength, and properly reflect the physics on a torus. ### Magnetic Fields on a Large Lattice Simple expressions can be obtained by considering uniform magnetic fields on a finite lattice. We take the lattice spacing, $a$, to have the same value in directions transverse to the magnetic field, with the remaining directions irrelevant to our discussion. As a result, the transverse coordinates can be indexed by lattice sites, $\vec{x}_\perp = a \, \vec{n}_\perp$. The restriction to finite lattice spacing leads to another possible quantization condition for the magnetic field $$Q B = \frac{2 \pi N_\Phi}{a L} \label{eq:otherB} .$$ On typical lattices, such magnetic fields are prohibitively large; however, in the present context, they eliminate complications from the gauge holonomies. For magnetic fields obeying the quantization condition in Eq. , we have $$\begin{aligned} W_1(x_2) &=& e^{- 2 \pi i N_\Phi x_2 / a} = 1 ,\notag \\ W_2(x_1) &=& e^{ 2 \pi i N_\Phi x_1 / a} = 1 ,\end{aligned}$$ because $\vec{x}_\perp / a$ indexes the lattice sites. The effect of topology is almost entirely eliminated. The Aharonov-Bohm phases still remain, but depend on the lattice volume $$e^{i Q B L^2 \nu_1 \nu_2 / 2} = (-)^{N_\Phi \nu_1 \nu_2 L / a} ,$$ where $L / a$ is the number of lattice sites in a transverse direction. When the magnetic field is quantized according to Eq. , the Aharonov-Bohm phases disappear on even-site lattices. In the following, we assume that the lattice has an even number of sites in the transverse directions for ease. Due to the absence of holonomies, the finite volume corrections in the spin-independent and spin-dependent cases are given simply by $V_0 (\vec{0}_\perp)$ and $V_1(\vec{0}_\perp)$, respectively. If we demand to identify finite volume corrections to the magnetic moment and magnetic polarizability from these results, then we must additionally assume the weak-field limit, $|e B| / m_\pi^2 \ll 1$. Because the spin-dependent energy is already linear in the magnetic field, we can evaluate the finite volume correction $V_1( \vec{0}_\perp)$ to zeroth order in the weak-field limit. This leads us to the finite-volume correction to the magnetic moment given above in Eq. , provided the lattice has an even number of sites in the transverse directions. We can additionally deduce the finite volume correction to the magnetic polarizability by expanding the spin-independent energy shift in powers of the magnetic field. Because we temporarily adopt the quantization condition in Eq. , there are no holonomies complicating the expansion in powers of the magnetic field. Additionally the assumption of an even number of lattice sites removes Aharonov-Bohm phases. Writing the expansion of the spin-independent finite volume effect in the form $$V_0 (\vec{0}_\perp) = \Delta M (L) - \frac{1}{2} 4 \pi \, \Delta \beta_M (L) \, B^2 + \cdots \label{eq:weakV0} ,$$ we find the finite volume correction to the magnetic polarizability is given by $$\begin{aligned} \Delta \beta_M (L) &=& \frac{e^2}{4\pi} \frac{g_A^2 m_\pi L^2}{144 \pi f^2} \sum_{\vec{\nu} - \{ \vec{0} \} } |\vec{\nu} |^2 e^{ - | \vec{\nu} | m_\pi L} \notag \\ && \times \Bigg[ 1 - \frac{21}{2} \frac{1}{ | \vec{\nu} | m_\pi L} + \frac{3}{2} \frac{1}{(\vec{\nu} \, m_\pi L)^2} \Bigg] \label{eq:betaFV} .\end{aligned}$$ We have been unable to relate this result to the one claimed in [@Detmold:2006vu]. That analysis, however, neglected breaking of rotational invariance. It is conceivable that a computation of the Compton scattering tensor with twisted boundary conditions taken in the limit of vanishing twist angles will reproduce the finite volume effect in Eq. . An investigation along these lines is left to future work. Blindly using the weak-field result to compute finite volume corrections to the magnetic polarizability yields unreasonably large corrections on the order of $\sim 500 \%$ at the physical pion mass, for a lattice size of $L = 4 \, \texttt{fm}$. These volume effects can be softened to $\sim 50 \%$ by retaining terms of all orders in $e B / m_\pi^2$; however, the underlying problem concerns the size of magnetic fields arising from the quantization condition in Eq.  employed to derive the above finite-size effect. The strong-field condition, $| e B | / m_\pi^2 \sim 1$, in conjunction with the field quantization in Eq. , translates into the restriction $$m_\pi L \sim \frac{2 \pi N_\Phi}{a \, m_\pi} \sim 200 .$$ To arrive at this estimate for the required size of $m_\pi L$, we assume the smallest magnetic flux quantum possible for QCD, $N_\Phi = 3$, which is due to the fractional electric charges of quarks, and take a reasonable lattice spacing of $a = 0.1 \, \texttt{fm}$, with a typical pion mass of $m_\pi = 200 \, \texttt{MeV}$. While the quantization condition in Eq.  has been utilized to arrive at desirably simple expressions for finite volume effects, namely those given in Eqs.  and , lattice practitioners should be wary of their use. Magnetic moments and magnetic polarizabilities do not have unique definitions on a torus, where rotational invariance is lost. As a consequence, finite volume corrections to these quantities must intrinsically depend on the lattice method employed in their calculation. Universal corrections are not possible except in extreme limits: asymptotically large lattices where magnetic fields and momentum transfers are infinitesimal, or finite size lattices with infinitesimal twisted boundary conditions. These extreme limits provide an environment where conventional definitions of moments and polarizabilities can be realized, albeit not necessarily in practice. ### Beyond the Static Limit Having detoured to extreme cases, let us return to the strong-field power counting with the magnetic field quantized according to Eq. . To evaluate the finite volume effects derived above, we must confront the Wilson loops that appear from charged pions winding around the torus. In turn, this requires us to go beyond the static limit in order to handle the coordinate dependence of the effective action. With the addition of the non-relativistic kinetic term, the neutron effective action takes the form $$\begin{aligned} \mathcal{L}_{\text{eff}} &=& N^\dagger \Bigg[ \partial_4 - \frac{\vec{\nabla}_\perp^2}{2 M} + E_0 + V_0 ( \vec{x}_\perp) \notag \\ && \phantom{space} + e B \sigma_3 \left[ E_1 + V_1( \vec{x}_\perp) \right] \Bigg] N ,\end{aligned}$$ where we have projected the neutron onto vanishing $\hat{x}_3$-component of momentum. With the coordinate dependence introduced by finite volume effects, however, the transverse momentum is no longer a good quantum number. Beyond the static limit, the neutron propagator has an eigenfunction expansion in terms of eigenmodes of the above action. Consequently the Fourier projection of neutron correlation functions onto $\vec{P}_\perp = \vec{0}_\perp$ will receive contributions from the entire set of energy eigenstates. In lattice QCD, this complication occurs in addition to hadronic excited-state contributions, which is reminiscent of the Landau level problem for charged particles in magnetic fields, see [@Tiburzi:2012ks]. The effect of finite volume is one of the complicating features for neutron spectroscopy in lattice QCD with external fields. To discuss effects of the coordinate-dependent potentials, we begin by isolating the coordinate dependence with a subtraction of images having $\vec{\nu}_\perp = \vec{0}_\perp$. To this end, we write the terms of the neutron effective action in the form $$\begin{aligned} V_0(\vec{x}_\perp) &=& \Delta M(L) + \Delta E_0(L) + U_0(\vec{x}_\perp), \notag \\ V_1(\vec{x}_\perp) &=& \Delta E_1(L) + U_1(\vec{x}_\perp),\end{aligned}$$ where all of the coordinate dependence has been relegated to the residual potentials, $U_0(\vec{x}_\perp)$ and $U_1(\vec{x}_\perp)$. These potentials arise from charged pions winding around the compact directions transverse to the magnetic field, as such windings produce Wilson loops. In this way, the finite volume corrections $\Delta E_0(L)$ and $\Delta E_1(L)$ arise solely from charged pions winding around the $\hat{x}_3$-direction. We additionally separate the magnetic field independent piece of the spin-independent finite volume effect. This is just the finite volume correction to the neutron mass, $\Delta M(L)$, arising from the charged pion loop. Because the coordinate-independent finite volume effects arise from winding only in the $\hat{x}_3$-direction, they are given by relatively simple expressions $$\begin{aligned} \Delta E_0(L) &=& \frac{g_A^2 m_\pi^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \left( \frac{e B s}{\sinh e B s} -1 \right) \notag \\ && \phantom{space} \times e^{- m_\pi^2 s} [ \vartheta_3 (0, e^{- L^2 / 4 s}) -1] \label{eq:E0} ,\end{aligned}$$ for the spin-independent case, and $$\begin{aligned} \Delta E_1(L) &=& - \frac{g_A^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \frac{e B s}{\sinh e B s} e^{ - m_\pi^2 s} \notag \\ && \phantom{spacingspa} \times [ \vartheta_3 (0, e^{- L^2 / 4 s}) -1] \label{eq:E1} ,\end{aligned}$$ for the spin-dependent case. The Jacobi elliptic-theta functions appearing above are defined in Appendix \[s:A\]. The residual potentials can both be written in the form $$\begin{aligned} U_j (\vec{x}_\perp) &=& \sum_{\vec{\nu} - \{ \vec{0}, \vec{0}_\perp \}} [W^\dagger_1(x_2)]^{\nu_1} [W^\dagger_2(x_1)]^{\nu_2} f_j (\vec{\nu}) \label{eq:Uj} ,\quad\end{aligned}$$ where the coordinate dependence is made explicit through the appearance of Wilson loops. One-dimensional integral representations for these potentials can be found from the expressions collected in Appendix \[s:A\]. These residual potentials isolate the non-trivial coordinate dependence of the neutron effective action. Utilizing the long Euclidean time limit of the neutron two-point correlation function, one will arrive at contributions from the lowest eigenstate of the Hamiltonian $$H = - \frac{\vec{\nabla}_\perp^2}{2 M} + U_0 (\vec{x}_\perp) + e B \sigma_3 \, U_1 ( \vec{x}_\perp) \label{eq:ham} ,$$ which is the non-trivial part of the neutron effective action. The corresponding lowest energy eigenvalue, $\mathcal{E}_0$, we write in terms of spin-independent and spin-dependent contributions $$\mathcal{E}_0 = \lambda_{0,0} (L) + e B \sigma_3 \, \lambda_{1,0} (L) .$$ These eigenvalues lead to additional finite volume corrections to the neutron energy, and hence to the magnetic moment and magnetic polarizability. We will argue, however, that finite volume effects from coordinate-dependent potentials can be treated in perturbation theory, and are suppressed by a power of $p$ because they require neutron recoil. As a result, $\Delta E_0$ and $\Delta E_1$ can be identified as the finite volume corrections to the neutron energy up to recoil corrections. To argue that the effect of residual potentials on the neutron energy is suppressed, we scrutinize the power counting of the effective Hamiltonian in Eq. . The kinetic energy operator represents an $\cO(p^2)$ contribution that ordinarily can be dropped by considering a neutron at rest. The coordinate dependence of the residual potentials, however, forces us to reconsider the Fourier modes of the neutron. The spin-independent residual potential $U_0(\vec{x}_\perp)$ scales as $\cO(p^3)$ while the spin-dependent residual potential $U_1(\vec{x}_\perp)$ scales as $\cO(p)$. The effect of both potentials on the neutron energy is thus at $\cO(p^3)$. Consequently we can consider the unperturbed neutron states as Fourier modes at $\cO(p^2)$ and address the effect of the residual potentials in perturbation theory. Because the unperturbed ground state is a neutron at rest, $\vec{P} = \vec{0}$, we have the vanishing of the unperturbed energy eigenvalue of the Hamiltonian, $\mathcal{E}_0^{(0)} = 0$. The leading correction to the non-degenerate ground-state energy, $\mathcal{E}_0^{(1)}$, scales as $\cO(p^3)$ and is determined from the matrix elements $$\begin{aligned} \lambda_{j,0}^{(1)} &=& \langle \vec{P} = \vec{0} \, | U_j | \vec{P} = \vec{0} \, \rangle \notag \\ &=& \sum_{\vec{\nu} - \{ \vec{0}, \vec{0}_\perp \} } \int_0^L d\vec{x}_\perp e^{ 2 \pi i N_\Phi ( \nu_1 x_2 - \nu_2 x_1 ) / L} f_j (\vec{\nu}) = 0,\notag \\\end{aligned}$$ where in the second line we have used the fact that images with $\vec{\nu}_\perp = \vec{0}_\perp$ have been explicitly removed from the summation over winding number. The vanishing of the ground-state matrix elements has a simple physical interpretation. In the Fourier basis, Wilson loops produced by the $\vec{\nu}$-th image insert momentum $$\vec{q}_\perp = \frac{2 \pi N_\Phi}{L} ( - \nu_2, \nu_1),$$ into the matrix elements of residual potentials. Thus all forward matrix elements of the perturbation vanish. One must look to degenerate Fourier modes to arrive at a non-vanishing first-order perturbation that scales as $\cO(p^3)$ in our power counting. As a consequence, the ground state energy $\mathcal{E}_0$ has finite volume corrections arising from the residual coordinate-dependent potentials at $\cO(p^4)$, which is beyond the order we are working. \ \ To assess the size of finite volume corrections on the ground-state neutron, we first consider the weak-field regime, $|e B| / m_\pi^2 \ll 1$. In this regime, we can evaluate the $\cO(p^3)$ finite volume effect on the spin-dependent neutron energy $\Delta E_1 (L)$ appearing in Eq.  at vanishing magnetic field. Notice all of the topological effects that complicate taking the weak-field limit reside in the coordinate-dependent potentials that affect the ground-state energy at $\cO(p^4)$. The leading finite-volume correction to the magnetic moment can easily be deduced by performing the sum over winding number. We arrive at the expression $$\Delta \kappa_n (L) = \frac{g_A^2 M}{2 \pi f^2 L} \log ( 1 - e^{ - m_\pi L} ) \label{eq:kappa3} .$$ Focusing on the spin-independent contribution, we can write $$\Delta E_0 (L) = - \frac{1}{2} 4 \pi \, \Delta \beta_M(L) \, B^2 + \cdots ,$$ in the weak-field regime. This simple expansion is valid due to the lack of topological contributions at this order, and enables us to identify the finite volume correction to the magnetic polarizability as $$\Delta \beta_M (L) = \frac{e^2}{4 \pi} \frac{g_A^2 L}{48 \pi f^2} \frac{e^{ - m_\pi L}}{( 1 - e^{ - m_\pi L} )^2} \left[ 1 + \frac{ 1 - e^{ - m_\pi L}}{m_\pi L} \right] \label{eq:beta3} .$$ The size of finite volume corrections to the neutron magnetic moment and magnetic polarizability is explored in Figs. \[f:kappaNP\] and \[f:beta\]. Included in the plot are the effects from proton-pion intermediate states given in this section, and effects from delta-pion intermediate states. The computation including delta degrees of freedom has been relegated for simplicity to Appendix \[s:D\]. Specifically relevant are the finite volume results shown in Eqs.  and . A practical consideration concerns the applicability of weak-field results to lattice QCD computations. While lattice volumes have increased in physical units leading to smaller values of magnetic fields, the lattice pion masses have decreased towards the physical point. As a result, the expansion parameter $ e B / m_\pi^2$ has roughly remained the same. The weak-field results determined above are thus contrasted with the more realistic strong-field results, for which $| e B | / m_\pi^2 \sim 1$. The latter results are determined from the energy shifts shown in Eqs.  and , via the relations $$\begin{aligned} \Delta \kappa_n (B, L) &=& 2 M \Delta E_1 (L),\end{aligned}$$ and $$\begin{aligned} \Delta \beta_M (B, L) &=& \Delta E_0 (L) \Big/ \left( - \frac{1}{2} 4 \pi B^2 \right) .\end{aligned}$$ These strong-field generalizations of finite-volume corrections to the magnetic moment and magnetic polarizability properly reduce in the weak-field limit. In order to investigate the finite volume corrections in the strong-field regime, we evaluate the effects using magnetic fields available on the torus, $e B L^2 = 2 \pi N_\Phi$. For reference, the size of the weak-field expansion parameter, $\xi = e B / m_\pi^2$, is shown as a function of lattice size $L$ in Fig. \[f:xi\]. Rather large volumes are needed for strong-field results to apply at the physical pion mass. For this reason, we plot the strong-field finite volume effects in Figs. \[f:kappaNP\] and \[f:beta\] choosing the smallest possible magnetic field for QCD, which requires $N_\Phi = 3$. Doubling the pion mass of course allows one to utilize magnetic fields with up to four times the strength. Despite the size of magnetic fields allowed on small volumes, the finite volume corrections to the magnetic moment remain considerably small. For this reason, we have optimistically plotted the results for the magnetic moment in Fig. \[f:kappaNP\] down to lattice sizes of $L = 3 \, \texttt{fm}$. On the other hand, larger finite-size effects are seen for the magnetic polarizability. As a result, we have opted to be more conservative about the minimum-size lattice for which our results for the spin-independent energy apply, and have limited the plots for the polarizability in Fig. \[f:beta\] to sizes $L > 4 \, \texttt{fm}$. For both spin-independent and spin-dependent finite volume corrections, we see a sizeable reduction in finite volume effects when strong-field results are utilized. This signals a breakdown of the weak-field expansion for the lattice sizes and pion masses considered. The need to treat the magnetic field non-perturbatively is further evidenced by Fig. \[f:xi\]. ### Asymptotic Volume Limit We entertain a final limit characterized by the possibility to obtain analytic results for the neutron wave functions. This limit is that of asymptotically large volumes. In the large volume limit, we can restrict our attention to sectors of minimal winding number, with neglected terms being exponentially suppressed. In this limit, the magnetic field becomes small for fixed magnetic flux quantum, $e B L^2 = 2 \pi N_\Phi$, so we can safely assume that the weak-field limit, $| e B | / m_\pi^2 \ll 1$, will additionally be satisfied. For the neutron effective action, we focus exclusively on the coordinate dependence, and take into account contributions from the sectors with winding number $\vec{\nu} = ( \pm 1, 0, 0)$ and $\vec{\nu} = ( 0, \pm 1, 0)$. In the weak-field limit, only terms from the spin-independent potential in Eq.  survive at leading order. These images produce the potential $$\begin{aligned} \cW (\vec{x}_\perp) = \cW \left[ \cos ( e B L x_1) + \cos ( e B L x_2) \right] ,\end{aligned}$$ with the overall strength of the potential set by the parameter $$\begin{aligned} \cW &=& \frac{g_A^2 m_\pi^2}{4 \pi f^2 L} e^{ - m_\pi L} ,\end{aligned}$$ which scales as $\cO(p^3)$ in the power counting. Further images yield contributions that are exponentially suppressed. As a consequence, the transverse motion of the neutron is determined by the effective Hamiltonian, $H_\text{eff} = - \frac{\vec{\nabla}_\perp^2}{2 M} + \cW ( \vec{x}_\perp)$, and the solutions factorize into products of solutions to the one-dimensional Hamiltonian, $$H = - \frac{1}{2 M} \frac{d^2}{d x^2} + \cW \cos ( e B L x) \label{eq:M}.$$ The corresponding Schrödinger equation that determines the energy eigenvalues is the Mathieu differential equation, which has the canonical form $$\left[\frac{d^2}{dv^2} + \alpha - 2 q \cos ( 2 v ) \right] \psi(v) = 0 .$$ Translating the parameters of our problem into Mathieu’s equation, we have the transverse coordinates given by $\vec{v}_\perp = \pi N_\Phi \vec{x}_\perp / L$, and $q = M L^2 \cW / \pi^2 N_\Phi^2$. Given the periodicity of neutron interpolating operators in lattice QCD, and the form of the potential in Eq. , we must seek solutions that are $\mathbb{Z}_{N_\Phi}$ translationally invariant, which corresponds to $\pi$-periodicity in the coordinates $\vec{v}_\perp$. The ground-state solution is thus characterized by the wave function $$\psi_0 (\vec{x}_\perp ) = N \texttt{ce}_0 \left( v_1, q \right) \texttt{ce}_0 \left( v_2, q \right) ,$$ where $\texttt{ce}_r (v, q)$ denotes a Mathieu function, and $N$ is an overall normalization factor. The corresponding ground-state energy is given by $$E(L) = \frac{\pi^2 N_\Phi^2}{M L^2} \, \texttt{a}_0 \left( q \right) ,$$ where $\texttt{a}_r (q)$ denotes a Mathieu characteristic. Using the series expansion, $\texttt{a}_0(q) = - \frac{1}{2} q^2 + \cO(q^4)$, we see the energy eigenvalue can be written as $$E(L) = - \frac{1}{2} M L^2 \frac{\cW^2}{\pi^2 N_\Phi^2} + \cdots .$$ The leading term in this expansion scales as $\cO(p^4)$, which is consistent with our general analysis above. Conclusion ---------- Above we consider chiral dynamics of the neutron in a strong external magnetic field, including a derivation of finite volume corrections relevant for uniform magnetic fields on a torus. In order to perform these computations, a number of theoretical ingredients are established. Magnetic periodicity is enforced on charged scalar Green’s functions, resulting in the coordinate-space propagator in Eq. . The features of this propagator allow us to anticipate many general properties of the finite volume computation. In particular, Wilson loops lead to the breaking of translational invariance down to the discrete $\mathbb{Z}_{N_\Phi}$ magnetic translation group. These Wilson loops must accompany the winding of charged particles around the torus due to gauge invariance. Furthermore, a phase factor in the infinite volume propagator, Eq. , underlies the appearance of the Aharonov-Bohm effect in our later computations. We are careful to show that the static charged fermion propagator is unaffected by magnetic periodic images. The charged scalar and static fermion propagators are utilized to compute chiral corrections to the correlation functions of the neutron. This is accomplished for the first time using a combination of heavy nucleon chiral perturbation theory, $p$-regime chiral perturbation theory for finite volume effects, and strong-field chiral counting to handle non-perturbative effects from the magnetic field. The finite volume computations are performed using a direct coordinate-space method that circumvents redundant application of the Poisson formula. The coordinate-space computation is applied to a simple example to recover the familiar finite volume corrections to the nucleon mass. After this demonstration, computations in magnetic fields are performed. In infinite volume, the dependence of the neutron energy on the magnetic field is considered for strong magnetic fields satisfying the condition $| e B | / m_\pi^2 \gtrsim 1$. In Fig. \[f:E0\], we show the spin-independent neutron energy shift in a strong magnetic field. While the perturbative expansion in powers of the magnetic field breaks down, our results demonstrate that a model incorporating the leading perturbative effect from the magnetic polarizability may still capture the non-perturbative behavior, however, the polarizability extracted from a perturbative model fit will differ from the true value. Comparison of perturbative and non-perturbative magnetic field fits will be necessary to assess this effect. A very similar observation is made about the neutron Zeeman splitting shown in Fig. \[f:E1\], which shows surprisingly linear behavior with respect to the magnetic field well beyond the perturbative magnetic field regime. In this case, higher-order chiral corrections are needed from the effective theory to verify the behavior at the largest magnetic fields shown in the plot. It would be quite fortuitous if this behavior persists. Finite volume corrections are derived for the neutron in the form of terms appearing in the neutron effective action, which arises after integrating out charged pions. The proton-pion (and delta-pion) intermediate states lead to coordinate dependence of the neutron effective action. Winding of charged pions around the torus must be accompanied by Wilson loops, and these gauge invariant quantities reflect the non-trivial holonomies of the torus. A topological phase can be acquired from pions winding in the plane transverse to the field, which is a manifestation of the Aharonov-Bohm effect. Exposing these features requires non-perturbative treatment of the magnetic field at finite volume. Consequently these features are not susceptible to perturbative expansion in powers of the magnetic field except in rather extreme limits. Considering such limits, we obtain formulae for the finite volume effects on the magnetic moment and magnetic polarizability. The former result agrees with that derived using twisted boundary conditions in the limit of a vanishing twist angle. We stress that the magnetic moments and magnetic polarizabilities do not have unique definitions on a torus. As such they are subject to finite volume effects that depend on the method employed in their determination. In the case of three- and four-point function computations, the breaking of rotational invariance is a complication that has often been overlooked in finite volume. In the external field problem, topological effects from holonomies and Aharonov-Bohm phases are the new features at finite volume. To evaluate the effect of finite volume on the neutron correlation function, we are led beyond the static approximation. The coordinate dependence of the neutron potential requires that the neutron kinetic term be retained. While Fourier modes no longer exactly describe the propagation of the neutron in finite volume, the finite volume corrections are suppressed relative to the kinetic term. Standard Rayleigh-Schrödinger perturbation theory can be employed to determine the effect on neutron energy levels. Focusing on the ground-state neutron energy, we determine the finite volume corrections, and uncover that the coordinate-dependent potentials do not affect results at the leading order of perturbation theory. As a result, the leading finite volume corrections for neutron magnetic moments and polarizabilities originate solely from pions winding in the direction of the magnetic field. Physically we might expect suppression of winding transverse to the field direction because the magnetic field provides a confining length scale, $\sim | e B |^{-1/2}$, that is characteristic of Landau levels. In the evaluation of finite volume effects, Figs. \[f:kappaNP\] and \[f:beta\], we consider results for both the weak-field limit and for strong magnetic fields. The latter are readily encountered on present-day lattices, and we see non-perturbative magnetic field effects decrease the size of finite volume corrections. Various further investigations have been suggested throughout. Here we describe a few that have yet to be mentioned. Central to our discussion has been the static approximation for the neutron. As the neutron effective potential on a torus is coordinate dependent, the static approximation must be reconsidered. Away from the static limit, strong magnetic fields in infinite volume additionally probe hadron structure, and allow for further coordinate dependence in neutron correlation functions beyond what is considered here. Such dependence can be investigated in the context of chiral perturbation theory with strong-field power counting. A study of this coordinate dependence in conjunction with the coordinate dependence exposed above may help in the construction of better neutron interpolating operators for lattice computations. Finally our computation has been limited to correlation functions of the neutron only in the interest of simplicity. Proton correlation functions can be treated in an entirely similar manner to that developed in this work. In the case of the proton, one must understand the modifications to Landau levels that arise both from the torus, and from charged pions winding around the torus. Nevertheless, we imagine that the methodology of strong-field chiral perturbation theory explored here will be of use to lattice QCD computations of hadrons in electromagnetic fields. Work supported in part by a joint City College of New York–RIKEN BNL Research Center fellowship, a grant from the Professional Staff Congress of the CUNY, the Alfred P. Sloan foundation through a CUNY-JFRASE award, and by the U. S. National Science Foundation, under Grant No. PHY$12$-$05778$. Including the Delta Resonance {#s:D} ============================= Of the low-lying baryon resonances, the nearby delta resonance gives rise to important virtual corrections to nucleon observables. The delta-nucleon mass splitting, $\Delta = M_\Delta - M_N$, introduces a new dimensionful parameter, and it is common to treat it as $\Delta / M \sim p$ in the power counting, see [@Hemmert:1997ye]. The $\cO(p)$ chiral Lagrange density describing the delta resonance and its interactions with the nucleon is given by $$\begin{aligned} \cL &=& \ol T_\mu \left( D_4 + \Delta \right) T_\mu + g_{\Delta N} \left[ \ol T_\mu \cA_\mu N + \ol N \cA_\mu T_\mu \right], \label{eq:T}\end{aligned}$$ where the delta-nucleon axial coupling is $g_{\Delta N} \sim 1.5$. The magnetic polarizability receives a contribution from the nucleon-delta magnetic dipole transition operator. As the leading contribution is from a tree-level diagram with intermediate-state delta, there is no volume correction associated with it. To derive the leading finite volume corrections to the neutron two-point function, we evaluate the four sunset diagrams shown in Fig. \[f:sunset\], where the intermediate-state baryon is now taken to be a delta resonance. From the Lagrange density in Eq. , we see the static delta propagator is given by $$[D_{FV} (x',x)]_{ij} = \delta_{L}^{(3)} (\vec{x} \, ' - \vec{x} \, ) \theta( x'_4 - x_4) e^{ - \Delta ( x'_4 - x_4)} \cP_{ij} \label{eq:D1.5} ,$$ where the spin-$\frac{3}{2}$ projection matrix has the form $\cP_{ij} = \frac{2}{3} \delta_{ij} - \frac{i}{3} \epsilon_{ijk} \sigma_k$. The proper-time formulation makes accounting for delta-resonance contributions straightforward. By virtue of the propagator, Eq. , the effect of the mass splitting $\Delta$ is to alter the relative-time integration used to project onto $P_4 = 0$, where we encounter the integrals $$\begin{aligned} \int_0^\infty dT \, e^{ - \frac{T^2}{4 s} - \Delta \, T} &=& \sqrt{\pi s} \, e^{\Delta^2 s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right), \qquad \int_0^\infty dT \, \frac{T}{2s} e^{ - \frac{T^2}{4 s} - \Delta \, T} = 1 - \Delta \sqrt{\pi s} \, e^{\Delta^2 s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) .\end{aligned}$$ Here $\operatorname{Erfc}(x)$ is the complement of the standard error function, for which $\operatorname{Erfc}(0) = 1$, and the second integral arises from utilizing an integration by parts on the relative-time integral. Such contribution only occurs for the spin-independent part of the sunset diagrams. Combining the result of the relative-time integration with the spin and flavor algebra, we have the delta-resonance contributions to the infinite volume quantities: the spin-independent energy $$\begin{aligned} E_0^{(\Delta)} &=& \frac{8}{9} \frac{g_{\D N}^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \left( \frac{e B s}{\sinh e B s} - 1 \right) e^{ - m_\pi^2 s} \left[ \frac{\Delta}{\sqrt{\pi s}} + \left( m_\pi^2 - \Delta^2 \right) e^{ \Delta^2 s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \right], $$ so that the sum $E_0 + E_0^{(\D)}$ takes into account intermediate-state nucleons and deltas, with $E_0$ is given in Eq. ; and the spin-dependent energy $$\begin{aligned} E_1^{(\Delta)} &=& \frac{2}{9} \frac{g_{\D N}^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \left( \frac{e B s}{\sinh e B s} - 1 \right) e^{ - ( m_\pi^2 - \Delta^2) s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) ,\end{aligned}$$ so that similarly $E_1 + E_1^{(\D)}$ takes into account both intermediate-state baryons, with $E_1$ given in Eq. . The differing numerical constants for spin-independent and spin-dependent contributions stem from the underlying spin $\times$ flavor $\times$ charge factors of intermediate states. In each case, we have subtracted the divergent zero-field results. The dimensionally regulated zero-field results produce the charged pion loop corrections to the neutron mass and magnetic moment, in the spin-independent and spin-dependent cases, respectively. Furthermore, the finite $\cO(B^2)$ term in the spin-independent energy leads to the correct contribution to the neutron’s magnetic polarizability, originally determined in [@Hemmert:1996rw]. The finite volume corrections with intermediate-state delta-resonance contributions can be similarly found. We can write their contribution in the form $$\begin{aligned} V_0^{(\D)} (\vec{x}_\perp) &=& \frac{8}{9} \frac{g_{\D N}^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \frac{e B s}{\sinh e B s} e^{ - m_\pi^2 s} \Bigg\{ \left[ \frac{\Delta}{\sqrt{\pi s}} + \left( m_\pi^2 - \Delta^2 \right) e^{ \Delta^2 s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \right] \mathcal{U}^{(A)} (s, B, \vec{x}_\perp) \notag \\ && \phantom{spacingspacing} + \frac{1}{2} ( e B L)^2 e^{ \Delta^2 s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \mathcal{U}^{(B)} (s, B, \vec{x}_\perp) \Bigg\} \label{eq:V0Delta} ,\end{aligned}$$ for the spin-independent part, and $$\begin{aligned} V_1^{(\D)} (\vec{x}_\perp) &=& \frac{2}{9} \frac{g_{\D N}^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \frac{e B s}{\sinh e B s} e^{ - ( m_\pi^2 - \Delta^2) s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \mathcal{U}_1 (s, B, \vec{x}_\perp) ,\label{eq:V1Delta}\end{aligned}$$ for the spin-dependent part. The functions $\mathcal{U}^{(A)} (s, B, \vec{x}_\perp)$, $\mathcal{U}^{(B)} (s, B, \vec{x}_\perp)$, and $\mathcal{U}_1 (s, B, \vec{x}_\perp)$ are given below in Eqs. –. In the main text, we present expressions for finite volume corrections to the neutron magnetic moment and magnetic polarizability arising from proton-pion intermediate states. Here we quote the analogous expressions taking into account contributions from delta-pion intermediate states. For the magnetic moment defined by taking extreme limits, we have $$\begin{aligned} \Delta \kappa_n^{(\Delta)} (L) = \frac{4}{9} \frac{g_{\D N}^2 M \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} e^{ - ( m_\pi^2 - \Delta^2) s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \mathcal{U}_1 (s, 0, 0) \label{eq:FVmomentNDelta} ,\end{aligned}$$ where $\mathcal{U}_1 (s, 0, 0)$ is defined without dependence on the magnetic flux quantum $N_\Phi$, and appears in Eq. . For the magnetic polarizability defined by taking extreme limits, we have $$\begin{aligned} \Delta \beta_M^{(\Delta)} (L) &=& \frac{e^2}{4 \pi} \frac{8}{9} \frac{g_{\D N}^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} e^{ - m_\pi^2 s} \Bigg\{ \frac{1}{3} s^2 \left[ \frac{\Delta}{\sqrt{\pi s}} + \left( m_\pi^2 - \Delta^2 \right) e^{ \Delta^2 s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \right] \mathcal{U}^{(A)} (s, 0, 0) \notag \\ && \phantom{spacingspace} + \frac{1}{6} s L^2 \left[ \frac{\Delta}{\sqrt{\pi s}} + \left( m_\pi^2 - \Delta^2 - \frac{6}{s} \right) e^{ \Delta^2 s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \right] \mathcal{U}^{(B)} (s, 0, 0) \Bigg\} \label{eq:betaFVDelta} ,\end{aligned}$$ with expressions for the functions $\mathcal{U}^{(A)} (s, 0, 0)$ and $\mathcal{U}^{(B)} (s, 0, 0)$ given in Eq. . In strong-field power counting, we obtain simple expressions for spin-independent and spin-dependent finite volume effects on the ground-state neutron. The simple expressions arise because non-vanishing corrections from coordinate-dependent potentials in the neutron effective action occur at $\cO(p^4)$. The argument applies equally for contributions arising from delta-pion intermediate states. For strong-field corrections to the neutron ground-state energy in finite volume, we thus have the results $$\begin{aligned} \Delta \kappa_n^{(\Delta)} (B, L) = - \frac{4}{9} \frac{g_{\D N}^2 M \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \frac{e B s}{\sinh e B s} e^{ - ( m_\pi^2 - \Delta^2) s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \left[ \vartheta_3( 0, e^{ - \frac{L^2}{4s}}) - 1 \right] \label{eq:FVmomentNDeltaFV} ,\end{aligned}$$ from delta-pion contributions to the magnetic moment, and $$\begin{aligned} \Delta \beta_M^{(\Delta)} (B, L) &=& \frac{8}{9} \frac{g_{\D N}^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \left( \frac{e B s}{\sinh e B s} - 1 \right) e^{ - m_\pi^2 s} \left[ \frac{\Delta}{\sqrt{\pi s}} + \left( m_\pi^2 - \Delta^2 \right) e^{ \Delta^2 s} \operatorname{Erfc}\left( \Delta \sqrt{s} \right) \right] \notag \\ && \phantom{spacingspacingspacing} \times \left[ \vartheta_3( 0, e^{ - \frac{L^2}{4s}}) - 1 \right] \Big/ \left( - \frac{1}{2} 4 \pi B^2 \right) \label{eq:betaFVDeltaFV} ,\end{aligned}$$ for the magnetic polarizability. Jacobi elliptic-theta functions appear in this expressions, and are defined in Appendix \[s:A\]. Jacobi Elliptic-Theta Functions {#s:A} =============================== Displayed below are one-dimensional, proper-time integral representations for the finite volume effects. The sums over winding number are cast in terms of three out of the four Jacobi elliptic-theta functions, namely $$\vartheta_2(z,q) = 2 \sum_{n = 0}^\infty q^{(n + \frac{1}{2})^2} \cos [( 2 n +1) z ] , \quad \vartheta_3(z,q) = 1 + 2 \sum_{n = 0}^\infty q^{n^2} \cos ( 2 n z ) , \quad \vartheta_4(z,q) = 1 + 2 \sum_{n = 0}^\infty (-)^n q^{n^2} \cos ( 2 n z ) .$$ As is customary, primes denote derivatives with respect to the first argument, $\vartheta'_j (z,q) = \frac{d}{dz} \vartheta_j ( z, q)$. For both the spin-independent $(j = 0)$ and spin-dependent $(j = 1)$ neutron potentials, we write $$\begin{aligned} V_j(\vec{x}_\perp) &=& \frac{g_A^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} \frac{e B s}{\sinh e B s} e^{ - m_\pi^2 s} \, \mathcal{U}_j (s, B, \vec{x}_\perp) .\end{aligned}$$ For both the spin-independent and spin-dependent cases, we separate out two recurrent terms $$\begin{aligned} \mathcal{U}_0 ( s, B, \vec{x}_\perp ) &=& m_\pi^2 \, \mathcal{U}^{(A)} (s , B, \vec{x}_\perp) + \frac{1}{2} ( e B L)^2 \, \mathcal{U}^{(B)} (s , B, \vec{x}_\perp), \\ \mathcal{U}_1 (s, B, \vec{x}_\perp) &=& - \mathcal{U}^{(A)} (s, B, \vec{x}_\perp) + \frac{e B L^2}{2 \tanh e B s} \mathcal{U}^{(B)} (s, B, \vec{x}_\perp) \label{eq:spindepU1} ,\end{aligned}$$ with the ancillary functions $\mathcal{U}^{(A)} (s, B, \vec{x}_\perp)$ and $\mathcal{U}^{(B)} (s, B, \vec{x}_\perp)$ having the definitions $$\begin{aligned} \mathcal{U}^{(A)} (s, B, \vec{x}_\perp) = \begin{cases} \vartheta_3(\frac{1}{2} e B L x_1 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \vartheta_3(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \vartheta_3(0,e^{- \frac{L^2}{4s}}) -1 , & N_\Phi = \text{even} \\ \\ \vartheta_3(e B L x_1 ,e^{- \frac{e B L^2}{\tanh e B s}}) \vartheta_3(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \vartheta_3(0,e^{- \frac{L^2}{4s}}) -1 \\ \phantom{mm} + \vartheta_2(e B L x_1 ,e^{- \frac{e B L^2}{\tanh e B s}}) \vartheta_4(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \vartheta_3(0,e^{- \frac{L^2}{4s}}) , & N_\Phi = \text{odd}, \end{cases} \label{eq:UA}\end{aligned}$$ along with $$\begin{aligned} \mathcal{U}^{(B)} (s, B, \vec{x}_\perp) = \begin{cases} \Big[ - \frac{1}{4} \vartheta''_3(\frac{1}{2} e B L x_1 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \vartheta_3(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \\ \phantom{mm} - \frac{1}{4} \vartheta_3(\frac{1}{2} e B L x_1 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \vartheta''_3(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \Big] \vartheta_3(0,e^{- \frac{L^2}{4s}}) , & N_\Phi = \text{even} \\ \\ \Big[ - \vartheta''_3(e B L x_1 ,e^{- \frac{e B L^2}{\tanh e B s}}) \vartheta_3(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \\ \phantom{mm} - \frac{1}{4} \vartheta_3(e B L x_1 ,e^{- \frac{e B L^2}{\tanh e B s}}) \vartheta''_3(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \\ \phantom{mm} - \vartheta''_2(e B L x_1 ,e^{- \frac{e B L^2}{\tanh e B s}}) \vartheta_4(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \\ \phantom{mm} - \frac{1}{4} \vartheta_2(e B L x_1 ,e^{- \frac{e B L^2}{\tanh e B s}}) \vartheta''_4(\frac{1}{2} e B L x_2 ,e^{- \frac{e B L^2}{4 \tanh e B s}}) \Big] \vartheta_3(0,e^{- \frac{L^2}{4s}}) , & N_\Phi = \text{odd}. \end{cases} \label{eq:UB}\end{aligned}$$ As defined, these finite volume effects include those for vanishing fields, such as the mass renormalization. These field-independent contributions can be removed by subtraction $$\Delta V_j(\vec{x}_\perp) = V_j (\vec{x}_\perp) - \frac{g_A^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} e^{ - m_\pi^2 s} \, \mathcal{U}_j (s, 0, 0) ,$$ where the zero-field functions are considerably simpler; and, are implicitly defined without dependence on the magnetic flux quantum, $N_\Phi \to 0$. Because these finite-volume functions occur frequently, we quote them here for ease of reference: $$\mathcal{U}_0 (s, 0, 0) = m_\pi^2 \, \mathcal{U}^{(A)} (s, 0, 0) , \qquad \mathcal{U}_1 (s, 0, 0) = - \mathcal{U}^{(A)} (s, 0, 0) + \frac{L^2}{2 s} \mathcal{U}^{(B)} (s, 0, 0) \label{eq:U0s} ,$$ with $$\begin{aligned} \mathcal{U}^{(A)} (s, 0, 0) = \vartheta_3(0,e^{- \frac{L^2}{4s}})^3 -1 , \quad \text{and} \quad \mathcal{U}^{(B)} (s, 0, 0) = - \frac{1}{2} \vartheta''_3(0, e^{- \frac{L^2}{4 s}}) \vartheta_3(0, e^{- \frac{L^2}{4s}})^2 \label{eq:UAB0s} .\end{aligned}$$ An alternate way to write the finite volume corrections to the magnetic moment and magnetic polarizability shown in Eqs.  and , respectively, is in terms of the ancillary functions $\mathcal{U}^{(A)} (s, 0, 0)$ and $\mathcal{U}^{(B)} (s, 0, 0)$, namely $$\begin{aligned} \Delta \kappa_n (L) = \frac{2 g_A^2 M \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} e^{ - m_\pi^2 s} \, \mathcal{U}_1 (s,0,0) ,\end{aligned}$$ for the magnetic moment, and $$\begin{aligned} \Delta \beta_M (L) &=& \frac{e^2}{4 \pi} \frac{g_A^2 \sqrt{\pi}}{(4 \pi f)^2} \int_0^\infty \frac{ds}{s^{3/2}} e^{ - m_\pi^2 s} \left[ \frac{1}{3} m_\pi^2 s^2 \, \mathcal{U}^{(A)} (s, 0, 0) - L^2 \left( 1 - \frac{1}{6} m_\pi^2 s \right) \mathcal{U}^{(B)} (s,0,0) \right] ,\end{aligned}$$ for the magnetic polarizability. Recall that these results only apply in extreme limits. Equivalence of Finite Volume Corrections to the Magnetic Moment {#s:B} =============================================================== Here we demonstrate the equivalence of finite volume corrections to the nucleon magnetic moment computed in two different ways. The first way employs an external magnetic field, and is the result derived above in the limit of an extremely weak magnetic field, namely that appearing in Eq. . The second way employs isospin twisted boundary conditions in the computation of the nucleon isovector magnetic form factor, for which the result can be found in [@Tiburzi:2006px]. To compare these two methods, we take the limit of zero twist angle of the latter computation. In order to compare results, we must be careful to multiply those obtained in [@Tiburzi:2006px] by a factor of $- \frac{1}{2}$ in order to convert loop contributions for the isovector magnetic moment into loop contributions for the neutron magnetic moment. In this context, it is useful to recall that the one-loop chiral corrections to the nucleon magnetic moment are exactly isovector. The isovector spin-flip matrix element between nucleon states is sensitive to the Pauli form factor. For a relative twist angle of $\theta$ between up and down quarks, this matrix element is non-vanishing for zero Fourier momentum transfer. In finite volume, the result $$\begin{aligned} F^{v}_2 = F_2^v (q^2) + \Delta F_2^v (L) ,\end{aligned}$$ was obtained in [@Tiburzi:2006px], where the effective momentum transfer $q$ arises from the twist angle $\theta$, through the simple relation $q = \theta / L$. In the limit, $\theta \to 0$, the Pauli form factor reduces to the isovector magnetic moment, $F_2^v (0) = \kappa^v$. The contribution to the matrix element $\Delta F_2^v (L)$ arises from finite volume effects, and was determined to be $$\begin{aligned} \Delta F_2^v (L) &=& \frac{3 g_A^2 M}{4 \pi^2 f^2} \int_0^1 dx \, \cL_{33} \Big( m_\pi P_\pi (x,q^2), x q, 0 \Big) - \frac{g_A^2 M}{4 \pi^2 f^2} \frac{ \cK_2 ( m_\pi, q, 0)}{q} \label{eq:TWBCs} .\end{aligned}$$ For simplicity, we drop contributions with intermediate-state delta resonances. One can also verify that the finite volume effects which include delta resonances and are computed in a background magnetic field agree with the zero twist angle limit of the expressions derived in [@Tiburzi:2006px]. The kinematic function appearing in the first term in Eq.  is given by $P_\pi (x, q^2) = [ 1 + x ( 1-x) q^2 / m_\pi^2]^{1/2}$, while the finite volume functions $\cL_{33}$ and $\cK_2$ were defined to be $$\begin{aligned} \cL_{33} (m, q, 0) &=& \frac{\sqrt{\pi}}{3} \int_0^\infty \frac{ds}{s^{3/2}} e^{- m^2 s} \Bigg[ \vartheta_3 ( 0, e^{ - \frac{L^2}{4 s}} )^2 \vartheta_3 (\frac{q L}{2}, e^{ - \frac{L^2}{4 s}} ) - 1 + \frac{L^2}{8 s} \vartheta''_3 ( 0, e^{ - \frac{L^2}{4 s}} ) \vartheta_3 ( 0, e^{ - \frac{L^2}{4 s}} ) \vartheta_3 (\frac{q L}{2}, e^{ - \frac{L^2}{4 s}} ) \Bigg] \notag \\ \\ \mathcal{K}_2 (m, q, 0) &=& - \frac{\sqrt{\pi}L}{4} \int_0^\infty \frac{ds}{s^{5/2}} e^{ - m^2 s} \vartheta'_3 (\frac{q L}{2}, e^{ - \frac{L^2}{4 s}} ) \vartheta_3 ( 0, e^{ - \frac{L^2}{4 s}} )^2 .\end{aligned}$$ The Jacobi elliptic-theta functions appearing in these definitions are given in Appendix \[s:A\]. To take the $\theta \to 0$ limit of the finite volume correction $\Delta F_2^v (L)$, we observe that the first term in Eq.  can merely be evaluated at $q =0$, while the second term requires a derivative, $\lim_{q \to 0} \cK_2 ( m, q, 0) / q = \frac{\partial}{\partial q} \cK_2 ( m, 0, 0)$. Upon taking the limit, we thus find $$\begin{aligned} \Delta \kappa^v (L) \equiv \lim_{\theta \to 0} \Delta F_2^v ( L) = - \frac{g_A^2 \sqrt{\pi} M}{4 \pi^2 f^2} \int_0^\infty \frac{ds}{s^{3/2}} e^{ - m_\pi^2 s} \, \mathcal{U}_1 ( s, 0 ,0) = - \frac{g_A^2 M m_\pi}{3 \pi f^2} \sum_{\vec{\nu} - \{\vec{0} \}} \left[ 1 - \frac{1}{2 |\vec{\nu}| m_\pi L} \right] e^ {- |\vec{\nu}| m_\pi L} ,\end{aligned}$$ where $\mathcal{U}_1 (s, 0, 0)$ is the function appearing in the context of finite volume corrections to the magnetic moment, and is given in Eq. . The second equality arises from performing the proper-time integration. Multiplying this finite volume correction to the isovector magnetic moment by $- \frac{1}{2}$ leads to the finite volume correction to the neutron magnetic moment, which agrees with Eq. . [^1]: Dirichlet boundary conditions for quarks have been proposed to overcome the difficulties associated with uniform electromagnetic fields on a torus. This choice introduces finite-size effects that are unfortunately rather complicated to address, see [@Tiburzi:2013vza] for a discussion of one such effect. [^2]: Recently an analysis of finite volume corrections to the neutron magnetic polarizability appeared [@Hall:2013dva], however, their approach does not account for effects addressed in our work. [^3]: One can also consider the uniform magnetic field to arise from gauge potentials having the form $A_\mu = \Big( - B (x_2 - X) , 0 , 0, 0 \Big)$, for example, where $X$ is a constant. In finite volume, results at $X = 0$ are no longer gauge equivalent to those at $X \neq 0$. This inequivalence is best expressed in terms of gauge holonomies. In this example, the Wilson loop $W_1(x_2)$ now must appear as $W_1(x_2) = \exp [- i Q B L ( x_2 - X) ]$. As a majority of our results will be expressed in terms of the Wilson loops, it is largely trivial to account for alternate ways to implement the external field. [^4]: Away from this limit, one must deal with the first-order correction which is linear in $P_4$. This leads to the wavefunction renormalization factor, but is not needed to the order we are working in the chiral expansion as it gives rise to terms at $\cO(p^4)$. [^5]: Unfortunately there are mismatched factors of $2$ in Eqs. (95) and (96) of [@Tiburzi:2008ma]. These transcription errors arose from incomplete rescaling of the proper-time by a factor of $2$ in Eqs. (A.5) and (A.7). None of the other equations suffer from this error. [^6]: For reference, the result claimed in [@Beane:2004tw] for the neutron magnetic moment is given by $\frac{g_A^2 M m_\pi}{12 \pi f^2} \sum_{\vec{\nu} - \{\vec{0} \}} \left[ 1 - \frac{2}{|\vec{\nu} \, | m_\pi L} \right] e^ {- |\vec{\nu} \, | m_\pi L}$.

ArXiv

--- author: - | [Farid Nouioua]{}\ Laboratoire d’informatique de Paris-Nord\ \ \ bibliography: - '/export/home/LIPN-SF1/users/RCLN/nouiouaf/articlelatex/biblio.bib' title: '**Norm Based Causal Reasoning in Textual Corpus**' --- Motivation ========== Truth based entailments are not sufficient for a good comprehension of NL. In fact, it can not deduce implicit information necessary to understand a text. On the other hand, norm based entailments are able to reach this goal. Let us consider this text [@Kayser:04a][@Kayser:04b]: “the vehicle in front of me braked”. Using a truth based approach; we can obtain all the logical consequences of a formula such as: $(\exists v, t)~ Vehicle (v) \wedge Instant (t) \wedge In-Front-Of (v,~ 'me',~ t) \wedge break (v, ~t).$ While norms provide further conclusions like: $v$ and me were in the same direction, no vehicle was between $v$ and me, I had to brake when $v$ braked …\ This idea was behind the development of Frames [@Minsky:75] and Scripts [@Schank:77][@Schank:79] in the 70’s. But these theories are not formalized enough and their adaptation to new situations is far from being obvious.\ Actually, no repository of norms is available for a given domain. Moreover, norms are seldom made explicit in texts, because as Schank noticed, texts do not describe the normal course of events but focus rather on the description of abnormal situations. The motivation of the present work is to extract norms by detecting their violations in the texts.\ We are working on a corpus of 60 texts describing car crashes. For each text, we are searching the cause of the accident as perceived by a standard reader. We hypothesize that the perceived cause of an abnormal event is the violation of a norm (anomaly). Among all the anomalies evoked by a text, one of them is considered as ’primary’. It represents the most plausible cause of the accident. The other anomalies result from the primary one and are called derived anomalies. The representation language =========================== In this work, we use a first order logic (FOL) representation language which takes into account modalities, time and non monotonicity.We only give here its main principles (see [@Kayser:04b] for details):\ In order to quantify over properties, we use the technique of reification commonly used in AI: a binary predicate $P(X, ~Y)$ is written: $Holds(P, ~X,~ Y)$[^1].\ We decompose the scene into a succession of discrete states. Each one is characterized by a set of literals keeping a stable truth value. Thus, what we represent explicitly is the linear time of the events, as they really occurred. The form of the predicates is: $Holds (P, A, t)$ where $P$ is a property, $A$ is an agent and $t$ is a state number.\ In addition to truth-values, the texts introduce modalities. The technique of reification enables us to treat modalities as first order predicates. In our work, we use two main modalities: The first one is a kind of necessity. It expresses agent duties. $Must (P, ~A,~ t)$ means that at state $t$, agent $A$ must reach the property $P$. The second one is a kind of possibility. It expresses the capacities of the agents. $Able-To (P, ~A,~ T)$ means that at state $t$, agent $A$ is able to reach the property $P$.\ Inference rules are written in Reiter’s default logic [@Reiter:80]. Material implications are written $(A \rightarrow B)$, normal defaults $\frac{A:B}{B}$ are written for short $A : B$ and semi-normal ones $\frac{A:B\wedge C}{C}$ are written A : B\[C\].\ A primary anomaly comes under one of the two following forms: $Must(P, ~A,~ t) \wedge Able-To(P,~ A,~ t) \wedge Holds(P',~ A,~ t+1) \wedge Incompatible(P,~ P')\rightarrow An$ $Holds(Combine(Disruptive\_Factor,~ C),~ A,~ t) \rightarrow An$ Derived anomalies are expressed by: $Must(P,~ A,~ t) \wedge \neg Able-To(P,~ A,~ t) \wedge Holds(P',~ A,~ t+1) \wedge Incompatible(P,~ P')\rightarrow D-An$ From the text to the cause of the accident ========================================== From the text to the semantic predicates ---------------------------------------- First of all, a tagger and a syntactico-semantic analysis are applied to the text. The result of this step is a set of linguistic relations between relevant words of the text. The list of this type of relations, called linguistic predicates, is very short, namely:\ $Subject(V,~ N), Object(V,~ N)$ : $N$ is the subject (resp. object) of the verb $V$.\ $Qualif-N(N,~ A), Qualif-V(V,~ A)$ : $A$ is a qualification for the noun $N$ (resp. the verb $V$). It is the case for example for adjectives and adverbs.\ $Compl-N(X, ~N,~ Z), Compl-V(X,~ V,~ Z)$ : $Z$ is a complement for the noun $N$ (resp. the verb $V$). It is introduced by $X$.\ $Support(X,~ Y)$ : $X$ is a support for $Y$. For instance in “A est venu heurter B”(A comes to run up against B), we have the relation $Support(venir,~ heurter)$. Let us consider the text :\ $"J'\acute{e}tais~\grave{a}~ l'arr\hat{e}t~ au~ feu~ rouge~ lorsque~ le~ v\acute{e}hicule~ A~ m'a~ percut\acute{e}~\grave{a}~ l'arri\grave{e}re"~$(I was stopped on a red light, when vehicle A bumped on me with the back). We obtain : $Subject(\hat{e}tre, J'),~ Compl-V(\grave{a}, \hat{e}tre, arr\hat{e}t),~ Compl-N(\grave{a}, arr\hat{e}t, feu),~ Qualif-N(feu, rouge),$ $Compl-V(lorsque,~ \hat{e}tre,~percuter),~ Subject(percuter,~v\acute{e}hicule),~Qualif-N(v\acute{e}hicule,~ A),$ $Object(percuter,~ m'),~ Compl-V(\grave{a},~ percuter, ~arri\grave{e}re).$ After that, a non monotonic linguistic reasoning process transforms the linguistic predicates into a set of semantic predicates. For our example we obtain here: $Holds(Stop,~B,~1), ~Holds(Combine(Light,~Red),~A,~1),$ $Holds(Combine(Bump,~B),~A,~2),~Holds(Combine(Shock,~ back),~B,~2)$. From the semantic predicates to the kernel ------------------------------------------ Semantic predicates are supposed to represent the explicit semantic content of the text. The semantic reasoning process uses inference rules to enrich the set of semantic predicates extracted from a text by adding further implicit conclusions. The inference rules are based on our common knowledge about the norms of the domain of car crashes.\ The kernel contains six (reified) predicates: $Stop,~Control,~Run\_Slowly\_Enough,~Start,~Move\_Back ,~ Combine(Disruptive\_Factor,C)$. Computing extensions of a first order semi-normal default theory is intractable in the general case. To overcome this difficulty, one has to consider sub-sets of the theory in which some constraints must be satisfied. In the present work, predicates and rules are designed so that they can be stratified i.e. organized in layers such that the derivation of a predicate belonging to a given layer depends only on the upper layers. Formally, the stratification constraints is verified if : (L(P) denotes the number of the layer containing the predicate P).\ Each implication $A \rightarrow B$, (resp. a normal default $A : B$) verifies : $L(A) ~\geq ~L(B)$.\ Semi normal defaults $A : B [C]$ verify : $L(A)~ \geq~ L(B)$ and $L(C)~ > ~Max (L(A),~ L(B))$\ Notice that rules belonging to the layer of number L(B) are those having B as conclusion.\ In our system, the stratification is applied in two levels. In the first level the stratification is based on the modalities.Four layers are identified. The first layer contains predicates with the empty modality ($Holds(P,~ A,~ t)$). The second one is constituted by duty predicates ($Must(P,~ A,~ t)$). In the third one we find predicates of capacity ($Able-To(P,~ A,~t)$). Finally the last layer contains the two predicates $An$ (primary anomaly) and $D-An$ (derived anomaly).\ The second level of stratification concerns the predicates of the two first layers (corresponding to empty and duty modalities). In each of these layers, we establish an order to the predicates so that the constraints of stratification are verified. We obtained 10 sub-layers in the first layer, and 2 sub-layers in the second one. We have checked manually the validity of this method on a significant part of the corpus, and we are developing an automatic reasoning system to validate automatically the obtained results.\ Considering our example, in the first layer, the involved predicates are ordered in the following 8 sub layers (two sub layers are not used in this example): $\{Bump, ~Shock\},~\{Shock\},~ \{Stop, ~Avoid, ~Obstacle\},\{Control\},~\{predictable\},$ $\{Same\_File, ~Follow\},~ \{Stop\_Cause\},~ \{Cause\_Later\_Stop\}$. Among others, the following inference rules of this layer are applied: $Holds(Combine(Bump, ~V), ~W,~ t) ~\rightarrow \neg Holds(Stop,~ W,~ t)~~ (sub layer~ 3).$ $Holds(Combine(Shock,~V),~ W,~ t) ~\wedge~ Holds(Combine( Shock\_Pos, ~V), ~Back,~ t) : $ $Holds(Combine(Follow, ~V),~ W, ~t-1)[Holds(Control,~W,~t-1)]~~ (sub layer~ 6).$ The first rule means that if $V$ bumps into $W$ at state $t$, then $V$ is not stopped at this state $t$. It enables us to infer $\neg Holds(Stop, ~A, ~2) (V = B,~ W = A,~ t = 2)$. The second rule states that, in general, if there is a shock between $V$ and $W$ at state $t$, and the position of the shock of $V$ is its back, then generally $W$ was following $V$ in the same file. From this rule we can obtain $Holds(Combine(Follow,~ B),~ A,~ 2) (V = B,~ W = A,~ t = 2)$.\ We deduce: $Holds(Combine(Follow,~B),~ A,~ 1)$ by applying the backward persistence rule for the predicate Follow: $Holds(Combine(Follow,~ V),~ W,~ t) : Holds(Combine(Follow,~ V),~ W,~ t-1)$ An example of inferring duties in layer 2, is the rule : $Holds(Combine(Follow,~ V),~ W,~ t)~ \wedge ~Holds(Stop,~ V,~ t)~ \rightarrow ~Must(Stop,~ W,~ t)~~ (sub layer 2)$ If $W$ follows $V$ in a file at state $t$, and $V$ stops at this state, then $W$ must stops at state $t$ too. From this rule, we infer: $Must(Stop,~ A,~ 1) (V = B,~ W = A,~ t = 1)$. Finally to determine if $A$ is able or not to avoid the shock, use the general rule [@Kayser:04b]: $Able-To(P,~ V,~ t)~ \leftrightarrow~ (\exists Act)~ Action(Act)~ \wedge~ Pcb(P,~ Act)~ wedge~ Available(Act,~ P,~ V,~ t)$ At state $t$, $V$ is able to reach $P$ if and only if there is an action $Act$ which is a potential cause for $P (Pcb(P,~ Act))$ and $Act$ is available for $V$ to reach $P$ at state $t$. Knowing the fact $Pcb(Brake,~ Stop)$ and that by default any potential cause of an effect is available, we obtain $Available(Brake,~ Stop,~ A,~ 1)$. Consequently we deduce $Able-To(Stop,~ A,~ 1)$.\ Finally, by applying the first form of primary anomalies, we can detect the cause of the accident: “ A did not stop in a situation in which it had to do. ” Conclusion and Perspectives =========================== In the present work, we propose a norm based reasoning system able to detect the causes of the accidents from their textual descriptions. The cause is seen as a violation of a norm. The study we have done enabled us to determine a limited number of semantic predicates (50) and inference rules (currently 150). In a short and medium term perspective, we will finish the implementation of the automatic reasoning system based on the idea of stratification, complete the design of remaining inference rules and validate our approach on new car crash reports. In a longer term perspective, we will try to generalize our methodology to other domains and we will explore the possibility of applying our approach to propose a norm base indexation of textual documents. [^1]: A predicate $Q(X,~ Y,~ Z)$ is written : $Holds(Combine(Q, ~X),~ Y,~ t)$. $Combine(X,~ t)$ is a complex property

ArXiv

--- abstract: 'For over twenty years, ultra-cold atomic systems have formed an almost perfect arena for simulating different quantum many-body phenomena and exposing their non-obvious and very often counterintuitive features. Thanks to extremely precise controllability of different parameters they are able to capture different quantum properties which were previously recognized only as theoretical curiosities. Here, we go over the current experimental progress in exploring the curious one-dimensional quantum world of fermions from the perspective of three subjectively selected trends being currently under vigorous experimental validation: ([*i*]{}) unconventional pairing in attractively interacting fermionic mixtures, ([*ii*]{}) fermionic systems subjected to the artificial spin-orbit coupling, ([*iii*]{}) fermionic gases of atoms with high $\mathrm{SU}({\cal N})$ symmetry of internal states.' author: - Jacek Dobrzyniecki - Tomasz Sowiński bibliography: - '\_Biblio.bib' title: | Simulating artificial one-dimensional physics\ with ultra-cold fermionic atoms: three exemplary themes --- Introduction ============ Physical sciences were born mainly to deliver an appropriate and understandable description of the observable world. Typically, the laws of physics were formulated [*after*]{} discovery of related phenomena and eventually then their general consequences were disclosed. However, from time to time, some artificial theoretical models (frequently additionally reduced to one spatial dimension) were introduced without any experimental motivation – just to expose some intriguing properties of the underlying theory [@1968CooperBook]. This kind of approach was intensified when quantum mechanics was born since then many realistic problems were considered as too complicated to be explained in a full quantum-mechanical treatment. This is how many interesting simplified theoretical models were introduced. Let us mention here only a few celebrated examples: the Dicke model [@1954DickePR], the Schwinger model [@1962SchwingerPR], the Hubbard model [@1963HubbardProcRSocLondon], the Jaynes-Cummings model [@1963JaynesProcIEEE], the Lieb-Liniger model [@1963LiebPhysRev1; @1963LiebPhysRev2], the Calogero-Sutherland model [@1971CalogeroJMP; @1971SutherlandJMP], the Harper-Hofstadter lattice model [@1976HofstadterPRB], the Sachdev-Ye-Kitaev model [@1993SachdevPRL], [*etc.*]{} The situation changed drastically along with tremendous development of experimental methods of precisely controlling interactions between light and matter on a subatomic scale. It turns out that these simplified theoretical models and a variety of their extensions and modifications are possible to engineer with atomic systems, [*i.e.*]{}, appropriately prepared and controlled ultra-cold atomic systems may serve as almost perfect realizations of systems described by desired Hamiltonians [@LewensteinBook]. In this way, Richard Feynman’s brilliant dream of creating [*quantum simulators*]{} [@1982FeynmanIJTP] can be realized and a new era of [*quantum engineering*]{} has started. One example of the exciting developments in the field is the ability to engineer effectively one-dimensional quantum many-body systems. This idea has long been of theoretical interest, including such important models as the Tonks model of a gas of impenetrable rods [@1936TonksPR] or the Gaudin-Yang general solution for the ground state of fermions [@1967GaudinPLA; @1967YangPRL]. Now, by utilizing the quantum simulator concept, many of the exotic physical phenomena characterizing one-dimensional systems can be explored experimentally for the first time. With advanced trapping techniques, atoms can be confined in traps of effectively varying dimensionality by controlling the strength of perpendicular confinement. This includes optical lattices [@2005BlochNatPhys] and single- and few-site potentials [@2011CazalillaRevModPhys; @2013GuanRevModPhys; @2019SowinskiRepProgPhys]. In particular, different one-dimensional systems of ultra-cold fermionic mixtures have been experimentally created in this way [@2005MoritzPRL; @2010LiaoNature]. In this review, we describe recent achievements in the domain of one-dimensional fermionic ultra-cold atom systems. Our focus is on the developments that have occurred in the past few years, since the last comprehensive review from 2013 by Guan *et al* [@2013GuanRevModPhys]. Since the progress in the entire field of 1D fermionic systems is exceedingly broad and rapid, a full catalogue of all the major advancements would be a tremendous undertaking. Therefore we concentrate on three subjectively selected main research directions which are currently being heavily explored and, in our opinion, will have significant importance for the future capability of quantum technologies. The organization of this article is as follows. In Section \[sec-fflo\] we describe the search for unconventional superconducting states in one-dimensional systems, which are of considerable theoretical interest but are still difficult to pin down experimentally. In Section \[sec-soc\] we describe systems under the influence of artificial spin-orbit coupling. In one-dimensional settings, this kind of coupling presents an interesting picture, since there is in fact no “orbit” in the usual sense. Furthermore, it has important applications, such as the simulation of topologically nontrivial models requiring higher dimensionality. In Section \[sec-sun\] we describe the research on atomic systems with higher-spin internal symmetries, which in the case of one spatial dimension offer a fascinating arena for exploring various exotic many-body phases. Section \[sec-concl\] is the conclusion. Unconventional pairing phases {#sec-fflo} ============================= The simplest properties of superconducting materials are typically described by the pairing mechanism of spontaneous formation of correlated pairs of opposite-spin fermions. The mechanism is appropriately described by the theory of superconductivity of Bardeen, Cooper, and Schrieffer (BCS) [@1957BardeenPR]. However, when there is no direct symmetry between opposite-spin components, certain more exotic pairing phases are possible. One of them, the *Fulde-Ferrell-Larkin-Ovchinnikov* (FFLO) pairing phase, has recently attracted significant interest due to the possibility of its realization in quasi-one-dimensional ultra-cold fermionic systems. The FFLO phase was first predicted in the 1960s by Fulde and Ferrell [@1964FuldePR] and independently by Larkin and Ovchinnikov [@1965LarkinSoviet], who considered the ground state of a solid-state superconductor subjected to an external magnetic field. The magnetic field causes a relative shift of the Fermi surfaces of both electron spin components. If this shift is too high, the Cooper pairing is destroyed and the transition from the superconducting to the normal state occurs. However, Fulde, Ferrell, Larkin, and Ovchinnikov showed that close to the transition (still in superconducting phase) an FFLO state can be formed. In this unusual region the pairing of fermions with momenta $\vec{k}$ and $-\vec{k}+\vec{Q}$ is favored over the standard BCS pairing of fermions with momenta $\vec{k}$ and $-\vec{k}$ [@2004CasalbuoniRevModPhys]. The resulting Cooper pairs have a nonzero center-of-mass momentum $\vec{Q}$ which is (in general) proportional to the magnitude of the mismatch between the Fermi wave vectors [@2018KinnunenRepProgPhys]. A signature feature of the FFLO phase is that the pairing order parameter $\Delta(\vec{x})$ is no longer constant in space as in the standard BCS phase, but rather has an oscillatory character, $\Delta(\vec{x}) \propto \cos(\vec{Q} \cdot \vec{x})$ [@2007MatsudaJpn]. In the decades since the original proposals, the FFLO phase has been extensively investigated. It has been the subject of several reviews [@2004CasalbuoniRevModPhys; @2007MatsudaJpn; @2013BeyerLowTemp], including very recent ones [@2018KinnunenRepProgPhys; @2018WosnitzaAnnPhys; @2019Agterberg]. The FFLO state is currently invoked to explain the behavior of several superconducting systems, including heavy-fermion and organic materials, as well as the cores of neutron stars [@2001AlfordJPhysG; @2004CasalbuoniRevModPhys]. However, in spite of its significance, universally accepted experimental evidence for the FFLO pairing has still not been obtained, although a number of experiments have shown results highly suggestive of the FFLO state [@2017CroitoruCondMat; @2019Kasahara]. Quantum simulators in one-dimensions ------------------------------------ A significant difficulty in the experimental search for the FFLO phase is that it is highly unstable in three dimensions, only occupying a small sliver of the phase diagram of fermionic systems [@2006SheehyPRL]. Any external magnetic field will typically cause orbital effects destructive for superconductivity, significantly limiting the possibilities of observing the FFLO state [@2007MatsudaJpn]. This problem can be overcome, however, by using lower-dimensional superconductors, for which the circulating eddy currents around magnetic field lines are blocked by geometric constraints [@2007MatsudaJpn; @2014MiyawakiJPhys; @2018KinnunenRepProgPhys]. In particular, the FFLO phase occupies a large part of the phase diagram in fermionic systems in one spatial dimension [@1987BuzdinSoviet; @2007HuPRL; @2007ParishPRL; @2012FeiguinChapter], making (quasi-)one-dimensional systems a very good environment to search for the elusive FFLO state. For this reason, the unconventional FFLO pairing in ultra-cold one-dimensional systems has recently been deeply investigated theoretically from various perspectives, for both confined [@2007OrsoPRL; @2008CasulaPRA; @2019Pecak; @2019Lydzba] and lattice systems [@2007FeiguinPRB; @2010HeidrichPRA; @2012FrancaPRA; @2017FrancaPhysA; @2018DePicoliBrazJPhys]. Quasi-one-dimensional quantum simulators created with ultra-cold neutral atoms constitute a highly controllable environment, where the Fermi surface mismatch can be precisely tuned by changing the spin composition of the initial population, rather than with external magnetic fields [@2006ZwierleinScience; @2006PartridgeScience; @2006ZwierleinNature; @2006ShinPRL; @2007SchunckScience]. The relative spin populations can be tuned, for example, by driving radio-frequency sweeps between the states at different powers [@2010LiaoNature]. Shown in Fig. \[Fig1\] is the theoretical phase diagram of a homogeneous 1D Fermi gas at $T = 0$, in the plane of the chemical potential, $(\mu_\uparrow + \mu_\downarrow)/2$, and spin population imbalance, $(\mu_\uparrow - \mu_\downarrow)/2$ (which is equivalent to the strength of the effective magnetic field). At low spin imbalance, the ground state of the system is the standard BCS paired phase. When the imbalance is increased, the system transitions into the FFLO-paired phase. At a high enough imbalance, the superfluid phase is destroyed and the system is in the normal, unpaired phase. Fig. \[Fig2\] shows an analogous phase diagram for a 1D lattice system described by an appropriate Hubbard-like model. Despite the differences, the overall structure of the phase diagram is similar to the homogenous 1D gas case, with the standard BCS-paired phase transitioning to the FFLO phase at a finite spin imbalance, and a subsequent transition to the normal phase beyond a critical imbalance value. ![The ground state phase diagram of a homogeneous Fermi gas, in the plane of average chemical potential $\mu = (\mu_\uparrow + \mu_\downarrow)/2$ vs. effective magnetic field $h = (\mu_\uparrow - \mu_\downarrow)/2$, where $\mu_\uparrow,\mu_\downarrow$ are the chemical potentials of individual spin components. $\mu$ and $h$ are given in units of binding energy $\epsilon_B$. $\mu_c$ and $\mu_s$ designate the critical values of $\mu$ that separate the distinct phases. One can distinguish between the “partially polarized” phase (i.e. the FFLO-type phase), the “fully paired” phase (standard BCS Cooper pairing), and the “fully polarized” phase (unpaired state). The arrow shows the trajectory of local chemical potential for a system in a harmonic trap. The shaded region corresponds to the vacuum state. Inset: zoom of the phase diagram near the point $O = (\epsilon_B/2,-\epsilon_B/2)$, with asymptotic behavior of phase diagram boundaries marked with dashed lines. Reproduced with permission from [@2007OrsoPRL]. Copyright 2007, American Physical Society. \[Fig1\] ](Fig1.jpg){width="1\linewidth"} ![Grand canonical phase diagram of the one-dimensional attractive Hubbard model (for fixed interaction $U=-5t$) in the plane of average chemical potential $\mu$ vs. effective magnetic field $h$. $\mu$ and $h$ are given in units of tunneling amplitude $t$. The distinct phases visible are as follows: V – vacuum, ED – unpolarized phase (standard BCS pairing), PP – partially polarized phase (FFLO-type pairing), FP$_1$ – fully polarized (unpaired) phase with low filling $n < 1$, FP$_2$ – fully polarized (unpaired) phase with filling $n = 1$. The remaining phases correspond to the states where at least one of the components forms a band insulator. Reproduced with permission from [@2010HeidrichPRA]. Copyright 2010, American Physical Society. []{data-label="Fig2"}](Fig2.jpg){width="1\linewidth"} One should remember that in the presence of external trapping, due to inhomogeneity, different configurations predicted by these phase diagrams may simultaneously co-exist at different locations in the trap. This phase separation can be understood in the local density approximation picture: the local value of the chemical potential $\mu$ varies along the system (in particular, for a harmonic trap, $\mu$ becomes smaller towards the trap edges). As a result, different phases are realized at different locations. Their arrangement reflects a trajectory across the phase diagram, parallel to the $\mu$ axis. Examples can be seen in Fig. \[Fig1\] and Fig. \[Fig2\], where the trajectories of the local chemical potential across the trap length are shown as vertical lines crossing several different phases. For a harmonically trapped 1D spin-imbalanced gas (whether with or without a lattice), a typical configuration is a two-shell structure, where the center of the system exhibits an FFLO phase and the edges of the system are in the unpaired phase or the standard BCS phase [@2007FeiguinPRB; @2007OrsoPRL; @2010HeidrichPRA]. Depending on the structure of the phase diagram, more complex phase-separation configurations can be obtained. This possibility was explored in [@2019CichySciRep], where it was shown how, by appropriately modifying the parameters of the confining trap, one can engineer different trajectories in the phase diagram. In this way, a desired configuration of separated phases can be created. ![(a) Theoretical phase diagram of a homogeneous Fermi gas at $T = 0$, in the plane of average chemical potential $\mu = (\mu_\uparrow + \mu_\downarrow)/2$ vs. effective magnetic field $h = (\mu_\uparrow - \mu_\downarrow)/2$, where $\mu_\uparrow,\mu_\downarrow$ are the chemical potentials of the individual spin species. The distinct phases shown are: the vacuum (white), fully paired (standard BCS-type) phase (green), partially polarized (FFLO-type) phase (orange), and the fully polarized (unpaired) phase (blue). The arrows show two different local chemical potential trajectories for a gas in a harmonic trap, for $h$ below or above the critical value $P_c$. (b) Phase diagram of a trapped 1D gas with infinitely strong attractive interactions, in the plane of radius from the trap centre vs. polarization of the central 1D trap. Reproduced with permission from [@2010LiaoNature]. Copyright 2010, Nature.[]{data-label="Fig3"}](Fig3.jpg){width="1\linewidth"} This theoretical phase separation picture was confirmed experimentally in the group of Randall G. Hulet and Erich J. Mueller at Rice University [@2010LiaoNature]. The experiment studied the nature of phase separation which occurred in fermionic systems confined to 1D tubes as the total spin imbalance of the system was tuned. At low spin polarization, below a critical value $P_c$, the center of the cloud was occupied by a partially polarized state. Towards the edges of the cloud, the system was fully paired. This corresponds to the trajectory labelled as “low P” in the phase diagram in Fig. \[Fig3\]. For values of polarization close to $P_c$, the partially polarized phase extended across the entire trap. Finally, at high polarization $P > P_c$ the state at the edge of the cloud changed from fully paired to fully polarized, in agreement with the theoretical prediction (the trajectory labelled as “high P” in Fig. \[Fig3\]). Undoubtedly, the experiment showed the validity of the predicted phase separation. Nevertheless, it did not provide direct evidence that the partially polarized state in the center of the trap was indeed the elusive FFLO state. In anticipation of future experimental work, several potential experimental signatures of the FFLO state have been proposed. A well-established possibility is measuring the pair momentum distribution of the partially polarized phase, with a peak at finite momentum $q$ providing an unambiguous signature of FFLO with pair momentum $q$ [@2008CasulaPRA; @2010LiaoNature]. In recent years, there have been proposals based on surveying the expansion dynamics of the cloud after switching off the trapping potential [@2011KajalaPRA; @2012LuPRL]. Recently, it was also suggested [@2016YinPRA] that the FFLO state subject to a sudden quench of the interaction strength should display characteristic, experimentally detectable post-quench features. It has also been shown that the visibility of the FFLO state should be greatly enhanced in one-dimensional boson-fermion mixtures with strong boson-fermion repulsion [@2019Singh]. Dimensionality crossover ------------------------ For future experimental work, promising perspectives are opened by the implementation of systems with an “intermediate” dimensionality. Although here we focus on 1D systems, the dimensional crossover technique offers interesting perspectives for the observation of FFLO and it is worth looking at recent developments in this area. Experimentally, a system of this kind can be implemented with an array of 1D tubes, where the amplitude of tunneling between neighboring tubes $t_\perp$ can be tuned, thus controlling the effective dimensionality. In particular, in lattice systems the relevant parameter is the ratio $t_\perp/t_\parallel$ (where $t_\parallel$ is the tunneling between sites of a single tube), which can range from 0 (fully 1D) to 1 (isotropic 3D) [@2012KimPRB; @2013HeikkinenPRB]. A quasi-1D regime, where $t_\perp$ is small but nonzero, is expected to be even better suited to the observation of the FFLO state even more than a purely 1D system. It comes from the fact that, in such a regime, there can exist long-range order absent from purely 1D systems, stabilizing the FFLO phase [@2007ParishPRL; @2013HeikkinenPRB]. Furthermore, correlations induced by the weak intertube tunnelings could synchronize the FFLO density modulations across different tubes and thus enhance the overall experimental signal from the array [@2007ParishPRL]. The difference between different dimensionalities manifests itself in the phase separation of a trapped Fermi gas. As noted above, in a quasi-one-dimensional trap one typically obtains a two-shell structure where the center of the trap is occupied by a partially polarized FFLO state while the edges are taken up by a polarized normal state. On the other hand, in a spherical 3D trap, a shell structure is predicted with the standard BCS superfluid occupying the trap center [@2007LiuPRA-Mean]. Theoretical and experimental research has confirmed that the crossover between 1D and 3D dimensionalities is indeed reflected in the phase separation structure of the system. For example, in [@2012KimPRB; @2013HeikkinenPRB] the phase diagram of a 3D array of 1D lattices with harmonic trapping was studied. The zero-temperature phase diagram of the relevant system can be seen in Fig. \[Fig4\]. At $t_\perp/t_\parallel > 0$, in addition to the familiar two-shell structure with FFLO in the center (region “III” in Fig. \[Fig4\]), there appears the possibility of obtaining three-shell structures, in which the gas in the center of the trap separates into two shells displaying FFLO and standard BCS phases (regions “I” and “II” in Fig. \[Fig4\]). As the transverse coupling increases, the structure with standard BCS pairing occupying the trap center (region “I” in the trap center), characteristic of a quasi-three-dimensional regime, becomes preferred [@2012KimPRB]. It is argued that the approximate crossover point between quasi-one-dimensional and quasi-three-dimensional physics, $t_\perp/t_\parallel \approx 0.3$, is a “sweet spot” where the FFLO state displays a highly uniform oscillation amplitude across the entire 1D tube [@2013HeikkinenPRB]. Above a critical temperature, which is approximately 1/3 that of the critical temperature for BCS superconductivity, the FFLO phase becomes fragile to losing its FFLO character and melting into standard BCS pairing [@2013HeikkinenPRB]. ![Phase diagram of the model representing an array of coupled one-dimensional Hubbard chains, in the plane of interchain coupling vs. spin polarization. Interchain coupling is given as the ratio $t_\perp/t_\parallel$ between transverse and inter-site tunnelings. Phase III represents a two-shell structure, with an FFLO core and fully polarized, unpaired edges. Phase II represents a three-shell structure with an FFLO core, standard BCS superfluid in the shoulders, and fully polarized edges. Phase I represents a three-shell structure with a standard BCS superfluid core, FFLO shoulders, and fully polarized edges. Reproduced with permission from [@2012KimPRB]. Copyright 2012, American Physical Society.[]{data-label="Fig4"}](Fig4.jpg){width="1\linewidth"} Recently the 1D-3D crossover scenario was successfully realized experimentally with $^6$Li atoms confined in an array of 1D tubes [@2016RevellePRL]. The array of 1D traps was formed with a 2D optical lattice and the transverse tunneling rate could be tuned by changing the 2D lattice depth (Fig. \[Fig5\]a). The quasi-1D and quasi-3D regimes could be distinguished by the local spin polarization at the midpoint of the central 1D tube (Fig. \[Fig5\]b): a partially polarized core corresponded to the quasi-1D regime, and an unpolarized core indicated a quasi-3D regime. It was found that the critical tunneling value, corresponding to the transition between the quasi-1D and quasi-3D regime, was approximately $t_c \approx 0.025\epsilon_b$ (where $\epsilon_b$ is the pair binding energy). ![(a) An array of 1D tubes formed by a 2D optical lattice. By decreasing the optical lattice depth, the intertube tunneling rate $t$ is increased. In this way the system can be gradually tuned from a quasi-1D to a quasi-3D regime. (b) Phase separation in a trapped Fermi gas in quasi-1D (top) and quasi-3D (bottom) regimes, at zero temperature and a small spin imbalance. The phases are: $\mathrm{SF_P}$ – FFLO superfluid, $\mathrm{SF_0}$ – standard BCS superfluid, $\mathrm{N_{PP}}$ – an unpaired phase with spin imbalance, $\mathrm{N_{FP}}$ – an unpaired normal phase. Arrows indicate phase boundaries at the different radii $R$. Reproduced with permission from [@2016RevellePRL]. Copyright 2016, American Physical Society. []{data-label="Fig5"}](Fig5.jpg){width="1\linewidth"} Finally, it is worth noting an alternative approach to the 1D-3D crossover was proposed in [@2016DuttaPRA]. Here, only a single 1D tube is considered, and the parameter controlling the dimensionality is the chemical potential $\mu$. When $\mu$ is small enough, transverse movement is confined to the lowest oscillator level, [*i.e.*]{}, the usual condition for quasi-1D dimensionality is fulfilled. For large enough $\mu$ transverse modes are accessible and the dynamics become locally 3D. The authors find that strong interactions, which mix single-particle levels, cause 3D-like behavior to occur at all densities. Mass-imbalanced mixtures ------------------------ A parallel direction of research on unconventional pairing concentrates on the relationship between mass-imbalanced 1D systems and FFLO. In the typically considered ultra-cold systems, the source of mismatch between the two spin Fermi momenta in the ultra-cold system is the imbalance between spin populations, which leads to a difference in the chemical potential $\mu$ and thus, the magnitude of the Fermi momenta. However, an alternate way to induce the difference between Fermi momenta is by using components with different masses. The most straightforward approach is creating a mixture of different atomic species with varying masses [@2018KinnunenRepProgPhys]. Alternatively, one can create a system confined in spin-dependent optical lattices, where the two spin components exhibit different tunneling amplitudes and thus different “effective mass”. Such an effect can be achieved, for example, by the use of a magnetic field gradient modulated in time [@2015JotzuPRL]. The occurrence of the FFLO phase in mass-imbalanced 1D fermionic systems was theoretically investigated in a number of past works [@2005CazalillaPRL; @2009WangPRA; @2009BatrouniEPL; @2009BurovskiPRL; @2010OrsoPRL; @2011RouxPRA; @2012RoscildeEPL]. Among recent works, a three-dimensional phase diagram as a function of the mass imbalance, spin imbalance and temperature was studied in [@2014RoscherPRA] for a many-body system of attractive free fermions, finding that FFLO-type phases occupy a large region of the parameter space. Another recent work [@2017ChungPRA] theoretically studied the zero-temperature phase diagram for an attractive $^6$Li-$^{40}$K mixture confined in a 1D harmonic trap. When the two mass-imbalanced atomic species are treated as distinct pseudospin components, a greater richness of phases emerges: one can now distinguish between “light FFLO” or “heavy FFLO” phases, depending on whether the heavier or the lighter species is an excess species in the partially polarized phase. Dynamical response technique ---------------------------- Finally, we note several recent works that focus on the dynamical properties of the system. Although most theoretical work focuses on the ground-state properties of the system, a significant area of research focuses on the time evolution of dynamical systems after the sudden change (quench) of some parameter, such as the external potential. For instance, an FFLO Fermi gas with initial harmonic confinement which is suddenly switched off can be considered [@2011KajalaPRA]. The resulting cloud expansion dynamics shows a clear two-fluid behavior, where the cloud expansion velocity of one of the two components (consisting of unpaired majority fermions) is related to the FFLO momentum. This provides an experimental signature of FFLO pairing with nonzero center-of-mass momentum of the pairs. Another often considered area is the post-quench dynamics after a change of interaction strength. Such a scenario was considered in [@2015RieggerPRA] for a 1D lattice system. In particular, after a quench from zero to attractive interactions, the post-quench state shows characteristic FFLO oscillations of the pair correlation, although with exponential decay of spatial correlations. On the other hand, after a sufficiently fast quench from attractive to repulsive interactions, the initial state’s FFLO correlations can be imprinted onto repulsively bound pairs if the final interaction strength is high enough. A different case was considered in [@2016YinPRA], which analyzed the dynamics of standard BCS and FFLO states quenched from attractive to zero interactions, and analyzed the dynamics of spin and charge correlations in the post-quench system. For a quench from an initial standard BCS state, the spin correlations eventually thermalize to those of a free Fermi gas at a temperature $T \sim U_{ini}$, while the charge component does not. On the other hand, for a quench from the FFLO state, neither component thermalizes. Although experimental implementation of such schemes is yet to be achieved, they are realizable with currently available techniques. For example, cloud expansion experiments with a clear resolving between clouds of single and paired atoms have been demonstrated recently [@2018SchergPRL]. Therefore, this approach may in the future provide the long-sought clear experimental evidence for the FFLO state. Spin-orbit coupling {#sec-soc} =================== With current experimental techniques, it is possible to engineer scenarios mimicking the existence of external gauge fields [@2011DalibardRevModPhys; @2014GoldmanRepProgPhys]. In particular, artificial generation of spin-orbit (SO) coupling, i.e., the coupling between the internal and the motional degrees of freedom of a particle, has been attracting increased interest in recent years. In condensed matter systems, SO coupling plays a crucial role in the formation of exotic, topologically nontrivial phases associated closely with the quantum Hall effect [@2010HasanRevModPhys; @2011QiRevModPhys; @2013GalitskiNature]. Recent progress in spintronics has also contributed to the interest in SO coupling [@2004ZuticRevModPhys; @2013GalitskiNature]. Typically SO coupling is understood as a purely relativistic effect [@1980ItzyksonBook; @1998JacksonBook3rd] which can be explained directly from the movement of a spinful particle in the intrinsic electric field of the sample. The particle, in its comoving reference frame, experiences a magnetic field that couples to the spin. It means that the resulting spin-orbit coupling is determined by the intrinsic properties of the material and is not easily tunable. However, ultra-cold systems of neutral atoms subject to synthetic SO coupling open an alternate way to investigate this phenomenon, providing an experimentally controllable environment where the SO coupling can be precisely engineered and tuned [@2011LinNature; @2014GoldmanRepProgPhys; @2018ZhangChapter]. General reviews concerning the realization of synthetic SO coupling in ultra-cold atoms can be found in [@2013GalitskiNature; @2015YiScienceChina; @2014GoldmanRepProgPhys; @2015ZhaiRepProgPhys; @2019ZhangJPhysChem]. Experimental implementation of artificial SO coupling in ultra-cold gases is already well established. The first realization of synthetic SO coupling in ultra-cold atoms came in 2011, with the realization of a SO coupling in a bosonic condensate at NIST [@2011LinNature]. Very soon it was also successfully engineered in 3D fermionic gases [@2012WangPRL; @2012CheukPRL; @2013WilliamsPRL; @2014FuNature]. In this section, we will describe the recent experiments with synthetic SO coupling in 1D fermionic gases. Experimental methods -------------------- ![A typical three-level Raman scheme for the generation of artificial spin-orbit coupling in ultra-cold atoms. Two internal states of the atoms, differing in energy by $\hbar \Omega$, represent the two pseudospin states $|e\rangle$ and $|g\rangle$. A pair of laser beams couples the two pseudospin states through an intermediate excited state $|E\rangle$. The lasers are detuned by $\delta$ from the Raman resonance.[]{data-label="Fig6"}](Fig6.jpg){width="0.8\linewidth"} First, let us describe the current experimental techniques for the generation of artificial SO coupling in 1D systems. A well-established technique is the Raman laser scheme, originally proposed in [@2009LiuPRL] and used in the earliest experimental realizations [@2011LinNature; @2012WangPRL; @2012CheukPRL; @2013WilliamsPRL; @2014FuNature]. In this approach, two internal states of the ultra-cold atoms are chosen to represent pseudospin states. Then a pair of counter-propagating lasers is shined on the ultra-cold atom system, inducing a two-photon Raman coupling between the two states (see Fig. \[Fig6\]). Due to the conservation of momentum during the absorption and reemission of photons, the transition of an atom between the internal states is accompanied by a change in the momentum. As a result, the motion of the particle becomes coupled to the spin [@2013GalitskiNature]. The magnitude of the transferred momentum depends on the wavelength of the Raman beams, but it can be tuned by changing the relative angle of their intersection [@2014GoldmanRepProgPhys]. ![Changes of the energy-momentum dispersion of a homogeneous ultra-cold Fermi gas in the presence of the SO coupling. Upper and bottom panels correspond to theoretical predictions and experimental results from [@2012CheukPRL], respectively. With no SO coupling, the spectrum consists of two degenerate, parabolic energy bands corresponding to the two spin states. The off-diagonal coupling term (linear in momentum) causes a shift (leftmost plots), which together with the Zeeman splitting terms leads to coupling between the two energy bands and opening a gap in the spectrum (middle and rightmost plots). Colors indicate the spin composition of the states. Experimentally measured dispersions (for various Raman couplings $\Omega$) reproduced with permission from [@2012CheukPRL]. Copyright 2012, American Physical Society. []{data-label="Fig7"}](Fig7a.jpg "fig:"){width="1\linewidth"} ![Changes of the energy-momentum dispersion of a homogeneous ultra-cold Fermi gas in the presence of the SO coupling. Upper and bottom panels correspond to theoretical predictions and experimental results from [@2012CheukPRL], respectively. With no SO coupling, the spectrum consists of two degenerate, parabolic energy bands corresponding to the two spin states. The off-diagonal coupling term (linear in momentum) causes a shift (leftmost plots), which together with the Zeeman splitting terms leads to coupling between the two energy bands and opening a gap in the spectrum (middle and rightmost plots). Colors indicate the spin composition of the states. Experimentally measured dispersions (for various Raman couplings $\Omega$) reproduced with permission from [@2012CheukPRL]. Copyright 2012, American Physical Society. []{data-label="Fig7"}](Fig7b.jpg "fig:"){width="1\linewidth"} This laser coupling scheme results in the realization of a one-dimensional SO coupling, equivalent to an additional term in the single-particle Hamiltonian of the general form $\propto q \hat{\sigma}_y$. Here $q$ is the momentum of the atom along the SO coupling direction, and $\hat{\sigma}_y$ is the spin Pauli matrix. Additionally, effective Zeeman terms appear in the Hamiltonian, which can be written in the general form $(\Omega/2)\hat{\sigma}_z + (\delta/2)\hat{\sigma}_y$. They are parametrized by the Raman coupling $\Omega$ and the two-photon detuning $\delta$ from the bare transition frequency (for details see [@2011LinNature; @2014GoldmanRepProgPhys]). The SO coupling has a characteristic effect on the energy-momentum dispersion relation. First, due to the counter-effect of momentum transfer for opposite spins, the two bands are split and relatively shifted. Secondly, the Zeeman splitting term causes a characteristic split around zero momentum and opens a gap in the spectrum [@2015YiScienceChina; @2018ZhangChapter]. Importantly, the resulting characteristic dispersion and spin texture of the spectrum can be probed experimentally, for example by means of a spectroscopic spin-injection technique [@2012CheukPRL] (see Fig. \[Fig7\]). For atoms confined in a 1D lattice, an alternative technique of synthesizing SO coupling has been developed in recent years. In this approach, the atoms are subjected to an optical clock laser, which induces a single-photon coupling between the ground atomic state and a long-lived, metastable excited state. When the trapping lattice is set to an appropriately selected “magic wavelength”, such that the trapping is identical for both these pseudospin states, a SO coupling emerges (Fig. \[Fig8\]). The SO coupling results from the fact that when the laser drives a transition between the ground and excited state, it imprints on the atom wave function an additional site-dependent phase, exactly as for an atom in an external magnetic field. Compared to the Raman technique, the advantage of this method is its simpler configuration (only one laser beam). It also avoids the detrimental effect of near-resonant intermediate states that would otherwise induce strong heating and hinder the observation of many-body effects [@2016WallPRL]. Much like in the case of the Raman laser scheme, the quasimomentum-energy dispersion undergoes a characteristic modification in the spin-orbit coupled 1D lattice system. It can be regarded as two bands, shifted with respect to each other and coupled (Fig. \[Fig9\]a). This SO-coupled spectrum is characterized by divergences (the Van Hove singularities) in the density of states at specific values of the quasimomentum. This results in the appearance of characteristic peaks in the excitation spectrum, at detunings comparable to the bandwidth (Fig. \[Fig9\]b). They can be used as a spectroscopic signature of the SO coupling [@2016WallPRL; @2016LiviPRL; @2017KolkowitzNature; @2018BromleyNatPhys]. ![Clock transition SO coupling setup in a 1D optical lattice, as shown in [@2016LiviPRL]. The fermionic $^{173}\mathrm{Yb}$ atoms are confined in an optical lattice with wavelength $\lambda_L$. The clock laser with wavelength $\lambda_C$, applied at angle $\theta$ to the optical lattice axis, drives a single-photon transition between the two states $|g\rangle ={}^1\mathrm{S}_0$ and $|e\rangle ={}^3\mathrm{P}_0$ (treated as pseudospin states). The momentum transfer $\delta_k = 2 \pi \cos(\theta)/\lambda_C$ causes a coupling between the momentum and the two pseudo-spin states. Reproduced with permission from [@2016LiviPRL]. Copyright 2016, American Physical Society.[]{data-label="Fig8"}](Fig8.jpg){width="1\linewidth"} ![(a) The momentum-energy dispersion spectrum of an ultra-cold Fermi gas in a lattice under SO coupling. Similarly to the free gas (Fig. \[Fig7\]), the bands corresponding to the two spin states are shifted and coupled. The band splitting is given by the Rabi frequency $\Omega$ of the clock transition, and the bandwidth is equal to $4J$ where $J$ is the lattice tunneling rate. The Van Hove singularities in the density of states occur at quasimomenta $q \sim 0$ and $q \sim \pi$, indicated by the yellow and blue arrows. (b) The $|g\rangle \rightarrow |e\rangle$ excitation spectrum of the clock transition, as a function of the detuning $\delta$ from the bare atomic transition (in units of $J$). The Van Hove singularities are manifested as peaks at the values $\delta \sim \pm 4J$ (yellow and blue arrows). Reproduced with permission from [@2018BromleyNatPhys]. Copyright 2018, Nature.[]{data-label="Fig9"}](Fig9.jpg){width="1\linewidth"} Apart from the two techniques described above, other methods have been proposed for SO generation. One proposal involves generating the effective SO coupling by periodic spin-dependent driving of atoms trapped in a lattice via a time-dependent magnetic field. In this way, the atom tunneling amplitudes become spin-dependent, and as a result, the characteristic SO splitting of the energy spectrum appears. The strength of the resulting SO coupling can be tuned by adjusting the driving amplitude [@2014StruckPRA]. Another example is the so-called Raman lattice scheme proposed in [@2013LiuPRL] and later implemented experimentally in [@2018SongSciAdv]. In this approach, two laser beams are used. One laser beam generates an optical lattice. The other perpendicular beam overlays the lattice with a periodic Raman potential inducing spin-flipping hopping between the lattice sites. It thus leads to effective spin-orbit coupling. In this approach, both beams can be generated by a single laser source, which simplifies the experimental setup. Experimental realizations ------------------------- We now proceed to describe the recent experimental achievements of spin-orbit coupled 1D Fermi gases. We start by listing the recent successful implementations of the clock lattice technique for generating the SO coupling. In an experiment by the Fallani group in LENS [@2016LiviPRL], a gas of ultra-cold $^{173}$Yb atoms was confined in a 1D magic wavelength lattice potential, with identical band structures for both internal states $|g\rangle = {}^1\mathrm{S}_0$ and $|e\rangle = {}^3\mathrm{P}_0$ chosen as the spin states. A clock laser along the lattice direction generated coherent coupling between the $|g\rangle$ and $|e\rangle$ states (Fig. \[Fig8\]). In this case, the clock laser transition was used both to implement the SO coupling and to probe the system spectroscopically. In particular, the authors confirmed that – with increasing SO coupling strength – the excitation spectrum of the clock transition displays a pair of characteristic peaks, corresponding to the Van Hove singularities. A similar experiment with SO-coupled 1D Fermi gas has been performed with $^{87}$Sr atoms in JILA [@2017KolkowitzNature]. In particular, the authors demonstrated that it is possible to selectively prepare atoms with particular quasimomenta $q$, thanks to the $q$-dependence of the clock transition frequency. In another recent experiment in JILA [@2018BromleyNatPhys] the authors focused on the effects of strong many-body interactions in the SO-coupled system, analyzing the influence of the interactions on the collective spin dynamics. Artificial dimensions --------------------- An interesting aspect of such one-dimensional lattice experiments is that the spin degree of freedom can be interpreted as a “synthetic dimension”, and transitions between the spin states can be interpreted as hoppings along this dimension [@2012BoadaPRL; @2014CeliPRL]. In this framework, a 1D lattice loaded with fermions of ${\cal N}$ spin components is interpreted as a 2D “ladder“ with ${\cal N}$ “legs” (Fig. \[Fig10\]). If the atoms are subject to an artificial spin-orbit coupling, the hopping in this synthetic dimension becomes complex, with a phase that depends on the lattice site index. The phase imprinted by the spin coupling varies between neighboring sites, with a value dependent on details of the 1D lattice potential and the artificial SO coupling gauge field. Then, in the synthetic dimension picture, the SO coupling corresponds to an effective magnetic field flux piercing each plaquette of the ladder [@2016WallPRL]. ![A synthetic two-leg ladder structure, realized by atoms in a 1D lattice which have two pseudospin states $|g\rangle$ and $|e\rangle$ coupled by a clock laser transition. Hopping between the sites of the lattice takes place with a real amplitude $J_n$ (dependent on the transverse mode $n$). The coupling between states $|g\rangle$ and $|e\rangle$ is equivalent to hopping along the synthetic spin dimension (from one ladder leg to the other), with a complex amplitude with magnitude $\Omega$ and a phase dependent on the lattice site and the artificial magnetic flux. Reproduced with permission from [@2016WallPRL]. Copyright 2016, American Physical Society.[]{data-label="Fig10"}](Fig10.jpg){width="1\linewidth"} Such synthetic ladders with accompanying SO coupling have attracted interest due to their potential application to study topologically nontrivial states of matter in ultra-cold atoms that are not attainable in standard condensed-matter systems. For example, the two-leg ladder offers a means to realize the Creutz ladder model, one of the most important minimal models that can realize topological insulator phases [@1999CreutzPRL; @2017JunemannPRX]. Detailed reviews on the realization of topological phases with ultra-cold atoms can be found in [@2016GoldmanNatPhys; @2019CooperRevModPhys; @2018ZhangAdvPhys]. However, an even more interesting possibility offered by the synthetic dimension framework is emulating 2D systems with 1D lattices [@2014CeliPRL]. The synthetic ladder can be interpreted as a fragment (a strip) of a larger 2D lattice. With the addition of the effective magnetic flux from SO coupling, the ladder system can emulate the physics of the topologically nontrivial Harper-Hofstadter lattice model [@1976HofstadterPRB], which describes charged particles in a 2D lattice in a uniform magnetic field. In fact, it can be shown theoretically that a two-leg ladder with SO coupling can accurately reproduce the energies and wave functions of the edge states of a real Hofstadter lattice [@2014HugelPRA]. ![The realization of a Harper-Hofstadter strip by means of a 1D lattice with a synthetic spin dimension. See the main text for details. Reproduced with permission from [@2015ManciniScience]. Copyright 2015, Science.[]{data-label="Fig11"}](Fig11.jpg){width="1\linewidth"} An excellent demonstration of these possibilities is given by the experiment by the Fallani group in LENS [@2015ManciniScience]. Using $^{173}$Yb atoms in a 1D lattice, with Raman laser coupling between two or three distinct spin states, the authors realized a two- or three-leg ladder geometry (see Fig. \[Fig11\] for a pictorial view). For the two-leg ladder case, spin-resolved measurement of momentum distributions revealed the presence of edge chiral currents, travelling in opposite directions along the two legs (Fig. \[Fig12\]). These currents, which can be detected by analyzing spin-resolved momentum distributions, are analogous to the topological chiral modes running along the edge of the 2D Hofstadter lattice. In fact, it can be easier to experimentally investigate such edge-localized phenomena in such a 1D simulator as opposed to a real 2D structure, since the momentum distribution for each spin component can be measured individually [@2019CooperRevModPhys]. By increasing the number of coupled spin states from two to three, one obtains a three-leg ladder geometry, which is an even closer approximation of a strip of a 2D system. Compared to a two-leg ladder, which is “all edge and no bulk”, the three-leg ladder has a “bulk” in the form of the “central” leg. Momentum distribution measurements reveal that no net chiral current is present in this “bulk” leg (Fig. \[Fig13\]). The experiment serves as a remarkable demonstration of how the physics in a 1D spin-orbit coupled lattice system can be mapped onto those of a 2D system. ![Uncovering chiral edge currents of fermionic atoms in a two-leg ladder with effective magnetic flux. (a) Top: Time-of-flight images representing the momentum distribution of atoms in the two pseudospin states $m=-5/2$ and $m=-1/2$. Middle: Integrated momentum distributions $n(k)$. Bottom: The imbalances $h(k) = n(k) - n(-k)$. The nonzero imbalance reveals the presence of a chiral current for the atoms in a given pseudospin state, with the opposite directions for both pseudospins. (b) The momentum distribution and imbalance $h(k)$ for atoms in the $m = -1/2$ state, for two opposite directions of the effective magnetic field. It can be seen that the direction of the chiral current is inverted as the effective field is turned in the opposite direction. (c) Visualization of the two chiral currents (orange arrows) along the two legs of the ladder that correspond to the two spin states. Reproduced with permission from [@2015ManciniScience]. Copyright 2015, Science.[]{data-label="Fig12"}](Fig12a.jpg "fig:"){width="\linewidth"} ![Uncovering chiral edge currents of fermionic atoms in a two-leg ladder with effective magnetic flux. (a) Top: Time-of-flight images representing the momentum distribution of atoms in the two pseudospin states $m=-5/2$ and $m=-1/2$. Middle: Integrated momentum distributions $n(k)$. Bottom: The imbalances $h(k) = n(k) - n(-k)$. The nonzero imbalance reveals the presence of a chiral current for the atoms in a given pseudospin state, with the opposite directions for both pseudospins. (b) The momentum distribution and imbalance $h(k)$ for atoms in the $m = -1/2$ state, for two opposite directions of the effective magnetic field. It can be seen that the direction of the chiral current is inverted as the effective field is turned in the opposite direction. (c) Visualization of the two chiral currents (orange arrows) along the two legs of the ladder that correspond to the two spin states. Reproduced with permission from [@2015ManciniScience]. Copyright 2015, Science.[]{data-label="Fig12"}](Fig12b.jpg "fig:"){width="\linewidth"} ![Same as Fig. \[Fig12\], but for a ladder with three legs corresponding to the pseudospin states $m=-5/2$, $m=-1/2$ and $m=+3/2$. Chiral currents are only present for atoms on the two “edge” legs ($m=-5/2,+3/2$), while the “middle leg” ($m=-1/2$) is characterized by a net zero current. Reproduced with permission from [@2015ManciniScience]. Copyright 2015, Science.[]{data-label="Fig13"}](Fig13a.jpg "fig:"){width="\linewidth"} ![Same as Fig. \[Fig12\], but for a ladder with three legs corresponding to the pseudospin states $m=-5/2$, $m=-1/2$ and $m=+3/2$. Chiral currents are only present for atoms on the two “edge” legs ($m=-5/2,+3/2$), while the “middle leg” ($m=-1/2$) is characterized by a net zero current. Reproduced with permission from [@2015ManciniScience]. Copyright 2015, Science.[]{data-label="Fig13"}](Fig13b.jpg "fig:"){width="\linewidth"} With regard to artificial ladder geometries, it is worth mentioning that a more complex ladder structure was recently achieved experimentally in Seul [@2018KangPRL]. In a 1D lattice with $^{173}$Yb atoms, the authors realized a three-leg *cross-linked ladder*: a ladder that allows hopping between lattice sites with a simultaneous change of orbital, corresponding to diagonal hopping across the ladder plaquettes (Fig. \[Fig14\]b). The system was implemented by overlaying the trapping optical lattice with a periodically oscillating lattice potential, generated by a pair of Raman lasers with different frequencies. This induced couplings between the first few excited orbitals of the optical lattice sites, which played the role of pseudospin, so that an “orbital-momentum” coupling played the role of spin-orbit coupling (Fig. \[Fig14\]a). The diagonal hopping was achieved by ensuring a significant overlap of the orbital wave functions corresponding to each site. ![(a) Schematic of the experimental setup realizing orbital-momentum coupling in a 1D optical lattice. The stationary 1D lattice potential $V(x)$ with lattice constant $\pi/k_L$ is overlaid with an oscillating lattice potential $\delta V(x,t)$ with lattice constant $\pi/k_R$. The moving lattice induces two-photon Raman transitions between the different orbital states of the lattice. (b) The lattice as a ladder, with the orbital states $s,p,d$ playing the role of the synthetic dimension. The particles can hop along the real dimension $x$ (black solid lines) and along the orbital dimension $\alpha$ (green solid lines), but “diagonal” hopping is possible as well (dashed lines). Because of the spatial modulation of the complex tunneling amplitude, an effective magnetic flux $\Phi = 2\pi(k_R/k_L + 1)$ per plaquette is created. The diagonal hopping additionally divides each plaquette into four sub-plaquettes with the magnetic flux distributed between them. Reproduced with permission from [@2018KangPRL]. Copyright 2018, American Physical Society.[]{data-label="Fig14"}](Fig14.jpg){width="1\linewidth"} The rapid experimental development in this area has been accompanied by theoretical developments as well. The ladder structures can also be used for more involved applications. For example, as proposed in [@2015MazzaNJP], a two-leg ladder with the two legs interpreted as “particle” and “hole” states can exhibit properties similar to that of a topologically nontrivial superconducting wire. Another interesting concept has been presented in [@2016MugelPRA] where it was shown how a topologically nontrivial system can be implemented via a quantum walk of ultra-cold atoms on a 1D lattice. It was argued that in certain parameter regimes the system can be mapped onto the Creutz ladder [@1999CreutzPRL]. Topological superfluids ----------------------- We now move on to another possibility opened by the application of SO coupling to 1D fermions, namely, the creation of topological superfluid phases. A topological superfluid phase features Cooper pairing between the fermions (analogously to the BCS phase) but also displays nontrivial topological characteristics. In particular, it can host zero-energy edge states with properties analogous to properties of the famous Majorana fermions – non-existing, but theoretically possible realizations of neutral particles obeying fermionic statistics being compatible with the relativistic quantum mechanics [@1937MajoranaNCimento; @1996RyderBook2nd]. In contrast to standard Dirac particles, Majorana fermions are their own antiparticles. Although Majorana particles were never observed as quantum particles, it is commonly argued that in some specific scenarios they may give an effective and appropriate description of excitations of many-body systems. In such cases, they are particularly interesting from a quantum information perspective, as they are highly resistant to decoherence and have been suggested as a vital element in fault-tolerant quantum computation [@2001KitaevPhysUsp; @2008NayakRevModPhys]. Topological superfluids represent a significant opportunity to generate Majorana fermions controllably. Majorana fermions are known to occur effectively in certain 2D superconductors characterized by $p$-wave interparticle interactions [@2001IvanovPRL]. A conventional 2D $s$-wave superconductor under artificial spin-orbit coupling can also harbor Majorana fermions [@2008ZhangPRL; @2009SatoPRL; @2010SatoPRB]. One-dimensional topological superfluids have been successfully created in heterostructures, consisting of a 1D semiconductor wire subject to strong spin-orbit coupling and brought in the proximity of a bulk $s$-wave superconductor [@2010OregPRL; @2011LutchynPRL; @2011FidkowskiPRB]. Experiments with such structures have uncovered evidence for the appearance of Majorana fermions in the wire [@2012MourikScience; @2012RokhinsonNatPhys; @2012DengNanoLetters; @2012DasNatPhys; @2013FinckPRL]. Schemes for creating Majorana fermions in ultra-cold atoms have been proposed as well, for example in systems of spin-orbit coupled 1D fermions inside a background 3D BEC [@2011JiangPRL]. In recent years, a number of theoretical studies have explored the possibility of using SO coupling to obtain a topological superfluid phase in a solitary 1D system of attracting ultra-cold fermions, without the need to couple to external systems. In [@2012LiuPRA-Topological; @2012WeiPRA] the case of a 1D fermionic gas in a harmonic trap, subjected to spin-orbit coupling and a Zeeman field, was analyzed. It was found that when the Zeeman field and the spin-orbit coupling are strong enough, the system can pass from a topologically trivial BCS superfluid phase into a topological superfluid phase (Fig. \[Fig15\]). This phase supports several zero-energy edge states, which have the Majorana-like symmetry. Analogous results were obtained in [@2015YangCommunTheorPhys] for 1D gas in a lattice. Additionally, it has been shown that a topological FFLO superfluid state, with a non-uniform pairing order parameter, can be obtained in this setup as well [@2013LiuPRA-Topological; @2013ChenPRL; @2017WangPRA]. ![The smallest eigenenergy $\min\{|E_\eta|\}$ of the Bogoliubov quasiparticle spectrum of a one-dimensional Fermi gas as a function of the Zeeman field $h$ under presence of the SO coupling of a fixed strength. Along with increasing Zeeman field the system transitions from a topologically trivial BCS superfluid first to a topological superfluid, and finally to a normal state. Insets show the quasiparticle energy spectrum at $h/E_F = 0.3$ and $0.5$. It can be seen that near-zero-energy edge modes are present in the topological superfluid phase. Reproduced with permission from [@2012LiuPRA-Topological]. Copyright 2012, American Physical Society.[]{data-label="Fig15"}](Fig15.jpg){width="1\linewidth"} Other works have investigated the possibility of manipulating the Majorana fermions generated in the topological superfluid. In particular, it has been proposed that Majorana fermions could be moved through the trap by manipulating the Zeeman field strength. Bound Majorana-like states can be generated at desired locations as well, by inserting impurities into the system [@2012LiuPRA-Manipulating; @2013LiuPRA-Impurity]. Other theoretical proposals involve dark solitons in the superfluid, which can also support Majorana fermions bound to their locations. Thus, they provide an indirect way to manipulate Majorana fermions or to identify the topological nature of the state through the filling status of the solitons [@2014XuPRL; @2015LiuPRA; @2019FanPRA]. It is also worth noting an interesting proposal for obtaining topological superfluids which was made in [@2015YanSciRep]. The work considers a 1D lattice with SO coupling that realizes a ladder geometry. It is proposed that two chiral edge states on the opposite legs of the ladder can undergo Cooper pairing, leading to a BCS-like superfluid phase with zero-energy Majorana modes localized at ends of the lattice. Finally, we wish to draw attention to a recent experimental realization of a symmetry-protected topological (SPT) phase with fermionic $^{173}$Yb atoms in a 1D lattice in [@2018SongSciAdv]. SPT phases are a subset of topological phases, distinguished by the fact that, while ordinary topologically ordered phases are robust against any local perturbations, SPT phases remain intact only against perturbations that respect specific protecting symmetries. Theoretical schemes for obtaining SPT phases in 1D Fermi systems were considered for spin-orbit coupled fermions in a Raman lattice [@2013LiuPRL] as well as for 1D fermions with SO coupling induced by Raman laser couplings [@2017ZhouPRL]. From the experimental point of view, the Raman lattice scheme was implemented and an SPT phase was successfully created in the 1D Fermi system [@2018SongSciAdv]. Strikingly, when the confining lattice potential was spin-dependent, the obtained topological phase was one of a new, exotic type, outside of the traditional Altland-Zirnbauer classification [@1997AltlandPRB] which is typically used to classify 1D SPT phases. Higher-spin systems {#sec-sun} =================== Due to the obvious historical reasons, most research on fermionic systems concerns spin-1/2 systems with only two distinct spin states, governed by a SU(2) symmetry. In this way, a very close analogy to the electronic systems is kept. However, current experimental achievements in the field of atomic physics allow the exploration of higher-spin systems, which can be used to realize a rich variety of interesting phases being completely beyond the range of solid-state physics [@2014CazalillaRepProgPhys; @2016CapponiAnnPhys; @2019SowinskiRepProgPhys]. The physics of higher $\mathrm{SU}({\cal N})$ symmetries are of interest to many branches of physics and can lead to new connections with high-energy physics. For instance, an $\mathrm{SU}(3)$ symmetry underlies the description of quarks in quantum chromodynamics [@2007GreinerBook], while an $\mathrm{SU}(6)$ symmetry has been used to describe the flavor symmetry of spinful quarks [@1964SakitaPR]. Study of $\mathrm{SU}(4)$ symmetry can be applied to study electron systems with orbital degeneracy [@1998LiPRL; @1999FrischmuthPRL; @1999AzariaPRL]. From a theoretical point of view, the study of 1D $\mathrm{SU}({\cal N})$ fermions dates back to the work by Sutherland [@1968SutherlandPRL], who extended Gaudin and Yang’s 1D fermionic gas model [@1967GaudinPLA; @1967YangPRL] to an arbitrary number of spin components, $\cal N$, giving the solution in terms of ${\cal N}$ nested Bethe ansatzes. The ground-state solution of the attractive case, which has the form of a ${\cal N}$-particle bound state, was given in [@1970TakahashiProgTheorPhys]. Since then, 1D $\mathrm{SU}({\cal N})$ multicomponent fermionic systems have been thoroughly explored theoretically. Examples include the three-component $\mathrm{SU}(3)$ Fermi gas [@2008GuanPRL; @2010HePRA; @2012KuhnNJP], $\mathrm{SU}(4)$ spin-3/2 fermions [@2008CapponiPRA; @2009GuanEPL; @2012SchlottmannPRB-Spin32], and systems with even higher symmetries [@2005SzirmaiPRB; @2007BuchtaPRB; @2008LiuPRA; @2008SzirmaiPRB; @2010GuanPRA; @2011ManmanaPRA; @2011YinPRA; @2012SchlottmannPRB-Arbitrary]. General reviews concerning ultra-cold fermionic systems with higher spin symmetries can be found in [@2014CazalillaRepProgPhys; @2016CapponiAnnPhys]. Achieving enlarged $\mathrm{SU}({\cal N})$ symmetry in condensed matter systems is usually very difficult, as it requires fine-tuning of the interaction parameters. Ultra-cold fermionic systems, thanks to their tunability, offer a very good environment for studying the higher $\mathrm{SU}({\cal N})$ physics experimentally. Towards $\mathrm{SU}({\cal N})$ symmetry ---------------------------------------- Let us look closely at the conditions necessary for obtaining $\mathrm{SU}({\cal N})$ symmetry with spin-$F$ ultra-cold atoms. A two-body interaction between two spin-$F$ fermions depends on their total spin $f$, which can assume possible values $f = 0,1,2... 2F$. Due to fermionic quantum statistics, only even $f$ values allow for interactions via $s$-wave collisions. As a result, the system exhibits $F + 1/2$ distinct $s$-wave scattering lengths $a_f = a_0, a_2, ... a_{2F-1}$. The $\mathrm{SU}({\cal N})$ symmetry is obtained only when all these scattering lengths are simultanously equal [@2014CazalillaRepProgPhys]. Alkaline-earth atoms, and atoms with a similar electronic structure such as ytterbium, are particularly well suited to this purpose [@2009CazalillaNJP; @2010GorshkovNatPhys; @2014CazalillaRepProgPhys]. In the ground state ${}^1\mathrm{S}_0$ of alkaline-earth atoms, as well as in the metastable excited state ${}^3\mathrm{P}_0$, the total electronic angular momentum is zero. As a result, the hyperfine interaction vanishes and the electronic shell configuration becomes decoupled from the nuclear spin. Since the differences in $a_f$ depend mainly on the electronic wave functions of the colliding atoms, this decoupling causes $a_f$ to become almost independent of the nuclear spin. More precisely, the nuclear-spin-dependent correction of the scattering lengths is on the order of $\sim 10^{-9}$ in the $^1\mathrm{S}_0$ state and on the order of $\sim 10^{-3}$ in the $^3\mathrm{P}_0$ state [@2010GorshkovNatPhys]. Thanks to this independence of scattering lengths on the spin, the system effectively exhibits a SU(2F+1) symmetry [@2016CapponiAnnPhys]. Alkaline-earth and alkaline-earth-like atoms have been successfully used to experimentally realize systems with higher $\mathrm{SU}({\cal N})$ symmetries. In particular, experiments in 3D and 2D setups have realized SU(6) symmetry with $^{173}$Yb atoms [@2010TaiePRL; @2012TaieNatPhys; @2014ScazzaNatPhys] as well as SU(10) symmetry with $^{87}$Sr atoms [@2013StellmerPRA; @2014ZhangScience]. One-dimensional realizations ---------------------------- ![Experimental creation of one-dimensional $\mathrm{SU}({\cal N})$ fermionic systems with tunable number of spin components. (a) A 2D optical lattice is used to create an array of independent 1D tubes of ultra-cold $^{173}$Yb atoms with up to six different nuclear spin orientations. (b) The number of spin components is fully tunable and can be determined via optical Stern-Gerlach detection. Reproduced with permission from [@2014PaganoNatPhys]. Copyright 2014, Nature.[]{data-label="Fig16"}](Fig16.jpg){width="1\linewidth"} For 1D systems, a breakthrough experimental achievement was performed by the Fallani group [@2014PaganoNatPhys]. In this experiment, a one-dimensional liquid of repulsively interacting $^{173}$Yb atoms with an arbitrarily tunable number of spin components was obtained (Fig. \[Fig16\]a). The number of spin components $\cal N$ was set during the preparation of the sample, by means of optical spin manipulation and detection techniques. The authors have explored the physics of this system for a varying number of components, from ${\cal N} = 1$ to ${\cal N} = 6$ (Fig. \[Fig16\]b), while keeping the number of atoms per spin component constant. It was found that with an increasing number of spin components, the system properties deviate from those of a spin-1/2 Luttinger liquid that is typically used to describe 1D fermionic systems [@2003GiamarchiBook]. In particular, as $\cal N$ increases, the Pauli principle is increasingly less important and the system gradually takes on the properties of a system of spinless bosons (confirming earlier theoretical predictions [@2011YangCPL; @2012GuanPRA]). This was experimentally confirmed by measuring the frequency of breathing oscillations of the cloud after a sudden change of the trap frequency. The authors have also analyzed the momentum distribution of the system and showed that it broadens monotonically as the number of components is increased. This can be explained qualitatively: as the number of spin components increases, the role of repulsions between the atoms is increased, which decreases the space available to the atoms (in a manner similar to the Pauli repulsion) and forces the occupation of higher-momentum states. The authors also probed the excitation spectra by means of the Bragg spectroscopy, finding that, for larger numbers of components, the results for the excitation frequency deviate from the predictions of the Luttinger liquid theory. Plethora of various phases -------------------------- ![The ground state phase diagram of a spin-3/2 $\mathrm{SU}(4)$ (upper plot) and spin-5/2 $\mathrm{SU}(6)$ (bottom plot) homogeneous one-dimensional Fermi gas with attractive interactions, in the plane of chemical potential $\mu$ vs. magnetic field $H$. The roman numbers indicate different possible phases made up of bound states of the corresponding number of fermions. The unlabelled region is the vacuum state. Multiple roman numbers added together indicate mixed phases with coexistence of different states. The vertical dashed lines indicate the trajectories of local chemical potential for systems in a harmonic trap. For clarity, compare to the $\mathrm{SU}(2)$ phase diagram in Fig. \[Fig1\]. Both figures reproduced with permission from [@2012SchlottmannPRB-Spin32] and [@2012SchlottmannPRB-Arbitrary], respectively. Copyright 2012, American Physical Society.[]{data-label="Fig17"}](Fig17a.jpg "fig:"){width="1\linewidth"} ![The ground state phase diagram of a spin-3/2 $\mathrm{SU}(4)$ (upper plot) and spin-5/2 $\mathrm{SU}(6)$ (bottom plot) homogeneous one-dimensional Fermi gas with attractive interactions, in the plane of chemical potential $\mu$ vs. magnetic field $H$. The roman numbers indicate different possible phases made up of bound states of the corresponding number of fermions. The unlabelled region is the vacuum state. Multiple roman numbers added together indicate mixed phases with coexistence of different states. The vertical dashed lines indicate the trajectories of local chemical potential for systems in a harmonic trap. For clarity, compare to the $\mathrm{SU}(2)$ phase diagram in Fig. \[Fig1\]. Both figures reproduced with permission from [@2012SchlottmannPRB-Spin32] and [@2012SchlottmannPRB-Arbitrary], respectively. Copyright 2012, American Physical Society.[]{data-label="Fig17"}](Fig17b.jpg "fig:"){width="1\linewidth"} In anticipation of future experiments, we will now point out the various interesting phases possible to create in high-spin $\mathrm{SU}({\cal N})$ systems. One characteristic aspect of multicomponent systems is the richness of their phase diagrams, which admits new types of binding beyond pair formation. Systems with ${\cal N}>2$ components allow the possibility of three-fermion (trions), four-fermion (quartets) and even larger clusters, as well as mixed phases with various combinations of such clusters [@2009GuanEPL]. To see an example of the rich possibilities, we may consider the higher-spin equivalent of the FFLO system considered in Section \[sec-fflo\]. Let us consider a homogeneous gas with attractive interactions and a spin imbalance, subjected to a magnetic field $H$. The phase diagram for such a system with spin-1/2 was considered in the previous sections (Fig \[Fig1\]). The phase diagrams for equivalent systems with higher spin symmetry are shown in Fig. \[Fig17\]. The spin-1/2 system admits three phases – a polarized phase of singlet atoms, a fully paired phase, and a partially polarized phase which contains both pairs and unpaired atoms. However, a system with ${\cal N}>2$ spin components admits more phases – a polarized phase of singlet atoms, a phase consisting of $\cal N$-fermion clusters, and numerous mixed phases in which various combinations of clusters with different particle numbers coexist. The resulting phase diagram is highly complex [@2012SchlottmannPRB-Spin32; @2012SchlottmannPRB-Arbitrary]. Such a complex phase diagram also allows for highly intricate phase separation in inhomogeneous systems, since a cut across the phase diagram can cross many phase boundaries [@2012SchlottmannPRB-Arbitrary]. Fig. \[Fig18\] shows an example of a complicated phase separation structure in a trapped spin-3/2 attractive gas. The phase-separated system can be described as a four-shell structure, displaying four distinct phases. It is useful to compare this case with the case of a trapped spin-1/2 attractive gas, which, as noted in Section \[sec-fflo\], can be described by a two-shell structure with one phase in the center and another in the wings. ![Radial structure of a spin-3/2 $\mathrm{SU}(4)$ one-dimensional Fermi gas with attractive interactions trapped in a harmonic trap. Roman numbers indicate the phases present in the different regions of the trap. Black lines labelled with $n_q$ indicate densities of $q+1$-particle states, line labelled with $N_p/L$ is the total particle density, line labelled with $M/L$ is the magnetization density. A complex phase separation pattern is clearly visible. Reproduced with permission from [@2012SchlottmannPRB-Spin32]. Copyright 2012, American Physical Society.[]{data-label="Fig18"}](Fig18.jpg){width="1\linewidth"} Higher-spin systems also allow the possibility of extending the idea of the FFLO phase to larger fermion clusters. In the $\mathrm{SU}(4)$ system made up of an array of tubes with weak tunneling between them, a theoretical calculation shows that the mixed phases (where clusters of different length coexist) can display characteristic FFLO-like oscillations in the order parameters [@2012SchlottmannPRB-Spin32]. A recent theoretical investigation [@2017SzirmaiPRA] of the phase diagram of the $\mathrm{SU}(4)$ attractive Hubbard model at quarter filling has found two distinct FFLO-like phases, a “paired-FFLO” and a “quartet-FFLO” phase. The latter is an equivalent of the normal FFLO state, for bound particle quartets as opposed to pairs. The “quartet FFLO” phase appears at lower interaction strengths, but at higher interactions, it transitions into a phase-separated state where quartets and pairs coexist (see [@2017SzirmaiPRA] for a detailed discussion). Orbital physics of higher-spin fermions --------------------------------------- Intriguing physics can be revealed in models which, in addition to the nuclear spin degree of freedom, explicitly take into account an additional orbital degree of freedom. For alkaline-earth atoms, the most natural candidates for this orbital degree of freedom are the electronic ground and excited states $|g\rangle ={}^1\mathrm{S}_0$ and $|e\rangle ={}^3\mathrm{P}_0$ [@2010GorshkovNatPhys]. Without breaking the $\mathrm{SU}({\cal N})$ symmetry, such a system allows for four distinct interaction strengths depending on the orbital states of the interacting fermions. Systems with this kind of two-orbital dynamics have been studied experimentally in 3D settings [@2014ScazzaNatPhys; @2014ZhangScience]. The physics of such a two-orbital higher-spin system in the one-dimensional case has been theoretically explored in [@2013SzirmaiPRB] where the authors analyzed the case of atoms with $\mathrm{SU}(10)$ symmetry. At incommensurate filling, the phase diagram in the plane of different interaction strengths is quite intricate. Interestingly, the system presents the possibility of realizing a novel form of an FFLO state, where the finite momentum of the pairs does not come from the spin imbalance but rather from the difference of Fermi momenta of the two orbital states. Instead of using the two $|g\rangle = {}^1\mathrm{S}_0$ and $|e\rangle = {}^3\mathrm{P}_0$ states, an alternative way to realize a two-orbital system is to exploit the transverse single-particle modes of the trapping potential which is used to realize quasi-one-dimensional geometry. Specifically, if the atoms may occupy the first-excited degenerate states $p_x$, $p_y$ of the transverse potential, these states can play the role of the two orbitals [@2012KobayashiPRL; @2014KobayashiPRA]. The phase diagram for a two-orbital fermionic system confined in a 1D lattice with incommensurate filling was explored theoretically in [@2016BoisPRB], for both the ${}^1\mathrm{S}_0/{}^3\mathrm{P}_0$ and the $p_x/p_y$ two-orbital models. Under certain circumstances, two-orbital higher-spin systems mentioned above may support topologically nontrivial phases, including symmetry-protected ones. For example, as shown in [@2013NonneEPL], the interplay between the orbital and nuclear spin degree of freedom for a one-dimensional optical lattice system at half-filling may lead to an interesting analogue of the Haldane phase [@2011CarrBook]. Creating symmetry-protected topological phases in SU(N) 1D lattice systems was also explored in [@2015BoisPRB-Phase; @2018UedaPRB]. This path of exploration is still ongoing and awaits experimental confirmation. Conclusion {#sec-concl} ========== It is a matter of fact that the one-dimensional many-body quantum systems are no longer only theoretical divagations. Due to the rapid experimental progress in controlling atoms and molecules in the ultra-cold regime, they become realistic systems having their own and very often exotic properties. It is highly possible that these unique features will find unconventional applications in the nearest future and will change many technological aspects. With this short review, we have summarized current progress in three directions which, in our opinion, are important not only from the point of technological exploitation but also have fundamental importance for developing our understanding of quantum many-body systems. This work is supported by the (Polish) National Science Center through Grant No. 2016/22/E/ST2/00555. [**Conflict of interest**]{} The authors declare no conflict of interest. [**Keywords**]{} quantum simulators, one-dimensional systems, unconventional pairing, spin-orbit coupling, higher-spin systems

ArXiv

--- abstract: 'Using the second order Israel-Stewart hydrodynamics we discuss the effect of viscosity on photon production in a parton plasma created in relativistic heavy ion collisions. We find that photon production rates can enhance by several factors due to the viscous effect in a chemically nonequilibrated plasma.' author: - 'Jitesh R. Bhatt[[^1]]{} and V. Sreekanth[[^2]]{}' title: Photon emission from out of equilibrium dissipative parton plasma --- A strongly coupled quark-gluon plasma (sGGP) or a matter in a perfect fluid state is widely expected to be produced in recent Relativistic Heavy Ion Collider (RHIC) experiments. Recent measurements has shown that the matter flows very rapidly at the time of its breakup into the freely streaming hadronic matter. Also the measurements of the elliptical flow parameter $v_2$ show a strong collectivity in the flow [@star; @phen; @phobos]. This would imply that the QGP can have a very low shear viscous stress and the ratio of its shear viscosity $\eta$ to the entropy density $s$ i.e. $\eta/s$ should not be much larger than the lower bound $1/4\pi$ [@kss05]. This led to a conjecture that the QGP formed at RHIC is the most perfect-fluid found in nature [@Hirano:2005wx]. There has been a lot of attempts to determine the viscosity of sQGP [@Hirano:2005wx; @bmpr90; @hk85; @heisel94; @AYM; @XUPRL]. The first order theory of viscous hydrodynamics is known to give unphysical results. For example, when the Navier-Stokes equations were applied to the one dimensional boost invariant expanding flow [@bjor], one finds the expression for the temperature to be $$T(\tau)=T_0 \left(\frac{\tau_0}{\tau}\right)^{1/3}\left[1+\frac{2\eta}{3 s \tau_0 T_0}\left(1-\left(\frac{\tau_0}{\tau}\right)^{2/3}\right)\right]. \nonumber$$ This describes reheating of the flow and $T$ has a maximum at time $\tau_{\rm max} = \tau_0 \left( \frac{1}{3} + \frac{s}{\eta} \frac{\tau_0\, T_0}{2} \right)^{-3/2} $. This is an unphysical behaviour and the first order viscous dynamics is known to have such problems [@lindblom; @Baier:2006um]. The second order hydrodynamics approach developed in the spirit of Israel and Stewarts [@Israel:1979wp] removes such artifacts. The second order viscous hydrodynamics later developed and applied in the context of heavy-ions collisions by several authors [@br07; @Teaney03; @SongHeinzCh06; @hs08; @Hama01; @AM207; @R97; @DT08]. It would be interesting to study the role that viscosity can play on the plasma signals. Hard photons are one such promising source that can provide information about the thermodynamical state of the plasma at time of their production. The plasma created in the heavy-ion collisions is expected to be in a state of chemical nonequilibrium. The photon emission from such a plasma has been studied within the framework of ideal hydrodynamics by earlier workers [@thoma; @dmkc00; @bdrs1997; @long05]. In this paper we study the photon production using causal hydrodynamics of Israel-Stewart [@Israel:1979wp]. In the center of the fireball in a nuclear collision the viscous stress-energy tensor in the local comoving frame has the form [@Teaney03; @Muronga:2001zk; @Muronga:2004sf]: $$T^{\mu\nu} = \left( \begin{array}{cccc} \varepsilon & 0 & 0 & 0 \\ 0 & P_{\perp} & 0 & 0 \\ 0 & 0 & P_{\perp} & 0 \\ 0 & 0 & 0 & P_{z} \end{array} \right) \label{eq:Tmunu}$$ with the transverse and longitudinal pressure $$\begin{aligned} P_{\perp} &=& P + \frac{1}{2}\Phi \nonumber \\ P_{z} &=& P - \Phi \label{eq:stress}\end{aligned}$$ Here $P$ denotes the (isotropic) pressure in thermal equilibrium, $\Phi$ denotes the non-equilibrium contributions to the pressure coming from shear stress. We ignore the bulk viscosity in the relativistic limit when the equation of state $p=\epsilon/3$ is obeyed [@wein]. However, the bulk viscosity can be important near the critical temperature [@fms08; @karsch07]. The shear tensor in that frame takes the form $\pi^{ij} = \mathrm{diag}(\Phi/2, \Phi/2,-\Phi)$ consistent with the symmetries in the transverse directions. To describe evolution of the energy density and the viscous stress $\phi$ we use second order dissipative hydrodynamics of Israel-Stewart [@Israel:1979wp; @Heinz:2005zi; @Muronga:2003ta; @emxg08]: $$\begin{aligned} \frac{\partial\varepsilon}{\partial\tau} &=& - \frac{1}{\tau}(\varepsilon + P - \Phi) \, , \label{eq:evol1} \\ \frac{\partial\Phi}{\partial\tau} &=& -\frac{\Phi}{\tau_{\pi}}-\frac{1}{2}\Phi \left( \frac{1}{\tau} + \frac{1}{\beta_2}T\frac{\partial}{\partial\tau}(\frac{\beta_2}{T}) \right) + \frac{2}{3}\frac{1}{\beta_2\tau} \,,\end{aligned}$$ where $\beta_2=9/(4\varepsilon)$ and $\tau_{\pi}=2\beta_2\eta$ denotes the relaxation time. Equations(3-4) are written in the local rest frame using hydrodynamic velocity $u^\mu=\frac{1}{\tau}(t,0,0,z)$, where $\tau=\sqrt{t^2-z^2}$ [@bjor]. Equation of state is required to solve these equations. We use ultra-relativistic equation of state : $P=\frac{1}{3}\epsilon$. To describe the chemical non-equilibration while maintaining the kinetic equilibrium, one can use the parton distribution [@levai], $$f(k,T)_{q,g}\,=\, \lambda_{q,g}(\tau)\frac{1}{e^{{\mathbf {u\cdot k}}/T(\tau)}\,\pm\,1}$$ where, $u^\mu$ is the four-velocity of the local comoving reference frame. The temperature $T$ is a time-dependent quantity and the distribution is multiplied by time and another dependent quantity called fugacity $\lambda_{q,g}(\tau)$ to describe deviations from the chemical equilibrium. The fugacity parameter become unity when the chemical-equilibrium is reached and in general it has the range $ 0\le \lambda_{q,g}\le 1$. The scattering processes $gg\leftrightarrow ggg $ and $gg\leftrightarrow q\bar{q}$ give the most dominant mechanism for the chemical equilibration. The master equations describing evolution the parton density are given by $$\begin{aligned} \partial_\mu(n_gu^\mu)&=&\frac{1}{2}\sigma_3n^2_g\left(1-\frac{n_g}{ \tilde{n}_g}\right) -\frac{1}{2}\sigma_2n^2_g\left(1-\frac{n^2_q\tilde{n}^2_g}{\tilde{n}^2_qn^2_g}\right),\\ \partial_\mu(n_qu^\mu)&=&\frac{1}{2}\sigma_2n^2_g \left(1-\frac{n^2_q\tilde{n}^2_g}{\tilde{n}^2_qn^2_g}\right),\end{aligned}$$ where $\tilde{n}_i$($i=q,g$) is parton density with unit fugacity [@biro] and $\sigma_2\,=\langle\sigma(gg\leftrightarrow q\bar{q})\rangle$ and $\sigma_3\langle\sigma(gg\leftrightarrow ggg)\rangle$ are thermally averaged scattering cross sections. It should be noted here that when equation for $n_g$ and $n_q$ are added one gets the total number density $n$ and the term with $\frac{1}{2}\sigma_2n^2_g \left(1-\frac{n^2_q\tilde{n}^2_g}{\tilde{n}^2_qn^2_g}\right)$ will drop out. This is because due to the the scattering process $gg\leftrightarrow q\bar{q}$ loss in the gluon density is equal to the gain in quark density and vice verse. $\epsilon$ and $n$ can be calculated using equation (5) as given below $$\begin{aligned} n=(\lambda_g a_1 + \lambda_qb_1)T^3 , \,\,\,\,\epsilon=(\lambda_g a_2 + \lambda_q b_2)T^4\end{aligned}$$ where $a_1={16\xi(3)}/{\pi^2}$, $a_2=8\pi^2/15$ for the gluons and $b_1=9\xi(3)N_f/\pi^2$, $b_2=7\pi^2N_f/20$ for the quarks. Using equations (3-4,6-8) following evolution equations for $T$, $\lambda_{q,g}$ and $\Phi$ can be obtained $$\begin{aligned} \frac{\dot{T}}{T}+\frac{1}{3\tau}&=&-\frac{1}{4} \frac{\dot{\lambda}_g+b_2/a_2\dot{\lambda}_q} {{\lambda}_g+b_2/a_2{\lambda}_q} + \frac{\Phi}{4\tau}\frac{1}{(a_2{\lambda}_g+b_2{\lambda}_q)T^4} \, , \label{eq:tempevol1} \\ \dot{\Phi}+\frac{\Phi}{\tau_\pi}&=&\frac{8}{27\tau} \left[ a_2\lambda_g+b_2\lambda_q\right]T^4-\frac{\Phi}{2} \left[\frac{1}{\tau}-5\frac{\dot{T}}{T}- \frac{\dot{\lambda}_g+b_2/a_2\dot{\lambda}_q} {{\lambda}_g+b_2/a_2{\lambda}_q}\right] \, ,\label{eq:shearvisco4}\\ \frac{\dot{\lambda_g}}{\lambda_g}+3\frac{\dot{T}}{T}+\frac{1}{\tau} &=&R_3\left(1-\lambda_g\right)-R_2\left(1-\frac{\lambda^2_q} {\lambda^2_g}\right) \, , \label{eq:gluonfugacityevol2}\\ \frac{\dot{\lambda_q}}{\lambda_q}+3\frac{\dot{T}}{T}+\frac{1}{\tau} &=&R_2\frac{a_1}{b_1} \left( \frac{\lambda_g}{\lambda_q}-\frac{\lambda_q}{\lambda_g} \right) \label{eq:quarkfugacityevol3} \end{aligned}$$ where, the rates $R_2=0.24N_f\alpha^2_s\lambda_gTln(5.5/\lambda_g)$ and $R_3=2.1\alpha^2_sT(2\lambda_g-\lambda^2_g)^{1/2}$ are defined as in Ref. [@levai; @biro]. We would like to note that our gluon fugacity equation (\[eq:gluonfugacityevol2\]) differs from that given in Ref. [@levai; @biro] by a fctor of two in second term in right hand side. We believe this is a typographical error. In equation (\[eq:tempevol1\]) the first term on left hand side is due to expansion of the plasma, while on the right hand side the first term describes effect of chemical nonequilibrium and second term is due to the presence of (causal) viscosity. The last term in parenthesis of equation (\[eq:shearvisco4\]) arises because of the chemical nonequilibrium process. It should be noted that equation (9) differ from that considered in Ref. [@chau00]. In their treatment first order viscous hydrodynamics is used which does not require time evolution of $\Phi$. However such treatment give unphysical results like reheating artifact [@Baier:2006um] as mentioned before. Elastic $(gg\leftrightarrow gg)$ as well as nonelastic processes like $gg\leftrightarrow ggg$ can contribute to the shear viscosity. Shear viscosity coefficient was recently calculated for the inelastic process in the presence of chemical nonequilibrium in Ref. [@emxg08]. It was shown that $\eta/T^3\simeq\,n_g/T^3\simeq \lambda_g$. From this one can write [@Muronga:2003ta] $$\tau_{\pi}=\frac{9}{2\varepsilon}\lambda_gT^3 .\label{taupimronga}$$ It ought to be mentioned that this viscosity prescription was not considered considered in Ref.[@chau00]. Kinetic theory without invoking nonequilibrium process gives $\tau_{\pi}=3/{2\pi T}$. Real photons are produced from the annihilation of a quark-antiquark pair into a photon and a gluon ($q\bar q\rightarrow g\gamma$) and by absorption of a gluon by a quark emitting a photon ($qg\rightarrow q\gamma$). Another source of photon production could be the bremsstrahlung but its effect can be ignored in the lowest order of a perturbation theory. In order to compute the photon production rates one needs to know the underlying amplitude $\mathcal{M}$ of the basic process involving the annihilation or Compton scattering process and the parton distribution functions given by [@peitzmann; @gh03] $$\begin{aligned} %\hspace*{-1cm} \frac{dN}{d^4xd^3p}&=&\frac{1}{(2\pi)^32E}\> \int \frac{d^3p_1}{(2\pi)^32E_1} \frac{d^3p_2}{(2\pi)^32E_2} \frac{d^3p_3}{(2\pi)^32E_3}\> \nonumber\\ && \times n_1(E_1)n_2(E_2)[1\pm n_3(E_3)] \\ \nonumber && \times \sum_{i} \langle |\mathcal{M}|^2\rangle \> (2\pi)^4\> \delta(P_1+P_2-P_3-P). \label{rate1}\end{aligned}$$ Here $P_1$ and $P_2$ are the 4-momenta of the incoming partons, $P_3$ of the outgoing parton, and $P$ of the produced photon. In equilibrium, the distribution functions $n_i(E_i)$ are given by the Bose-Einstein distribution, $n_B(E_i)=1/[\exp(E_i/T)-1]$, for gluons and by the Fermi-Dirac distribution, $n_F(E_i)=1/[\exp(E_i/T)+1]$, for quarks, respectively. The factor $\langle |\mathcal{M}|^2\rangle$ is the matrix element of the basic process averaged over the initial states and summed over the final states. The $\sum_{i}$ indicates the sum over the initial parton states. The fugacity factors can enter equation (14) when equation (5) is considered\ $$n_1(E)n_2(E)(1\pm\,n_3(e))\mapsto\, \lambda_1n_1\lambda_2n_2(1\pm\,\lambda_3n_3) \nonumber.$$ This is can be rewritten as $$\begin{aligned} \lambda_1n_1\lambda_2n_2(1\pm\,\lambda_3n_3)&=& \lambda_1\lambda_2\lambda_3 n_1n_2(1\pm\,n_3)\\ \nonumber &+&\lambda_1\lambda_2(1-\lambda_3)n_1n_2\end{aligned}$$ In carrying out the momentum integration it is useful to introduce a parameter $k_c$ to distinguish between soft and hard momenta of the quark [@kapu91]. For the hard part of the photon rate following [@thoma] we take $k^2_c=2m^2_q=0.22g^2T^2\left(\lambda_g+\lambda_q/2\right)$, where, $m_q$ is the quark-thermal-mass which can be obtained from zero momentum limit of quark self-energy in the high temperature limit. The first term on the right hand side of equation (15) can lead to the following photon rate [@thoma] using the Boltzmann distribution functions instead of a quantum mechanical ones: $$\left(2E\frac{dn}{d^3pd^4x}\right)_1= \frac{5\alpha\alpha_s\lambda^2_q\lambda_g}{9\pi^2}T^2e^{-E/T} \left[ln\left(\frac{4ET}{k^2_c}\right)-1.42\right].$$ Here $\alpha$ and $\alpha_s$ are the electromagnetic and the strong interaction coupling constants. The second term in equation (15) will give, under the Boltzmann approximation, the following contribution to the photon rate: $$\begin{aligned} \hspace{-1cm} \left(2E\frac{dn}{d^3pd^4x}\right)_2 &=&\\ \nonumber && \frac{10\alpha\alpha_s}{9\pi^4}T^2e^{-E/T} \left\lbrace \lambda_q\lambda_g\left(1-\lambda_q\right) \left[1-2\gamma+2ln\left(4ET/k^2_c\right)\right] +\lambda_q\lambda_q\left(1-\lambda_g\right) \left[-2-2\gamma+2ln\left(4ET/k^2_c\right)\right]\right\rbrace,\end{aligned}$$ The total photon production rate $2E\frac{dn}{d^3pd^4x}$ can be obtained by adding equations (16-17), is required to be convoluted with the space time evolution of the heavy-ion collision. We define [@thoma] $$\begin{aligned} \left(2\frac{dn}{d^2p_\bot dy}\right)_{y,p_\bot}&=&\int d^4x\left(2E\frac{dn}{d^3pd^4x}\right)\\ &=&Q\int_{\tau_0}^{\tau_1} d\tau ~\tau \int_{-y_{nuc}}^{y_{nuc}}dy^{'} \left(2E\frac{dn}{d^3pd^4x}\right)\nonumber\end{aligned}$$ where the times after the maximum overlap of the nuclei are $\tau_0$ and and $\tau_1$ and $y_{nuc}$ is the rapidity of the nuclei. $Q$ is the transverse cross-section of the nuclei and $p_\bot$ is the photon momentum in direction perpendicular to the collision axis. For a $Au$ nucleus $Q \sim 180 fm^2$. The quantity $\left(2E\frac{dn}{d^3pd^4x}\right)$ is Lorentz invariant and it is evaluated in the local rest frame in equation (18). The photon energy in this frame, i.e in the frame comoving with the plasma, can be written as $p_\bot cosh(y-y^\prime)$. -- -- -- -- In Figure (1), we have shown $T$,$\lambda_g,\lambda_q$ as function of time. We have solved the equations (9-12) together with the initial conditions at $\tau_{iso}$ from HIJING Monte Carlo model [@HIJING]. Which are $\lambda^{0}_{g}\,=\,0.09$, $\lambda^{0}_{q}\,=\,0.02$ and $T_{o}\,=\,0.57 GeV$ for RHIC with $\tau_{iso}\,=\,0.31 fm/c$ and $\lambda^{0}_{g}\,=\,0.14$, $\lambda^{0}_{q}\,=\,0.03$ and $T_{o}\,=\,0.83 GeV$ for LHC with $\tau_{iso}\,=\,0.23 fm/c$. Presence of the causal viscosity decreases the fall of temperature due to expansion and the chemical nonequilibrium. However if one considers the first order theory, there can be unphysical instability. Fugacity of gluons and quarks increase more slowly due to the presence of the viscosity compared to the cases when no viscous effects were included. This is because the chemical equilibration is reached here with falling of the temperature. The temperature can decrease due to the expansion and chemical nonequilibration. The lowering of $T$ can help in attaining chemical equilibrium and which in turn will increase the rate at which the fugacities reach their equilibrium values. Inclusion of the viscosity will slowdown the falling rate of the temperature. Consequently the fugacities will take more time to reach their equilibrium values. -- -- -- -- We plot photon spectra by using equation (\[taupimronga\]) for $\tau_{\pi}$ in solving equations (9-12). The figures (2-3) compare the case without viscosity with the case of finite shear viscosity. Figure (2) shows the photon spectra emitted at fixed rapidities as a function of transverse momenta $p_\bot$. The photon flux is normalized with the transverse size of the colliding nuclei($Q$). For LHC we take: $\tau_0\,=\,0.5 fm/c$, $\tau_1\,=6.25 fm/c$ and $y_{nuc}\,=\,8.8$. We use equation (\[taupimronga\]) for $\tau_{\pi}$ in solving equations (9-12). The figure compares the case without viscosity with the case of finite shear viscosity [@emxg08]. Figure (3) shows the comparison similar to that of figure (2) but with a set of initial conditions for RHIC: $\tau_0\,=\,0.7 fm/c$,$\tau_1 \,=\, 4 fm/c$ and $y_{nuc}\,=\,6.0$. For $\alpha_s = 0.3$, shear viscosity to entropy density ratio $\eta/s\sim 0.29$. Figures (2-3) show that viscous effects enhance the photon flux by a factor (1.5-2). -- -- -- -- Finally, we compare the photon fluxes calculated using equation (\[taupimronga\]) with the fluxes calculated using the kinetic viscosity ($\tau_{\pi}=3/{2\pi T}$). Figure (4) shows the photon flux calculated using the kinetic viscosity prescription for LHC and RHIC. However, we do not find any significant change in the flux for the results obtained using equation (\[taupimronga\]). -- -- -- -- In conclusions, we have studied the second order dissipative hydrodynamics with chemical nonequilibration. We find that the effect of viscosity enhancing the photon flux by a factor ranging between 1.5-2 for the parameter space relevant for LHC and RHIC. Our results are in a broad qualitative agreement with the results obtained in Ref.[@chau00] using the first order theory. We also find that the two viscosity prescriptions with inelastic scattering [@emxg08] and the one involving elastic collisions only using kinetic theory give similar results for the photon production rate. [10]{} K. H. Ackermann et al. \[STAR Collaboration\], Phys. Rev. Lett. [**86**]{}, 402 (2001); C. Adler et al. \[STAR Collaboration\], ibid. [**87**]{} (2001) 182301; [**89**]{}, 132301 (2002); Phys. Rev.  [**C 66**]{}, 034904 (2002); J. Adams et al. \[STAR Collaboration\], nucl-ex/0409033; J. Adams et al. \[STAR Collaboration\], Phys. Rev. Lett. [**92**]{}, 052302 (2004). K. Adcox et al. \[PHENIX Collaboration\], Phys. Rev. Lett. [**89**]{}, 212301 (2002); S. S. Adler et al. \[PHENIX Collaboration\], ibid. [**91**]{}, 182301(2003); B. B. Back et al. \[PHOBOS Collaboration\], Phys. Rev. Lett. [**89**]{}, 222301(2002), B. B. Back et al. \[PHOBOS Collaboration\], Phys. Rev. Lett. [**94**]{}, 122303 (2005), arXiv:nucl-ex/0406021, B. B. Back et al. Phys. Rev. [**C 72**]{}, 051901 (2005), arXiv:nucl-ex/0407012. P.K. Kovtun, D.T. Son and A.O. Starinets, Phys. Rev. Lett. [**94**]{}, 111601 (2005). T. Hirano and M. Gyulassy, Nucl. Phys. [**A769**]{}, 71 (2006). G. Baym, H. Monien, C. J. Pethick and D. G. Ravenhall, Phys. Rev. Lett. [**64**]{}, 1867 (1990). A. Hasoya and K. Kajantie, Nucl. Phys. [**B 250**]{}, 666 (1985). H. Heiselberg, Phys. Rev. [**D 49**]{}, 4739 (1994). P. Arnold, G. D. Moore, L. G. Yaffe, J. High Energy Phys. [**11**]{}, 001 (2000); J. High Energy Phys. [**0305**]{}, 051 (2003). Z. Xu, C. Greiner, Phys. Rev. Lett. [**100**]{}, 172301 (2008). J. D. Bjorken, Phys. Rev. [**D 27**]{}, 140 (1983). R. Baier, P. Romatschke, and U. A. Wiedemann, Phys. Rev. [**C 73**]{}, 064903 (2006). W. A. Hiscock and L. Lindblom, Phys. Rev. [**D 31**]{}, 725 (1985). W. Israel, Annals Phys.  [**100**]{}, 310 (1976); W. Israel and J. M. Stewart, Ann. Phys. [**118**]{}, 341 (1979). R. Baier, P. Romatschke, Eur. Phys. J.  [**C 51**]{}, 677 (2007). D. Teaney, Phys. Rev.  [**C 68**]{}, 034913 (2003). U. W. Heinz, H. Song, A. K. Chaudhuri, Phys. Rev. [**C 73**]{}, 034904 (2006). H. Song, U.W. Heinz, Phys. Rev. [**C 77**]{}, 064901 (2008). C. E. Aguiar, T. Kodama, T. Osada, Y. Hama, J. Phys. [**G 27**]{}, 75 (2001). A. Muronga, Phys. Rev. [**C 76**]{}, 014909 (2007). J. Brachmann, A. Dumitru, J. A. Maruhn, H. Stöcker, W. Greiner, D. H. Rischke, Nucl. Phys. [**A 619**]{}, 391 (1997). K. Dusling, D. Teaney, Phys. Rev. [**C 77**]{}, 034905 (2008). C. T. Traxler and M. H. Thoma, Phys. Rev. [**C 53**]{}, 1348 (1996). D. Dutta, A.K. Mohanty, K. Kumar and R.K. Choudhury, Phys. Rev. [**C 61**]{}, 06491 (2000). R. Baier, M. Dirks, K. Redlich and D. Schiff, Phys. Rev. [D 56]{}, 2548 (1997). J.L. Long, Z.J. He, Y.G. Ma and B.Liu, Phys. Rev. [**C 72**]{}, 064907 (2005). A. Muronga, Phys. Rev. Lett.  [**88**]{}, 062302 (2002), \[Erratum-ibid.  [**89**]{}, 159901 (2002)\]. A. Muronga and D. H. Rischke, (2004) arXiv:nucl-th/0407114. S. Weinberg, [*Gravitation and Cosmology*]{}, (John Wiley & Sons, 1972). R. J. Fries, B. Müller and A. Schäffer, Phys. Rev. [**C 78**]{}, 034913 (2008). F. Karsch, D. Kharzeev, and K. Tuchin, Phys. Lett. [**B 663**]{}, 217 (2008). D. Kharzeev and K. Tuchin, arXiv: hep-ph/0705.4280. U. W. Heinz, (2005), arXiv: nucl-th/0512049. A. Muronga, Phys. Rev. [**C 69**]{}, 034903 (2004). A. El, A. Muronga, Z. Xu and C. Greiner, arXiv: hep-ph/0812.27620. P. Lévai, B. Müller and X. -N. Wang, Phys. Rev. [**C 51**]{}, 3326 (1995). T. S. Biró, E. van Doorn, B. Müller and X. -N. Wang, Phys. Rev. [**C 48**]{}, 1275 (1993). A.K. Chaudhuri, J. Phys. G: Nucl. Part. Phys.[**26**]{} 1433 (2000). T. Peitzmann and M. H. Thoma, Phys. Rept. [**364**]{}, 175-246 (2002),arXiv: hep-ph/0111114. C. Gale and K. L. Haglin, arXiv: hep-ph/0306098v3. J. Kapusta, P. Lichard and D. Seibert, Phys. Rev. [**D 44**]{}, 2774, (1991). X. N. Wang and M. Gyluassy, Phys. Rev. [**D 44**]{}, 3501, (1991); Comput. Phys. Commun. [**83**]{}, 307 (1994). [^1]: email: [email protected] [^2]: email: [email protected]

ArXiv

--- abstract: 'We find all 2-Bridge links up to 11 crossings and locate them in Thistlethwaite’s link table. The splitting numbers of some links are calculated as a consequence of this identification.' address: 'department of mathematics, istanbul technical university, maslak, istanbul 34469, turkey' author: - Ali Sait Demir title: 'Identification of 2-Bridge Links' --- [^1] introduction ============ Various properties of links are computed and listed in KnotAtlas [@katlas] or Knotilus [@knotilus] or Knotinfo [@knotinf] databases. The aim of this study is to identify 2-bridge links inside the list of links, so that the information about 2-bridge links can be related to their representatives in Thistlethwaite’s table of links and vice versa. Two-bridge links are links with two components which can be put into the form as in Figure \[fig:2brd\], where the integers $a_1,\dots,a_n$ denote the number of overcrossings (or undercrossings). ![A 2-bridge knot or link.[]{data-label="fig:2brd"}](2br-ipe.pdf){width="0.7\linewidth"} They are also called rational links since they can be classified by the rational number $$\frac{p}{q}=a_1+\frac{1}{a_2+\frac{1}{\ddots+\frac{1}{a_n}}}$$ given by the continued fraction on the right, where $gcd(p,q)=1$. The Conway normal form $C(a)=C(a_1,\dots,a_n)$ can also be used to identify this link. Note that, for $\frac{p}{q}$ to represent a link $p$ must be even. See [@mur96] for more details on 2-bridge links. An algorithm for identifying 2-bridge links of a certain number of crossing is given in the next section. As an application, in section \[sec:split\] we calculate the splitting numbers of some 2-bridge links that have certain type of Conway normal forms. The following theorems, due to Schubert [@schu56], are used frequently in this study to eliminate equivalent links from the list of possible combinations. **(Schubert) (Theorem 9.3.3 of [@mur96])**\[thm:sch\] The 2-bridge links $L(p,q)$ and $L(p',q')$ are equivalent as unoriented links if and only if $p=p'$ and $qq'\equiv 1\ (mod\ p)$. **(Schubert) (Theorem 9.4.1 of [@mur96])**\[thm:sch2\] If the orientation of one component of the 2-bridge link $L(p,q)$, where both $p>0$ and $q>0$, is reversed the resulting link is equivalent to $L(p,q-p)=L^*(p,p-q)$. In Example \[ex1\] below, we exhibit how these theorems are used to find equivalent (up to orientation) links of 7 crossings. Identifying and listing the links ================================= In this section an algorithm to find all 2-bridge links with $n$ crossings is outlined and the results for links up to 11 crossing are listed in section \[sec:lists\]. These links are matched with their Thistlethwaite’s Id to make other data about these links relatable to their Conway normal forms or rational representations. For a similar study on the 2-bridge knots, see De Wit’s paper[@DeWit]. In order to find and identify 2-bridge links up to a certain number of crossings $n$:\ 1. Find all permutations of positive integers that add up to $n$, 2. Calculate the rational number $\frac{p}{q}$ for each such permutation, 3. Rule out the permutations giving knots instead of links, by checking the parity of $p$, 4. Shorten the list by picking one representative of permutations with equal continued fraction or equivalent link according to Schubert’s criteria (Theorem \[thm:sch\] and \[thm:sch2\]), 5. Identify the Gauss code of the link by looking at its Conway form and locate it in Thistlethwaite’s link table. The Thistlethwaite’s link table lists links up to orientations and mirror images. This means in the tables below a link may represent 2 (if there is only one orientation which is the case the link has a palindromic Conway form) or 4 (two orientations of both the link and its mirror) links. ID Link $(p,q)$ Conway Form ID Link $(p,q)$ Conway Form ----------- --------------------- ------------------------------------------------------ -- ----------- --------------------- ------------------------------------------------------ $E_{1 }$ $( 14 , 3 )$ $[ 4 , 1 , 1 , 1 ]$ $E_{ 11}$ $( 16 , 9 )$ $[ 1 , 1 , 3 , 1 , 1 ]$ $E_{ 2}$ $( 14 , 3 )$ $[ 4 , 1 , 2 ]$ $E_{ 12}$ $( 16 , 9 )$ $[ 1 , 1 , 3 , 2 ]$ $E_{ 3}$ $( 14 , 5 )$ $[ 2 , 1 , 3 , 1 ]$ $E_{ 13}$ $( 18 , 5 )$ $[ 3 , 1 , 1 , 1 , 1 ]$ $E_{ 4}$ $( 14 , 5 )$ $[ 2 , 1 , 4 ]$ $E_{ 14}$ $( 18 , 5 )$ $[ 3 , 1 , 1 , 2 ]$ $E_{ 5}$ $( 14 , 9 )$ $[ 1 , 1 , 1 , 3 , 1 ]$ $E_{ 15}$ $( 18 , 7 )$ $[ 2 , 1 , 1 , 2 , 1 ]$ $E_{ 6}$ $( 14 , 9 )$ $[ 1 , 1 , 1 , 4 ]$ $E_{ 16}$ $( 18 , 7 )$ $[ 2 , 1 , 1 , 3 ]$ $E_{ 7}$ $( 14 , 11 )$ $[ 1 , 3 , 1 , 1 , 1 ]$ $E_{ 17}$ $( 18 , 11 )$ $[ 1 , 1 , 1 , 1 , 2 , 1 ]$ $E_{ 8}$ $( 14 , 11 )$ $[ 1 , 3 , 1 , 2 ]$ $E_{ 18}$ $( 18 , 11 )$ $[ 1 , 1 , 1 , 1 , 3 ]$ $E_{ 9}$ $( 16 , 7 )$ $[ 2 , 3 , 1 , 1 ]$ $E_{ 19}$ $( 18 , 13 )$ $[ 1 , 2 , 1 , 1 , 1 , 1 ]$ $E_{10 }$ $( 16 , 7 )$ $[ 2 , 3 , 2 ]$ $E_{ 20}$ $( 18 , 13 )$ $[ 1 , 2 , 1 , 1 , 2 ]$ : All possible combinations of 7 crossings[]{data-label="tab:tab"} \[ex1\] In Table \[tab:tab\] we list all 20 possible 2-bridge links with 7 crossings and explain why these are represented by only 3 links in Table \[tab:tab7\]. Consider the links $E_1$ to $E_8$. The links with continued fractions leading to equal rational number are equivalent, hence we have $E_1=E_2$, $E_3=E_4$, $E_5=E_6$ and $E_7=E_8$. By Theorem \[thm:sch\] $E_1=E_3$ since $3\cdot5\equiv 1\ (mod \ 14)$. Similarly $E_5=E_7$. These equivalences assume that the links are not oriented. Suppose $E_3$ is given the standard orientation. According to Theorem \[thm:sch2\], if the orientation of one of the components of $E_3$ is reversed the resulting link is equivalent (as oriented links) to $E_5$. Since Thistlethwaite’s link table lists links up to orientations all these 8 links are represented by one link, namely $L_7A_6$. Similarly the links $E_9$ to $E_{12}$ and $E_{13}$ to $E_{20}$ are represented by $L_7A_4$ and $L_7A_5$, respectively. splitting numbers {#sec:split} ================= Although the classification of 2-bridge links is complete, various characteristics or local properties of these links are actively studied. In this section we will mention one of such invariants, namely the splitting number of a link. The splitting number $sp(L)$ of the link $L$ is defined to be the minimum number of crossing changes between different components of $L$ to convert $L$ into a split link. For more information about splitting numbers see [@ccz16] and [@cfp17] and references therein. In [@cfp17] authors calculate the splitting number of links up to 9 crossings. They use 5 methods based on covering properties or Alexander invariants, case by case, for determining the splitting numbers. The other study [@ccz16], due to Cimasoni et. al., calculates the splitting number by looking at the signature and nullity of the links. As a consequence of their main result the following theorem is about the splitting number of certain 2-bridge links: **(Theorem 4.7 of [@ccz16])** The splitting number of the 2-bridge link $C(2a_1,b_1,\dots,2a_{n-1},b_{n-1},2a_n)$ is $a_1+a_2+\dots+a_n$, where all $a_i$ and $b_i$ are positive integers. \[thm4.7\] Link $sp(L)$ Link $sp(L)$ Link $sp(L)$ ----------------- --------- -- ----------------- --------- -- ----------------- --------- $L_{10}A_{48}$ 2 $L_{11}A_{132}$ 2 $L_{11}A_{299}$ 4 $L_{10}A_{64}$ 3 $L_{11}A_{194}$ 3 $L_{11}A_{312}$ 4 $L_{10}A_{75}$ 3 $L_{11}A_{206}$ 3 $L_{11}A_{319}$ 4 $L_{10}A_{87}$ 3 $L_{11}A_{222}$ 3 $L_{11}A_{360}$ 5 $L_{10}A_{89}$ 4 $L_{11}A_{263}$ 4 $L_{11}A_{372}$ 5 $L_{10}A_{98}$ 4 $L_{11}A_{278}$ 4 $L_{10}A_{102}$ 4 $L_{11}A_{289}$ 4 : Splitting numbers of some 2-bridge links[]{data-label="tab:split"} Our calculations below reveal which alternating two component links are 2-bridge links. Therefore one can also calculate the splitting numbers of the following links, which turn out to be 2-bridge links with the desired Conway form, using Theorem \[thm4.7\]: $$L_5A_1,\ L_6A_1,\ L_7A_4,\ L_7A_6,\ L_8A_6,\ L_8A_8,\ L_8A_{11},\ L_9A_{18},\ L_9A_{26},\ L_9A_{30},\ L_9A_{36},\ L_9A_{40}$$ Besides the above links that already appear with splitting numbers in [@cfp17], we calculate the splitting numbers of those links in table \[tab:split\] with 10 and 11 crossings that fit into the Conway form of Theorem \[thm4.7\]: We note that the splitting number of $L_{11}A_{372}$ was also calculated in [@ccz16] (Example 4.4), but not making use of the fact that Theorem \[thm4.7\] also applies to this link. acknowledgments {#acknowledgments .unnumbered} =============== The author would like to express his gratitude to Alexander Degtyarev for introducing the subject and commenting on an earlier draft of the paper. Two-bridge links up to 11 crossings {#sec:lists} =================================== The Link $\frac{p}{q}$ or $(p,q)$ Conway Form Thistlethwaite’s Id ----------------------------------- ------------------ --------------------- (4,1) \[4\] or \[3,1\] $L_4A_1$ (8,3) \[2,1,2\] $L_5A_1$ : 2-Bridge Links of 4 and 5 Crossings The Link $\frac{p}{q}$ or $(p,q)$ Conway Form Thistlethwaite’s Id ----------------------------------- ------------------ --------------------- (6,1) \[6\] or \[5,1\] $L_6A_3$ (10,3) \[3,3\] $L_6A_2$ (12,5) \[2,2,2\] $L_6A_1$ : 2-Bridge Links of 6 Crossings The Link $\frac{p}{q}$ or $(p,q)$ Conway Form Thistlethwaite’s Id ----------------------------------- ------------- --------------------- (14,5) \[2,1,4\] $L_7A_6$ (16,7) \[2,3,2\] $L_7A_4$ (18,5) \[3,1,1,2\] $L_7A_5$ : 2-Bridge Links of 7 Crossings[]{data-label="tab:tab7"} The Link $\frac{p}{q}$ or $(p,q)$ Conway Form Thistlethwaite’s Id ----------------------------------- ------------------ --------------------- (8,1) \[8\] or \[7,1\] $L_8A_{14}$ (16,3) \[5,2,1\] $L_8A_{12}$ (20,9) \[2,4,2\] $L_8A_6$ (22,5) \[4,2,2\] $L_8A_{11}$ (24,7) \[3,2,3\] $L_8A_{13}$ (26,7) \[3,1,2,2\] $L_8A_{10}$ (30,11) \[2,1,2,1,2\] $L_8A_8$ (34,13) \[2,1,1,1,1,2\] $L_8A_9$ : 2-Bridge Links of 8 Crossings The Link $\frac{p}{q}$ or $(p,q)$ Conway Form Thistlethwaite’s Id ----------------------------------- ----------------- --------------------- (20,7) \[2,1,6\] $L_9A_{36}$ (24,5) \[4,1,4\] $L_9A_{40}$ (24,11) \[2,5,2\] $L_9A_{18}$ (28,11) \[2,1,1,5\] $L_9A_{39}$ (30,7) \[4,3,2\] $L_9A_{30}$ (32,7) \[4,1,1,3\] $L_9A_{38}$ (34,9) \[3,1,3,2\] $L_9A_{25}$ (36,11) \[3,3,1,2\] $L_9A_{34}$ (40,11) \[3,1,1,1,3\] $L_9A_{35}$ (44,13) \[3,2,1,1,2\] $L_9A_{37}$ (46,17) \[2,1,2,2,2\] $L_9A_{26}$ (50,19) \[2,1,1,1,2,2\] $L_9A_{27}$ : 2-Bridge Links of 9 Crossings The Link $\frac{p}{q}$ or $(p,q)$ Conway Form Thistlethwaite’s Id ----------------------------------- ------------------- --------------------- (10,1) \[10\] or \[9,1\] $L_{10}A_{118}$ (22,3) \[7,3\] $L_{10}A_{114}$ (26,5) \[5,5\] $L_{10}A_{120}$ (28,13) \[2,6,2\] $L_{10}A_{48}$ (32,5) \[6,2,2\] $L_{10}A_{98}$ (38,9) \[4,4,2\] $L_{10}A_{75}$ (40,7) \[5,1,2,2\] $L_{10}A_{97}$ (40,9) \[4,2,4\] $L_{10}A_{102}$ (42,11) \[3,1,4,2\] $L_{10}A_{73}$ (42,13) \[3,4,3\] $L_{10}A_{115}$ (48,11) \[4,2,1,3\] $L_{10}A_{100}$ (48,17) \[2,1,4,1,2\] $L_{10}A_{89}$ (52,11) \[4,1,2,1,2\] $L_{10}A_{99}$ (56,15) \[3,1,2,1,3\] $L_{10}A_{101}$ (56,17) \[3,3,2,2\] $L_{10}A_{94}$ (58,17) \[3,2,2,3\] $L_{10}A_{116}$ (60,13) \[4,1,1,1,1,2\] $L_{10}A_{93}$ (62,23) \[2,1,2,3,2\] $L_{10}A_{64}$ (64,19) \[3,2,1,2,2\] $L_{10}A_{96}$ (64,23) \[2,1,3,1,1,2\] $L_{10}A_{90}$ (66,25) \[2,1,1,1,3,2\] $L_{10}A_{65}$ (68,19) \[3,1,1,2,1,2\] $L_{10}A_{92}$ (70,29) \[2,2,2,2,2\] $L_{10}A_{87}$ (74,31) \[2,2,1,1,2,2\] $L_{10}A_{83}$ (76,21) \[3,1,1,1,1,1,2\] $L_{10}A_{88}$ (80,31) \[2,1,1,2,1,1,2\] $L_{10}A_{91}$ : 2-Bridge Links of 10 Crossings The Link $\frac{p}{q}$ or $(p,q)$ Conway Form Thistlethwaite’s Id ----------------------------------- ------------------------------------------------------------------------------------ --------------------- $( 26 , 3 )$ $[ 8 , 1 , 2 ]$ $L_{11}A_{ 360 }$ $( 32 , 15 )$ $[ 2 , 7 , 2 ]$ $L_{11}A_{ 132 }$ $( 34 , 5 )$ $[ 6 , 1 , 4 ]$ $L_{11}A_{ 372 }$ $( 38 , 5 )$ $[ 7 , 1 , 1 , 2 , ]$ $L_{11}A_{ 367 }$ $( 44 , 7 )$ $[ 6 , 3 , 2 ]$ $L_{11}A_{ 278 }$ $( 46 , 7 )$ $[ 6 , 1 , 1 , 3 , ]$ $L_{11}A_{ 364 }$ $( 46 , 11 )$ $[ 4 , 5 , 2 , ]$ $L_{11}A_{ 206 }$ $( 50 , 9 )$ $[ 5 , 1 , 1 , 4 , ]$ $L_{11}A_{ 365 }$ $( 50 , 13 )$ $[ 3 , 1 , 5 , 2 , ]$ $L_{11}A_{ 192 }$ $( 52 , 9 )$ $[ 5 , 1 , 3 , 2 , ]$ $L_{11}A_{ 275 }$ $( 54 , 17 )$ $[ 3 , 5 , 1 , 2 , ]$ $L_{11}A_{ 355 }$ $( 56 , 13 )$ $[ 4 , 3 , 4 ]$ $L_{11}A_{ 319 }$ $( 58 , 11 )$ $[ 5 , 3 , 1 , 2 ]$ $L_{11}A_{ 371 }$ $( 62 , 11 )$ $[ 5 , 1 , 1 , 1 , 3 ]$ $L_{11}A_{ 358 }$ $( 62 , 13 )$ $[ 4 , 1 , 3 , 3 ]$ $L_{11}A_{ 369 }$ $( 64 , 15 )$ $[ 4 , 3 , 1 , 3 ]$ $L_{11}A_{ 302 }$ $( 70 , 13 )$ $[ 5 , 2 , 1 , 1 , 2 ]$ $L_{11}A_{ 362 }$ $( 72 , 19 )$ $[ 3 , 1 , 3 , 1 , 3 ]$ $L_{11}A_{ 298 }$ $( 74 , 23 )$ $[ 3 , 4 , 1 , 1 , 2 ]$ $L_{11}A_{ 368 }$ $( 76 , 23 )$ $[ 3 , 3 , 3 , 2 ]$ $L_{11}A_{ 260 }$ $( 76 , 27 )$ $[ 2 , 1 , 4 , 2 , 2 ]$ $L_{11}A_{ 263 }$ $( 78 , 17 )$ $[ 4 , 1 , 1 , 2 , 3 ]$ $L_{11}A_{ 366 }$ $( 78 , 29 )$ $[ 2 , 1 , 2 , 4 , 2 ]$ $L_{11}A_{ 194 }$ $( 80 , 17 )$ $[ 4 , 1 , 2 , 2 , 2 ]$ $L_{11}A_{ 312 }$ $( 82 , 23 )$ $[ 3 , 1 , 1 , 3 , 3 ]$ $L_{11}A_{ 356 }$ $( 82 , 31 )$ $[ 2 , 1 , 1 , 1 , 4 , 2 ]$ $L_{11}A_{ 196 }$ $( 84 , 19 )$ $[ 4 , 2 , 2 , 1 , 2 ]$ $L_{11}A_{ 299 }$ $( 84 , 25 )$ $[ 3 , 2 , 1 , 3 , 2 ]$ $L_{11}A_{ 271 }$ $( 86 , 25 )$ $[ 3 , 2 , 3 , 1 , 2 ]$ $L_{11}A_{ 361 }$ $( 88 , 19 )$ $[ 4 , 1 , 1 , 1 , 2 , 2 ]$ $L_{11}A_{ 280 }$ $( 92 , 21 )$ $[ 4 , 2 , 1 , 1 , 1 , 2 ]$ $L_{11}A_{ 305 }$ $( 92 , 33 )$ $[ 2 , 1 , 3 , 1 , 2 , 2 ]$ $L_{11}A_{ 264 }$ $( 94 , 39 )$ $[ 2 , 2 , 2 , 3 , 2 ]$ $L_{11}A_{ 222 }$ $( 98 , 27 )$ $[ 3 , 1 , 1 , 1 , 2 , 3 ]$ $L_{11}A_{ 359 }$ $( 98 , 41 )$ $[ 2 , 2 , 1 , 1 , 3 , 2 ]$ $L_{11}A_{ 221 }$ $( 100 , 27 )$ $[ 3 , 1 , 2 , 2 , 1 , 2 ]$ $L_{11}A_{ 297 }$ $( 100 , 39 )$ $[ 2 , 1 , 1 , 3 , 2 , 2 ]$ $L_{11}A_{ 284 }$ $( 104 , 29 )$ $[ 3 , 1 , 1 , 2 , 2 , 2 ]$ $L_{11}A_{ 272 }$ $( 106 , 31 )$ $[ 3 , 2 , 2 , 1 , 1 , 2 ]$ $L_{11}A_{ 363 }$ $( 108 , 29 )$ $[ 3 , 1 , 2 , 1 , 1 , 1 , 2 ]$ $L_{11}A_{ 300 }$ $( 112 , 31 )$ $[ 3 , 1 , 1 , 1 , 1 , 2 , 2 ]$ $L_{11}A_{ 262 }$ $( 112 , 41 )$ $[ 2 , 1 , 2 , 1 , 2 , 1 , 2 ]$ $L_{11}A_{ 289 }$ $( 116 , 45 )$ $[ 2 , 1 , 1 , 2 , 1 , 2 , 2 ]$ $L_{11}A_{ 266 }$ $( 128 , 47 )$ $[ 2 , 1 , 2 , 1 , 1 , 1 , 1 , 2 ]$ $L_{11}A_{ 247 }$ $( 144 , 55 )$ $[ 2 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 ]$ $L_{11}A_{ 248 }$ : 2-Bridge Links of 11 Crossings [10]{} D. Cimasoni, A. Conway, and K. Zacharova, *Splitting numbers and signatures*, Proc. Amer. Math. Soc., **144**, (2016), 479-497. Cha, J., Friedl, S., and Powell, M., *Splitting Numbers of Links*,Proceedings of the Edinburgh Mathematical Society,**60**(3), (2017), 587-614. doi:10.1017/S0013091516000420 H. Schubert, *Knoten mit zwei Brücken*, Mathematische Zeitschrift, **65**, (1956), no. 3, 133-170. K. Murasugi, *Knot Theory and its Applications*, Birkh¨auser, 1996. Translated by Bohdan Kurpita. J. C. Cha and C. Livingston, LinkInfo: Table of Knot Invariants,<http://www.indiana.edu/~linkinfo>, July 23, 2019. The Knot Atlas, <http://katlas.org/wiki/The_Thistlethwaite_Link_Table>, July 23, 2019. Knotilus <http://knotilus.math.uwo.ca/>, July 23, 2019. D. De Wit, *The 2-bridge knots of up to 16 crossings*, Journal of Knot Theory and Its Ramifications, **16**, 997-1019, (2007). [^1]: The author was supported in part by TÜBİTAK (The Scientific and Technological Research Council of Turkey) Grant 117F341.

ArXiv

--- abstract: 'We study the dynamics of spin flipping at first order transitions in zero temperature two-dimensional random-field Ising model driven by an external field. We find a critical value of the disorder strength at which a discontinuous sharp jump in magnetization first occurs. We discuss growth morphology of the flipped-spin domains at and away from criticality.' address: | The Institute of Mathematical Sciences,\ Chennai-600 113 ,\ India.\ $^\[email protected]\ $^\[email protected] author: - 'RATNADEEP ROY$^\dagger$ and PURUSATTAM RAY$^\ddagger$' title: RESPONSE OF RANDOM FIELD ISING MODEL DRIVEN BY AN EXTERNAL FIELD --- Introduction ============ The [*Random Field Ising Model*]{} $(RFIM)$ has been widely studied in recent years in the context of hysteresis \[1-3\], return-point memory \[4,5\] and Barkhausen noise \[6-11\]. Hysteresis is a common phenomenon, exhibited, for example, by most magnetic materials. In hysteresis the magnetization lags behind the applied field and the phenomenon has received lots of attention in the past. Moreover, if one observes carefully, the change in magnetization in systems like that in ferromagnetic alloys and amorphous ferromagnets, takes place in a series of irregular pulses as the external field is slowly varied. In experiments these pulses manifest themselves as acoustic emissions and is known as Barkhausen noise. Interestingly, Barkhausen signals show scale-free behavior similar to that of systems at criticality. Barkhausen noise has been studied much in recent years in the context of driven disordered systems far from equilibrium example of which include martensite transformation in shape-memory alloys \[12\], eartquakes \[13\], deformation of granular materials \[14\] and the breakdown of solids \[15\]. Various experiments \[8\] show that the distribution of size, duration and energy associated with Barkhausen noise show power-law behavior. Sethna et al, in order to account for the hysteresis and Barkhausen noise incorporated disorder in the form of random fields into the otherwise pure spin systems such as the Ising model \[3\]. In the [*RFIM*]{}, the Ising spins $\{s_i\}$ with nearest neighbor ferromagnetic interaction [*J*]{} are coupled to the on-site quenched random magnetic fields $\{h_i\}$ and the external field $h_a$. The Hamiltonian of the system is $$H = -J \sum_{<ij>}{s_i}{s_j} - \sum_{i}{h_i}{s_i} - {h_a}\sum_{i}{s_i}$$ where $\{h_i\}$ are quenched independent random variable drawn from a distribution of zero mean and variance $\Delta$. In the study of hysteresis and Barkhausen jumps the applied field is ramped up or down adiabatically (so that the rate of spin flips is much larger than the rate of change of $h_a$). In the absence of disorder ($\{h_i\} = 0$), the model exhibits a sharp first order transition taking the magnetization $m(h_a)$ from -1 to +1 as $\it h_a$ passes through [*zJ*]{}, [*z*]{} being the coordination number of the underlying lattice. In presence of random field, the sharpness of the transition gets smeared out and the change in magnetization with field takes place in a series of “spin-cluster” flips and the magnetization changes in sporadic jumps as is observed in [*BN*]{}. Simulations of [*RFIM*]{} in three-dimensions show that for large values of $\Delta$ (greater than a critical value $\Delta_c$), the spins flip in small clusters resulting in avalanches of small sizes. The sizes of these avalanches are distributed in a power-law. At $\Delta = \Delta_c$ avalanches of all sizes occur. For $\Delta < \Delta_c$, the magnetization shows a first order jump (corresponding to an infinite avalanche) as is the case if there is no disorder. The behavior of avalanches and avalanche size distribution at and near $\Delta_c$ has been studied in the mean field limit \[3\] and the power law relating to the avalanche size distribution is known exactly in [*RFIM*]{} in one-dimension and on Bethe lattices \[11\]. For a linear chain and a Bethe lattice of coordination number z =3 the avalanche size (s) disribution decreases exponentially for large [*s*]{}. No first order jump has been observed for any finite amount of disorder (unbounded) for $z \leq 3$. Thus the low $\Delta$ behavior depends on the coordination number of the underlying lattice. For $z \geq 4$ and for small disorder the magnetization shows a first order discontinuity for several continuous and unimodal distributions of the random fields. The avalanche distribution [*P(s)*]{} varies as $s^{-\frac{5}{2}}$ for large s near the discontinuity. This growth of magnetization is essentially due to the motion of an interface of a cluster of positive spins through a disordered media. Renormalization group studies of the interface (starting from a continuum model and an interface without any backbends) motion in disordered medium shows that the interface at the critical value of the strength of the disorder becomes rough with a roughness exponent $\zeta = (5-d)/3$ \[16\] for a d-dimensional system. According to this relation the upper critical dimension is five and the lower critical dimension is two. At $d = 2$, it is considered that the interface may have backbends and can encompass bubbles as it moves through the medium. Simulation results show the following generic features of the interface growth in random media \[17-24\]: (i) For large strengths of the disorder the interface growth is percolation-like.The fractal dimension of the magnetic-interface corresponds to that ordinary site percolation. (ii) At the critical value of the disorder the interface becomes self-affine with a roughness exponent which closely matches the RG result. (iii) For low strength of the disorder the interface growth is faceted. We study the spin flipping dynamics in [*RFIM*]{} in two-dimensions at zero temperature as the external field is ramped up slowly from a high negative value. Our intention is to see if there is any finite value of the disorder strength $\Delta_c$ in two-dimension. At $\Delta_c$ and at a particular field $h_a = h_{a}^{*}$, the spin starts flipping from a site and does not end till the flipped spins span the entire system. The domain wall that separates the flipped spin regions from the rest is highly tortuous (fractal like). The spin flipped region has bubbles of unflipped spins, so that $m(h_a)$ does not jump from -1 to +1, instead attains a value $\sim 0.76$ irrespective of system size. In the regime when the flipped spin domain grows, the number [*M*]{} of flipped spins and the interface length [*I*]{} both grow with time [*t*]{} as $M \sim I \sim t^2$. For $\Delta < \Delta_c$, [*M*]{} jumps from -1 to +1 at a particular field and the flipped spin domain remains compact as it grows following $M \sim I^2 \sim t^2$. For $\Delta > \Delta_c$, $m(h_a)$ grows continuously with the field in jumps of different sizes. These jumps correspond to flipping of small spin clusters at various places in the lattice. The Model and Simulation ======================== We have studied RFIM in two-dimensions on a square lattice of size [*L*]{} with a classical Ising spin $s_i = \pm 1$ and a quenched random field $h_i$ on every site $i$ of the lattice. Each spin interacts with each of its nearest neighbor spin through a ferromagnetic interaction $J = 1$. The random field $h_i$ is drawn from a uniform distribution $$\begin{aligned} p(h_i) &=& {} \frac{1}{2\Delta}\qquad {\textrm{if} -\Delta \leq h_i \leq\Delta}\nonumber \\ &=& ~ 0 \qquad~~~ {\textrm{otherwise .}}\end{aligned}$$ The spins are subjected to an applied field $h_a$ which is varied. We study the model at zero temperature. We start with a configuration of all down spins which corresponds to a high negative value of $h_a$. The field $h_a$ is then slowly raised and the resulting spin flips are recorded. The spin flipping process follows the rule of zero temperature Glauber dynamics \[25\], where a spin is flipped only if the flipping lowers the energy of the system as calculated according to the Hamiltonian given by “Eq (1)". We use periodic boundary conditions in our simulation. For a certain [*L*]{}, $\Delta$ and a configuration of the random field $h_i$ we first find out the spins $\{\alpha\}$ which can be most easily flipped and the corresponding value of the applied field $h_{a}^{*}$ required to flip the spins. We then set $h_a = h_{a}^{*}$ and flip all the spins in $\{\alpha\}$ simultaneously. This constitutes one time step in our simulation. Next we check if this primary set of spin flips introduce secondary spin flips and so on, always flipping all the flippable spins simultaneously. If there is no spin that can be flipped, we increase $h_a$ to trigger next generation of spin flipping. We continue this process till all the spins of the lattice are flipped. We have taken results for [*L*]{} = 500 to 7000 and for various values $\Delta$ from 2.0 to 2.8. For a certain $\Delta$, our results are averaged over 50 initial random configurations of the on-site fields. Results and Discussion ====================== Our simulation results show that there is a critical value $\Delta_c$ of the strength of the disorder. The growth morphology of domains of flipped spins are distinctly different for $\Delta < \Delta_c$, $\Delta > \Delta_c$ and for $\Delta = \Delta_c$. We discuss the three cases below: For $\Delta < \Delta_c$ as the external field is ramped up there is a sharp first order transition (like that happens in the absence of disorder) at a particular field ($h_a = 4J - \Delta$)that takes the magnetization from -1 to +1 as is shown in Fig.1 This infinite avalanche nucleates from a single spin and invades the system in course of time. The flipped spin cluster remains compact while it grows: the mass [*M*]{} (the number of upturned spins) of the cluster grows with time [*t*]{} (in Monte Carlo steps) as $ M \sim t^2$ (Fig.2). The number of spins [*I*]{} along the interface grows as $I \sim t$ (Fig.2). For $\Delta > \Delta_c$ an infinite avalanche never happens. The magnetization increases in irregular steps with the increase of external field as shown in Fig.3. In Fig.4 we show a snapshot of the upturned spins at a particular field when the flipping of the spins has stopped. In this regime we see nucleation of many domains as opposed to the nucleation of a single domain for $\Delta < \Delta_c$. Such a feature is also observed in dilute [*RFIM*]{} \[10\]. For any field, when the system has stopped evolving, a typical distribution of the random fields over the unturned spins along the interface is as shown in Fig.5. It consists of two steps. The higher step is for those sites that have large negative random fields that would require three upturned neighbors for its flipping. The lower step corresponds to the random fields of those sites which can flip if two of its neighbors have already flipped. This shows that the the interface spins are strongly pinned. The non-existence of sites (at the interface) that can flip when one of its neighbor is up is responsible for the freezing of further spin flipping at the particular field. At the critical value of $\Delta = \Delta_c$ we see the growth of a single domain and a discontinuous first order jump in the magnetization at a particular value of the applied field. However, unlike the situation for $\Delta < \Delta_c$ here the domain of flipped spins encompasses bubbles of unturned spins of various sizes at different stage of its growth and the interface has overhangs and is fractal like. As a result magnetization does not jump from -1 to +1 but to (around) 0.76 irrespective of the system size (we have checked it for [*L*]{} = 500, 800, 2000, 4000, 5000 and 7000). The change in magnetization with field is shown in Fig.6. A typical up-spin cluster at a certain time of its developement is shown in Fig.7. It shows the tortuous interface with bubbles of unturned spins (white region). In this case both the mass and interface develops as $M \sim I \sim t^2$ as is shown in Fig.8. At any time during the growth of the up-spin domain the probability distribution of the random fields of the unturned spins along the interface has three steps as is shown in Fig.9. The lowest step corresponds to the random fields on those sites that can flip if one of its neighbor is up. This guarantees the growth of domains as along the interface one neighbor will always be up. However when the up-spins span the system the probability distribution of the random fields along the interface corresponds to those for the bubbles only and is similar to that for $\Delta > \Delta_c$. This shows that the spins along the interface of the bubbles are strongly pinned. The external field has to be continuously increased for the magnetization to change from 0.76 to +1 which will eventually fill up the bubbles (Fig.6). Conclusion ========== We find a finite,non-zero $\Delta_c$ for bounded distribution. This $\Delta_c$ demarcates a region ($\Delta > \Delta_c$) where infinite avalanches never happens to that from a region where magnetization shows a jump discontinuity from -1 to +1. Earlier studies indicate $\Delta_c = 0$ for unbounded distribution. We find at $\Delta_c$ the interface around the flipped domains have backbends and is fractal like. It encloses bubbles of unflipped spins in all length scales. [27]{} E. C. Stoner, [*Rev. Mod. Phys.*]{} [**25**]{}, 2 (1953). B. K. Chakrabarti and M. Acharyya, [*Rev. Mod. Phys.*]{} [**71**]{}, 847 (1999). J. P Sethna, K. Dahmen, S. Kartha, J. A. Krumhansl, B. W. Roberts and J. D. Shore, [*Phys. Rev. Lett.*]{} [**70**]{}, 3347 (1993). P. Shukla, [*Phys. Rev.*]{} [**E62**]{}, 4725 (2000). P. Shukla, [*Phys. Rev.*]{} [**E63**]{}, 027102 (2001). H. Barkhausen, [*Z. Phys.*]{} [**20**]{}, 401 (1919). O. Perković, K. Dahmen and J. P Sethna, [*Phys. Rev. Lett.*]{} [**75**]{}, 4528 (1995). D. Spasojević, S. Bukvić, S. Milošević and H. E. Stanley, [*Phys. Rev.*]{} [**E54**]{}, 2531 (1996). B. Drossel and K. Dahmen, [*Euro. Phys. J.*]{} [**B3**]{}, 485 (1998). B. Tadić, [*Physica*]{} [**A270**]{}, 125 (1999). S. Sabhapandit, P. Shukla and D. Dhar, [*J. Stat. Phys*]{} [**98**]{}, 103 (2000). L. Carrillo, L. Manosa, J. Ortin, A. Planes and E. Vives, [*Phys. Rev. Lett.*]{} [**81**]{}, 1889 (1998). M. Sahimi and S. Arbabi, [*Phys. Rev. Lett.*]{} [**77**]{}, 3689 (1996). I. Main, [Rev. Geophys.]{} [**34**]{}, 433 (1996). S. Zapperi, P. Ray, H. E. Stanley and A. Vespignani, [*Phys. Rev. Lett.*]{} [**78**]{}, 1408 (1997). O. Narayan and D. S. Fisher, [*Phys. Rev.*]{} [**B48**]{}, 7030 (1992). R. Bruinsma and G. Aeppli, [*Phys. Rev. Lett.*]{} [**52**]{}, 1547 (1984). J. Koplik and H. Levine, [*Phys. Rev.*]{} [**B32**]{}, 280 (1985). H. Leschhorn and L. Tang, [*Phys. Rev.*]{} [**E49**]{}, 1238 (1994). H. Ji and M. O. Robbins, [*Phys. Rev.*]{} [**A44**]{}, 2538 (1991). H. Ji and M. O. Robbins, [*Phys. Rev.*]{} [**A46**]{}, 14519 (1992). M. R. Swift, A. Maritan and J. R. Banavar, [*Phys. Rev. Lett*]{} [**77**]{}, 5288 (1996). Z. Jiang and H. G. E. Hentschel, [*Phys. Rev.*]{} [**A45**]{}, 4169 (1992). B. Koiller, H. Ji and M. O. Robbins, [*Phys. Rev.*]{} [**B46**]{}, 5258 (1992). R. J. Glauber, [*J. Math. Phys.*]{} [**4**]{}, 294 (1963).

ArXiv

--- abstract: '\[abst\] Statistical mechanics is applied to lossy compression using multilayer perceptrons for unbiased Boolean messages. We utilize a tree-like committee machine (committee tree) and tree-like parity machine (parity tree) whose transfer functions are monotonic. For compression using committee tree, a lower bound of achievable distortion becomes small as the number of hidden units $K$ increases. However, it cannot reach the Shannon bound even where $K\to\infty$. For a compression using a parity tree with $K \ge 2$ hidden units, the rate distortion function, which is known as the theoretical limit for compression, is derived where the code length becomes infinity.' author: - Kazushi Mimura - Masato Okada title: Statistical mechanics of lossy compression using multilayer perceptrons --- introduction ============ Cross-disciplinary fields that combine information theory with statistical mechanics have developed rapidly in recent years and achievements in these have become the center of attention. The employment of methods derived from statistical mechanics has resulted in significant progress in providing solutions to several problems in information theory, including problems in error correction [@Sourlas1989; @Kabashima2000; @Nishimori1999; @Montanari2000], spreading codes [@Tanaka2001; @Tanaka2005] and compression codes [@Murayama2003; @Murayama2004; @Hosaka2002; @Hosaka2005]. Above all, data compression plays an important role as one of the base technologies in many aspects of information transmission. Data compression is generally classified into lossless compression and lossy compression [@Shannon1948; @Shannon1959; @Cover1991]. Lossless compression is aimed at reducing the size of message under the constraint of perfect retrieval. In lossy compression, on the other hand, the length of message can be reduced by allowing a certain amount of distortion. The theoretical framework for lossy compression scheme is called rate distortion theory, which consists partly of Shannon’s information theory [@Shannon1948; @Shannon1959]. Several lossy compression codes, whose schemes saturate the rate distortion function that represents an optimal performance, were discovered in the case where the code length becomes infinity. For instance, Low Density Generator Matrix (LDGM) code [@Murayama2003; @Murayama2004] and using a nonmonotonic perceptron [@Hosaka2002; @Hosaka2005; @Hosaka2005b] were proposed. In these compression codes, a decoder is first defined to retrieve a reproduced message from a codeword. In the encoding problem, for a given source message, we must find a codeword that minimizes the distortion between the reproduced message and the source message. Therefore, fundamentally, the computational cost of compressing a message is of exponential order of a codeword length. It is important to understand properties of various lossy compression codes saturating the optimal performance for the development of more useful codes. Since a multilayer network includes a nonmonotonic perceptron as a special case, we employ tree-like committee machine and parity machine as typical multilayer networks [@Barkai1990; @Barkai1991; @Barkai1992] to lossy compression and analytically evaluate their performance. lossy compression ================= Let us start by defining the concepts of the rate distortion theory [@Cover1991]. Let $y$ be a discrete random variable with source alphabet ${\cal Y}$. We will assume that the alphabet is finite. An source message of $M$ random variables, ${\mbox{\boldmath $y$}}={}^t(y^1,\cdots,y^M)\in{\cal Y}^M$, is compressed into a shorter expression, where the operator ${}^t$ denotes the transpose. Here, the encoder describes the source sequence ${\mbox{\boldmath $y$}}\in{\cal Y}^M$ by a codeword ${\mbox{\boldmath $s$}}={\cal F}({\mbox{\boldmath $y$}})\in{\cal S}^N$. The decoder represents ${\mbox{\boldmath $y$}}$ by a reproduced message $\hat{{\mbox{\boldmath $y$}}}={\cal G}({\mbox{\boldmath $s$}})\in\hat{\cal Y}^M$, as illustrated in Fig. \[fig:framework\]. Note that $M$ represents the length of a source sequence, while $N$ represents the length of a codeword. The code rate is defined by $R=N/M$ in this case. A distortion function is a mapping $d:{\cal Y}\times\hat{\cal Y}\to\mathbb{R}^+$ from the set of source alphabet-reproduction alphabet pair into the set of non-negative real numbers. In most cases, the reproduction alphabet $\hat{\cal Y}$ is the same as the source alphabet ${\cal Y}$. After this, we set ${\cal Y}=\hat{\cal Y}$. An example of common distortion functions is Hamming distortion given by $$d(y,\hat{y}) = \biggl\{ \begin{array}{ll} 0, & y=\hat{y}, \\ 1, &y\ne\hat{y}, \end{array}$$ which results in the probability of error distortion, since $E[d(y,\hat{y})]=P[y\ne\hat{y}]$, where $E$ and $ P$ represent the expectation and the probability of its argument respectively. The distortion between sequences ${\mbox{\boldmath $y$}}, \hat{{\mbox{\boldmath $y$}}} \in {\cal Y}^M$ is defined by $d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}})=\sum_{\mu=1}^M d(y^\mu,\hat{y}^\mu)$. Therefore, the distortion associated with the code is defined as $D=E[\frac 1M d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}})]$, where the expectation is with respect to the probability distribution on ${\cal Y}$. A rate distortion pair $(R,D)$ is said to be [*achievable*]{} if there exists a sequence of rate distortion codes $({\cal F},{\cal G})$ with $E[\frac 1M d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}})] \le D$ in the limit $M\to\infty$. We can now define a function to describe the boundary called the [*rate distortion function*]{}. The rate distortion function $R(D)$ is the infimum of rates $R$ such that $(R,D)$ is in the rate distortion region of the source for a given distortion $D$ and all rate distortion codes. The infimum of rates $R$ for a given distortion $D$ and given rate distortion codes $({\cal F},{\cal G})$ is called the [*rate distortion property*]{} of $({\cal F},{\cal G})$. We restrict ourselves to a Boolean source ${\cal Y}=\{0,1\}$. We assume that the source sequence is not biased to rule out the possibility of compression due to redundancy. The non-biased Boolean message in which each component is generated independently from an identical distribution $P(y^\mu=1)=P(y^\mu=0)=1/2$. For this simple source, the rate distortion function for an unbiased Boolean source with Hamming distortion is given by $$R(D)=1-h_2(D), \label{eq:RDF}$$ where $h_2(x)=-x\log_2(x)-(1-x)\log_2(1-x)$ called the binary entropy function. ![Rate distortion encoder and decoder.[]{data-label="fig:framework"}](fig1.eps){width="0.45\linewidth"} compression using multilayer perceptrons ======================================== To simplify notations, let us replace all the Boolean representations $\{0,1\}$ with the Ising representation $\{1,-1\}$ throughout the rest of this paper. We set ${\cal Y}={\cal S}=\hat{\cal Y}=\{1,-1\}$ as the binary alphabets. We consider an unbiased source message in which a component is generated independently from an identical distribution: $$P(y^\mu)=\frac 12 \delta (y^\mu -1) + \frac 12 \delta (y^\mu +1),$$ for simplicity. First let us define a decoder. We can construct a nonlinear map ${\cal G} : {\cal S}^N \to \hat{\cal Y}^M$ from codeword ${\mbox{\boldmath $s$}}\in{\cal S}^N$ to reproduced message $\hat{{\mbox{\boldmath $y$}}}=(\hat{y}^\mu)\in\hat{\cal Y}^M$. For a given source message ${\mbox{\boldmath $y$}}=(y^\mu)\in{\cal Y}^M$, the role of the encoder is to find a codeword ${\mbox{\boldmath $s$}}\in{\cal S}^N$ that minimizes the distortion between its reproduced message ${\cal G}({\mbox{\boldmath $s$}})$ and the source message ${\mbox{\boldmath $y$}}$. We choose a nonlinear map ${\cal G}$ utilizing tree-like multilayer perceptrons, i.e., a tree-like committee machine (committee tree) and a tree-like parity machine (parity tree). Figure \[fig:tree\] shows its architecture. The codeword ${\mbox{\boldmath $s$}}$ is divided into $N/K$-dimensional $K$ disjoint vectors ${\mbox{\boldmath $s$}}_1, \cdots ,{\mbox{\boldmath $s$}}_K\in{\cal S}^{N/K}$ as ${\mbox{\boldmath $s$}}={}^t({\mbox{\boldmath $s$}}_1, \cdots, {\mbox{\boldmath $s$}}_K)$. The $l$th hidden unit receives the vector ${\mbox{\boldmath $s$}}_l$. The outputs of the committee tree and the parity tree are a majority decision and a parity of hidden unit outputs, respectively. ![The architecture of tree-like multilayer perceptrons with $N$ input units and $K$ hidden units. []{data-label="fig:tree"}](fig2.eps){width=".6\linewidth"} The $\mu$th bit of the reproduced message $\hat{y}^\mu$ is defined by utilizing the committee tree as $$\hat{y}^\mu({\mbox{\boldmath $s$}}) \equiv {\mbox{sgn}}\biggl( \sum_{l=1}^K f \biggl( \sqrt{\frac KN}{\mbox{\boldmath $s$}}_l \cdot {\mbox{\boldmath $x$}}_l^\mu \biggr) \biggr), \label{eq:decoder_CM}$$ where ${\mbox{\boldmath $x$}}_l^{\mu} \sim {\cal N}({\mbox{\boldmath $0$}},{\mbox{\boldmath $1$}})$ are fixed $N/K$-dimensional vectors and the map $f:\mathbb{R}\to{\cal Y}$ is a transfer function. Function ${\mbox{sgn}}(x)$ denotes the sign function taking 1 for $x \ge 0$ and -1 for $x<0$. Similarly, the $\mu$th bit $\hat{y}^\mu$ of the reproduced message is also defined by utilizing the parity tree as $$\hat{y}^\mu({\mbox{\boldmath $s$}}) \equiv \prod_{l=1}^K f \biggl( \sqrt{\frac KN}{\mbox{\boldmath $s$}}_l \cdot {\mbox{\boldmath $x$}}_l^\mu \biggr). \label{eq:decoder_PM}$$ The decoder ${\cal G}$ from the codeword ${\mbox{\boldmath $s$}}$ to the reproduced message $\hat{{\mbox{\boldmath $y$}}}=(\hat{y}^\mu)$ is described as $${\cal G}({\mbox{\boldmath $s$}}) \equiv \hat{{\mbox{\boldmath $y$}}}({\mbox{\boldmath $s$}})={}^t(\hat{y}^1({\mbox{\boldmath $s$}}),\cdots ,\hat{y}^M({\mbox{\boldmath $s$}})),$$ In this framework, the encoder ${\cal F}$ from the original message ${\mbox{\boldmath $y$}}$ to the codeword ${\mbox{\boldmath $s$}}$ can be written as $${\cal F}({\mbox{\boldmath $y$}}) \equiv \mathop{\rm argmin}_{\hat{{\mbox{\boldmath $s$}}}} d({\mbox{\boldmath $y$}},{\cal G}(\hat{{\mbox{\boldmath $s$}}})), \label{eq:encode}$$ with respect to the case of both the committee tree and the parity tree. Employing the Ising representation, where the length of the codeword is infinite, the average Hamming distortion can be represented as $$E[d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}})] = \sum_{\mu=1}^M [ 1- \Theta ( y^\mu \hat{y}^\mu) ],$$ where the function $\Theta(x)$ denotes the step function taking 1 for $x \ge 0$ and 0 otherwise. Since we assume the unbiased source message in this paper, we set $f(x)={\mbox{sgn}}(x)$. This encoding scheme is essentially the same as a learning of the multilayer perceptrons because of a following reason. We first assign the random input vector ${\mbox{\boldmath $x$}}^\mu={}^t({\mbox{\boldmath $x$}}_1^\mu,\cdots,{\mbox{\boldmath $x$}}_K^\mu) \in \mathbb{R}^N$ to each bit of the original message $y^\mu$. The encoder must find a weight vector ${\mbox{\boldmath $s$}}$ which satisfies input-output relations ${\mbox{\boldmath $x$}}^\mu \mapsto y^\mu$ as much as possible. Then we use this optimal weight vector ${\mbox{\boldmath $s$}}$ as a codeword. Therefore, in a lossless case of $D=0$, an evaluation of the rate distortion property of these codes is entirely identical to the calculation of the storage capacity [@Gardner1988; @Krauth1989]. analytical evaluation ===================== We analytically evaluate the typical performance, according to Hosaka et al [@Hosaka2002], for the proposed compression scheme using the replica method. The minimum permissible average distortion $D$ is calculated, when the code rate $R$ is fixed. For a given original message ${\mbox{\boldmath $y$}}$ and the input vectors $\{{\mbox{\boldmath $x$}}_l^{\mu}\}$, the number of codewords ${\mbox{\boldmath $s$}}$, which provide a fixed Hamming distortion $MD=d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}})$, can be expressed as $${\cal N}(D,R) = \mathop{\rm Tr}_{{\mbox{\boldmath $s$}}} \delta \biggl( MD ; d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}}({\mbox{\boldmath $s$}})) \biggr),$$ where $\delta(m;n)$ denotes Kronecker’s delta taking 1 if $m=n$ and 0 otherwise. Since the original message ${\mbox{\boldmath $y$}}$ and the input vectors $\{{\mbox{\boldmath $x$}}_l^\mu\}$ are randomly generated predetermined variables, the quenched average of the entropy per bit over these parameters, $$S(D,R)= \frac 1N <\ln {\cal N}(D,R)>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}}, \label{eq:def_S}$$ is naturally introduced for investigating the typical properties, where $< \; >_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}}$ denotes the average over ${\mbox{\boldmath $y$}}$ and $\{{\mbox{\boldmath $x$}}_l^{\mu}\}$. We calculate the entropy $S(D)$ by the replica method (see Appendix \[appendix.ReplicaMethod\]). The rate-distortion region can be represented by $\{(D,R) | S(D,R) \ge 0 \}$. Therefore, a minimum code rate $R$ for a fixed distortion $D$ is given by a solution of $S(D,R)=0$. Note that a minimum code rate $R$ for $D=0$ coincides with a reciprocal of the critical storage capacity of a multilayer perceptron, i.e., the critical storage capacity $\alpha_c(\equiv M/N)$ can be obtained by $S(0,1/\alpha_c)=0$. Replica symmetric theory of lossy compression using committee tree ------------------------------------------------------------------ ### For general $K$ In the lossy compression using the committee tree, we obtain average entropy $S_{CT}(D,R)$ as $$\begin{aligned} S_{CT}(D,R) &=& \mathop{\rm extr}_{\beta,q,\hat{q}} \biggl( R^{-1} \biggl< \int \biggl( \prod_{l=1}^K Dt_l \biggr) \nonumber \\ & & \times \ln \{ e^{-\beta} + (1-e^{-\beta}) \Sigma (\{t_l\};y) \} \biggr>_y \nonumber \\ & & + \int Du \ln 2 \cosh \sqrt{\hat{q}} u - \frac{\hat{q}(1-q)}2 \nonumber \\ & & + R^{-1}\beta D \biggr), \label{eq:S}\end{aligned}$$ where $$\Sigma (\{t_l\};y) \equiv \mathop{\rm Tr}_{\{ \tau_l =\pm 1\}} \Theta \biggl( -y\sum_{l=1}^K \tau_l \biggr) \prod_{l=1}^K H(Q \tau_l t_l), \label{eq:Sigma}$$ with $Q \equiv \sqrt{q/(1-q)}$ (see Appendix \[appendix.RM.CT.generalK\]). For any $K$, we can obtain a minimum code rate $R$, which gives $S_{CT}(D,R)=0$ for a fixed distortion $D$. ### For large $K$ We concentrate in the following on the simple case of large $K$, where the $K$-multiple integrals can be reduced to a single Gaussian integral. We assume that the number of hidden units $K$ is large but still $K \ll N$. Using the central limit theorem, the averaged entropy is given by $$\begin{aligned} S_{CT}(D,R) &=& \mathop{\rm extr}_{\beta,q,\hat{q}} \biggl( R^{-1} \biggl< \int Dt \ln \{ e^{-\beta} \nonumber \\ & & + (1-e^{-\beta}) H \biggl( \sqrt{\frac{q_{eff}}{1-q_{eff}}}t \biggr) \} \biggr>_y \nonumber \\ & & + \int Du \ln 2 \cosh \sqrt{\hat{q}} u - \frac{\hat{q}(1-q)}2 \nonumber \\ & & + R^{-1}\beta D \biggr) , \label{eq:S_CT}\end{aligned}$$ where $q_{eff} \equiv \int Dt [1-2H(Qt)]^2 = \frac 2{\pi} \sin^{-1} q$ and $Q_{eff} \equiv \sqrt{q_{eff}/(1-q_{eff})}$ (see Appendix \[appendix.RM.CT.largeK\]). Figure \[fig:CT\] shows that the limit of achievable code rate $R$ expected for $N \to \infty$ plotted versus the distortion $D$ for $K=1, 3$ and $K\to\infty$. For a fixed code rate $R$, the achievable distortion decreases as the number of hidden units $K$ increases. However, it does not saturate Shannon’s limit even if in the limit $K\to\infty$. For large $K$, the EA order parameter $q$, which means the average overlap between different codewords, does not converge to zero. Since this means that codewords are correlated, the distribution of codewords is biased in ${\cal S}^N$. Note that a nonzero EA order parameter does not mean that the reproduced message has a nonzero average due to the random input vector which have a zero average. ![ The rate distortion property of lossy compression using a committee tree. The limit of achievable code rate $R$ expected for $N \to \infty$ plotted versus the distortion $D$ for $K=1$ (dotted line), $K=3$ (short dashed line) and $K\to\infty$ (long dashed line). Solid line denotes rate-distortion function $R(D)$ for binary sequences by Shannon. []{data-label="fig:CT"}](fig3.eps){width="0.9\linewidth"} Replica symmetric theory of lossy compression using parity tree --------------------------------------------------------------- In the lossy compression using the parity tree, on the other hand, we hence obtain averaged entropy $S_{PT}(D,R)$ as $$\begin{aligned} S_{PT}(D,R) &=& \mathop{\rm extr}_{\beta,q,\hat{q}} \biggl( R^{-1} \biggl< \int \biggl( \prod_{l=1}^K Dt_l \biggr) \nonumber \\ & & \times \ln \{ e^{-\beta} + (1-e^{-\beta}) \Pi (\{t_l\};y) \} \biggr>_y \nonumber \\ & & + \int Du \ln 2 \cosh \sqrt{\hat{q}} u - \frac{\hat{q}(1-q)}2 \nonumber \\ & & + R^{-1}\beta D \biggr), \label{eq:S_PT}\end{aligned}$$ where $$\Pi (\{t_l\};y) \equiv \frac 12 \biggl( 1+y\prod_{l=1}^K [1-2H(Q t_l)] \biggr) .$$ For cases utilizing a committee tree and a parity tree, only terms $\Sigma (\{t_l\};y)$ and $\Pi (\{t_l\};y)$ are different. Since both the order parameters $q$ and $\hat{q}$ at the saddle-point of (\[eq:S\_PT\]) are less than one, the average entropy can be expanded with respect to $\prod_{l=1}^K [1-2H(Q t_l)] (<1)$. Solutions of the saddle-point equation derived from the expanded form of average entropy are obtained as $$\left\{ \begin{array}{l} q=0, \\ \hat{q}=0, \\ D=\displaystyle{\frac{e^{-\beta}}{1+e^{-\beta}}}, \end{array} \right. \label{eq:K>=2PTspe}$$ in the case $K\ge 2$ (see Appendix \[appendix.RM.PT\]). For $K=1$, $q>0$ holds. Note that for $K=1$, a parity tree is equivalent to a committee tree. For $K\ge 2$, the order parameter $q$ becomes zero, namely all codewords are uncorrelated and distributed all round in ${\cal S}^N$. Where $K \ge 2$, substituting (\[eq:K&gt;=2PTspe\]) into (\[eq:S\_PT\]), average entropy is obtained as $$\begin{aligned} S_{PT}(D,R) &=& -R^{-1}\ln 2+ \ln 2 -R^{-1}D\ln D \nonumber \\ & & -R^{-1}(1-D)\ln (1-D). \end{aligned}$$ A minimum code rate $R$ for a fixed distortion $D$ and $K \ge 2$ is given by $S_{PT}(D,R)=0$. Solving this equation with respect to $R$, we obtain $$R = 1-h_2(D) \equiv R_{RS}(D),$$ which is identical to the rate-distortion function for uniformly unbiased binary sources (\[eq:RDF\]). However, since calculation is based on the RS ansatz, we verify the AT stability to confirm the validity of this solution. As the RS solution to lossy compression using a parity tree with $K=2$ hidden units can be simply expressed as (\[eq:K&gt;=2PTspe\]), the stability condition is analytically obtained as $$R>\frac 8{\pi^2}(1-2D)^2 \equiv R_{AT}(D), \label{eq:ATline}$$ where boundary $R=R_{AT}(D)$ is called the AT line (see Appendix \[appendix.ATline\]). For $K\ge 3$, the RS solution does not exhibit the AT instability throughout the achievable region of the rate-distortion pair $(R,D)$. Figure \[fig:PT\] shows the limit of achievable distortion $D$ expected for $N \to \infty$ plotted versus code rate $R$ for $K=1$ and $K\ge 2$. In the case $K\ge 2$, the limit of achievable distortion is identical to the rate-distortion function. The dash-dotted line in Fig. \[fig:PT\] denotes the AT line for $K=2$. The region above the AT line denotes that the RS solution is stable. For $K=2$, we found that for distortion $0.126 \alt D \le 0.5$, $R_{RS}(D)$ can become smaller than $R_{AT}(D)$. Nevertheless this instability may not be serious in practice, because the region where the RS solution becomes unstable is narrow. The annealed approximation of the entropy (\[eq:def\_S\]) gives a lower bound to the rate distortion property. It coincides with the rate distortion function. According to Opper’s discussion [@Opper1995], the entropy (\[eq:def\_S\]) can be represented by the information entropy formally. The annealed information entropy can give a upper bound to the rate distortion property. However, its evaluation is difficult (see Appendix \[appendix.lower\_bound\]). ![ The rate distortion property of lossy compression using a parity tree. The limits of achievable code rate $R$ expected for $N \to \infty$ plotted versus the distortion $D$ for $K=1$ (dashed line) and $K \ge 2$ (solid line). Solid line also denotes rate-distortion function, which is identical to limit of achievable distortion for $K \ge 2$. Dash-dotted line denotes AT line for $K=2$. For $K\ge 3$, RS solution does not exhibit AT instability throughout achievable region. []{data-label="fig:PT"}](fig4.eps){width="0.9\linewidth"} distribution of codewords ========================= It has already been shown that both compression using a sparse matrix and compression using a nonmonotonic perceptron also achieve optimal performance known as Shannon’s limit [@Murayama2003; @Hosaka2002]. All these schemes and compression using a parity tree with $K\ge 2$ hidden units becomes the common EA order parameter $q=0$. In compression using a nonmonotonic perceptron, the $\mu$th bit of the reproduced message is defined as $\hat{y}^\mu({\mbox{\boldmath $s$}}) \equiv \hat{f} ( N^{-1/2} {\mbox{\boldmath $s$}} \cdot {\mbox{\boldmath $x$}}^\mu )$, where $\hat{f}$ is the transfer function with mirror symmetry, i.e., $\hat{f}(-x)=\hat{f}(x)$ [@Hosaka2002]. Due to the mirror symmetry of $\hat{f}$, both ${\mbox{\boldmath $s$}}$ and $-{\mbox{\boldmath $s$}}$ provide identical output for any ${\mbox{\boldmath $x$}}^\mu$. Hence, the EA order parameter is likely to become zero. The transfer function $\hat{f}$ with parameter $\kappa$ is defined as taking 1 for $|x| \le \kappa$ and $-1$ otherwise. Figure \[fig:distribution\] shows the relationship between a codeword and a bit of the reproduced message. Figure \[fig:distribution\] (a) is the case of compression using a nonmonotonic perceptron. In compression using a parity tree, on the other hand, the $\mu$th bit of the reproduced message is $$\hat{y}^\mu(-{\mbox{\boldmath $s$}}) = \prod_{l=1}^K {\rm sgn} \biggl( \sqrt{\frac KN} {\mbox{\boldmath $x$}}_l^{\mu} \cdot (-{\mbox{\boldmath $s$}}_l)\biggr) = (-1)^K \hat{y}^\mu({\mbox{\boldmath $s$}}). \label{eq:mirror}$$ For $K=1$, i.e., a parity tree is identical to a monotonic perceptron, $\hat{y}^\mu(-{\mbox{\boldmath $s$}}) = - \hat{y}^\mu({\mbox{\boldmath $s$}})$ holds. Here, the EA order parameter becomes $q>0$. Therefore, the distribution of codewords is biased in ${\cal S}^N$. Compression using a parity tree with $K=1$ hidden unit cannot achieve Shannon’s limit. Figure \[fig:distribution\] (b) shows the case of compression using a monotonic perceptron, i.e., a committee tree and a $K=1$ parity tree. However, for an even number of hidden units $K$, a parity tree also has the same effect as mirror symmetry. We will next discuss the case of $K\ge 2$. Let ${\cal V}({\mbox{\boldmath $s$}}) \subset {\cal S}^N$ be a set of vectors that reversed the signs of an arbitrary even number of blocks of a codeword ${\mbox{\boldmath $s$}}={}^t({\mbox{\boldmath $s$}}_1, \cdots , {\mbox{\boldmath $s$}}_K)$, e.g., ${}^t(-{\mbox{\boldmath $s$}}_1, -{\mbox{\boldmath $s$}}_2 ,{\mbox{\boldmath $s$}}_3, \cdots , {\mbox{\boldmath $s$}}_K) \in {\cal V}({\mbox{\boldmath $s$}})$. The cardinality of the set ${\cal V}({\mbox{\boldmath $s$}})$ is $$||{\cal V}({\mbox{\boldmath $s$}})|| = \sum_{n=0}^{\lfloor K/2 \rfloor} {}_K C_{2n}=2^{K-1},$$ where $\lfloor x \rfloor$ is the largest integer $\le x$. According to (\[eq:decoder\_PM\]), all $\hat{{\mbox{\boldmath $s$}}} \in {\cal V}({\mbox{\boldmath $s$}})$ provide identical output for any ${\mbox{\boldmath $x$}}_l^\mu$. The summation of all $\hat{{\mbox{\boldmath $s$}}} \in {\cal V}({\mbox{\boldmath $s$}})$ becomes $$\sum_{\hat{{\mbox{\boldmath $s$}}} \in {\cal V}({\mbox{\boldmath $s$}})} \hat{{\mbox{\boldmath $s$}}} = {}^t (\cdots , 2^{K-2}{\mbox{\boldmath $s$}}_l + 2^{K-2}(-{\mbox{\boldmath $s$}}_l),\cdots ) ={\mbox{\boldmath $0$}}.$$ This means that $2^{K-1}$ vectors with the same distortion as codeword ${\mbox{\boldmath $s$}}$ are distributed throughout ${\cal S}^N$. For instance, Fig. \[fig:distribution\] (c) shows the distribution of codewords obtained by compression using a $K=2$ parity tree. The set ${\cal S}^N$ is divided by two $N-1$-dimensional hyperplanes whose normal vectors are orthogonal to each other. For the $\mu$th bit of the reproduced message, the normal vectors of hyperplanes are ${}^t({\mbox{\boldmath $x$}}_1^\mu,{\mbox{\boldmath $0$}})$ and ${}^t({\mbox{\boldmath $0$}},{\mbox{\boldmath $x$}}_2^\mu) \in \mathbb{R}^N$. Figure \[fig:distribution\] (d) shows the case of compression using a $K=3$ parity tree. Here, although the same effect as mirror symmetry cannot be seen, nevertheless, EA order parameter $q$ becomes zero for the reason mentioned above. This situation is the same for $K \ge 4$. With respect to LDGM code [@Murayama2003], Murayama succeeded in developing a practical encoder using the Thouless-Anderson-Palmer (TAP) approach which introduced inertia term heuristically [@Murayama2004]. The TAP approach is called belief propagation (BP) in the field of information theory. Hosaka et al applied this inertia term introduced BP to compression using a nonmonotonic perceptron [@Hosaka2005b]. In compression using a parity tree with $K$ hidden units, the number of codewords which give a minimun distortion is $2^{K-1}$. Therefore, it may become easy to find codewords as the number of hidden units $K$ becomes large. But, in a practical encoding problem, it may not be easy to use a large $K$ since $K \ll N$ is needed. ![ Relationship between codeword and bit of reproduced message in lossy compression using parity tree with $K$ hidden units. Symbol $+$ denotes bit of the reproduced message is $1$ and $-$ denotes that it is $-1$. Set ${\cal S}^N$ is divided by $K$ hyperplanes, whose normal vectors are orthogonal each other. For $K\ge 2$, vectors with same distortion as codeword ${\mbox{\boldmath $s$}}$ are distributed throughout ${\cal S}^N$. (a) a nonmonotonic perceptron, $q=0$, (b) a $K=1$ parity tree, $q>0$, (c) a $K=2$ parity tree, $q=0$, and (d) a $K=3$ parity tree, $q=0$. []{data-label="fig:distribution"}](fig5.eps){width=".8\linewidth"} conclusion ========== We investigated a lossy compression scheme for unbiased Boolean messages employing a committee tree and a parity tree, whose transfer functions were monotonic. The lower bound for achievable distortion in using a committee tree became small when the number of hidden units $K$ was large. It did not reach Shannon’s limit even in the case where $K\to\infty$. However, lossy compression using a parity tree with $K \ge 2$ hidden units could achieve Shannon’s limit where the code length became infinity. We assumed the RS ansatz in our analysis using the replica method. In using a parity tree with $K \ge 2$, the RS solution was unstable in the narrow region. For $K\ge 3$, the RS solution did not exhibit the AT instability throughout the achievable region. There is generally more than one code with the same distortion as a codeword. The EA order parameter, which means an average overlap between different codewords, need to be zero to reach Shannon’s limit like several known schemes which saturate this limit. Therefore, it may be a necessary condition that the EA order parameter becomes zero to reach Shannon’s limit. Since the encoding with our method needs exponential-time, we need to employ various efficient polynominal-time approximation encoding algorithms. It is under way to investigate the influence of the number of hidden units on the accuracy of approximation encoding algorithms. In future work, we intend to evaluate the upper bound to the rate distortion property without replica. [**ACKNOWLEDGEMENTS**]{} This work was partially supported by a Grant-in-Aid for Scientific Research on Priority Areas No. 14084212, and for Scientific Research (C) No. 16500093, and for Encouragement of Young Scientists (B) No. 15700141 from the Ministry of Education, Culture, Sports, Science and Technology of Japan. Analytical Evaluation using the replica method {#appendix.ReplicaMethod} ============================================== The entropy $S(D,R)$ can be evaluated by the replica method: $$S(D,R) = \lim_{n \to 0} \frac 1{nN} \ln <{\cal N}^n (D,R)>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}}.$$ A moment ${\cal N}^n(D,R)$, which is the number of codewords with respect to an $n$-replicated system, can be represented as $${\cal N}^n(D,R) = \mathop{\rm Tr}_{{\mbox{\boldmath $s$}}^1,\cdots ,{\mbox{\boldmath $s$}}^n} \prod_{a=1}^n \delta \biggl( MD ; d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}}({\mbox{\boldmath $s$}}^a)) \biggr),$$ where ${\mbox{\boldmath $s$}}^a ={}^t({\mbox{\boldmath $s$}}_1^a,\cdots ,{\mbox{\boldmath $s$}}_K^a)$ and the superscript $a$ denotes a replica index. Inserting an identity $$\begin{aligned} 1 &=& \prod_{a<b} \prod_{l=1}^K \int_{-\infty}^\infty dq_l^{ab} \delta \biggl( {\mbox{\boldmath $s$}}_l^a \cdot {\mbox{\boldmath $s$}}_l^b -\frac NK q_l^{ab} \biggr) \nonumber \\ &=& \biggl( \frac 1{2\pi i} \biggr)^{n(n-1)K/2} \int \biggl( \prod_{a<b} \prod_l dq_l^{ab} d\hat{q}_l^{ab} \biggr) \nonumber \\ & & \times \exp \biggl[ \sum_{a<b} \sum_l \hat{q}_l^{ab} \biggl({\mbox{\boldmath $s$}}_l^a \cdot {\mbox{\boldmath $s$}}_l^b -\frac NK q_l^{ab} \biggr) \biggr] ,\end{aligned}$$ into this expression to separate the relevant order parameter. Utilizing the Fourier expression of Kronecker’s delta, $$\begin{aligned} \delta \biggl( MD;d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}}({\mbox{\boldmath $s$}}^a)) \biggr) =& \int_{i(c-\pi)}^{i(c+\pi)} \frac{d\beta^a}{2\pi i} e^{\beta^a(D-d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}}({\mbox{\boldmath $s$}}^a)))}, \nonumber \\ &\qquad \qquad \forall c \in \mathbb{R},\end{aligned}$$ we can calculate the average moment $<{\cal N}^n (D,R)>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}}$ for natural numbers $n$ as $$\begin{aligned} <{\cal N}^n (D,R)>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} &\simeq& \int \biggl( \prod_a d\beta^a \biggr) \int \biggl( \prod_{a<b} \prod_l dq_l^{ab} d\hat{q}_l^{ab} \biggr) \exp N \biggl[ \nonumber \\ & & R^{-1} \ln \biggl< \int \biggl( \prod_l d{\mbox{\boldmath $u$}}_l d{\mbox{\boldmath $v$}}_l \biggr) \prod_l e^{-\frac 12 {}^t{\mbox{\boldmath $v$}}_l Q_l {\mbox{\boldmath $v$}}_l + i{\mbox{\boldmath $v$}}_l\cdot{\mbox{\boldmath $u$}}_l} \prod_a \biggl\{ e^{-\beta^a} + (1-e^{-\beta^a}) \Theta ( y,\{u_l^a\}) \biggr\} \biggr>_y \nonumber \\ & & + \frac 1K \ln \mathop{\rm Tr}_{\{s_l^a\}} \exp \biggl( \sum_{a<b} \sum_l \hat{q}_l^{ab} s_l^a s_l^b \biggr) - \frac 1K \sum_{a<b} \sum_l q_l^{ab} \hat{q}_l^{ab} + R^{-1} D \sum_a \beta^a \biggr], \label{eq:Nn}\end{aligned}$$ where $Q_l$ is an $n \times n$ matrix having matrix elements $\{ q_l^{ab} \}$ and $< h(y) >_y = \sum_{y\in\{-1,1\}} [\frac 12\delta(y-1)+\frac 12\delta(y+1)]h(y)$. Function $\Theta ( y,\{u_l^a\})$ included in the right hand side of (\[eq:Nn\]) depends on the decoder (details are discussed in the following sections). We analyze a system in the thermodynamic limit $N,M \to \infty$, while code rate $R$ is kept finite. This integral (\[eq:Nn\]) will be dominated by the saddle-point of the extensive exponent and can be evaluated via a saddle-point problem with respect to $\beta^a, q_l^{ab}$ and $\hat{q}_l^{ab}$. Here, we assume the replica symmetric (RS) ansatz: $$\left\{ \begin{array}{l} \beta_a = \beta, \\ q_l^{ab} = (1-q) \delta_{ab} + q, \\ \hat{q}_l^{ab} = (1-\hat{q}) \delta_{ab} + \hat{q}, \end{array} \right.$$ where $\delta_{k,k'}$ is Kronecker’s delta taking 1 if $k=k'$ and 0 otherwise. This ansatz means that all the hidden units are equivalent after averaging over the disorder. Lossy compression using committee tree\ for general $K$ {#appendix.RM.CT.generalK} --------------------------------------- In the lossy compression using the committee tree, the $\Theta ( y,\{u_l^a\})$ included in (\[eq:Nn\]) is obtained as $$\Theta ( y,\{u_l^a\}) = \Theta \biggl( y\sum_{l=1}^K {\mbox{sgn}}(u_l^a) \biggr).$$ Therefore, we obtain average entropy $S_{CT}(D,R)$ as $$\begin{aligned} S_{CT}(D,R) &=& \mathop{\rm extr}_{\beta,q,\hat{q}} \biggl( R^{-1} \biggl< \int \biggl( \prod_{l=1}^K Dt_l \biggr) \nonumber \\ & & \times \ln \{ e^{-\beta} + (1-e^{-\beta}) \Sigma (\{t_l\};y) \} \biggr>_y \nonumber \\ & & + \int Du \ln 2 \cosh \sqrt{\hat{q}} u - \frac{\hat{q}(1-q)}2 \nonumber \\ & & + R^{-1}\beta D \biggr), \label{eq:app.S}\end{aligned}$$ where $$\Sigma (\{t_l\};y) \equiv \mathop{\rm Tr}_{\{ \tau_l =\pm 1\}} \Theta \biggl( -y\sum_{l=1}^K \tau_l \biggr) \prod_{l=1}^K H(Q \tau_l t_l), \label{eq:app.Sigma}$$ with $Q \equiv \sqrt{q/(1-q)}$. Utilizing the Fourier expression of the step function $\Theta(x)=\int_0^\infty d\lambda \int_{-i\infty}^{i\infty} \frac{d\hat{\lambda}}{2\pi i}e^{\hat{\lambda}(\lambda -x)}$, the saddle-point equations $\frac{\partial S}{\partial \beta}=\frac{\partial S}{\partial q}=\frac{\partial S}{\partial \hat{q}}=0$ become $$\begin{aligned} q &=& \int Du \tanh^2 \sqrt{\hat{q}} u, \\ \hat{q} &=& 2R^{-1} \biggl< \int \biggl( \prod_{l=1}^K Dt_l \biggr) \frac{-(1-e^{-\beta})\Sigma '(\{ t_l \};y)}{e^{-\beta}+(1-e^{-\beta})\Sigma (\{ t_l \};y)} \biggr>_y, \nonumber \\ & & \\ D &=& \biggl< \int \biggl( \prod_{l=1}^K Dt_l \biggr) \frac{e^{-\beta}-e^{-\beta}\Sigma (\{ t_l \};y)}{e^{-\beta}+(1-e^{-\beta})\Sigma (\{ t_l \};y)} \biggr>_y, \end{aligned}$$ where $\Sigma '(\{t_l\};y) \equiv \partial \Sigma (\{t_l\};y)/\partial q$. Substituting the solutions to the saddle-point equations into (\[eq:app.S\]), average entropy $S_{CT}(D,R)$ is obtained. Thus, for any $K$, we can obtain a minimum code rate $R$, which gives $S_{CT}(D,R)=0$ for a fixed distortion $D$. Lossy compression using committee tree\ for large $K$ {#appendix.RM.CT.largeK} --------------------------------------- We concentrate in the following on the simple case of large $K$, where the $K$-multiple integrals can be reduced to a single Gaussian integral. We assume that the number of hidden units $K$ is large but still $K \ll N$. Here, the term $\Sigma (\{t_l\};y)$ included in (\[eq:app.S\]) does not depend on all the individual integration variables $t_l$ but only on the combination $\sum_{l=1}^K [2H(Qt_l)-1]$. With the central limit theorem, the term is given by $$\begin{aligned} \Sigma (\{t_l\};y) &=& \int_0^\infty d\lambda \int_{-\infty}^\infty \frac{d\hat{\lambda}}{2\pi} \exp \biggl\{ i \hat{\lambda} \lambda \nonumber \\ & & + i \hat{\lambda} y \frac 1{\sqrt{K}}\sum_l [2H(Qt_l)-1] \nonumber \\ & & -\hat{\lambda}^2 \biggl( 1- \frac 1K \sum_l [2H(Qt_l)-1]^2 \biggr) \biggr\}. \nonumber \\\end{aligned}$$ Therefore, we obtain averaged entropy as $$\begin{aligned} S_{CT}(D,R) &=& \mathop{\rm extr}_{\beta,q,\hat{q}} \biggl( R^{-1} \biggl< \int Dt \ln \{ e^{-\beta} \nonumber \\ & & + (1-e^{-\beta}) H \biggl( \sqrt{\frac{q_{eff}}{1-q_{eff}}}t \biggr) \} \biggr>_y \nonumber \\ & & + \int Du \ln 2 \cosh \sqrt{\hat{q}} u - \frac{\hat{q}(1-q)}2 \nonumber \\ & & + R^{-1}\beta D \biggr) , \label{eq:app.S_CT}\end{aligned}$$ where $q_{eff} \equiv \int Dt [1-2H(Qt)]^2 = \frac 2{\pi} \sin^{-1} q$ and the saddle-point equations are $$\begin{aligned} q &=& \int Du \tanh^2 \sqrt{\hat{q}} u, \\ \hat{q} &=& 2R^{-1} \biggl< \int Dt \frac{-(1-e^{-\beta}) H'(Q_{eff}t)}{e^{-\beta}+(1-e^{-\beta})H(Q_{eff}t)} \biggr>_y, \\ D &=& \biggl< \int Dt \frac{e^{-\beta}-e^{-\beta}H(Q_{eff}t)}{e^{-\beta}+(1-e^{-\beta})H(Q_{eff}t)} \biggr>_y, \end{aligned}$$ with $Q_{eff} \equiv \sqrt{q_{eff}/(1-q_{eff})}$ and $H'(Q_{eff}t) \equiv \partial H(Q_{eff}t)/\partial q$. Lossy compression using parity tree\ for general $K$ {#appendix.RM.PT} ------------------------------------ In the lossy compression using the parity tree, on the other hand, the $\Theta ( y,\{u_l^a\})$ included in (\[eq:Nn\]) is obtained as $$\Theta ( y,\{u_l^a\}) = \Theta \biggl( y\prod_l {\mbox{sgn}}(u_l^a) \biggr),$$ Hence, we obtain averaged entropy $S_{PT}(D,R)$ as $$\begin{aligned} S_{PT}(D,R) &=& \mathop{\rm extr}_{\beta,q,\hat{q}} \biggl( R^{-1} \biggl< \int \biggl( \prod_{l=1}^K Dt_l \biggr) \nonumber \\ & & \times \ln \{ e^{-\beta} + (1-e^{-\beta}) \Pi (\{t_l\};y) \} \biggr>_y \nonumber \\ & & + \int Du \ln 2 \cosh \sqrt{\hat{q}} u - \frac{\hat{q}(1-q)}2 \nonumber \\ & & + R^{-1}\beta D \biggr), \label{eq:app.S_PT}\end{aligned}$$ where $$\Pi (\{t_l\};y) \equiv \frac 12 \biggl( 1+y\prod_{l=1}^K [1-2H(Q t_l)] \biggr) .$$ For cases utilizing a committee tree and a parity tree, only terms $\Sigma (\{t_l\};y)$ and $\Pi (\{t_l\};y)$ are different. Since both the order parameters $q$ and $\hat{q}$ at the saddle-point of (\[eq:app.S\_PT\]) are less than one, the average entropy $S_{PT}(D,R)$ can be expanded with respect to $\prod_{l=1}^K [1-2H(Q t_l)] (<1)$ as $$\begin{aligned} S_{PT}(D,R) &=& \mathop{\rm extr}_{\beta,q,\hat{q}} \biggl( R^{-1} \biggl\{ \ln \frac {1+e^{-\beta}}2 \nonumber \\ & & - \sum_{m=1}^\infty \frac 1{2m} \biggl( \frac {1-e^{-\beta}}{1+e^{-\beta}} \biggr)^{2m} \nonumber \\ & & \times \biggl[ \int Dt [1-2H(Qt)]^{2m} \biggr]^K \biggr\} \nonumber \\ & & + \int Du \ln 2 \cosh \sqrt{\hat{q}} u - \frac{\hat{q}(1-q)}2 \nonumber \\ & & + R^{-1}\beta D \biggr). \end{aligned}$$ We obtain saddle-point equations using this expanded form of the averaged entropy: $$\begin{aligned} q &=& \int Du \tanh^2 \sqrt{\hat{q}} u, \label{eq:app.q.PM} \\ \hat{q} &=& 2R^{-1} K \sum_{m=1}^\infty \biggl( \frac{1-e^{-\beta}}{1+e^{-\beta}} \biggr)^{2m} \nonumber \\ & & \times \biggl[ \int Dt (1-2H(Qt))^{2m} \biggr]^{K-1} \nonumber \\ & & \times \int Dt (1-2H(Qt))^{2m-1} \frac{te^{-(Qt)^2/2}}{\sqrt{2\pi q}(1-q)^{3/2}}, \nonumber \\ & & \label{eq:app.q^.PM} \\ D &=& \frac{e^{-\beta}}{1+e^{-\beta}} + \sum_{m=1}^\infty \frac{2e^{-\beta}}{(1+e^{-\beta})^2} \biggl( \frac{1-e^{-\beta}}{1+e^{-\beta}} \biggr)^{2m-1} \nonumber \\ & & \times \biggl[ \int Dt (1-2H(Qt))^{2m} \biggr]^K. \label{eq:app.D.PM} \end{aligned}$$ For $K\ge 2$, because of the existence of term $[ \int Dt (1-2H(Qt))^{2m} ]^{K-1}$ in (\[eq:app.q\^.PM\]), solutions to the saddle-point equations can become $q=\hat{q}=0$. We can find no other solutions except for $q=\hat{q}=0$ by solving (\[eq:app.q.PM\])-(\[eq:app.D.PM\]) numerically for $K\ge 2$. Substituting this into (\[eq:app.D.PM\]), we obtain $D=e^{-\beta}/(1+e^{-\beta})$. Almeida-Thouless instability of replica symmetric solution {#appendix.ATline} ========================================================== General case {#appendix.ATline.general} ------------ The Hessian computed at the replica symmetric saddle-point characterizes fluctuations in the order parameters $\beta^a$, $q_l^{ab}$ and $\hat{q}_l^{ab}$ around the RS saddle-point. Instability of the RS solution is signaled by a change of sign of at least one of the eigenvalues of the Hessian. Let ${\cal M}(\{\beta^a\},\{ q_l^{ab} \},\{\hat{q}_l^{ab}\})$ be the exponent of the integrand of the integral (\[eq:Nn\]). Equation (\[eq:Nn\]) can be represented as $$\begin{aligned} & & <{\cal N}^n(D,R)>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ &=& \int \biggl( \prod_a d\beta^a \biggr) \int \biggl( \prod_{a<b} \prod_l dq_l^{ab} d\hat{q}_l^{ab} \biggr) \nonumber \\ & & \times \exp ( N {\cal M}(\{\beta^a\},\{ q_l^{ab} \},\{\hat{q}_l^{ab}\}) ). \end{aligned}$$ We expand ${\cal M}$ around $\beta$, $q$ and $\hat{q}$ in $\delta\beta^a$, $\delta q_l^{ab}$ and $\delta \hat{q}_l^{ab}$ and then find up to second order $$\begin{aligned} & & {\cal M}(\{\beta+\delta\beta^a\},\{ q+q_l^{ab} \},\{\hat{q}+\delta\hat{q}_l^{ab}\}) \nonumber \\ &=& {\cal M}(\{\beta\},\{q\},\{\hat{q}\}) + \frac 12 {}^t{\mbox{\boldmath $\nu$}} G {\mbox{\boldmath $\nu$}} + {\cal O}(||{\mbox{\boldmath $\nu$}}||^3), \end{aligned}$$ where $${\mbox{\boldmath $\nu$}}={}^t(\{ \delta \beta^a \} , \{ \delta q_1^{ab} \} , \{ \delta \hat{q}_1^{ab} \} , \cdots , \{ \delta q_K^{ab} \} , \{ \delta \hat{q}_K^{ab} \} ),$$ is the perturbation to the RS saddle-point. The Hessian $G$ is the following $[n+Kn(n-1)] \times [n+Kn(n-1)]$ matrix: $$G= \left( \begin{array}{ccccc} S & T & T & \cdots & T \\ {}^tT & U & V & \cdots & V \\ {}^tT & V & U & \cdots & V \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ {}^tT & V & V & \cdots & U \end{array} \right),$$ where $n \times n$ matrix $S$, $n \times n(n-1)$ matrix $T$ and $n(n-1) \times n(n-1)$ matrices $U, V$ are $$\begin{aligned} S &=& (\{S^{a,b}\}), \nonumber \\ T &=& (\{T^{a,bc}\},\{\hat{T}^{a,bc}\}), \nonumber \\ U &=& \left( \begin{array}{cc} \{U^{ab,cd}\} & \{\tilde{U}^{ab,cd}\} \\ \{\tilde{U}^{ab,cd}\} & \{\hat{U}^{ab,cd}\} \end{array} \right), \nonumber \\ V &=& \left( \begin{array}{cc} \{V^{ab,cd}\} & \{\tilde{V}^{ab,cd}\} \\ \{\tilde{V}^{ab,cd}\} & \{\hat{V}^{ab,cd}\} \end{array} \right), \end{aligned}$$ with $$\begin{aligned} S^{a,b} &=& \partial^2 {\cal M}/\partial \beta^a \partial \beta^b, \nonumber \\ T^{a,bc} &=& \partial^2 {\cal M}/\partial \beta^a \partial q_l^{bc}, \nonumber \\ \hat{T}^{a,bc} &=& \partial^2 {\cal M}/\partial \beta^a \partial \hat{q}_l^{bc}, \nonumber \\ U^{ab,cd} &=& \partial^2 {\cal M}/\partial q_l^{ab} \partial q_l^{cd}, \nonumber \\ \hat{U}^{ab,cd} &=& \partial^2 {\cal M}/\partial \hat{q}_l^{ab} \partial \hat{q}_l^{cd}, \nonumber \\ \tilde{U}^{ab,cd} &=& \partial^2 {\cal M}/\partial q_l^{ab} \partial \hat{q}_l^{cd}, \nonumber \\ V^{ab,cd} &=& \partial^2 {\cal M}/\partial q_l^{ab} \partial q_{l'}^{cd} \quad (l\ne l'), \nonumber \\ \hat{V}^{ab,cd} &=& \partial^2 {\cal M}/\partial \hat{q}_l^{ab} \partial \hat{q}_{l'}^{cd} \quad (l\ne l'), \nonumber \\ \tilde{V}^{ab,cd} &=& \partial^2 {\cal M}/\partial q_l^{ab} \partial \hat{q}_{l'}^{cd} \quad (l\ne l'). \end{aligned}$$ For $(\beta,q,\hat{q})$ to be a local maximum of ${\cal M}$, it is necessary for the Hessian $G$ to be negative definite, i.e., all of its eigenvalues must be negative. Matrices $U$ and $V$ contain the quadratic fluctuations of the order parameters in the same and different hidden units, respectively. Because of the block form of $G$, the eigenproblem splits into an uncoupled diagonalization of the two matrices: $U-V$ and $$\hat{G}= \left( \begin{array}{ccc} S & \; & T \\ K{}^tT & \; & U+(K-1)V \end{array} \right).$$ The eigenvectors of $U-V$ correspond to fluctuations in directions that break the permutation symmetry (PS). The eigenvectors of $\hat{G}$ represent fluctuations that do not break this symmetry. The most unstable mode corresponds to an eigenvector of $\hat{G}$ that breaks the replica symmetry (RS). We can write the eigenvalue equation as $$\hat{G} {\mbox{\boldmath $\mu$}} = \lambda {\mbox{\boldmath $\mu$}},$$ with $${\mbox{\boldmath $\mu$}}={}^t(\{ \epsilon^a \} , \{ \eta^{ab} \} , \{ \hat{\eta}^{ab} \} ).$$ There are three types of eigenvectors, i.e., ${\mbox{\boldmath $\mu$}}_1$, ${\mbox{\boldmath $\mu$}}_2$ and ${\mbox{\boldmath $\mu$}}_3$ [@Almeida1978]. The first ${\mbox{\boldmath $\mu$}}_1$ has the form: $$\epsilon^a=\epsilon, \quad \eta^{ab}=\eta, \quad \hat{\eta}^{ab}=\hat{\eta}.$$ Using the orthogonality of ${\mbox{\boldmath $\mu$}}_1$ and ${\mbox{\boldmath $\mu$}}_2$, the second type of eigenvector ${\mbox{\boldmath $\mu$}}_2$ has the form: $$\begin{aligned} \epsilon^a &=& \left\{ \begin{array}{ll} (1-n) \epsilon, & (a=\theta), \\ \epsilon, & ({\rm otherwise}), \end{array} \right. \nonumber \\ \eta^{ab} &=& \left\{ \begin{array}{ll} \frac 12(2-n) \eta, & (a=\theta \;{\rm or}\; b=\theta), \\ \eta, & ({\rm otherwise}), \end{array} \right. \nonumber \\ \hat{\eta}^{ab} &=& \left\{ \begin{array}{ll} \frac 12(2-n) \hat{\eta}, & (a=\theta \;{\rm or}\; b=\theta), \\ \hat{\eta}, & ({\rm otherwise}), \end{array} \right. ,\end{aligned}$$ for a specific replica $\theta$. In the limit $n\to 0$ this eigenvector ${\mbox{\boldmath $\mu$}}_2$ converges to ${\mbox{\boldmath $\mu$}}_1$, therefore the eigenvalue of the eigenvector ${\mbox{\boldmath $\mu$}}_2$ becomes degenerate with ${\mbox{\boldmath $\mu$}}_1$’s. Similarly, using the orthogonality of ${\mbox{\boldmath $\mu$}}_2$ and ${\mbox{\boldmath $\mu$}}_3$, the third type of eigenvector ${\mbox{\boldmath $\mu$}}_3$ has the form: $$\begin{aligned} \epsilon^a &=& 0, \nonumber \\ \eta^{ab} &=& \left\{ \begin{array}{l} \frac 12(2-n)(3-n) \eta, \; (a=\theta,b=\nu), \\ \frac 12(3-n) \eta, \\ \quad (a=\theta \;{\rm or}\; a=\nu \;{\rm or}\; b=\theta \;{\rm or}\; b=\nu), \\ \eta, \; ({\rm otherwise}), \end{array} \right. \nonumber \\ \hat{\eta}^{ab} &=& \left\{ \begin{array}{l} \frac 12(2-n)(3-n) \hat{\eta}, \; (a=\theta,b=\nu), \\ \frac 12(3-n) \hat{\eta}, \\ \quad (a=\theta \;{\rm or}\; a=\nu \;{\rm or}\; b=\theta \;{\rm or}\; b=\nu), \\ \hat{\eta}, \; ({\rm otherwise}), \end{array} \right. , \nonumber \\ \label{eq:app.repliconmode}\end{aligned}$$ for two specific replicas $\theta$ and $\mu$. In the limit $n\to 0$, perturbations keep symmetry of the eigenvectors ${\mbox{\boldmath $\mu$}}_1$ and ${\mbox{\boldmath $\mu$}}_2$ across the replicas. Therefore, ${\mbox{\boldmath $\mu$}}_1$ and ${\mbox{\boldmath $\mu$}}_2$ are irrelevant to replica symmetry breaking (RSB) but only determines the stability within the RS ansatz. Hence, the third eigenvector ${\mbox{\boldmath $\mu$}}_3$, which is called the replicon mode, causes RSB. The eigenvalue equation $\hat{G}{\mbox{\boldmath $\mu$}}_3=\lambda_3{\mbox{\boldmath $\mu$}}_3$ with respect to (\[eq:app.repliconmode\]) splits into $T{\mbox{\boldmath $\mu$}}_3={\mbox{\boldmath $0$}}$ and $[U+(K-1)V]{\mbox{\boldmath $\mu$}}_3'=\lambda_3{\mbox{\boldmath $\mu$}}_3'$, where ${\mbox{\boldmath $\mu$}}_3 = {}^t({\mbox{\boldmath $0$}}, {\mbox{\boldmath $\mu$}}_3')$. Therefore, the eigenproblem of $\hat{G}$ is equivalent to that of $U+(K-1)V$. Let us calculate the elements of $U$ and $V$. The second derivative ${\cal M}$ by $q_l^{ab}$ related to the $U^{ab,cd}, V^{ab,cd}$ is $$\begin{aligned} \frac{\partial^2 {\cal M}}{\partial q_l^{ab} \partial q_{l'}^{cd}} &=& R^{-1}<v_l^a v_l^b v_{l'}^c v_{l'}^d>_{u,v} \nonumber \\ & & - R^{-1}<v_l^a v_l^b>_{u,v} <v_{l'}^c v_{l'}^d>_{u,v}, \label{eq:app.PQRP'Q'R'}\end{aligned}$$ where $$<g(\{ v_l^a \})>_{u,v}=\frac {\displaystyle \biggl< \int \biggl( \prod_l d{\mbox{\boldmath $u$}}_l d{\mbox{\boldmath $v$}}_l e^{-\frac 12 {}^t{\mbox{\boldmath $v$}}_l Q_l {\mbox{\boldmath $v$}}_l + i{\mbox{\boldmath $v$}}_l\cdot{\mbox{\boldmath $u$}}_l} \biggr) g(\{ v_l^a \}) \prod_a \biggl\{ e^{-\beta^a} + (1-e^{-\beta^a}) \Theta ( y,\{u_l^a\}) \biggr\} \biggr>_y} {\displaystyle \biggl< \int \biggl( \prod_l d{\mbox{\boldmath $u$}}_l d{\mbox{\boldmath $v$}}_l e^{-\frac 12 {}^t{\mbox{\boldmath $v$}}_l Q_l {\mbox{\boldmath $v$}}_l + i{\mbox{\boldmath $v$}}_l\cdot{\mbox{\boldmath $u$}}_l} \biggr) \prod_a \biggl\{ e^{-\beta^a} + (1-e^{-\beta^a}) \Theta ( y,\{u_l^a\}) \biggr\} \biggr>_y},$$ for any function $g(\{ v_l^a \})$. The second derivative ${\cal M}$ by $\hat{q}_l^{ab}$ related to the $\hat{U}^{ab,cd}, \hat{V}^{ab,cd}$ is $$\begin{aligned} \frac{\partial^2 {\cal M}}{\partial \hat{q}_l^{ab} \partial \hat{q}_{l'}^{cd}} &=& \left\{ \begin{array}{ll} K^{-1}<s^a s^b s^c s^d>_s & \\ \quad - K^{-1}<s^a s^b>_s <s^c s^d>_s, & \; (l=l'), \\ 0, & \; (l\ne l'), \end{array} \right. \nonumber \\ & & \label{eq:app.P^Q^R^P^'Q^'R^'}\end{aligned}$$ where $$<g(\{ s^a \})>_s=\frac {\displaystyle \int Dz \mathop{\rm Tr}_{\{ s^a\}} g(\{ s^a \}) \exp \biggl( \sqrt{\hat{q}}\,z \sum_a s^a \biggr)} {\displaystyle \int Dz \mathop{\rm Tr}_{\{ s^a\}} \exp \biggl( \sqrt{\hat{q}}\,z \sum_a s^a \biggr)},$$ for any function $g(\{ s^a \})$. The second derivative ${\cal M}$ by $q_l^{ab},\hat{q}_l^{ab}$ related to the $\tilde{U}^{ab,cd}, \tilde{V}^{ab,cd}$ is $$\frac{\partial^2 {\cal M}}{\partial q_l^{ab} \partial \hat{q}_{l'}^{cd}} = \left\{ \begin{array}{ll} K^{-1}, & \; (l=l', a=c, b=d), \\ 0, & \; ({\rm otherwise}). \end{array} \right. \label{eq:app.PtQtRtPt'Qt'Rt'}$$ Using Gardner’s method [@Gardner1988], we find that the RS stability criterion is $$K \gamma <1, \label{eq:app.AT}$$ where $$\begin{aligned} \gamma &\equiv& \gamma_0 + (K-1) \gamma_1, \nonumber \\ \gamma_0 &\equiv& P-2Q+R, \nonumber \\ \gamma_1 &\equiv& P'-2Q'+R', \nonumber \\ P &\equiv& U^{ab,ab}, \nonumber \\ Q &\equiv& U^{ab,ac} \quad (b\ne c), \nonumber \\ R &\equiv& U^{ab,cd} \quad (a\ne c,b\ne d), \nonumber \\ P' &\equiv& V^{ab,ab}, \nonumber \\ Q' &\equiv& V^{ab,ac} \quad (b\ne c), \nonumber \\ R' &\equiv& V^{ab,cd} \quad (a\ne c,b\ne d). \end{aligned}$$ The line $K \gamma =1$ is called the AT line. Setting $K=0$, on the other hand, the matrix $U+(K-1)V$ is equal to $U-V$. When $K=0$, inequality $K\gamma =0<1$ of (\[eq:app.AT\]) always holds. Therefore, permutation symmetry breaking (PSB) does not occur in this system. For lossy compression using a parity tree with $K = 2$ hidden units {#appendix.ATline.K=2PT} ------------------------------------------------------------------- Let us consider the RS stability of lossy compression using a parity tree with $K=2$ hidden units. Here, $\Theta ( y,\{u_l^a\})$ is given by $\Theta ( y,\{u_l^a\}) = \Theta ( y\prod_l {\mbox{sgn}}(u_l^a))$, therefore solutions to the saddle-point equations are $$q=\hat{q}=0, \; D=\frac{e^{-\beta}}{1+e^{-\beta}}. \label{eq:app.K>=2PTspe}$$ Substituting (\[eq:app.K&gt;=2PTspe\]) into (\[eq:app.PQRP’Q’R’\]) and (\[eq:app.P\^Q\^R\^P\^’Q\^’R\^’\]), we obtain $$\begin{aligned} P'&=& R^{-1}\frac 4{\pi^2}(1-2D)^2, \nonumber \\ P &=& Q=R=Q'=R'=0. \end{aligned}$$ Therefore, using (\[eq:app.AT\]), the RS stability can be obtained as $$R > \frac 8{\pi^2}(1-2D)^2 \equiv R_{AT}(D).$$ This proves (\[eq:ATline\]). For lossy compression using a parity tree with $K \ge 3$ hidden units {#appendix.ATline.K>=3PT} --------------------------------------------------------------------- Next, let us consider the RS stability of lossy compression using a parity tree with $K \ge 3$ hidden units. Here, the solutions to the saddle-point equations are $q=\hat{q}=0, D=e^{-\beta}/(1+e^{-\beta})$ as well as for $K=2$. Substituting (\[eq:app.K&gt;=2PTspe\]) into (\[eq:app.PQRP’Q’R’\]) and (\[eq:app.P\^Q\^R\^P\^’Q\^’R\^’\]), we obtain $$P=Q=R=P'=Q'=R'=0.$$ Since the inequality $K \gamma =0<1$ of (\[eq:app.AT\]) always holds, the RS solution does not exhibit the AT instability throughout the achievable region for $K\ge 3$. A lower bound to the rate distortion property of lossy compression using a parity tree {#appendix.lower_bound} ====================================================================================== In order to derive a lower bound to the rate distortion property, an upper bound to the entropy is necessary. Using Jensen’s inequality, an upper bound to the entropy $S^{\; upper}(D,R)$ is given by $$\begin{aligned} S(D,R) &= \frac 1N <\ln {\cal N}(D,R)>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ &\le \frac 1N \ln <{\cal N}(D,R)>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ &\equiv S^{\; upper}(D,R). \end{aligned}$$ After a simple calculation, we obtain the upper bound to the entropy of lossy compression using a parity tree $S_{PT}^{\; upper}(D,R)$ as $$\begin{aligned} S_{PT}^{\; upper}(D,R) =& \ln 2 + \mathop{\rm extr}_{\beta} \biggl( R^{-1} \ln \frac{1+e^{-\beta}}2 + \beta R^{-1}D \biggr) \nonumber \\ =& -R^{-1}\ln 2+ \ln 2 -R^{-1}D\ln D \nonumber \\ & -R^{-1}(1-D)\ln (1-D). \end{aligned}$$ Note that this annealed entropy $S_{PT}^{\; anneal}(D,R)$ is not depend on the number of hidden units $K$. Solving $S_{PT}^{\; anneal}(D,R)=0$ with respect to $R$, we obtain $$R = 1-h_2(D).$$ This shows that the rate distortion function for uniformly unbiased binary sources (\[eq:RDF\]) can be also derived as a lower bound to the rate distortion property of compression using a parity tree. We next discuss a upper bound to the rate distortion property. In order to derive a upper bound to the rate distortion property, we need an lower bound to the entropy. Using Jensen’s inequality, an upper bound to the entropy $S^{\; upper}(D,R)$ is represented by $$\begin{aligned} S(D,R) &= \frac 1N < \ln {\cal N}(D,R)>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ &= \frac 1N \biggl< - \ln \frac 1{{\cal N}(D,R)} \biggr>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ &\ge - \frac 1N \ln \biggl< \frac 1{{\cal N}(D,R)} \biggr>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ &\equiv S^{\; lower}(D,R). \end{aligned}$$ This inequality can be also obtained by an annealed information entropy as follows. According to Opper’s discussion [@Opper1995], we first define a function that characterizes a version space as follows: $$\begin{aligned} \rho({\mbox{\boldmath $s$}}) \equiv \frac{\delta \biggl( MD ; d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}}({\mbox{\boldmath $s$}})) \biggr)}{\displaystyle{\mathop{\rm Tr}_{{\mbox{\boldmath $s$}}}\delta \biggl( MD ; d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}}({\mbox{\boldmath $s$}})) \biggr)}}. \end{aligned}$$ Since this function $\rho({\mbox{\boldmath $s$}})$ is non-negative and normalized to $\mathop{\rm Tr}_{{\mbox{\boldmath $s$}}} \rho({\mbox{\boldmath $s$}})=1$, it defines a probability with respect to ${\mbox{\boldmath $s$}}$. Therefore we obtain the information entropy per bit ${\cal H}(D,R)$ as $$\begin{aligned} {\cal H}(D,R) \equiv& \frac 1N \biggl< \mathop{\rm Tr}_{{\mbox{\boldmath $s$}}} \rho({\mbox{\boldmath $s$}}) \ln \frac 1{\rho({\mbox{\boldmath $s$}})} \biggr>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ = & \frac 1N \biggl< \ln \frac 1{\rho({\mbox{\boldmath $s$}})} \biggr>_{{\mbox{\boldmath $s$}},{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ \ge & - \frac 1N \ln < \rho({\mbox{\boldmath $s$}}) >_{{\mbox{\boldmath $s$}},{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ = & - \frac 1N \ln < \mathop{\rm Tr}_{{\mbox{\boldmath $s$}}} \rho({\mbox{\boldmath $s$}})^2 >_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \nonumber \\ = & - \frac 1N \ln \biggl< \frac 1{{\cal N}(D,R)} \biggr>_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}}, \end{aligned}$$ where $< g({\mbox{\boldmath $s$}}) >_{{\mbox{\boldmath $s$}}} = \mathop{\rm Tr}_{{\mbox{\boldmath $s$}}} \rho({\mbox{\boldmath $s$}}) g({\mbox{\boldmath $s$}})$. Using the identity $$\begin{aligned} \rho({\mbox{\boldmath $s$}}) \ln \frac 1{\rho({\mbox{\boldmath $s$}})}= \left\{ \begin{array}{l} 0, \quad {\rm if} \; \delta ( MD ; d({\mbox{\boldmath $y$}},\hat{{\mbox{\boldmath $y$}}}({\mbox{\boldmath $s$}})) )=0, \\ {\cal N}(D,R)^{-1} \ln {\cal N}(D,R), \; {\rm otherwise}, \end{array} \right.\end{aligned}$$ we can easily confirm ${\cal H}(D,R)=S(D,R)$. However, it is difficult to evaluate the lower bound $S^{\; lower}(D,R)$ directly because $< {{\cal N}(D,R)}^{-1} >_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}} \ge < {{\cal N}(D,R)} >_{{\mbox{\boldmath $y$}},{\mbox{\boldmath $x$}}}^{-1}$. This difficulty is caused by a limitation of the version space due to the distortion. This limitation complicates the probability $\rho({\mbox{\boldmath $s$}})$. [12]{} N. Sourlas, Nature, [**339**]{}, 693 (1989). Y. Kabashima, T. Murayama and D. Saad, Phys. Rev. Lett., [**84**]{}, 1355 (2000). H. Nishimori and K. Y. M. Wong, Phys. Rev. E, [**60**]{}, 132 (1999). A. Montanari and N. Sourlas, Eur. Phys. J. B, [**18**]{}, 107 (2000). T. Tanaka, Europhys. Lett., [**54**]{}, 540 (2001). T. Tanaka and M. Okada, IEEE Trans. Inform. Theory, [**51**]{}, 2, 700 (2005). T. Murayama and M. Okada, J. Phys. A: Math. Gen.,[**36**]{}, 11123 (2003). T. Murayama, Phys. Rev. E, [**69**]{}, 035105(R) (2004). T. Hosaka, Y. Kabashima and H. Nishimori, Phys. Rev. E, [**66**]{}, 066126 (2002). T. Hosaka and Y. Kabashima, J. Phys. Soc. Jpn., [**74**]{}, 1, 488 (2005). T. Hosaka and Y. Kabashima, [*Preprint*]{}, arXiv.org, cs.IT/0509086. C. E. Shannon, Bell Syst. Tech. J., [**27**]{}, 379 (1948). C. E. Shannon, IRE Nat. Conv. Rec., [**4**]{}, 142 (1959). T. M. Cover and J. A. Thomas, Elements of Information Theory (Wiley, New York, 1991). E. Barkai, D. Hansel and I. Kanter, Phys. Rev. Lett., [**65**]{}, 18, 2312 (1990). E. Barkai and I. Kanter, Europhys. Lett., [**14**]{}, 2, 107 (1991). E. Barkai, D. Hansel and H. Sompolinsky, Phys. Rev. A, [**45**]{}, 6, 4146 (1992). E. Gardner, J. Phys. A: Math. Gen., [**21**]{}, 257 (1988). W. Krauth and M. Mézard, J. Phys. (France), [**50**]{}, 3057 (1989). M. Opper, Phys. Rev. E, [**51**]{}, 4, 3613 (1995). J. R. L. de Almeida and D. J. Thouless, J. Phys. A: Math. Gen., [**11**]{}, 5, 983 (1978).

ArXiv

--- abstract: 'An inverse scattering problem for the 3D acoustic equation in time domain is considered. The unknown spatially distributed speed of sound is the subject of the solution of this problem. A single location of the point source is used. Using a Carleman Weight Function, a globally strictly convex cost functional is constructed. A new Carleman estimate is proven. Global convergence of the gradient projection method is proven. Numerical experiments are conducted.' author: - 'Michael V. Klibanov[^1]' - 'Jingzhi Li[^2]' - 'Wenlong Zhang[^3]' title: 'Convexification for the Inversion of a Time Dependent Wave Front in a Heterogeneous Medium[^4]' --- Coefficient Inverse Problem, convexification, Carleman estimate, global strict convexity, numerical examples 35R30 Introduction {#sec:1} ============ We develop analytically and test numerically a globally convergent numerical method for a Coefficient Inverse Problem (CIP) for the 3D acoustic equation in the time domain. The spatially distributed speed of sound is the subject of the solution of this CIP. Only a single location of the point source is used here. For reasons explained below, we call this method convexification". In the publication [@Klib97] of the first author a globally strictly convex cost functional was constructed for the same CIP for the first time. This paper is a deep revision of [Klib97]{}. We significantly update the theory of [@Klib97]. Our goal is to make the theory of [@Klib97] computationally feasible. New results of this paper are (also, see Remark 4.1 in section 4): 1. We prove a new Carleman estimate for the Laplace’s operator (Theorem 4.1). This one is well suitable for computations. On the other hand, the Carleman estimate of [@Klib97] uses two large parameters and, therefore, changes too rapidly, which makes it inconvenient for computations. 2. Using the Carleman estimate of Theorem 4.1, we convexify“ our CIP, i.e. we construct a new weighted Tikhonov-like functional, which is strictly convex on a set $B\left( m,R\right) $ in a certain Hilbert space of an arbitrary size” $R>0$, where the number $m\in \left( 0,R\right) $ is independent on $R$. Since the number $R>0$ is an arbitrary one, then this is the *global strict convexity*. The key element of the above functional is the presence of the Carleman Weight Function (CWF) in it. Unlike [@Klib97], our functional contains a regularization parameter $\beta \in \left( 0,1\right) .$ It is the presence of this parameter which allows us to work without an inconvenient assumption of [@Klib97] (the last paragraph of page 1379 of [@Klib97]) which is about working on a certain compact set. Unlike the latter, we consider here a more convenient case of the bounded set $B\left( m,R\right) $. We call our approach convexification". 3. We prove the existence and uniqueness of the minimizer of that functional on $B\left( m,R\right) $ and, most importantly, the *global convergence* of the gradient projection method to the correct solution of our CIP. In other words, the convergence of the gradient projection method to the correct solution is guaranteed if it starts from an arbitrary point of the set $B\left( m,R\right) $ and if the noise in the data tends to zero. 4. This theory is applied then to computational studies of the proposed numerical method. Such studies are not a part of [@Klib97]. We call a numerical method for a CIP *globally* convergent if a theorem is proved, which guarantees that this method delivers at least one point in a sufficiently small neighborhood of the exact solution without any advanced knowledge of this neighborhood. Thus, since the convergence to the exact solution of the gradient projection method of the minimization of the above mentioned weighted Tikhonov-like functional is guaranteed for any starting point of the set $B\left( m,R\right) $ and since its size" $R>0$ is an arbitrary number, then the numerical method proposed in this paper is a globally convergent one. A conventional numerical methods for a CIP relies on the minimization of a Tikhonov least squares cost functional, see, e.g. [Chav,Gonch,Lakhal,Rizutti]{}. However, since, as a rule, such a functional is non convex, then the fundamental and still not addressed problem of such a method is the presence of multiple local minima and ravines in that functional. Therefore, convergence of an optimization method for such a functional can be guaranteed only if its starting point is located in a sufficiently small neighborhood of the exact solution, i.e. the *local* convergence. There exists a vast literature about numerical solutions of inverse scattering problems, which are close to the CIP we consider, see, e.g. [Am1,Am2,Cakoni,Kar,Gonch,It,Ito,Kab1,Kab2,Liu1,Liu2,Lakhal,Rizutti]{} and references cited therein. They consider either problems of finding unknown coefficients or problems of finding shapes of obstacles. In the case of unknown coefficients, these cited papers consider the case of either multiple locations of the point source or multiple directions of the incident plane wave, which is unlike our case of a single point source. Initially globally strictly convex cost functionals with CWFs in them were constructed in [@Klib95; @Klib97; @Klib971] for CIPs for hyperbolic and parabolic PDEs. Numerical studies were not conducted in these works. In the past few years an interest to this topic was renewed in publications of the first author with coauthors. New analytical results were combined with computational ones for some CIPs [KlibThanh,KlibKol1,KlibKol2,KlibKol3,KEIT]{}, including a quite challenging case of experimental data [@KlibKol4]. However, the convexification was not studied numerically for CIPs in time domain in the case when the initial condition is vanishing, as we have here. In the case of a non vanishing initial condition, a different version of the convexification was recently proposed in [@Baud] for the PDE $u_{tt}=\Delta u+a(x)u$ with the unknown coefficient $a(x)$. The case of a non vanishing initial condition is less challenging one than our case of the $\delta -$function in the initial condition. An indication of the latter is that the uniqueness theorem for the non vanishing case can be proved by the Bukhgeim-Klibanov method, which was originated in [@BukhKlib]. On the other hand, it is still unclear how to prove uniqueness theorem for our case of the $\delta -$function in the initial condition. Thus, we just assume uniqueness here for computational purposes. Since this paper is not about the Bukhgeim-Klibanov method, we cite now, in addition to the above first publication, only some limited relevant references [@BK; @Bell; @KT; @Ksurvey]. In section 2 we state the CIP we consider. In section 3 we derive a boundary value problem for a system of coupled elliptic PDEs. In section 4 we derive the above mentioned weighted globally strictly convex Tikhonov-like functional with a CWF in it and also formulate our Theorems 4.1-4.6. In section 5 we prove these theorems. In section 6 we present results of numerical experiments. Statement of the Inverse Problem {#sec:2} ================================ Below $\mathbf{x=}\left( x,y,z\right) =\left( x_{1},x_{2},x_{3}\right) \in \mathbb{R}^{3}.$ Let $A>0$ be a number. Since the domain of interest $\Omega $ is a cube in our computations, then it is convenient to set $\Omega \subset \mathbb{R}^{3}$ as$$\Omega =\left\{ \mathbf{x=}\left( x,y,z\right) :-A/2<x,y<A/2,z\in \left( 0,A\right) \right\} . \label{2.1}$$Let the number $a>0.$ We set the single point source we use as $\mathbf{x}_{0}=\left( 0,0,-a\right) $. Hence, this source is located below the domain $\Omega .$ Let $\Gamma _{0}$ be the upper boundary of $\Omega $ and $\Gamma _{1}$ be the rest of this boundary, $$\Gamma _{0}=\left\{ \mathbf{x=}\left( x,y,z\right) :-A/2<x,y<A/2,z=A\right\} ,\Gamma _{1}=\partial \Omega \diagdown \Gamma _{0}. \label{2.01}$$Thus, $\Gamma _{0}$ is the transmitted" side of $\Omega .$ Let the function $c\left( \mathbf{x}\right) $ be such that$$c\in C^{13}\left( \mathbb{R}^{3}\right) , \label{2.2}$$$$c\left( \mathbf{x}\right) \geq c_{0}=const.>0\text{ in }\overline{\Omega }, \label{2.3}$$$$c\left( \mathbf{x}\right) =1,\forall \mathbf{x}\in \mathbb{R}^{3}\diagdown \Omega . \label{2.4}$$ **Remark 2.1**. *We assume in (\[2.1\]) the* $C^{13}-$*smoothness of the function* $c\left( \mathbf{x}\right) $* since the representation (\[2.9\]) of the solution of the Cauchy problem (\[2.7\]), (\[2.8\]) works only under this assumption. Indeed, this smoothness was carefully calculated in Theorem 4.1 of the book [@Rom].* The physical meaning of the function $c\left( \mathbf{x}\right) $ is that $1/\sqrt{c\left( \mathbf{x}\right) }$ is the speed of sound. Consider the conformal Riemannian metric generated by the function $c\left( \mathbf{x}\right) ,$$$d\tau =\sqrt{c\left( \mathbf{x}\right) }\sqrt{\left( dx\right) ^{2}+\left( dy\right) ^{2}+\left( dz\right) ^{2}}. \label{2.5}$$The metric (\[2.5\]) generates geodesic lines $\Gamma \left( \mathbf{x},\mathbf{x}_{0}\right) $, $\mathbf{x}\in \mathbb{R}^{3}$. Let $\tau \left( \mathbf{x}\right) $ is the travel time along the geodesic line $\Gamma \left( \mathbf{x},\mathbf{x}_{0}\right) .$ Then [@Rom] the function $\tau \left( \mathbf{x}\right) $ is the solution of the eikonal equation$$\left\vert \nabla \tau \left( \mathbf{x}\right) \right\vert ^{2}=c\left( \mathbf{x}\right) , \label{2.6}$$with the condition $\tau \left( \mathbf{x}\right) =O\left( \left\vert \mathbf{x}-\mathbf{x}_{0}\right\vert \right) $ as $\left\vert \mathbf{x}-\mathbf{x}_{0}\right\vert \rightarrow 0.$ Furthermore,$$\tau \left( \mathbf{x}\right) =\mathop{\displaystyle \int}\limits_{\Gamma \left( \mathbf{x},\mathbf{x}_{0}\right) }\sqrt{c\left( \mathbf{y}\right) }d\sigma .$$Everywhere below we rely on the following assumption without further comments [@Rom]: **Regularity Assumption**. *For the above specific point source* $\mathbf{x}_{0},$ *geodesic lines generated by the function* $c\left( \mathbf{x}\right) $* are regular. In other words, for any points* $\mathbf{x}\in \mathbb{R}^{3}$* there exists unique geodesic line* $\Gamma \left( \mathbf{x},\mathbf{x}_{0}\right) $* connecting points* $\mathbf{x}$* and* $\mathbf{x}_{0}$*.* A sufficient condition of the regularity of geodesic lines was derived in [@Rom2]. This condition is$$\mathop{\displaystyle \sum }\limits_{i,j=1}^{3}\frac{\partial ^{2}\ln c\left( \mathbf{x}\right) }{\partial x_{i}\partial x_{j}}\xi _{i}\xi _{j}\geq 0,\text{ }\forall \mathbf{x,\xi }\in \mathbb{R}^{3}.$$ As the forward problem, we consider** **the following Cauchy problem for the acoustic equation [@Colton] of the hyperbolic type for the function $u\left( \mathbf{x},\mathbf{x}_{0},t\right) :$ $$c\left( \mathbf{x}\right) u_{tt}=\Delta u,\mathbf{x}\in \mathbb{R}^{3},t>0, \label{2.7}$$$$u\left( \mathbf{x},0\right) =0,u_{t}\left( \mathbf{x},0\right) =\delta \left( \mathbf{x}-\mathbf{x}_{0}\right) . \label{2.8}$$It was proven in Theorem 4.1 of [@Rom] that, given the above conditions (\[2.2\])-(\[2.4\]) as well as the Regularity Assumption, the problem (\[2.7\]), (\[2.8\]) has unique solution $u\left( \mathbf{x},\mathbf{x}_{0},t\right) $ which can be represented as $$u\left( \mathbf{x},t\right) =A\left( \mathbf{x}\right) \delta \left( t-\tau \left( \mathbf{x}\right) \right) +H\left( t-\tau \left( \mathbf{x}\right) \right) \widehat{u}\left( \mathbf{x},t\right) , \label{2.9}$$where $\tau \left( \mathbf{x}\right) ,A\left( \mathbf{x}\right) \in C^{12}\left( \mathbb{R}^{3}\right) $, the function $\widehat{u}\left( \mathbf{x},t\right) \in C^{2}\left( t\geq \tau \left( \mathbf{x}\right) \right) $ and $A\left( \mathbf{x}\right) >0,\forall \mathbf{x}\in \mathbb{R}^{3}.$ In (\[2.9\]) $H\left( z\right) $ is the Heaviside function,$$H\left( z\right) =\left\{ \begin{array}{c} 1,z>0, \\ 0,z<0.\end{array}\right.$$Let the number $T>0$. Denote $S_{T}=\partial \Omega \times \left( 0,T\right) $ and $\Gamma _{0,T}=\Gamma _{0}\times \left( 0,T\right) .$ Since we work with only a single location $\mathbf{x}_{0}$ of the point source, we will omit below indications of dependencies on $\mathbf{x}_{0}.$ **Coefficient Inverse Problem (CIP)**. *Let the domain* $\Omega $* be as in (\[2.1\]). Suppose that the following two functions are given:*$$u\left( \mathbf{x},t\right) \mid _{\left( \mathbf{x,}t\right) \in S_{T}}=f_{0}\left( \mathbf{x},t\right) ,\partial _{z}u\left( \mathbf{x},t\right) \mid _{\left( \mathbf{x,}t\right) \in \Gamma _{0,T}}=f_{1}\left( \mathbf{x},t\right) , \label{2.10}$$*Determine the function* $c\left( \mathbf{x}\right) $* for* $\mathbf{x}\in \Omega .$ The knowledge of the normal derivative of the function $u\left( \mathbf{x},t\right) $ at the upper boundary of the cube $\Omega $ can be justified as follows. Suppose that measurements $\varphi \left( x,y,t\right) $ of the amplitude $u\left( \mathbf{x},t\right) $ of acoustic waves are conducted on the surface $\Gamma _{1}$ as well as on the plane $\left\{ z=A\right\} .$ Using (\[2.4\]), (\[2.7\]) and (\[2.8\]), we obtain in the half space $\left\{ z>A\right\} $$$u_{tt}=\Delta u,\mathbf{x}\in \left\{ z>A\right\} ,t>0,$$$$u\left( \mathbf{x},0\right) =u_{t}\left( \mathbf{x},0\right) =0,$$$$u\left( x,y,A\right) =\varphi \left( x,y,t\right) ,\left( x,y\right) \in \mathbb{R}^{2},t>0.$$Solving this initial boundary value problem, we uniquely obtain the function $u\left( x,y,z,t\right) $ for $z>A,t>0.$ Next, we obtain $\partial _{z}u\left( x,y,A,t\right) .$ A System of Coupled Quasilinear Elliptic Equations {#sec:3} ================================================== In this section we reduce the CIP (\[2.7\])-(\[2.10\]) to the Cauchy problem for* *a system of coupled elliptic PDEs. The function $w\left( \mathbf{x},t\right) $ {#sec:3.1} ------------------------------------------- Integrate equation (\[2.7\]) twice with respect to $t$ for points $\mathbf{x}\in \Omega .$ Hence, we consider the function $p\left( \mathbf{x},t\right) ,$$$p\left( \mathbf{x},t\right) =\mathop{\displaystyle \int}\limits_{0}^{t}dy\mathop{\displaystyle \int}\limits_{0}^{y}u\left( \mathbf{x},s\right) ds,\text{ }\mathbf{x}\in \Omega . \label{3.1}$$Hence, $p_{tt}\left( \mathbf{x},t\right) =u\left( \mathbf{x},t\right) $ for $\mathbf{x}\in \Omega .$ Since $\mathbf{x}_{0}\notin \overline{\Omega },$ then $\delta \left( \mathbf{x}-\mathbf{x}_{0}\right) =0$ for $\mathbf{x}\in \Omega .$ Hence, by (\[2.8\]) $$\mathop{\displaystyle \int}\limits_{0}^{t}dy\mathop{\displaystyle \int}\limits_{0}^{y}u_{tt}\left( \mathbf{x},s\right) ds=\mathop{\displaystyle \int}\limits_{0}^{t}\left( u_{t}\left( \mathbf{x},y\right) -u_{t}\left( \mathbf{x},0\right) \right) dy=u\left( \mathbf{x},t\right) =p_{tt}\left( \mathbf{x},t\right) ,\text{ }\mathbf{x}\in \Omega . \label{3.2}$$Next, by (\[2.7\]) and (\[3.1\]) $$\mathop{\displaystyle \int}\limits_{0}^{t}dy\mathop{\displaystyle \int}\limits_{0}^{y}u_{tt}\left( \mathbf{x},s\right) ds=\mathop{\displaystyle \int}\limits_{0}^{t}dy\mathop{\displaystyle \int}\limits_{0}^{y}\Delta u\left( \mathbf{x},s\right) ds=\Delta p\left( \mathbf{x},t\right) ,\text{ }\mathbf{x}\in \Omega . \label{3.3}$$Comparing (\[3.2\]) and (\[3.3\]), we obtain $$p_{tt}\left( \mathbf{x},t\right) =\Delta p\left( \mathbf{x},t\right) \text{ for }\left( \mathbf{x},t\right) \in \Omega \times \left( 0,T\right) . \label{3.4}$$ Next, by (\[2.9\]) and (\[3.1\])$$p\left( \mathbf{x},t\right) =A\left( \mathbf{x}\right) \left( t-\tau \left( \mathbf{x}\right) \right) H\left( t-\tau \left( \mathbf{x}\right) \right) +O\left( \left( t-\tau \left( \mathbf{x}\right) \right) ^{2}\right) H\left( t-\tau ^{0}\left( \mathbf{x}\right) \right) , \label{3.5}$$where$$\left\vert O\left( \left( t-\tau \left( \mathbf{x}\right) \right) ^{2}\right) \right\vert \leq B\left( t-\tau \left( \mathbf{x}\right) \right) ^{2}\text{ as }t\rightarrow \tau ^{+}\left( \mathbf{x}\right) , \label{3.6}$$$$\left\vert \partial _{t}O\left( \left( t-\tau \left( \mathbf{x}\right) \right) ^{2}\right) \right\vert \leq B\left( t-\tau \left( \mathbf{x}\right) \right) \text{ as }t\rightarrow \tau ^{+}\left( \mathbf{x}\right) \label{3.7}$$with a certain constant $B>0$ independent on $\left( \mathbf{x},t\right) \in \Omega \times \left( 0,T\right) .$ Consider the function $w\left( \mathbf{x},t\right) $ defined as$$w\left( \mathbf{x},t\right) =p\left( \mathbf{x},t+\tau \left( \mathbf{x}\right) \right) ,\text{ for }\left( \mathbf{x},t\right) \in \Omega \times \left( 0,T\right) . \label{3.8}$$Then it follows from the above that $w\in C^{2}\left( \overline{\Omega }\times \left[ 0,T\right] \right) $ and by (\[3.5\])-(\[3.7\])$$w\left( \mathbf{x},0\right) =0, \label{3.9}$$$$w_{t}\left( \mathbf{x},0\right) =A\left( \mathbf{x}\right) >0. \label{3.10}$$Substituting (\[3.8\]) in (\[2.7\]) and using (\[2.6\]), we obtain$$\Delta w-2\mathop{\displaystyle \sum }\limits_{i=1}^{3}w_{x_{i}t}\tau _{x_{i}}-w_{t}\Delta \tau =0,\mathbf{x}\in \Omega ,t\in \left( 0,T_{1}\right) , \label{3.11}$$$$T_{1}=T-\max_{\overline{\Omega }}\tau \left( \mathbf{x}\right) . \label{3.110}$$ Denote $\widetilde{f}_{0}\left( \mathbf{x},t\right) =f_{0}\left( \mathbf{x},t+\tau \left( \mathbf{x}\right) \right) $ and $\widetilde{f}_{1}\left( \mathbf{x},t\right) =f_{1}\left( \mathbf{x},t+\tau \left( \mathbf{x}\right) \right) .$ Then by (\[2.10\])$$w\left( \mathbf{x},t\right) \mid _{\left( \mathbf{x,}t\right) \in S_{T_{1}}}=\widetilde{f}_{0}\left( \mathbf{x},t\right) ,\text{ }\partial _{z}w\left( \mathbf{x},t\right) \mid _{\left( \mathbf{x,}t\right) \in \Gamma _{0,T_{1}}}=\widetilde{f}_{1}\left( \mathbf{x},t\right) . \label{3.12}$$ Thus, our goal below is to construct a numerical method, which would approximately find the functions $w\left( \mathbf{x},t\right) ,\tau \left( \mathbf{x}\right) $ for $\mathbf{x}\in \Omega ,t\in \left( 0,T_{1}\right) $ from conditions (\[3.9\])-(\[3.12\]). Suppose that these two functions are approximated. Then the corresponding approximation for the target coefficient $c\left( \mathbf{x}\right) $ can be easily found via the backwards calculation, $$c\left( \mathbf{x}\right) =\left\vert \nabla \tau \left( \mathbf{x}\right) \right\vert ^{2}. \label{3.120}$$ The system of coupled quasilinear elliptic PDEs {#sec:3.2} ----------------------------------------------- **Lemma 3.1**. *Consider the set of functions* $$\left\{ t,t^{2},...,t^{n},...\right\} =\left\{ t^{n}\right\} _{n=1}^{\infty }. \label{3.13}$$* Then this set is complete in* $L_{2}\left( 0,T_{1}\right) .$ **Proof**. Let a function $f\left( t\right) \in L_{2}\left( 0,T_{1}\right) $ be such that$$\mathop{\displaystyle \int}\limits_{0}^{T_{1}}f\left( t\right) t^{n}dt=0,n=1,2,...$$Consider the function $\widetilde{f}\left( t\right) =f\left( t\right) t.$ Then $$\mathop{\displaystyle \int}\limits_{0}^{T_{1}}\widetilde{f}\left( t\right) t^{m}dt=0,m=0,1,2,... \label{3.14}$$It is well known that (\[3.14\]) implies that $\widetilde{f}\left( t\right) \equiv 0.$ $\square $ Orthonormalize the set (\[3.13\]) using the Gram-Schmidt orthonormalization procedure. Then Lemma 3.1 implies that we obtain a basis $\left\{ P_{n}\left( t\right) \right\} _{n=1}^{\infty }$ of orthonormal polynomials in $L_{2}\left( 0,T_{1}\right) $ such that $$P_{n}\left( 0\right) =0,\forall n=1,2,... \label{3.15}$$By (\[3.15\]) this is not a set of standard orthonormal polynomials. Let the integer $N>1.$ Approximate the function $w\left( \mathbf{x},t\right) $ as$$w\left( \mathbf{x},t\right) =\mathop{\displaystyle \sum }\limits_{n=1}^{N}w_{n}\left( \mathbf{x}\right) P_{n}\left( t\right) . \label{3.16}$$Here and below we use $=$" instead of $\approx "$ for convenience. Substitute (\[3.16\]) in the left hand side of (\[3.11\]) and assume that the resulting left hand side equals zero. We obtain for $\mathbf{x}\in \Omega $ $$\mathop{\displaystyle \sum }\limits_{n=1}^{N}\Delta w_{n}\left( \mathbf{x}\right) P_{n}\left( t\right) -2\mathop{\displaystyle \sum }\limits_{i=1}^{3}\tau _{x_{i}}\left( \mathbf{x}\right) \mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( t\right) \partial _{x_{i}}w_{n}\left( \mathbf{x}\right) -\Delta \tau \left( \mathbf{x}\right) \mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( t\right) w_{n}\left( \mathbf{x}\right) =0. \label{3.17}$$By (\[3.10\]) and (\[3.16\]) we can assume that $$\mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( 0\right) w_{n}\left( \mathbf{x}\right) =A\left( \mathbf{x}\right) >0,\text{ }\forall \mathbf{x}\in \overline{\Omega }. \label{3.18}$$Set in (\[2.19\]) $t=0.$ Hence, we obtain the first elliptic equation,$$\Delta \tau \left( \mathbf{x}\right) +2\left[ \mathop{\displaystyle \sum }\limits_{i=1}^{3}\tau _{x_{i}}\mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( 0\right) \partial _{x_{i}}w_{n}\left( \mathbf{x}\right) \right] \left[ \mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( 0\right) w_{n}\left( \mathbf{x}\right) \right] ^{-1}=0,\text{ }\mathbf{x}\in \Omega . \label{3.180}$$We rewrite this equation as$$\Delta \tau =F_{1}\left( \nabla \tau ,\nabla \widetilde{W},\widetilde{W}\right) ,\mathbf{x}\in \Omega , \label{3.19}$$where $\widetilde{W}\left( x\right) =\left( w_{1}\left( \mathbf{x}\right) ,...,w_{N}\left( \mathbf{x}\right) \right) ^{T}.$ Next, for $n=1,...,N$ multiply both sides of (\[3.17\]) by $P_{n}\left( t\right) $ and integrate with respect to $t\in \left( 0,T_{1}\right) .$ Replace in the resulting equation $\Delta \tau $ with the right hand side of (\[3.19\]). We obtain $$\Delta \widetilde{W}=F_{2}\left( \nabla \tau ,\nabla \widetilde{W},\widetilde{W}\right) ,\mathbf{x}\in \Omega . \label{3.20}$$ Consider the $\left( N+1\right) -$dimensional vector function $$W\left( \mathbf{x}\right) =\left( \tau \left( \mathbf{x}\right) ,\widetilde{W}\left( \mathbf{x}\right) \right) ^{T}. \label{3.200}$$ Thus, (\[3.12\]), (\[3.19\]) and (\[3.20\]) lead to the following Cauchy problem for a system of coupled quasilinear elliptic equations$$\Delta W+F\left( \nabla W,W\right) =0,\mathbf{x}\in \Omega , \label{3.22}$$$$W\mid _{\partial \Omega }=q^{0}\left( \mathbf{x}\right) ,\partial _{z}W\mid _{\Gamma _{0}}=q^{1}\left( \mathbf{x}\right) , \label{3.23}$$$$q^{0}\left( \mathbf{x}\right) =\left( \tau \left( \mathbf{x}\right) ,q_{1}^{0}\left( \mathbf{x}\right) ,...,q_{N}^{0}\left( \mathbf{x}\right) \right) ^{T},\text{ }q_{n}^{0}\left( \mathbf{x}\right) =\mathop{\displaystyle \int}\limits_{0}^{T_{1}}\widetilde{f}_{0}\left( \mathbf{x},t\right) P_{n}\left( t\right) dt,\text{ }\mathbf{x}\in \partial \Omega , \label{3.24}$$$$q^{1}\left( \mathbf{x}\right) =\left( \partial _{z}\tau \left( \mathbf{x}\right) ,q_{1}^{1}\left( \mathbf{x}\right) ,...,q_{N}^{1}\left( \mathbf{x}\right) \right) ^{T},q_{n}^{1}\left( \mathbf{x}\right) =\mathop{\displaystyle \int}\limits_{0}^{T_{1}}\widetilde{f}_{1}\left( \mathbf{x},t\right) P_{n}\left( t\right) dt,\text{ }\mathbf{x}\in \Gamma _{0}. \label{3.25}$$ In (\[3.24\]) and (\[3.25\]) $n=1,...,N.$ In (\[3.22\]) the $\left( N+1\right) -$dimensional vector function $F\in C^{1}\left( \mathbb{R}^{3N+5}\right) $. Thus, we have obtained the system (\[3.22\]) of coupled quasilinear elliptic PDEs with the Cauchy data (\[3.23\])-(\[3.25\]). Unknowns in this problem are the function $\tau \left( \mathbf{x}\right) $ and Fourier coefficients $w_{n}\left( \mathbf{x}\right) $ of the function $w\left( \mathbf{x},t\right) $ in (\[3.16\]). Therefore, we solve below the problem (\[3.22\])-(\[3.25\]) of finding the $\left( N+1\right) -$dimensional vector function $W\in C^{2}\left( \overline{\Omega }\right) .$ In fact, however, we find below $W\in H^{3}\left( \Omega \right) .$ **Remarks 3.1**: 1. *The number* $N$* in (\[3.16\]) should be chosen computationally, see section 6. Since (\[3.16\]) is an approximation of the function* $w\left( \mathbf{x},t\right) ,$* then (\[3.17\])-([3.23]{}) are also understood in the approximate sense. Thus, (\[3.16\])-(\[3.23\]) form our **approximate mathematical model**. Since the original CIP is a very challenging one, it is hard to anticipate that it would be effectively solved numerically without such approximations. Since we develop a numerical method here, then it is fine to introduce an approximate mathematical model.* 2. *We cannot prove convergence of our resulting solutions to the correct one as* $N\rightarrow \infty .$* The underlying reason of this is the ill-posed nature of our CIP combined with the nonlinearity. On the other hand, truncated Fourier series are considered quite often in the field of Inverse and Ill-Posed Problems, whereas convergencies of resulting solutions at* $N\rightarrow \infty $* are proven only very rarely. Nevertheless, corresponding approximate mathematical models usually provide quite good numerical results: we refer, e.g. to [Kab1,Kab2,KlibKol3,KEIT]{} for corresponding publications.* Globally Strictly Convex Tikhonov-like Functional {#sec:4} ================================================= The functional {#sec:4.1} -------------- All Banach spaces considered below are spaces of real valued functions. If we say below that a vector function belongs to a certain Banach space, then this means that all its components belong to this space, and the norm of this function in that space is defined as the square root of the sum of squares of norms of its components. To arrange a certain projection operator for the gradient projection method below, the best way is to have zero Cauchy data. Hence, we assume that there exists an $\left( N+1\right) -$dimensional vector function $G=\left( g_{0}\left( \mathbf{x}\right) ,...,g_{N}\left( \mathbf{x}\right) \right) ^{T}\in H^{3}\left( \Omega \right) $ satisfying boundary conditions ([3.23]{}), i.e. such that $$G\mid _{\partial \Omega }=q^{0}\left( \mathbf{x}\right) ,\partial _{z}G\mid _{\Gamma _{0}}=q^{1}\left( \mathbf{x}\right) . \label{4.1}$$Let $$W-G=Q\in H^{3}\left( \Omega \right) ,Q\left( \mathbf{x}\right) =\left( q_{0}\left( \mathbf{x}\right) ,...,q_{N}\left( \mathbf{x}\right) \right) ^{T}. \label{4.100}$$ Then (\[3.22\]), (\[3.23\]) and (\[4.1\]) imply that $$\Delta Q+\Delta G+F\left( Q+G,\nabla \left( Q+G\right) \right) =0,\mathbf{x}\in \Omega , \label{4.2}$$$$Q\mid _{\partial \Omega }=0,\partial _{z}Q\mid _{\Gamma _{0}}=0. \label{4.3}$$ Let $H_{0}^{3}\left( \Omega \right) $ be the subspace of the space $H^{3}\left( \Omega \right) $ defined as$$H_{0}^{3}\left( \Omega \right) =\left\{ v\in H^{3}\left( \Omega \right) :v\mid _{\partial \Omega }=0,\partial _{z}v\mid _{\Gamma _{0}}=0\right\} .$$Choose an arbitrary number $R>0$ and also choose another number $m\in \left( 0,R\right) ,$ which is independent on $R$. Consider the set $B\left( m,R\right) $ of $\left( N+1\right) -$dimensional vector functions $Z\left( \mathbf{x}\right) =\left( z_{0}\left( \mathbf{x}\right) ,...,z_{N}\left( \mathbf{x}\right) \right) ^{T}$ such that$$B\left( m,R\right) =\left\{ \begin{array}{c} Z\in H_{0}^{3}\left( \Omega \right) ,\text{ }\left\Vert Z\right\Vert _{H^{3}\left( \Omega \right) }<R, \\ \mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( 0\right) \left( z_{n}\left( \mathbf{x}\right) +g_{n}\left( \mathbf{x}\right) \right) >m,\forall \mathbf{x}\in \overline{\Omega }.\end{array}\right. \label{4.4}$$ The second condition (\[4.3\]) is generated by (\[3.18\]). By embedding theorem  $H^{3}\left( \Omega \right) \subset C^{1}\left( \overline{\Omega }\right) .$ This implies that $\overline{B\left( m,R\right) }\subset C^{1}\left( \overline{\Omega }\right) $ and also that there exist numbers $D_{1}\left( R\right) >0$ and $D_{2}\left( G\right) >0$ depending only on listed parameters such that $$\left\Vert Z\right\Vert _{C^{1}\left( \overline{\Omega }\right) }\leq D_{1}\left( R\right) ,\text{ }\forall Z\in \overline{B\left( m,R\right) }, \label{4.40}$$$$\left\Vert G\right\Vert _{C^{1}\left( \overline{\Omega }\right) }\leq D_{2}\left( G\right) . \label{4.400}$$Temporary replace the vector functions $Q\left( \mathbf{x}\right) =$ $\left( q_{0}\left( \mathbf{x}\right) ,...,q_{N}\left( \mathbf{x}\right) \right) ^{T} $ and $\nabla Q\left( \mathbf{x}\right) =\left( \nabla q_{0}\left( \mathbf{x}\right) ,...,\nabla q_{N}\left( \mathbf{x}\right) \right) ^{T}$ with the vector of real variables $\left( y_{0},y_{1},...,y_{4N+3}\right) ^{T}=y\in \mathbb{R}^{4N+4}.$ Consider the set $Y\subset \mathbb{R}^{4N+4}$, $$Y=\left\{ y\in \mathbb{R}^{4N+4}:\mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( 0\right) \left( y_{n}+g_{n-1}\left( \mathbf{x}\right) \right) >m,\forall \mathbf{x}\in \overline{\Omega }\right\} .$$Obviously $Y$ is an open set in $\mathbb{R}^{4N+4}$. Denote $p_{1}=\left( y_{0},...,y_{N}\right) ,p_{2}=\left( y_{0,1},y_{0,2},y_{0,3},y_{1,1},...,y_{N,3}\right) .$ Then $y=\left( p_{1},p_{2}\right) ^{T}\in \mathbb{R}^{4N+4}.$ It follows from (\[3.180\])-(\[3.22\]) that, as the function of $y,$ $$F\left( p_{1}+G\left( \mathbf{x}\right) ,p_{2}+\nabla G\left( \mathbf{x}\right) \right) \in C^{2}\left( \overline{Y}\right) ,\forall \mathbf{x}\in \overline{\Omega }. \label{4.42}$$ **Lemma 4.1**. *The set* $B\left( m,R\right) $* is convex.* **Proof**. Let the number $\alpha \in \left( 0,1\right) $ and vector functions $Z,Y\in B\left( m,R\right) .$ Consider the function $\alpha Z+\left( 1-\alpha \right) Y.$ Then $$\left\Vert \alpha Z+\left( 1-\alpha \right) Y\right\Vert _{H^{3}\left( \Omega \right) }\leq \alpha \left\Vert Z\right\Vert _{H^{3}\left( \Omega \right) }+\left( 1-\alpha \right) \left\Vert Y\right\Vert _{H^{3}\left( \Omega \right) }<\alpha R+\left( 1-\alpha \right) R=R.$$Next, let $Z\left( \mathbf{x}\right) =\left( \tau _{1}\left( \mathbf{x}\right) ,z_{1}\left( \mathbf{x}\right) ,...,z_{N}\left( \mathbf{x}\right) \right) ^{T},Y\left( \mathbf{x}\right) =\left( \tau _{2}\left( \mathbf{x}\right) ,y_{1}\left( \mathbf{x}\right) ,...,y_{N}\left( \mathbf{x}\right) \right) ^{T}.$ Then $$\alpha \mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( 0\right) z_{n}\left( \mathbf{x}\right) +\left( 1-\alpha \right) \mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( 0\right) y_{n}\left( \mathbf{x}\right) >\alpha m+\left( 1-\alpha \right) m=m.\text{ \ }\square$$ Our weighted Tikhonov-like cost functional is$$J_{\lambda ,\beta }\left( Q+G\right) =e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta Q+\Delta G+F\left( \nabla \left( Q+G\right) ,Q+G\right) \right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x} \label{4.5}$$$$+\beta \left\Vert Q+G\right\Vert _{H^{3}\left( \Omega \right) }^{2}.$$In (\[4.5\]) the numbers $\lambda \geq 1,b>0,\beta \in \left( 0,1\right) .$ Here, $\lambda $ is the parameter of our CWF $e^{2\lambda \left( z+b\right) ^{2}}$ and $\beta $ is the regularization parameter. The multiplier $e^{-2\lambda b^{2}}$ is introduced to balance two terms in the right hand side of (\[4.5\]). Indeed, by (\[2.1\]) $$\min_{\mathbf{x}\in \overline{\Omega }}\left( e^{-2\lambda b^{2}}e^{2\lambda \left( z+b\right) ^{2}}\right) =1. \label{4.50}$$ **Minimization Problem.** *Minimize the functional* $J_{\lambda ,\beta }\left( Q\right) $* in (\[4.5\]) on the set* $B\left( m,R\right) $* defined in (\[4.4\]).* Theorems {#sec:4.2} -------- **Theorem 4.1**. *There exists a sufficiently large number* $\lambda _{0}=\lambda _{0}\left( \Omega ,b\right) \geq 1$* and a constant* $C_{1}=C_{1}\left( \Omega ,b\right) >0,$* both depending only on* $\Omega $* and* $b,$* such that for all* $\lambda \geq \lambda _{0}$* and for all functions* $u\in H^{2}\left( \Omega \right) $* such that* $u\mid _{\partial \Omega }=u_{z}\mid _{\Gamma _{0}}=0$* the following Carleman estimate holds*$$\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}\geq \frac{C_{1}}{\lambda }\mathop{\displaystyle \sum }\limits_{i,j=1}^{3}\mathop{\displaystyle \int}\limits_{\Omega }u_{x_{i}x_{j}}^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x} \label{4.6}$$$$+C_{1}\lambda \mathop{\displaystyle \int}\limits_{\Omega }\left( \left( \nabla u\right) ^{2}+\lambda ^{2}u^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x.}$$ Below $C_{2}=C_{2}\left( F,\left\Vert G\right\Vert _{H^{3}\left( \Omega \right) },m,R,\Omega ,b\right) >0$ denotes different constants depending only on listed parameters. **Theorem 4.2** (global strict convexity). *For all* $Q\in B\left( m,2R\right) $*,* $\lambda ,\beta >0$* there exists Frechét derivative* $J_{\lambda ,\beta }^{\prime }\left( Q+G\right) \in H_{0}^{3}\left( \Omega \right) .$* Let* $\lambda _{0}$* be the number of Theorem 4.1. There exists a number* $\lambda _{1}=\lambda _{1}\left( F,\left\Vert G\right\Vert _{H^{3}\left( \Omega \right) },m,R,\Omega ,b\right) \geq \lambda _{0}$* depending only on listed parameters such that for any* $\lambda \geq \lambda _{1}$* and any* $\beta >0$* the functional* $J_{\lambda ,\beta }\left( Q\right) $* is strictly convex on* $B\left( m,R\right) ,$* i.e. the following estimate holds for all* $Q_{1},Q_{2}\in B\left( m,R\right) $$$J_{\lambda ,\beta }\left( Q_{2}+G\right) -J_{\lambda ,\beta }\left( Q_{1}+G\right) -J_{\lambda ,\beta }^{\prime }\left( Q_{1}+G\right) \left( Q_{2}-Q_{1}\right)$$$$\geq \frac{C_{2}}{\lambda }\mathop{\displaystyle \sum }\limits_{i,j=1}^{3}\left\Vert \left( Q_{2}-Q_{1}\right) _{x_{i}x_{j}}\right\Vert _{L_{2}\left( \Omega \right) }^{2}+C_{2}\lambda \left\Vert Q_{2}-Q_{1}\right\Vert _{H^{1}\left( \Omega \right) }^{2}+\beta \left\Vert Q_{2}-Q_{1}\right\Vert _{H^{3}\left( \Omega \right) }^{2}. \label{4.7}$$ **Theorem 4.3**. *The* *Frechét derivative* $J_{\lambda ,\beta }^{\prime }\left( Q+G\right) \in H_{0}^{3}\left( \Omega \right) $ *of the functional* $J_{\lambda ,\beta }\left( Q\right) $*satisfies the Lipschitz continuity condition in* $B\left( m,2R\right) $* for all* $\lambda ,\beta >0.$* In other words, there exists a number* $L=L\left( F,\left\Vert G\right\Vert _{H^{3}\left( \Omega \right) },m,R,\Omega ,b,\lambda ,\beta \right) $* depending only on listed parameters such that* $$\left\Vert J_{\lambda ,\beta }^{\prime }\left( Q_{2}+G\right) -J_{\lambda ,\beta }^{\prime }\left( Q_{1}+G\right) \right\Vert _{H^{3}\left( \Omega \right) }\leq L\left\Vert Q_{2}-Q_{1}\right\Vert _{H^{3}\left( \Omega \right) },\text{ }\forall Q_{1},Q_{2}\in B\left( m,2R\right) .$$ **Theorem 4.4**. *For each pair* $\lambda \geq $* *$\lambda _{1},\beta >0$* there exists unique minimizer* $Q_{\min ,\lambda ,\beta }\in \overline{B\left( m,R\right) }$* of the functional* $J_{\lambda ,\beta }\left( Q\right) $* on the set* $\overline{B\left( m,R\right) }.$* Furthermore,* $$J_{\lambda ,\beta }^{\prime }\left( Q_{\min ,\lambda ,\beta }+G\right) \left( Q_{\min ,\lambda ,\beta }-p\right) \leq 0,\text{ }\forall p\in H_{0}^{3}\left( \Omega \right) . \label{4.8}$$ We now arrange the gradient projection method of the minimization of the functional $J_{\lambda ,\beta }\left( Q+G\right) $ on the set $\overline{B\left( m,R\right) }.$ Let the number $\gamma \in \left( 0,1\right) .$ Let $P_{\overline{B}}:H_{0}^{3}\left( \Omega \right) \rightarrow \overline{B\left( m,R\right) }$  be the projection operator of the space $H_{0}^{3}$ on the set $\overline{B\left( m,R\right) }.$ Let $Q_{0}\in B\left( m,R\right) $ be an arbitrary point of this set. The gradient projection method amounts to the following sequence:$$Q_{n}=P_{\overline{B}}\left( Q_{n-1}-\gamma J_{\lambda ,\beta }^{\prime }\left( Q_{n-1}+G\right) \right) ,n=1,2,... \label{4.9}$$ **Theorem 4.5**. *Let* $\lambda _{1}$* and* $\beta $*be parameters of Theorem 4.2. Choose a number* $\lambda \geq \lambda _{1}.$* Let* $Q_{\min ,\lambda ,\beta }\in \overline{B\left( m,R\right) }$* be the unique minimizer of the functional* $J_{\lambda ,\beta }\left( Q\right) $* on the set* $\overline{B\left( m,R\right) }$* (Theorem 4.3). Then there exists a sufficiently small number* $\gamma _{0}=\gamma _{0}\left( F,\left\Vert G\right\Vert _{H^{3}\left( \Omega \right) },m,R,\Omega ,b,\beta \right) \in \left( 0,1\right) $*depending only on listed parameters such that the sequence (\[4.9\]) converges to* $Q_{\min ,\lambda ,\beta }$* in the space* $H^{3}\left( \Omega \right) .$* More precisely, there exists a number* $\theta =\theta \left( F,\left\Vert G\right\Vert _{H^{3}\left( \Omega \right) },m,R,\Omega ,b,\beta \right) \in \left( 0,1\right) $* such that the following estimate holds*$$\left\Vert Q_{n}-Q_{\min ,\lambda ,\beta }\right\Vert _{H^{3}\left( \Omega \right) }\leq \theta ^{n}\left\Vert Q_{\min ,\lambda ,\beta }-Q_{0}\right\Vert _{H^{3}\left( \Omega \right) }. \label{4.90}$$ Following the Tikhonov concept for ill-posed problems [@BK; @T], we now assume the existence of the exact solution $Q^{\ast }\left( \mathbf{x}\right) =\left( q_{0}^{\ast }\left( \mathbf{x}\right) ,...,q_{N}^{\ast }\left( \mathbf{x}\right) \right) ^{T}\in B\left( m,R\right) $ of the problem (\[4.2\]), (\[4.3\]) with the noiseless data $G^{\ast }\left( \mathbf{x}\right) =\left( g_{0}^{\ast }\left( \mathbf{x}\right) ,...,g_{N}^{\ast }\left( \mathbf{x}\right) \right) ^{T}\in H^{3}\left( \Omega \right) .$ In particular, this means that $$\mathop{\displaystyle \sum }\limits_{n=1}^{N}P_{n}^{\prime }\left( 0\right) \left( q_{n}^{\ast }\left( \mathbf{x}\right) +g_{n}^{\ast }\left( \mathbf{x}\right) \right) >m,\forall \mathbf{x}\in \overline{\Omega }.$$Also, let the number $\delta \in \left( 0,1\right) $ be the level of the error in the data $G,$ i.e.$$\left\Vert G-G^{\ast }\right\Vert _{H^{3}\left( \Omega \right) }<\delta . \label{4.10}$$Since $\delta \in \left( 0,1\right) ,$ then (\[4.10\]) implies that we can regard in Theorem 4.6 that constants $\lambda _{1},C_{2},\gamma _{0},\theta $ introduced above depend on $\left\Vert G^{\ast }\right\Vert _{H^{3}\left( \Omega \right) }$ rather than on $\left\Vert G\right\Vert _{H^{3}\left( \Omega \right) }.$ We are doing so both in the formulation and in the proof of Theorem 4.6. **Theorem 4.6** (error estimates and convergence).** ***Let* $\lambda _{1}=\lambda _{1}\left( F,\left\Vert G\right\Vert _{H^{3}\left( \Omega \right) },m,R,\Omega ,b\right) $* be the number of Theorem 4.2. Define the number* $\eta $* as* $\eta =\left[ 4\left( A+b\right) ^{2}\right] ^{-1}.$* Choose a sufficiently small number* $\delta _{0}\in \left( 0,1\right) $* such that* $\ln \delta _{0}^{-\eta }\geq \lambda _{1}.$* Let in (\[4.10\])* $\delta \in \left( 0,\delta _{0}\right) $*. Choose* $\lambda =\lambda \left( \delta \right) =\ln \delta ^{-\eta }>\lambda _{1}$ *implying that* $\exp \left[ 2\lambda \left( \delta \right) \left( A+b\right) ^{2}\right] =1/\sqrt{\delta }.$ *Let* $Q_{\min ,\lambda ,\beta }\in $* *$\overline{B\left( m,R\right) }$* be the unique minimizer of the functional* $J_{\lambda ,\beta }\left( Q\right) $* on the set* $\overline{B\left( m,R\right) }$*, the existence of which is guaranteed by Theorem 4.4. Let* $\left\{ Q_{n}\right\} _{n=0}^{\infty }$* *$\subset \overline{B\left( m,R\right) }$* be the sequence of the gradient projection method (\[4.9\]) with an arbitrary starting point* $Q_{0}\in \overline{B\left( m,R\right) }.$* Then the following estimates hold for all* $\beta \in \left( 0,1\right) $$$\left\Vert Q^{\ast }-Q_{\min ,\lambda ,\beta }\right\Vert _{H^{1}\left( \Omega \right) }\leq C_{2}\left( \delta ^{\eta /2+1/4}+\sqrt{\beta }\delta ^{\eta /2}\right) , \label{1}$$$$\left\Vert Q^{\ast }-Q_{\min ,\lambda ,\beta }\right\Vert _{H^{2}\left( \Omega \right) }\leq C_{2}\left( \delta ^{1/4}+\sqrt{\beta }\right) \sqrt{\ln \delta ^{-\eta }}, \label{2}$$$$\left\Vert Q^{\ast }-Q_{n}\right\Vert _{H^{1}\left( \Omega \right) }\leq C_{2}\left( \delta ^{\eta /2+1/4}+\sqrt{\beta }\delta ^{\eta /2}\right) +\theta ^{n}\left\Vert Q_{\min ,\lambda ,\beta }-Q_{0}\right\Vert _{H^{3}\left( \Omega \right) }, \label{3}$$$$\left\Vert Q^{\ast }-Q_{n}\right\Vert _{H^{2}\left( \Omega \right) }\leq C_{2}\left( \delta ^{1/4}+\sqrt{\beta }\right) \sqrt{\ln \delta ^{-\eta }}+\theta ^{n}\left\Vert Q_{\min ,\lambda ,\beta }-Q_{0}\right\Vert _{H^{3}\left( \Omega \right) }, \label{4}$$$$\left\Vert c^{\ast }-c_{n}\right\Vert _{L_{2}\left( \Omega \right) }\leq C_{2}\left( \delta ^{1/4}+\sqrt{\beta }\right) \sqrt{\ln \delta ^{-\eta }}+\theta ^{n}\left\Vert Q_{\min ,\lambda ,\beta }-Q_{0}\right\Vert _{H^{3}\left( \Omega \right) }, \label{5}$$*In particular, if the regularization parameter* $\beta =\sqrt{\delta }$*, as required by the regularization theory [@T], then estimates (\[1\])-(\[4\]) become* $$\left\Vert Q^{\ast }-Q_{\min ,\lambda ,\beta }\right\Vert _{H^{1}\left( \Omega \right) }\leq C_{2}\delta ^{\eta /2+1/4}$$$$\left\Vert Q^{\ast }-Q_{\min ,\lambda ,\beta }\right\Vert _{H^{2}\left( \Omega \right) }\leq C_{2}\delta ^{1/4}\sqrt{\ln \delta ^{-\eta }},$$$$\left\Vert Q^{\ast }-Q_{n}\right\Vert _{H^{1}\left( \Omega \right) }\leq C_{2}\delta ^{\eta /2+1/4}+\theta ^{n}\left\Vert Q_{\min ,\lambda ,\sqrt{\delta }}-Q_{0}\right\Vert _{H^{3}\left( \Omega \right) },$$$$\left\Vert Q^{\ast }-Q_{n}\right\Vert _{H^{2}\left( \Omega \right) }\leq C_{2}\delta ^{1/4}\sqrt{\ln \delta ^{-\eta }}+\theta ^{n}\left\Vert Q_{\min ,\lambda ,\sqrt{\delta }}-Q_{0}\right\Vert _{H^{3}\left( \Omega \right) },$$$$\left\Vert c^{\ast }-c_{n}\right\Vert _{L_{2}\left( \Omega \right) }\leq C_{2}\delta ^{1/4}\sqrt{\ln \delta ^{-\eta }}+\theta ^{n}\left\Vert Q_{\min ,\lambda ,\sqrt{\delta }}-Q_{0}\right\Vert _{H^{3}\left( \Omega \right) }.$$*Here* $c_{n}\left( \mathbf{x}\right) $ *is defined via ([4.100]{}) with* $Q=Q_{n}$*, next (\[3.200\]), and next (\[3.120\]). Further,* $c^{\ast }\left( \mathbf{x}\right) $ *is defined as the exact target coefficient, which corresponds to the noiseless data* $G^{\ast }$* with* $Q^{\ast }$* and* $W^{\ast }=Q^{\ast }+G^{\ast }$*in (\[4.100\]), and next similarly as for* $c_{n}\left( \mathbf{x}\right) . $ **Remarks 4.1:** 1. *The CWF of [@Klib97] depends on two large parameters, unlike the one in (\[4.5\]). This makes the CWF of [@Klib97] to be quite difficult for calculations, see a similar remark on page 1579 of [@Baud] for the case of a CWF for the hyperbolic operator* $\partial _{t}^{2}-\Delta .$* Thus, we establish in Theorem 4.1 a new Carleman estimate for the CWF* $e^{2\lambda \left( z+b\right) ^{2}}$*.* 2. *It follows from (\[3.180\])-(\[3.22\]) and (\[4.4\]) that condition (\[4.42\]) holds true.* 3. *Analogs of Theorems 4.3-4.6 were not proven in [@Klib97]. On the other hand, Theorems 4.4-4.6 are computationally oriented.* 4. *The regularization parameter* $\beta $* was not used in [@Klib97]. On the other hand, the presence of the regularization term* $\beta \left\Vert Q+G\right\Vert _{H^{3}\left( \Omega \right) }^{2}$*in the functional* $J_{\lambda ,\beta }\left( Q+G\right) $* is important since this term ensures that in the gradient projection method (\[4.9\]) all functions* $Q_{n}\in H_{0}^{3}\left( \Omega \right) .$ *Since* $H^{3}\left( \Omega \right) \subset C^{1}\left( \overline{\Omega }\right) ,$* and since we use estimates of* $C^{1}\left( \overline{\Omega }\right) -$*norms of some functions in the proof of Theorem 4.2, then we indeed need* $Q_{n}\in H_{0}^{3}\left( \Omega \right) $*. However, if using the functional of [@Klib97], then there is no guarantee that* $Q_{n}\in H_{0}^{3}\left( \Omega \right) $.* On the other hand, final estimates (\[1\])-(\[4\]) are valid for all values* $\beta \in \left( 0,1\right) .$ Proofs {#sec:5} ====== We now refer to the publication [@Bak] where some theorems of convex analysis are established. The proof of Theorem 4.3 is very similar with the proof of Theorem 3.1 of [@Bak]. As soon as Theorems 4.2 and 4.3 are proven, the proof of Theorem 4.4 is quite similar with the proof of Lemma 2.1 of [@Bak]. Next, as soon as Theorems  4.2 and 4.4 are proven, the proof of Theorem 4.5 is again quite similar with the proof of Theorem 2.1 of [@Bak]. Hence, we prove here only Theorems 4.1, 4.2 and 4.6. Proof of Theorem 4.1 {#sec:5.1} -------------------- In this proof, the function $u\in C^{3}\left( \overline{\Omega }\right) .$ The case $u\in H^{2}\left( \Omega \right) $ follows from density arguments. Consider the function $v=ue^{\lambda \left( z+b\right) ^{2}}.$ Then $u=ve^{-\lambda \left( z+b\right) ^{2}}$. Hence, $u_{xx}=v_{xx}e^{-\lambda \left( z+b\right) ^{2}},u_{yy}=v_{yy}e^{-\lambda \left( z+b\right) ^{2}},$ $$u_{zz}=\left( v_{zz}-2\lambda \left( z+b\right) v_{z}+4\lambda ^{2}\left( z+b\right) ^{2}\left( 1+O\left( 1/\lambda \right) \right) v\right) e^{-\lambda \left( z+b\right) ^{2}}.$$In this proof, $C_{1}=C_{1}\left( \Omega ,b\right) >0$ denotes different constants depending only on $\Omega $ and $b$ and $O\left( 1/\lambda \right) $ denotes different $z-$dependent functions satisfying $\left\vert O\left( 1/\lambda \right) \right\vert ,\left\vert \nabla O\left( 1/\lambda \right) \right\vert \leq C_{1}/\lambda ,$ Hence,$$\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}=\left[ \left( v_{xx}+v_{yy}+v_{zz}+4\lambda ^{2}\left( z+b\right) ^{2}\left( 1+O\left( 1/\lambda \right) \right) v\right) -2\lambda \left( z+b\right) v_{z}\right] ^{2}$$$$\geq -4\lambda \left( z+b\right) v_{z}\left( v_{xx}+v_{yy}+v_{zz}+4\lambda ^{2}\left( z+b\right) ^{2}\left( 1+O\left( 1/\lambda \right) \right) v\right)$$$$=\left( -4\lambda \left( z+b\right) v_{z}v_{x}\right) _{x}+4\lambda \left( z+b\right) v_{zx}v_{x}+\left( -2\lambda \left( z+b\right) v_{z}v_{y}\right) _{y}+4\lambda \left( z+b\right) v_{zy}v_{y}$$$$+\left( -2\lambda \left( z+b\right) v_{z}^{2}\right) _{z}+2\lambda v_{z}^{2}+\left( -8\lambda ^{3}\left( z+b\right) ^{3}\left( 1+O\left( 1/\lambda \right) \right) v^{2}\right) _{z}+24\lambda ^{3}\left( z+b\right) ^{3}\left( 1+O\left( 1/\lambda \right) \right) v^{2}$$$$=-2\lambda \left( v_{x}^{2}+v_{y}^{2}\right) +2\lambda v_{z}^{2}+24\lambda ^{3}\left( z+b\right) ^{3}\left( 1+O\left( 1/\lambda \right) \right) v^{2}$$$$+\left( -2\lambda \left( z+b\right) v_{z}^{2}+2\lambda \left( z+b\right) v_{x}^{2}+2\lambda \left( z+b\right) v_{y}^{2}-8\lambda ^{3}\left( z+b\right) ^{3}\left( 1+O\left( 1/\lambda \right) \right) v^{2}\right) _{z}$$$$-2\lambda \left( v_{x}^{2}+v_{y}^{2}\right) +2\lambda v_{z}^{2}+24\lambda ^{3}\left( z+b\right) ^{2}\left( 1+O\left( 1/\lambda \right) \right) v^{2}.$$Since $v\mid _{\partial \Omega }=v_{z}\mid _{\Gamma _{0}}=0$ and $2\lambda v_{z}^{2}\geq 0,$ then integrating the above over $\Omega $ going back from $v$ to $u$ and using Gauss’ formula, we obtain for sufficiently large $\lambda \geq C_{1}$$$\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}\geq -2\lambda \mathop{\displaystyle \int}\limits_{\Omega }\left( u_{x}^{2}+u_{y}^{2}\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+23\lambda ^{3}\mathop{\displaystyle \int}\limits_{\Omega }u^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x.} \label{5.1}$$ Next,$$-u\Delta ue^{2\lambda \left( z+b\right) ^{2}}=\left( -u_{x}ue^{2\lambda \left( z+b\right) ^{2}}\right) _{x}+u_{x}^{2}e^{2\lambda \left( z+b\right) ^{2}}+\left( -u_{y}ue^{2\lambda \left( z+b\right) ^{2}}\right) _{y}+u_{y}^{2}e^{2\lambda \left( z+b\right) ^{2}}$$$$+\left( -u_{z}ue^{2\lambda \left( z+b\right) ^{2}}\right) _{z}+u_{z}^{2}e^{2\lambda \left( z+b\right) ^{2}}+4\lambda \left( z+b\right) u_{z}ue^{2\lambda \left( z+b\right) ^{2}}$$$$=\left( u_{x}^{2}+u_{y}^{2}+u_{z}^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}+\left( 2\lambda \left( z+b\right) u^{2}e^{2\lambda \left( z+b\right) ^{2}}\right) _{z}-8\lambda ^{2}\left( z+b\right) ^{2}\left( 1+O\left( 1/\lambda \right) \right) u^{2}e^{2\lambda \left( z+b\right) ^{2}}$$$$+\left( -u_{x}ue^{2\lambda \left( z+b\right) ^{2}}\right) _{x}+\left( -u_{y}ue^{2\lambda \left( z+b\right) ^{2}}\right) _{y}.$$Integrating this over $\Omega $ and using Gauss’ formula, we obtain for sufficiently large $\lambda \geq C_{1}$$$-\mathop{\displaystyle \int}\limits_{\Omega }u\Delta ue^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}=\mathop{\displaystyle \int}\limits_{\Omega }\left( u_{x}^{2}+u_{y}^{2}+u_{z}^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x} \label{5.2}$$$$-9\lambda ^{2}\mathop{\displaystyle \int}\limits_{\Omega }\left( z+b\right) ^{2}u^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}.$$Multiply (\[5.2\]) by $5\lambda /2$ and sum up with (\[5.1\]). Since $23\lambda ^{3}-\left( 9\cdot 5/2\right) \lambda ^{3}=\lambda ^{3}/2,$ then $$\mathbf{-}\frac{5}{2}\lambda \mathop{\displaystyle \int}\limits_{\Omega }u\Delta ue^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x+}\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}$$$$\geq \frac{1}{2}\lambda \mathop{\displaystyle \int}\limits_{\Omega }\left( u_{x}^{2}+u_{y}^{2}+u_{z}^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x+}\frac{1}{2}\lambda ^{3}\mathop{\displaystyle \int}\limits_{\Omega }u^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x.} \label{5.3}$$Next, applying Cauchy-Schwarz inequality, we obtain$$\mathbf{-}\frac{5}{2}\lambda \mathop{\displaystyle \int}\limits_{\Omega }u\Delta ue^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x+}\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}\leq \frac{25}{4}\lambda ^{2}\mathop{\displaystyle \int}\limits_{\Omega }u^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+\frac{1}{2}\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x.}$$Combining this with (\[5.3\]), we obtain for sufficiently large $\lambda _{0}=\lambda _{0}\left( \Omega ,b\right) >0$ and for $\lambda \geq \lambda _{0}$$$\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}\geq C_{1}\lambda \mathop{\displaystyle \int}\limits_{\Omega }\left( \left( \nabla u\right) ^{2}+\lambda ^{2}u^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x.} \label{5.4}$$ The next step is to incorporate the term with second derivatives in ([4.6]{}). We have $$\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}=\left( u_{xx}+u_{yy}+u_{zz}\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}=\left( u_{xx}^{2}+u_{yy}^{2}+u_{zz}^{2}\right) e^{2\lambda \left( z+b\right) ^{2}} \label{5.5}$$$$+2\left( u_{xx}u_{yy}+u_{xx}u_{zz}+u_{yy}u_{zz}\right) e^{2\lambda \left( z+b\right) ^{2}}.$$The second line of (\[4.6\]) gives:$$\left( 2u_{xx}u_{yy}+2u_{xx}u_{zz}+2u_{yy}u_{zz}\right) e^{2\lambda \left( z+b\right) ^{2}}=\left( 2u_{xx}u_{y}e^{2\lambda \left( z+b\right) ^{2}}\right) _{y}-2u_{xxy}u_{y}e^{2\lambda \left( z+b\right) ^{2}}$$$$+\left( 2u_{xx}u_{z}e^{2\lambda \left( z+b\right) ^{2}}\right) _{z}-2u_{xxz}u_{z}e^{2\lambda \left( z+b\right) ^{2}}-8\lambda \left( z+b\right) u_{xx}u_{z}e^{2\lambda \left( z+b\right) ^{2}}$$$$+\left( 2u_{yy}u_{z}e^{2\lambda \left( z+b\right) ^{2}}\right) _{z}-2u_{yyz}u_{z}e^{2\lambda \left( z+b\right) ^{2}}-8\lambda \left( z+b\right) u_{yy}u_{z}e^{2\lambda \left( z+b\right) ^{2}} \label{5.6}$$$$=2\left( u_{xy}^{2}+u_{xz}^{2}+u_{yz}^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}+\left[ 2\left( u_{xy}u_{y}-u_{xz}u_{z}\right) e^{2\lambda \left( z+b\right) ^{2}}\right] _{x}$$$$+\left[ 2\left( u_{xx}u_{y}-u_{xz}u_{z}\right) e^{2\lambda \left( z+b\right) ^{2}}\right] _{y}-8\lambda \left( z+b\right) u_{xx}u_{z}e^{2\lambda \left( z+b\right) ^{2}}-8\lambda \left( z+b\right) u_{yy}u_{z}e^{2\lambda \left( z+b\right) ^{2}}.$$Using Cauchy-Schwarz inequality, we estimate from the below the last line of (\[5.6\]) as$$-8\lambda \left( z+b\right) u_{xx}u_{z}e^{2\lambda \left( z+b\right) ^{2}}-8\lambda \left( z+b\right) u_{yy}u_{z}e^{2\lambda \left( z+b\right) ^{2}} \label{5.7}$$$$\geq -\frac{1}{2}\left( u_{xx}^{2}+u_{yy}^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}-64\lambda ^{2}\left( z+b\right) ^{2}u_{z}^{2}e^{2\lambda \left( z+b\right) ^{2}}.$$Combining (\[5.6\])-(\[5.7\]), we obtain$$\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}\geq \frac{1}{2}\mathop{\displaystyle \sum }\limits_{i,j=1}^{3}\mathop{\displaystyle \int}\limits_{\Omega }u_{x_{i}x_{j}}^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}-64\lambda ^{2}\mathop{\displaystyle \int}\limits_{\Omega }\left( z+b\right) ^{2}u_{z}^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}.$$Multiply this estimate by $C_{1}/\left( 128\lambda \right) $ and sum up with (\[5.4\]). We obtain$$\left( 1+\frac{C_{1}}{128\lambda }\right) \mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta u\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}\geq \frac{C_{1}}{256\lambda }\mathop{\displaystyle \sum }\limits_{i,j=1}^{3}\mathop{\displaystyle \int}\limits_{\Omega }u_{x_{i}x_{j}}^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x} \label{5.8}$$$$\mathbf{+}\frac{C}{2}\lambda \mathop{\displaystyle \int}\limits_{\Omega }\left( \left( \nabla u\right) ^{2}+\lambda ^{2}u^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x.}$$Since $C_{1}>0$ denotes different constants, then the target estimate ([4.6]{}) follows from (\[4.8\]) immediately. $\square $ Proof of Theorem 4.2 {#sec:5.2} -------------------- Denote $h=Q_{2}-Q_{1}$ implying that $Q_{2}=Q_{1}+h.$ Also, $h\in H_{0}^{3}\left( \Omega \right) ,$ $\left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }<2R$. Hence, by (\[4.40\]) $$\left\Vert h\right\Vert _{C^{1}\left( \overline{\Omega }\right) }<D_{1}\left( 2R\right) . \label{5.9}$$Using the multidimensional analog of Taylor formula (see, e.g. [@V] for this formula)* *and (\[4.42\]), we obtain $$\Delta h+\left( \Delta Q_{1}+\Delta G\right) +F\left( h+Q_{1}+G,\nabla \left( h+Q_{1}+G\right) \right)$$$$=\left[ \Delta h+F^{\left( 1\right) }\left( Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) h+F^{\left( 2\right) }\left( Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \nabla h\right] \label{5.10}$$$$+F_{nonlin}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) + \left[ \left( \Delta Q_{1}+\Delta G\right) +F\left( Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \right] ,$$where elements of $\left( N+1\right) \times \left( N+1\right) $ matrix $F^{\left( 1\right) }$ and $\left( 3N+3\right) \times \left( 3N+3\right) $ matrix are bounded for $\mathbf{x}\in \overline{\Omega }$, i.e.$$\left\vert F_{i,j}^{\left( 1\right) }\left( Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \right\vert ,\left\vert F_{i,j}^{\left( 2\right) }\left( Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \right\vert \leq C_{2},\text{ }\forall \mathbf{x}\in \overline{\Omega }, \label{5.11}$$where the subscript $i,j"$ denotes an arbitrary element of the corresponding matrix indexed as $\left( i,j\right) .$ Next, the $\left( N+1\right) -$dimensional vector function $F_{nonlin}$ depends nonlinearly on $h,\nabla h.$ Furthermore, the following estimate follows from (\[4.40\])-(\[4.42\])$$\left\vert F_{nonlin}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \right\vert \leq C_{2}\left( \left\vert h\right\vert ^{2}+\left\vert \nabla h\right\vert ^{2}\right) ,\text{ }\forall \mathbf{x}\in \overline{\Omega }. \label{5.12}$$Next, (\[5.9\]) and (\[5.12\]) imply with a different constant $C_{2}$ $$\left\vert F_{nonlin}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \right\vert \leq C_{2}\left( \left\vert h\right\vert +\left\vert \nabla h\right\vert \right) ,\text{ }\forall \mathbf{x}\in \overline{\Omega }. \label{5.13}$$It follows from (\[5.10\])-(\[5.13\]) that$$\left[ \Delta h+\left( \Delta Q_{1}+\Delta G\right) +F\left( h+Q_{1}+G,\nabla \left( h+Q_{1}+G\right) \right) \right] ^{2}$$$$-\left[ \left( \Delta Q_{1}+\Delta G\right) +F\left( Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \right] ^{2} \label{5.14}$$$$=Lin_{1}\left( \Delta h\right) +Lin_{2}\left( \nabla h\right) +Lin_{3}\left( h\right)$$$$+\left( \Delta h\right) ^{2}+M_{1}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \Delta h+M_{2}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) ,$$where expressions $Lin_{1}\left( \Delta h\right) ,Lin_{2}\left( \nabla h\right) $ and $Lin_{3}\left( h\right) $ are linear with respect to $\Delta h,\nabla h$ and $h$ respectively and $$\left\vert Lin_{1}\left( \Delta h\right) +Lin_{2}\left( \nabla h\right) +Lin_{3}\left( h\right) \right\vert \leq C_{2}\left( \left\vert \Delta h\right\vert +\left\vert \nabla h\right\vert +\left\vert h\right\vert \right) ,\text{ }\forall \mathbf{x}\in \overline{\Omega }. \label{5.15}$$Next, the following estimates are valid for $M_{1}$ and $M_{2}$ $$\left\vert M_{1}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \right\vert \leq C_{2}\left( \left\vert \nabla h\right\vert +\left\vert h\right\vert \right) ,\text{ }\forall \mathbf{x}\in \overline{\Omega }, \label{5.16}$$$$\left\vert M_{2}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \right\vert \leq C_{2}\left( \left\vert \nabla h\right\vert ^{2}+\left\vert h\right\vert ^{2}\right) ,\forall \mathbf{x}\in \overline{\Omega }. \label{5.17}$$In particular, (\[5.16\]), (\[5.17\]) and Cauchy-Schwarz inequality imply $$\left( \Delta h\right) ^{2}+M_{1}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right) \Delta h+M_{2}\left( h,\nabla h,Q_{1}+G,\nabla \left( Q_{1}+G\right) \right)$$$$\geq \frac{1}{2}\left( \Delta h\right) ^{2}-C_{2}\left( \left\vert \nabla h\right\vert ^{2}+\left\vert h\right\vert ^{2}\right) ,\forall \mathbf{x}\in \overline{\Omega }. \label{5.18}$$Using (\[4.5\]) and (\[5.14\])-(\[5.17\]), we obtain$$J_{\lambda ,\beta }\left( Q_{1}+h\right) -J_{\lambda ,\beta }\left( Q_{1}\right) =X_{lin}\left( h\right) +X_{nonlin}\left( h\right) , \label{5.19}$$where $X_{lin}\left( h\right) $ can be extended from $\left\{ \left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }<2R\right\} \subset H_{0}^{3}\left( \Omega \right) $ to the entire space $H^{3}\left( \Omega \right) $ as a bounded linear  functional, $$X_{lin}\left( h\right) =e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( Lin_{1}\left( \Delta h\right) +Lin_{2}\left( \nabla h\right) +Lin_{3}\left( h\right) \right) \left( \mathbf{x}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+2\beta \left[ h,Q_{1}+G\right] , \label{5.20}$$where $\left[ .,.\right] $ is the scalar product in $H^{3}\left( \Omega \right) .$ As to $X_{nonlin}\left( h\right) $ in (\[5.19\]), it follows from (\[4.5\]), (\[5.14\]), (\[5.16\]) and (\[5.17\]) that $$\lim_{\left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }\rightarrow 0}\frac{X_{nonlin}\left( h\right) }{\left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }}=0. \label{5.21}$$Using (\[5.15\]) and (\[5.19\])-(\[5.21\]), we obtain that $X_{lin}\left( h\right) $ is the Frechét derivative $J_{\lambda ,\beta }^{\prime }\left( Q\right) $ of the functional $J_{\lambda ,\beta }\left( Q\right) $ at the point $Q$, i.e. $X_{lin}\left( h\right) =J_{\lambda ,\beta }^{\prime }\left( Q_{1}\right) \left( h\right) $. Thus, the existence of the Frechét derivative is established. Next, using (\[4.5\]) and (\[5.14\])-(\[5.20\]), we obtain$$J_{\lambda ,\beta }\left( Q_{1}+G+h\right) -J_{\lambda ,\beta }\left( Q_{1}+G\right) -J_{\lambda ,\beta }^{\prime }\left( Q_{1}+G\right) \left( h\right) \label{5.22}$$$$\geq \frac{1}{2}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta h\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}-C_{2}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \left\vert \nabla h\right\vert ^{2}+\left\vert h\right\vert ^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+\beta \left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }^{2}.$$We now apply Carleman estimate (\[4.6\]), assuming that $\lambda \geq \lambda _{0},$$$\frac{1}{2}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta h\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}-C_{2}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \left\vert \nabla h\right\vert ^{2}+\left\vert h\right\vert ^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+\beta \left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }^{2}$$$$\geq \frac{C_{1}}{\lambda }\mathop{\displaystyle \sum }\limits_{i,j=1}^{3}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }h_{x_{i}x_{j}}^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x+}C_{1}\lambda e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \left( \nabla h\right) ^{2}+\lambda ^{2}h^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}$$$$-C_{2}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \left\vert \nabla h\right\vert ^{2}+\left\vert h\right\vert ^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+\beta \left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }^{2}.$$Choosing sufficiently large $\lambda _{1}=\lambda _{1}\left( F,\left\Vert G\right\Vert _{H^{3}\left( \Omega \right) },m,R,\Omega ,b\right) \geq \lambda _{0}$ and letting $\lambda \geq \lambda _{1},$ we obtain with a different constant $C_{2}$$$\frac{1}{2}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta h\right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}-C_{2}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \left\vert \nabla h\right\vert ^{2}+\left\vert h\right\vert ^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+\beta \left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }^{2}$$$$\geq \frac{C_{2}}{\lambda }\mathop{\displaystyle \sum }\limits_{i,j=1}^{3}e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }h_{x_{i}x_{j}}^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x+}C_{2}\lambda e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \left( \nabla h\right) ^{2}+\lambda ^{2}h^{2}\right) e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+\beta \left\Vert h\right\Vert _{H^{3}\left( \Omega \right) }^{2}.$$This, (\[5.22\]) and (\[4.50\]) imply (\[4.7\]).  $\square $ Proof of Theorem 4.6 {#sec:5.3} -------------------- We rewrite the functional $J_{\lambda ,\beta }\left( Q\right) $ in (\[4.5\]) as $$J_{\lambda ,\beta }\left( Q+G\right) =J_{\lambda ,\beta }^{0}\left( Q+G\right) +\beta \left\Vert Q+G\right\Vert _{H^{3}\left( \Omega \right) }^{2}. \label{5.23}$$Since the vector function $Q^{\ast }\in B\left( m,R\right) $ is the exact solution of the problem (\[4.2\]), (\[4.3\]) with the noiseless data $G^{\ast },$ then $J_{\lambda ,\beta }^{0}\left( Q^{\ast }+G^{\ast }\right) =0.$ Hence, $$J_{\lambda ,\beta }\left( Q^{\ast }+G^{\ast }\right) \leq C_{2}\beta . \label{5.24}$$Next, $J_{\lambda ,\beta }\left( Q^{\ast }+G\right) =\left( J_{\lambda ,\beta }\left( Q^{\ast }+G\right) -J_{\lambda ,\beta }\left( Q^{\ast }+G^{\ast }\right) \right) +J_{\lambda ,\beta }\left( Q^{\ast }+G^{\ast }\right) .$ Hence, applying (\[5.24\]), we obtain$$J_{\lambda ,\beta }\left( Q^{\ast }+G\right) \leq \left\vert J_{\lambda ,\beta }\left( Q^{\ast }+G\right) -J_{\lambda ,\beta }\left( Q^{\ast }+G^{\ast }\right) \right\vert +C_{2}\beta . \label{5.25}$$Using (\[4.10\]) and (\[5.23\]), we estimate the first term in the right hand side of (\[5.25\]),$$\left\vert J_{\lambda ,\beta }\left( Q^{\ast }+G\right) -J_{\lambda ,\beta }\left( Q^{\ast }+G^{\ast }\right) \right\vert \leq \left\vert J_{\lambda ,\beta }^{0}\left( Q^{\ast }+G\right) -J_{\lambda ,\beta }^{0}\left( Q^{\ast }+G^{\ast }\right) \right\vert +C_{2}\beta \delta$$$$\leq C_{2}\delta \exp \left( 2\lambda \left( A+b\right) ^{2}\right) +C_{2}\beta \delta . \label{5.26}$$Recall that due to our choice $\lambda =\lambda \left( \delta \right) =\ln \delta ^{-\eta },$ where $\eta =\left[ 4\left( A+b\right) ^{2}\right] ^{-1},$ we have $\delta \exp \left( 2\lambda \left( A+b\right) ^{2}\right) =1/\sqrt{\delta }.$ Hence, (\[5.26\]) implies $$\left\vert J_{\lambda ,\beta }\left( Q^{\ast }+G\right) -J_{\lambda ,\beta }\left( Q^{\ast }+G^{\ast }\right) \right\vert \leq C_{2}\sqrt{\delta },\text{ }\forall \beta \in \left( 0,1\right) .$$Combining this with (\[5.24\]), we obtain$$J_{\lambda ,\beta }\left( Q^{\ast }+G\right) \leq C_{2}\left( \sqrt{\delta }+\beta \right) . \label{5.27}$$ Until now we have not used in this proof the strict convexity of the functional $J_{\lambda ,\beta }\left( Q+G\right) $ for $Q\in B\left( m,R\right) .$ But we will use this property in the rest of the proof. Recall that by Theorem 4.3 $Q_{\min ,\lambda ,\beta }\in \overline{B\left( m,R\right) }$ is the unique minimizer on the set $\overline{B\left( m,R\right) }$ of the functional $J_{\lambda ,\beta }\left( Q+G\right) $ on the set $\overline{B\left( m,R\right) }$. By Theorem 4.2 $$J_{\lambda ,\beta }\left( Q^{\ast }+G\right) -J_{\lambda ,\beta }\left( Q_{\min ,\lambda ,\beta }+G\right) -J_{\lambda ,\beta }^{\prime }\left( Q_{\min ,\lambda ,\beta }\right) \left( Q^{\ast }-Q_{\min ,\lambda ,\beta }\right) \label{5.28}$$$$\geq \frac{C_{2}}{\lambda }\mathop{\displaystyle \sum }\limits_{i,j=1}\left\Vert \left( Q^{\ast }-Q_{\min ,\lambda ,\beta }\right) _{x_{i}x_{j}}\right\Vert _{L_{2}\left( \Omega \right) }^{2}+C_{2}\lambda \left\Vert Q^{\ast }-Q_{\min ,\lambda ,\beta }\right\Vert _{H^{1}\left( \Omega \right) }^{2}+\frac{\beta }{2}\left\Vert Q^{\ast }-Q_{\min ,\lambda ,\beta }\right\Vert _{H^{3}\left( \Omega \right) }^{2}.$$Since by (\[4.8\]) $-J_{\lambda ,\beta }^{\prime }\left( Q_{\min ,\lambda ,\beta }\right) \left( Q^{\ast }-Q_{\min ,\lambda ,\beta }\right) \leq 0,$ then (\[5.27\]) implies that the left hand side of (\[5.28\]) can be estimated as$$J_{\lambda ,\beta }\left( Q^{\ast }+G\right) -J_{\lambda ,\beta }\left( Q_{\min ,\lambda ,\beta }+G\right) -J_{\lambda ,\beta }^{\prime }\left( Q_{\min ,\lambda ,\beta }+G\right) \left( Q^{\ast }-Q_{\min ,\lambda ,\beta }\right) \leq C_{2}\left( \sqrt{\delta }+\beta \right) .$$Hence, using our choice of $\lambda =\lambda \left( \delta \right) =\ln \delta ^{-\eta }$ and (\[5.28\]), we obtain estimates (\[1\]) and ([2]{}). Estimates (\[3\]) and (\[4\]) are obtained from (\[1\]) and ([2]{}) respectively using (\[4.90\]) and the triangle inequality. Estimate (\[5\]) obviously follows from estimate (\[4\]). $\ \square $ Numerical Studies {#sec:6} ================= The single point source is now $\mathbf{x}_{0}=(0,0,-5).$ We choose in ([2.1]{}) the numbers $A=1$. Hence, below $$\Omega =\left\{ -1/2<x,y<1/2,z\in \left( 0,1\right) \right\} ,$$$$\Gamma _{0}=\left\{ \mathbf{x=}\left( x,y,z\right) :-1/2<x,y<1/2,z=1\right\} ,\Gamma _{1}=\partial \Omega \diagdown \Gamma _{0}. \label{6.0}$$We have introduced the vector function $G$ in section 4.1 and thus obtained the problem (\[4.2\]), (\[4.3\]) for the vector function $Q=W-G$ from the problem (\[3.22\]), (\[3.23\]) for the vector function $W$ in order to obtain zero boundary conditions (\[4.3\]) for $Q$. The latter was convenient for proofs of Theorems 4.3-4.6. However, it follows from Theorems 4.2-4.6 that their obvious analogs hold true for the functional $$J_{\lambda ,\beta }\left( W\right) =e^{-2\lambda b^{2}}\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta W+F\left( \nabla W,W\right) \right) ^{2}e^{2\lambda \left( z+b\right) ^{2}}d\mathbf{x}+\beta \left\Vert W\right\Vert _{H^{3}\left( \Omega \right) }^{2}, \label{6.1}$$$$W\in B_{W}\left( m,R\right) =\left\{ W:W=Q-G,\forall Q\in B\left( m,R\right) \right\} . \label{6.10}$$Furthermore, we use in (\[6.1\]) $\beta =0,b=0$. Therefore, we ignore the multiplier $e^{-2\lambda b^{2}},$ which was used above to balance first and second terms in the right hand side of (\[4.5\]), see (\[4.50\]). Hence, we minimize the weighted cost functional $$J_{\lambda }\left( W\right) =\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta W+F\left( \nabla W,W\right) \right) ^{2}e^{2\lambda z^{2}}d\mathbf{x} \label{6.100}$$on the set (\[6.10\]). We conjecture that the case $\beta =0$ works probably because the minimal mesh size of 1/32 in the finite differences we use to minimize this functional is not too small, and all norms in finite dimensional spaces are equivalent, see item 4 of Remark 4.1. In addition, recall that, by the same item, one can choose any value of $\beta \in \left( 0,1\right) $ in convergence estimates (\[1\])-(\[4\]). Also, we use the gradient descent method (GD) instead of a more complicated gradient projection method. We have observed that GD works well for our computations. The latter coincides with observations in all earlier publications about numerical studies of the convexification [KlibThanh,KlibKol1,KlibKol2,KlibKol3,KlibKol4,KEIT]{}. As to our choice $b=0,$ one can derive from the proof of Theorem 4.1 that a slightly modified Carleman estimate of this theorem works in a little bit smaller domain $\Omega ^{\prime }=\Omega \cap \left\{ z>\varepsilon \right\} $ for any small number $\varepsilon >0.$ Finally, we believe that simplifications listed in this section work well numerically due to a commonly known observation that numerical studies are usually less pessimistic than the theory is. Some details of the numerical implementation {#sec:6.1} -------------------------------------------- To solve the inverse problem, we should first computationally simulate the data (\[2.10\]) at $\partial \Omega $ via the numerical solution of the forward problem (\[2.7\]), (\[2.8\]). To solve the problem (\[2.7\]), (\[2.8\]) computationally, we have used the standard finite difference method. To avoid the use of the infinite space $\mathbb{R}^{3}$ in the solution of the forward problem, we choose the cube $\Omega _{f}=\left\{ -6.5<x,y<6.5,z\in \left( -6,7\right) \right\} .$ So that $\Omega \subset \Omega _{f}$, $\partial \Omega \cap \partial \Omega _{f}=\varnothing $ and $\mathbf{x}_{0}=\left( 0,0,-5\right) \in \Omega _{f}.$ We choose a sufficiently large number $T_{0}=6.5$. Then we solve equation (\[2.7\]) with the initial condition (\[2.8\]) and zero Dirichlet boundary condition at $\partial \Omega _{f}$ for $\left( \mathbf{x},t\right) \in \Omega _{f}\times \left( 0,T_{0}\right) $ via finite differences. Indeed, the wave originated at $\mathbf{x}_{0}$ cannot reach neither vertical sides of $\Omega _{f}$ nor the upper side $\left\{ z=7\right\} \cap \overline{\Omega _{f}}$ of $\Omega $ for times $t\in \left( 0,6.5\right) .$ However, it reaches the upper side $\left\{ z=1\right\} \cap \overline{\Omega }$ of $\Omega $ at $t=6.$ This wave reaches the lower side $\left\{ z=-6\right\} \cap \overline{\Omega _{f}}$  of $\Omega _{f}$ at $t=1,$ which means that the zero Dirichlet boundary condition on the lower side is incorrect. Still, for $t\in \left( 0,6.5\right) ,$ the wave reflected from the lower side of of $\Omega _{f}$ does not reach the upper side $\left\{ z=1\right\} \cap \overline{\Omega }$ of $\Omega ,$ where the  data for our CIP are given. Hence, this reflected wave does not affect our data. We use the explicit scheme. The grid step size in each spatial direction is $\Delta x=1/32$ and in time direction $\Delta t=0.002.$ To avoid a substantial increase of the computational time, we do not decrease these step sizes. When solving the forward problem, we model the $\delta \left( \mathbf{x}-\mathbf{x}_{0}\right) -$function in (\[2.8\]) as$$\widetilde{\delta }\left( \mathbf{x}-\mathbf{x}_{0}\right) =\left\{ \begin{array}{l} \frac{1}{\varepsilon }\exp \left( -\frac{1}{1-\left\vert \mathbf{x}-\mathbf{x}_{0}\right\vert ^{2}/\varepsilon }\right) ,\text{ if }\left\vert \mathbf{x}-\mathbf{x}_{0}\right\vert ^{2}<\varepsilon =0.01, \\ 0,\text{ otherwise.}\end{array}\right.$$ In computations of the inverse problem, for each test we use, we choose in the data (\[2.10\]) $T=\max_{\overline{\Omega }}\tau \left( \mathbf{x}\right) +0.1.$ We have observed that $T<T_{0}$ in all our tests. Next, we set $T_{1}=T-\max_{\overline{\Omega }}\tau \left( \mathbf{x}\right) =0.1,$ see (\[3.11\]), (\[3.110\]). An important question is on how do we figure out boundary conditions at $\partial \Omega $ for the function $\tau \left( \mathbf{x}\right) ,$ i.e. $\tau \left( \mathbf{x}\right) \mid _{\partial \Omega }$ and also $\partial _{z}\tau \left( \mathbf{x}\right) \mid _{\Gamma _{0}}.$ In principle, for $\mathbf{x}\in \partial \Omega ,$ one should choose such a number $\tau _{0}\left( \mathbf{x}\right) $ that $\tau _{0}\left( \mathbf{x}\right) =\min_{t}\left\{ t:u\left( \mathbf{x,}t\right) >0\right\} .$ However, it is hard to choose in practice the number $\tau _{0}\left( \mathbf{x}\right) $ satisfying this criterion. Therefore, we calculate such a number $\widetilde{t}\left( \mathbf{x}\right) $ at which the first wave with the largest maximal value arrives at the point $\mathbf{x}\in \partial \Omega $, see Figure 2. Next, we set $\tau _{0}\left( \mathbf{x}\right) \mid _{\partial \Omega }:=\widetilde{t}\left( \mathbf{x}\right) \mid _{\partial \Omega }.$ To calculate the derivative $\partial _{z}\tau _{0}\left( \mathbf{x}\right) \mid _{\Gamma _{0}},$ we compute the discrete normal derivative of $\tau _{0}\left( \mathbf{x}\right) $ over the mesh in the forward problem. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![(a) A schematic diagram of domains $\Omega ,$ source and detectors. (b) This figure explains how do we approximately choose the boundary condition $\protect\tau \left( \mathbf{x}\right) \mid _{\partial \Omega }.$ We have chosen here a selected point ${\mathbf{x}}\in \Gamma _{0}$[]{data-label="domain"}](Figures/domain.png "fig:"){width="6cm"} ![(a) A schematic diagram of domains $\Omega ,$ source and detectors. (b) This figure explains how do we approximately choose the boundary condition $\protect\tau \left( \mathbf{x}\right) \mid _{\partial \Omega }.$ We have chosen here a selected point ${\mathbf{x}}\in \Gamma _{0}$[]{data-label="domain"}](Figures/figure2.png "fig:"){width="6cm"} (a) (b) ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ To minimize the weighted cost functional $J_{\lambda }\left( W\right) $ in (\[6.10\]), we act similarly with the previous above cited works about numerical studies of the convexification for a number of other CIPs. More precisely, we write the differential operators involved in $J_{\lambda }\left( W\right) $ via finite differences and minimize with respect to the values of the discrete analog of the vector function $W$ at grid points. As to the choice of the parameter $\lambda ,$even though the above theory works only for sufficiently large values of $\lambda $, we have established in our computational experiments that the choice $\lambda =1$ is an optimal one for all tests we have performed. This again repeats observation of all above cited works on numerical studies of the convexification, in which the optimal choice was $\lambda \in \left[ 1,3\right] .$ We have also tested two different values of the number $N$ terms in the series (\[3.16\]): $N=1$ and $N=3.$ Our computational results indicate that $N=3$ provides results of a good quality. In all tests below, the starting point of GD is the vector function $W\left( \mathbf{x}\right) ,$ which is generated by the coefficient $c\left( \mathbf{x}\right) \equiv 1$ in equation (\[2.7\]), $W_{c\equiv 1}\left( \mathbf{x}\right) .$ A multi-level minimization method of the functional $J_{\protect\lambda }\left( W\right) $ {#sec:6.2} ------------------------------------------------------------------------------------------ We have found in our computational experiments that the gradient descent method for our weighted cost functional $J_{\lambda ,0}\left( W\right) $ converges rapidly on a coarse mesh. This provides us with a rough image. Hence, we have implemented a multi-level method [@LZ]. Let $M_{h_{1}}\subset M_{h_{2}}...\subset M_{h_{K}}$ be nested finite difference meshes, i.e. $M_{h_{k}}$ is a refinement of $M_{h_{k-1}}$ for $k\leq K$. Let $P_{h_{k}}$ be the corresponding finite difference functional space. One the first level $M_{h_{1}}$, we solve the discrete optimization problem. In other words, let $W_{h_{1},\min }$ be the minimizer on the finite difference analog of the set (\[6.10\]) of the following functional $$J_{\lambda }\left( W_{h_{1}}\right) =\mathop{\displaystyle \int}\limits_{\Omega }\left( \Delta W_{h_{1}}+F\left( \nabla W_{h_{1}},W_{h_{1}}\right) \right) ^{2}e^{2\lambda z^{2}}d\mathbf{x.} \label{6.4}$$In (\[6.4\]) the integral and the derivatives are understood in the discrete sense, and $W_{h_{1},\min }$ is found via the GD. Then we interpolate the minimizer $W_{h_{1},\min }$ on the finer mesh $M_{h_{2}}$ and use the resulting vector function $W_{h_{2},\text{int}}$ as the starting point of the gradient descent method of the optimization of the direct analog of functional (\[6.4\]) in which $h_{1}$ is replaced with $h_{2}$ and $W_{h_{1}}$ is replaced with $W_{h_{2}}.$ This process is repeated until we obtain the minimizer $W_{h_{K},\min }$ on the $K_{th}$ level on the mesh $M_{h_{K}}$. Since $(x,y,z)\in (-1/2,1/2)\times (-1/2,1/2)\times (0,1)$, then our first level $M_{h_{1}}$ is set to be the uniform mesh with the grid step $h_{1}=1/8 $. For each mesh refinement, we will refine the mesh via setting the new grid step of the refined mesh in all directions to be 1/2 of the previous grid step. On each level $M_{h_{k}}$, we stop iterations as soon as we see that $\Vert \nabla J_{\lambda }^{(h_{k})}(W_{h_{k}})\Vert <2\times 10^{-2}$. Next, we refine the mesh and compute the solution on the next level $M_{h_{k+1}}$. In the end, we compute our approximation for the target coefficient $c(\mathbf{x})$ using the final minimizer $W_{h_{K},\min }$. Numerical testing {#sec:6.3} ----------------- In the tests of this section, we demonstrate the efficiency of our numerical method for imaging of small inclusions as well as for imaging of a smoothly varying function $c\left( \mathbf{x}\right) $. In all tests the background value of $c_{bkgr}=1.$ Note that a postprocessing of images was not applied in our numerical tests. All necessary derivatives of the data were calculated using finite differences, just as in all above cited previous publications about numerical studies of the convexification, including two noisy experimental data [@KlibKol3; @KlibKol4]. Just as in all those works, we have not observed instabilities due to the differentiation, most likely because the step sizes of finite differences were not too small. On Figures 2-7 slices are depicted to demonstrate the values of the computed function $c\left( \mathbf{x}\right) .$ **Test 1**. First, we test the reconstruction by our method of a single ball shaped inclusion depicted on Figure \[example1\] a). $c=2$ inside of this inclusion and $c=1$ outside. Hence, the inclusion/background contrast is 2:1. We show the 3D image and slices for $N=1,3$, see Figures [example1]{}. **Test 2**. Second, we test the reconstruction by our method of a single elliptically shaped inclusion depicted on Figure \[example2\] a). $c=2$ inside of this inclusion and $c=1$ outside. Hence, the inclusion/background contrast is 2:1. We show the 3D image and slices for $N=1,3$, see Figures \[example2\]. **Test 3**. We now test the performance of our method for imaging of two ball shaped inclusions depicted on Figure \[example3\] a). $c=2$ inside of each inclusion and $c=1$ outside of these inclusions. Figures [example3]{} display results. **Test 4**. We now test our method for the case when the function $c\left( \mathbf{x}\right) $ is smoothly varying within an abnormality and with a wide range of variations approximately between 0.6 and 1.7. The results are shown in Figure \[example4\]. Thus, our method can accurately image not only sharp" inclusions as in Tests 1-3, but abnormalities with smoothly varying functions $c\left( \mathbf{x}\right) $ in them as well. **Test 5**. In this example, we test the reconstruction by our method of a single ball shaped inclusion with a high inclusion/background contrast, see Figure \[example5\] a). $c=5$ inside of this inclusion and $c=1$ outside. Hence, the inclusion/background contrast is 5:1. See Figures [example5]{} for results. **Test 6**. In this example we test the stability of our algorithm with respect to the random noise in the data. We test the stability for the case of the function $c\left( \mathbf{x}\right) $ described in Test 4. The noise is added for $\mathbf{x}\in \Gamma _{0}$ (see (\[6.0\])) as: $$g_{0,\text{noise}}(\mathbf{x},t)=g_{0}(\mathbf{x},t)(1+\epsilon \xi _{t})\text{ and }g_{1,\text{noise}}(\mathbf{x},t)=g_{1}(\mathbf{x},t)(1+\epsilon \xi _{t}), \label{6.5}$$where functions $g_{0}(\mathbf{x},t),g_{1}(\mathbf{x},t)$ are defined in (\[2.10\]), $\epsilon $ is the noise level and $\xi _{t}$ is a random variable depending only on the time $t$ and uniformly distributed on $[-1,1]$. We took $\epsilon =5\%$ which is 5% noise. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![*Results of Test 1. Imaging of one ball shaped inclusion with* $c=2$* in it and $c=1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N $. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.8.*[]{data-label="example1"}](Figures/1inclu_circle_true_3d.png "fig:"){width="5cm"} ![*Results of Test 1. Imaging of one ball shaped inclusion with* $c=2$* in it and $c=1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N $. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.8.*[]{data-label="example1"}](Figures/1inclu_circle_true_slice.png "fig:"){width="5cm"} (a)3D image of true $c$ \(b) Slice image of true $c$ ![*Results of Test 1. Imaging of one ball shaped inclusion with* $c=2$* in it and $c=1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N $. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.8.*[]{data-label="example1"}](Figures/1inclu_circle_N1_3d.png "fig:"){width="5cm"} ![*Results of Test 1. Imaging of one ball shaped inclusion with* $c=2$* in it and $c=1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N $. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.8.*[]{data-label="example1"}](Figures/1inclu_circle_N1_slice.png "fig:"){width="5cm"} \(c) $N=1$ \(d) $N=1$ ![*Results of Test 1. Imaging of one ball shaped inclusion with* $c=2$* in it and $c=1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N $. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.8.*[]{data-label="example1"}](Figures/1inclu_circle_N3_3d.png "fig:"){width="5cm"} ![*Results of Test 1. Imaging of one ball shaped inclusion with* $c=2$* in it and $c=1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N $. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.8.*[]{data-label="example1"}](Figures/1inclu_circle_N3_slice.png "fig:"){width="5cm"} \(e) $N=3$ \(f) $N=3$ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![*Results of Test 2. Imaging of one elliptically shaped inclusion with* $c =2$* in it and $c =1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.6.*[]{data-label="example2"}](Figures/1inclu_ellipse_true_3d.png "fig:"){width="5cm"} ![*Results of Test 2. Imaging of one elliptically shaped inclusion with* $c =2$* in it and $c =1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.6.*[]{data-label="example2"}](Figures/1inclu_ellipse_true_slice.png "fig:"){width="5cm"} (a)3D image of true $c$ \(b) Slice image of true $c$ ![*Results of Test 2. Imaging of one elliptically shaped inclusion with* $c =2$* in it and $c =1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.6.*[]{data-label="example2"}](Figures/1inclu_ellipse_N1_3d.png "fig:"){width="5cm"} ![*Results of Test 2. Imaging of one elliptically shaped inclusion with* $c =2$* in it and $c =1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.6.*[]{data-label="example2"}](Figures/1inclu_ellipse_N1_slice.png "fig:"){width="5cm"} \(c) $N=1$ \(d) $N=1$ ![*Results of Test 2. Imaging of one elliptically shaped inclusion with* $c =2$* in it and $c =1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.6.*[]{data-label="example2"}](Figures/1inclu_ellipse_N3_3d.png "fig:"){width="5cm"} ![*Results of Test 2. Imaging of one elliptically shaped inclusion with* $c =2$* in it and $c =1$ outside. Hence, the inclusion/background contrast is 2:1. We have stopped at the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.6.*[]{data-label="example2"}](Figures/1inclu_ellipse_N3_slice.png "fig:"){width="5cm"} \(e) $N=3$ \(f) $N=3$ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![*Results of Test 3. Imaging of two ball shaped inclusions with* $c =2$* in each of them and $c =1$ outside. We have stopped on the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. In each imaged inclusion, the maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.9.*[]{data-label="example3"}](Figures/2inclu_circle_true_3d.png "fig:"){width="5cm"} ![*Results of Test 3. Imaging of two ball shaped inclusions with* $c =2$* in each of them and $c =1$ outside. We have stopped on the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. In each imaged inclusion, the maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.9.*[]{data-label="example3"}](Figures/2inclu_circle_true_slice.png "fig:"){width="5cm"} (a)3D image of true $c$ \(b) Slice image of true $c$ ![*Results of Test 3. Imaging of two ball shaped inclusions with* $c =2$* in each of them and $c =1$ outside. We have stopped on the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. In each imaged inclusion, the maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.9.*[]{data-label="example3"}](Figures/2inclu_circle_N1_3d.png "fig:"){width="5cm"} ![*Results of Test 3. Imaging of two ball shaped inclusions with* $c =2$* in each of them and $c =1$ outside. We have stopped on the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. In each imaged inclusion, the maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.9.*[]{data-label="example3"}](Figures/2inclu_circle_N1_slice.png "fig:"){width="5cm"} \(c) $N=1$ \(d) $N=1$ ![*Results of Test 3. Imaging of two ball shaped inclusions with* $c =2$* in each of them and $c =1$ outside. We have stopped on the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. In each imaged inclusion, the maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.9.*[]{data-label="example3"}](Figures/2inclu_circle_N3_3d.png "fig:"){width="5cm"} ![*Results of Test 3. Imaging of two ball shaped inclusions with* $c =2$* in each of them and $c =1$ outside. We have stopped on the 3rd mesh refinement for all three values of $N$. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. In each imaged inclusion, the maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 1.9.*[]{data-label="example3"}](Figures/2inclu_circle_N3_slice.png "fig:"){width="5cm"} \(e) $N=3$ \(f) $N=3$ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![*Results of Test 4. Imaging of a smoothly varying coefficient. The function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.4 and 1.6. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies approximately between 0.7 and 1.6.*[]{data-label="example4"}](Figures/smmoth4_true_3d.png "fig:"){width="5cm"} ![*Results of Test 4. Imaging of a smoothly varying coefficient. The function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.4 and 1.6. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies approximately between 0.7 and 1.6.*[]{data-label="example4"}](Figures/smooth4_true_slice.png "fig:"){width="5cm"} \(a) 3D image of true $c$ \(b) Slice image of true $c$ ![*Results of Test 4. Imaging of a smoothly varying coefficient. The function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.4 and 1.6. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies approximately between 0.7 and 1.6.*[]{data-label="example4"}](Figures/smooth4_N1_3d.png "fig:"){width="5cm"} ![*Results of Test 4. Imaging of a smoothly varying coefficient. The function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.4 and 1.6. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies approximately between 0.7 and 1.6.*[]{data-label="example4"}](Figures/smooth4_N1_slice.png "fig:"){width="5cm"} \(c) $N=1$ \(d) $N=1$ ![*Results of Test 4. Imaging of a smoothly varying coefficient. The function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.4 and 1.6. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies approximately between 0.7 and 1.6.*[]{data-label="example4"}](Figures/smooth4_N3_3d.png "fig:"){width="5cm"} ![*Results of Test 4. Imaging of a smoothly varying coefficient. The function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.4 and 1.6. a) and b) Correct images. c) and d) Computed images for $N=1$. e) and f) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies approximately between 0.7 and 1.6.*[]{data-label="example4"}](Figures/smooth4_N3_slice.png "fig:"){width="5cm"} \(e) $N=3$ \(f) $N=3$ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![*Results of Test 5. Imaging of one ball shaped inclusion with* $c =5$* in it and* $c =1$ *outside. Hence, the inclusion/background contrast is 5:1. We have stopped at the 3rd mesh refinement. a) and b) Correct images. c) and d) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 4.*[]{data-label="example5"}](Figures/1inclu_high_circle_true_3d.png "fig:"){width="5cm"} ![*Results of Test 5. Imaging of one ball shaped inclusion with* $c =5$* in it and* $c =1$ *outside. Hence, the inclusion/background contrast is 5:1. We have stopped at the 3rd mesh refinement. a) and b) Correct images. c) and d) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 4.*[]{data-label="example5"}](Figures/1inclu_high_circle_true_slice.png "fig:"){width="5cm"} \(a) 3D image of true $c$ \(b) Slice image of true $c$ ![*Results of Test 5. Imaging of one ball shaped inclusion with* $c =5$* in it and* $c =1$ *outside. Hence, the inclusion/background contrast is 5:1. We have stopped at the 3rd mesh refinement. a) and b) Correct images. c) and d) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 4.*[]{data-label="example5"}](Figures/1inclu_high_circle_N3_3d.png "fig:"){width="5cm"} ![*Results of Test 5. Imaging of one ball shaped inclusion with* $c =5$* in it and* $c =1$ *outside. Hence, the inclusion/background contrast is 5:1. We have stopped at the 3rd mesh refinement. a) and b) Correct images. c) and d) Computed images for $N=3$. The maximal value of the computed coefficient $c\left( \mathbf{x}\right) $ is approximately 4.*[]{data-label="example5"}](Figures/1inclu_high_circle_N3_slice.png "fig:"){width="5cm"} \(e) $N=3$ \(f) $N=3$ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![*Results of Test 6. We test the reconstruction of the same function $c\left( \mathbf{x}\right) $ as in Test 4 (Figures 5) but with the noise in the data. The level of noise in is $\protect\epsilon=5\% $. We have stopped at the 3rd mesh refinement for $N=3$. a) and b) Correct images. c) and d) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.7 and 1.6.*[]{data-label="example6"}](Figures/smmoth4_true_3d.png "fig:"){width="5cm"} ![*Results of Test 6. We test the reconstruction of the same function $c\left( \mathbf{x}\right) $ as in Test 4 (Figures 5) but with the noise in the data. The level of noise in is $\protect\epsilon=5\% $. We have stopped at the 3rd mesh refinement for $N=3$. a) and b) Correct images. c) and d) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.7 and 1.6.*[]{data-label="example6"}](Figures/smooth4_true_slice.png "fig:"){width="5cm"} \(a) 3D image of true $c$ \(b) Slice image of true $c$ ![*Results of Test 6. We test the reconstruction of the same function $c\left( \mathbf{x}\right) $ as in Test 4 (Figures 5) but with the noise in the data. The level of noise in is $\protect\epsilon=5\% $. We have stopped at the 3rd mesh refinement for $N=3$. a) and b) Correct images. c) and d) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.7 and 1.6.*[]{data-label="example6"}](Figures/smooth4_N3_3d_5noise.png "fig:"){width="5cm"} ![*Results of Test 6. We test the reconstruction of the same function $c\left( \mathbf{x}\right) $ as in Test 4 (Figures 5) but with the noise in the data. The level of noise in is $\protect\epsilon=5\% $. We have stopped at the 3rd mesh refinement for $N=3$. a) and b) Correct images. c) and d) Computed images for $N=3$. The computed function $c\left( \mathbf{x}\right) $ in the inclusion varies between 0.7 and 1.6.*[]{data-label="example6"}](Figures/smooth4_N3_slice_5noise.png "fig:"){width="5cm"} \(c) $N=3$ \(d) $N=3$ --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ [99]{} <span style="font-variant:small-caps;">H. Ammari, J. Garnier, W. Jing, H. Kang, M. Lim, K. Solna, and H. Wang</span>, *Mathematical and Statistical Methods for Multistatic Imaging*, vol. 2098 of Lecture Notes in Mathematics, Springer, Cham, 2013. <span style="font-variant:small-caps;">H. Ammari, Y.T. Chow, and J. Zou</span>, *The concept of heterogeneous scattering and its applications in inverse medium scattering*, SIAM J. Math. Anal., 46 (2014), 2905-2935. <span style="font-variant:small-caps;">A. B. Bakushinskii, M. V. Klibanov, and N. A. Koshev</span>, *Carleman weight functions for a globally convergent numerical method for ill-posed Cauchy problems for some quasilinear PDEs*, Nonlinear Analysis: Real World Applications, 34 (2017), pp. 201–224. <span style="font-variant:small-caps;">L. Baudouin, M. de Buhan and S. Ervedoza</span>, *Convergent algorithm based on Carleman estimates for the recovert of a potential in the wave equation*, SIAM J. on Numerical Analysis, 55 (2017), 1578-1613. <span style="font-variant:small-caps;">L. Beilina and M. V. Klibanov</span>, *Approximate global convergence and adaptivity for coefficient inverse problems*, Springer, 2012. <span style="font-variant:small-caps;">M. Bellassoued and M. Yamamoto</span>, *Carleman Estimates and Applications to Inverse Problems for Hyperbolic Systems*, Springer Japan KK, 2017. <span style="font-variant:small-caps;">A. Bukhgeim and M. Klibanov</span>, *Uniqueness in the large of a class of multidimensional inverse problems*, Soviet Math. Doklady, 17 (1981), pp. 244–247. <span style="font-variant:small-caps;">F. Cakoni and D. Colton</span>, *Qualitative Methods in Inverse Scattering Theory. An Introduction*, Springer, Berlin, 2006. <span style="font-variant:small-caps;">G. Chavent</span>, *Nonlinear Least Squares for Inverse Problems - Theoretical Foundations and Step-by-Step Guide for Applications*, Springer, 2009. <span style="font-variant:small-caps;">D. Colton and R. Kress</span>, *Inverse Acoustic and Electromagnetic Scattering Theory*, Springer, New York, 3rd ed., 2013. <span style="font-variant:small-caps;">A. A. Duchkov and A. L. Karchevskii,</span> *Application of temperature monitoring to estimate the heat flux and thermophysical properties of bottom sediments*, Doklady Earth Sciences, 2014, Vol. 458, Part 2, pp. 1285–1288. <span style="font-variant:small-caps;">A.V. Goncharsky and S.Y. Romanov</span>, *Iterative methods for solving coefficient inverse problems of wave tomography in models with attenuation*, Inverse Problems, 33 (2017), 025003. <span style="font-variant:small-caps;">K. Ito, B. Jin, and J. Zou</span>, *A direct sampling method to an inverse medium scattering problem*, Inverse Problems 28 (2012), 025003. <span style="font-variant:small-caps;">K. Ito, B. Jin, and J. Zou</span>, *A direct sampling method for inverse electromagnetic medium scattering*, Inverse Problems 29 (2013), 095018. <span style="font-variant:small-caps;">S. I. Kabanikhin, A. D. Satybaev, and M.A. Shishlenin</span>, *Direct Methods of Solving Multidimensional Inverse Hyperbolic Problem*, VSP, Utrecht, 2004. <span style="font-variant:small-caps;">S.I. Kabanikhin, K.K. Sabelfeld, N.S. Novikov, and M.A. Shishlenin</span>, *Numerical solution of the multidimensional Gelfand-Levitan equation*, J. Inverse and Ill-Posed Problems 23 (2015), pp. 439-450. <span style="font-variant:small-caps;">M.V. Klibanov and O.V. Ioussoupova</span>, Uniform strict convexity of a cost functional for three-dimensional inverse scattering problem, SIAM J. Math. Anal., 26 (1995), pp. 147–179. <span style="font-variant:small-caps;">M.V. Klibanov</span>, *Global convexity in a three-dimensional inverse acoustic problem*, SIAM J. Math. Anal., 28 (1997), pp. 1371–1388. <span style="font-variant:small-caps;">M.V. Klibanov</span>, *Global convexity in diffusion tomography*, Nonlinear World, 4 (1997), pp. 247–265. <span style="font-variant:small-caps;">M. V. Klibanov and A. Timonov</span>, *Carleman Estimates for Coefficient Inverse Problems and Numerical Applications*, de Gruyter, Utrecht, 2004. <span style="font-variant:small-caps;">M. V. Klibanov</span>, *Carleman estimates for global uniqueness, stability and numerical methods for coefficient inverse problems*, J. Inverse and Ill-Posed Problems, 21 (2013), pp. 477–560. <span style="font-variant:small-caps;">M. V. Klibanov</span>, *Carleman weight functions for solving ill-posed Cauchy problems for quasilinear PDEs*, Inverse Problems, 31 (2015), 125007. <span style="font-variant:small-caps;">M. V. Klibanov and N. T. Th[à]{}nh</span>, *Recovering dielectric constants of explosives via a globally strictly convex cost functional*, SIAM J. Appl. Math., 75 (2015), pp. 518–537. <span style="font-variant:small-caps;">M. V. Klibanov, A. E. Kolesov, L. Nguyen, and A. Sullivan</span>, *Globally strictly convex cost functional for a 1-D inverse medium scattering problem with experimental data*, SIAM J. Appl. Math., 77 (2017), pp. 1733-1755. <span style="font-variant:small-caps;">M. V. Klibanov and A. E. Kolesov,</span> *Convexification of a 3-D coefficient inverse scattering problem*, Computers and Mathematics with Applications, published online, https://doi.org/10.1016/j.camwa.2018.03.016, 2018. <span style="font-variant:small-caps;">M. V. Klibanov, A. E. Kolesov, L. Nguyen, and A. Sullivan</span>, *A new version of the convexification method for a 1D coefficient inverse problem with experimental data*, Inverse Problems, 34 (2018), 115014. <span style="font-variant:small-caps;">M. V. Klibanov, A. E. Kolesov, D.-L. Nguyen,</span> *Convexification method for an inverse scattering problem and its performance for experimental backscatter data for buried targets*, arXiv: 1805.07618, 2018. <span style="font-variant:small-caps;">M.V. Klibanov, J. Li, and W. Zhang,</span> Electrical impedance tomography with restricted Dirichlet-to-Neumann map data, arXiv:1803.11193, 2018. <span style="font-variant:small-caps;">A. Lakhal</span>, *A decoupling-based imaging method for inverse medium scattering for Maxwell’s equations*, Inverse Problems, 26 (2010), 015007. <span style="font-variant:small-caps;">J. Li and J. Zou</span>, *A multilevel model correction  method for parameter identification*, Inverse Problems 23 (2007), 1759. <span style="font-variant:small-caps;">J. Li, H. Liu, and Q. Wang</span>, *Enhanced multilevel linear sampling methods for inverse scattering problems*, J. Comput. Phys., 257 (2014), pp. 554–571. <span style="font-variant:small-caps;">J. Li, P. Li, H. Liu, and X. Liu</span>, *Recovering multiscale buried anomalies in a two-layered medium*, Inverse Problems, 31 (2015), 105006. <span style="font-variant:small-caps;">G. Rizzuti and A. Gisolf</span>, *An iterative method for 2D inverse scattering problems by alternating reconstruction of medium properties and wavefields: theory and application to the inversion of elastic waveforms*, Inverse Problems 33 (2017), 035003. <span style="font-variant:small-caps;">V.G. Romanov</span>, *Inverse Problems of Mathematical Physics*, VNU Press, Utrecht, 1986. <span style="font-variant:small-caps;">V.G. Romanov,</span> *Inverse problems for differential equations with memory*, Eur. J. Math. Comput. Appl., 2 (2014), pp. 51–80. <span style="font-variant:small-caps;">A.N. Tikhonov, A.V. Goncharsky, V.V. Stepanov and A.G. Yagola</span>, *Numerical Methods for the Solution of Ill-Posed Problems*, Kluwer, London, 1995. <span style="font-variant:small-caps;">M.M. Vajnberg</span>, *Variational Method and Method of Monotone Operators in the Theory of Nonlinear Equations*, John Wiley& Sons, Washington, DC, 1973. [^1]: Department of Mathematics and Statistics, University of North Carolina at Charlotte, Charlotte, NC 28223, USA ([email protected]) [^2]: Corresponding author. Department of Mathematics, Southern University of Science and Technology (SUSTech), 1088 Xueyuan Boulevard, University Town of Shenzhen, Xili, Nanshan, Shenzhen, Guangdong Province, P.R.China ([email protected]) [^3]: Department of Mathematics, Southern University of Science and Technology (SUSTech), 1088 Xueyuan Boulevard, University Town of Shenzhen, Xili, Nanshan, Shenzhen, Guangdong Province, P.R.China ([email protected]) [^4]: Submitted to the editors DATE.

ArXiv

--- abstract: 'We study the feasibility of a long–baseline neutrino experiment from [[CERN]{}]{} to Gran Sasso LNGS Laboratories using the [[CERN]{}]{} PS accelerator. Baseline and neutrino energy spectrum are suitable to explore a region of the (${{\Delta m}^2}$, ${\sin^2 2\theta}$) parameters space which is not reached by K2K, the first experiment that will test at accelerator the atmospheric neutrino anomaly put in evidence by [Super–Kamiokande]{}.' author: - | P. F. Loverre, R. Santacesaria, F. R. Spada\ [*Università “La Sapienza” and Istituto Nazionale di Fisica Nucleare (INFN)*]{}\ [*Rome, Italy*]{}\ –\ [Submitted to *The European Physical Journal C*]{} title: 'About testing ${\nu_{\mu}}$ oscillation with ${{\Delta m}^2}$ smaller than 0.001 ${{\mathrm{eV}}^2}$ with the [[CERN]{}]{} Proton Synchrotron' --- The recent [Super–Kamiokande]{} measurements of atmospheric neutrino fluxes [@skam:1998] favour ${\nu_{\mu}}\rightarrow {\nu_\tau}$ (or ${\nu_{\mu}}\rightarrow {\nu_x}$) oscillations, with almost maximal mixing and ${{\Delta m}^2}$ in the range $(5 \div 60) \cdot 10^{-4} \, {{\mathrm{eV}}^2}$. The first test of this atmospheric neutrino anomaly at accelerator will be performed in Japan by the K2K [@k2k:1997] experiment. K2K has recently started taking data using a neutrino beam generated by the KEK 12–GeV Proton Synchrotron directed toward the [Super–Kamiokande]{} detector, which is placed about 250 Km away from KEK. The K2K experiment, owing to an L/E ratio of order 250/1 (Km/GeV), explores via disappearance ${\nu_{\mu}}$ oscillations down to ${{\Delta m}^2}\sim 2 \cdot 10^{-3} \, {{\mathrm{eV}}^2}$. The same ${{\Delta m}^2}$ region can be explored with higher sensitivity with the high energy neutrino beams of [[FNAL]{}]{} (NuMI – MINOS experiment [@minos:web; @ables:1995]) and [[CERN]{}]{} ([[CERN]{}]{} – Gran Sasso LNGS beam [@cern-gs:1998]). The higher energy will also provide the opportunity for a direct search of ${\nu_\tau}$ charged current interactions. However, if the K2K experiment will not find any attenuation or distortion of the ${\nu_{\mu}}$ flux, it will become of primary importance to perform an accelerator experiment capable to extend the oscillation search to lower values of ${{\Delta m}^2}$, down to the limit ${{\Delta m}^2}= 5 \cdot 10^{-4} \, {{\mathrm{eV}}^2}$ of the region suggested by the atmospheric measurements of [Super–Kamiokande]{}. The use of a low energy neutrino beam from the Cern Protosynchrotron (PS) allows this measurement. The importance of an experiment based on a neutrino beam from the PS has already been underlined by F. Dydak et al. [@dydak:nufact]. In that paper, however, the accent has been put on a long term, detailed exploration of the oscillation mechanism; the project would be based, already in a first phase, on a major upgrade of the PS machine, allowing to improve the intensity by a factor 100. In our opinion, the existing machine could directly be exploited to test ${{\Delta m}^2}$ values down to $5 \cdot 10^{-4} \, {{\mathrm{eV}}^2}$. The intensity of the PS is in fact high enough to make possible a long–baseline, [[CERN]{}]{} – Gran Sasso, ${\nu_{\mu}}$ disappearance experiment. A neutrino beam from the PS would have an energy similar to that of the K2K beam, while the baseline of the PS experiment would be 732 Km instead of 250 Km. If it is possible to collect at Gran Sasso a statistics similar to that of the K2K experiment, the factor 3 in distance converts directly in a sensitivity to 3 times smaller ${{\Delta m}^2}$, i.e. ${{\Delta m}^2}\simeq 6~\cdot~10^{-4} \, {{\mathrm{eV}}^2}$. In a relatively short term, the only other project which could allow to study oscillations with ${{\Delta m}^2}$ below 0.001 ${{\mathrm{eV}}^2}$ is the use by MINOS of a low energy beam, the PH2(low) option described in ref. . We have not made a detailed comparison of that beam with the PS beam. We just note that the characteristics of the two beams are quite different. The Fermilab low energy beam has higher intensity, but the energy spectrum is peaking at somewhat higher energies (2–3 GeV compared to the 1.2 GeV of the PS beam) with significant contributions up to 50 GeV. These features might complicate the analysis of the data with respect to the case of the PS beam. Our calculations of neutrino flux, event rates and corresponding sensitivity to neutrino oscillations for a PS [[CERN]{}]{} – Gran Sasso experiment are presented in the following. We have simulated the PS neutrino beam in the setup used by the BEBC – PS 180 [@bebc:1986] experiment (proton energy of 19.2 GeV and horn focusing). Assuming a decay tunnel length of 300 m, we obtain at a distance of 732 Km the ${\nu_{\mu}}$ flux displayed in fig. \[fig:flux\] for $10^{21}$ protons on target. An experiment designed to search for ${\nu_{\mu}}$ disappearance would have to measure the absolute rate and the energy spectrum of ${\nu_{\mu}}$ charged current interactions, and compare them with the prediction based on the measurement performed in a similar detector located close to the neutrino source. In our evaluation of the sensitivity to oscillation of the experiment, the event rates have been calculated assuming, as in ref. , a cross section of $0.85 \cdot 10^{-38} \, \mathrm{cm}^2$ for the quasi–elastic reaction ${\nu_{\mu}}\, n \rightarrow \mu^- \, p$. In the simulation, ${\nu_{\mu}}\, n \rightarrow \mu^- \, p$ reactions represent $50 \%$ of the total number of charged current interactions. The effect of a finite experimental resolution was taken into account by imposing the energy cut $E_\nu > 0.6$ GeV. We assume that with this cut it will be possible to identify and measure events with a $\mu$ in the final state with 100% efficiency for quasi–elastic interactions and with 50% efficiency for the remaining charged current interactions. Event rates were then computed for - a total flux of $1.0 \cdot 10^{21}$ protons on target (POTs) - a fiducial mass of 20 Ktons for the detector. We will justify these two assumptions later. The resulting energy spectrum of the interactions with a muon in the final state is shown in fig. \[fig:mu\]. The histogram contains 337 events. If oscillations occurred, their effect would be a reduction of the number of interactions and, for ${{\Delta m}^2}< 0.01 \, {{\mathrm{eV}}^2}$, a distortion of the energy spectrum. We assume that the total rate at the [*far*]{} detector can be predicted with a 10% systematic error from a measurement at the [*close*]{} detector. Fig. \[fig:osc\] then shows the 90% $C.L.$ exclusion curve which could be obtained in absence of oscillation, adding quadratically systematic and statistical errors. The whole region of the parameter space suggested by the Kamiokande and [Super–Kamiokande]{} measurements would be covered. Of course, the experiment should aim at detecting oscillation with more than 1.28 $\sigma$ significance. This will be possible because at low ${{\Delta m}^2}$  the oscillation would also produce a distortion of the energy spectrum. As an example, the inset in fig. \[fig:mu\] shows the energy spectrum of the interacting ${\nu_{\mu}}$ neutrinos for the case of a ${\nu_{\mu}}\rightarrow {\nu_\tau}$ oscillation with ${\sin^2 2\theta}=1$ and ${{\Delta m}^2}=7 \cdot 10^{-4} \, {{\mathrm{eV}}^2}$. With respect to the no-oscillation spectrum, there are 76 events less, and 50 of them are missing from the first three bins of the histogram, yielding an evidence for oscillation of more than three standard deviations. We conclude that, in spite of the limited statistics, the experiment would be capable to detect oscillations in the whole range of the (${{\Delta m}^2}$, ${\sin^2 2\theta}$) parameters suggested by [Super–Kamiokande]{}. The experiment will need a very large [*far*]{} detector, a similar, but much smaller [*close*]{} detector, and a very intense use of the PS. Our calculations assumed an integrated flux corresponding to $1.0 \cdot 10^{21}$ POTs. Presently, the radiation constraints limit the PS intensity to an average value of $1.2 \cdot 10^{13}$ protons on target per second; therefore, four years of data taking with a detector of 20 Ktons would be required. This is a very heavy but achievable task; a better investigation of the beam design – optimisation of beam energy, target and horn design – may well lead to a substantial improvement of the neutrino flux and to a consequent reduction of running time and/or detector mass. If no evidence for oscillation will be found by the K2K experiment, the discussed option deserves great attention. [99]{} Y. Fukuda et al., Phys. Rev. Lett. **81** (1998), 1562–1567. K. Nishikawa, Nucl. Phys. B (Proc. Supp.) **59** (1997), 289–296. The MINOS Collaboration, Tech. report, NuMI–L–337TDR, 1998. E. Ables et al., FERMILAB Proposal P–875, 1995. Tech. report, CERN/98–02 and INFN/AE–98/05, May 1998. F. Dydak et al., CERN–SPSC/98–30, October 1998. C. Angelini et al., Phys. Lett. B **179** (1986), 307. N. Armenise et al., Letter of Intent, CERN–SPSC/97–21, October 1997. S. Hatakeyama et al., Phys. Rev. Lett. **81** (1998), 2016–2019.

ArXiv

--- abstract: 'New approach to exact solvability of dilaton gravity theories is suggested which appeals directly to structure of field equations. It is shown that black holes regular at the horizon are static and their metric is found explicitly. If a metric possesses singularities the whole spacetime can be divided into different sheets with one horizon on each sheet between neighboring singularities with a finite value of dilaton field (addition horizons may arise at infinite value of it), neighboring sheets being glued along the singularity. The position of singularities coincide with the values of dilaton in solutions with a constant dilaton field. Quantum corrections to the Hawking temperature vanish. For a wide subset of these models the relationship between the total energy and the total entropy of the quantum finite size system is the same as in the classical limit. For another subset the metric itself does not acquire quantum corrections. The present paper generalizes Solodukhin’s results on the RST model in that instead of a particular model we deal with whole classes of them. Apart from this, the found models exhibit some qualitatively new properties which are absent in the RST model. The most important one is that there exist quantum black holes with geometry regular everywhere including infinity.' address: | Department of Physics, Kharkov State University, Svobody Sq.4, Kharkov\ 310077, Ukraine\ E-mail: [email protected] author: - 'O. B. Zaslavskii' title: 'Exactly solvable models of two-dimensional dilaton gravity and quantum eternal black holes' --- introduction ============ Two-dimensional theories of gravity count on better understanding the role of quantum effects in black hole physics in a more realistic four-dimensional case. In particular, new insight was gained due to reducing the problem of black hole evaporation to solving differential equations of the semiclassical theory [@callan]. However, these equations remain too complicated in the sense that exact solutions cannot be found. This obstacle was overcome in the approach based on modifying the form of the semiclassical action in such a way that solvability is restored. The well-known example is the Russo-Susskind-Thorlacius (RST) model [@rst]. In particular, this enabled one to give exhausting analysis of either geometry or thermodynamics of eternal black holes in the framework of RST dilaton gravity [@solod]. Examples of exactly solvable theories of dilaton gravity were discussed in [@bil], [@alw]. As was shown by Kazama, Satoh and Tsuichiya (KST), these models as well as the RST one can be found as particular cases of the unified approach [@kaz] based on the symmetries of the nonlinear sigma model. A number of other exactly solvable models were suggested later [@rob], [@fub], [@bose], [@cruz]. The aim of the present paper is two-fold. First, we suggest new approach to finding criteria for exact solvability and establish its equivalence to that of [@kaz]. We also show that our approach encompasses all particular models mentioned above. Second, we solve the field equations in a closed form and study general properties of the found spacetimes. In this point we generalize the Solodukhin’s results for RST eternal black holes. The essential feature of models considered below consists in that we specify not the explicit dependence of the action coefficients on dilaton but, rather, relationship between these coefficients. This means that instead of one or several particular exactly solvable models we deal at once with the whole class of them. The paper is organized as follows. In Sec. II we describe our approach and derive the basic equation which selects exactly-solvable models of two-dimensional dilaton gravity among all possible forms. We find the relation between action coefficients and demonstrate how it enables one to cast the theory into the Liouville form. We also briefly show that the this relation is equivalent to that obtained in [@kaz] in a quite different approach. Further, we prove that all black holes with a regular geometry are static and find their metric explicitly. In Sec. III we discuss properties of found solutions - spacetime structure, thermodynamics, the existence of a special class of solutions with a constant dilaton field value, etc. In Sec. IV we show that all exactly solvable models mentioned above fall in the found class. In Sec. V we summarize the main features of our approach and properties of obtained solutions and outline some perspectives for future researches. basic equations =============== Conditions of exact solvability ------------------------------- Let us consider the system described by the action $$I=I_{0}+I_{PL} \label{1}$$ where $$I_{0}=\frac{1}{2\pi }\int_{M}d^{2}x\sqrt{-g}[F(\phi )R+V(\phi )(\nabla \phi )^{2}+U(\phi )]+\frac{1}{\pi }\int_{\partial M}dskF(\phi ) \label{2}$$ Here the boundary term with the second fundamental form $k$ makes the variational problem self-consistent, $ds$ is the line element along the boundary $\partial M$ of the manifold $M.$ $I_{PL\text{ }}$is the Polyakov-Liouville action [@polyakov] incorporating effects of Hawking radiation and its back reaction on the black hole metric for a multiplet of N scalar fields. It is convenient to write it down in the form [@solod], [@israel] $$I_{PL}=-\frac{\kappa }{2\pi }\int_{M}d^{2}x\sqrt{-g}[\frac{(\nabla \psi )^{2}% }{2}+\psi R]-\frac{\kappa }{\pi }\int_{\partial M}\psi kds \label{3}$$ The function $\psi $ obeys the equation $$\Box \psi =R \label{4}$$ where $\Box =\nabla _{\mu }\nabla ^{\mu }$, $\kappa =N/24$ is the quantum coupling parameter. We imply that all coefficients $F,V,U$ in (\[2\]) may, in general, contain terms with $\kappa .$ For instance, in the RST model $F=e^{-2\phi }-\kappa \phi .$ Varying the action with respect to metric gives us $(T_{\mu \nu }=2\frac{% \delta I}{\delta g^{\mu \nu }})$: $$T_{\mu \nu }\equiv T_{\mu \nu }^{(0)}+T_{\mu \nu }^{(PL)}=0 \label{6}$$ where $$T_{\mu \nu }^{(0)}=\frac{1}{2\pi }\{2(g_{\mu \nu }\Box F-\nabla _{\mu }\nabla _{\nu }F)-Ug_{\mu \nu }+2V\nabla _{\mu }\phi \nabla _{\nu }\phi -g_{\mu \nu }V(\nabla \phi )^{2}\}\text{,} \label{7}$$ $$T_{\mu \nu }^{(PL)}=-\frac{\kappa }{2\pi }\{\partial _{\mu }\psi \partial _{\nu }\psi -2\nabla _{\mu }\nabla _{\nu }\psi +g_{\mu \nu }[2R-\frac{1}{2}% (\nabla \psi )^{2}]\} \label{8}$$ Variation of the action with respect to $\phi $ gives rise to the equation $$RF^{^{\prime }}+U^{\prime }=2V\Box \phi +V^{\prime }(\nabla \phi )^{2}\text{,% } \label{9}$$ where prime denotes derivative with respect to $\phi $. In general, eqs. (\[4\]) - (\[9\]) cannot be solved exactly. In particular, the auxiliary function $\psi $ is a rather complicated nonlocal functional of dilaton field $\phi $ and its derivatives, the explicit form of which cannot be found. The key idea for finding exactly solvable models consists in two assumptions. First, we select among all variety of models such a subset for which the above equations admit the solution $\psi =\psi (\phi )$, i.e. a local connection between $\psi $ and $\phi .$ As was shown by Solodukhin [@solod], for eternal black holes in the RST model $\psi =-2\phi .$ Now the dependence in question can be, generally speaking, nonlinear. Second, we impose such a constraint on the coefficients of the action that enables us to single out the terms with derivatives of $\phi $ with respect to coordinates. As a result, instead of solving the original (rather complicated equations) it is sufficient to ensure the cancellation of the coefficients at $\Box \phi $ and $(\nabla \phi )^{2}$ which themselves already do not contain derivatives with respect to coordinates. It turns out that such a constraint is not very tight and includes a large variety of models which may be of physical interest. Inasmuch as the auxiliary function $\psi $ can be expressed in terms of $% \phi $ directly, the action $I_{0\text{ }}$and the Polyakov-Liouville action are combined in such a way that field equations (\[6\]) - (\[8\]) can be formally obtained from the action $I_{0}$ only but with the ”renormalized” coefficients which receive some shifts: $F\rightarrow \tilde{F},V\rightarrow \tilde{V}$ where $$\tilde{F}=F-\kappa \psi \text{,} \label{10}$$ $$\tilde{V}=V-\frac{\kappa }{2}\psi ^{\prime 2} \label{11}$$ Taking the trace of eq. (\[6\]) we get the relation $$U=\Box \tilde{F} \label{12}$$ Then eq. (\[9\]), with eqs. (\[4\]), (\[10\]) - (\[12\]) taken into account, reads $$\begin{aligned} &&A_{1}\Box \phi +A_{2}(\nabla \phi )^{2}=0\text{,} \nonumber \\ &&A_{1}=(u-\kappa \omega )\psi ^{\prime }+\omega u-2V\text{,} \label{13} \\ &&A_{2}=(u-\kappa \omega )\psi ^{\prime \prime }+\omega u^{\prime }-V^{\prime }\text{,} \nonumber\end{aligned}$$ where $\omega \equiv U^{\prime }/U,$ $u\equiv F^{\prime }.$ For arbitrary coefficients $A_{1}(\phi )$, $A_{2}(\phi )$ eq.(\[13\]) cannot be solved exactly. This can be done, however, if we demand that both coefficients in eq.(\[13\]) turn into zero. Such a demand represents the [*sufficient* ]{}condition for eq.(\[13\]) to be satisfied. Then it follows from eq. $% A_{1}=0$ that $$\psi ^{\prime }=\frac{2V-\omega u}{u-\kappa \omega }\text{,} \label{14}$$ which enables us to find at once $\psi $ in terms of known functions $u$, $V$, $\omega $ by direct integration. Demanding that both equations $A_{1}=0$ and $A_{2}=0$ be consistent with each other, we differentiate eq. $A_{1}=0$ and compare the result with $A_{2}=0$. Then we have $$u^{\prime }(2V-\omega u)+u(u\omega ^{\prime }-V^{\prime })+\kappa (\omega V^{\prime }-2V\omega ^{\prime })=0 \label{15}$$ Thus, we made some selections among all possible models. Eq.(\[15\]) is the only constraint on the relationship between these three coefficients that leaves enough freedom in choosing a model. Both eqs.(\[14\]), (\[15\]) represent direct consequences of our assumptions $\psi =\psi (\phi )$, $A_{1}=A_{2}=0$ and do not hold true in an arbitrary model. In particular, for the original form of the action in the string-inspired gravity [@string] $F=e^{-2\phi },V=4e^{-2\phi },\omega =-2$ these equation are satisfied in the zero order in $\kappa $ only. However, for the RST action $$F=e^{-2\phi }-\kappa \phi ,V=4e^{-2\phi },\omega =-2 \label{model}$$ and this model does obey eq.(\[15\]). Then the integration of (\[14\]) leads to the relationship $\psi =-2\phi +const$ that agrees with [@solod]. It follows from eq. (\[15\]) that the function $\omega $ expressed in terms of $V$ and $u$ reads $$\omega =\frac{u-D\sqrt{u^{2}-2V\kappa }}{\kappa } \label{w}$$ Reverting this formula, we have $$V=\omega (u-\frac{\kappa \omega }{2})+C(u-\kappa \omega )^{2} \label{v}$$ where $C=(2\kappa )^{-1}(1-D^{-2})$. We must assume that the constant $% D\rightarrow 1$ when $\kappa \rightarrow 0$ in order to have the well-defined classical limit. In what follows we mainly restrict ourselves to the simplest case $C\equiv 0$, $D\equiv 1$, when $$\omega =\frac{u-\sqrt{u^{2}-2V\kappa }}{\kappa }\text{, }V=\omega (u-\frac{% \kappa \omega }{2}) \label{d=1}$$ Comparison with KST action -------------------------- In eq. (3.18) of Ref.[@kaz] the following relation between different action coefficients was obtained: $$a_{,q}(1-\frac{\kappa }{2}W_{,q})+\kappa W_{,qq}(a+\frac{1}{2}W_{,q})=0 \label{k}$$ where (in our notations) $W=\int d\phi \omega $ and by definition $% V=aF^{\prime 2}+F^{\prime }\omega \equiv au^{2}+u\omega $. Substituting these expressions into (\[k\]) we obtain after simple but rather lengthy calculation the equation which coincides with our eq. (\[15\]) exactly. It was pointed out in [@kaz] that results of [@bil], [@alw] are contained in the general formula following from (\[k\]). In our scheme the models of Ref. [@alw] (which include those of [@bil] as particular cases) follow from (\[d=1\]) if one writes down $V=4e^{-2\phi }[1+h(\phi )] $, $u=-2e^{-2\phi }[1+\tilde{h}]$. KST approach has the advantage of elucidating the hidden symmetry of the action in terms of a nonlinear $\sigma $ model. On the other hand, the present approach, in our view, is much simpler in that it operates directly with the original action coefficients and demonstrates explicitly the origin of solvability as the cancellation of coefficients at $(\nabla \phi )^{2}$ and $\Box \phi $ in (\[13\]). It also enables us to establish the staticity of black hole solutions (in the absence of external origins and matter fields) and find their explicit form. It is this issue that we now turn to. Form of the metric ------------------ Let us return to general formulae without specifying the gauge. With eq. (\[12\]) taken into account, the field equations (\[6\]) - (\[8\]) can be rewritten in the form $$\lbrack \xi _{1}\Box \phi +\xi _{2}(\nabla \phi )^{2}]g_{\mu \nu }=2(\xi _{1}\nabla _{\mu }\nabla _{\nu }\phi +\xi _{2}\nabla _{\mu }\phi \nabla _{\nu }\phi ) \label{16}$$ where $\xi _{1}=\tilde{F}^{\prime }$, $\xi _{2}=\tilde{F}^{\prime \prime }-% \tilde{V}$. Let us multiply this equation by the factor $\eta $ chosen in such a way that $(\xi _{1}\eta )^{\prime }=\xi _{2}\eta $ whence $\eta =\exp (-\int d\phi \tilde{V}/\tilde{F}^{\prime })$. Then eq. (\[16\]) turns into $$g_{\mu \nu }\Box \mu =2\nabla _{\mu }\nabla _{\nu }\mu \label{17}$$ where by definition $\mu ^{\prime }=\xi _{1}\eta .$ This equation takes the same form as eq.(2.24) from [@solod] and entails the same conclusions about properties of the geometry. It is convenient to choose the space-like coordinate as $x=\mu (\phi )/B$ where $B$ is some constant. Then it follows from eq.(\[17\]) that the metric takes the Schwarzschild-like form and is static: $$ds^{2}=-gdt^{2}+g^{-1}dx^{2} \label{18}$$ In this gauge $R=-\frac{d^{2}g}{dx^{2}}$ . Substituting it into eq.(\[4\]) we get after integration: $$g=A\int_{\phi _{h}}^{\phi }d\tilde{\phi}\frac{\partial \mu }{\partial \tilde{% \phi}}e^{\psi (\tilde{\phi})-\psi (\phi )} \label{19}$$ $$\begin{aligned} x &=&B^{-1}\mu \text{, }\mu =\int^{\phi }d\tilde{\phi}\frac{\partial \tilde{F% }}{\partial \tilde{\phi}}e^{\int^{\tilde{\phi}}d\phi ^{\prime }\alpha (\phi ^{\prime })}\text{,} \nonumber \\ \alpha &=&-\frac{\tilde{V}}{\tilde{F}^{\prime }}\text{,} \label{20}\end{aligned}$$ where $A$ is a constant. Here it is supposed that a spacetime has an event horizon at $x=x_{h}$, $\phi =\phi _{h}$. Now we will see that the obtained expressions for the metric can be simplified further. Substituting eq. (\[v\]) into (\[19\]), (\[20\]) we find after some rearrangement that $\alpha =-[\omega +C(u-\kappa \omega )] $ and $$\begin{aligned} g &=&aC^{-1}(e^{CH}-e^{CH_{h}})e^{-\psi }\text{, }a\equiv AC(1-2\kappa C)% \text{, }\psi =(1-2\kappa C)\int d\phi \omega +2CF\text{, } \label{c} \\ x &=&B^{-1}\mu \text{, }\mu ^{\prime }=(1-2\kappa C)H^{\prime }\exp [-CF-(1-\kappa C)\int d\phi \omega ]\text{, }H\equiv F-\kappa \int d\phi \omega \nonumber\end{aligned}$$ It would seem that eqs. (\[c\]) do not represent the full solution of the problem since the functions under discussion should obey one more equation - eq. (\[12\]). It is remarkable, however, that, as follows from the substitution of (\[c\]) into (\[12\]), this equation is satisfied provided the constants entering formulae for the metric obey the relation $$aB^{2}e^{CH_{h}}(1-2\kappa C)=4\lambda ^{2} \label{const}$$ where $U=4\lambda ^{2}\exp (\int d\phi \omega )$ and the limit of integration in the integral $\int^{\phi }d\phi \omega $ is chosen equally in all formulae above. In the case $C=0$ this gives us $$g=(\tilde{F}-\tilde{F}_{h})\exp [-\int d\tilde{\phi}\omega ]\text{, }\mu ^{\prime }=\tilde{F}^{\prime }e^{-\psi }\text{, }\tilde{F}=H\text{, }\psi =\int \omega d\phi \text{, }B^{2}=4\lambda ^{2} \label{g}$$ where we put $a=1$ for definiteness. For the RST model (\[model\]) we obtain in accordance with [@solod] $\mu =\phi -\frac{\kappa }{4}e^{2\phi }$, $B=-2\lambda $. In the above formulae it was tacitly assumed that the function $\psi $ as well as the metric itself is regular at the horizon (when the temperature is equal to its Hawking value). As explained in detail in Ref. [@solod] (Sec. 2B), such a choice of boundary conditions corresponds to a black hole in equilibrium with Hawking radiation but does not describe formation of a hole from a flat space due to incoming matter. Liouville theory ---------------- The suggested approach enables one to cast the original action into the Liouville form in a very simple way. In the conformal gauge $$ds^{2}=-e^{2\rho }dx^{+}dx^{-} \label{conf}$$ Bearing in mind that in this gauge the Polyakov action [@polyakov] reads $I_{PL}=-\frac{2\kappa }{\pi }\int d^{2}x\partial _{+}\rho \partial _{-}\rho =\frac{\kappa }{\pi }\int d^{2}x\sqrt{g}(\nabla \rho )^{2}$ and integrating in (\[2\]) the term with curvature by parts we have (we omit now all boundary terms) $$I=\frac{1}{2\pi }\int d^{2}x\sqrt{g}[V(\nabla \phi )^{2}+2\nabla \rho \nabla F+2\kappa (\nabla \rho )^{2}+4\lambda ^{2}e^{\eta }]\text{,} \label{i}$$ where by definition $\eta =\int d\phi \omega $. Let now the potential $V$ obey the condition (\[v\]). Then the action can be rewritten in the form $$I=\frac{1}{2\pi }\int d^{2}x\sqrt{g}[\nabla \eta \nabla F-\frac{\kappa }{2}% (\nabla \eta )^{2}+C(\nabla H)^{2}+2\nabla \rho \nabla F+2\kappa (\nabla \rho )^{2}+4\lambda ^{2}e^{\eta }] \label{i2}$$ where $H=F-\kappa \eta $. Now introduce new fields $\Omega $ and $\chi $ instead of $\phi $ and $\rho $ according to $H=2\kappa \Omega $, $\eta (\phi )=2(\chi -\Omega -\rho )$. Then after simple rearrangement we obtain $$I=\frac{1}{\pi }\int d^{2}x\sqrt{g}\{\kappa [(\nabla \chi )^{2}-(\nabla \Omega )^{2}(1-2C\kappa )]+2\lambda ^{2}e^{2(\chi -\Omega -\rho )}\} \label{il}$$ Using the explicit formula for the conformal gauge (\[conf\]) we obtain the action $$I=\frac{1}{\pi }\int d^{2}x\{2\kappa [(\partial _{+}\Omega \partial _{-}\Omega (1-2C\kappa )-\partial _{+}\chi \partial _{-}\chi ]+\lambda ^{2}e^{2(\chi -\Omega )}\} \label{li}$$ which in the case $C=0$ takes the familiar form. Meanwhile, it is instructive to rederive in the conformal gauge the obtained exact solutions with an arbitrary $C\neq 0$. In this gauge the equations of motion which follow from (\[li\]) read $$\begin{aligned} 2\kappa (1-2C\kappa )\partial _{+}\partial _{-}\Omega &=&-\lambda ^{2}e^{2(\chi -\Omega )}\text{,} \label{motion} \\ 2\kappa \partial _{+}\partial _{-}\chi &=&-\lambda ^{2}e^{2(\chi -\Omega )} \nonumber\end{aligned}$$ As usual, they should be supplemented by the constraint equations $% T_{++}=T_{--}=0$. The expressions for the classical part of $T_{++\text{ }}$ and $T_{--}$ follow directly from (\[7\]). The formula for the quantum contribution can be obtained from the conservation law and conformal anomaly and has the form $$T_{\pm \pm }^{(PL)}=\frac{2\kappa }{\pi }[(\partial _{\pm }\rho )^{2}-\partial _{\pm }^{2}\rho +t_{\pm }] \label{stress}$$ The functions $t_{\pm }$ depend on the choice of boundary conditions. For the Hartle-Hawking state we deal with $t_{\pm }=0$. With the potential $V$ from (\[v\]) the constraint equations take the form $$-\partial _{\pm }^{2}F+2\partial _{\pm }\rho \partial _{\pm }\rho +(\partial _{\pm }\phi )^{2}V+2\kappa [(\partial _{\pm }\rho )^{2}-\partial _{\pm }^{2}\rho ]=0 \label{mot+}$$ As seen from eq.(\[motion\]), it is convenient to use the gauge in which $% \chi =\Omega (1-2C\kappa )$, so $\chi -\Omega =-CH$. Then after some algebra we find that the factor $(1-2C\kappa )$ in a remarkable way is singled out in constraints and we arrive at equations obtained from (\[mot+\]) and (\[motion\]) $$\begin{aligned} \partial _{\pm }^{2}H+C(\partial _{\pm }H)^{2} &=&0\text{,} \label{eqh} \\ (1-2\kappa C)\partial _{+}\partial _{-}H &=&-\lambda ^{2}e^{-2CH} \nonumber\end{aligned}$$ It is convenient to make the substitution $z=e^{C(H-H_{h})}$ where $H_{h}$ is some constant. Then the first equation in (\[eqh\]) turns into $% \partial _{\pm }^{2}z=0$ whence $z=bx^{+}x^{-}+d$ (linear terms can always be removed by shifts in coordinates). By substitution in the second eq.(\[eqh\]) we obtain the restriction $bd=-\tilde{\lambda}^{2}$ where $\tilde{% \lambda}^{2}=\lambda ^{2}C(1-2\kappa C)^{-1}e^{-2CH_{h}}$. Let us choose $% d=1 $, $b=-\tilde{\lambda}^{2}$ and make transformation to new coordinates according to $\tilde{\lambda}x^{+}=e^{\tilde{\lambda}\sigma ^{+}}$, $-\tilde{% \lambda}x^{-}=-e^{-\tilde{\lambda}x^{-}}$. Then the metric (\[conf\]) turns into $$ds^{2}=-gd\sigma ^{+}d\sigma ^{-} \label{sigma}$$ where $$g=e^{-\psi }\frac{(e^{CH}-e^{CH_{h}})}{e^{CH_{h}}}\text{, }\psi =(1-2\kappa C)\int d\phi \omega +2CF \label{met}$$ To rewrite the metric in the Schwarzschild gauge (\[18\]), let us introduce coordinates $\sigma ^{\pm }=t\pm \sigma $ and $d\sigma =dxg^{-1}$. Then $$x^{\prime }(\phi )=\frac{CH^{\prime }}{2\tilde{\lambda}}e^{-CH_{h}-CF-(1-% \kappa C)\int d\phi \omega } \label{coord}$$ . We see that (\[met\]), (\[coord\]) coincide with (\[c\]) if the constants are identified according to $a=Ce^{-CH_{h}}$, $B=2\tilde{\lambda}% C^{-1}(1-2\kappa C)e^{CH_{h}}$. properties of solutions ======================= Structure of spacetime ---------------------- Now we can make general conclusions about the structure of spacetime. Let us restrict ourselves by the case $C=0$. It follows from (\[g\]), (\[const\]) that the curvature $R=-\frac{d^{2}g}{dx^{2}}$ is equal to $$R=\frac{4\lambda ^{2}}{\tilde{F}^{\prime }}[\frac{\omega (\tilde{F}-\tilde{F}% _{h)}}{\tilde{F}^{\prime }}]^{\prime }\exp (\int^{\phi }d\tilde{\phi}\omega ) \label{r}$$ Let, by definition, $\tilde{F}^{\prime }(\phi _{c})=0$, $x_{c}=x(\phi _{c})$. Near $\phi _{c}$ the function $\tilde{F}^{\prime }\propto \phi -\phi _{c},x-x_{c}$ $\propto (\phi -\phi _{c})^{2}$ and, in general, $R\propto (\phi -\phi _{c})^{-3}\propto (x-x_{c})^{-3/2}$. The exceptional case arises when $\phi _{c}=\phi _{h}.$ Then the above expression in square brackets is finite and $R\propto (\phi -\phi _{c})^{-1}\propto (x-x_{c})^{-1/2}$. Thus, singularity becomes weaker but does not disappear. Such behavior, found earlier for the RST model [@solod], is inherent to any model under consideration. Thus, the metric possesses singularities in the points $\phi =\phi _{c}$ where $\tilde{F}^{\prime }=0.$ The spacetime splits into intervals between zeros of $\tilde{F}^{\prime }$ which can be viewed as different sheets that generalizes the corresponding feature of the RST model [@solod]. Within each of them the function $\tilde{F}^{\prime }$ does not alter its sign, so the function $\tilde{F}(\phi )$ is monotonic and the equation $\tilde{F}=% \tilde{F}_{h\text{ }}$has only one root. Then, according to (\[g\]), there is only one horizon at $\phi =\phi _{h}$ on every sheet between any two singularities with a finite value of $\phi $. In principle, it may happen that on a sheet between infinity and a singularity nearest to it there exists an additional horizon due to the factor $e^{-\psi }$ in which case the coordinate $x$ calculated according to (\[g\]) takes, generally speaking, a finite value in this limit. To obtain the maximally extended analytical continuation of spacetime, one is led to accept the possibility of complex dilaton field values [@accel]. We will not, however, discuss such possibilities further (a more detailed description of spacetime structure will be done elsewhere[@prep]). For the RST model there exist only two sheets but, depending on properties of the function $\tilde{F}(\phi )$, the number of sheets in a general case can be made arbitrary. Any two neighboring sheets are separated by the singularity located at $\phi =\phi _{c}$. As the singular points of the curvature represent zeros of the function $% \tilde{F}^{\prime }$, the case under consideration admits solutions regular everywhere, if the function $\tilde{F}^{\prime }$ does not turn into zero. Let, for instance, $F=$ $e^{\omega \phi }+\delta \kappa \omega \phi $ where $% \delta $ is a pure number and $\omega $ is constant. Then $\tilde{F}^{\prime }=\omega [e^{\omega \phi }+\kappa (\delta -1)]$. If $\delta >1$, the expression in square brackets changes its sign nowhere and does not tend to zero at infinity. It follows from eq. (\[g\]) that the coefficient $V$ for such solutions is everywhere positive, so the action is well-defined. As far as relationship between dilaton and coordinates is concerned, it follows from (\[g\]) that $\mu ^{\prime }=\omega +\omega \kappa (\delta -1)e^{-\omega \phi }$, so $x=(2\lambda )^{-1}[\omega \phi -\kappa (\delta -1)e^{-\omega \phi }]$ and $x\rightarrow \pm \infty $ when $\omega \phi \rightarrow \pm \infty $. The metric function $g=1+e^{-\omega \phi }[(\delta -1)\kappa \omega \phi -b]$, $b=[e^{\omega \phi }+\delta -1)\kappa \omega \phi ]_{\phi =\phi _{h}}$ has one zero at $\phi =\phi _{h}$, $g\rightarrow 1$ at $\omega \phi \rightarrow \infty $ and $g\rightarrow -\infty $ at $\omega \phi \rightarrow -\infty $. The curvature $R\simeq 4\lambda ^{2}[\kappa (\delta -1)]^{-1}e^{-\left| \omega \phi \right| }\rightarrow 0$ when $\omega \phi \rightarrow -\infty $ and $R\simeq -4\lambda ^{2}(\delta -1)\omega \kappa \phi e^{-\omega \phi }\rightarrow 0$ when $\omega \phi \rightarrow \infty $. Thus, not only the curvature is finite at both infinities but, moreover, the spacetime is flat there. If $\kappa \rightarrow 0$ we return to the string inspired dilaton gravity for which the singularity lies at $% \omega \phi \rightarrow -\infty $. In this sense, it is quantum effects which are responsible for removing the singularity. (On the other hand, one can obtain the geometry regular everywhere already on a classical level due to the choice, for example, $F=e^{\omega \phi }+K\omega \phi $ with $K>0$.) Following general rules [@walk], [@tom] we can draw the Penrose diagram of this nonsingular spacetime starting from fundamental building blocks. In so doing, the form of such a block depends crucially on whether or not $f(x)=\int^{x}dyg(y)^{-1}$remains finite at the boundary: it is triangle if $f$ is finite and square if $f$ diverges. In our case the function $f$ can be calculated exactly from (\[g\]): $f=(2\lambda )^{-1}\ln (\tilde{F}-\tilde{F}_{h})$. Collecting all this information we obtain the Penrose diagram which is depicted at Fig. \[pen1\]. The geodesic distance $\tau =\int dx(c-g)^{-1/2}$ where $c$ is a constant, so timelike geodesics can reach either on plus or minus infinity only for an infinite interval of time$.$ It is seen from Fig. \[pen1\] that a horizon is rather acceleration horizon than a true black hole one. The structure of spacetime is similar to that of Rindler (with extension to the complete Minkowski spacetime). In the limit $\omega \rightarrow 0$, $\delta \rightarrow \infty $ with $\omega \delta =const$ this spacetime turns into the Rindler one directly as it immediately follows from the above formulae for the metric and coordinate. It is worth noting that, as we shall see now, there exists also quite another type of solutions regular everywhere and having a black hole horizon: those for which $\tilde{F}^{\prime }\rightarrow 0$ at infinity in such a way that cancellation of $\tilde{F}^{\prime }$ is compensated by $% \exp (\int \omega d\phi )$, so that the whole expression (\[r\]) remains finite. Consider the following example. Let $\tilde{F}$ be monotonic function of $\phi $ such that $\tilde{F}\simeq e^{\gamma _{\pm }\phi }$ when $\phi \rightarrow \pm \infty $ with $\gamma _{\pm }<0$. Let also the function $\psi (\phi )=\int^{\phi }d\phi \omega $ have the asymptotics $\psi \simeq \omega _{\pm }\phi +\psi _{\pm }$ when $\phi \rightarrow \pm \infty $. Then at both infinities the curvature (\[r\]) behaves like $R\simeq A_{\pm }\exp [(\omega _{\pm }-2\gamma _{\pm })\phi ]$ where $A_{\pm }$ are constants. We choose $\omega _{+}-2\gamma _{+}<0$ and $\omega _{-}-2\gamma _{-}>0,$ so $R\rightarrow 0$ when $\phi \rightarrow \pm \infty $. Let us also choose, for definiteness, $\gamma _{-}=\omega _{-}$. Then $g\rightarrow 1$ when $\phi \rightarrow -\infty $ and $g\sim -e^{-\omega _{+}\phi }\rightarrow -\infty $ when $\phi \rightarrow \infty $. The coordinate $% x\simeq -2\lambda \left| \gamma _{-}\right| e^{-\psi _{-}}\phi \rightarrow \infty $ at $\phi \rightarrow -\infty $ and $x\simeq -\left| \gamma _{+}\right| (\gamma _{+}-\omega _{+})^{-1}e^{(\gamma _{+}-\omega _{+})\phi }\rightarrow -\infty $ at $\phi \rightarrow \infty $ if $\gamma _{+}>\omega _{+}$. Thus, the coordinate $-\infty <x<\infty $. All restrictions imposed on parameters of the solutions read $\omega _{+}<2\gamma _{+}<\gamma _{+}<0$ and $\gamma _{-}=\omega _{-}<0$ and are self-consistent. With all these properties, the Penrose diagrams looks like Fig. \[pen2\] where the horizontal lines represent not a singularity (as it would be for the Schwarzschild metric) but regular spatial infinities which can be reached along time-like geodesics only for a infinite proper time. The boundaries of the spacetime under discussion are null- and time-like complete [@tom], as $x\rightarrow \pm \infty $ and $\tau $ diverges. As a matter of fact, constructing both types of diagrams relies only on the asymptotic behavior of the functions $\tilde{F}$ and $\omega $, so they describe the whole classes of spacetimes. Now we will show that our metrics (\[g\]) contain also the third type of regular spacetimes - those with a constant curvature. Indeed, let $\tilde{F}=\exp (\frac{1}{2}\int_{0}^{\phi }\omega d\phi )$. Substituting this into (\[r\]) we find after simple manipulations that $R=8\lambda ^{2}\tilde{F}_{h}=const>0$, so spacetime in question is of de Sitter type. This generalizes the observation made in [@accel] for the particular choice $\omega =-2$ and $\kappa =0$. The metrics (\[g\]) corresponding to the choice $C=0$ possess the following interesting property. If the functions $\omega (\phi )$ and $% \tilde{F}(\phi )$, whatever their form would be, do not contain $\kappa $, the metric also does not contain $\kappa $. Thus, within one-loop accuracy, we obtained the whole classes of models for which a [*classical*]{} geometry is the [*exact*]{} solution of field equations derived from [*quantum*]{} Lagrangians. Solutions with a constant dilaton value --------------------------------------- Apart from solutions discussed above, there is one more class of them. It is seen from eq. (\[13\]) that this equation turns into identity when $\phi =const\equiv \phi _{0}$. For such solutions field eq.(\[12\]) gives us $% U=-\kappa R$. Substituting it into eq.(\[9\]) we have $R\tilde{F}^{\prime }=0$ where we have taken into account that $\tilde{F}^{\prime }=u-\kappa \omega $. This means that nontrivial solutions $(R\neq 0)$ exist only for values of the dilaton field $\phi _{0}=\phi _{c}$. Let me recall that this is just the point where the curvature for solutions described by eq.(\[g\] ) diverges. This gives nontrivial interplay between two branches of solutions, also found for the particular case of the RST model [@solod]: the values of the dilaton field for constant dilaton solutions coincides with the singularity of non-constant ones. In particular, it follows from the contents of the present paragraph that the class of models under discussion does not contain constant dilaton solutions with two horizons found in [@zasl98]. It is not surprising since the latter solutions exist only under the presence of an electromagnetic field which is now absent. Thermodynamics -------------- Let us now return to the solutions with $\phi \neq const$ and discuss their thermodynamic properties. From eqs. (\[c\]), (\[const\]) we obtain the formula for the Hawking temperature $T_{H}$ which turns out surprisingly simple: $$T_{H}=(4\pi )^{-1}(\frac{dg}{dx})_{x=x_{h}}=\frac{aB}{4\pi (1-2\kappa C)} \label{temp}$$ If the spacetime is not asymptotically flat, the choice of the Euclidean time is ambiguous and this is reflected in the appearance of the constants $% a $, $B$ in the metric (\[c\]) and temperature. More interesting is, however, the case when a spacetime is flat at infinity where there is a clear definition of time and a distant observer measures a temperature in an unambiguous way. Let us assume that $C=0$ and, for definiteness, the asymptotically flat region corresponds to $\phi \rightarrow -\infty $ and $% \tilde{F}\rightarrow \infty $, $\exp (-\int^{\phi }d\phi \omega )\rightarrow 0$ in such a way that their product is constant. Adjusting the limit of integration, one can always achieve $\lim_{\phi \rightarrow -\infty }\tilde{F% }\exp (-\int^{\phi }d\phi \omega )=a=1$ to have $g=1$ at infinity. Then it follows from (\[g\]), (\[temp\]) that the temperature in this case $$T_{H}=(2\pi )^{-1}\lambda \label{htemp}$$ and does not acquire quantum corrections. Moreover, as this temperature is a constant, it turns out that all horizons present in the solution have the same temperature. Both properties generalize the similar feature of the RST model [@solod]. In the latter case $\omega =const$, $\tilde{F}\propto e^{\omega \phi }$. It turns out that the general structure of the theories under consideration enables one to relate entropy and energy of a system in a rather simple manner. If the Hamiltonian constraint $T_{0}^{0}=0$ is taken into account one can infer that the action takes the thermodynamic form $I=\beta E-S$. This expression for the zero loop action can be obtained by direct generalization of the procedure elaborated in [@action] for the string-inspired dilaton gravity. Here $\beta =T_{H}^{-1}\sqrt{g}$ is the inverse Tolman temperature on the boundary, the energy $E$ $=E_{o}-(2\pi )^{-1}(\frac{\partial F}{\partial l})_{B}$ where $E_{o\text{ }}$is the subtraction constant irrelevant for our purposes, $dl$ is the line element, the index ”B” indicates that the corresponding quantities are calculated at the boundary. If quantum effects are taken into account the entropy entering this formula is to be understood as $S$ $=S_{0}+S_{q}$ where zero loop contribution $S_{0}=2F(\phi _{h})$ and that of Hawking radiation $% S_{q}=2\kappa [\psi (\phi _{B})-\psi (\phi _{h})]$ [@qent] (we choose the constant in the definition of entropy in such a way that $S_{q}=0$ in the limit $\phi _{B}=\phi _{h}$ when there is no room for radiation). For the total entropy we have $S=S_{0}+S_{q}=2[\tilde{F}(\phi _{h})+\kappa \psi (\phi _{B})]$. It follows directly from (\[g\]) with $a=1$, $B=2\lambda $ that $E-E_{0}=-% \frac{\lambda }{\pi }\exp (\frac{\psi _{B}}{2})\sqrt{F_{B}-S/2}$. The general first law $\beta \delta E=\delta S$ then tells us that the Hawking temperature $T_{H}=\lambda /2\pi $ does not acquire quantum correction in agreement with the above formula (\[htemp\]). It was pointed out in [@solod] that not only the temperature but the energy and entropy as well do not acquire quantum corrections in the RST model. Strictly speaking, however, the physical meaning of this statement is not quite clear since it refers to characteristics of a black hole itself and is based on subtracting contributions of a hot gas [@solod] whereas in the microcanonical ensemble approach the total energy (but not its constituents) should be fixed. Meanwhile, it is seen from the above formula directly that if the function $F(\phi )$ is such that it does not contain $\kappa $, the dependence of the energy on total entropy retains the same form either in the classical or quantum domain, so there are no quantum corrections for characteristic of the whole system in this sense. The difference between this situation and that discussed in [@solod] consists in that in the RST model $F$ contains $\kappa $ but $V$ does not, whereas in our case the situation is reverse according to eqs. (\[d=1\]), (\[g\]). It is worth noting that the expression for $V$ can be rewritten as $V=-2\tilde{F}% ^{\prime }+2\kappa $. Therefore, although this coefficient may change its sign it happens only after $\tilde{F}^{\prime }$ so does, i.e. beyond the singularity. On the whole sheet with $\tilde{F}^{\prime }\leq 0$ the quantity $V$ is positive as it should be for the action principle to be well-defined. It is also worth noting that if $\omega =const$ the formula for the entropy of Hawking radiation can be written as $S_{q}=2\kappa \omega (\phi _{B}-\phi _{h})$ that coincides with the entropy of hot gas in a flat space in the container of the length $L=\left| \phi _{B}-\phi _{h}\right| =\lambda \left| x_{B}-x_{h}\right| $ with a temperature $T=\lambda /2\pi $. This generalizes the corresponding observation [@solod] for the RST model. One can say that, in some sense, the entropy of thermal radiation in the backgrounds under discussion does not acquire curvature corrections. Factorization ------------- It is instructive to look at the approach in question from somewhat another view point. Eqs. (\[9\]), (\[12\]) lead to the relation $$R(u-\kappa \omega )=(2V-\omega u)\Box \phi +(V^{\prime }-\omega u^{\prime })(\nabla \phi )^{2} \label{28}$$ If the potential $V$ satisfies eq. (\[g\]), eq. (\[28\]) takes the form $$(u-\kappa \omega )[R-\Box (\int d\phi \omega )]=0 \label{fact}$$ Thus, our choice of relationship between action coefficients gives rise to the important property of factorization, generalizing observation made for the RST model in [@solod]. If $u=\kappa \omega $ we return to the constant dilaton field solutions discussed above. Otherwise $R=\Box \omega $ and, according to (\[4\]), $\omega =\psi ^{\prime }$ in agreement with (\[g\]). One can try to gain the factorization property without referring to eq. (\[14\]) and the form of $V$ following from it. Let us choose the functions, entering the action of the model, in such a way that the coefficient at ($% \nabla \phi )^{2}$ cancel and, besides, the coefficient at $\Box \phi $ be proportional to that at curvature: $$V^{\prime }=\omega u^{\prime },\text{ }2V-\omega u=-\omega _{0}(u-\kappa \omega ) \label{29}$$ where $\omega _{0\text{ }}$is some constant. Then eq.(\[28\]) turns into $$(u-\kappa \omega )(R+\omega _{0}\Box \phi )=0 \label{30}$$ However, one can check directly that eq. (\[14\]) is satisfied in this case automatically, so we do not obtain new solutions. It is worth noting that now $D\neq 1$ where $D$ is a constant entering (\[w\]). There is also the choice $V^{\prime }-\omega u^{\prime }=p(u-\kappa \omega )$, $2V-\omega u=q(u-\kappa \omega )$ with arbitrary functions $p$,$q$ that ensures factorization but we will not discuss this possibility further. comparison with known models ============================ We demonstrated above that the RST model enters our scheme as a particular case. Here we show that the same is true for other known exactly solvable models. BPP model --------- This model [@bose] is characterized in our notations by $$F=e^{-2\phi }-2\kappa \phi \text{, }\tilde{F}=e^{-2\phi }\text{, }\omega =-2% \text{, }\psi =2\phi \text{, }V=4e^{-2\phi }+2\kappa \label{bpp}$$ It is seen from (\[bpp\]) immediately that (\[d=1\]) is satisfied. Substituting (\[bpp\]) into (\[g\]) we obtain $$g=1-e^{2\phi -2\phi _{h}}\text{, }-\lambda x=\phi \label{gbpp}$$ Thus, this model possesses rather unexpected feature: the metric has the same form in terms of $\phi $ and $\phi _{h}$ (or $x$ and $x_{h}$) as its classical counterpart! In other words, not only quantum corrections to the Hawking temperature vanish but also so do quantum correction to the metric itself. It is worth stressing that although this property sharply contrasts with the explicit form of the metric listed in [@bose] there is no contradiction here: authors of [@bose] consider solutions which are radiationless at infinity and have a singular horizon (analog of the Boulware state) whereas we deal with the Hartle-Hawking state that implies that the stress-energy tensor of radiation does not vanish at infinity, an event horizon being regular. In fact, as was explained in Sec. IIIA, any model from our set for which $F=\tilde{F}(\phi )+\kappa \int d\phi \omega $ where $\tilde{F}(\phi )$ does not contain $\kappa $ will give the metric function without quantum corrections as follows from (\[g\]). Other models ------------ The model discussed by Michaud and Myers [@rob] is characterized by $% F=e^{-2\phi }-\frac{\kappa }{2}(\alpha +\sum_{n=2}^{K}a_{n}\phi ^{n})$, $% V=4e^{-2\phi }-\frac{\kappa }{2}(\beta +\sum_{n=2}^{K}b_{n}\phi ^{n-1})$, $% \omega =-2$ where $b_{n}=-2na_{n}$, $\beta =4-2\alpha $. For the Fabbri and Russo (FR) model [@fub] $F=\exp (-\frac{2\phi }{n}% )+\kappa \frac{(1-2n)}{n}\phi $, $V=\frac{4}{n}\exp (-\frac{2\phi }{n}% )+2\kappa \frac{(n-1)}{n}$, $\omega =-2$. When $n=1$ the RST result is reproduced. The model which interpolates between the RST and BPP ones is considered in [@cruz] (CN model). In this case $F=\exp (-2\phi )+2\kappa (a-1)\phi $, $% V=4\exp (-2\phi )+2(1-2a)\kappa $, $\omega =-2$. It reduces to the RST model when $a=\frac{1}{2}$ and to the BPP one when $a=0$. It is easily seen that eq. (\[d=1\]) is satisfied in all these cases. Moreover, one can suggest, for instance, a new model which incorporates at once features of the RST, FR, CN and BPP: $F=\exp (-\frac{2\phi }{n}% )+2\kappa (a-1)$, $V=4\exp (-\frac{2\phi }{n})+2\kappa (1-2a)$, $\omega =-2$. When $a=2n^{-1}$ we return to the BPP, $n=1$ corresponds to the CN model. For the exponential models [@cr2] $F=\phi $, $u=1$, $V=0$, $\omega =const $. This case is described by eq. (\[w\]) with $D\neq 1$. summary and outlook =================== Thus, the models considered in the present paper have the following properties: 1\) they are exactly solvable in the sense that either the metric or dilaton field are found in a closed form; 2) their geometry is static; 3) quantum corrections to the Hawking temperature vanish; 4) for wide subsets of these models one of the following properties holds: a) the relationship between the total energy and the total entropy of the quantum finite size system is the same as in the classical limit, b) the metric itself does not contain quantum corrections and has the same form either in the classical or the quantum domain; 5) there exists the special class of solutions with a constant dilaton field $\phi =\phi _{c}$, the geometry of non-constant dilaton solutions become singular in the point $\phi _{c}$; 6) all spacetime for $\phi \neq \ const$ can be divided to separate sheets with one and only one horizon on every sheet between two neighboring singularities with finite $\phi _{c}$ (plus, perhaps, additional horizons due to $\phi =\infty $ or $% \phi =-\infty $), different sheets are glued in the singular points; all horizons on different sheets share the same temperature; 7) there exists the solution with one horizon and without singularities. The most part of these properties is inherent to the RST model [@solod]. It does not possess, however, the properties 4a) and 4b). Besides, in 6) the number of sheets for the RST model is equal to two, whereas in general it can be arbitrary. Thus, the RST model turns out to be only a representative of a more wide class of models sharing common features. Moreover, we gained also the qualitatively new property which was absent in the RST model: the existence of quantum black holes without singularities whose metric is found explicitly. In this respect the corresponding solutions resemble those in string theory obtained in exact (non-perturbative) approach [@dijk], [@perry]. On the other hand, regular black holes solutions found in the present paper differ from similar ones in [@banks] in that solutions discussed in [@banks] represent the extreme black holes whereas in our case they are essentially nonextreme since their temperature is nonzero constant. The crucial point in which generalization of the RST model is performed consists in that we do not specify the form of the action coefficients and only impose one restriction (\[15\]) on them, so instead of particular models we operate with whole classes of them. In this respect our approach is similar to that of [@kaz] but, in contrast to it, we appeal directly to properties of field equations and do not rely on the general structure of the nonlinear sigma model from which the dilaton-gravity action can be obtained. In so doing, we only assumed (i) the local connection between two unknown functions $\psi $ and $\phi $; (ii) cancellations of coefficients at first and second derivatives of $\phi $ in eq.(\[13\]) that saved us from the trouble to solve a generic complicated differential equation for $\phi (x)$ and enabled us to introduce variables in terms of which the field equations are greatly simplified. It is of interest to generalize the elaborated approach to theories with higher derivatives (in particular, to generalize exact solutions for black holes found at the classical level [@od]), additional scalar and gauge fields, interactions between black holes spacetimes and shock waves, etc. The problem deserving separate attention is detailed desription and classification of different types of spacetime structure of considered quantum black holes similarily to what has been done for classical dilaton black holes [@accel], [@jose]. Of special interest is the issue of regular black holes including their formation by gravitation collapse starting form the vacuum. It would be temting to derive general criteria for Lagrangians admitting nonsingular black hole solutions either in the nonextreme or extreme case and select among them those with exact solutions. Besides the issues connected with black hole physics, exactly solvable models can be useful for the analysis of conceptual problems in different schemes of quantization of dilaton gravity theories [@jakiw]. acknowledgments =============== I am grateful to Sergey Solodukhin for stimulating remarks and to J. V. Cruz for drawing my attention to a number of papers on exactly solvable models in dilaton gravity. This work is supported by International Science Education Program (ISEP), grant No. QSU082068. C. G. Callan, Giddings, J. A. Harvey, and A. Strominger, Phys. Rev. D[** 45**]{} , R1005 (1992); T.Banks, A. Dabholkar, M. R. Douglas, and M. O. ’Loughlin, Phys. Rev.[** **]{}D[** 45**]{}, 3607 (1992). J. G. Russo, L. Susskind, and L. Thorlacius, Phys. Rev. D[** 46**]{}, 3444 (1992); Phys. Rev. D[** 47**]{}, 533 (1992). S. N. Solodukhin, Phys. Rev. D[** 53**]{}, 824 (1996). A. Bilal and C. G. Callan, Nucl. Phys. [**B**]{} 394, 73 (1993). S. P. de Alwis, Phys. Rev. D [**46**]{}, 5429 (1992). Y. Kazama, Y. Satoh, and A. Tsuichiya, Phys. Rev. D [**51**]{}, 4265 (1995). G. Michaud and R. C. Myers, Two-Dimensional Dilaton Black Holes, gr-qc/9508063. A. Fabbri and J. G. Russo, Phys. Rev. D [**53**]{}, 6995 (1995). S. Bose, L. Parker, and Y. Peleg, Phys. Rev. D [**52**]{}, 3512 (1995). J. Cruz and J. Navarro-Salas, Phys. Lett. B 375, 47 (1996). J. Cruz, J. Navarro-Salas, M. Navarro and C. F. Talavera, Phys. Lett. B [**402**]{}, 270 (1997). A. M. Polyakov, Phys. Lett. B [**103**]{}, 207 (1981). V. P. Frolov, W. Israel, and S. N. Solodukhin, Phys. Rev. D[** 54,**]{} 2732 (1996). G. Mandal, A. Sengupta, and S. Wadia, Mod. Phys. Lett. A[** 6**]{}, 1685 (1991); E. Witten, Phys. Rev. D[** 44**]{}, 314 (1991). R. Balbinot and A. Fabbri, Class. Quant. Grav. [**13**]{}, 2457 (1996) O. B. Zaslavskii, in preparation. M. Walker, J. Math. Phys. [**11**]{} 2280 (1970). T. Klösch and T. Strobl, Class. Quant. Grav. [**13**]{}, 2395 (1996). O. B. Zaslavskii, Phys. Lett.[** **]{}B[** 424**]{}, 271 (1998) (hep-th/9802117). V. P. Frolov, Phys. Rev. D[** 46**]{}, 5383 (1992); G.W.Gibbons and M.J.Perry, Int.J.Mod.Phys.[** **]{}D[** 1**]{} 335 (1992). R. C. Myers, Phys. Rev. [**D**]{} 50, 6412 (1994); D. V. Fursaev and S. N. Solodukhin, Phys. Rev. D [**52**]{} (1995) 2133. R. Dijkkgraaf, H. Verlinde, and E. Verlinde, Nucl. Phys. B[** 371**]{}, 269 (1992). M. J. Perry and E. Teo, Phys. Rev. Lett. [**70**]{}, 2669 (1993). T. Banks and M. O’ Loughlin, Phys. Rev. D [**48**]{}, 698 (1993). E. Elizalde, P. Fosibla-Vela, S. Naftulin, and S. Odintsov, Phys. Lett. [**B**]{} 352, 237 (1995). J. P. S. Lemos and P. Sá, Phys. Rev. D [**49**]{}, 2897 (1994), Erratum Phys. Rev. D [**51**]{}, 5967 (1995). E.Benedict, R. Jackiw, and H.-J.Lee, Phys. Rev. D [**54**]{}, 6213 (1996); R. Jakiw and C. Teitelboim, in [*Quantum Theory of Gravity*]{}, edited by S. Christensen (Hilger, Bristol, UK, 1984).

ArXiv

--- abstract: 'Prof. Boris I. Plotkin [@SevLect; @AlgGeom] drew attention to the question when an equivalence between two categories is isomorphic as a functor to an isomorphism between them. It turns out that it is quite important for universal algebraical geometry and concerns mainly the categories $\Theta \sp 0 (X) $ of free universal algebras of some variety $\Theta $ free generated by finite subsets of $X$. In the paper, a complete answer to the Plotkin’s question is given: there are no proper autoequivalences of the category $\Theta \sp 0 (X) $. Also some connected problems are discussed.' author: - | Grigori I. Zhitomirski[^1]\ Department of Mathematics and Statistics\ Bar Ilan University\ Ramat Gan 52900, Israel title: Some remarks on autoequivalences of categories --- Introduction {#introduction .unnumbered} ============ Prof. Boris I. Plotkin set a question arose by studying of the universal algebraical geometry [@SevLect; @AlgGeom]. Without going into details, this question can be formulated in the following way. Let $\mathcal V$ be a variety of universal algebras. Consider the category $\Theta$ whose objects are all algebras from $\mathcal V$ and whose morphisms are all homomorphisms of them. Fix an infinite set $X$. Let $\Theta \sp 0(X) $ be the full subcategory of $\Theta $ containing only free $\mathcal V-$algebras over finite subsets of the set $X$. The question is: if there are autoequivalences of $\Theta \sp 0 (X) $ that are not isomorphic to any automorphism of this category? Below, we give a negative answer to this question, i. e. we show that every autoequivalence of the category $\Theta \sp 0 (X) $ is isomorphic to an automorphism of this category, and present some results concerning that theme. Properly to say Prof. B.I. Plotkin regards the mentioned question to be important for the universal algebraical geometry. The reasons for this opinion are the following ones. Let $\mathcal C$ be an arbitrary small category. Let $End{\mathcal C}$ be the monoid of all endofunctors $F: {\mathcal C} \to {\mathcal C}$ and $Aut {\mathcal C}$ be the group of all endofunctors that are automorphisms of this category. The relation “two functors are isomorphic (natural equivalent)” is a congruence relation on the monoid $ End {\mathcal C}$. The corresponding quotient monoid is denoted by $End\sp 0 {\mathcal C}$ and the group of all invertible elements of it is denoted by $Aut \sp {0}\ \mathcal C$. We have the monoid homomorphism: $\eta : End {\mathcal C} \to End\sp 0 {\mathcal C}$ and the induced homomorphism $\eta \sp * : Aut {\mathcal C }\to Aut \sp {0}\ {\mathcal C}$. The following diagram is commutative. $$\CD Aut {\mathcal C} @> \eta \sp * >> Aut \sp 0 {\mathcal C} \\ @VVV @VVV \\ End {\mathcal C} @ >> \eta > End\sp 0 {\mathcal C} \endCD$$ Here the arrows on each side are the inclusion maps. The case the homomorphism $\eta\sp *$ being surjective is very important. It is easy to prove that this holds if and only if every autoequivalence of the category $\mathcal C$ is isomorphic to an automorphism of this category. Therefore we are interested to know if the category $\Theta \sp 0 (X)$ satisfies this condition. There is one more situation that leads to the same question but this time for two categories. Let $H$ be an object of the category $\Theta $, i.e. it is an algebra from $\mathcal V$. Let $W$ be an object of the category $\Theta \sp 0 =\Theta \sp 0 (X)$, i.e. it is a free algebra in $\mathcal V$ over a finite subset of $X$. The set $Hom(W,H)$ is regarded as a an affine space and its elements are called the points. There is a Galois correspondence between binary relations in $W$ regarded as sets of equations and the sets of points. The closed subsets under this correspondence are called algebraical sets and $H$-closed congruences respectively. Also a notion of algebraical variety can be introduced. Thus with every algebra $H$, some geometry is connected, and the question appears what it means that two algebras have the same geometry. This question leads to the problem of recognizing if the corresponding two categories merely equivalent or what is more isomorphic. In the Section 2 we give a necessary and sufficient condition for an equivalence of two small categories to be isomorphic to an automorphism of them (Theorem \[bijection\]) and obtain from it the mentioned answer to the Plotkin’s question. In the section 3 we obtain some similar results for arbitrary categories represented in the category of sets. Preliminaries ============= We use the usual notions and notation of category theory (see for example \[3\]). Let $\mathcal {C}$ and $\mathcal {D}$ be two arbitrary categories. A functor $F: \mathcal {C} \to \mathcal {D}$ is called an [*equivalence* ]{} if 1\) it is faithful and full, i.e. for every two objects $A,B $ from $\mathcal {C}$ the functor $F$ determines bijection of $Hom (A,B)$ onto $Hom(F(A),F(B))$; 2\) every object from $\mathcal {D}$ is isomorphic to $F(A)$ for some object $A$ of $\mathcal {C}$. These two conditions mean, that $F$ induces an isomorphism between skeletons of these two categories. An equivalence $F: \mathcal {C} \to \mathcal {C}$ is called an [*autoequivalence* ]{} of the category $\mathcal {C}$. We recall that a morphism (or a natural transformation) from a functor $F: \mathcal {C} \to \mathcal {D}$ to a functor $G: \mathcal {C} \to \mathcal {D}$ is a map $\alpha$ which assigns to every object $A$ of $\mathcal {C}$ a morphism $\alpha {\sb A }: F(A) \to G(A)$ such that for every morphism $f: A\to B$ of $\mathcal {C}$ the following diagram commutes: $$\CD F(A)@> \alpha \sb A >> G(A) \\ @VF(f)VV @VVG(f)V \\ F(B) @ >> \alpha \sb B > G(B) \endCD$$ A functor from a category to itself is called an endofunctor. The category of all functors from a category $\mathcal {C}$ to a category $\mathcal {D}$ is denoted by $Funct({\mathcal {C}},{\mathcal {D}} )$. The category of endofunctors of a category $\mathcal {C}$ is denoted by $Funct\sb {\mathcal {C}}$ . Equivalences and isomorphisms from $\mathcal {C}$ to $\mathcal {D}$ (if they exist) are objects of the category $Funct({\mathcal {C}},{\mathcal {D}} )$, and an equivalence which is not isomorphic ( in sense of this category ) to any isomorphism is called a proper equivalence. Every autoequivalence of $\mathcal {C}$ that is isomorphic to the identity functor $\bold Id \sb {\mathcal {C}}$ is called an inner autoequivalence as well an automorphism (that certainly is an autoequivalence) is called an inner automorphism if it is isomorphic to [**Id**]{}. For every object $A$ of $\mathcal {C}$, we denote by $[A]$ the class of all $\mathcal {C}$-objects that are isomorphic to $A$. Let $Sk = Sk\sb {\mathcal C}$ be a skeleton of $\mathcal {C}$. For every object $A$ there exists an unique common object $\overline {A}$ from $Sk$ and $[A]$. Chose an isomorphism $u\sb A : A\to \overline {A}$ for every object $A$ . The next result is well known. \[onSc\] Let $\nu \sb {sk}$ be a map that assigns to every object $A$ the object $\overline {A}$ and to every morphism $f:A\to B$ the morphism $\overline {f} : \overline {A} \to \overline {B}$ by the following way: $\overline {f} =u\sb A \sp {-1} f u\sb B $. Then the map $\nu \sb {sk}$ is an inner autoequivalence of the category $\mathcal {C}$. It is easy to see that there are categories admitting a proper autoequivalence. As an example, every category $\mathcal {C}$ is suitable if it satisfies the condition: its skeleton $Sk = Sk\sb {\mathcal C}$ admits some automorphism $\psi$ that does not fix at least one object but there are no non-trivial automorphisms of the category $\mathcal {C}$ itself. Indeed, let $\nu \sb {sk}$ be the autoequivalence defined in the Lemma above. Thus $\psi \circ \nu \sb {sk}$ is an autoequivalence but if $\psi (X)\not = X$ for some $X\in Sk$ then $\psi \circ \nu \sb{sk}(X)$ is not isomorphic to $X$. Hence $\psi \circ \nu \sb {sk}$ is a proper autoequivalence. Consider, for instance, four items $a,b,c,d$ as objects and pre-order relation defined by following arrows: $a\to b, a\to c\leftrightarrows d$ and of course loops for every object. Let $\mathcal {C}$ be the corresponding category. There is non-trivial automorphism of $Sk\sb {\mathcal C}$ but there are no non-trivial automorphisms of $ {\mathcal C}$. From this point of view the following fact is of interest and it answers the mentioned Plotkin’s question. The proofs of results presented in the next section are quite simple but the author has not met them in the literature and therefore explains them explicitly. There are no proper autoequivalences of the category $\Theta \sp 0 (X) $ ======================================================================== \[bijection\] Let $\mathcal {C}$ and $\mathcal {D}$ be two arbitrary small categories. An equivalence $\alpha :\mathcal C \to \mathcal D$ is isomorphic to an isomorphism if and only if for every to objects $X$ and $Y$ such that $\alpha (X) = Y$ the sets $[X] $ and $[Y]$ are of equal cardinality. Let $\pi :\mathcal C \to \mathcal D$ be an equivalence. Let $\nu =\nu \sb {sk} :\mathcal {D} \to Sk \sb {\mathcal D }$. Consider $\alpha =\nu \circ \pi$. If $\pi (X)=Y\sb 1$ and $\alpha (X)=Y\sb 2$ then objects $Y\sb 1$ and $Y\sb 2$ are isomorphic and hence $[Y\sb 1 ]= [Y\sb 2]$. On the other hand, since $\nu $ is isomorphic to the identity functor $\bold Id \sb {\mathcal D}$, an equivalence $\pi$ is isomorphic to an isomorphism if and only if so $\nu \circ \pi$ is. These two facts mean that it is sufficient to proof our statement for the equivalence $\alpha$. If $\alpha$ is isomorphic to an isomorphism $\varphi$, then for every object $Y$ of $Sk \sb {\mathcal D }$, $\varphi$ determines an one-to-one map of $\alpha \sp {-1} (Y) $ onto $[Y]$. It means that for every to objects $X$ and $Y$ such that $\alpha (X) = Y$ the sets $[X] $ and $[Y]$ are of equal cardinality. If this last condition is satisfied then we have an one-to-one map $\varphi \sb Y :\alpha \sp {-1} (Y) \to [Y]$ for every object $Y$ of $Sk \sb {\mathcal D }$ . So we can define for every object $A$ the object $\varphi (A) =\varphi \sb Y (A)$ if $A \in \alpha \sp {-1} (Y)$. Let $A \in \alpha \sp {-1} (Y\sb 1), B \in \alpha \sp {-1} (Y\sb 2)$ and $f:A\to B$ be a morphism. Then one has the morphism $\alpha (f): \alpha (A) \to \alpha (B)$. The morphism $\varphi (f): \varphi (A) \to \varphi (B)$ can be uniquely defined from the following diagram in order to make it commutative: $$\CD \varphi (A)@> u\sb {\varphi (A)} >> Y\sb 1 \\ @V\varphi (f)VV @VV\alpha (f)V \\ \varphi (B) @ >> u\sb {\varphi (B)} > Y\sb 2 \endCD$$ It is oblivious that the defined map $\varphi$ is an isomorphism of the categories $\mathcal C \to \mathcal D$ and it is isomorphic to $\alpha$. The next result follows easily from the theorem above. \[mainCond\] 1\. In order to a small category $\mathcal C$ has no proper autoequivalences it is necessary and sufficient that for every automorphism $\gamma$ of $Sk$ and for every two objects $X$ and $Y$ of $Sk$ such that $\gamma (X)= Y$, the sets $[X]$ and $[Y]$ are of equal cardinality. Particularly this condition is satisfied if all automorphisms of the category $Sk\sb {\mathcal C} $ leave fixed its objects. 2\. If a small category $\mathcal {C}$ satisfies the condition that for every two its objects $A,B$ the set $[A]$ and $[B]$ have the same cardinality then every autoequivalence of $\mathcal {C}$ is isomorphic to an automorphism of this category. Now consider the main question mentioned in Introduction. Let $\mathcal V $ be a variety of universal algebras and $X$ be an infinite set. The free $\mathcal V $-algebra $W=W(X\sb 0)$ corresponds to every finite subset $X\sb 0$ of $X$ being free generated by this subset. Let $\Theta\sp 0 (X)$ be the category which objects are all such free algebras an which morphisms are homomorphisms of $\mathcal V $-algebras. For every finite subset $X\sb 0 \subset X$ the set $[W(X\sb 0)]$ of $\Theta\sp 0 (X)$-objects has the same cardinality that the set $X$. Hence we have \[free\] Let $\mathcal V \sb 1$ and $\mathcal V \sb 2$ be two varieties of universal algebras. Let $X$ be an infinite set and $\Theta \sb 1 \sp 0 (X),\; \Theta \sb 2 \sp 0 (X) $ be two corresponding categories of free algebras. Then every equivalence of this categories is isomorphic to an isomorphism. Thus if such categories are equivalent they are isomorphic. Further, for a categories of the kind $\Theta \sp 0 (X)$, every its autoequivalence is isomorphic to an automorphism of this category. Now we give more example of a category which has proper autoequivalences. Let $E$ be an ordered set such that there is a non-trivial automorphism $\varphi $ of it. Let $e\in E$ be an element such that $\varphi (e) \neq e$. Add to $E$ a new element $a$ and consider on the set $E'=E\cup \{a\}$ the pre-order relation $\rho$ that will be obtained by adding two pairs $(a,e)$ and $(e,a)$ to the existing order relation on $E$. Let $\mathcal C$ be the category defined on the base of this pre-order relation. The map $\varphi '$ that acts like $\varphi $ on $E$ and takes $a$ to $\varphi (e)$ (i.e. we have $\varphi '(a)=\varphi '(e)$) determines an proper autoequivalence according to Theorem  \[bijection\] because $[e]=\{e,a\}$ and $[\varphi ' (e)] =\{\varphi ' (e)\}$. The proof of Theorem \[bijection\] uses the Axiom of Choice but it is possible to remove references to this axiom in some simple cases, for example for the category $\Theta \sp 0 (X)$ over a countable set $X$ . One can try to extend Theorem \[bijection\] up to non-small categories. But this way leads into jungle of non-sets cardinalities. It concerns axioms which are independent from usual set-theoretic axioms. This reason explains why it is impossible to give an example of a big category (like the category of universal algebras for some variety) with proper autoequivalences. Because the equality of cardinalities for every two sets of kind $[A]$ is not so easy to see we give in the next section some other conditions for an autoequivalence to be isomorphic to an automorphism. Categories with a forgetful functor =================================== We consider such categories $\mathcal C$ that are represented in the category $Set$ (the category of all sets and maps), that is, there exists a faithful functor $Q: \mathcal C \to Set$. Such a functor is called a forgetful functor. If $\mathcal C$ is a category of universal algebras, then the forgetful functor is usually the natural forgetful functor, which assigns to every algebra $A$ the underlying set $\vert A \vert$ and to every homomorphism itself as a mapping, but not only this case. Additionally we assume that the category $\mathcal C$ satisfies the following condition for every two objects $A$ and $B$: (\*) If $Q(A)$ and $Q(B)$ have the same cardinality then, for every bijection $u :Q(A)\to Q(B)$, there exists an unique object $C$ of $\mathcal {C}$ and an isomorphism $u \sp * :C\to B$ such that $Q(C)= Q(A)$ and $Q(u \sp *) = u$. This requirement is a weakening of the condition to be abstract for a subcategory of a category of sets supplied with some structure (for example, an algebraic structure). Indeed, every bijection $u :Q(A)\to Q(B)$ induces the unique structure on the set $Q(A)$ that the map $u$ is an isomorphism of the obtained object $C$ onto $B$. Our requirement is that this object belongs to the subcategory under consideration. \[generConD\] Let $\pi :\mathcal {C} \to \mathcal {C}$ be an autoequivalence. If for every object $A$ of $\mathcal {C}$ there exists a bijection $u \sb A : Q (A) \to Q\pi (A)$ such that for every morphism $f:A \to B$ of $\mathcal {C}$ the following diagram commutes: $$\CD Q(A)@> u \sb A >> Q\pi (A) \\ @VQ(f)VV @VVQ\pi (f)V \\ Q(B) @ >> u \sb B > Q\pi (B) \endCD$$ then $\pi $ is isomorphic to an automorphism $F $ of $\mathcal {C}$. According to hypothesis, we have for every object $A$ the bijection $u \sb A : Q (A) \to Q\pi (A)$. Because of the property (\*), there exists an unique object $F(A)$ and an unique isomorphism $u \sp * \sb A :F(A)\to \pi (A)$ such that : $$Q(F(A)) =Q(A)\qquad {\rm and}\qquad Q(u \sp * \sb A )= u \sb A .$$ Next, define for every morphism $f:A \to B$ of $\mathcal {C}$ the morphism $F (f): F(A) \to F(B)$ by the following way: $F(f)= u \sp * \sb A \pi (f)(u \sb B \sp * )\sp {-1}$. Thus we have the following commutative diagram: $$\CD F(A)@> u \sp * \sb A >> \pi (A) \\ @VF(f)VV @VV\pi (f)V \\ F(B) @ >> u \sp * \sb B > \pi (B) \endCD$$ It means that $F$ is an endofunctor of the category $\mathcal C $ and it is isomorphic to the given autoequivalence $\pi$. It remains to proof that $F$ is an automorphism of $\mathcal C $. Since $F$ is autoequivalence, we have only to show that $F$ is an bijection with respect to objects. Let $F(A) =F(B)$. Then we have an isomorphism $g=(u \sp * \sb A )\sp {-1}u \sp * \sb B : \pi (A) \to \pi (B)$. Thus there exists an unique isomorphism $f: A\to B$ such that $\pi (f) =g$. According to this definition of the morphism $f$ and hypothesis of the theorem, the following diagram commutes: $$\CD Q(A)@> u \sb A >> Q\pi (A) \\ @VQ(f)VV @VVQ(g)V \\ Q(B) @ >> u \sb B > Q\pi (B) \endCD .$$ Having in the mind that $Q(A)=Q(F(A))=Q(F(B))=Q(B)$, we obtain: $$Q(f)= u \sb A u \sb A \sp {-1}u \sb B u \sb B \sp {-1} =1\sb {Q(A)}.$$ Because of the condition (\*) it means that $A=B$. Now let $S$ be an arbitrary object of $\mathcal C$. Since $\pi$ is an autoequivalence, there exists an object $T$ such that there exists an isomorphism $i:S\to \pi (T)$. Thus we obtain a bijection $g =Q(i)u \sb T \sp {-1}:Q( S) \to Q(T) $. Let $V$ be an object from $\mathcal C$ such that $Q(V)=Q(S)$, $g\sp * : V\to T$ is an isomorphism, and $Q(g\sp *)= g$. Then the following diagram commutes: $$\CD Q(V)@> u \sb V >> Q\pi (V) \\ @Vg VV @VVQ\pi (g \sp *)V \\ Q(T) @ >> u \sb T > Q(\pi (T)) \endCD .$$ Let $h=u \sp * \sb V \ \pi (g \sp *)\ i \sp {-1} :F(V)\to S$. It is clear that $h$ is an isomorphism and $Q(F(V))=Q(S)$. It follows from the diagram above and the definition of the bijection $g$ that $$Q(h)=u \sb V \ u \sb V \sp {-1} \ g \ u \sb T \ Q(i\sp {-1}) = Q(i)\ u \sb T \sp {-1}\ u \sb T \ Q(i\sp {-1}) =1\sb {Q(S)}$$ and hence $F(V) =S$. This completes the proof. The hypotheses of Theorem \[generConD\] are satisfied in many important cases. \[saferepr\] If the functor $Q: \mathcal C \to Set $ is represented by an object $W$ and if an autoequivalence $\pi$ satisfies the condition that $W$ is isomorphic to $\pi (W)$ then $\pi$ satisfies the hypotheses of Theorem \[generConD\] and hence it is isomorphic to an automorphism of $\mathcal C$. Let $W$ and $ \pi (W)$ are isomorphic according to hypotheses. Thus the functor $Q$ is represented by $\pi (W)$ too. We have two isomorphisms in the category $Funct(\mathcal C ,Set )$, namely $\bold g :Q \to Hom(W,-)$ and $\bold h :Q \to Hom(\pi (W), -)$. Since $\pi $ induces a bijection between sets $Hom(A,B)$ and $Hom(\pi (A),\pi (B))$ for every pair $(A,B)$ of objects, there is a bijection $u \sb A : Q (A) \to Q\pi (A)$ for every object $A$ defined as the composition of bijections: $$\CD Q(A)@>\bold g \sb a >>Hom (W,A)@>\pi >>Hom(\pi (W),\pi (A)) @> \bold h \sp {-1} \sb {\pi (A)}>>Q\pi (A). \endCD$$ The fact that the family $u \sb A : Q (A) \to Q\pi (A)$ of bijections satisfies the hypotheses of Theorem \[generConD\] one can see from the following diagram in which every of three squares is commutative for every morphism $f:A\to B$: $$\CD Q(A)@>\bold g \sb a >>Hom (W,A)@>\pi >>Hom(\pi (W),\pi (A)) @> \bold h \sp {-1} \sb {\pi (A)}>>Q\pi (A) \\ @Vf VV @VVHom(W,f)V @VVHom(\pi (W),\pi (f))V @VVQ\pi (f)V \\ Q(B)@>\bold g \sb a >>Hom (W,B)@>\pi >>Hom(\pi (W),\pi (B)) @> \bold h \sp {-1} \sb {\pi (B)}>>Q\pi (B) \endCD .$$ The first consequence from Theorem \[saferepr\] is the following result: \[set\] Every full subcategory of $Set$ containing an one-element set as an object has no proper equivalences, moreover all its autoequivalences are inner. Let $\mathcal V$ be a variety of universal algebras and $\Theta $ be the corresponding category. Let $Q: \Theta \to Set$ be the forgetful functor. It is clear that condition (\*) of the previous section is satisfied. Moreover, the functor $Q$ is represented by the free algebra $W$ in $\mathcal V $ over an one-element set. Thus according to Theorem  \[saferepr\], if $W$ is isomorphic to $\pi (W)$ for an autoequivalence $\pi$ of $\Theta $, then $\pi $ is isomorphic to an automorphism of this category. To be sure, that the same fact is valid for every full subcategory of $\Theta $ if it satisfies the condition (\*) and contains a free algebra $W$ over a singleton. So the question is if for an autoequivalence $\pi$ the algebras $W$ and $\pi (W)$ are isomorphic. Since $\pi$ induces an isomorphism between the monoids $END(W)$ and $END(\pi (W))$ of endomorphisms of these two algebras it is sufficient to check if the fact that these monoids are isomorphic implies that the mentioned algebras are isomorphic too. If it is true we can conclude that every autoequivalence is isomorphic to an automorphism of this category. Hence we have \[freeAlg\] Let $\mathcal C $ be the category of all free algebras in a variety $\mathcal V $ of universal algebras. Let $W$ be the free algebra $W$ in $\mathcal V $ over an one-element set. If any free algebra $F$ in $\mathcal V $ is isomorphic to $W$ provided that monoids $END(W)$ and $END(F)$ are isomorphic then the category $\mathcal C $ has no proper autoequivalence. For many known varieties, the monogenic free algebras satisfy the condition mentioned above, in particular, the categories of all free semigroups, of all free inverse semigroups or of all free groups have no proper autoequivalence. I am very thankful to Prof. Boris I. Plotkin who has called my attention to the mentioned above problems and has discussed the results with me. B. Plotkin, [*Seven lectures on the universal algebraic geometry*]{}. Preprint, (2002), Arxiv:math, GM/0204245, 87pp. B. Plotkin, [*Algebras with the same (algebraic) geometry*]{}, Proceedings of the Steklov Institute of Mathematics. [**242**]{} (2003), 165-196. S. MacLane, [*Categories for the working mathematicians*]{}, Springer, 1971. [^1]: This research is partially supported by THE ISRAEL SCIENCE FOUNDATION founded by The Israel Academy of Sciences and Humanities - Center of Excellence Program.

ArXiv

--- abstract: 'We classify the biharmonic non-Legendre curves in a Sasakian space form for which the angle between the tangent vector field and the characteristic vector field is constant and obtain explicit examples of such curves in $\mathbb{R}^{2n+1}(-3)$.' address: | Department of Mathematics\ “Gh. Asachi” Technical University of Iasi\ Bd. Carol I no. 11\ 700506 Iasi, Romania author: - Dorel Fetcu title: A note on biharmonic curves in Sasakian space forms --- [^1] [^2] Introduction ============ In 1964, J. Eells and J.H. Sampson introduced the notion of poly-harmonic maps as a natural generalization of harmonic maps ([@Eells]). Thus, while *harmonic maps* between Riemannian manifolds $\phi:(M,g)\to(N,h)$ are the critical points of the *energy functional* $E(\phi)=\frac{1}{2}\int_{M}|d\phi|^{2} \ v_{g}$, the *biharmonic maps* are the critical points of the *bienergy functional* $E_{2}(\phi)=\frac{1}{2}\int_{M}|\tau(\phi)|^{2} \ v_{g}$. On the other hand, B.-Y. Chen defined the biharmonic submanifolds in an Euclidean space as those with harmonic mean curvature vector field ([@Chen]). If we apply the characterization formula of biharmonic maps to Riemannian immersions into Euclidean spaces, we recover Chen’s notion of biharmonic submanifold. The Euler-Lagrange equation for the energy functional is $\tau(\phi)=0$, where $\tau(\phi)=\operatorname{trace}\nabla d\phi$ is the tension field, and the Euler-Lagrange equation for the bienergy functional was derived by G. Y. Jiang in [@Jiang]: $$\begin{array}{cl} \tau_{2}(\phi)&=-\Delta\tau(\phi)-\operatorname{trace}\ R^{N}(d\phi,\tau(\phi))d\phi\\ \\ &=0.\end{array}$$ Since any harmonic map is biharmonic, we are interested in non-harmonic biharmonic maps, which are called *proper-biharmonic*. There are several classification results and some methods to construct biharmonic submanifolds in space forms ([@MontaldoOniciuc], [@BMO]). In a natural way, the next step is the study of biharmonic submanifolds in Sasakian space forms. Thus, J. Inoguchi classified in [@Ino] the proper-biharmonic Legendre curves and Hopf cylinders in a 3-dimensional Sasakian space form $M^{3}(c)$, and in [@FetcuOniciuc] the explicit parametric equations were obtained. In [@CIL], J.T. Cho, J. Inoguchi and J.-E. Lee studied the biharmonic curves in a 3-dimensional Sasakian space forms and T. Sasahara studied the biharmonic integral surfaces in 5-dimensional Sasakian space forms ([@Sasahara1]). New classification results for biharmonic Legendre curves and examples of proper-biharmonic submanifolds in any dimensional Sasakian space form were obtained in [@FetcuOniciuc2]. Recent results on biharmonic submanifolds in spaces of nonconstant sectional curvature were obtained by T. Ichiyama, J. Inoguchi and H. Urakawa in [@Ura], by Y.-L. Ou and Z.-P. Wang in [@Ou], and by W. Zhang in [@Zhang]. Biharmonic submanifolds in pseudo-Euclidean spaces were also studied, and many examples and classification results were obtained (for example, see [@Arv], [@Chen]). The goals of our paper are to obtain new classification results for biharmonic non-Legendre curves in any dimensional Sasakian space form and to obtain explicit equations for some of such curves in $\mathbb{R}^{2n+1}(-3)$. For a general account of biharmonic maps see [@MontaldoOniciuc] and *The Bibliography of Biharmonic Maps* [@bibl]. **Conventions.** We work in the $C^{\infty}$ category, that means manifolds, metrics, connections and maps are smooth. The Lie algebra of the vector fields on $M$ is denoted by $C(TM)$. Preliminaries ============= A triple $(\varphi,\xi,\eta)$ is called a *contact structure* on a manifold $N^{2n+1}$, where $\varphi$ is a tensor field of type $(1,1)$ on $N$, $\xi$ is a vector field and $\eta$ is an 1-form, if $$\begin{array}{c} \varphi^{2}=-I+\eta\otimes\xi,\ \ \eta(\xi)=1,\ \ \forall X,Y\in C(TN).\end{array}$$ A Riemannian metric $g$ on $N$ is said to be an associated metric and then $(N,\varphi,\xi,\eta,g)$ is a *contact metric manifold* if $$g(\varphi X,\varphi Y)=g(X,Y)-\eta(X)\eta(Y),\ \ \ g(X,\varphi Y)=d\eta(X,Y),\ \ \forall X,Y\in C(TN).$$ A contact metric structure $(\varphi,\xi,\eta,g)$ is called [*normal*]{} if $$N_{\varphi}+2d\eta\otimes\xi=0,$$ where $$N_{\varphi}(X,Y)=[\varphi X,\varphi Y]-\varphi \lbrack \varphi X,Y]-\varphi \lbrack X,\varphi Y]+\varphi ^{2}[X,Y],\ \ \forall X,Y\in C(TN),$$ is the Nijenhuis tensor field of $\varphi$. A contact metric manifold $(N,\varphi,\xi,\eta,g)$ is a *Sasakian manifold* if it is normal or, equivalently, if $$(\nabla_{X}\varphi)(Y)=g(X,Y)\xi-\eta(Y)X,\ \ \forall X,Y\in C(TN).$$ The [*contact distribution*]{} of a Sasakian manifold $(N,\varphi,\xi,\eta,g)$ is defined by $\{X\in TN:\eta(X)=0\}$, and an integral curve of the contact distribution is called [*Legendre curve*]{}. Let $(N,\varphi,\xi,\eta,g)$ be a Sasakian manifold. The sectional curvature of a 2-plane generated by $X$ and $\varphi X$, where $X$ is an unit vector orthogonal to $\xi$, is called *$\varphi$-sectional curvature* determined by $X$. A Sasakian manifold with constant $\varphi$-sectional curvature $c$ is called a *Sasakian space form* and it is denoted by $N(c)$. The curvature tensor field of a Sasakian space form $N(c)$ is given by $$\label{eccurv} \begin{array}{ll} R(X,Y)Z=&\frac{c+3}{4}\{g(Z,Y)X-g(Z,X)Y\}+\frac{c-1}{4}\{\eta(Z)\eta(X)Y-\\ \\ &-\eta(Z)\eta(Y)X+g(Z,X)\eta(Y)\xi-g(Z,Y)\eta(X)\xi+\\ \\&+g(Z,\varphi Y)\varphi X-g(Z,\varphi X)\varphi Y+2g(X,\varphi Y)\varphi Z\}. \end{array}$$ Biharmonic non-Legendre curves in Sasakian space forms ====================================================== Let $(N^{m},g)$ be a Riemannian manifold and $\gamma:I\to N$ a curve parametrized by arc length, that is $\vert\gamma'\vert=1$. Then $\gamma$ is called a *Frenet curve of osculating order r*, $1\leq r\leq m$, if there exists orthonormal vector fields $E_{1},E_{2},...,E_{r}$ along $\gamma$ such that $E_{1}=\gamma'=T$ and $$\nabla_{T}E_{1}=\kappa_{1}E_{2},\ \ \nabla_{T}E_{2}=-\kappa_{1}E_{1}+\kappa_{2}E_{3},..., \nabla_{T}E_{r}=-\kappa_{r-1}E_{r-1},$$ where $\kappa_{1},...,\kappa_{r-1}$ are positive functions on $I$. A Frenet curve of osculating order 1 is a geodesic; a Frenet curve of osculating order 2 with $\kappa_{1}=\operatorname{constant}$ is called a *circle*; a Frenet curve of osculating order $r$, $r\geq 3$, with $\kappa_{1},...,\kappa_{r-1}$ constants is called a *helix of order r* and helix of order $3$ is called, simply, helix. Let $(N^{2n+1},\varphi,\xi,\eta,g)$ be a Sasakian space form with constant $\varphi$-sectional curvature $c$ and $\gamma:I\to N$ a non-Legendre Frenet curve of osculating order $r$ with $\eta(T)=f$, where $f$ is a function defined along $\gamma$ and $f\neq 0$. Since $$\begin{array}{llc} \nabla_{T}^{3}T=&(-3\kappa_{1}\kappa_{1}')E_{1}+(\kappa_{1}''-\kappa_{1}^{3}-\kappa_{1}\kappa_{2}^{2})E_{2}+(2\kappa_{1}'\kappa_{2}+\kappa_{1}\kappa_{2}')E_{3}\\ \\ &+\kappa_{1}\kappa_{2}\kappa_{3}E_{4} \end{array}$$ and $$\begin{array}{llc} R(T,\nabla_{T}T)T=&\Big(-\frac{(c+3)\kappa_{1}}{4}+\frac{(c-1)\kappa_{1}}{4}f^{2}\Big)E_{2}-\frac{(c-1)}{4}ff'T+\frac{(c-1)}{4}f'\xi\\ \\&-\frac{3(c-1)\kappa_{1}}{4}g(E_{2},\varphi T)\varphi T, \end{array}$$ we get $$\label{eq:1} \begin{array}{rl} \tau_{2}(\gamma)=&\nabla_{T}^{3}T-R(T,\nabla_{T}T)T\\ \\ =&(-3\kappa_{1}\kappa_{1}'+\frac{(c-1)}{4}ff')E_{1}+\Big(\kappa_{1}''-\kappa_{1}^{3}-\kappa_{1}\kappa_{2}^{2}+\frac{(c+3)\kappa_{1}}{4}-\frac{(c-1)\kappa_{1}}{4}f^{2}\Big)E_{2}\\ \\&+(2\kappa_{1}'\kappa_{2}+\kappa_{1}\kappa_{2}')E_{3}+\kappa_{1}\kappa_{2}\kappa_{3}E_{4}-\frac{(c-1)}{4}f'\xi+\frac{3(c-1)\kappa_{1}}{4}g(E_{2},\varphi T)\varphi T. \end{array}$$ If $c=1$ then $\gamma$ is proper-biharmonic if and only if $$\left\{\begin{array}{l} \kappa_{1}=\operatorname{constant}>0,\ \ \kappa_{2}=\operatorname{constant}\\ \\ \kappa_{1}^{2}+\kappa_{2}^{2}=1\\ \\ \kappa_{2}\kappa_{3}=0\end{array}\right.$$ and we can state the following Theorem. \[teocase1\]If $c=1$ then $\gamma$ is proper-biharmonic if and only if either $\gamma$ is a circle with $\kappa_{1}=1$, or $\gamma$ is a helix with $\kappa_{1}^{2}+\kappa_{2}^{2}=1$. Now, assume that $c\neq 1$. Then $\gamma$ is proper-bihrmonic if and only if $$\label{eq:2} \left\{\begin{array}{l} \kappa_{1}=\operatorname{constant}>0,\\ \\ \kappa_{1}^{2}+\kappa_{2}^{2}=\frac{c+3}{4}-\frac{c-1}{4}f^{2}-\frac{1}{\kappa^{2}_{1}}\frac{c-1}{4}(f')^{2}+\frac{3(c-1)}{4}(g(E_{2},\varphi T))^{2}\\ \\ \kappa_{2}'-\frac{1}{\kappa_{1}}\frac{c-1}{4}f'\eta(E_{3})+\frac{3(c-1)}{4}g(E_{2},\varphi T)g(E_{3},\varphi T)=0\\ \\\kappa_{2}\kappa_{3}-\frac{1}{\kappa_{1}}\frac{c-1}{4}f'\eta(E_{4})+\frac{3(c-1)}{4}g(E_{2},\varphi T)g(E_{4},\varphi T)=0\end{array}\right.,$$ since, from $\eta(T)=g(T,\xi)=f$ and the first Frenet equation, we get $g(\nabla_{T}T,\xi)=g(\kappa_{1}E_{2},\xi)=f'$. Obviously, the above equations are simpler when $f=\eta(T)=\cos\beta_{1}$ is a constant, where $\beta_{1}\in(0,\pi)\setminus\{\frac{\pi}{2}\}$ is the angle between the tangent vector field $T$ and the characteristic vector field $\xi$. In the following, we will study only this special case. We have Let $c\neq 1$ and $\gamma$ a Frenet curve of osculating order $r$ such that $\eta(T)=\cos\beta_{1}=\operatorname{constant}\notin\{-1,0,1\}$. Then $\gamma$ is proper-biharmonic if and only if either a\) $\gamma$ is a circle with $\varphi T=\pm\sin\beta_{1}E_{2}$ and $ \kappa^{2}_{1}=1+(c-1)\sin^{2}\beta_{1}>0$, or b\) $\gamma$ is a helix with $\varphi T=\pm\sin\beta_{1}E_{2}$ and $\kappa^{2}_{1}+\kappa_{2}^{2}=1+(c-1)\sin^{2}\beta_{1}>0$, or c\) $\gamma$ is a Frenet curve of osculating order $r$, where $r\geq 4$, with $$\varphi T=\sin\beta_{1}\cos\beta_{2}E_{2}+\sin\beta_{1}\sin\beta_{2}E_{4}$$ and $$\left\{\begin{array}{l}\kappa_{1}=\operatorname{constant}>0,\ \ \kappa_{2}=\operatorname{constant}\\ \\ \kappa_{1}^{2}+\kappa_{2}^{2}=\frac{c+3}{4}-\frac{c-1}{4}\cos^{2}\beta_{1}+\frac{3(c-1)}{4}\sin^{2}\beta_{1}\cos^{2}\beta_{2}\\ \\\kappa_{2}\kappa_{3}=-\frac{3(c-1)}{8}\sin^{2}\beta_{1}\sin(2\beta_{2}) \end{array}\right.,$$ where $\beta_{2}\in (0,2\pi)$ is a constant such that $c+3-(c-1)\cos^{2}\beta_{1}+3(c-1)sin^{2}\beta_{1}\cos^{2}\beta_{2}$ $>0$, $3(c-1)\sin(2\beta_{2})<0$ and $n\geq 2$. First, we see that $\eta(E_{2})=g(E_{2},\xi)=\frac{1}{\kappa_{1}}f'=0$. Next, assume that $g(E_{2},\varphi T)$ $=\alpha$, where $\alpha$ is a function defined along $\gamma$. Then, using the second Frenet equation, one obtains $$\alpha'=g(\nabla_{T}E_{2},\varphi T)+g(E_{2},\nabla_{T}\varphi T)=\kappa_{2}g(E_{3},\varphi T)+g(E_{2},\kappa_{1}\varphi E_{2}+\xi-fT),$$ and, since the second term in the right side vanishes, it follows $\kappa_{2}g(E_{3},\varphi T)=\alpha'$. Now, if $\gamma$ is proper-biharmonic, replacing into the third equation of (\[eq:2\]) we obtain $$\kappa_{2}\kappa_{2}'+\frac{3(c-1)}{4}\alpha\alpha'=0$$ and then $$\kappa^{2}_{2}+\frac{3(c-1)}{4}\alpha^{2}+\omega_{0}=0,$$ where $\omega_{0}$ is a constant. The second equation of (\[eq:2\]) becomes $$\kappa_{1}^{2}+\kappa_{2}^{2}=\frac{c+3}{4}-\frac{c-1}{4}f^{2}-\kappa_{2}^{2}-\omega_{0}.$$ Hence $\kappa_{2}=\operatorname{constant}$ and $\alpha=\operatorname{constant}$. If $\kappa_{2}=0$ then, from the biharmonic equation $\tau_{2}(\gamma)=0$, we get $E_{2}\parallel\varphi T$ and, since $g(\varphi T,\varphi T)=1-f^{2}=\sin^{2}\beta_{1}$ it follows $\varphi T=\pm\sin\beta_{1}E_{2}$. Hence $\gamma$ is a circle with $$\kappa^{2}_{1}=\frac{c+3}{4}-\frac{c-1}{4}\cos^{2}\beta_{1}+\frac{3(c-1)}{4}\sin^{2}\beta_{1}=1+(c-1)\sin^{2}\beta_{1}.$$ Assume that $\kappa_{2}\neq 0$. Then $g(E_{3},\varphi T)=0$ and, if $\kappa_{3}=0$ then $\gamma$ is a helix with $$\kappa^{2}_{1}+\kappa_{2}^{2}=\frac{c+3}{4}-\frac{c-1}{4}\cos^{2}\beta_{1}+\frac{3(c-1)}{4}\sin^{2}\beta_{1}=1+(c-1)\sin^{2}\beta_{1},$$ since, using again the biharmonic equation, one obtains $E_{2}\parallel \varphi T$ in this case too. Next, let $\gamma$ be a proper-biharmonic Frenet curve of osculating order $r$ with $r\geq 4$. Then $g(E_{3},\varphi T)=0$ and, from the biharmonic equation we have $\varphi T\in \operatorname{span}\{E_{2},E_{4}\}$. Since $$g(\varphi T,\varphi T)=1-f^{2}=\sin^{2}\beta_{1}$$ it follows $$\varphi T=\sin\beta_{1}\cos\beta_{2}E_{2}+\sin\beta_{1}\sin\beta_{2}E_{4},$$ where $$g(E_{2},\varphi T)=\alpha=\sin\beta_{1}\cos\beta_{2}\ \ \ \text{and}\ \ \ g(E_{4},\varphi T)=\sin\beta_{1}\sin\beta_{2}$$ with $\beta_{2}=\operatorname{constant}\in(0,2\pi)$. Finally, if the dimension of $N$ is equal to $3$ we can consider an orthogonal system of vectors $\{E=T-f\xi,\varphi T,\xi\}$ along $\gamma$ and, since $f$ is a constant, it follows easily that $\nabla_{T}T\parallel\varphi T$. Hence $E_{2}\parallel\varphi T$ in this case. A special role in the biharmonic equation $\tau_{2}(\gamma)=0$ is played by $g(E_{2},\varphi T)$. In the following, we consider $\gamma$ to be a Frenet curve of osculating order $r$, with $\eta(T)=f(s)=\cos\beta(s)$ not necessarily constant, such that $E_{2}\perp\varphi T$ or $E_{2}\parallel \varphi T$. **Case I: $\mathbf{c\neq 1,\ E_{2}\perp\varphi T}$.** In this case $\gamma$ is proper-biharmonic if and only if $$\label{eq:3} \left\{\begin{array}{l} \kappa_{1}=\operatorname{constant}>0,\\ \\ \kappa_{1}^{2}+\kappa_{2}^{2}=\frac{c+3}{4}-\frac{c-1}{4}f^{2}-\frac{1}{\kappa^{2}_{1}}\frac{c-1}{4}(f')^{2}\\ \\ \kappa_{2}'-\frac{1}{\kappa_{1}}\frac{c-1}{4}f'\eta(E_{3})=0\\ \\\kappa_{2}\kappa_{3}-\frac{1}{\kappa_{1}}\frac{c-1}{4}f'\eta(E_{4})=0\end{array}\right.,$$ From $g(E_{2},\xi)=\frac{1}{\kappa_{1}}f'$ one obtains $g(\nabla_{T}E_{2},\xi)-g(E_{2},\varphi T)=\frac{1}{\kappa_{1}}f''$ and then $\kappa_{2}\eta(E_{3})$ $=\frac{1}{\kappa_{1}}f''+\kappa_{1}f$. Replacing into the third equation of one obtains $$\kappa_{2}\kappa_{2}'-\frac{1}{\kappa^{2}_{1}}\frac{c-1}{4}(f'f''+\kappa^{2}_{1}ff')=0,$$ and then $$\kappa_{2}^{2}-\frac{1}{\kappa^{2}_{1}}\frac{c-1}{4}((f')^{2}+\kappa^{2}_{1}f^{2})+\omega_{1}=0,$$ where $\omega_{1}$ is a constant. Now, from the second equation of we have $$\kappa_{1}^{2}+\kappa_{2}^{2}=\frac{c+3}{4}-\kappa^{2}_{2}-\omega_{1}.$$ Hence $\kappa_{2}=\operatorname{constant}$ and $(f''+\kappa^{2}_{1}f)f'=0$. Now, using the Frenet equations, from $g(E_{2},\varphi T)=0$ one obtains $\kappa_{2}g(E_{3},\varphi T)=-\frac{1}{\kappa_{1}}f'$ and then, from $\kappa_{2}g(E_{3},\xi)=\frac{1}{\kappa_{1}}f''+\kappa_{1}f$, we get $$\kappa_{2}\kappa_{3}g(E_{4},\xi)=\frac{1}{\kappa_{1}}(f'''+(\kappa^{2}_{1}+\kappa^{2}_{2})f').$$ Since $\tau_{2}(\gamma)=0$ implies $\eta(\tau_{2}(\gamma))=0$ one obtains, after a straightforward computation that $f'f'''=0$. Using this result and differentiating $(f''+\kappa^{2}_{1}f)f'=0$ along $\gamma$ we have $$\kappa^{2}_{1}(f')^{2}+(f''+\kappa^{2}_{1}f)f''=0.$$ We just obtained that $f=\operatorname{constant}$. We can state \[tperp\] Assume that $c\neq 1$, $n\geq 2$ and $\nabla_{T}T\perp\varphi T$. Then $\gamma$ is proper-biharmonic if and only if either a\) $\gamma$ is a circle with $\eta(T)=\cos\beta_{0}$ and $\kappa_{1}^{2}=\frac{c+3}{4}-\frac{c-1}{4}\cos^{2}\beta_{0}$, or b\) $\gamma$ is a helix with $\eta(T)=\cos\beta_{0}$ and $\kappa_{1}^{2}+\kappa_{2}^{2}=\frac{c+3}{4}-\frac{c-1}{4}\cos^{2}\beta_{0}$, where $\beta_{0}\in(0,2\pi)\setminus\{\frac{\pi}{2},\pi,\frac{3\pi}{2}\}$ is a constant such that $\frac{c+3}{4}-\frac{c-1}{4}\cos^{2}\beta_{0}>0$. We note that the biharmonic equation $\tau_{2}(\gamma)=0$ for curves $\gamma$ with $\nabla_{T}T\perp\varphi T$ is equivalent to $$\Delta H=\frac{1}{4}(c+3-(c-1)\cos^{2}\beta_{0})H,$$ i.e. $H$ is an eigenvector of $\Delta$, where $H=\nabla_{T}T$ is the mean curvature vector field of $\gamma$. **Case II: $\mathbf{c\neq 1,\ E_{2}\parallel\varphi T}$.** In this case $g(E_{2},\xi)=\frac{1}{\kappa_{1}}f'=0$ and then $f=\cos\beta_{0}=\operatorname{constant}$. Since $g(\varphi T,\varphi T)=1-(g(T,T))^{2}=\sin^{2}\beta_{0}$ we have $\varphi T=\pm\sin\beta_{0}E_{2}$. We obtain \[p1\] Assume that $c\neq 1$ and $\nabla_{T}T\parallel\varphi T$. Then $\gamma$ is proper-biharmonic if and only if either a\) $\gamma$ is a circle with $\eta(T)=\cos\beta_{0}$ and $\kappa_{1}^{2}=c-(c-1)\cos^{2}\beta_{0}$, or b\) $\gamma$ is a helix with $\eta(T)=\cos\beta_{0}$ and $\kappa_{1}^{2}+\kappa_{2}^{2}=c-(c-1)\cos^{2}\beta_{0}$, where $\beta_{0}\in(0,2\pi)\setminus\{\frac{\pi}{2},\pi,\frac{3\pi}{2}\}$ is a constant such that $c-(c-1)\cos^{2}\beta_{0}>0$. Next, let $\gamma$ be a proper-biharmonic non-Legendre curve with $\nabla_{T}T\parallel\varphi T$. As $\varphi T=\pm\sin\beta_{0}E_{2}$ one obtains after a straightforward computation that $$\nabla_{T}E_{2}=-\frac{1}{\sin\beta_{0}}\Big(\frac{\kappa_{1}}{\sin\beta_{0}}\pm\cos\beta_{0}\Big)T+ \frac{1}{\sin\beta_{0}}\Big(\frac{\kappa_{1}\cos\beta_{0}}{\sin\beta_{0}}\pm 1\Big)\xi.$$ Using the second Frenet equation we have $$\kappa_{2}^{2}=\frac{(\kappa_{1}\cos\beta_{0}\pm\sin\beta_{0})^{2}}{\sin^{2}\beta_{0}}.$$ Thus $\gamma$ is a circle if and only if $\kappa_{1}=\mp\tan\beta_{0}>0$. From Proposition \[p1\] we easily get that $\gamma$ is a proper-biharmonic circle if and only if $$\kappa_{1}^{2}=\frac{c-1+\sqrt{c^{2}-2c+5}}{2}\ \ \text{and}\ \ \cos^{2}\beta_{0}=\frac{c+1-\sqrt{c^{2}-2c+5}}{2(c-1)}.$$ If $\kappa_{2}\neq 0$, from the expression of $\kappa_{2}$ and the third Frenet equation it follows that $\kappa_{3}=0$. Hence $\gamma$ is a helix. Now, $\gamma$ is proper-biharmonic if and only if $\kappa_{1}$ satisfies $$\kappa_{1}^{2}\pm\cos(2\beta_{0})\kappa_{1}+(1-c)\sin^{4}\beta_{0}=0$$ and $\beta_{0}\in(0,2\pi)\setminus\{\frac{\pi}{2},\pi,\frac{3\pi}{2}\}$ if $c>1$ or $\beta_{0}\in(0,2\pi)\setminus\{\frac{\pi}{2},\pi,\frac{3\pi}{2}\}$ such that $\cos\beta_{0}\in\Big(-\sqrt{\frac{c-1}{c-2}},\sqrt{\frac{c-1}{c-2}}\Big)$ if $c<1$. We conclude with the following \[tpar\] If $c\neq 1$ and $\nabla_{T}T\parallel\varphi T$, then $\gamma$ is a Frenet curve of osculating order $r\leq 3$ and it is proper-biharmonic if and only if either a\) $\gamma$ is a circle with $\eta(T)=\pm\sqrt{\frac{c+1-\sqrt{c^{2}-2c+5}}{2(c-1)}}$ and $\kappa_{1}^{2}=\frac{c-1+\sqrt{c^{2}-2c+5}}{2}$, or b\) $\gamma$ is a helix with $\eta(T)=\cos\beta_{0}$ and $\kappa_{1}$ satisfies $$\kappa_{1}^{2}\pm\cos(2\beta_{0})\kappa_{1}+(1-c)\sin^{4}\beta_{0}=0,$$ where $\beta_{0}=\operatorname{constant}\in(0,2\pi)\setminus\{\frac{\pi}{2},\pi,\frac{3\pi}{2}\}$ if $c>1$ or $\beta_{0}=\operatorname{constant}\in(0,2\pi)\setminus\{\frac{\pi}{2},\pi,\frac{3\pi}{2}\}$ such that $\cos\beta_{0}\in\Big(-\sqrt{\frac{c-1}{c-2}},\sqrt{\frac{c-1}{c-2}}\Big)$ if $c<1$. In the last case $\kappa^{2}_{2}=(\kappa_{1}\cot\beta_{0}\pm 1)^{2}$. A curve $\gamma$ with $\nabla_{T}T\parallel\varphi T$ is proper-biharmonic if and only if $$\Delta H=(c-(c-1)\cos^{2}\beta_{0})H,$$ where $H$ is the mean curvature vector field of $\gamma$. Biharmonic curves in $\mathbb{R}^{2n+1}(-3)$ ============================================ While proper-biharmonic Legendre curves exist only in a Sasakian space form $N^{2n+1}(c)$ with constant $\varphi$-sectional curvature $c$ bigger than $1$ if $n=1$, or $-3$ if $n>1$ (see [@Ino], [@FetcuOniciuc2]), proper-biharmonic non-Legendre curves can be found in Sasakian space forms with any $\varphi$-sectional curvature. We mention that, in the case when $c=-3$, T. Sasahara studied in [@Sas1] the submanifolds in the Sasakian space form $\mathbb{R}^{2n+1}(-3)$ whose $\varphi$-mean curvature vectors are eigenvectors of the Laplacian and in [@Sas2] the Legendre surfaces in $\mathbb{R}^{5}(-3)$ for which mean curvature vectors field are eigenvectors of the Laplacian. In this section we obtain the explicit equations for proper-biharmonic circles with $E_{2}\perp\varphi T$ and for all proper-biharmonic curves with $E_{2}\parallel\varphi T$ in $\mathbb{R}^{2n+1}(-3)$. First, let us recall briefly some notions and results about the structure of the Sasakian space form $\mathbb{R}^{2n+1}(-3)$ as they are presented in [@Blair]. Consider on $\mathbb{R}^{2n+1}(-3)$, with elements of the form $(x^{1},...,x^{n},y^{1},...,y^{n},z)$, its standard contact structure defined by the 1-form $\eta=\frac{1}{2}(dz-\sum_{i=1}^{n}y^{i}dx^{i})$, the characteristic vector field $\xi=2\frac{\partial}{\partial z}$ and the tensor field $\varphi$ given by the matrix $$\left(\begin{array}{ccc}0&\delta_{ij}&0\\ -\delta_{ij}&0&0\\ 0&y^{j}&0\end{array}\right).$$ Then $g=\eta\otimes\eta+\frac{1}{4}\sum_{i=1}^{n}((dx^{i})^{2}+(dy^{i})^{2})$ is an associated Riemannian metric and $(\mathbb{R}^{2n+1},\varphi,\xi,\eta,g)$ is a Sasakian space form with constant $\varphi$-sectional curvature equal to $-3$, denoted $\mathbb{R}^{2n+1}(-3)$. The vector fields $X_{i}=2\frac{\partial}{\partial y^{i}}$, $X_{n+i}=\varphi X_{i}=2(\frac{\partial}{\partial x^{i}}+y^{i}\frac{\partial}{\partial z})$, $i=1,...,n$, and $\xi=2\frac{\partial}{\partial z}$ form an orthonormal basis in $\mathbb{R}^{2n+1}(-3)$ and after straightforward computations one obtains $$[X_{i},X_{j}]=[X_{n+i},X_{n+j}]=[X_{i},\xi]=[X_{n+i},\xi]=0,\ \ \ [X_{i},X_{n+j}]=2\delta_{ij}\xi$$ and $$\nabla_{X_{i}}X_{j}=\nabla_{X_{n+i}}X_{n+j}=0,\ \ \nabla_{X_{i}}X_{n+j}=\delta_{ij}\xi,\ \ \nabla_{X_{n+i}}X_{j}=-\delta_{ij}\xi,$$ $$\nabla_{X_{i}}\xi=\nabla_{\xi}X_{i}=-X_{n+i},\ \ \nabla_{X_{n+i}}\xi=\nabla_{\xi}X_{n+i}=X_{i}$$ for any $i,j=1,...,n$. Now, let $\gamma:I\to \mathbb{R}^{2n+1}(-3)$ be a Frenet curve of osculating order $r>1$, parametrized by arc length, with the tangent vector field $T=\gamma'$ given by $$\label{eq:4} T=\sum_{i=1}^{n}(T_{i}X_{i}+T_{n+i}X_{n+i})+\cos\beta_{0}\xi,$$ where $\cos\beta_{0}$ is a constant. Using the above formulas for the Levi-Civita connection we have $$\label{eq:5} \nabla_{T}T=\sum_{i=1}^{n}((T'_{i}+2\cos\beta_{0}T_{n+i})X_{i}+(T'_{n+i}-2\cos\beta_{0}T_{i})X_{n+i})$$ From Theorems \[tperp\] and \[tpar\], using the same techniques as in [@Cad1], [@Cad2] and [@CIL], we get The parametric equations of proper-biharmonic circles parametrized by arc length in $\mathbb{R}^{2n+1}(-3)$, $n\geq 2$, with $\nabla_{T}T\perp\varphi T$, are $$\label{eq:6} \left\{\begin{array}{lll} x^{i}(s)&=&\pm\frac{1}{\kappa_{1}}(2\sin(\kappa_{1}s)c_{1}^{i}\mp 2\cos(\kappa_{1}s)c_{2}^{i}-\cos(2\kappa_{1}s)d_{1}^{i}\\ \\&&-\sin(2\kappa_{1}s)d_{2}^{i})+a^{i}\\ \\ y^{i}(s)&=&\frac{1}{\kappa_{1}}(2\cos(\kappa_{1}s)c_{1}^{i}\pm 2\sin(\kappa_{1}s)c_{2}^{i}+\sin(2\kappa_{1}s)d_{1}^{i}\\ \\&&-\cos(2\kappa_{1}s)d_{2}^{i})+b^{i}\\ \\ z(s)&=&\pm\frac{2}{\kappa_{1}}(1+\sum_{i=1}^{n}((c_{1}^{i})^{2}+(c_{2}^{i})^{2}))s\\ \\&&+\frac{1}{2\kappa_{1}^{2}}\sum_{i=1}^{n}(\pm\cos(4\kappa_{1}s)d_{1}^{i}d_{2}^{i}-2\cos(2\kappa_{1}s)c_{1}^{i}c_{2}^{i}\\ \\ &&+4\cos(3\kappa_{1}s)c_{2}^{i}d_{2}^{i}-4\sin(3\kappa_{1}s)c_{1}^{i}d_{2}^{i})\\ \\&&\mp\frac{1}{\kappa_{1}}\sum_{i=1}^{n}b^{i}(-2\sin(\kappa_{1}s)c_{1}^{i}\pm 2\cos(\kappa_{1}s)c_{2}^{i}\\ \\ &&+\cos(2\kappa_{1}s)d_{1}^{i}+\sin(2\kappa_{1}s)d_{2}^{i})+e \end{array}\right.,$$ where $\kappa_{1}^{2}=\cos^{2}\beta_{0}$, $\beta_{0}\in(0,2\pi)\setminus\{\frac{\pi}{2},\pi,\frac{3\pi}{2}\}$ is a constant, and $a^{i}$, $b^{i}$, $c_{1}^{i}$, $c_{2}^{i}$, $d_{1}^{i}$, $d_{2}^{i}$ and $e$ are constants such that the $n$-dimensional constant vectors $c_{j}=(c_{j}^{1},...,c_{j}^{n})$ and $d_{j}=(d_{j}^{1},...,d_{j}^{n})$, $j=1,2$, satisfy $$\left\{\begin{array}{lc}\vert c_{1}\vert^{2}+\vert c_{2}\vert^{2}+\vert d_{1}\vert^{2}+\vert d_{2}\vert^{2}=\sin^{2}\beta_{0}\\ \\ \langle c_{1},d_{1}\rangle\pm\langle c_{2},d_{2}\rangle=0,\ \langle c_{1},d_{2}\rangle\mp\langle c_{2},d_{1}\rangle=0\end{array}\right..$$ Let $\gamma:I\to \mathbb{R}^{2n+1}(-3)$ be a circle parametrized by arc length, with the tangent vector field $T=\gamma'$ given by and $\nabla_{T}T\perp\varphi T$. From the equation one obtains $$E_{2}=\frac{1}{\kappa_{1}}\sum_{i=1}^{n}((T'_{i}+2\cos\beta_{0}T_{n+i})X_{i}+(T'_{n+i}-2\cos\beta_{0}T_{i})X_{n+i})$$ and, using $g(E_{2},\varphi T)=0$, a direct computation shows that $$\begin{array}{lll} \nabla_{T}E_{2}&=&\frac{1}{\kappa_{1}}(\sum_{i=1}^{n}((T'_{i}+2\cos\beta_{0}T_{n+i})'+(T'_{n+i}-2\cos\beta_{0}T_{i})\cos\beta_{0})X_{i}\\ \\&&+((T'_{n+i}-2\cos\beta_{0}T_{i})'-(T'_{i}+2\cos\beta_{0}T_{n+i})\cos\beta_{0})X_{n+i}) \end{array}$$ and, since $\gamma$ is a circle, it follows $$\label{eq:7} \left\{\begin{array}{c} A'_{i}+B_{i}\cos\beta_{0}=0\\ \\ B'_{i}-A_{i}\cos\beta_{0}=0 \end{array}\right.,$$ where $A_{i}=\frac{1}{\kappa_{1}}(T'_{i}+2\cos\beta_{0}T_{n+i})$ and $B_{i}=\frac{1}{\kappa_{1}}(T'_{n+i}-2\cos\beta_{0}T_{i})$. Solving and imposing for $\gamma$ to be proper-biharmonic, according to Theorem \[tperp\] that is $\kappa_{1}=\pm\cos\beta_{0}>0$, we get the following equations $$\left\{\begin{array}{c} T'_{i}\pm 2\kappa_{1}T_{n+i}=\kappa_{1}\cos(\kappa_{1}s)c_{1}^{i}\pm\kappa_{1}\sin(\kappa_{1}s)c_{2}^{i}\\ \\ T'_{n+i}\mp 2\kappa_{1}T_{i}=\pm\kappa_{1}\sin(\kappa_{1}s)c_{1}^{i}-\kappa_{1}\cos(\kappa_{1}s)c_{2}^{i} \end{array}\right.,$$ which general solutions are $$\left\{\begin{array}{c} T_{i}=-\sin(\kappa_{1}s)c_{1}^{i}\pm\cos(\kappa_{1}s)c_{2}^{i}+\cos(2\kappa_{1}s)d_{1}^{i}+\sin(2\kappa_{1}s)d_{2}^{i}\\ \\ T_{n+i}=\pm\cos(\kappa_{1}s)c_{1}^{i}+\sin(\kappa_{1}s)c_{2}^{i}\pm\sin(2\kappa_{1}s)d_{1}^{i}\mp\cos(2\kappa_{1}s)d_{2}^{i} \end{array}\right.,$$ where $c_{1}^{i}$, $c_{2}^{i}$, $d_{1}^{i}$ and $d_{2}^{i}$ are constants, such that $$\left\{\begin{array}{lc}\sum_{i=1}^{n}((c_{1}^{i})^{2}+(c_{2}^{i})^{2}+(d_{1}^{i})^{2}+(d_{2}^{i})^{2})=\sin^{2}\beta_{0}\\ \\ \sum_{i=1}^{n}((c_{1}^{i})(d_{1}^{i})\pm(c_{2}^{i})(d_{2}^{i}))=0,\ \ \ \sum_{i=1}^{n}((c_{1}^{i})(d_{2}^{i})\mp(c_{2}^{i})(d_{1}^{i}))=0\end{array}\right.,$$ since $g(T,T)=1$. Finally, replacing into expression of $\gamma'$ and integrating we get . In order to find explicit examples of proper-biharmonic curves with $\nabla_{T}T\perp\varphi T$ in $\mathbb{R}^{2n+1}(-3)$ we will stick at proper-biharmonic circles since the computations in the case of helices are rather complicated. Proper-biharmonic curves in $\mathbb{R}^{2n+1}(-3)$, with $\nabla_{T}T\parallel\varphi T$, are either a\) Proper-biharmonic circles given by $$\label{eq:8} \left\{\begin{array}{ll} x^{i}(s)=&(\sqrt{5}+1)\Big(\cos\Big(\frac{\sqrt{5}-1}{2}s\Big)c_{1}^{i}+\sin\Big(\frac{\sqrt{5}-1}{2}s\Big)c_{2}^{i}\Big)+a^{i}\\ \\ y^{i}(s)=&(\sqrt{5}+1)\Big(\sin\Big(\frac{\sqrt{5}-1}{2}s\Big)c_{1}^{i}-\cos\Big(\frac{\sqrt{5}-1}{2}s\Big)c_{2}^{i}\Big)+b^{i}\\ \\ z(s)=&\frac{1-\sqrt{5}\pm 2\sqrt{1+\sqrt{5}}}{2}s+\frac{3+\sqrt{5}}{2}\sum_{i=1}^{n}(((c_{1}^{i})^{2}-(c_{2}^{i})^{2})\sin((\sqrt{5}-1)s)\\ \\&-2\cos((\sqrt{5}-1)s)c_{1}^{i}c_{2}^{i})+(1+\sqrt{5})\sum_{i=1}^{n}b_{i}\Big(\sin\Big(\frac{\sqrt{5}-1}{2}s\Big)c_{2}^{i}\\ \\&+ \cos\Big(\frac{\sqrt{5}-1}{2}s\Big)c_{1}^{i}\Big)+d \end{array}\right.,$$ where $a^{i}$, $b^{i}$, $c_{1}^{i}$, $c_{2}^{i}$ and $d$ are constants such that the $n$-dimensional constant vectors $c_{j}=(c_{j}^{1},...,c_{j}^{n})$, $j=1,2$, satisfy $$\vert c_{1}\vert^{2}+\vert c_{2}\vert^{2}=\frac{3-\sqrt{5}}{4}.$$ or b\) Proper-biharmonic helices given by $$\label{eq:9} \left\{\begin{array}{ll} x^{i}(s)=&-\frac{2\kappa_{1}}{\kappa_{1}\pm\sin(2\beta_{0})}\Big(\cos\Big(\frac{\kappa_{1}\pm\sin(2\beta_{0})}{\kappa_{1}}s\Big)c_{1}^{i} +\sin\Big(\frac{\kappa_{1}\pm\sin(2\beta_{0})}{\kappa_{1}}s\Big)c_{2}^{i}\Big)+a^{i}\\ \\ y^{i}(s)=&\frac{2\kappa_{1}}{\kappa_{1}\pm\sin(2\beta_{0})}\Big(\sin\Big(\frac{\kappa_{1}\pm\sin(2\beta_{0})}{\kappa_{1}}s\Big)c_{1}^{i} -\cos\Big(\frac{\kappa_{1}\pm\sin(2\beta_{0})}{\kappa_{1}}s\Big)c_{2}^{i}\Big)+b^{i}\\ \\z(s)=&2\Big(\cos\beta_{0}+\frac{\kappa_{1}\sin^{2}\beta_{0}}{\kappa_{1}\pm\sin(2\beta_{0})}\Big)s+ \frac{\kappa^{2}_{1}}{(\kappa_{1}\pm\sin(2\beta_{0}))^{2}}\\ \\&\cdot\Big(\sin\Big(\frac{2(\kappa_{1}\pm\sin(2\beta_{0}))}{\kappa_{1}}s\Big)\sum_{i=1}^{n}((c_{1}^{i})^{2}-(c_{2}^{i})^{2})\\ \\&+\cos\Big(\frac{2(\kappa_{1}\pm\sin(2\beta_{0}))}{\kappa_{1}}s\Big)\sum_{i=1}^{n}(c_{1}^{i}c_{2}^{i})\Big)\\ \\&-\frac{2\kappa_{1}}{\kappa_{1}\pm\sin(2\beta_{0})}\sum_{i=1}^{n}b^{i}\Big(\cos\Big(\frac{\kappa_{1}\pm\sin(2\beta_{0})}{\kappa_{1}}s\Big)c_{1}^{i} \\ \\&+\sin\Big(\frac{\kappa_{1}\pm\sin(2\beta_{0})}{\kappa_{1}}s\Big)c_{2}^{i}\Big)+d \end{array}\right.,$$ where $\beta_{0}\in(0,2\pi)\setminus\{\frac{\pi}{2},\pi,\frac{3\pi}{2}\}$ is a constant such that $\cos\beta_{0}\in\Big(-1,-\frac{2\sqrt{5}}{5}\Big)\cup\Big(\frac{2\sqrt{5}}{5},1\Big)$, $\kappa_{1}$ is a positive solution of the equation $$\kappa_{1}^{2}\pm\sin(2\beta_{0})\kappa_{1}+4\sin^{4}\beta_{0}=0$$ and $a^{i}$, $b^{i}$, $c_{1}^{i}$, $c_{2}^{i}$ and $d$ are constants such that $$\vert c_{1}\vert^{2}+\vert c_{2}\vert^{2}=\sin^{2}\beta_{0}.$$ We will prove only the first statement because the second one can be obtained in a similar way by the meaning of Theorem \[tpar\]. Assume that $\gamma$ is a proper-biharmonic circle in $\mathbb{R}^{2n+1}(-3)$ parametrized by arc length, such that $\nabla_{T}T\parallel\varphi T$. Then, from and since $\varphi T=\sum_{i=1}^{n}(-T_{n+i}X_{i}+T_{i}X_{n+i})$, $g(\varphi T,\varphi T)=\sin^{2}\beta_{0}$, where $\eta(T)=\cos\beta_{0}$, one obtains $$T_{i}'=\Big(\mp\frac{\sin(2\beta_{0})}{\kappa_{1}}-1\Big)T_{n+i},\ \ \ \ T_{n+i}'=\Big(\pm\frac{\sin(2\beta_{0})}{\kappa_{1}}+1\Big)T_{i}$$ Now, since $\gamma$ is a proper-biharmonic circle we get, from Theorem \[tpar\], $\kappa_{1}=\mp\tan\beta_{0}>0$ and $\cos^{2}\beta_{0}=\frac{1+\sqrt{5}}{4}$ and hence the above equations become $$T_{i}'=\frac{\sqrt{5}-1}{2}T_{n+i},\ \ \ \ T_{n+i}'=\frac{1-\sqrt{5}}{2}T_{i},$$ with general solutions $$T_{i}=\cos\Big(\frac{\sqrt{5}-1}{2}\Big)c_{1}^{i}+\sin\Big(\frac{\sqrt{5}-1}{2}\Big)c_{2}^{i}, \ \ T_{n+i}=\cos\Big(\frac{\sqrt{5}-1}{2}\Big)c_{2}^{i}-\sin\Big(\frac{\sqrt{5}-1}{2}\Big)c_{1}^{i},$$ where $c_{1}^{i}$ and $c_{2}^{i}$, $i=1,...,n$, are constants. Replacing in the expression of $T=\gamma'$, integrating and imposing $g(T,T)=1$ we obtain the conclusion. [99]{} A. Arvanitoyeorgos, F. Defever, G. Kaimakamis, V.J. Papantoniou. *Biharmonic Lorentz hypersurfaces in $E_{1}^{4}$*, Pacific J. Math., 229(2007), 293–305. A. Balmuş, S. Montaldo, C. Oniciuc. *Classification results for biharmonic submanifolds in spheres*, Israel J. Math., to appear. D.E. Blair. *Riemannian Geometry of Contact and Symplectic Manifolds*, Birkhäuser Boston, Progress in Mathematics, Volume 203, 2002. R. Caddeo, C. Oniciuc, P. Piu. *Explicit formulas for non-geodesic biharmonic curves of the Heisenberg group*, Rend. Sem. Mat. Univ. e Politec. Torino, 62(2004), 265–278. R. Caddeo, S. Montaldo, C. Oniciuc, P. Piu. *The clasification of biharmonic curves of Cartan-Vranceanu 3-dimensional spaces*, Modern Trends in Geometry and Topology, Deva, September 5-11, 2005, 121–131. B.Y. Chen. *A report on submanifolds of finite type*, Soochow J. Math., 22(1996), 117–337. J.T. Cho, J. Inoguchi, J.-E. Lee. *Biharmonic curves in 3-dimensional Sasakian space form*, Ann. Math. Pura Appl., to appear. J. Eells, J.H. Sampson. *Harmonic mappings of Riemannian manifolds*, Amer. J. Math., 86(1964), 109–160. D. Fetcu, C. Oniciuc. *Explicit formulas for biharmonic submanifolds in non-Euclidean 3-spheres*, Abh. Math. Semin. Univ. Hamburg, 77(2007), 179–190. D. Fetcu, C. Oniciuc. *Explicit formulas for biharmonic submanifolds in Sasakian space forms*, Pacific J. Math., 240(1)(2009), 85–107. T. Ichiyama, J. Inoguchi, H. Urakawa. *Bi-harmonic maps and bi-Yang-Mills fields*, preprint. J. Inoguchi. *Submanifolds with harmonic mean curvature in contact 3-manifolds*, Colloq. Math., 100(2004), 163–179. G.Y. Jiang. *2-harmonic maps and their first and second variational formulas*, Chinese Ann. Math. Ser. A7(4)(1986), 389–402. S. Montaldo, C. Oniciuc. *A short survey on biharmonic maps between Riemannian manifolds*, Rev. Un. Mat. Argentina, 47(2)(2006), 1–22. Y-L. Ou, Z-P. Wang. *Biharmonic maps into Sol and Nil spaces*, `arXiv:math.DG/0612329v1`. T. Sasahara. *Submanifolds in a Sasakian manifold $\mathbb{R}^{2n+1}(-3)$ whose $\phi$-mean curvature vectors are eigenvectors*, J. Geom., 75(2002), 166–178. T. Sasahara. *Spectral decomposition of mean curvature vector fields of surfaces in a Sasakian manifold $\mathbb{R}^{2n+1}(-3)$*, Results Math., 43(1–2)(2003), 168–180. T. Sasahara. *Legendre surfaces in Sasakian space forms whose mean curvature vectors are eigenvectors*, Publ. Math. Debrecen, 67(3–4)(2005), 285–303. The Bibliography of Biharmonic Maps. `http://people.unica.it/ biharmonic/publication/`. W. Zhang. *New examples of biharmonic submanifolds in $\mathbb{C}P^{n}$ and $\mathbb{S}^{2n+1}$*, `arXiv:math.DG/0705.3961v1`. [^1]: The author wants to thank to Prof. C. Oniciuc for many discussions and useful suggestions. [^2]: The author was partially supported by the Grant CEEX, ET, 5871/2006, Romania.

ArXiv

--- abstract: 'We studied ferromagnetism in the one-dimensional Hubbard model with doubly degenerate atomic orbitals by means of the density-matrix renormalization-group method and obtained the ground-state phase diagrams. It was found that ferromagnetism is stable from low to high ($0< n < 1.75$) electron density when the interactions are sufficiently strong. Quasi-long-range order of triplet superconductivity coexists with the ferromagnetic order for a strong Hund coupling region, where the inter-orbital interaction $U''-J$ is attractive. At quarter-filling ($n=1$), the insulating ferromagnetic state appears accompanying orbital quasi-long-range order. For low densities ($n<1$), ferromagnetism occurs owing to the ferromagnetic exchange interaction caused by virtual hoppings of electrons, the same as in the quarter-filled system. This comes from separation of the charge and spin-orbital degrees of freedom in the strong coupling limit. This ferromagnetism is fragile against variation of band structure. For high densities ($n>1$), the phase diagram of the ferromagnetic phase is similar to that obtained in infinite dimensions. In this case, the double exchange mechanism is operative to stabilize the ferromagnetic order and this long-range order is robust against variation of the band-dispersion. A partially polarized state appears in the density region $1.68 \nearl n \nearl 1.75$ and phase separation occurs for $n$ just below the half-filling ($n=2$).' author: - 'Harumi Sakamoto[@address], Tsutomu Momoi' - Kenn Kubo title: | Ferromagnetism in the one-dimensional Hubbard model with orbital degeneracy:\ From low to high electron density --- Introduction ============ The Hubbard model has been employed for a long time as a standard model for metallic ferromagnetism of itinerant electrons[@Kanamori; @Gutzwiller; @Hubbard]. It is, however, not easy to show that the Hubbard model simulates a metallic ferromagnet. Many investigations revealed that a simple single-band Hubbard model on a hyper-cubic lattice may not exhibit ferromagnetism. Some additional features in the model seem to be necessary in order to stabilize ferromagnetism. One of such features is a special lattice or band structure. Lieb first showed that a half-filled flat band induces a net magnetization[@lieb]. Then Mielke and Tasaki introduced other examples of flat-band ferromagnets[@mielke; @tasaki1]. Tasaki proved, furthermore, that a class of models with half-filled lowest bands with finite band dispersions exhibits ferromagnetism for sufficiently strong intra-atomic repulsions[@tasaki3]. Müller-Hartmann found ferromagnetism in a one-dimensional system where the band dispersion has two minima[@muller-hartmann]. It was also shown by perturbational as well as numerical methods that partially filled bands in such systems can realize ferromagnetism[@penc; @sakamoto; @pieri; @Daul]. Another important factor realizing ferromagnetism was discussed to be degeneracy of atomic orbitals[@Slater; @Zener; @Van]. Actually magnetic carriers of typical metallic ferromagnets originate from the degenerate 3$d$-orbitals, though they are mixed with $s$- or $p$-orbitals. In the presence of degenerate orbitals, Hund coupling favors local triplet spin configurations of two electrons occupying different orbitals at the same site. Following mechanisms were proposed which lead to bulk ferromagnetism from intra-atomic Hund coupling. Zener proposed the so-called [*double exchange*]{} mechanism originally for doped manganites, which works when the electron density per site is not an integer[@Zener; @Anderson-H], e.g., the density of electrons per site ($n$) is between $I$ and $I+1$. If electrons are strongly correlated, each lattice site is occupied by either $I$ or $I+1$ electrons, and Hund coupling forces their spins to be parallel in an atom. The magnitude of the formed atomic spin is $I/2$ or $(I+1)/2$ for $I\le N_{\rm d}-1$, and $N_{\rm d}-I/2$ or $N_{\rm d}-(I+1)/2$ for $I\ge N_{\rm d}$, where $N_{\rm d}$ denotes the degree of degeneracy. An electron can hop from a site occupied by $I+1$ electrons to a site occupied by $I$ electrons. Since the hopping matrix is diagonal with respect to the spin indices, an electron can hardly hop between sites whose spins are directed antiparallelly, while it can freely propagate among sites with parallel spins. As a result, ferromagnetic spin alignment that lowers the kinetic energy is favored. Another mechanism leading to ferromagnetism is most effective when the electron density per site is an integer, $n=I$, where $I$ is an integer which satisfies $1\le I\le N_{\rm d}-1$ or $N_{\rm d}+1\le I \le 2N_{\rm d}-1$[@Roth]. Each site is occupied by $I$ electrons in the strong coupling regime and hoppings of electrons occur as virtual processes. The intermediate state of a second order process has the lowest energy if the spins are aligned parallelly. This leads to an effective ferromagnetic exchange interaction that induces orbital ordering at the same time. Van Vleck argued that the ferromagnetic exchange due to virtual hoppings might be the origin of the metallic ferromagnetism in Ni[@Van]. Though the mechanisms favoring ferromagnetism can be qualitatively understood as above, it is far from trivial whether ferromagnetic long-range order occurs in bulk systems. We need to examine whether the models with orbital degeneracy really simulate ferromagnets by reliable methods. For this purpose, we study the simplest model with orbital degeneracy, the Hubbard model with doubly degenerate atomic orbitals. The model is described by the Hamiltonian $$\begin{aligned} H &=& -\sum_{i<j} t_{ij} \sum_{m=A,B} \sum_{\sig=\uparrow,\downarrow} (c_{im\sigma}^{\dagger}c_{jm\sigma}+{\rm H.c.}) \nonumber \\ & & +U\sum_{i}\sum_{m=A,B} n_{im\uparrow}n_{im\downarrow} +U'\sum_{i}\sum_{\sigma=\uparrow,\downarrow}n_{iA\sigma}n_{iB-\sigma} \nonumber \\ & & +(U'-J)\sum_{i}\sum_{\sigma=\uparrow,\downarrow} n_{iA\sigma}n_{iB\sigma} \nonumber \\ & & -J\sum_{i} (c_{iA\uparrow}^{\dagger}c_{iA\downarrow} c_{iB\downarrow}^{\dagger}c_{iB\uparrow}+{\rm H.c.}) \nonumber \\ & & -J'\sum_{i} (c_{iA\uparrow}^{\dagger}c_{iA\downarrow}^{\dagger} c_{iB\uparrow}c_{iB\downarrow}+{\rm H.c.}), \label{eq:deghub}\end{aligned}$$ where $c_{im\sigma}^{\dagger}$ $(c_{im\sigma})$ creates (annihilates) an electron at the site $i$ with spin $\sigma\,(=\uparrow,\,\downarrow$) in the orbital $m\,(=A,\,B)$. Here we assumed that only the hoppings between the same orbitals exist, and the Coulomb interaction works only between electrons on a same site. The interaction parameters $U$, $U'$, $J$ and $J'$ represent the intra- and the inter-orbital repulsion, the exchange (Hund-rule) coupling, and the pair hopping, respectively. Particle-hole symmetry allows us to study only the case with $0\le n\le 2$. Quite a few previous studies exist on the doubly degenerate Hubbard model. One must be careful in comparing them, since different assumptions were employed for the interaction parameters. Many assumed that $J'=0$. In fact, these parameters are dependent on each other if we calculate them from orbital wave functions. The relation $$J = J' \label{eq:rel1}$$ holds when the orbital wave functions are real. In addition, the interaction parameters $U,$ $U'$ and $J$ satisfy the relation $$U = U'+2J \label{eq:rel2}$$ for $e_{g}$ orbitals which is relevant for the cubic symmetry. This Hamiltonian has the spin SU(2) and charge U(1) symmetries. In addition, orbital degrees of freedom have U(1) rotational symmetry since $$[H,\sum_i \tau_i^y]=0.$$ Here the orbital pseudo-spin operator is defined as $\tau_i^y=\frac{1}{2}\sum_{\sigma=\uparrow,\downarrow} {\bf C}_{i\sigma}^\dagger \sigma_i^y {\bf C}_{i\sigma}$, where $\sigma_i^y$ denotes Pauli’s matrix and ${\bf C}_{i\sigma}^\dagger=(c_{iA\sigma}^\dagger, c_{iB\sigma}^\dagger)$. The last symmetry holds if the coupling parameters satisfy the relations (\[eq:rel1\]) and (\[eq:rel2\]), and it could be easily seen in the effective Hamiltonian derived in the strong coupling limit[@MomoiK]. Because of this rotational symmetry, a rigorous argument for one-dimensional systems excludes oscillating orbital LRO corresponding to $\tau_i^z$-correlations[@Momoi]. In the limit of $J(=J')=0$, the symmetry of orbital degrees of freedom becomes SU(2) invariant and the whole Hamiltonian has the high symmetry SU(4). We assume the relations (\[eq:rel1\]) and (\[eq:rel2\]), and assume all interaction parameters to be positive throughout this work. In realistic situations the parameters satisfy $U > U' > J $, but we study the strong Hund coupling case $J \ge U'$ as well, for theoretical interest. Let us explain the level structure of a doubly occupied atomic site since it rules the physics of the model (\[eq:deghub\]). The spin-triplet states, where two electrons occupy different orbitals, have the lowest energy $U'-J$. In one of three spin-singlet states, electrons occupy different orbitals and the energy is $U'+J$. In the other two singlets with the energies $U-J'(=U'+J)$ and $U+J'(= U'+3J)$, electrons occupy the same orbital. In the strongly correlated regime where $U' > J \gg |t_{ij}|$, almost all sites are empty or singly occupied for $n<1$, while they are singly or doubly occupied for $n>1$ and spins in a doubly occupied site form a triplet. Thus the ground state properties for $U'>J$ strongly depend on whether $n$ is less than, equal to or more than unity. If $J$ is larger than $U'$, a doubly occupied site has negative interaction energy and hence electrons tend to form bound triplet pairs. In this case, sites are empty or doubly occupied for $0<n<2$ if $N_{\rm e}$ is even. If $N_{\rm e}$ is odd, there can be an unpaired electron, which will cause a strong finite-size effect. Roth examined the model in three dimensions in the quarter-filled case ($n=1$)[@Roth]. She found that the ferromagnetic ground state has an orbital superlattice structure in which two sublattices are occupied by electrons of different orbitals when the interaction is strong. This kind of ground state was studied by using an effective Hamiltonian in the strong coupling limit[@Kugel; @Cyrot; @Inagaki] and by means of the mean field theory for general electron density[@Inagaki-K]. Several exact diagonalization studies were performed for small clusters in one dimension[@kusakabe3; @kusakabe2; @Gill; @Kuei; @Hirsch]. Most of them investigated the quarter-filled case ($n=1$) and found the ferromagnetic ground state for $U'\nearg J$. Hirsch studied the less-than-quarter-filled case and found the ferromagnetic ground state. Their results, however, depend strongly on the boundary conditions and the number of lattice sites. There are rigorous proofs for the ferromagnetic ground state in strong coupling limits in one dimension[@Kubo; @kusakabe2; @Shen]. One must be careful for that these proofs are valid in different limits of strong coupling. For the strong Hund coupling case ($U\rightarrow \infty$ and $U'>J\rightarrow \infty$), existence of ferromagnetism is proved for $1<n<2$[@Kubo; @kusakabe2]. In the special limit $J=U' \rightarrow \infty$ and $U\rightarrow \infty$, ferromagnetism is stable for $0<n<2$[@kusakabe2]. Shen obtained a rather general result that the ground state is fully spin-polarized for any $n$ between 0 and 2 except for 1 if $U=\infty$, and $U'(>0)$ and $J=J'(>0)$ are finite[@Shen]. The present model (\[eq:deghub\]) was also studied in infinite dimensions using the dynamical mean-field theory[@MomoiK; @Held]. Under the assumptions (\[eq:rel1\]) and (\[eq:rel2\]), the ferromagnetic ground state with and without orbital order were found for $n=1$ and $n>1$, respectively, while ferromagnetism was not found for $n<1$ up to the interactions of $U=40$ and $J\le U'=20$[@MomoiK; @difference]. Bünemann [*et al.*]{} studied ferromagnetism in the three-dimensional two-band model with rather realistic DOS for Ni using a Gutzwiller approximation. Their results showed that increase of DOS at the Fermi level could stabilize ferromagnetism for electron-doped cases ($2>n>1$)[@Bunemann]. In this paper, we studied the ground state of the model (\[eq:deghub\]) in one dimension under open boundary condition, using the density-matrix renormalization group (DMRG) method[@white]. This method enables us to study larger systems than those studied previously. Owing to the open boundary condition, we found remarkably little size-dependence. If we apply the periodic boundary condition, electrons can exchange their positions turning around the chain and it causes very large size-dependence of the ground state (e.g. even-odd oscillations). Preliminary results of this study were reported previously in Ref. . This paper is constructed as follows: In Section \[sec:method\], we briefly review the DMRG method. In Section \[sec:main-result\], we study the present model with only nearest-neighbor hoppings and report our numerical results, such as magnetic phase diagrams. Ferromagnetic phases appear for a wide parameter region and for all electron density ($0<n<2$). In most of density region, the ferromagnetic states are fully polarized, whereas partially polarized ferromagnetic states are found in a small region ($1.68\le n \le1.75$). We also found coexistence of metallic ferromagnetism and quasi-long-range order of triplet superconductivity when the Hund coupling is stronger than the inter-orbital Coulomb repulsion ($J>U'$). Near the half-filling case ($1.75<n<2$), phase separation occurs, and the paramagnetic (Haldane) phase and ferromagnetic one coexist. In Section \[sec:sub-result\], we examine the stability of the ferromagnetic states obtained in Section \[sec:main-result\], adding perturbation to the density of states. The metallic ferromagnetic order for more than quarter-filling ($n > 1$) is stable against this perturbation, but, for less than quarter-filling ($n<1$), it easily becomes unstable by this variation. Finally, we give summary and discussion in Section \[sec:summary\]. Appendixes contain rigorous results in the strong coupling limit. Using charge and spin-orbital separation, we show that the ferromagnetic spin state for less than quarter-filling is the same as one at quarter-filling. The ferromagnetism at less than quarter-filling is created by the exchange interaction produced by virtual second-order hopping processes. Method {#sec:method} ====== We obtained the ground state of the model (\[eq:deghub\]) with up to 62 sites using the density matrix renormalization group (DMRG) method[@white]. The method is standard and we will not describe it in detail. Readers should refer to White’s paper for the DMRG method. We note only several points particular to this study. We employed the open boundary condition throughout this paper. In the DMRG calculations, we always kept rotational symmetry in the spin space so that the obtained ground state is an eigenstates of the total spin[@sakamoto]. We used the finite-system method of DMRG to obtain the ground-state phase diagrams of the model with up to 16 sites. This method is superior to the infinite-system method in terms of accuracy of the ground-state quantities of finite systems. Another merit of this method is that the final result does not depend so strongly on the ground state of the initial small system and on the path of the renormalization process as in the infinite-system method. Repeated sweeps of the renormalization processes in the finite-system method can usually remedy the failure due to a wrong initial state. Near the phase boundary between a paramagnetic phase and a ferromagnetic one with saturated magnetization, however, the final results sometimes depend on the choice of the renormalization paths. For example, two different choices of the paths $(N,N_{\rm e})=(4, 2) (6, 3) (8, 4) (10, 5) (12, 6) (14, 7) (16, 8)$ and $(N,N_{\rm e})=(4, 2) (6, 4) (8, 4) (10, 6) (12, 6) (14, 8) (16, 8)$ give a perfect ferromagnetic ground state and a paramagnetic one, respectively, at $(N, N_{\rm e}) = (16, 8)$ for $U'=J=2|t|$, where $N_{\rm e}$ and $N$ are the number of electrons and lattice sites, respectively. Since the transition is of first order, the energy levels of a paramagnetic state and a ferromagnetic one naturally exist very close to each other near the phase boundary. To study the phase diagram carefully, we searched the ground state through many renormalization paths. Finally, we regard the lowest energy state as the ground state. Hence, we need to try several paths and initial states and compare their results in order to obtain a reliable phase diagram as long as the number $m_{\rm D}$ of reserved eigenstates of the density matrix is finite. The maximum number of $m_{\rm D}$ was chosen to be 200 in this study for technical reasons. In order to calculate longer correlations we also studied the systems with up to 62 sites using the infinite-system DMRG method. Results of short chains were obtained by using the exact diagonalization method. Results for the nearest-neighbor model {#sec:main-result} ====================================== In this section, we present the results for the model with only nearest neighbor hoppings $(t_{ij}=t\delta_{i,j\pm 1})$. Quarter-filled system --------------------- We first exhibit the results in the quarter-filled case ($n=1$). Figure \[fig:mgd-q\] shows the $(U'/t,J/t)$ phase diagram of the ground state of the quarter-filled system obtained for $N=4$ and 8. The difference between the results for $N=4$ and $N=8$ is amazingly small. We hence believe that the present results are representing the phase diagram of an infinite system, and did not study larger systems at this electron density. A ferromagnetic ground state with full spin polarization appears around $J \simeq U'$ for $U'/t \nearg 5$. The parameter regions $J/U'\gg 1$ and $J/U' \ll 1$ are paramagnetic. No partially polarized state was found at this filling. The present result is similar to those by previous studies with open boundary conditions though different assumptions for parameters were employed[@kusakabe3; @kusakabe2]. In previous studies employing periodic or anti-periodic boundary conditions, the phase boundaries of their ferromagnetic phases depend on the system size greatly[@kusakabe3; @Gill; @Hirsch]. ![Ground-state phase diagram at quarter-filling ($n=1$) of the model with nearest-neighbor hoppings. Interaction parameters are set as $U=U'+2J$ and $J=J'$. The symbol “$S_{\rm tot}={\rm Max.}$” denotes the perfect ferromagnetic phase, and “$S_{\rm tot}=0$” the paramagnetic one.[]{data-label="fig:mgd-q"}](diagram-q){width="7.5cm"} The appearance of the ferromagnetism in the region $U'>J$ can be well understood in terms of the effective Hamiltonian in the strong coupling limit ($U'-J \gg t$). At quarter-filling, every site is singly occupied in this limit, and only the spin and orbital degrees of freedom remain. The second order effects of virtual hoppings between singly occupied sites lead to an effective Hamiltonian ${\cal H}_{\rm eff}$ for the spin and orbital degrees of freedom[@Kugel; @Cyrot; @Inagaki; @MomoiK]: $$\begin{aligned} \lefteqn{{\cal H}_{\rm eff} = \sum_{i=1}^{N-1} h(i,i+1)}, \label{eqn:quarter}\\ & &h(i,j) \equiv -2t^{2}\left\{\frac{4U}{U^{2}-J'^{2}} \left(\frac{1}{4}+\tau_{i}^{z}\tau_{j}^{z}\right) \left(\frac{1}{4}-\mbox{\bold$S$}_{i}\cdot\mbox{\bold$S$}_{j}\right) \right. \nonumber \\ & & -\frac{2J'}{U^{2}-J'^{2}} \left(\tau_{i}^{-}\tau_{j}^{-}+\tau_{i}^{+}\tau_{j}^{+}\right) \left(\frac{1}{4}-\mbox{\bold$S$}_{i}\cdot\mbox{\bold$S$}_{j}\right) \label{eqn:h-func} \\ & & +\frac{2U'}{U'^{2}-J^{2}} \left[\frac{1}{4}-\tau_{i}^{z}\tau_{j}^{z} -2\left(\mbox{\bold$\tau$}_{i}\cdot\mbox{\bold$\tau$}_{j} -\tau_{i}^{z}\tau_{j}^{z}\right) \left(\frac{1}{4}+\mbox{\bold$S$}_{i}\cdot\mbox{\bold$S$}_{j}\right)\right] \nonumber \\ & & +\left.\frac{2J}{U'^{2}-J^{2}} \left[\tau_{i}^{z}\tau_{j}^{z} -\mbox{\bold$\tau$}_{i}\cdot\mbox{\bold$\tau$}_{j} +2\left(\frac{1}{4}-\tau_{i}^{z}\tau_{j}^{z}\right) \left(\frac{1}{4}+\mbox{\bold$S$}_{i}\cdot\mbox{\bold$S$}_{j}\right)\right] \right\}. \nonumber\end{aligned}$$ Here $\mbox{\bold $S$}_i$ is a spin operator and $\mbox{\bold$\tau$}_{i}=(\tau_{i}^{x},\tau_{i}^{y},\tau_{i}^{z})$ is a pseudo-spin operator representing the orbital degrees of freedom, e.g., $\tau_{i}^{+}$ is defined as $\sum_{\sig}c_{iA\sig}^{+}c_{iB\sig}$. According to (\[eqn:h-func\]) the exchange energy between the sites $i$ and $i+1$ is $-2t^{2}/(U'-J)$, $-2t^{2}/(U'+J)$ and $-2Ut^{2}/(U^{2}-J^{2})$ for the states $|m,\sig\ra_{i}|\barm, \sig\ra_{i+1}$, $|m,\sig\ra_{i}|\barm,-\sig\ra_{i+1}$ and $|m,\sig\ra_{i}|m,-\sig\ra_{i+1}$, respectively, where $|m,\sig\ra_{i}$ denotes the state of the site $i$ occupied by an electron of orbital $m$ with spin $\sigma$, and $\barm$ labels the complement of $m$, e.g., ${\bar A} = B$. As a result, ${\cal H}_{\rm eff}$ favors ferromagnetism with an orbital antiferromagnetic superlattice structure, where electrons on two sublattices occupy different orbitals. Numerical results of the four sites system described by ${\cal H}_{\rm eff}$ exhibited that the ground state of ${\cal H}_{\rm eff}$ is ferromagnetic for $J>0.35U'$. This agrees with our observation that the lower boundary of the ferromagnetic phase in Fig. \[fig:mgd-q\] approaches the line $J=0.35U'$ for large $U'/t$. We may consider that the system is well described by ${\cal H}_{\rm eff}$ for $U'/t \nearg 10$ and $U' > J$. A numerical study by two of present authors in infinite dimensions showed that the ferromagnetism with antiferromagnetic pseudo-spin order is stable for $J>0$ in the strong coupling limit[@MomoiK], which agrees with the result in the classical limit of ${\cal H}_{\rm eff}$. The ferromagnetic phase is smaller in one dimension than that in infinite dimensions due to stronger quantum fluctuations. In one dimension, the antiferromagnetic orbital correlations, which supports the ferromagnetism, do not grow to a real long-range order but remains as a quasi-long-range order (QLRO), since orbital degrees of freedom have U(1) rotational symmetry and a rigorous argument excludes oscillating orbital LRO in one dimension[@Momoi]. In the ferromagnetic phase at $n=1$, we found only fully polarized states. Note that the charge and orbital degrees of freedom of fully polarized states can be described by a single-band Hubbard model with intra-atomic repulsion $U'-J$, where the conventional spin degrees of freedom are replaced with the orbital ones[@Shen]. Hence, we can conclude that this ferromagnetic state has orbital antiferromagnetic QLRO for $U'-J>0$. In addition, for general filling $n$, one can write down the asymptotic form of the orbital correlation function in the fully polarized ferromagnetic phase in the form $$\label{eq:orbital_correlation} \langle \tau_0^z \tau_l^z \rangle \sim \cos(n\pi l)|l|^{-\alpha_S}$$ for $U'-J>0$. The exponent $\alpha_S(=1+K_\rho)$ is obtained from the single-band Hubbard model with coupling $U'-J$ at filling $n$[@Schulz; @Kawakami]. We also checked the orbital ordering by the DMRG method and presented the results in Fig.\[fig:orbital\]. ![Correlation functions of orbital ordering $\la \tau_{i} \tau_{j} \ra$ in the ferromagnetic ground state for $J/t=15$ and $U'/t=20$ at filling $n=1$ and $n=0.5$. The system size is 34 and $i$ is fixed at $N/2=17$.[]{data-label="fig:orbital"}](orbital){width="7.5cm"} ![Electron density $n$ per site as a function of the chemical potential $\mu$ for $U'/t=15$ and $J/t=10$ for $N=16$. The flat parts indicate insulating states. Abbreviations indicate the characters of the ground states as FM: ferromagnetic metal, FI: ferromagnetic insulator, FM(P): partially ferromagnetic metal and PI: paramagnetic insulator.[]{data-label="fig:m-in"}](n-myu){width="7.5cm"} Also, from the above mapping, we can see that a metal-insulator transition occurs in the ferromagnetic phase on the line $U'=J$, i.e., the ferromagnetic state is insulating for $U'>J$ and metallic for $U'\le J$. (Figure \[fig:m-in\] shows chemical potential dependence of the electron density for $U'/t=15$ and $J/t=10$ with $N=16$.) On the other hand, in the paramagnetic phase, since each band is quarter-filled and the perfect nesting no longer exists, a metal-insulator transition may not occur with an infinitesimal repulsion $U'-J$, which works between electrons with the same spin in different orbitals. Indeed, for paramagnetic states at $J=0$, Assaraf [*et al*]{}. numerically showed that a metal-insulator transition occurs at $U_{\rm c}/t\sim 2.8$[@Assaraf]. We expect that, in the paramagnetic phase with finite $J(>0)$, a metal-insulator transition may occur at a finite positive value of $U'-J$. Figure \[fig:n-m2\] shows numerical results of the electron density as a function of the chemical potential in the paramagnetic phase with $J/t=3$ and $U'/t=3$, 4, and 8. The data for $U'/t=8$ clearly show that the quarter-filled state is insulating. The data for $U'/t=3$ and 4 seem to indicate that the systems are metallic at quarter filling, though it is hard to estimate the critical point from the present calculation because of finite-size effects. ![Electron density ($n$) per site as a function of the chemical potential ($\mu$) for $U'/t=3$, 4, and 8, and $J/t=3$ for $N=16$. The flat parts indicate insulating states. All states are paramagnetic.[]{data-label="fig:n-m2"}](n16j3ud348-2){width="7.5cm"} For $J=0$, the effective Hamiltonian (\[eqn:quarter\]) reduces to a model with $SU(4)$ symmetry. It is known that the ground state of this model is a spin-singlet state $(S=\tau=0)$ in one dimension[@Sutherland]. The paramagnetic state obtained for $ 0.35U' > J >0$ in Fig.\[fig:mgd-q\] may be interpreted as an extension of this $SU(4)$ symmetric state. We find that, for the $U'=J$ case, the ferromagnetic phase extends down to a weak coupling region. At $U'=J$, double occupation of a site does not cost energy and the ferromagnetic ground state is metallic. The ferromagnetism in the strong coupling region is stabilized by a kind of the double exchange mechanism as was discussed rigorously for $U'=J=\infty$[@kusakabe2]. The ferromagnetic phase extends to the parameter region $J>U'$. In this region, the fully polarized states can be described by the attractive single-band Hubbard model[@Shen], whose ground state is known to have QLRO of the pairing correlations[@Bogoliubov]. Hence the ferromagnetic ground state for $J>U'$ has QLRO of triplet superconductivity as well. For strong attraction ($J-U' \gg t$), all electrons are coupled into triplet pairs if $N_{\rm e}$ is even and the system is described by an effective Hamiltonian ${\cal H}_{\rm b}$ in terms of hard-core bosons with spin unity: $$\begin{aligned} {\cal H}_{\rm b} &=& -\sum_{i=1}^{N-1} \nyo{t} P \left [\sum_{s=0,\pm1}({b}_{is}^{\da}{b}_{i+1 s}+\mbox{H.c.}) +{n}_{i}^{b}+{n}_{i+1}^{b}\right]P \nonumber \\ & & +\sum_{i=1}^{N-1} P\left[ \nyo{J}{\mbox{\bold$S$}}_{i}^{b}\cdot {\mbox{\bold$S$}}_{i+1}^{b} +(2\nyo{t}-\nyo{J}){n}_{i}^{b}{n}_{i+1}^{b}\right]P, \label{eqn:Jlarge}\end{aligned}$$ where $\nyo{t}=2t^{2}/(J-U')$ and $\nyo{J}=2t^{2}/(J+U)$. The summation in ${\cal H}_{\rm b}$ runs over all the nearest neighbor sites. The creation and spin operators of a boson at a site $i$ are defined as $$\begin{aligned} b_{i\,1}^{\da} &=& c_{iA\up}^{\da}c_{iB\up}^{\da}, \nonumber \\ b_{i\,0}^{\da} &=& \frac{1}{\sqrt{2}} (c_{iA\up}^{\da}c_{iB\down}^{\da} +c_{iA\down}^{\da}c_{iB\up}^{\da}), \\ b_{i\,-1}^{\da} &=& c_{iA\down}^{\da}c_{iB\down}^{\da}, \nonumber\end{aligned}$$ and $$\begin{aligned} (S_{i}^{b})^{z} & = & b_{i\,1}^{\da}b_{i\,1}-b_{i\,-1}^{\da}b_{i\,-1}, \\ (S_{i}^{b})^{-} & = & \sqrt{2}(b_{i\,-1}^{\da}b_{i\,0} +b_{i\,0}^{\da}b_{i\,1}), \\ (S_{i}^{b})^{+} & = & \sqrt{2}(b_{i\,1}^{\da}b_{i\,0} +b_{i\,0}^{\da}b_{i\,-1}),\end{aligned}$$ respectively. The number operator $n_{i}^{b}$ is defined as $\sum_{s=0,\pm 1}{b}^{\da}_{is}{b}_{is}$ and $P$ projects out states in which sites are occupied by more than one bosons. Since $\nyo{J}$ is positive, ${\cal H}_{\rm b}$ does not favor ferromagnetism and the ferromagnetic phase should not extend to this region. We examined the phase boundary in the small system with $N=4$ and found that the slope of the phase boundary approaches unity for large $U'/t$. The appearance of the ferromagnetism in the region $J \nearg U'$ is thus interpreted as a smearing out of the ferromagnetic ground state at $J=U'$ to this region. Namely, this ferromagnetism is also caused by the double exchange mechanism at least in the strong coupling region. The above arguments should hold not only for the quarter-filled case but also for the hole-doped and electron-doped cases. Actually the slope of the phase boundary for $J>U'$ approaches unity for large $U'/t$ also in the systems with $(N,N_{\rm e})=(6,4)$ and $(4,6)$. Shen showed coexistence of ferromagnetism and triplet pairing correlations in the whole region where $U=\infty$ and $1 < J/U' < \infty$[@Shen]. On the other hand the present result shows that such a phase appears only in a small region close to the line $J=U'$. This discrepancy comes from the difference in the assumed relations between parameters. ![Correlation functions $\la n_{i}^{b}n_{j}^{b}\ra$ and $\la S_{i}^{b}\cdot S_{j}^{b}\ra$ in the paramagnetic ground state for $J/t=50$ and $U'/t=10$ at quarter-filling. The system size is 36 and $i$ is fixed at $N/2=18$.[]{data-label="fig:xxz"}](ninj "fig:"){width="7cm"}\ ![Correlation functions $\la n_{i}^{b}n_{j}^{b}\ra$ and $\la S_{i}^{b}\cdot S_{j}^{b}\ra$ in the paramagnetic ground state for $J/t=50$ and $U'/t=10$ at quarter-filling. The system size is 36 and $i$ is fixed at $N/2=18$.[]{data-label="fig:xxz"}](sisj "fig:"){width="7cm"} Next, we consider the spin-singlet ground state for $J\gg U'$ where the system is described by ${\cal H}_{\rm b}$. Figure \[fig:xxz\] shows the density correlation function $\la n_{i}^{b}n_{j}^{b}\ra$ of bosons and the spin correlation function $\la \mbox{\bold$S$}_{i}^{b}\cdot\mbox{\bold$S$}_{j}^{b}\ra$ in the ground state of the system with 36 sites obtained by the infinite-system DMRG method. The correlations were measured from the center, i.e. $i$ is fixed to be $N/2$. We observe an even-odd oscillation in the density correlations. This oscillation is due to repulsive interactions between bosons. Actually the second term in ${\cal H}_{\rm b}$ works as repulsions since $||\nyo{J}{\mbox{\bold$S$}}_{i}^{b}\cdot {\mbox{\bold$S$}}_{i+1}^{b} +2\nyo{t}-\nyo{J}||> 0$ holds when ${n}_{i}^{b}{n}_{i+1}^{b}=1$. On the other hand, $\la\mbox{\bold$S$}_{i}^{b}\cdot\mbox{\bold$S$}_{j}^{b}\ra$ shows an oscillation with the period four with slow decay. This oscillation is understood as a superposition of the density correlation with period two and the antiferromagnetic spin correlations caused by the exchange coupling $\nyo{J}$. Hole- and electron-doped systems -------------------------------- Next we study the hole-doped case ($n<1$), where holes are doped into the quarter-filled system. Figure \[fig:mgd-h\] shows the $(U'/t, J/t)$ phase diagram of the ground states for $n=0.5$ and $0.75$ in the system with $N=16$. A ferromagnetic phase appears for a wide parameter region and the magnetization is saturated in the whole phase. Size-dependence among the results for $N=8$, 12 and $16$ was very weak, as in the case of quarter-filling. Comparing with the result for $n=1$ we note that the ferromagnetic phase has expanded in the weak-coupling region down to $U'/t\simeq 3$ as well as to a region where $J \ge U'$. ![\[fig:mgd-h\]Ground-state phase diagram for densities $n=0.5$ and $0.75$ in the $N=16$ system. Interaction parameters satisfy $U=U'+2J$ and $J=J'$.](diagram-h){width="7.5cm"} We note that the lower boundary of the ferromagnetic phase for large $U'/t\,(>J/t)$ almost agrees with that for quarter-filling. This result is a manifestation of the separation of the charge and spin-orbital degrees of freedom in one dimension. We can derive this result by using the same argument as that employed by Ogata and Shiba for the single-band Hubbard model[@Ogata]. In the limit $U'-J=\infty$, the motion of electrons is described by that of spinless fermions on a single chain, and the ground-state wave function is decomposed into a spinless fermion (charge) part and a spin-orbital part. The spin-orbital part is described in terms of spins and pseudo-spins in a [*squeezed*]{} system with $N_{\rm e}$ sites where empty sites are omitted. Interactions among the spins and pseudo-spins are caused by virtual hopping processes and described by the following Hamiltonian (see Appendix \[sec:hole-effec\]) $$\label{eqn:hole-eff0} {\cal H'}_{\rm eff} = n\left( 1-\frac{\sin 2n\pi}{2n\pi}\right) {\cal H}_{\rm eff},$$ where ${\cal H}_{\rm eff}$ is the Hamiltonian given by Eq.(\[eqn:quarter\]) where the site-labels of the spin-orbital quantities are those of the squeezed system. Since ${\cal H'}_{\rm eff}$ differs from ${\cal H}_{\rm eff}$ only by a multiplicative factor, it leads to the $(U'/t, J/t)$ phase diagram identical to that for quarter-filling in the strong coupling limit. This argument explains the present result that the lower boundaries of the ferromagnetic phases for $n=0.5$, $0.75$, and $1$ agree very well for $U'/t\nearg 10$. We may conclude that the origin of the ferromagnetism in this region is the effective exchange coupling due to virtual hoppings. Orbital degrees of freedom are also similar to those at quarter-filling except for the difference of the characteristic wavelength. This behavior can be seen in the orbital correlation function in the ferromagnetic phase obtained by the DMRG method (Fig. \[fig:orbital\]) and the asymptotic form given by Eq. (\[eq:orbital\_correlation\]). Note that the exponent $\alpha_S$ in Eq.(\[eq:orbital\_correlation\]) does not depend on the filling $n$ in the strong coupling limit. The above result greatly depends on the fact that the wave function is decoupled to the charge and spin-orbital parts in the limit $U'-J=\infty$, and that the ground-state energy is independent of the spin-orbital degrees of freedom in the first order of $t$. It should be noted that this is not the case in either higher dimensions or one-dimensional models with far-neighbor hoppings. In these cases the motion of electrons leads to effective spin-orbital interactions in the first order of $t$, as in the case of Nagaoka’s ferromagnetism. In fact, ferromagnetism was not found for $n<1$ in infinite dimensions[@MomoiK; @difference]. We will discuss the effects of far-neighbor hoppings in Section \[sec:sub-result\]. It is remarkable that the ferromagnetism appears for rather weak Hund-rule coupling, that is, $J/t\simeq 3$ for $U'/t\simeq 5$. Hirsch argued based on a diagonalization study of finite-size systems that the Hund-rule coupling is not effective enough to realize ferromagnetism in systems with low density $(n<1)$ and that ferromagnetic exchange interaction between different sites are necessary to explain ferromagnetism in systems such as Ni[@Hirsch]. The present result, however, shows that a moderate Hund-rule coupling is sufficient to realize the ferromagnetism in the one-dimensional model with nearest-neighbor hoppings. Finally we examined the electron-doped case where $n>1$. The phase diagram for $n=1.25$ in the systems with $N=8$ and 16 is shown in Fig. \[fig:mgd-e\]. Size-dependence was very weak, as in the case of quarter-filling. In the systems with $N_{\rm e}=4m+2$, where $m$ is an integer, ground states in the weak-coupling region show small magnetization $S_{\rm tot}=1$ instead of $S_{\rm tot}=0$. This weak spin polarization occurs when one-electron states at the Fermi level are half-filled, i.e. $N_{\rm e}=4m+2$, under the open-boundary condition. The Hund coupling aligns the two electrons at the Fermi level parallel and produces the $S_{\rm tot}=1$ ground state. We regard the appearance of $S_{\rm tot}=1$ weak polarization as a finite-size effect and not as an evidence of unsaturated ferromagnetism in bulk systems. In the ferromagnetic phase, the magnetization is always saturated at this electron density. The ferromagnetism appears in a larger region than in the quarter-filled case in particular for small $J/t$ and large $U'/t$. This enhancement of the stable ferromagnetic state is a result of the double exchange mechanism, which works effectively for $1<n<2$. The appearance of ferromagnetism for $U>U'>J\gg t$ is expected from a rigorous proof which holds in the limit where $U>U'>J \rightarrow \infty$[@Kubo]. Our numerical results agree with this argument and assure that the ferromagnetism extends to a finite parameter region. ![\[fig:mgd-e\]Ground-state phase diagram at density $n=1.25$ for the system with 8 and 16 sites. Interaction parameters satisfy $U=U'+2J$ and $J=J'$. In the case of $N=8$, paramagnetic states show small magnetization $S_{\rm tot}=1$ because of a finite-size effect under open boundary condition. (See text.)](diagram-e2){width="7.5cm"} We show the electron density $n$ as a function of the chemical potential $\mu$ for $U'/t=15$ and $J/t=10$ in Fig. \[fig:m-in\] for the system with 16 sites. For $n>1$ and $n<1$, the ferromagnetic ground state is metallic, since the compressibility ${\rm d}n/{\rm d}\mu$ is finite. For $n=1$ and $2$, the value of ${\rm d}n/{\rm d}\mu$ is vanishing and hence the ground state is insulating. The system is a fully polarized ferromagnetic insulator for the quarter-filling case. On the other hand, the ground state is paramagnetic at the half-filling ($n=2$) where triplet electron pairs occupy all the lattice sites. Triplet pairs on nearest neighbor sites are interacting antiferromagnetically due to the second-order effect of virtual hoppings. As a result the ground state at $n=2$ should be a Haldane singlet with an excitation gap. There is a jump in the density between $n=1.75$ and $2$ in Fig. \[fig:m-in\], which implies the occurrence of phase separation. The occurrence of phase separation was found also in infinite dimensions[@Kubo2]. We found partially polarized ground states in a small region of density where $1.6875 \le n \le 1.75$. As the density increases and approaches the half-filling, the antiferromagnetic interactions between the triplet pairs become operative and suppress the ferromagnetism due to the double exchange mechanism. This may cause the appearance of the partially polarized states. We show the magnetization of a small system where $(N, N_{\rm e})=(8, 14)$ as a function of $J$ in Fig. \[fig:partial\]. The figure shows that the partially polarized state appears in a finite range of the parameter, i.e., $8 < U'/t < 23$ with $J/t=U'/t-1$. This result may imply that the transition from a paramagnetic to ferromagnetic ground state is of the second order for $n\nearl 2$ in contrast to the discontinuous change to a fully polarized state observed in other lower density region. ![Total spin of the ground state as a function of $U'/t$ in the system where $J/t=U'/t-1$ and $(N,N_{\rm e})=(8,14)$.[]{data-label="fig:partial"}](n8ne14judm1-2){width="7.5cm"} Effect of the band-dispersion {#sec:sub-result} ============================= It is widely known that the profile of the density of states (DOS) of the band plays an important role in the realization of itinerant ferromagnetism. In the Hartree-Fock theory of the Hubbard model, for example, the ferromagnetism is stabilized if $UD_{\rm F}>1$ where $D_{\rm F}$ denotes the DOS at the Fermi level. In fact, the whole profile of the DOS affects the stability of the ferromagnetism[@Kanamori; @Sakurai]. The DOS of the nearest-neighbor model in one dimension, which we have studied above, has a minimum at the center of the band and diverges at the band-edges. These features are quite different from those of DOS in higher dimensions. In this section, we examine the effect of the band-dispersion (or DOS) on ferromagnetism, studying a one-dimensional model with far-neighbor hoppings. The model we employ has a linear dispersion of the band as $$\epsilon_{k} = \left\{ \begin{array}{ll} -2t+(4tk/\pi) & (0 < k < \pi) \\ -2t-(4tk/\pi) & (-\pi < k < 0) \end{array} \right. \label{eqn:line}$$ instead of $-2t\cos k$ of the nearest-neighbor model, and hence its DOS is independent of the band energy (see Fig. \[fig:dos\]). For this model, the hopping integrals are calculated from the dispersion (\[eqn:line\]) by $$t_{ij} = \frac{2}{N+1}\sum_k \epsilon_k \sin(ki)\sin(kj),$$ where the Fourier transform of $c_{im\sig}^{\da}$ is given by $$c_{k m \sig}^{\da} = \sqrt{\frac{2}{N+1}}\sum_{i=1}^{N}\sin(ki) c_{i m \sig}^{\da} \label{eqn:fourier}$$ and $k = \ell\pi/(N+1)$ with $\ell=1,2,..,N$. Due to the electron-hole symmetry of the dispersion (\[eqn:line\]) the system is bipartite, i.e. the hopping integrals $t_{ij}$ vanish for $|i-j|=2\ell$ with integer $\ell$, and are always negative for $|i-j|=2\ell + 1$ for positive $t$. The hopping integral $t_{ij}$ is given by $-t[1-(-1)^{i-j}]/\pi(i-j)^{2}$ in the limit $N\rightarrow\infty$. We call this model [*linear-band*]{} model in the following. We study small clusters with $N=4$, 5 and 6 under the open boundary conditions, exactly diagonalizing them, and compare the results with those of the nearest-neighbor model. ![Density of states per site and spin in the nearest-neighbor model (dotted line) and in the linear-band model (solid line).[]{data-label="fig:dos"}](dos){width="7.5cm"} We show the ground state phase diagram for the quarter-filling obtained from the systems with $(N,N_{\rm e})=(4,4)$ in Fig.\[fig:lmgd-q\]. The phase boundaries for two models are in good agreement in the strong coupling region where $U'>J$, while the ferromagnetism is slightly suppressed in the linear-band model in the weak-coupling region where $J\ge U'$. The ferromagnetism is assisted by antiferromagnetic orbital QLRO when $U'>J$. The long-range hoppings in the linear-band model do not destroy the orbital correlations since $t_{ij}$ satisfies the bipartite condition. The present result suggests that the phase boundary between the ferromagnetic and paramagnetic phases does not depend strongly on the spatial dependence of the effective exchange interactions. ![Ground-state phase diagram at quarter-filling obtained by the exact diagonalization of the system with $N=N_{\rm e}=4$. Filled circles and filled triangles display the phase boundary for the linear-band model and the nearest-neighbor model, respectively.[]{data-label="fig:lmgd-q"}](band-q){width="7.5cm"} Next we show the phase diagram for $(N,N_{\rm e})=(6,4)$ in Fig.\[fig:lmgd-h\] as an example of the hole-doped case. It is striking that the ferromagnetic region for the linear-band model is strongly suppressed compared to that for the nearest-neighbor model. The ferromagnetic phase appears only for $U'/t \ge 25$ and exists in a very narrow region along the line $J/U'=1$. This strong reduction of the ferromagnetic phase for $U'>J$ is interpreted as a result of the breakdown of the separation of the charge and spin-orbital degrees of freedom. An electron can pass the other electrons by far-neighbor hoppings, and they lead to effective interactions among the spin and orbital degrees of freedom in the first-order of $t_{ij}$. The effective Hamiltonian in the limit $U'>J\rightarrow \infty$, which contains only hopping terms and prohibits double occupancy, leads to a unique paramagnetic ground state ($S=0$) in the linear-band model for $(N,N_{\rm e})=(6,4)$. Consequently, the ferromagnetic phase does not appear for $U'>J \gg t$ in this model and occurs only in the region $J\simeq U'$ and $U'/t\nearg 25$. ![Ground-state phase diagram for the system with $(N,N_{\rm e})=(6,4)$. Filled circles and filled triangles display the phase boundary for the linear-band model and the nearest-neighbor model, respectively.[]{data-label="fig:lmgd-h"}](band-h){width="7.5cm"} Figure \[fig:lmgd-e\] displays the phase diagram for $(N,N_{\rm e})=(5,6)$ as an example of the electron-doped case. In this case, the phase boundary for the linear-band model is similar to that for the nearest-neighbor model. The ferromagnetism for $U' > J \gg t$ is caused by the double exchange mechanism. Far-neighbor hoppings are expected to favor ferromagnetism as well as nearest-neighbor ones. We note that rigorous proofs[@kusakabe2; @Kubo; @Shen] for the ferromagnetism in the strong coupling limit do not hold for the linear-band model. The present result implies that the ferromagnetic state for $n >1$ is robust against the variation of the DOS and would survive in higher dimensions. Actually, ferromagnetic state was found to be stable for a wide range of interaction parameters for $n>1$ in infinite dimensions[@MomoiK]. ![Ground-state phase diagram for $(N,N_{\rm e})=(5,6)$. Filled circles and filled triangles display the phase boundary for the linear-band model and the nearest-neighbor model, respectively.[]{data-label="fig:lmgd-e"}](band-e){width="7.5cm"} Though we have studied only small systems, the above results clearly show that the band-dispersion has a strong effect on the stability of ferromagnetism for the hole-doped system. On the other hand, it gives only a minor effect on the ferromagnetism for the quarter-filled and the electron-doped systems. Summary and Discussion {#sec:summary} ====================== We studied the ferromagnetism in the one-dimensional Hubbard model with orbital degeneracy by numerical methods. We first studied the model with only nearest-neighbor hoppings by the DMRG calculations of clusters with up to 16 sites and obtained the following results: 1. In the case of the quarter-filling $(n=1)$, a fully polarized ferromagnetic phase appears in a region $U'/t \ge 4$ and $J/t \ge 4$. The ferromagnetic state is insulating for $U'>J$ and metallic for $J\ge U'$. The ferromagnetism for $U'-J\gg t$ is caused mainly by the second-order hopping processes. The phase boundary between the ferromagnetic and paramagnetic phases approaches the asymptote $J= 0.35U'$ for large $U'\,(>J)$. The ferromagnetic phase extends considerably to a region with $J>U'$ for intermediate coupling regime, but the phase boundary on this side becomes parallel to the line $J=U'$ for large $U'/t$. 2. In the hole-doped cases ($n=0.5$ and $0.75$), metallic ferromagnetism appears in a similar parameter region to that of insulating ferromagnetism at the quarter-filling. The phase boundary between the ferromagnetic and paramagnetic phases for large $U'(>J)$ agrees with that for $n=1$, which results from decoupling of the ground-state wave function into the charge and spin-orbital parts in the limit $U'-J=\infty$. The ferromagnetic phase expands to the weak-coupling region more than the quarter-filled case. 3. In the electron-doped case $(n=1.25)$, the ferromagnetic phase dominates the region with $U'>J$. This ferromagnetism is metallic and caused mainly by the double exchange mechanism. Near the half-filled case ($1.6875 \le n \le 1.75$), this ferromagnetic order becomes partially saturated. 4. In a large-Hund-coupling region, $J > U'$, QLRO of triplet superconductivity coexists with the metallic ferromagnetism, which appears for all electron densities, i.e., $0<n<1.75$. It is important to note that the present results for the phase diagram showed little system-size dependence. We hence claim that the present results represent the properties in the thermodynamic limit. The critical values of the interaction parameters for the appearance of ferromagnetism are shown in Fig. \[fig:n-dep\]. In the case of $U'>J$, the critical Hund coupling $J_{\rm c}/t$ is almost constant for $n\le 1$, while it suddenly drops for $n=1.25$. Though we have not examined for other $n$ values for $n>1$, the data suggest a clear discrepancy between $n>1$ and $n\le 1$. For $U'=J$ and $U'<J$, the critical inter-orbital repulsion $U'_{\rm c}/t$ is apparently a smooth function of the density. For the case $U'=J$, $U'_{\rm c}/t$ seems to vanish for decreasing $n$. It would be interesting to examine whether the system exhibits ferromagnetism with infinitesimal interactions in the low-density limit. ![Critical values of the interaction parameters for the appearance of ferromagnetism as functions of the density: $J_{\rm c}/t$ for $U'/t=20$ (top), $U'_{\rm c}/t$ for $J/t=20$ (middle) and $U'_{\rm c}/t$ $(=J_{\rm c}/t)$ for $U'=J$ (bottom).[]{data-label="fig:n-dep"}](critical-n-dep){width="7.5cm"} The double exchange mechanism causes ferromagnetism by lowering the kinetic energy, and hence the strength of the double exchange mechanism may be measured by the difference $\Delta K$ of the kinetic energy in the ferromagnetic state $K_{\rm F}$ from that in the paramagnetic one $K_{\rm P}$, i.e., $\Delta K=K_{\rm F}-K_{\rm P}$. Largely negative $\Delta K$ implies that the double exchange mechanism is effective. We show $\Delta K$ for $U'=J$ in the small systems with $(N,N_{\rm e})=(6,4)$, $(4,4)$ and $(4,6)$ in Fig.\[fig:deltaK-1\]. We see that $\Delta K$ is a decreasing function of $U'/t$ in all the cases and becomes negative for large $U'/t$. The value $U'_{0}/t $ where $\Delta K =0$ decreases with decreasing $n$ which implies that the double exchange mechanism is more effective in the system with lower density for $U'=J$. The values of $U'_{0}/t $ themselves are, however, much larger than the critical value $U'_{\rm c}/t$ for the appearance of the ferromagnetism. For example, $U'_{0}/t \simeq 8$ and $U'_{\rm c}/t \simeq 4$ for $(N,N_{\rm e})=(6,4)$. This result shows that the ferromagnetism in the weak-coupling region is not caused by the double exchange mechanism. The ferromagnetism occurs for a gain of the interaction energy. The situation is similar in the system with $(N,N_{\rm e})=(4,6)$ for $U'>J$ as well. The value $J_{0}/t $ where $\Delta K =0$ is much larger than the critical value $J_{\rm c}/t$. For example, $J_{0}/t \simeq 7$ and $J_{\rm c}/t \simeq 3$ for $U'/t=20$. The system-size dependence of $U'_{0}/t $ (and $J_{0}/t$) is left for a future study; whether it agrees with $U'_{\rm c}/t$ (and $J_{\rm c}/t$) or not in the thermodynamic limit. ![Difference $\Delta K$ of the kinetic energy in ferromagnetic states $K_{\rm F}$ from that in paramagnetic states $K_{\rm P}$, i.e., $\Delta K=K_{\rm F}-K_{\rm P}$. The numbers of system sizes and electrons are set as $(N,N_{\rm e})=(6,4)$, $(4,4)$ and $(4,6)$.[]{data-label="fig:deltaK-1"}](dk-1){width="7.5cm"} We found a partially polarized ferromagnetic state for a density region adjoining the phase-separation region just below half-filling, where the magnetization gradually changes depending on the strength of interactions. In these states, ferromagnetic spin correlation decays especially near edges of open systems. We may need further study in order to confirm that this partial ferromagnetism remains in the thermodynamic limit. There is also another possibility that a partially polarized state might apparently appear because of phase separation into a perfectly polarized part and a paramagnetic part. Recently phase separation is found in other systems and is thought to be a common feature of strong correlations[@Yunoki; @Kubo2]. Next, we examined the effect of the band-dispersion on the ferromagnetism using diagonalization of small clusters for a model with far-neighbor hoppings. The ferromagnetism is greatly destabilized at less than quarter-filling $n<1$. In this case, the second-order effects of hoppings, which are dominating in the nearest-neighbor hopping model, are overwhelmed by the first-order effect. The phase diagram for the quarter-filling does not change greatly by far-neighbor hoppings that satisfy the bipartite condition. This result will be modified if the frustration is introduced by non-bipartite hopping integrals. The ferromagnetism for the electron-doped case is quite robust against variation of the DOS shape. We may extract some presumptions on the ferromagnetism due to the orbital degeneracy in two and three dimensions both from the present results and those in infinite dimensions[@MomoiK; @difference]. At the quarter-filling ($n=1$) the ferromagnetism is supported by the orbital antiferromagnetic correlations for $U'>J$. This mechanism is common in all dimensions and robust if the hopping integrals satisfy the bipartite condition. We hence expect that the insulating ferromagnetic phase occurs in a similar parameter region in all dimensions. The lower phase boundary will approach the line $J=\alpha U'$ $(0.35 >\alpha >0)$ for large $U'/t$, and $\alpha $ will decrease with increasing the dimensionality. For the electron-doped system ($n>1$), it is natural to expect that the double-exchange mechanism works in any dimension. The metallic ferromagnetism for $U'>J$ will appear in a similar parameter region in any dimension for $2>n>1$, since this ferromagnetic order is insensitive to the variation of the band-dispersion. The situation is subtle in the case of the hole-doped system ($n<1$). The double-exchange mechanism hardly works for $n<1$ because of inter-orbital Coulomb repulsion. In this density region, indeed, variation of band-dispersion strongly destabilizes ferromagnetism in one dimension and no ferromagnetic phase was found in infinite dimensions, though only sparse parameter points were examined[@MomoiK; @difference]. In three (and two) dimensions, hence, the ferromagnetism might not appear in the simple nearest-neighbor hopping model on hyper-cubic lattices for $n<1$. To confirm this argument, we need to take account of the interplay between the electron correlations and the band-dispersion and need to accomplish serious calculations for two- and three-dimensional systems. For $J>U'$, QLRO of triplet superconductivity appears, which is supported by ferromagnetic order. In this region, a ferromagnetic state is stabilized by the double exchange mechanism. However, this ferromagnetic state is affected by far-neighbor hoppings and, in infinite dimensions, the boundary of the ferromagnetic phase does not extend beyond the $J=U'$ line up to the coupling $U'=10$[@MomoiK], though we (TM and KK) did not study the $J>U'$ region. The coexistence of ferromagnetism and triplet superconductivity for $J\ge U'$ might not appear in higher dimensions. We expect that a two-dimensional system may be a plausible candidate for showing ferromagnetic triplet superconductivity. Properties of paramagnetic ground states are of theoretical interest as well, though we did not study them in detail. The model has the $SU(4)$ symmetry for $J=0$ and solved exactly in one dimension[@Sutherland]. Recent numerical works clarified the properties of the quarter-filled system[@Yamashita; @Frischmuth]. It will be an interesting problem to investigate the spin and orbital correlations for general $n$ and $J\ge 0$. From the charge and spin-orbital separation as shown in Appendix \[sec:hole-effec\], we can say that the spin degrees of freedom for $n<1$ are equal to those at $n=1$ in the strong coupling limit. In the other paramagnetic region for $J \gg U'$, electrons are bound to form triplet pairs. We found strong charge correlations for the quarter-filling, but did not find LRO of charge density. It is an interesting problem to study whether a charge ordering occurs and coexists with antiferromagnetic LRO in higher dimensions. Numerical calculations were performed on Facom VPP500 at the ISSP, University of Tokyo, and facilities of the Information Center, University of Tsukuba. In the numerical calculations, we partly used subroutines of TITPACK Ver. 2 coded by H. Nishimori. This work was financially supported by Grant in Aid Nod. 09640453, 10203202 and 11640365 from the Ministry of Education, Science, Sports and Culture of Japan. One of the authors (K.K.) was supported by Center for Science and Engineering Research, Research Institute of Aoyama Gakuin University. The Effective Hamiltonian for $n<1$ in the Strong-coupling Limit {#sec:hole-effec} ================================================================ In this appendix, we derive the effective exchange Hamiltonian for $n<1$ in the limit $U' -J\gg t $ in the nearest-neighbor hopping model. We treat the hopping term in (\[eq:deghub\]) as a perturbation. In the unperturbed ground state, every site is singly occupied or empty. As the first-order effect of the hopping term, electrons hop from singly occupied sites to empty ones. The effective Hamiltonian in the first order of $t$ is written as $$H_{\rm eff}^{(1)} = -t\sum^{N}_{i=1}\sum_{m,\sig} \left(\nc_{im\sig}^{\da}\nc_{i+1m\sig}+\mbox{H.c.}\right), \label{eqn:hole-eff1} \\$$ where $\nc_{jm\sig}\equiv c_{jm\sig}(1-n_{jm\,-\sig}) (1-n_{j\barm\sig})(1-n_{j\barm\,-\sig})$ is the Gutzwiller-projected annihilation operator and $\nn_{j}$ is defined as $\sum_{m,\sig}\nc_{jm\sig}^{\da}\nc_{jm\sig}$. In the unperturbed ground-state subspace $\nn_{j}$ is 0 or 1. Since the electrons cannot exchange their positions under the open boundary conditions, the matrix elements of $H_{\rm eff}^{(1)}$ are independent of the spin-orbital degrees of freedom. Hence $H_{\rm eff}^{(1)}$ is equivalent to a Hamiltonian for free spinless fermions, i.e., $$H_{\rm eff}^{(1)} = -t \sum_{i=1}^{N}\left(a_{i+1}^{\da}a_{i}+{\rm H.c.}\right),$$ where $a_{i}^{\da}$ is the creation operator of a spinless fermion. Then the number operator $\nn_{i}$ is equivalent to $a_{i}^{\da}a_{i}$. We may write down the ground-state wave function $|\Psi_{\rm g}\ra$ as a direct product of the ground state of spinless fermions, $|\Phi_{\rm SF}\ra$, and a spin-orbital wave function $|\Omega\ra$ as[@Ogata] $$|\Psi_{\rm g}\ra = |\Phi_{\rm SF}\ra\bigotimes|\Omega\ra. \label{eqn:decom}$$ Here $|\Phi_{\rm SF}\ra$ is a simple Slater determinant for spinless fermions and describes only the charge degrees of freedom. The ground-state energy is degenerate for any spin-orbital wave function $|\Omega\ra$ up to the first order in terms of $t$. The second-order terms due to virtual hoppings determine $|\Omega\ra$. The effective Hamiltonian in the second order is given by $$H_{\rm eff}^{(2)} = H_{\rm eff}^{(2a)} +H_{\rm eff}^{(2b)} \label{eqn:hole-eff2},$$ where $$H_{\rm eff}^{(2a)} = \sum_{i=1}^{N}\nn_{i}\nn_{i+1}h(i,i+1) \label{eqn:hole-eff2a}$$ and $$\begin{aligned} \lefteqn{H_{\rm eff}^{(2b)} = \sum_{i=1}^{N}\sum_{\sig} \sum_{m}\times} \nonumber \\ & & \left[-\frac{t^{2}}{U'-J}\right. (\nc_{i-1\,m\sig}^{\da}n_{i\,\barm\sig}\nc_{i+1\,m\sig}+\mbox{H.c.}) \nonumber \\ & & +\frac{t^{2}}{U'-J} (\nc_{i-1\,\barm\sig}^{\da}\nc_{i\,m\sig}^{\da}\nc_{i\,\barm\sig}\nc_{i+1\,m\sig} +\mbox{H.c.}) \nonumber \\ & & -\frac{U't^{2}}{(U')^{2}-J^{2}} (\nc_{i-1\,m\sig}^{\da}n_{i\,\barm\,-\sig}\nc_{i+1\,m\sig}+\mbox{H.c.}) \nonumber \\ & & +\frac{U't^{2}}{(U')^{2}-J^{2}} (\nc_{i-1\,\barm\sig}^{\da}\nc_{i\,m\,-\sig}^{\da}\nc_{i\,\barm\sig} \nc_{i+1\,m\,-\sig} +\mbox{H.c.}) \nonumber \\ & & -\frac{Jt^{2}}{(U')^{2}-J^{2}} (\nc_{i-1\,m\sig}^{\da} \nc_{i\,\barm-\sig}^{\da} \nc_{i\,\barm\sig} \nc_{i+1\,m\,-\sig} +\mbox{H.c.}) \nonumber \\ & & +\frac{Jt^{2}}{(U')^{2}-J^{2}} (\nc_{i-1\,\barm\sig}^{\da}\nc_{i\,m\,-\sig}^{\da} \nc_{i\,\barm\,-\sig}\nc_{i+1\,m\,\sig} +\mbox{H.c.}) \nonumber \\ & & -\frac{Ut^{2}}{U^{2}-(J')^{2}} (\nc_{i-1\,m\sig}^{\da}n_{i\,m\,-\sig}\nc_{i+1\,m\sig} +\mbox{H.c.}) \nonumber \\ & & +\frac{Ut^{2}}{U^{2}-(J')^{2}} (\nc_{i-1\,m\sig}^{\da}\nc_{i\,m\,-\sig}^{\da}\nc_{i\,m\sig}\nc_{i+1\,m\,-\sig} +\mbox{H.c.}) \nonumber \\ & & -\frac{J't^{2}}{U^{2}-(J')^{2}} (\nc_{i-1\,m\sig}^{\da}\nc_{i\,m\,-\sig}^{\da} \nc_{i\,\barm\,\sig}\nc_{i+1\,\barm\,-\sig} +\mbox{H.c.}) \nonumber \\ & & +\left.\frac{J't^{2}}{U^{2}-(J')^{2}} (\nc_{i-1\,m\sig}^{\da}\nc_{i\,m\,-\sig}^{\da} \nc_{i\,\barm\,-\sig}\nc_{i+1\,\barm\,\sig} +\mbox{H.c.}) \right]. \label{eqn:hole-eff2b}\end{aligned}$$ The first term $H_{\rm eff}^{(2a)}$ represents the exchange interactions between electrons occupying nearest-neighbor sites, where $h(i,i+1)$ is defined in the effective Hamiltonian (\[eqn:h-func\]) for the quarter-filling. The second term $H_{\rm eff}^{(2b)}$ represents correlated hoppings between next-nearest-neighbor sites. The effective Hamiltonian $H_{\rm eff}^{(2)}$ contains both the charge and spin-orbital degrees of freedom. We may average charge degrees of freedom in $H_{\rm eff}^{(2)}$ out since $|\Phi_{\rm SF}\ra$ is fixed to minimize $H_{\rm eff}^{(1)}$. In order to accomplish the averaging we introduce the [*squeezed*]{} system where the empty sites are omitted from the original system. Hence it is a chain with $N_{\rm e}$ sites with spin and orbital degrees of freedom on each site. We label the sites in the squeezed system by $\eta$. Employing the notation $j_{\eta}$ for the position in the original system of the $\eta$-th electron in the squeezed system, we rewrite $H_{\rm eff}^{(2)}$ as $$H_{\rm eff}^{(2a)} =\sum_{\eta=1}^{N_{\rm e}}\sum_{i=1}^{N} \nn_{i}\nn_{i+1} \delta_{ij_{\eta}}h(\eta,\eta+1)$$ and $$H_{\rm eff}^{(2b)} = \sum_{\eta=1}^{N_{\rm e}}\sum_{i=1}^{N}\sum_{\sig_{0}=\pm}\sum_{m_{0}=\pm} \left(\nn_{i}a_{i-1}^{\da}a_{i+1}+{\rm H.c.}\right) \delta_{ij_{\eta}}\Theta(\eta,\sig_{0},m_{0}), \label{eqn:effec-2b0}$$ where $$\begin{aligned} \Theta(\eta,\sig_{0},m_{0}) &\equiv & \left\{\mbox{\rule{0pt}{0.6cm}}\right. -\frac{t^{2}}{U'-J} \lef(\frac{1}{2}+\sig_{0}S^{z}_{\eta}\ri)\tau^{m_{0}}_{\eta} \lef(\frac{1}{2}+\sig_{0}S^{z}_{\eta+1}\ri)\tau^{-m_{0}}_{\eta+1} \nonumber \\ & & +\frac{t^{2}}{U'-J} \lef(\frac{1}{2}+\sig_{0}S^{z}_{\eta}\ri) \lef(\frac{1}{2}-m_{0}\tau^{z}_{\eta}\ri) \lef(\frac{1}{2}+\sig_{0}S^{z}_{\eta+1}\ri) \lef(\frac{1}{2}+m_{0}\tau^{z}_{\eta+1}\ri) \nonumber \\ & & -\frac{U't^{2}}{(U')^{2}-J^{2}} S_{\eta}^{\sig_{0}}\tau_{\eta}^{m_{0}} S_{\eta+1}^{-\sig_{0}}\tau_{\eta+1}^{-m_{0}} \nonumber \\ & & +\frac{U't^{2}}{(U')^{2}-J^{2}} \lef(\frac{1}{2}+\sig_{0}S^{z}_{\eta}\ri) \lef(\frac{1}{2}-m_{0}\tau^{z}_{\eta}\ri) \lef(\frac{1}{2}-\sig_{0}S^{z}_{\eta+1}\ri) \lef(\frac{1}{2}+m_{0}\tau^{z}_{\eta+1}\ri) \nonumber \\ & & -\frac{Jt^{2}}{(U')^{2}-J^{2}} \lef(\frac{1}{2}+\sig_{0}S^{z}_{\eta}\ri)\tau^{m_{0}}_{\eta} \lef(\frac{1}{2}-\sig_{0}S^{z}_{\eta+1}\ri)\tau^{-m_{0}}_{\eta+1} \nonumber \\ & & +\frac{Jt^{2}}{(U')^{2}-J^{2}} S_{\eta}^{\sig_{0}}\lef(\frac{1}{2}-m_{0}\tau_{\eta}^{z}\ri) S_{\eta+1}^{-\sig_{0}}\lef(\frac{1}{2}+m_{0}\tau_{\eta+1}^{z}\ri) \nonumber \\ & & -\frac{Ut^{2}}{U^{2}-(J')^{2}} S_{\eta}^{\sig_{0}}\lef(\frac{1}{2}+m_{0}\tau_{\eta}^{z}\ri) S_{\eta+1}^{-\sig_{0}}\lef(\frac{1}{2}+m_{0}\tau_{\eta+1}^{z}\ri) \nonumber \\ & & +\frac{Ut^{2}}{U^{2}-(J')^{2}} \lef(\frac{1}{2}+\sig_{0}S^{z}_{\eta}\ri) \lef(\frac{1}{2}+m_{0}\tau^{z}_{\eta}\ri) \lef(\frac{1}{2}-\sig_{0}S^{z}_{\eta+1}\ri) \lef(\frac{1}{2}+m_{0}\tau^{z}_{\eta+1}\ri) \nonumber \\ & & -\frac{J't^{2}}{U^{2}-(J')^{2}} \lef(\frac{1}{2}+\sig_{0}S^{z}_{\eta}\ri)\tau^{m_{0}}_{\eta} \lef(\frac{1}{2}-\sig_{0}S^{z}_{\eta+1}\ri)\tau^{m_{0}}_{\eta+1} \nonumber \\ & & +\frac{J't^{2}}{U^{2}-(J')^{2}} S_{\eta}^{\sig_{0}}\tau_{\eta}^{m_{0}} S_{\eta+1}^{-\sig_{0}}\tau_{\eta+1}^{m_{0}} \left.\mbox{\rule{0pt}{0.6cm}} \right\}. \label{eqn:effec-h2}\end{aligned}$$ The symbols $\sig_{0}=+,\,-$ and $m_{0}=+,\,-$ correspond to the spin $\up,\,\down$ and the orbital $A,\,B$, respectively. The correlated hopping processes in the original system are mapped to the spin-orbital exchange processes in the squeezed system as illustrated in Fig. \[fig:hole-map\]. ![An examples of mapping from labelling of electrons by sites to labelling by counting only electrons from the left ($\eta$-th electron). A second-order term $\nn_{iB\down}\nc_{i+1\,A\up}^{\da}\nc_{i-1\,A\up}$ is mapped to the term $S_{\eta}^{-}\tau_{\eta}^{-}S_{\eta+1}^{+}\tau_{\eta+1}^{+}$.[]{data-label="fig:hole-map"}](map2){width="5cm"} In the limit where $N \rightarrow \infty$ with $n$ kept constant, the expectation values of $\nn_{i}\nn_{i+1}\delta_{ij_{\eta}}$ and $(\nn_{i}a_{i-1}^{\da}a_{i+1}+{\rm H.c.})\delta_{ij_{\eta}}$ in the state $|\Phi_{\rm SF}\ra$ are independent of both $\eta$ and $i$. Then we obtain $$\begin{aligned} \la H_{\rm eff}^{(2a)}\ra_{\rm SF} &=& \frac{N}{N_{\rm e}}\la \nn_{i}\nn_{i+1}\ra_{\rm SF} \sum_{\eta=1}^{N_{\rm e}}h(\eta,\eta+1) \label{eqn:hole-eff2a-ab} \\ \la H_{\rm eff}^{(2b)}\ra_{\rm SF} &=& \frac{2N}{N_{\rm e}}\la \nn_{i}a_{i-1}^{\da}a_{i+1}\ra_{\rm SF} \sum_{\eta=1}^{N_{\rm e}}\sum_{\sig_{0}=\pm} \sum_{m_{0}=\pm}\Theta(\eta,\sig_{0},m_{0}). \label{eqn:effec-2b}\end{aligned}$$ Here the notation $\la\cdot\cdot\cdot\ra_{\rm SF}$ means the expectation value in the state $|\Phi_{\rm SF}\ra$. We note that the matrix elements of $\Theta(\eta,\sig_{0},m_{0})$ are equal to those of $-h(\eta,\eta+1)/2$. We estimate the expectation values as $$\begin{aligned} \la \nn_{i}\nn_{i+1} \ra_{\rm SF} &=& n^{2}-\left(\frac{\sin n\pi}{\pi}\right)^{2}\end{aligned}$$ and $$\begin{aligned} \la \nn_{i}a_{i-1}^{\da}a_{i+1}\ra_{\rm SF} &=& n\frac{\sin 2n\pi}{2\pi}-\left(\frac{\sin n\pi}{\pi}\right)^{2},\end{aligned}$$ by using Wick’s theorem. Finally, we obtain the effective Hamiltonian for the spin-orbital degrees of freedom as $$\begin{aligned} \la H_{\rm eff}\ra_{\rm SF} &=& n\left(1-\frac{\sin 2n\pi}{2n\pi}\right){\cal H}_{\rm eff}. \label{eqn:hole-eff}\end{aligned}$$ Here ${\cal H}_{\rm eff}$ is equal to the effective Hamiltonian (\[eqn:quarter\]) for the quarter-filling, where the summation over lattice sites are taken over those in the squeezed system. The factor $n(1-\sin 2n\pi/2n\pi)$ is positive for all density ($0<n<1$). Hence, the effective spin-orbital exchange interactions in the hole-doped system are same with that for the quarter-filled system except for a single multiplicative factor. The present result suggests that the same fact holds quite generally for one-dimensional systems with any number of internal degrees of freedom. Ferromagnetism due to a single unpaired electron in the limit where $J-U'= \infty$ {#ap:J-large} ================================================================================== In finite-size calculations with odd number of electrons we find that the ferromagnetic state is stable in a large parameter region where $J>U'$ as shown in Fig. \[fig:Ne=odd\]. This ferromagnetism is caused by the motion of a single unpaired electron among bound triplet pairs. We consider a system with odd $N_{\rm e}$ system in the strong coupling limit where $J\rightarrow \infty$ with $J/U' (>1)$ kept constant under the open boundary conditions. We prove, in the following, that the ground state of this system is fully polarized. Only nearest-neighbor hoppings are assumed to exist. ![Ground-state phase diagram for $(N,N_{\rm e})=(4,3)$ and $(12,9)$ as examples of the odd $N_{\rm e}$ cases.[]{data-label="fig:Ne=odd"}](Ne_odd){width="7.5cm"} In this limit two electrons on a site is bound to form a triplet pair. As a result $M=(N_{\rm e}-1)/2$ sites are doubly occupied and a single unpaired electron exits. In this limit, only allowed processes in the Hamiltonian are hoppings of an electron from a doubly occupied site to a singly occupied one and those from a singly occupied to a vacant one. The effective Hamiltonian $H_{\rm eff}$ is written as $$\begin{aligned} H_{\rm eff} = -t\sum^{N}_{i=1} \{P^{D}_{i}c_{im\sig}^{\da}P_{i}^{S}P^{S}_{i+1}c_{i+1m\sig}P_{i+1}^{D} \nonumber\\ +P^{S}_{i}c_{im\sig}^{\da}P_{i}^{V}P^{V}_{i+1}c_{i+1m\sig}P_{i+1}^{S} +\mbox{H.c.}\},\end{aligned}$$ where $P_{i}^{D}$, $P_{i}^{S}$ and $P_{i}^{V}$ are the projection operators for the doubly occupied triplet, the singly-occupied and vacant states, respectively. We can divide the whole phase space into subspaces with fixed $N_{A}$, $N_{B}$ and $S^{z}$. Here $N_{A}$ and $N_{B}$ denote the number of electrons in the orbital $A$ and $B$, respectively ($|N_{A}-N_{B}|=1$), which are conserved. Each subspace is further divided into $_{N-1}C_{M}$ sectors according to the configuration of doubly-occupied and vacant sites. States in each sector have the same configuration of doubly-occupied and empty sites if we neglect the singly-occupied site. Different sectors are disconnected with respect to $H_{\rm eff}$. We call the sectors as $\Gamma_{j}(N_{A},N_{B},S^{z})$, where $j=1,2,..,_{N-1}C_{M}$. We can show for any two configurations $\alpha$ and $\beta$ of spins and electrons belonging to a same sector $\Gamma_{j}(N_{A},N_{B},S^{z})$ that: 1. there exists $n\ge 1$ such that $\la\Phi_{\alpha}|H_{\rm eff}^{n}|\Phi_{\beta}\ra\neq0 $ 2. $\la\Phi_{\alpha}|H_{\rm eff}|\Phi_{\beta}\ra\leq0. $ Here $|\Phi_{\alpha}\ra$ denotes a basis vector chosen as $$|\Phi_{\alpha}\ra = \left[\prod_{\ell=1}^{M}(-1)^{j_{\ell}}\right] A_{\alpha_{1}}A_{\alpha_{2}}\cdots A_{\alpha_{N}}|0\ra, \label{eqn:baseUd=J}$$ where $\alpha_{i}$ denotes the atomic state of the $i$-th site and $A_{\alpha_{i}}$ is the corresponding operator, i.e. $c_{iA\uparrow}^{\da}c_{iB\uparrow}^{\da}$, $c_{iA\downarrow}^{\da}c_{iB\downarrow}^{\da}$, $1/\sqrt{2}(c_{iA\uparrow}^{\da}c_{iB\downarrow}^{\da} + c_{iA\downarrow}^{\da}c_{iB\uparrow}^{\da})$ for doubly-occupied triplet states, $c_{im\sig}^{\da}$ for singly-occupied states, and unity for a vacant state. The number $j_{\ell}$ denotes the position of the $\ell$-th doubly-occupied site counted from an end. Here $|0\ra$ is vacuum state. The proof of (I) and (II) are straightforward[@Kubo]. Now that the two conditions (I) and (II) on the matrix elements of $H_{\rm eff}$ are satisfied, the Perron-Frobenius theorem ensures that the ground state in the sector $\Gamma_{j}(N_{A},N_{B},S^{z})$ is unique and can be written as a linear combination of all the basis vectors with strictly positive (or negative) coefficients. The positive (or negative) definiteness of the coefficients leads to that the ground state is an eigenstate of the maximum value of the total spin[@Kubo]. We thus proved that the ground state in each sector is fully polarized. As a result the ground state for subspace with fixed $N_{A}$ and $N_{B}$ is fully polarized. We understand that the large ferromagnetic region in the phase diagram in the case of odd $N_{\rm e}$ for $J- U'\gg t$ is caused by the motion of a single unpaired electron and is a finite-size effect. [99]{} Present address: Development & Techinology Division, Hitachi Maxell Ltd., 6-20-1 Kinuno-dai, Yawara-mura, Tsukuba-gun 300-2496, Japan. J. Kanamori, Prog. Theor. Phys. [**30**]{} (1963) 275. M. C. Gutzwiller, Phys. Rev. Lett. [**10**]{} (1963) 159. J. Hubbard, Proc. Roy. Soc. London [**A276**]{} (1993) 238. E. H. Lieb, Phys. Rev. Lett. [**62**]{} (1989) 1201; Errata Phys. Rev. Lett. [**62**]{} (1989) 1927. A. Mielke, J. Phys.[**A24**]{} (1991) L73; ibid. [ **24**]{} (1991) 3311; ibid [**25**]{} (1992) 4335; Phys. Lett. [**A174**]{} (1993) 443; A. Mielke and H. Tasaki, Commum. Math. Phys. [**158**]{} (1993) 341. H. Tasaki, Phys. Rev. Lett. [**69**]{} (1992) 1608. H. Tasaki, Phys. Rev. Lett. [**75**]{} (1995) 4678. E. Müller-Hartmann, J. Low. Temp. Phys. (1995) 349. K. Penc, H. Shiba, F. Mila and T. Tsukagoshi, Phys. Rev. B [**54**]{} (1996) 4056. H. Sakamoto and K. Kubo, J. Phys. Soc. Jpn. [**65**]{} (1996) 4056. P. Pieri, S. Daul, D. Baeriswyl, M. Dzierzawa and P. Fazekas, Phys. Rev. B [**54**]{} (1996) 9250. S. Daul and R. M. Noack, Z. Physik B [**103**]{} (1997) 293. J. C. Slater, Phys. Rev. [**49**]{} (1936) 537; ibid. [**49**]{} (1936) 931. C. Zener, Phys. Rev. [**81**]{} (1951) 440, [*ibid.*]{} [**82**]{} (1951) 403. J. H. Van Vleck, Rev. Mod. Phys. [**25**]{} (1953) 220. P.W. Anderson and H. Hasegawa, Phys. Rev. (1955) 675. L. M. Roth, Phys. Rev. [**149**]{} (1966) 306. K. I. Kugel and D. I. Khomskii, Sov. Phys. -JETP[**37**]{} (1974) 725. M. Cyrot and C. Lyon-Caen, J. Phys. C[**36**]{} (1975) 253. S. Inagaki, J. Phys. Soc. Jpn. [**39**]{} (1975) 596. S. Inagaki and R. Kubo, Int. J. Mag. [**4**]{} (1973) 139. K. Kusakabe and H. Aoki, Mol. Cryst. Liq. Cryst. [**233**]{} (1993) 71. K. Kusakabe and H. Aoki, Physica B [**194-196**]{} (1994) 217. W. Gill and D. J. Scalapino, Phys. Rev. B [**35**]{} (1987) 215. J. Kuei and R. T. Scaletter, Phys. Rev. [**55**]{} (1997) 14968. J. E. Hirsch, Rhys. Rev. B [**56**]{} (1997) 56. K. Kubo, J. Phys. Soc. Jpn. [**51**]{} (1982) 782. S. Q. Shen, Phys. Rev. B [**57**]{} (1998) 6474. T. Momoi and K. Kubo, Phys. Rev. B [**58**]{} (1998) R567. K. Held and D. Vollhardt, Eur. Phys. J. B [**5**]{} (1998) 473. In the dynamical mean-field theory, Held and Vollhardt (HV) found ferromagnetism for the hole-doped region ($n<1$)[@Held] at $U=8$, $U'=6$ and $J=2$, but two of us \[Momoi and Kubo (MK)\] did not discover it[@MomoiK] in the region up to $U\le 40$ and $U'\le 20$. One possibility of this discrepancy may be that ferromagnetic phase appears in a very narrow region for the hole-doped case and MK could not discover it. As another possibility, note that the systems studied are different in these two cases. In ref. , HV adopted a semi-elliptic DOS. On the other hand, in ref., MK used a Gaussian DOS, which is the exact one for the hyper-cubic lattice in infinite dimensions. As we showed in the present paper, the shape of DOS can greatly affect the appearance of ferromagnetism for $n<1$. Since the long tail of the Gaussian DOS is disadvantageous to ferromagnetism at low densities, the discrepancy of the phase diagrams presumably comes from the choice of DOS. Also, HV took an Ising anisotropic limit of the Hund coupling and neglected pair-hopping terms, but MK treated all terms correctly. J. Bünemann, W. Weber, and F. Gebhard, Phys. Rev. B [**57**]{}, 6896 (1998). T. Momoi, J. Stat. Phys. [**85**]{}, 193 (1996). S. R. White, Phys. Rev. Lett. [**69**]{} (1992) 2863; Phys. Rev. B [**48**]{} (1993) 10345. K. Kubo, D. M. Edwards, A. C. M. Green, T. Momoi and H. Sakamoto, in [*Physics of Manganites*]{} eds. by  T.A. Kaplan and S.D. Mahanti, (Kluwer Academic/Plenum Publ., New York, 1999) p.71. H. J. Schulz, Phys. Rev. Lett. [**64**]{}, 2831 (1990). N. Kawakami and S.-K. Yang, Phys. Lett. A [ **148**]{}, 359 (1990). R. Assaraf, P. Azaria, M. Caffarel, and P. Lecheminant, Phys. Rev. B [**60**]{}, 2299 (1999). B. Sutherland, Phys. Rev. B [**12**]{} (1975) 3795. N. M. Bogoliubov and V. E. Korepin, Inter. J. Mod. Phys. B [**3**]{} (1989) 427. M. Ogata and H. Shiba, Phys. Rev. B [**41**]{} (1990) 2326, H. Shiba and M. Ogata, Int. J. Mod. Phys. [ **B5**]{} (1991) 31. A. Sakurai, Prog. Theor. Phys. [**39**]{} (1968) 312. S. Yunoki, J. Hu, A.L. Malvezzi, A. Moreo, N. Furukawa and E. Dagotto, Phys. Rev. Lett. [**80**]{} (1998) 845. Y. Yamashita, N. Shibata and K. Ueda, Phys. Rev. B [**58**]{} (1998) 9114. B. Frischmuth, F. Mila and M. Troyer, Phys. Rev. Lett. [**82**]{} (1999) 835.

ArXiv

--- author: - 'J. Khoo , G. Hobbs , R. N. Manchester , D. Miller , J. Dempsey' date: 'Received  2012 month day; accepted  2012  month day' title: Using the Parkes Pulsar Data Archive --- cite\#1\#2[^\[[\#1@tempswa\ \#2]{}\]^]{} [GBK]{}[song]{} Introduction {#sec:intro} ============ The Parkes Pulsar Data Archive provides public access to 165,755 data files related to observations of pulsars recorded at the Parkes observatory. It is part of the Data Access Portal—an on-line interface (http://data.csiro.au) that provides access to published Commonweath Scientific and Industrial Research Organisation (CSIRO) data sets over a range of science disciplines. The Parkes Pulsar Data Archive is the largest publicly accessible database of pulsar data files. Approximately $ 10^{5} $ data files contain observations from surveys of the sky and the remainder are for observations of 775 known pulsars and their corresponding calibration signals. Improvements to the system have been made since the initial description of the archive in Hobbs et al. (2011)[@Hobbs; @et; @al.(2011)]. This paper focuses on how to use the current Parkes Pulsar Data Archive and provides walk-throughs of common use cases. § \[sec:contents\] describes the types of data that can be obtained from the data archive. § \[sec:nav-pulsar-search\] explains how to navigate to the Pulsar Search from the main Data Access Portal page. § \[sec:examples\] shows examples of how to obtain data used in complemetary data-processing tutorials [@k12; @h12] from the on-line interface. § \[sec:vo\] provides examples of how to use various Virtual Observatory tools to access contents of the data archive. ![CSIRO’s Data Access Portal (http://data.csiro.au)[]{data-label="fig:dap"}](dap.png){width="150mm"} ![Data Access Portal: Specific Search[]{data-label="fig:spec-search"}](spec-search.png){width="150mm"} ![Pulsar Search: main window[]{data-label="fig:pulsar-search-main"}](pulsar_search_main.png){width="130mm"} ![Search criteria for Hobbs et al. 2012[]{data-label="fig:hobbs-main"}](hobbs_main_search.png){width="110mm"} ![Facet search options for Hobbs et al. 2012[]{data-label="fig:hobbs-facet"}](hobbs_facet.png){width="150mm"} ![Thumbnail preview of PSR J1539$-$5628 for Hobbs et al. 2012[]{data-label="fig:hobbs-preview"}](hobbs_preview.png){width="150mm"} Archive Contents {#sec:contents} ================ There are two main observing modes that are used in pulsar astronomy: 1) “fold mode” that results in a single pulse profile for each observation; such files are used for calculating pulse arrival times and studies utilising pulse profiles, and 2) “search mode” a set of time series recording the signal from the telescope using a selected number of bits, sample interval, and frequency resolution. This mode is used, primarily, for discovering new pulsars and studying single pulses from pulsars. The data formats for both modes are based on the PSRFITS definition [@Hotan; @et; @al.(2004)]. In addition to the raw observations of fold-mode data, the Parkes Pulsar Data Archive also contains the corresponding compressed representation of the data; the compression performed is integration over the polarsation, time, and frequency domains. This gives the user the option to opt for downloading the compressed file—decreasing the download time—rather than the raw data. Pulsar data recorded at Parkes are automatically added to the Parkes Pulsar Data Archive; however, the archive complies to the 18-month embargo policy enforced by CSIRO such that only members of the observing project have access to observations recorded in the last 18 months. An exemption to this rule is data recorded by the observing project PULSE@Parkes [@Hobbs; @et; @al.(2009)]; since the project is for science outreach and the education of high-school students and teachers, the data are immediately available. Navigating to the Pulsar Search {#sec:nav-pulsar-search} =============================== Follow the instructions below to access the main pulsar search page: 1. Navigate a web browser to http://data.csiro.au. 2. From the Data Access Portal home page (Figure \[fig:dap\]), click ‘SPECIFIC SEARCH’ to access the various data access points. 3. From the ‘SPECIFIC SEARCH TOOLS’ page (Figure \[fig:spec-search\]), click ‘ATNF PULSAR OBSERVATION SEARCH’ to access the Pulsar Search window (Figure \[fig:pulsar-search-main\]). Examples {#sec:examples} ======== This section will provide two step-by-step procedures to obtain data. Hobbs et al. (2012): PSR J1539$-$5626 Preprocessed Fold-mode Data {#sec:example} ----------------------------------------------------------------- The data used in Hobbs et al. (these proceedings) are preprocessed (with the data integrated in time, frequency, and polarisation) observations of PSR J1539$-$5626. These observations are part of the P262 project: “Timing of young pulsars”. Data files can be obtained via the following instructions: 1. Proceed to the Pulsar Search webpage using a web browser. 2. Enter the following search criteria in the main pulsar search window: J1539$-$5626 for source name, P262 for project ID (as shown in Figure \[fig:hobbs-main\]), and then click “Search”. 3. New tables will appear on the screen showing a summary of query results. On the left-hand table, untick “raw” so that only preprocessed files are displayed (shown in Figure \[fig:hobbs-facet\]). This will refine the search and remove all raw data files (leaving only preprocessed observations of PSR J1539$-$5626 from project P262). 4. Press the icon under the “Preview” column to see a preview of the pulse profile. Select the desired files to download by ticking the toggle box on the left-hand side (shown in Figure \[fig:hobbs-preview\]). Click “Download” to copy the selected files onto your computer. Once fold-mode data have been obtained, PSRCHIVE [@Hotan; @et; @al.(2004)] can be used to calculate pulsar period variations. Once pulse arrival times have been calculated, TEMPO2 [@Hobbs; @et; @al.(2006)] can be used to study the pulse arrival times. ![Search critera for Keith et al. (2012)[]{data-label="fig:keith-main"}](keith_main_search.png){width="130mm"} ![Query results for Keith et al. (2012)[]{data-label="fig:keith-results"}](keith_results.png){width="150mm"} Keith et al. (2012): PSR J0835$-$4510 Search-mode Data {#sec:example} ------------------------------------------------------ The data file used in Keith et al. (2012) is a search-mode observation near the Vela Pulsar (PSR J0835$-$4510). These observations were recorded in the P050 project: “Parkes Southern Pulsar Survey”. It can be obtained via the following instructions 1. Proceed to the Pulsar Search webpage using a web browser (see § \[sec:nav-pulsar-search\]). 2. Enter the following search criteria in the main pulsar search window: G123070 for source name, P050 for project ID (as shown in Figure \[fig:keith-main\]), and then click “Search”. 3. Select the first result by ticking the checkbox on the left-hand side of the table (as shown in Figure \[fig:keith-results\]). Scroll down to the bottom of the page and click “Download” to copy the files onto your computer. Once search-mode data have been obtained, SIGPROC [@Lorimer(2011)] or PRESTO [@Ransom(2011)] can be used to process the data and search for pulsars. Virtual Observatory {#sec:vo} =================== Introduction {#sec:vo-intro} ------------ The Virtual Observatory (VO) is a set of astronomy-related data archives and software tools. VO software tools have the capability to perform complex operations on any combination of the accessible VO data archives. The contents of a VO data archive adhere to a defined set of protocols; this allows development of VO tools to follow the VO definition and, as a result, perform operations (such as statistical analysis, plotting, etc.) on contents of VO databases. This section provides the steps required to access the Parkes Pulsar Data Archive using two VO tools from the AstroGrid software package (http://www.astrogrid.org): VODestop and TopCat. VODesktop is used as the medium to communicate with the Parkes Pulsar Data Archive. You will be shown how to perform a example query to obtain a VOTable (a VO-formatted file representing the results of the query). We will then use TopCat to analyse and plot the results. VODesktop: Performing a query {#sec:vo-table} ----------------------------- ![Endpoint registry prefences for VODesktop.[]{data-label="fig:vodesktop-pref"}](vodesktop-pref.png){width="155mm"} ![Positions of the data in the Parkes Pulsar Data Archive.[]{data-label="fig:topcat-sphere"}](topcat-sphere.png){width="120mm"} ![Distribution of number of observations versus time.[]{data-label="fig:topcat-histogram"}](topcat-histogram.png){width="110mm"} 1. Go to http://www.astrogrid.org/wiki/Install/Downloads and download VODesktop. 2. Open VODesktop and go to Preferences -$>$ Registry. 3. Change the both registry endpoints to http://datanet.csiro.au/astrogrid-registry/services/RegistryQueryv1\_0 (as shown in Figure \[fig:vodesktop-pref\]). 4. Click Apply and restart VODesktop. 5. Select Recent Changes and click on ANDS Parkes Data Archive. 6. Under Actions, click ADQL Query. 7. In the Task Runner window, change the query text to: SELECT \* FROM observation 8. Choose to save the results, locally, as a VOTable. TopCat: Analysing the results {#sec:vo-topcat} ----------------------------- 1. Go to http://www.astrogrid.org/wiki/Install/Downloads and download TopCat. 2. To load the saved VOTable, File -$>$ Load Table -$>$ Filestore Browser and navigate to where you saved the VOTable. Once the VOTable has been read by TopCat, TopCat’s plotting functionality become available. Figure \[fig:topcat-sphere\] shows the pulsar positions on the 3D galactic coordinate sphere using the Spherical Plot. Figure \[fig:topcat-histogram\] shows a histogram of all pulsar observations and observation date. Future Development {#sec:future} ================== The Parkes Pulsar Data Archive will undergo expansion to include timing data from ATNF Parkes Swinburne Recorder and CASPER Parkes Swinburne Recorder; and (potentially) the Berkeley Parkes Swinburne Recorder recorded at the Parkes 64-m telescope. Further conversions between data formats will also be performed to include historical data from earlier projects. A related database of pulsar profiles is being developed at CSIRO Astronomy and Space Science (CASS). This tool will allow users to search pulse profiles of known pulsars and see visual representations when comparing parameters between pulsars. CASS is also developing a data access portal that provides access to various non-pulsar radio-astronomy files of interest to external collaborators of CASS staff members. These systems will use the Parkes Pulsar Data Archive as a guide and will, also, be compatible with the Virtual Observatory formats and protocols. Conclusion {#sec:conclusion} ========== The Parkes Pulsar Data Archive currently hosts 165,755 data files—search mode and fold mode. These data are publicly accessible. The data archive plays an important role in providing researchers worldwide with observations recorded at Parkes Observatory. The Parkes Pulsar Data Archive is an effective solution for our current data sets (pulsar observations from Parkes Observatory since 1991). With ASKAP and SKA being built, a different set of challenges stand ahead to produce similar software solutions. This project is supported by the Australian National Data Service (ANDS). ANDS is supported by the Australian Government through the National Collaborative Research Infrastructure Strategy Program and the Education Investment Fund (EIF) Super Science Initiative19. We acknowledge the software development provided by the CSIRO IM&T Software Services, the business process development by the CSIRO IM&T Data Management Service and project management through Citadel Systems. This research has made use of software provided by the UK’s AstroGrid Virtual Observatory Project, which is funded by the Science and Technology Facilities Council and through the EU’s Framework 6 programme. The Parkes radio telescope is part of the Australia Telescope, which is funded by the Commonwealth of Australia for operation as a National Facility managed by the Commonwealth Scientific and Industrial Research Organisation (CSIRO). GH is the recipient of an Australian Research Council QEII Fellowship (\#DP0878388). [1]{} Hobbs G. Pulsars as gravitational wave detectors \[C\]. High-Energy Emission from Pulsars and their Systems. Astrophysics and Space Science Proceedings. 2011, 229. Keith M J, Installation and Use of Pulsar Search Software \[J\]. Astronomical Research and Technology, 2012, 9(3): 219-228. Hobbs G, Using the Pulsar Timing Software Package, TEMPO2 \[J\]. Astronomical Research and Technology, 2012, 9(3): 257-276. Hotan A W, van Straten W, Manchester R N. PSRCHIVE and PSRFITS: an open approach to radio pulsar data storage and analysis \[J\]. Publications of the Astronomical Society of Australia, 2004, 21(3): 302-309. Hobbs G, Hollow R, Champion D, et al. The PULSE@Parkes project: a new observing technique for long-term pulsar monitoring \[J\]. Publications of the Astronomical Society of Australia, 2009, 26: 468-475. Hobbs G, Edwards R T, Manchester R N. TEMPO2: a new pulsar-timing package - I. An overview \[J\]. Monthly Notices of the Royal Astronomical Society, 2006, 369(2): 655-672. Lorimer D R. SIGPROC: pulsar signal processing programs \[C\]. Astrophysics Source Code Library, 2011, 7016. Ransom S. PRESTO: PulsaR Exploration and Search TOolkit \[C\]. Astrophysics Source Code Library, 2011, 7017. \[lastpage\]

ArXiv

--- abstract: 'We provide further non-trivial solutions to the recently proposed time-dependent Dyson and quasi-Hermiticity relation. Here we solve them for the generalized version of the non-Hermitian Swanson Hamiltonian with time-dependent coefficients. We construct time-dependent solutions by employing the Lewis-Riesenfeld method of invariants and discuss concrete physical applications of our results.' author: - 'Andreas Fring$^{\bullet}$ and Miled H. Y. Moussa$^{\bullet,\circ}$' title: 'The non-Hermitian Swanson model with a time-dependent metric' --- Introduction ============ PT symmetric (PTS) quantum mechanics has attracted increasing attention since is was demonstrated that PTS Hamiltonians possess real spectra [@BB1998] and allow for a unitary evolution with a redefined inner product [@Geyer; @Mostafa]. Phase transitions between the regimes of unbroken and broken PT symmetry, which are a key feature in the energy spectrum are well understood to occur when two real eigenvalues coalesce to form complex conjugate pair [@BB1998]. Many interesting new results have recently emerged from the application of PTS concepts to different areas of physics, in the classical and the quantum domain, on both fronts, theoretical as well as experimental. We mention here a few, such as the design of an ultralow-threshold phonon laser [@PL], the demonstration of defect states [@OL] and beam dynamics [@BD] in PTS optical lattices, and the fact that the Jarzynski equality generalizes to PTS domain [@J]. Reinforcing the practical features, there are optical structures described by PTS concepts that enable unprecedented control of light [@CL]. At a classical level, PTS properties have also been observed in a variety of experimental set-ups, ranging from quantum optics [@QO] to NMR [@NMR] and superconductivity [@S]. Although the grounds for treating non-Hermitian Hamiltonians using time-independent metric operators have been extensively studied and well established [@CB; @CA], the generalization to time-dependent (TD) metric operators has raised controversy [@M; @Z; @Wang; @MM]. In Ref. [@M], Mostafazadeh has demonstrated that using a TD metric operator one can not ensure the unitarity of the time-evolution simultaneously with the observability of the Hamiltonian. From this perspective, with which we agree, the authors of Refs. [@Z; @Wang; @MM] fail to ensure a unitary time-evolution by insisting on the observability of the Hamiltonian. However, we have recently suggested [@Andreas] that this is not an obstacle and certainly not a no-go theorem. It is very common in the context of PTS quantum mechanics that certain operators, such as position or momentum, may become non-observable auxiliary variables and only their quasi-Hermitian counterparts can be measured. In [@Andreas] we take the view that the Hamiltonian, meaning the operator that satisfies the TD Schrödinger equation (SE), joins this set of observables in the scenario where a TD metric operator is considered. For this proposal to be meaningful the TD quasi-Hermiticity relation and TD dyson relation need to possess non-trivial solutions. When this is the case, we have unitary time evolution and well defined observables. Here we provide new non-trivial solution to this set of equations for generalized time-dependent version of Swanson Hamiltonian [@Swanson] by solving its TDSE and by computing some observables. In order to solve the SE, we shall adapt a method presented in Ref. [@MBS; @SanjibA] for treating TD Hermitian Hamiltonians. This method takes advantage of a unitary TD transformation on the SE, here replaced by a non-unitary transformation to conform with non-Hermitian Hamiltonians, and the diagonalization of a TD Invariant on the Lewis and Riesenfeld framework [@LR]. The authors in Ref. [@MBS] pursued the solution of the SE governed by a general TD quadratic Hamiltonian in order to investigate the mechanism of squeezed states following from the nonlinear amplification terms of the Hamiltonian [@Walls; @Scully]. Here, we shall focus on the technique to treat a TD non-Hermitian Hamiltonian, leaving open the possibility of further analysis of the squeezing mechanism coming from the nonlinear terms of a TD non-Hermitian Hamiltonian. Non-Hermitian Hamiltonian systems with TD metric ================================================ Let us briefly review the scheme proposed in [@Andreas]: We consider a non-Hermitian TD Hamiltonian $H(t)$ whose associated SE, $i\partial _{t}\left\vert \psi(t)\right\rangle =H(t)\left\vert \psi(t)\right\rangle $, is mapped, by means of the Hermitian TD operator $\eta(t)$, into the SE $i\partial_{t}\left\vert \phi(t)\right\rangle =h(t)\left\vert \phi (t)\right\rangle $, where the corresponding wave functions are transformed as $\left\vert \phi(t)\right\rangle =\eta(t)\left\vert \psi(t)\right\rangle $ and the Hamiltonians are related by means of the *TD Dyson relation* $$h(t)=\eta(t)H(t)\eta^{-1}(t)+i\left[ \partial_{t}\eta(t)\right] \eta ^{-1}(t)\text{.} \label{1}$$ We set here $\hbar=1$. The key feature in this equation is the fact that $H(t)$ is no longer quasi-Hermitian, i.e. related to $h(t)$ by means of a similarity transformation, due to the presence of the last term. Thus $H(t)$ is not a self-adjoined operator and therefore not observable. Using the Hermiticity of $h(t)$ we then derived the *TD quasi-Hermiticity relation* $$H^{\dag}(t)\rho(t)-\rho(t)H(t)=i\partial_{t}\rho(t), \qquad\rho(t)=\eta ^{\dagger}\eta, \label{2}$$ replacing the standard quasi-Hermiticity relation for a time-independent $\rho$, given by $H^{\dag}\rho=\rho H$. In fact, the TD quasi-Hermiticity relation ensures the TD probability densities in the Hermitian and non-Hermitian systems to be related in the standard form $$\left\langle \psi(t)\left\vert \tilde{\psi}(t)\right. \right\rangle _{\rho }=\left\langle \psi(t)\left\vert \rho(t)\right\vert \tilde{\psi}(t)\right\rangle =\left\langle \phi(t)\left\vert \tilde{\phi}(t)\right. \right\rangle \text{.} \label{3}$$ With the assumption that $\rho(t)$ is a positive-definite operator, it plays the role of the TD metric and we conclude that any self-adjoined operator $o(t)$, i.e. observable, in the Hermitian system possesses a counterpart $O(t)$ in the non-Hermitian system given by $$O(t)=\eta^{-1}(t)o(t)\eta(t)\text{,} \label{4}$$ in complete analogy to the time-independent scenario. Thus as long as the generalized equations (\[1\]) and (\[2\]) posses non-trivial solutions for $\eta(t)$ and $\rho(t)$, respectively, we have a well defined physical system with TB observables and unitary time-evolution governed by a TD non-Hermitian Hamiltonian. Albeit we have the slightly unusual feature that the TD Hamiltonian $H(t)$ does not belong to the set of observables. The generalized time-dependent Swanson Hamiltonian ================================================== The system we wish to investigate here is related to the non-Hermitian TD Swanson Hamiltonian$$H(t)=\omega(t)\left( a^{\dagger}a+1/2\right) +\alpha(t)a^{2}+\beta (t)a^{\dagger2}\text{,} \label{5}$$ where $a$ and $a^{\dagger}$ are bosonic annihilation and creation operators, for instance of a light field mode. In comparison with time-independent case all parameters have acquired an explicit time-dependence $\omega (t),\alpha(t),\beta(t) \in\mathbb{C}$. Clearly when $\omega(t) \notin \mathbb{R}$ or $\alpha(t)\neq\beta^{\ast}(t)$ the Hamiltonian (\[5\]) is not Hermitian. It becomes PT-symmetric when demanding $\omega(t),\alpha (t),\beta(t)$ to be even functions in $t$ or generic functions of $it$. Let us now solve the TD Dyson equation by making the following general and for simplicity Hermitian Ansatz for the Dyson map $$\begin{aligned} \eta(a,a^{\dagger},t) & =\exp\left[ \epsilon(t)\left( a^{\dagger }a+1/2\right) +\mu(t)a^{2}+\mu^{\ast}(t)a^{\dagger2}\right] \label{6a}\\ & =\exp\left[ \lambda_{+}(t)K_{+}\right] \exp\left[ \ln\lambda_{0}(t) K_{0}\right] \exp\left[ \lambda_{-}(t)K_{-}\right] \text{.} \label{6}$$ We require here the variant (\[6\]) of our Ansatz to be able to compute the time-derivatives of $\eta$. The equality follows by recalling that $K_{+}=a^{\dagger2}/2$, $K_{-}=a^{2}/2$, $K_{0}=(a^{\dagger}a/2+1/4)$ form an $SU(1,1)$-algebra, such that the group element in (\[6a\]) can be Iwasawa decomposed according to [@Klimov]. The TD coefficients read \[7\]$$\begin{aligned} \lambda_{+} & =\frac{2\mu^{\ast}\sinh\Xi}{\Xi\cosh\Xi-\epsilon\sinh\Xi }\text{,}\label{7a}\\ \lambda_{-} & =\lambda_{+}^{\ast}\text{,}\label{7b}\\ \lambda_{0} & =\left( \cosh\Xi-\frac{\epsilon}{\Xi}\sinh\Xi\right) ^{-2}\text{,} \label{7c}$$ where we abbreviated the argument of the hyperbolic functions to $\Xi =\sqrt{\epsilon^{2}-4\left\vert \mu\right\vert ^{2}}$, demanding $\epsilon$ to be real and $\epsilon^{2}-4\left\vert \mu\right\vert ^{2}\geq0$. The notation may be simplified even further when introducing some new quantities. Similarly as in [@Musumbu] we define $z=2\mu/\epsilon =\left\vert z\right\vert e^{i\varphi}$ within the unit circle, such that we obtain $\Xi=\epsilon\sqrt{1-\left\vert z\right\vert ^{2}}$. Furthermore, we define $\Phi=\left\vert z\right\vert /\Gamma_{-}$ with $\Gamma_{\pm}=1\pm\tilde{\Xi}\coth\Xi$, $\tilde{\Xi}=\Xi/\epsilon$, $\tilde{\Gamma}_{\pm }=\Gamma_{\pm}/\tilde{\Xi}$, and finally $\chi=\tilde{\Gamma}_{+}/\tilde{\Gamma}_{-}=2/\Gamma_{-}-1=2\Phi/\left\vert z\right\vert -1$. The notation settled, the coefficients in (\[7a\])-(\[7c\]) simplify to \[8\]$$\begin{aligned} \lambda_{+} & =-\Phi e^{-i\varphi}\text{,}\label{8a}\\ \lambda_{-} & =-\Phi e^{i\varphi}\text{,}\label{8b}\\ \lambda_{0} & =\frac{1}{\tilde{\Gamma}_{-}^{2}\sinh^{2}\Xi}=\Phi^{2}-\chi\text{.}\label{8c}$$ where $\sinh^{2}\Xi=\tilde{\Xi}^{2}\Phi^{2}/\left[ \left\vert z\right\vert ^{2}\left( \Phi^{2}-\chi\right) \right] $ $=\tilde{\Xi}^{2}\lambda _{+}\lambda_{-}/\left\vert z\right\vert ^{2}\lambda_{0}$. Using the relations $$\eta(t)\begin{pmatrix} a\\ a^{\dagger}\end{pmatrix} \eta^{-1}(t)=\pm\frac{1}{\sqrt{\lambda_{0}}}\begin{pmatrix} -1 & \lambda_{+}\\ -\lambda_{-} & \chi \end{pmatrix}\begin{pmatrix} a\\ a^{\dagger}\end{pmatrix} \text{,} \label{9}$$ we obtain, after some algebra, the transformed Hamiltonian $$\begin{aligned} h(z,\epsilon,t) & =\eta(t)H(t)\eta^{-1}(t)+i\dot{\eta}(t)\eta^{-1}(t)\nonumber\\ & =W(z,\epsilon,t)(a^{\dagger}a+1/2)+V(z,\epsilon,t)a^{2}+T(z,\epsilon ,t)a^{\dagger2}\text{,} \label{10}$$ where the coefficient functions are \[11\]$$\begin{aligned} W(z,\epsilon,t) & =-\frac{1}{\lambda_{0}}\left[ \omega\left( \chi +\lambda_{+}\lambda_{-}\right) \right. \left. +2\left( \alpha\lambda _{+}+\beta\chi\lambda_{-}\right) -\frac{i}{2}\left( \dot{\lambda}_{0}-2\lambda_{+}\dot{\lambda}_{-}\right) \right] \text{,}\label{11a}\\ V(z,\epsilon,t) & =\frac{1}{\lambda_{0}}\left( \alpha+\omega\lambda _{-}+\beta\lambda_{-}^{2}+\frac{i}{2}\dot{\lambda}_{-}\right) \text{,}\label{11b}\\ T(z,\epsilon,t) & =\frac{1}{\lambda_{0}}\left[ \omega\chi\lambda_{+}+\alpha\lambda_{+}^{2}+\beta\chi^{2}+\frac{i}{2}\left( \lambda_{0}\dot{\lambda}_{+}+\lambda_{+}^{2}\dot{\lambda}_{-}-\lambda_{+}\dot{\lambda }_{0}\right) \right] \text{.} \label{11c}$$ As common the overhead dot denotes derivatives with respect to time. For the Hamiltonian $h(t)$ to be Hermitian we need to impose $W$ to be real and in addition $T=V^{\ast}$. From the first constraint we derive the equality $$\dot{\lambda}_{0}=2\left\vert \omega\right\vert \left( \chi+\Phi^{2}\right) \sin\varphi_{\omega}+2\Phi\left[ \dot{\Phi}+2\left\vert \alpha\right\vert \sin\left( \varphi-\varphi_{\alpha}\right) -2\left\vert \beta\right\vert \chi\sin\left( \varphi+\varphi_{\beta}\right) \right] \text{,} \label{12}$$ while the second one leads to the coupled nonlinear differential equations $$\begin{aligned} \dot{\Phi} & =\frac{2}{\chi-1}\left\{ \left[ \left\vert \omega\right\vert \Phi\sin\varphi_{\omega}+\left\vert \alpha\right\vert \sin\left( \varphi-\varphi_{\alpha}\right) \right] \left( 1-\Phi^{2}\right) \right. \left. +\left\vert \beta\right\vert \left[ \left( 2\chi-1\right) \Phi ^{2}-\chi^{2}\right] \sin\left( \varphi+\varphi_{\beta}\right) \right\} \text{,}\nonumber\\ \dot{\varphi} & =\frac{2}{\left( \chi-1\right) \Phi}\left[ \left\vert \alpha\right\vert \left( 1-\Phi^{2}\right) \cos\left( \varphi -\varphi_{\alpha}\right) \right. \left. +\left\vert \beta\right\vert \left( \Phi^{2}-\chi^{2}\right) \cos\left( \varphi+\varphi_{\beta}\right) \right] +2\left\vert \omega\right\vert \cos\varphi_{\omega} \text{.} \label{13}$$ Here $\varphi_{\alpha}$, $\varphi_{\beta}$ and $\varphi_{\omega}$ are the polar angles of $\alpha$, $\beta$ and $\omega$, respectively and $\chi$ is a function of $\Phi$ and $\left\vert z\right\vert $, as defined above. Therefore, in a similar way to that in Ref. [@Musumbu], we may consider $\left\vert z\right\vert $ as the only free parameter that determines the metric, with $\epsilon$ following from the relation $$\epsilon=\frac{1}{\sqrt{1-\left\vert z\right\vert ^{2}}}\operatorname{arctanh}\frac{\sqrt{1-\left\vert z\right\vert ^{2}}\Phi}{\Phi-\left\vert z\right\vert } =\frac{1}{2\sqrt{1-\left\vert z\right\vert ^{2}}}\ln\left[ \frac{\left( 1+\sqrt{1-\left\vert z\right\vert ^{2}}\right) \Phi-\left\vert z\right\vert }{\left( 1-\sqrt{1-\left\vert z\right\vert ^{2}}\right) \Phi-\left\vert z\right\vert } \right] \text{,} \label{14b}$$ as may be derived from the parameter $\Phi=\left\vert z\right\vert /\Gamma _{-}$, as defined above, which in turn depends, as well as on $\varphi$, also on the solution of the system (\[13\]) and the TD coefficients of the starting Hamiltonian (\[5\]). Evidently, a given pair ($\left\vert z\right\vert ,\Phi$), i.e., a given choice of $\left\vert z\right\vert $, this must be further corroborated by a real solution of $\epsilon$ in Eq. (\[14b\]), with the argument of the $\operatorname{arctanh}$ ($\ln$) being not greater than unity (greater than zero), thus demanding $\left\vert z\right\vert ^{2}>2\Phi/\left( 1+\Phi^{2}\right) $. We finally observe that $\left\vert z\right\vert $ can conveniently be considered as a time-independent parameter, constraining the time-dependence to the remaining parameters $\varphi$ and $\epsilon$. Solutions of the Schr[ö]{}dinger equation for the generalized time-dependent Swanson Hamiltonian ================================================================================================ In order to solve the SE for $H(t)$ we shall adapt to the case of TD non-Hermitian Hamiltonians a method presented in Ref. [@MBS] for solving the SE for TD Hermitian Hamiltonians. This technique takes advantage of a TD transformation on the SE for the desired Hamiltonian, here a nonunitary transformation to conform with non-Hermitian Hamiltonians, and the diagonalization of a TD Invariant within the Lewis and Riesenfeld framework [@LR]. The Lewis and Riesenfeld method ensures that a solution of the SE governed by a TD Hermitian Hamiltonian $\mathcal{H}(t)$ is an eigenstate of an associated Hermitian invariant $I(t)$, defined as $\partial_{t}I(t)+i\left[ \mathcal{H}(t),I(t)\right] =0$, apart from a TD global phase factor. The method in Ref. [@MBS] proposes that, instead of solving the SE for $\mathcal{H}(t)$ by deriving an invariant directly associated with this Hamiltonian, a transformation is performed on the SE for bringing the original Hamiltonian to another form which has already an associated invariant. The authors in Ref. [@MBS] pursued the solution of the SE governed by a general TD quadratic (Hermitian) Hamiltonian in order to investigate the mechanism of squeezed states [@Walls; @Scully] following from the nonlinear amplification terms of the Hamiltonian. They thus consider the unitary squeeze operator for transforming the SE for the TD quadratic Hamiltonian, reducing it to a form associated with a linear Hamiltonian which has already an associated invariant [@PLawande]. Here, we shall focus on the method to approach a TD non-Hermitian Hamiltonian, leaving open the analysis of the squeezing mechanism coming from the nonlinear terms of a TD non-Hermitian Hamiltonian. In the present contribution a similar strategy to that in Ref. [@MBS] will be used, starting from the non-Hermitian $H(t)$ and then deriving the transformed Hermitian $h(t)$ through the metric operator $\eta(t)$, instead of a unitary transformation. We further identify this transformed Hamiltonian with the Hermitian quadratic one treated in Ref. [@MBS], whose solutions have been derived. Evidently, we must disregard the linear amplification process considered in Ref. [@MBS] since it is absent from $h(t)$. To this end, we next rewrite the coefficients of the Hermitian (\[10\]) considering the Eqs. (\[12\]) and (\[13\]). Under the Eqs. (\[12\]) and (\[13\]) we obtain the real frequency$$W(\left\vert z\right\vert ,\varphi,t)=\left\vert \omega\right\vert \cos \varphi_{\omega}+\frac{2\Phi}{1-\chi}\left[ \left\vert \alpha\right\vert \cos\left( \varphi-\varphi_{\alpha}\right) -\left\vert \beta\right\vert \cos\left( \varphi+\varphi_{\beta}\right) \right] \text{.} \label{15}$$ From the system (\[13\]) we obtain $V(\left\vert z\right\vert ,\varphi ,t)=T^{\ast}(\left\vert z\right\vert ,\varphi,t)=V_{R}(\left\vert z\right\vert ,\varphi,t)+iV_{I}(\left\vert z\right\vert ,\varphi,t)=\kappa(\left\vert z\right\vert ,\varphi,t)e^{i\zeta(\left\vert z\right\vert ,\varphi,t)}$, with $\kappa=\left( V_{R}^{2}+V_{I}^{2}\right) ^{1/2}$, $\zeta=\arctan\left( V_{I}/V_{R}\right) $, and \[F\]$$\begin{aligned} V_{R}(\left\vert z\right\vert ,\varphi,t) & =\frac{1}{1-\chi}\left( \left\vert \omega\right\vert \Phi\sin\varphi_{\omega}\sin\varphi+\left\vert \alpha\right\vert \cos\varphi_{\alpha}-\left\vert \beta\right\vert \chi \cos\varphi_{\beta}\right) \text{,}\label{Fa}\\ V_{I}(\left\vert z\right\vert ,\varphi,t) & =\frac{1}{\chi-1}\left( \left\vert \omega\right\vert \Phi\sin\varphi_{\omega}\cos\varphi-\left\vert \alpha\right\vert \sin\varphi_{\alpha}-\left\vert \beta\right\vert \chi \sin\varphi_{\beta}\right) \text{.} \label{Fb}$$ Note that when starting with a Hermitian Hamiltonian (\[5\]), with real $\omega$ and $\alpha=\beta^{\ast}$, we verify from Eqs. (\[15\]) and Eq. (\[F\]) that $W(\left\vert z\right\vert ,\varphi,t)=\left\vert \omega\right\vert $ and $V(\left\vert z\right\vert ,\varphi,t)=\alpha(t)$, such that $h=H$. The solutions of the Schrödinger equation generated by Hamiltonian (\[10\]), given in Ref. [@MBS] as $$\left\vert v_{n}(t)\right\rangle =U(t)\left\vert n\right\rangle \text{,} \label{17}$$ define a complete set of states, $\left\vert n\right\rangle $ being the Fock states and $U(t)$ the unitary operator $$U(t)=\Upsilon(t)S\left[ \xi(t)\right] D\left[ \theta(t)\right] R\left[ \Omega(t)\right] \text{.} \label{18}$$ Here $S\left[ \xi(t)\right] =\exp\left\{ \left[ \xi(t)a^{\dagger2}-\xi^{\ast}(t)a^{2}\right] /2\right\} $ is the squeeze operator, with $\xi(t)=r(t)e^{i\phi(t)}$ defining the squeeze parameters, which follow from another set of coupled nonlinear differential equations \[19\]$$\begin{aligned} \dot{r}(t) & =-2\kappa(t)\sin\left[ \zeta(t)-\phi(t)\right] \text{,}\label{19a}\\ \dot{\phi}(t) & =-2W(t)-4\kappa(t)\coth\left[ 2r(t)\right] \cos\left[ \zeta(t)-\phi(t)\right] \text{.} \label{19b}$$ where $D\left[ \theta(t)\right] =\exp\left[ \theta(t)a^{\dagger}-\theta^{\ast}(t)a\right] $ is the displacement operator and $\theta(t)$ satisfies the equation $i\dot{\theta}(t)=\Omega(t)\theta(t)$, with $$\Omega(t)=W(t)+2\kappa(t)\tanh r(t)\cos\left[ \zeta(t)-\phi(t)\right] \text{.} \label{20}$$ Finally, $R\left[ \Omega(t)\right] =\exp\left[ -i\varpi(t)a^{\dagger }a\right] $ is the rotation operator, with $\varpi(t)=\int_{0}^{t}\Omega(t^{\prime})dt^{\prime}$, and $\Upsilon(t)=\exp\left( -i\varpi (t)/2\right) $ is a global phase factor. Having the wave vectors in Eq. (\[17\]), we directly obtain the solutions of the Schrödinger equation generated by Hamiltonian (\[5\]), given by $$\left\vert \psi_{n}(t)\right\rangle =\eta^{-1}(t)\left\vert v_{n}(t)\right\rangle =\eta^{-1}(t)U(t)\left\vert n\right\rangle \text{.}\label{21}$$ For a generic superposition $\left\vert \psi(t)\right\rangle ={\textstyle\sum\nolimits_{n}} c_{n}\left\vert \psi_{n}(t)\right\rangle $ it follows that $$\left\vert \psi(t)\right\rangle =\eta^{-1}(t)V(t)\left\vert v(0)\right\rangle \text{,}\label{22}$$ with the evolution operator$$V(t)=U(t)U^{\dagger}(0)=\Upsilon(t)S\left[ \xi(t)\right] D\left[ \theta(t)\right] R\left[ \Omega(t)\right] S^{\dagger}\left[ \xi(0)\right] D^{\dagger}\left[ \theta(0)\right] \text{,}.\label{23}$$ At this point it is worth mentioning a theorem which can be straightforwardly adapted from Ref. [@MBS] to the context of TD non-Hermitian quantum mechanics: *If* $I(t)$ *is an invariant associated with an non-Hermitian Hamiltonian* $H(t)$*, then* $I_{\eta}(t)=\eta (t)I(t)\eta^{-1}(t)$ *is also an invariant but associated with the transformed Hermitian Hamiltonian* $h(t)$*, both invariants* $I(t)$ *and* $I_{\eta}(t)$ *sharing the same eigenvalue spectrum. Moreover, the Lewis and Riesenfeld phase is invariant under the transformation* $\eta(t)$*.* It is not difficult to see that this theorem fully supports the solutions presented in Eqs. (\[21\]) and (\[22\]). Before analyzing the observables associated with the pseudo-Hermitian $H(t)$, it is worth addressing two particular cases: when the coefficients of $H(t)$ are real TD functions and when considering a time-independent metric operator. On the solutions for the TD coupled differential equations (\[13\]), (\[19\]) and (\[25\]) ------------------------------------------------------------------------------------------ Before addressing particular cases where the coefficients of the Hamiltonian (\[5\]) are real TD functions and/or a time-independent metric operator is considered, we add a few comment on the coupled equations ruling the evolution of the metric parameters $\Phi$ and $\varphi$ \[Eqs. (\[13\]) and (\[25\])\] and the squeezing parameters $r$ and $\phi$ \[Eq. (\[19\])\]. As advanced in Ref. [@MBS], despite its time dependence, the system (\[19\]) can be solved analytically, by quadrature, under particular constraints linking together its TD functions and thus leaving a lower degree of arbitrariness. Some solutions for system (\[19\]) have been presented in Ref. [@MBS], and reasoning by analogy with this reference it will be possible to find analytical solutions for the systems (\[13\]) and (\[25\]), at least for some specific demands on the TD functions. For example, considering a real TD function $$\omega(t)\equiv\frac{\dot{f}(t)}{2}+2\left\vert \beta\right\vert \Phi \cos\left( \varphi-\varphi_{\alpha}\right)$$ and $\varphi_{\beta}(t)=-\varphi_{\alpha}(t)$, we eliminate the parameter time from the system (\[13\]), to obtain, with $\varsigma(t)=\varphi(t)+f(t)$ and a constant $v=$ $\varphi_{\alpha}(t)+f(t)$, the first order differential equation$$\frac{d\Phi}{d\varsigma}=\frac{\Phi}{\tan\left( \varsigma-v\right) }\text{,}$$ whose integration leads to a constant of motion and thus to the solutions for $\Phi$ and $\varphi$. Particular cases ================ The generalized TD Swanson’s Hamiltonian with real coefficients $\omega(t),\alpha(t),\beta(t)$ ---------------------------------------------------------------------------------------------- When considering the TD coefficients $\omega(t),\alpha(t),\beta(t)$ to be real functions instead of complex ones, the equations in Sections III and IV considerably simplify. Let us start by demanding $h(t)$ in Eq. (\[10\]) to be Hermitian. By imposing $W$ to be real we now obtain$$\dot{\lambda}_{0}=2\Phi\left[ \dot{\Phi}+2\left( \alpha-\beta\chi\right) \sin\varphi\right] \text{,} \label{24}$$ while the imposition $T=V^{\ast}$ leads to the simplified coupled nonlinear differential equations \[25\]$$\begin{aligned} \dot{\Phi} & =\frac{2}{\chi-1}\left\{ \alpha\left( 1-\Phi^{2}\right) +\beta\left[ \left( 2\chi-1\right) \Phi^{2}-\chi^{2}\right] \right\} \sin\varphi\text{,}\label{25a}\\ \dot{\varphi} & =2\omega-\frac{2}{\left( 1-\chi\right) \Phi}\left[ \alpha\left( 1-\Phi^{2}\right) +\beta\left( \Phi^{2}-\chi^{2}\right) \right] \cos\varphi\text{.} \label{25b}$$ Again, $\left\vert z\right\vert $ can be taken as the only free parameter that determines the metric, with $\epsilon$ following from Eq. (\[14b\]). To further identify the transformed Hermitian $h(t)$ with the quadratic Hamiltonian whose SE is solved in Ref. [@MBS], we rewrite the coefficients of $h(t)$ considering the Eqs. (\[24\]) and (\[25\]). We thus obtain the real frequency $$W(\left\vert z\right\vert ,\varphi,t)=\omega+\frac{2\Phi}{1-\chi}\left[ \alpha-\beta\right] \cos\varphi\text{,} \label{26}$$ and the simplified real function \[16\]$$V(\left\vert z\right\vert ,\varphi,t)=T(\left\vert z\right\vert ,\varphi ,t)=\kappa(\left\vert z\right\vert ,\varphi,t)=\frac{\alpha-\beta\chi}{1-\chi }\text{.} \label{27}$$ From the above equations the solutions presented in Eqs. (\[21\]) and (\[22\]) follow straightforwardly. The generalized TD Swanson’s Hamiltonian with a time-independent metric operator -------------------------------------------------------------------------------- When a time-independent metric operator is considered, the coefficients of the transformed Hamiltonian $h(z,\epsilon,t)$ simplify to \[28\]$$\begin{aligned} W(z,\epsilon,t) & =-\frac{1}{\lambda_{0}}\left[ \omega\left( \chi +\lambda_{+}\lambda_{-}\right) +2\left( \alpha\lambda_{+}+\beta\chi \lambda_{-}\right) \right] \text{,}\label{28a}\\ V(z,\epsilon,t) & =\frac{1}{\lambda_{0}}\left( \alpha+\omega\lambda _{-}+\beta\lambda_{-}^{2}\right) \text{,}\label{28b}\\ T(z,\epsilon,t) & =\frac{1}{\lambda_{0}}\left( \omega\chi\lambda_{+}+\alpha\lambda_{+}^{2}+\beta\chi^{2}\right) \text{,}\label{28c}$$ For $h$ to be Hermitian we again impose $W$ to be real and $T=V^{\ast}$. The first constraint leads to the relation $$\left\vert \omega\right\vert \left( \chi+\Phi^{2}\right) \sin\varphi _{\omega}+2\Phi\left[ \left\vert \alpha\right\vert \sin\left( \varphi -\varphi_{\alpha}\right) -\left\vert \beta\right\vert \chi\sin\left( \varphi+\varphi_{\beta}\right) \right] =0\text{,}\label{29}$$ while the latter gives rise to the equations \[30\]$$\begin{aligned} \left\vert \omega\right\vert \left( 1-\chi\right) \Phi\cos\varphi_{\omega }-\left\vert \alpha\right\vert \left( 1-\Phi^{2}\right) \cos\left( \varphi-\varphi_{\alpha}\right) +\left\vert \beta\right\vert \left( \chi ^{2}-\Phi^{2}\right) \cos\left( \varphi+\varphi_{\beta}\right) & =0\text{,}\label{30a}\\ \left\vert \omega\right\vert \left( 1+\chi\right) \Phi\sin\varphi_{\omega }+\left\vert \alpha\right\vert \left( 1+\Phi^{2}\right) \sin\left( \varphi-\varphi_{\alpha}\right) -\left\vert \beta\right\vert \left( \chi ^{2}+\Phi^{2}\right) \sin\left( \varphi+\varphi_{\beta}\right) & =0\text{.}\label{30b}$$ From Eqs. (\[29\]) and (\[30b\]) we obtain the relation $$\left\vert \alpha\right\vert \left( 1-\Phi^{2}\right) \sin\left( \varphi-\varphi_{\alpha}\right) =\left\vert \beta\right\vert \left( \chi ^{2}-\Phi^{2}\right) \sin\left( \varphi+\varphi_{\beta}\right) \text{,}\label{31}$$ which, together with Eq. (\[30a\]), gives us \[32\]$$\begin{aligned} \sin\left( \varphi-\varphi_{\alpha}\right) & =\frac{\left\vert \beta\right\vert \left( \chi^{2}-\Phi^{2}\right) }{\left\vert \omega \right\vert \left( 1-\chi\right) \Phi\cos\varphi_{\omega}}\sin\left( \varphi_{\alpha}+\varphi_{\beta}\right) \text{,}\label{32a}\\ \sin\left( \varphi+\varphi_{\beta}\right) & =\frac{\left\vert \alpha\right\vert \left( 1-\Phi^{2}\right) }{\left\vert \omega\right\vert \left( 1-\chi\right) \Phi\cos\varphi_{\omega}}\sin\left( \varphi_{\alpha }+\varphi_{\beta}\right) \text{.}\label{32b}$$ By substituting Eq. (\[32\]) back into Eq. (\[29\]), we finally obtain the equation $$\left\vert z\right\vert \Phi^{3}+\left( 2-\left\vert z\right\vert ^{2}\right) \Phi^{2}-3\left\vert z\right\vert \Phi+\left\vert z\right\vert ^{2}=0\text{,}\label{33}$$ whose roots enable us to compute $\varphi$ from Eq. (\[32\]) and then $\epsilon$ from the relation given in Eq. (\[14b\]). Here, the real frequency $W(\left\vert z\right\vert ,\varphi,t)$ and the complex $V(\left\vert z\right\vert ,\varphi,t)=T^{\ast}(\left\vert z\right\vert ,\varphi,t)$ still follow from Eqs. (\[15\]) and (\[F\]), respectively, with time-independent $z$ and $\epsilon$. ### A time-independent metric operator with real TD coefficients $\omega(t),\alpha(t),\beta(t)$ When a time-independent metric operator is considered together with real TD parameters $\omega(t),\alpha(t),\beta(t)$, it follows from Eq. (\[29\]) that $\varphi=0$ and from Eq. (\[30a\]) we derive the equation $$\left( \left\vert \alpha\right\vert -\left\vert \beta\right\vert \right) \Phi^{2}+\left\vert \omega\right\vert \left( 1-\chi\right) \Phi-\left\vert \alpha\right\vert +\left\vert \beta\right\vert \chi^{2}=0 \label{34}$$ which leads to the relation$$\frac{\tanh(2\Xi)}{\tilde{\Xi}}=\frac{\alpha-\beta}{\alpha+\beta-z\omega }\text{,} \label{35}$$ and, consequently, to the metric parameter \[36\]$$\begin{aligned} \epsilon & =\frac{1}{2\sqrt{1-\left\vert z\right\vert ^{2}}}\operatorname{arctanh}\frac{\left( \left\vert \alpha\right\vert -\left\vert \beta\right\vert \right) \sqrt{1-\left\vert z\right\vert ^{2}}}{\left\vert \alpha\right\vert +\left\vert \beta\right\vert -\left\vert z\right\vert \left\vert \omega\right\vert }\label{36a}\\ & =\frac{1}{4\sqrt{1-\left\vert z\right\vert ^{2}}}\ln\frac{\left\vert \alpha\right\vert +\left\vert \beta\right\vert -\left\vert z\right\vert \left\vert \omega\right\vert +\left( \left\vert \alpha\right\vert -\left\vert \beta\right\vert \right) \sqrt{1-\left\vert z\right\vert ^{2}}}{\left\vert \alpha\right\vert +\left\vert \beta\right\vert -\left\vert z\right\vert \left\vert \omega\right\vert -\left( \left\vert \alpha\right\vert -\left\vert \beta\right\vert \right) \sqrt{1-\left\vert z\right\vert ^{2}}}\text{.} \label{36b}$$ However, a time-independent metric brings about the constraint on the TD parameters of the Hamiltonian $$\frac{\left\vert \dot{\alpha}\right\vert +\left\vert \dot{\beta}\right\vert -\left\vert z\right\vert \left\vert \dot{\omega}\right\vert +\left( \left\vert \dot{\alpha}\right\vert -\left\vert \dot{\beta}\right\vert \right) \sqrt{1-\left\vert z\right\vert ^{2}}}{\left\vert \alpha\right\vert +\left\vert \beta\right\vert -\left\vert z\right\vert \left\vert \omega\right\vert +\left( \left\vert \alpha\right\vert -\left\vert \beta\right\vert \right) \sqrt{1-\left\vert z\right\vert ^{2}}}=\frac{\left\vert \dot{\alpha}\right\vert +\left\vert \dot{\beta}\right\vert -\left\vert z\right\vert \left\vert \dot{\omega}\right\vert -\left( \left\vert \dot{\alpha}\right\vert -\left\vert \dot{\beta}\right\vert \right) \sqrt{1-\left\vert z\right\vert ^{2}}}{\left\vert \alpha\right\vert +\left\vert \beta\right\vert -\left\vert z\right\vert \left\vert \omega\right\vert -\left( \left\vert \alpha\right\vert -\left\vert \beta\right\vert \right) \sqrt{1-\left\vert z\right\vert ^{2}}}\text{,} \label{37}$$ where we have assume a time-independent $\left\vert z\right\vert $ as the only free parameter that determines the metric, with $\epsilon$ following from Eq. (\[38\]). The existence of a real solution for $\epsilon$ demands the argument of the $\operatorname{arctanh}$ ($\ln$) to be not greater than unity (to be greater than zero), and consequently, there is no real solution for $\left\vert z\right\vert \in\left[ \left\vert z_{-}\right\vert ,\left\vert z_{+}\right\vert \right] $, with $$\left\vert z_{\pm}\right\vert =\frac{\left( \left\vert \alpha\right\vert +\left\vert \beta\right\vert \right) \left\vert \omega\right\vert \pm\left( \left\vert \alpha\right\vert -\left\vert \beta\right\vert \right) \left( \left\vert \omega\right\vert ^{2}-4\left\vert \alpha\right\vert \left\vert \beta\right\vert \right) }{\left\vert \omega\right\vert ^{2}+\left( \left\vert \alpha\right\vert -\left\vert \beta\right\vert \right) ^{2}}\text{.} \label{38}$$ The roots $\left\vert z_{\pm}\right\vert $ present the same form as those in Ref. [@Musumbu], the difference here being that $\left\vert \omega \right\vert $, $\left\vert \alpha\right\vert $, and $\left\vert \beta \right\vert $ are TD functions instead of constant parameters, additionally constrained by Eq. (\[37\]), thus placing an additional difficulty for the observance of the requirements for a real solution for $\epsilon$. Finally, when we identify the Hamiltonian $h(z,\epsilon,t)$ with the Hermitian quadratic one in ([@MBS]) we obtain for $W(\left\vert z\right\vert ,\varphi,t)$ and $V(\left\vert z\right\vert ,\varphi,t)=T^{\ast}(\left\vert z\right\vert ,\varphi,t)=\kappa(\left\vert z\right\vert ,\varphi ,t)e^{i\zeta(\left\vert z\right\vert ,\varphi,t)}$, the same expressions as in Eqs. (\[26\]) and (\[27\]), respectively, with time-independent $\left\vert z\right\vert $ and $\epsilon$. Observables =========== The generalized TD Swanson Hamiltonian -------------------------------------- Considering the observables for the generalized TD Swanson Hamiltonian, we start by focusing on the derivation of all the Hermitian operators on the continuous variety of Hilbert spaces $\mathcal{H}_{z}$ for any $|z|\in \lbrack-1,1]$. As argued in [@Andreas], the Hamiltonian $H$ itself is not one of the Hermitian operator due the presence of the gauge-like term in Eq. (\[1\]). Using Eq. (\[14b\]) to rewrite the metric operator in Eq. (\[6\]) in the form [@Musumbu; @GHS] \[39\]$$\begin{aligned} \eta(t) & =\left( \frac{\left( 1+\sqrt{1-\left\vert z\right\vert ^{2}}\right) \Phi-\left\vert z\right\vert }{\left( 1-\sqrt{1-\left\vert z\right\vert ^{2}}\right) \Phi-\left\vert z\right\vert }\right) ^{\frac{a^{\dagger}a+\frac{1}{2}\left( za^{2}+z^{\ast}a^{\dagger2}\right) +\frac{1}{2}}{2\sqrt{1-\left\vert z\right\vert ^{2}}}}\label{39a}\\ & =\left( \frac{\left( 1+\sqrt{1-\left\vert z\right\vert ^{2}}\right) \Phi-\left\vert z\right\vert }{\left( 1-\sqrt{1-\left\vert z\right\vert ^{2}}\right) \Phi-\left\vert z\right\vert }\right) ^{\frac{\left( 1-\left\vert z\right\vert \cos\varphi\right) p^{2}+\left( 1+\left\vert z\right\vert \cos\varphi\right) \omega^{2}x^{2}-\left\vert z\right\vert \omega\sin \varphi\left\{ x,p\right\} }{4\omega\sqrt{1-\left\vert z\right\vert ^{2}}}} , \label{39b}$$ which we use to solve the quasi-Hermiticity condition $O^{\dagger}(t)\mu(t)=\mu(t)O(t)$. Given (\[39\]), we only find the observables $$O(t)=\left( 1-\left\vert z\right\vert \cos\varphi\right) p^{2}+\left( 1+\left\vert z\right\vert \cos\varphi\right) \omega^{2}x^{2}-\left\vert z\right\vert \omega\sin\varphi\left\{ x,p\right\} \text{,} \label{40}$$ demonstrating that neither the position $x=\frac{1}{\sqrt{2\omega}}\left( a+a^{\dagger}\right) $ nor the momentum $p=i\sqrt{\frac{\omega}{2}}\left( a^{\dagger}-a\right) $ operators remain Hermitian as they are in the standard $L^{2}$-metric, with regard to the TD $\eta(t)$-metric even for particular choices of $\left\vert z\right\vert $. Using the relation $O(t)=\eta ^{-1}(t)o\eta(t)$ together with Eq. (\[13\]) we may compute the quasi-Hermitian position $X(t)$ and momentum $P(t)$ operators \[41\]$$\begin{aligned} X(t) & =\frac{1}{\left\vert z\right\vert \sqrt{\Phi^{2}-\chi}}\left\{ \left[ \left( 1-i\left\vert z\right\vert \sin\varphi\right) \Phi-\left\vert z\right\vert \right] x+\frac{i}{\omega}\left( 1-\left\vert z\right\vert \cos\varphi\right) \Phi p\right\} \text{,}\label{41a}\\ P(t) & =\frac{1}{\left\vert z\right\vert \sqrt{\Phi^{2}-\chi}}\left\{ \left[ \left( 1+i\left\vert z\right\vert \sin\varphi\right) \Phi-\left\vert z\right\vert \right] p-i\omega\left( 1+\left\vert z\right\vert \cos \varphi\right) \Phi x\right\} \text{,} \label{41b}$$ corroborating the conclusion we have drawn from Eq. (\[40\]). Particular cases ---------------- The observables computed above in Eqs. (\[40\]) and (\[41\]) also apply to the cases where real coefficients $\omega(t),\alpha(t),\beta(t)$ are assumed and when a time-independent metric operator is considered, the difference being that $\Phi$ and $\varphi$ now follow, instead of Eq. (\[13\]), from the coupled Eqs. (\[25\]) in the former case, and from Eqs. (\[32\]) and (\[33\]) in the latter case. However, when a time-independent metric operator is considered simultaneously with real coefficients $\omega (t),\alpha(t),\beta(t)$, the Hermitian observables in Eq. (\[40\]) and those in Eq. (\[41\]) simplify to \[42\]$$\begin{aligned} O(t) & =\left( 1-\left\vert z\right\vert \right) p^{2}+\left( 1+\left\vert z\right\vert \right) \omega^{2}x^{2}\text{,}\label{42a}\\ X(t) & =\frac{1}{\left\vert z\right\vert \sqrt{\Phi^{2}-\chi}}\left[ \left( \Phi-\left\vert z\right\vert \right) x+\frac{i}{\omega}\left( 1-\left\vert z\right\vert \right) \Phi p\right] \text{,}\nonumber\\ & =\cosh\left( \Xi\right) x+\frac{i}{\omega}\frac{\left( 1-\left\vert z\right\vert \right) }{\sqrt{1-\left\vert z\right\vert ^{2}}}\sinh\left( \Xi\right) p\text{,}\label{42b}\\ P(t) & =\frac{1}{\left\vert z\right\vert \sqrt{\Phi^{2}-\chi}}\left[ \left( \Phi-\left\vert z\right\vert \right) p-i\omega\left( 1+\left\vert z\right\vert \right) \Phi x\right] \text{,}\nonumber\\ & =\cosh\left( \Xi\right) p-i\omega\frac{\left( 1+\left\vert z\right\vert \right) }{\sqrt{1-\left\vert z\right\vert ^{2}}}\sinh\left( \Xi\right) x\text{.} \label{42c}$$ The Eqs. (\[42\]) are exactly of the same form as those in Ref. [@Musumbu], the difference being that here we have TD parameters. Therefore, when considering the Hamiltonian (\[5\]) with time-independent real parameters together with a time-independent metric operator, it is straightforward to verify that all the above derivations are in complete agreement with those in [@Musumbu]. Conclusion ========== We have studied a generalized Swanson Hamiltonian allowing for TD complex coefficients and a TD metric operator. We treated the model within the framework introduced in Ref. [@Andreas] where, despite the lack of the observability of the non-Hermitian Hamiltonian under a TD metric operator, their associated observables are computed as in the case where a time-independent metric is considered. To solve the SE for the generalized TD Swanson’s Hamiltonian we have adapted a technique presented in Ref. [@MBS] which relies on the Lewis and Riesenfeld TD invariants. Apart from deriving the solutions of the SE for our TD non-Hermitian Hamiltonian we have thus computed their associated observables, analyzing particular cases where a time–independent metric operator is considered and TD real coefficients are assumed for the non-Hermitian Hamiltonian. From the results presented here we may next explore some interesting applications such as the generation of squeezing from a non-Hermitian parametric oscillator. Moreover, our TD Hamiltonian can be also considered to describe the non-Hermitian dynamical Casimir effect, and thus the rate of particles creations resulting, for example, from the accelerated movement of a cavity mirror can also be computed. The results for the generation of squeezing and the rate of photon creation derived from a non-Hermitian quadratic Hamiltonian can then be compared with the well-known results coming from the Hermitian Hamiltonians, thus delivering more timely hints on the physics of non-Hermitian Hamiltonians. As another application motivated by this work is the possibility of engineering effective non-Hermitian Hamiltonians within trapped ions, circuit or cavity QED, NMR and other systems presenting great flexibility of handling its internal interactions. By mastering not only the technique for treating non-Hermitian Hamiltonians, but also for constructing non-Hermitian interactions, we may seek to contribute with the implementation of processes such as quantum simulation and quantum logical implementation, bringing additional ingredients to the usual Hermitian quantum mechanics. [**Acknowledgements**]{} M.H.Y. Moussa wishes to express his thanks to CAPES, Brazilian financial agency, and City University London for kind hospitality. [99]{} C. M. Bender and S. Boettcher, Phys. Rev. Lett. **80**, 5243 (1998). F. G. Scholtz, H. B. Geyer and F. J. W. Hahne, Ann. Phys. (N.Y.) **213,** 74 (1992). A. Mostafazadeh A, J. Math. Phys. **43**, 205 (2002); *ibid.* **43**, 2814 (2002); *ibid.* **43**, 3944 (2002). H. Jing, S. K. Özdemir, X.-Y. Lü, J. Zhang, L. Yang, and F. Nori, Phys. Rev. Lett. **113**, 053604 (2014). A. Regensburger, M.-A. Miri, C. Bersch, J. Nager, G. Onishchukov, D. N. Christodoulides, and U. Peschel, Phys. Rev. Lett. **110**, 223902 (2013). K. G. Makris, R. El-Ganainy, and D. N. Christodoulides, Phys. Rev. Lett. **100**, 103904 (2008). S. Deffner and A. Saxena, Phys. Rev. Lett. **114**, 150601 (2015). B. Peng, Ş. K. Özdemir, F. Lei, F. Monifi, M. Gianfreda, G. Lu Long, S. Fan, F. Nori, C. M. Bender, and L. Yang, Nat. Phys. **10**, 394 (2014); A. Regensburger, C. Bersch, M.-A. Miri, G. Onishchukov, D. N. Christodoulides, and U. Peschel, Nature (London) **488**, 167 (2012); M. Zhang, G. S. Wiederhecker, S. Manipatruni, A. Barnard, P. McEuen, and M. Lipson, Phys. Rev. Lett. **109**, 233906 (2012); A. Guo, G. J. Salamo, M. Volatier-Ravat, V. Aimez, G. A. Siviloglou, and D. N. Christodoulides, Phys. Rev. Lett. **103**, 093902 (2009). C.E.Ruter, K. G. Makris, R. El-Ganainy, D. N. Christodoulides, M. Segev, and D. Kip, Nat. Phys. **6**, 192 (2010); L. Feng, M. Ayache, J. Huang, Y.-L. Xu, M. H. Lu, Y. F. Chen, Y. Fainman, and A. Scherer, Science 333, 729 (2011). C. Zheng, L. Hao, and G. L. Long, Philos. Trans. R. Soc. A **371**, 20120053 (2013). J. Rubinstein, P. Sternberg, and Q. Ma, Phys. Rev. Lett. **99**, 167003 (2007); N. M. Chtchelkatchev, A. A. Golubov, T. I. Baturina, and V. M. Vinokur, Phys.Rev.Lett. **109**, 150405 (2012). C. M. Bender, Contemp. Phys. **46**, 277 (2005); Rep. Prog. Phys. **70**, 947 (2007); C. M. Bender, Rept. Prog. Phys. **70**, 947 (2007); A. Mostafazadeh, Int. J. Geom. Meth. Mod. Phys. **7**, 1191 (2010); C. Bender, A. Fring, U. Günther, and H. Jones, J. Phys. A **45**, 440301 (2012); M. Znojil, Int. J. Theor. Phys. **54**, 1572 (2015). C. Figueira de Morisson Faria and A. Fring, J. Phys. A: Math. Theor. **39**, 9269 (2006); *ibid.*, Laser Physics **17**, 424 (2007). A. Mostafazadeh, Phys. Lett. B **650**, 208 (2007); *ibid.*, arXiv:0711.0137 (2007); *ibid.*, arXiv:0711.1078 (2007). M. Znojil, arXiv:0710.5653 (2007); *ibid.*, arXiv:0711.0514 (2007); *ibid*., Phys. Rev. D **78**, 085003 (2008). J. Gong and Q.-h. Wang, J. Phys. A: Math. Theor. **46** 485302 (2013). M. Maamache, Phys. Rev. A **92**, 032106 (2015). A. Fring and M. H. Y. Moussa, Phys. Rev. A **93**, 042114 (2016). M. S. Swanson, J. Math. Phys. **45**, 585 (2004). S. S. Mizrahi, M. H. Y. Moussa, and B. Baseia, Int. J. Mod. Phys. B **8**, 1563 (1994); B. Baseia, S. S. Mizrahi, and M. H. Y. Moussa, Phys. Rev. A **46**, 5885 (1992). See also C. J. Villas-Boas, F. R. de Paula, R. M. Serra, and M. H. Y. Moussa, Phys. Rev. A **68**, 053808 (2003); *ibid.* J. Opt. B **5**, 391 (2003); S. Dey and A. Fring, Phys. Rev. D **90**, 084005 (2014). H. R. Lewis, Jr. and W. B. Riesenfeld, J. Math. Phys. **10**, 1458 (1969). D. F. Walls and G. J. Milburn, Quantum Optics (Springer-Verlag, Berlin, 1995). M. O. Scully and M. S. Zubairy, Quantum Optics (Cambridge, London, 1997). Andrei B. Klimov and Sergei M. Chumakov, A Group-Theoretical Approach to Quantum Optics: Models of Atom-Field Interactions (Wiley-VCH, Weinheim, 2009); S. M. Barnett, P. Radmore: Methods in Theoretical Quantum Optics (Oxford. University Press, New York, Oxford, 1997). D.P. Musumbu, H.B. Geyer and W.D. Heiss, J. Phys. A **40**, F75 (2007). R. R. Puri and S. V. Lawande, Phys. Lett. A **70**, 69 (1979). H. B. Geyer, W. D. Heiss, and F. G. Scholtz, arXiv:0710.5593 (unpublished).

ArXiv

--- author: - | Dinh Tien Tuan Anh, Quach Vinh Thanh, Anwitaman Datta\ [email protected], {vtquach,anwitaman}@ntu.edu.sg title: 'CloudMine: Multi-Party Privacy-Preserving Data Analytics Service' ---

ArXiv

--- abstract: 'We develop an efficient algorithm for a spatially inhomogeneous matrix-valued quantum Boltzmann equation derived from the Hubbard model. The distribution functions are $2 \times 2$ matrix-valued to accommodate the spin degree of freedom, and the scalar quantum Boltzmann equation is recovered as special case when all matrices are proportional to the identity. We use Fourier discretization and fast Fourier transform to efficiently evaluate the collision kernel with spectral accuracy, and numerically investigate periodic, Dirichlet and Maxwell boundary conditions. Model simulations quantify the convergence to local and global thermal equilibrium.' address: - 'Departments of Mathematics, Physics, and Chemistry, Duke University, Box 90320, Durham, NC 27708 USA' - 'Mathematics Department, Technische Universität München, Boltzmannstra[ß]{}e 3, 85747 Garching bei München, Germany' author: - Jianfeng Lu - 'Christian B. Mendl' title: | Numerical scheme for a spatially inhomogeneous\ matrix-valued quantum Boltzmann equation --- Introduction ============ Boltzmann’s kinetic theory is widely used to describe the dynamics of rarified gases. Immediately after the discovery of quantum mechanics, a modification of the classical Boltzmann equation to take quantum interactions into account has been proposed by Nordheim [@Nordheim1928] (with a more systematic derivation by Peierls [@Peierls:29]), and by Uehling and Uhlenbeck [@UehlingUhlenbeck1933; @Uehling1934]. With the inclusion of quantum mechanical effects into the collision operator, the quantum Boltzmann equation has many applications, e.g., for the kinetic description of Bose-Einstein condensation [@SemikozTkachev:95; @Spohn:10], spintronics and decoherence theory in quantum computing [@Spintronics2001; @ElHajj2014; @PossannerNegulescu2011; @VacchiniHornberger2009], or kinetic modeling of semiconductor devices [@MarkowichRinghoferSchmeiser:90]. In recent works, starting from the Hubbard model in the weak interaction limit, a matrix-valued Boltzmann equation has been derived [@BoltzmannHubbard2012; @BoltzmannNonintegrable2013; @DerivationBoltzmann2013] for the spatially homogeneous setting without advection term. To describe spatially inhomogeneous systems, one combines the Boltzmann transport equation with the collision term derived in [@DerivationBoltzmann2013]: $$\label{eq:boltzmann} \partial_t W + v_x\,\partial_{x} W = \mathcal{C}[W] - i [ \vec{B} \cdot \vec{\sigma}, W],$$ where the state variable $W$ is the Wigner distribution of the spin-density matrix, $\vec{B}$ is an external magnetic field, $\vec{\sigma}$ the Pauli matrices, and the collision term $\mathcal{C}[W]$ will be specified below in Section \[sec:boltzmann\]. We emphasize that while the form of the equation follows the usual quantum Boltzmann equation, the collision term is quite different (which is systematically derived from a many-body quantum mechanics model), and the matrix-valued $W$ distinguishes the equation from the usual kinetic equations. The focus of this paper is devising an efficient algorithm for solving . Our goal in this work is twofold: First, we would like to develop a numerical scheme that systematically converges to the true solution; for that purpose, we use a spectral method in the velocity variable. The collision operator, albeit much more complicated than for the usual classical or scalar quantum Boltzmann equation, can be efficiently calculated using Carleman representation and fast Fourier transforms. Second, we want to investigate non-trivial boundary conditions, like Dirichlet and Maxwell boundary conditions, and the effect of external magnetic fields. These developments should lead to a better understanding of the physics modeled by these equations. Kinetic equations are traditionally solved by Monte Carlo methods (also known as particle methods). In recent years, the development of efficient real space or Fourier space methods to solve Boltzmann equations has been a very active research area. In particular, the line of research initiated by [@MouhotPareschi:06] and further developed in [@FilbertHuJin:12; @HuYing:12] is especially relevant for our approach. The paper [@MouhotPareschi:06] proposed a fast algorithm for computing the Boltzmann collision kernel based on Fourier discretization, and [@HuYing:12] further improved the efficiency of the algorithm. The method we develop in this work for the collision operator of the matrix-valued Boltzmann equation is closely related, albeit with some differences: (a) Since we are dealing with collision terms originating from quantum mechanics, the microscopic energy is not necessarily conserved (see the effective Hamiltonian in Eq.  below); some new ideas are required to treat the resulting terms. (b) To evaluate double convolutions exactly using fast Fourier transforms, we use a double padding approach to avoid aliasing issues. As a result, while maintaining the spectral accuracy, the scheme also nicely respects the conservation law of the continuous equation. Further details can be found in Section \[sec:collision\]. A numerical algorithm for the spatially homogeneous matrix-valued Boltzmann equation in one dimension was considered before in [@BoltzmannHubbard2012; @BoltzmannNonintegrable2013], which calculates the collision term directly using numerical quadrature. A subsequent work by one of the authors [@Mendl2013] considers a lattice Boltzmann method (LBM) for the spatially inhomogeneous equation with periodic boundary conditions, which can be understood as a discrete velocity method with very few velocity grid points. Due to the small number of grid points, the accuracy of the numerical result compared to the original equation is not guaranteed. In contrast to that, the method proposed here systematically approximates the original equation as we refine the grid. Considering the *matrix-valued* Boltzmann equation, it is useful to represent the spin-density Wigner distribution in the basis of Pauli matrices. In fact, the formulas for the collision terms are more compact in the new representation, which might be of independent interest for understanding the physics and mathematics of the equation. The rest of the paper is organized as follows. We introduce the spatially inhomogeneous matrix-valued Boltzmann equation and its associated boundary conditions in Section \[sec:boltzmann\]. We will focus on the development of the fast algorithm for the collision operator in Section \[sec:collision\]. To deal with the spatial degree of freedom, we use the finite volume method and a parallel implementation based on MPI; this is discussed in Section \[sec:parallel\]. We show some numerical results for validating the algorithm and for exploring interesting physical phenomena in Section \[sec:numerics\]. Finally we wrap up the paper with some conclusive remarks in Section \[sec:conclusion\]. The spatially inhomogeneous matrix-valued Boltzmann equation {#sec:boltzmann} ============================================================ The starting point for the derivation [@DerivationBoltzmann2013] of the matrix-valued Boltzmann equation is the Hubbard model with a weak pair potential $\lambda V$ such that $0 < \lambda \ll 1$. Consider a spin-$\frac{1}{2}$ Fermi field with annihilation operators $a_{s}(x)$, $x \in {\mathbb{Z}}^d$, $s \in \{\uparrow, \downarrow \}$, obeying the anti-commutation relations $$\big\{ a_{s}(x)^{\dagger}, a_{s'}(x') \big\} = \delta_{xx'} \delta_{ss'}, \quad \big\{ a_{s}(x), a_{s'}(x') \big\} = 0, \quad \text{and} \quad \big\{ a_{s}(x)^{\dagger}, a_{s'}(x')^{\dagger} \big\} = 0,$$ where $A^{\dagger}$ denotes the adjoint operator of $A$. Using the second quantization formulation, the (many-body) Hamiltonian of the Hubbard system is then given by $$\label{eq:HubbardHamiltonian} H = \sum_{x, y \in {\mathbb{Z}}^d} \sum_{s \in \{\uparrow, \downarrow\}} \alpha(x - y) a_{s}(x)^{\dagger} a_{s}(y) + \tfrac{1}{2} \sum_{x \in {\mathbb{Z}}^d} \sum_{s, s' \in \{\uparrow, \downarrow\}} \lambda V(x-y) a_{s}(x)^{\dagger}a_{s}(x) a_{s'}(y)^{\dagger} a_{s'}(y).$$ Here the first term on the right hand side in the Hamiltonian describes the hopping from site $y$ to $x$ with $\alpha$ the hopping amplitude, and the non-quadratic second term gives the interactions of two excitons with $V$ the interaction potential. The grid ${\mathbb{Z}}^d$ must be distinguished from the spatial dimension considered below (in some sense, the grid ${\mathbb{Z}}^d$ is on the microscopic scale while the spatial inhomogeneity is introduced on a mesoscopic scale). In Fourier representation, the time-dependent (Heisenberg picture) field operators $\hat{a}_s(t,v)$ adhere to the initial ($t = 0$) anti-commutation relation $\{ \hat{a}_s(v)^{\dagger}, \hat{a}_{s'}(v')\} = \delta_{s s'}\,\delta(v - v')$, with $v$, $v'$ denoting velocity variables. As discussed in [@DerivationBoltzmann2013], the time-dependent average Wigner matrix $W$ defined by $$\label{eq:defnW} \big\langle \hat{a}_s(t,v)^{\dagger} \, \hat{a}_{s'}(t,v') \big\rangle = \delta(v - v') W_{s s'}(t,v)$$ will approximately satisfy a Boltzmann kinetic equation $\partial_t W = \mathcal{C}[W]$ for times up to order $\lambda^{-2}$. Here, as in the Heisenberg picture, the average $\langle \cdot \rangle$ is taken with respect to the initial state of the system. The effective Boltzmann equation is much easier to solve compared to the original quantum many-body system, which is an extremely high-dimensional problem. Augmenting the Boltzmann equation with the usual transport term for the spatially inhomogeneous setting and including an external magnetic field $\vec{B}$, one arrives at Eq. , where $\vec{\sigma} = (\sigma_1, \sigma_2, \sigma_3)$ are the Pauli matrices: $$\sigma_1 = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}, \quad \sigma_2 = \begin{pmatrix} 0 & -i \\ i & 0 \end{pmatrix}, \quad \sigma_3 = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}.$$ Hence $\vec{B} \cdot \vec{\sigma} = \sum_{i=1}^3 B_i \sigma_i$. In , the Hermitian spin-density matrix Wigner distribution $W: {\mathbb{R}}_+ \times \Omega \times {\mathbb{R}}^2 \to {\mathbb{C}}^{2\times2}$ additionally depends on the spatial location $x \in \Omega$. When the system is spatially homogeneous, $W$ as defined in is a positive-semidefinite Hermitian matrix with eigenvalues in $[0, 1]$ at any $(t, v)$. In the case of where spatial inhomogeneity is introduced on a mesoscopic scale, we likewise assume that initially the Wigner matrix $W(0, x, v)$ is positive-semidefinite with eigenvalues in $[0, 1]$ at any $(x, v)$. This property is preserved by the evolution of . For simplicity, we will only consider the case that $\Omega$ is an open and bounded interval on ${\mathbb{R}}$; and assume that the velocity space is ${\mathbb{R}}^2$. In other words, we are considering the case with one space dimension and two velocity dimensions. Physically, this means that the solution is homogeneous with respect to one of the spatial variables for a full two-dimensional (two space and two velocity dimensions) model. Without loss of generality, we will also assume $\Omega = (0,1)$. The collision operator in consists of a conservative and dissipative part: $\mathcal{C} = \mathcal{C}_{\mathrm{c}} + \mathcal{C}_{\mathrm{d}}$, where [@BoltzmannHubbard2012; @BoltzmannNonintegrable2013] $$\label{eq:Cc} \mathcal{C}_{\mathrm{c}}[W](t, x, v) = -i \big[ H_{{\mathrm{eff}}}(t, x, v), W(t, x, v) \big]$$ with the effective Hamiltonian $$\begin{gathered} \label{eq:Heff} H_{{\mathrm{eff}}}(t,x,v_1) = \int {\,\mathrm{d}}v_2 {\,\mathrm{d}}v_3 {\,\mathrm{d}}v_4 \, \delta(\underline{v}) \mathcal{P}( 1 / \underline{\omega} ) (W_3 W_4 - W_2 W_3 - W_3 W_2 - \operatorname{tr}[W_4] W_3 + \operatorname{tr}[W_2] W_3 + W_2).\end{gathered}$$ Here $\mathcal{P}$ denotes the principal value, and we have used the shorthand notations $W_i = W(t, x, v_i)$ for $i = 1, \dots, 4$, the velocity difference $\underline{v} = v_1 + v_2 - v_3 - v_4$, and the energy difference $\underline{\omega} = \omega(v_1) + \omega(v_2) - \omega(v_3) - \omega(v_4)$. The energy (or dispersion relation) $\omega(v)$ is precisely the Fourier transform of the hopping amplitude $\alpha$ appearing in . The terms $W_i W_j$ are usual matrix products. Due to the invariance under $v_3 \leftrightarrow v_4$, the matrix product $W_3 W_4$ in the integrand could be replaced by $W_4 W_3$; in particular, $H_{{\mathrm{eff}}}$ is Hermitian. The dissipative part of the collision operator is given by $$\label{eq:Cd} \mathcal{C}_{\mathrm{d}}[W](t, x, v_1) = \pi \int {\,\mathrm{d}}v_2 {\,\mathrm{d}}v_3 {\,\mathrm{d}}v_4 \delta(\underline{v}) \delta(\underline{\omega}) \big( \mathcal{A}[W]_{1234} + \mathcal{A}[W]^{\ast}_{1234} \big)$$ with $$\begin{gathered} \mathcal{A}[W]_{1234} = - W_4 {\tilde{W}}_2 W_3 + W_4 \operatorname{tr}[{\tilde{W}}_2 W_3] \\ - \bigl({\tilde{W}}_4 W_3 - {\tilde{W}}_4 W_2 - {\tilde{W}}_2 W_3 + {\tilde{W}}_4 \operatorname{tr}[W_2] - {\tilde{W}}_4 \operatorname{tr}[W_3] + \operatorname{tr}[W_3 {\tilde{W}}_2]\bigr) W_1,\end{gathered}$$ where ${\tilde{W}} = \mathbbm{1} - W$. As explained in [@BoltzmannHubbard2012; @BoltzmannNonintegrable2013], the first two summands (plus their Hermitian conjugates) can be identified as gain term and $( \dots ) W_1$ (plus Hermitian conjugate) as loss term. Alternatively, by making use of the invariance under $v_3 \leftrightarrow v_4$, the integrand of $\mathcal{C}_{\mathrm{d}}[W]$ can be represented as $$\mathcal{A}[W]_{1234} + \mathcal{A}[W]^{\ast}_{1234} \equiv \mathcal{A}_{\mathrm{quad}}[W]_{1234} + \mathcal{A}_{\mathrm{tr}}[W]_{1234}$$ with $$\label{eq:AW_Hubbard} \begin{split} \mathcal{A}_{\mathrm{quad}}[W]_{1234} &= -\tilde{W}_1 W_3 \tilde{W}_2 W_4 - W_4 \tilde{W}_2 W_3 \tilde{W}_1 + W_1 \tilde{W}_3 W_2 \tilde{W}_4 + \tilde{W}_4 W_2 \tilde{W}_3 W_1, \\ \mathcal{A}_{\mathrm{tr}}[W]_{1234} &= \big(\tilde{W}_1 W_3 + W_3 \tilde{W}_1\big) \operatorname{tr}[\tilde{W}_2 W_4] - \big(W_1 \tilde{W}_3 + \tilde{W}_3 W_1\big) \operatorname{tr}[W_2 \tilde{W}_4]. \end{split}$$ The representation emphasizes the similarity to the scalar collision operator [@UehlingUhlenbeck1933], which is recovered when all $W_i$ are proportional to the identity matrix. For what follows, we always take $\omega(v) = \frac{1}{2} {\lvertv\rvert}^2$ as dispersion relation. Note that if a general dispersion relation is taken, the transport term in changes to $(\partial_{v_x} \omega(v))\cdot \partial_x W$. One may check that the collision operator $\mathcal{C} = \mathcal{C}_{\mathrm{c}} + \mathcal{C}_{\mathrm{d}}$ satisfies the density, momentum and energy conservation laws (at each $x$) $$\int \mathcal{C}[W](v) {\,\mathrm{d}}v = 0, \quad \int v \operatorname{tr}\bigl[\mathcal{C}[W](v)\bigr] {\,\mathrm{d}}v = 0, \quad \int \tfrac{1}{2}{\lvertv\rvert}^2 \operatorname{tr}\bigl[ \mathcal{C}[W](v) \bigr] {\,\mathrm{d}}v = 0.$$ As a result, the corresponding fluid dynamic moments, i.e., density $\rho(t, x) \in {\mathbb{C}}^{2 \times 2}$, velocity $u(t, x) \in {\mathbb{R}}^2$, and internal energy $\varepsilon(t, x) \in {\mathbb{R}}$ $$\begin{aligned} \rho(t, x) & = \int W(t, x, v) {\,\mathrm{d}}v; \label{eq:density}\\ \operatorname{tr}[\rho(t, x)] u(t, x) & = \int v \operatorname{tr}[W(t, x, v)] {\,\mathrm{d}}v; \label{eq:velocity}\\ \operatorname{tr}[\rho(t, x)] \varepsilon(t, x) & = \int \tfrac{1}{2}{\lvertv - u\rvert}^2 \operatorname{tr}[W(t, x, v)] {\,\mathrm{d}}v \label{eq:internal_energy}\end{aligned}$$ satisfy local conservation laws. The (local) entropy of the state $W$ is defined as $$\label{eq:entropy_def} S[W](t,x) = - \int \operatorname{tr}\bigl[ W(t,x,v) \log W(t,x,v) + {\tilde{W}}(t,x,v) \log {\tilde{W}}(t,x,v) \bigr] {\,\mathrm{d}}v.$$ The H-theorem states that the global entropy production rate is positive (see [@BoltzmannHubbard2012] for the matrix-valued case) $$\sigma[W](t) := \frac{{\,\mathrm{d}}}{{\,\mathrm{d}}t} \int S[W](t,x) {\,\mathrm{d}}x = - \int \int \operatorname{tr}\Bigl[ \bigl(\log W(t,x,v) - \log {\tilde{W}}(t,x,v) \bigr) \mathcal{C}[W](t,x,v) \Bigr] {\,\mathrm{d}}v {\,\mathrm{d}}x \geq 0$$ for all $W$ with eigenvalues in $[0, 1]$ and periodic boundary conditions. The advection term in the integrand vanishes since we integrate over the spatial domain. In the asymptotic long-time limit $t \to \infty$ for a closed system (with periodic boundary conditions) and in the absence of external fields, the solution of the Boltzmann equation is expected to converge to the Fermi-Dirac distribution $$\label{eq:W_FermiDirac} W_{{\mathrm{FD}}}(v) = \sum_{s \in \{ \uparrow, \downarrow \}} \bigl( \mathrm{e}^{( \omega(v) - \mu_{s})/(k_{\mathrm{B}}T)} + 1\bigr)^{-1} {\lverts\rangle}{\langles\rvert}$$ for a $v$-independent spin basis ${\lverts\rangle}$, temperature $T$, and chemical potentials $\mu_{\uparrow}$ and $\mu_{\downarrow}$ (see [@BoltzmannHubbard2012] for a proof of convergence in the spatially homogeneous case). Note that the Fermi-Dirac distribution maximizes the entropy among states with the same fluid dynamic moments. The moments of $W_{{\mathrm{FD}}}$ have analytical expressions: in two dimensions and for the dispersion $\omega(v) = \frac{1}{2} {\lvertv\rvert}^2$, $$\label{eq:momentsFermiDirac} \rho_{{\mathrm{FD}}} = 2\pi k_{\mathrm{B}} T \sum_{s \in \{ \uparrow, \downarrow \}} \log\bigl(1 + \mathrm{e}^{\mu_s / (k_{\mathrm{B}}T)} \bigr), \quad \varepsilon_{{\mathrm{FD}}} = k_{\mathrm{B}} T \frac{-\sum_{s \in \{ \uparrow, \downarrow \}} \mathrm{Li}_2\bigl(-\mathrm{e}^{\mu_s / (k_{\mathrm{B}}T)} \bigr)}{\sum_{s \in \{ \uparrow, \downarrow \}} \log\bigl(1 + \mathrm{e}^{\mu_s / (k_{\mathrm{B}}T)} \bigr)}$$ where $\mathrm{Li}_n$ is the polylogarithm function. The average velocity of $W_{{\mathrm{FD}}}$ in is zero. To complete the equation, we need to impose the boundary conditions. Let $\Sigma = \partial \Omega \times {\mathbb{R}}^2 = \{0, 1\} \times {\mathbb{R}}^2$, and denote by $n(x)$ be the outward unit normal vector at $x \in \partial \Omega$. We define the outgoing and incoming boundaries as $$\Sigma_{\pm} = \big\{ (x,v) \in \Sigma; \pm n(x) \cdot v > 0 \big\}.$$ Hence $$\begin{aligned} & \Sigma_+ = \big\{ (0,v); v_x < 0\} \cup \{(1, v); v_x > 0 \big\}; \\ & \Sigma_- = \big\{ (0,v); v_x > 0\} \cup \{(1, v); v_x < 0 \big\}.\end{aligned}$$ For the boundary condition on the incoming boundary $\Sigma_-$, we consider - periodic boundary conditions: for $(x, v) \in \Sigma_-$, $$W(t, x, v) = W(t, 1-x, v).$$ Note that $(1-x, v) \in \Sigma_+$. - Dirichlet boundary conditions: for $(x, v) \in \Sigma_-$, $$W(t, x, v) = \Phi(t, x, v)$$ where $\Phi: {\mathbb{R}}_+ \times \Sigma_- \to {\mathbb{C}}^{2\times2}$ is a given boundary state. - Maxwell boundary conditions: $$W(t, x, v) = \mathcal{R}_x\big( W(t, x, \cdot)\vert_{\Sigma_+^x}\big)(v),$$ where $\mathcal{R}_x$ is a Maxwell reflection operator: $$\mathcal{R}_x = (1 - \alpha) \mathcal{L}_x + \alpha \mathcal{D}_x.$$ Here $\alpha \in [0,1]$ is the accommodation coefficient. The local reflection operator $\mathcal{L}_x$ is given by $$\big(\mathcal{L}_x F\big)(v_x, v_y) = F(-v_x, v_y),$$ and the diffusive reflection is given by (for a specified Fermi-Dirac state depending on spin basis, $T$, $\mu_{\uparrow}$ and $\mu_{\downarrow}$) $$(\mathcal{D}_x F)(v) = Z_x^{-1} W_{{\mathrm{FD}}}(v) {\widebar{F}}(x),$$ where ${\widebar{F}}(x)$ is the total outgoing number flux $${\widebar{F}}(x) = \int_{v \cdot n(x)>0} \operatorname{tr}(F(v)) v\cdot n(x) {\,\mathrm{d}}v,$$ and $Z_x$ is a normalizing constant such that $$Z_x = \int_{v \cdot n(x)<0} \operatorname{tr}(W_{{\mathrm{FD}}}(v)) {\lvertv \cdot n(x)\rvert} {\,\mathrm{d}}v.$$ In the following, we introduce an efficient and accurate numerical scheme for the Boltzmann equation . Fast spectral method for the collision operator {#sec:collision} =============================================== Calculating the collision operator is the computationally most demanding step in solving . Here, we first represent the collision operator using Pauli matrices, and then develop a fast Fourier spectral method inspired by the ideas in [@MouhotPareschi:06; @FilbertHuJin:12; @HuYing:12]. Representation of the collision operator using Pauli matrices ------------------------------------------------------------- Since the spin-density matrix $W(t, x, v)$ is Hermitian, it can be represented in the basis of the identity matrix and the Pauli matrices: $$\label{eq:spin_repr} W_i = W(v_i) = w_{i, 0} \mathbbm{1} + \sum_{j = 1}^3 w_{i, j} \sigma_{j},$$ where the subscript $i$ specifies the velocity dependence, and we have suppressed the dependence on $(t, x)$ for concise notation. Moreover, we define the vector of components as $$w_i = (w_{i,0}, w_{i,1}, w_{i,2}, w_{i,3}) \in \mathbb{R}^4,$$ and introduce the notation ${\boldsymbol{\sigma}} = (\mathbbm{1}, \vec{\sigma})$ so that $$w_i \cdot {\boldsymbol{\sigma}} = w_{i, 0} \mathbbm{1} + \sum_{j=1}^3 w_{i, j} \sigma_j.$$ The $3$-vector part $\vec{w}_i = (w_{i,1}, w_{i,2}, w_{i,3}) \in \mathbb{R}^3$ is exactly the Bloch vector of $W_i$ (up to normalization), and the eigenvalues of $W_i$ are $w_{i,0} \pm {\lvert\vec{w}_i\rvert}$. We will also use the $4 \times 4$ “metric tensor” $\eta = \mathrm{diag}(1, -1, -1, -1)$ and set $$\langle w_i, w_j \rangle_{\eta} = w_i^T \eta\, w_j.$$ Since the eigenvalues of $W_i$ are in the interval $[0,1]$, one can verify that likewise $\langle w_i, w_j \rangle_{\eta} \in [0,1]$. Using the interchangeability of $v_3 \leftrightarrow v_4$ in the integral, the gain term of the dissipative collision operator can be written as $$- W_4 {\tilde{W}}_2 W_3 + W_4 \operatorname{tr}[{\tilde{W}}_2 W_3] + \mathrm{h.c.} \equiv 2\, \langle w_3, w_4 \rangle_{\eta} \big( \mathbbm{1} - (\eta\, w_2)\cdot{\boldsymbol{\sigma}} \big),$$ and the loss term as $$\begin{gathered} - \bigl({\tilde{W}}_4 W_3 - {\tilde{W}}_4 W_2 - {\tilde{W}}_2 W_3 + {\tilde{W}}_4 \operatorname{tr}[W_2] - {\tilde{W}}_4 \operatorname{tr}[W_3] + \operatorname{tr}[W_3 {\tilde{W}}_2]\bigr) W_1 + \mathrm{h.c.}\\ \equiv - \big( \langle w_3, w_4 \rangle_{\eta} \mathbbm{1} - (w_{3,0} + w_{4,0} - 1) \, (\eta\, w_2)\cdot{\boldsymbol{\sigma}} \big) W_1 + \mathrm{h.c.}\end{gathered}$$ For the Hermitian conjugate, one requires the anti-commutator of two Wigner matrices, which reads in the Pauli representation $$\bigl\{ W_1, W_2 \bigr\} \equiv W_1 W_2 + W_2 W_1 = 2\, \big( (w_{1,0} w_2 + w_{2,0} w_1) \cdot {\boldsymbol{\sigma}} - \langle w_1, w_2 \rangle_{\eta} \mathbbm{1} \big).$$ Again using the interchangeability of $v_3 \leftrightarrow v_4$ in the integral, the integrand in Eq.  of the conservative collision operator becomes $$\begin{gathered} (W_3 W_4 - W_2 W_3 - W_3 W_2 - \operatorname{tr}[W_4] W_3 + \operatorname{tr}[W_2] W_3 + W_2) \\ \equiv \big( \langle w_2, w_2 \rangle_\eta - \langle w_3 - w_2, w_4 - w_2 \rangle_\eta \big) \mathbbm{1} - (w_{3,0} + w_{4,0} - 1) W_2.\end{gathered}$$ Since the identity matrix does not contribute to the commutator in Eq. , it suffices to keep the second term $- (w_{3,0} + w_{4,0} - 1) W_2$ only. The commutator in Eq.  reads in the Pauli matrix representation $$-i\,\bigl[W_2, W_1\bigr] = 2\, (\vec{w}_2 \times \vec{w}_1) \cdot \vec{\sigma}.$$ To summarize, we have obtained the representation $$\label{eq:CcPauli} \mathcal{C}_{\mathrm{c}}[W]_1 = i \int {\,\mathrm{d}}v_2 {\,\mathrm{d}}v_3 {\,\mathrm{d}}v_4 \, \delta(\underline{v}) \mathcal{P}( 1 / \underline{\omega} ) (w_{3, 0} + w_{4, 0} - 1) \bigl[W_2, W_1\bigr],$$ and similarly $$\begin{gathered} \label{eq:CdPauli} \mathcal{C}_{\mathrm{d}}[W]_1 = \pi \int {\,\mathrm{d}}v_2 {\,\mathrm{d}}v_3 {\,\mathrm{d}}v_4 \, \delta(\underline{v}) \delta(\underline{\omega}) \Bigl( 2 {\left\langlew_3, w_4\right\rangle}_{\eta} \bigl(\mathbbm{1} - W_1 - (\eta w_2) \cdot {\boldsymbol{\sigma}}\bigr) \\ + (w_{3,0} + w_{4,0} - 1) \bigl\{ W_1, ( \eta w_2) \cdot {\boldsymbol{\sigma}} \bigr\} \Bigr).\end{gathered}$$ We remark that it is also possible to write the matrix-valued Boltzmann equation as a kinetic equation with multiple components, if we regard each matrix entry as a component. However, we prefer the more natural and physical representation in terms of Pauli matrices. Fast Fourier spectral method {#sec:FourierCollision} ---------------------------- To efficiently evaluate the collision terms and , we generalize the ideas in [@MouhotPareschi:06; @FilbertHuJin:12; @HuYing:12] for a Fourier spectral discretization of the velocity space. Let us discuss the conservative part first. Using the Carleman representation [@Carleman:57; @Wennberg:94], we perform a change of variables $v_1 \mapsto v, v_3 \mapsto v + u, v_4 \mapsto v + u'$, such that $$\begin{aligned} & v_2 = v_3 + v_4 - v_1 = v + u + u'; \\ & \underline{\omega} = \omega(v_1) + \omega(v_2) - \omega(v_3) - \omega(v_4) = u \cdot u'. \end{aligned}$$ Substituting into , we arrive at $$\label{eq:CcPauliCarleman} \mathcal{C}_{\mathrm{c}}[W](v) = i \int_{B_R} \int_{B_R} {\,\mathrm{d}}u {\,\mathrm{d}}u' \mathcal{P}( 1 / (u \cdot u')) (2 w_0(v+u) - 1) \bigl(W(v+u+u')W(v) - \text{h.c.}\bigr),$$ where we have used the symmetry between $u$ and $u'$. Here $R$ indicates the truncation of the collision integral, taken so that $B_R$ approximately covers the support of $W$ in the $v$ variable. Hence, we just need to deal with integrals of the kind $$I_1(v) = \int_{B_R} \int_{B_R} {\,\mathrm{d}}u {\,\mathrm{d}}u' \mathcal{P}(1 / (u \cdot u')) f(v+u) g(v + u + u') h(v).$$ To apply the Fourier method, we periodize the functions $f$, $g$ and $h$ on the domain $[-L, L]^2$ with $L \geq \frac{3 + \sqrt{2}}{2} R$, and define the Fourier grid $$\label{eq:fourier_grid} \Xi = [-N/2, -N/2 + 1, \ldots, N/2 - 1]^2.$$ Here the cut-off frequency $N$ controls the accuracy. Using the Fourier inversion formula, we approximate $$f(v) \approx \sum_{\xi \in \Xi} {\widehat{f}}(\xi) \exp( i \pi \xi \cdot v / L) \quad \text{and} \quad {\widehat{f}}(\xi) = \slashint_{[-L, L]^2} f(v) \mathrm{e}^{- i \frac{\pi}{L} \xi \cdot v} {\,\mathrm{d}}v$$ Then $$\begin{split} I_1(v) & = \sum_{\chi, \eta, \zeta} \int_{B_R} \int_{B_R} {\,\mathrm{d}}u {\,\mathrm{d}}u' \mathcal{P}(1 / (u \cdot u')) {\widehat{f}}(\chi) {\widehat{g}}(\eta) {\widehat{h}}(\zeta) \mathrm{e}^{i \frac{\pi}{L} v \cdot (\chi + \eta + \zeta)} \mathrm{e}^{i \frac{\pi}{L} u \cdot ( \chi + \eta)} \mathrm{e}^{i \frac{\pi}{L} u' \cdot \eta} \\ & = \sum_{\chi, \eta, \zeta} {\widehat{f}}(\chi) {\widehat{g}}(\eta) {\widehat{h}}(\zeta) \mathrm{e}^{i \frac{\pi}{L} v \cdot (\chi + \eta + \zeta)} G(\chi+\eta, \eta) \end{split}$$ where the matrix $G(\xi, \chi)$ is defined as $$G(\xi, \chi) = \int_{B_R} {\,\mathrm{d}}u \int_{B_R} {\,\mathrm{d}}u'\, \mathcal{P}(1 / (u\cdot u')) \exp(i \pi \xi \cdot u / L) \exp(i \pi \chi \cdot u' / L).$$ Changing to polar coordinates, one obtains $$G(\xi, \chi) = \int_0^R {\,\mathrm{d}}r \int_0^R {\,\mathrm{d}}r' \int_{S^1} {\,\mathrm{d}}\theta \int_{S^1} {\,\mathrm{d}}\theta' \, \mathcal{P}\bigl(1 / (\theta \cdot \theta')\bigr) \exp\bigl(i \pi \xi \cdot \theta r/ L \bigr) \exp\bigl(i \pi \chi \cdot \theta' r'/ L \bigr).$$ Since $\mathcal{P}(1/(\theta \cdot \theta'))$ is odd in both $\theta$ and $\theta'$, it suffices to take the odd part of the complex exponentials in the above integral, and we get $$G(\xi, \chi) = - \int_0^R {\,\mathrm{d}}r \int_0^R {\,\mathrm{d}}r' \int_{S^1} {\,\mathrm{d}}\theta \int_{S^1} {\,\mathrm{d}}\theta' \, \mathcal{P}\bigl(1 / (\theta \cdot \theta')\bigr) \sin\bigl(\pi \xi \cdot \theta r/ L \bigr) \sin\bigl(\pi \chi \cdot \theta' r' / L \bigr).$$ Note that $$\phi_R(\xi \cdot \theta) = \int_0^R {\,\mathrm{d}}r \sin\bigl(\pi \xi \cdot \theta r / L \bigr) = \frac{L}{\pi (\xi \cdot \theta)} \Bigl[ \cos\bigl(\pi \xi \cdot \theta R / L \bigr) - 1 \Bigr] = - \frac{2L}{\pi(\xi \cdot \theta)} \sin^2\bigl(\pi \xi \cdot \theta R / (2L)\bigr).$$ Hence, $$G(\xi, \chi) = - \int_{S^1} {\,\mathrm{d}}\theta \int_{S^1} {\,\mathrm{d}}\theta' \mathcal{P}\bigl(1 / (\theta \cdot \theta')\bigr) \phi_R(\xi \cdot \theta) \phi_R(\chi \cdot \theta').$$ We approximate $G$ using numerical quadrature with a trapezoidal rule. To deal with the singularity in the principal value integral, we take the grids of $\theta$ and $\theta'$ to be $$\label{eq:theta_grid} \theta_j = \exp(i(j-1) \pi / J), \quad \text{and} \quad \theta_j' = \exp(i(j-1/2) \pi / J), \qquad j = 1, \ldots, J$$ for some positive integer $J$. The quadrature rule converges exponentially [@TrefethenWeideman:13]\*[Section 6]{}. Due to symmetry, only quadrature points on half circles are required. We arrive at the final approximation $$\label{eq:approxG} G(\xi, \chi) \approx \sum_{j = 1}^{J} \sum_{j' = 1}^J \omega_{G, j, j'} \phi_{R, j}(\xi) \phi'_{R, j'}(\chi)$$ where the weights are given by $$\omega_{G, j, j'} = - \Bigl( \frac{2 \pi}{J}\Bigr)^2 \frac{1}{\theta_j \cdot \theta'_{j'}}$$ and we have used the shorthand notation $$\phi_{R, j}(\xi) = \phi_R(\xi \cdot \theta_j) \quad \text{and} \quad \phi'_{R, j'}(\chi) = \phi_R(\chi \cdot \theta'_{j'}).$$ In summary, we have obtained the approximation $$\begin{split} {\widehat{I}}_1(\xi) & \approx \sum_{\substack{\chi, \eta, \zeta, \\ \chi + \eta + \zeta = \xi}} \sum_{j, j'} \omega_{G, j, j'}\, {\widehat{f}}(\chi)\,{\widehat{g}}(\eta)\,{\widehat{h}}(\zeta)\,\phi_{R, j}(\chi + \eta)\, \phi'_{R, j'}(\eta) \\ & = \sum_{j, j'} \omega_{G, j, j'} \sum_{\zeta} \Bigl[ \sum_{\eta} {\widehat{f}}(\xi - \zeta - \eta)\, \bigl(\phi'_{R, j'}(\eta)\, {\widehat{g}}(\eta) \bigr) \Bigr] \phi_{R, j}(\xi - \zeta) \, {\widehat{h}}(\zeta). \end{split}$$ For each $j$ and $j'$, we first calculate the product $\phi'_{R, j'} {\widehat{g}}$ (complexity ${\mathcal{O}}(JN^2)$); the summation over $\eta$ is a convolution by FFT (${\mathcal{O}}(JN^2 \log N)$); we then multiply the result pointwise with $\phi_{R, j}$ (${\mathcal{O}}(J^2N^2)$). The summation over $\zeta$ is another convolution (${\mathcal{O}}(J^2N^2 \log N)$). The total complexity is thus ${\mathcal{O}}(J^2 N^2 \log N)$. We use double zero padding in the Fourier coefficients to avoid aliasing. The dissipative part in Carleman representation reads $$\begin{gathered} \mathcal{C}_{\mathrm{d}}[W](v) = \pi \int_{B_R} \int_{B_R} {\,\mathrm{d}}u {\,\mathrm{d}}u' \, \delta(u \cdot u') \Bigl( 2 {\left\langlew(v+u), w(v+u')\right\rangle}_{\eta} \bigl(\mathbbm{1} - W(v) - (\eta w(v+u+u')) \cdot {\boldsymbol{\sigma}}\bigr) \\ + (2 w_{0}(u+v) - 1) \bigl\{ W(v), ( \eta w(v+u+u')) \cdot {\boldsymbol{\sigma}} \bigr\} \Bigr), \end{gathered}$$ which follows from by the same change of variables leading to . Expanding the above expression, it is straightforward to check that it consists of the following three kinds of integrals: $$\begin{aligned} & I_2(v) = \int_{B_R} \int_{B_R} {\,\mathrm{d}}u {\,\mathrm{d}}u' \delta(u \cdot u') f(v+u) g(v+u') h(v); \\ & I_3(v) = \int_{B_R} \int_{B_R} {\,\mathrm{d}}u {\,\mathrm{d}}u' \delta(u \cdot u') f(v+u) g(v+u') h(v + u + u'); \\ & I_4(v) = \int_{B_R} \int_{B_R} {\,\mathrm{d}}u {\,\mathrm{d}}u' \delta(u \cdot u') f(v+u) g(v+u + u') h(v),\end{aligned}$$ where $f, g, h$ stand for certain components of $w$. To evaluate these integrals, we define $$H(\xi, \chi) = \int_{B_R} {\,\mathrm{d}}u \int_{B_R} {\,\mathrm{d}}u'\, \delta(u\cdot u') \exp(i \pi \xi \cdot u / L) \exp(i \pi \chi \cdot u' / L).$$ By similar steps as leading to , one obtains $$\label{eq:approxH-J} H(\xi, \chi) \approx \sum_{j = 1}^{J} \omega_{H, j}\, \psi_{R, j}(\xi)\, \psi'_{R, j}(\chi).$$ Here $$\omega_{H, j} = \frac{\pi}{J}, \quad \psi_{R, j}(\xi) = \frac{2L}{\pi \xi \cdot \theta_j} \sin\Bigl( \frac{\pi R}{L} \xi \cdot \theta_j \Bigr) \quad \text{and} \quad \psi'_{R, j}(\chi) = \frac{2L}{\pi \xi \cdot \theta_j} \sin\Bigl( \frac{\pi R}{L} \xi \cdot \mathcal{R}_{\pi/2}\theta_j \Bigr),$$ where $\mathcal{R}_{\pi/2}$ is the rotation by $\pi/2$. Then the Fourier representation of $I_2$ is $$\begin{split} {\widehat{I}}_2(\xi) & = \sum_{\substack{\chi, \eta, \zeta, \\ \chi + \eta + \zeta = \xi}} \sum_{j} \omega_{G,j} {\widehat{f}}(\chi)\, {\widehat{g}}(\eta)\, {\widehat{h}}(\zeta)\, \psi_{R, j}(\chi)\, {\tilde{\psi}}_{R, j}(\eta) \\ & = \sum_j \omega_{G, j} \sum_{\zeta} \sum_{\chi + \eta = \xi - \zeta} \bigl(\psi_{R, j}(\chi)\, {\widehat{f}}(\chi) \bigr) \bigl( {\tilde{\psi}}_{R, j}(\eta)\, {\widehat{g}}(\eta) \bigr)\, {\widehat{h}}(\zeta). \end{split}$$ For each $j$, the pointwise products in the brackets are computed (complexity ${\mathcal{O}}(JN^2)$); for each $j$ and $\xi$, summation over $\chi$, $\eta$ and $\zeta$ is a double convolution by FFT $({\mathcal{O}}(JN^2 \log N))$. The total complexity is $({\mathcal{O}}(J N^2 \log N))$. The integral type $I_4$ is similar to $I_1$, and hence we will omit the details. For $I_3$, we need another representation of $H(\xi, \chi)$, as suggested by Hu and Ying in [@HuYing:12]: $$\label{eq:approxH-JM} \begin{aligned} H(\xi, \chi) & \approx \frac{\pi}{J} \sum_{m=1}^M w_{R, m} \sum_{j=1}^J \Bigl[ \exp( i \pi \rho_{R, m} \xi \cdot \theta_j / L)\, \psi_{R, j}'(\chi) \Bigr], \\ & =: \sum_{m=1}^M \sum_{j=1}^J {\tilde{\omega}}_{H, j, m} \Bigl[ \exp( i \alpha_{j, m} \cdot \xi)\, \psi_{R, j}'(\chi) \Bigr] \end{aligned}$$ with $${\tilde{\omega}}_{H, j, m} = \frac{\pi}{J} w_{R, m} \quad \text{and} \quad \alpha_{j, m} = \pi \rho_{R, m} \theta_j / L.$$ In the above, $(\rho_{R, m}, w_{R, m})$ are the nodes and weights of a Gauss-Legendre quadrature on $[-R, R]$. Then, the integral $I_3$ can be calculated as $$\begin{aligned} {\widehat{I}}_3(\xi) & = \sum_{j, m} {\tilde{\omega}}_{H, j, m} \sum_{\chi, \eta, \zeta: \, \chi + \eta + \zeta = \xi} {\widehat{f}}(\chi)\, {\widehat{g}}(\eta)\, {\widehat{h}}(\zeta)\, \exp(i \alpha_{j, m} \cdot \chi)\, \exp(i \alpha_{j, m} \cdot \zeta)\, \psi'_{R, j}(\eta + \zeta) \\ & = \sum_{j, m} {\tilde{\omega}}_{H, j, m} \sum_{\chi, \eta, \zeta: \, \chi + \eta + \zeta = \xi} \bigl(\exp(i \alpha_{j, m} \cdot \chi) {\widehat{f}}(\chi)\bigr)\, {\widehat{g}}(\eta)\, \psi'_{R, j}(\xi - \chi)\, \bigl(\exp(i \alpha_{j, m} \cdot \zeta)\, {\widehat{h}}(\zeta) \bigr) \\ & = \sum_{j, m} {\tilde{\omega}}_{H, j, m} \sum_{\chi} \bigl(\exp(i \alpha_{j, m} \cdot \chi)\, {\widehat{f}}(\chi)\bigr)\, \psi'_{R, j}(\xi - \chi) \sum_{\eta, \zeta: \, \eta + \zeta = \xi - \chi} {\widehat{g}}(\eta)\, \bigl(\exp(i \alpha_{j, m} \cdot \zeta)\, {\widehat{h}}(\zeta) \bigr). \end{aligned}$$ We first calculate the pointwise products in the bracket for each $j$ and $m$ (${\mathcal{O}}(JMN^2)$); summation over $\chi$, $\eta$ and $\zeta$ is a double convolution $({\mathcal{O}}(JMN^2 \log N))$. Since $M$ is ${\mathcal{O}}(N)$, the total complexity is $({\mathcal{O}}(JN^3 \log N))$. We remark that while the above discussion is limited to the case of two dimensional velocity space, it is possible to extend the method to $3D$ by generalizing the method in [@HuYing:12] to our case. We will leave this to future works. The numerical results in this work are limited to one space dimension and two velocity dimensions. Time splitting algorithm and parallelization {#sec:parallel} ============================================ The numerical discretization of Eq.  is based on a time splitting algorithm to deal with the convection, collision and external magnetic terms separately. Specifically, we perform half a time step of convection in physical velocity space, then transform the Wigner state to velocity Fourier space using FFT for the collision integrals as discussed in Section \[sec:FourierCollision\] and for applying the external magnetic field, and finally switch back to physical velocity space for another half time step of convection. For the convection we use the finite volume method with minmod slope limiter [@LeVeque:92]\*[Ch. 16]{}; other slope limiters can also be applied. For completeness, we briefly recall the formulation with some comments on parallelization of the algorithm. In the advection step, each discrete velocity $v$ can be treated independently due to the time splitting. The minmod slope limiter method updates the solution as $$\label{eq:slopeLimiterStep} \begin{split} W_j^{n+1} &= W_j^n + \frac{\Delta t}{2\Delta x} \Bigg[ -v_x \big(W_{j+1}^n - W_{j-1}^n\big) + {\lvertv_x\rvert}\big(W_{j+1}^n - 2 W_j^n + W_{j-1}^n \big) \\ &\hspace{8em} + \tfrac{1}{2}\left(\mathrm{sgn}(\hat{\nu}) - \hat{\nu}\right) \left( -v_x \big(S_{j+1}^n - S_{j-1}^n \big) + {\lvertv_x\rvert}\big( S_{j+1}^n - 2 S_j^n + S_{j-1}^n \big) \right) \Bigg]. \end{split}$$ where $\Delta x$ denotes the spatial mesh width, $\Delta t$ the time step, $n$ the discretized time index, $j$ the finite volume cell index, and $\hat{\nu} = \Delta t\, v_x / \Delta x$. We use the shorthand notations $$W_j^n = W(n\,\Delta t, j\,\Delta x, v) \quad \text{and} \quad S_j^n = \operatorname{minmod}\big(W_{j+1}^n - W_j^n, W_j^n - W_{j-1}^n\big)$$ with the minmod function defined as $$\label{eq:minmod} \operatorname{minmod}(a,b) = \begin{cases} a & \mathrm{if}\ \ {\lverta\rvert} \le {\lvertb\rvert} ~~\text{and}~~ a\,b > 0; \\ b & \mathrm{if}\ \ {\lvertb\rvert} < {\lverta\rvert} ~~\text{and}~~ a\,b > 0; \\ 0 & \mathrm{if}\ \ a\,b < 0. \end{cases}$$ The first line in Eq.  is precisely Godunov’s method, and the second line originates from the additional slope limiter terms. As mentioned above, for the time splitting algorithm we actually perform two transport steps with $\Delta t/2$. Note that the Wigner state at the next time step depends on *two* neighbors on either side, i.e., on the five finite volumes with indices $j-2, \dots, j+2$. Concerning parallelization, each computing node handles a few adjacent finite volumes, that is, we parallelize the computation along the spatial $x$ dimension. This straightforward approach takes advantage of the locality of the computationally demanding collision step, which is independent of the neighboring finite volumes. Since the transport step depends on two neighbors on each side, every computing node handles at least two adjacent finite volumes to minimize inter-process communication. In our custom C implementation, we use MPI to transfer neighboring states during the transport step. Numerical examples {#sec:numerics} ================== Validation of the algorithm --------------------------- We first present several numerical tests to validate our algorithm. Let us start with the approximation of the kernels $G$ in and $H$ in and , which depend on the choice of the number $2 J$ of points on the unit circle in , and the total number $M$ of Gauss-Legendre quadrature nodes on $[-R, R]$ for the radial quadrature in . To test the dependence on these parameters, we calculate the dissipative and conservative collision kernels $\mathcal{C}_{\mathrm{d}}[W]$ and $\mathcal{C}_{\mathrm{c}}[W]$ for a fixed spin density matrix $W(v)$. Here, $W(v)$ is (somewhat arbitrarily) chosen as $$\begin{gathered} \label{eq:validationW} W(v) = \frac{1}{3\pi} (1 + v_x) \mathrm{e}^{-\tfrac{1}{2}(v_x - 1/2)^2 - \tfrac{1}{2}(v_y + v_x)^2} \mathbbm{1} + \frac{3\sqrt{6/5}}{55\pi} \left(1 - v_x - \tfrac{v_y}{6}\right)^2 \mathrm{e}^{-\tfrac{1}{12}(2 v_x + v_y)^2 - \tfrac{1}{10} v_y^2} \sigma_1 \\ + \frac{1}{4\pi \sqrt{2}} \mathrm{e}^{-\tfrac{1}{8} (v_x-1)^2 - \tfrac{1}{4} v_y^2} \sigma_2 + \frac{9}{56\pi} \left(\tfrac{1}{3} + v_x\right)^2 \mathrm{e}^{-\tfrac{1}{8} (v_x-v_y)^2 - \tfrac{1}{8} (v_x + 2 v_y)^2} \sigma_3 \ .\end{gathered}$$ The Fourier representation determines the discretization in physical velocity space, i.e., $v \in (2 L/N)\,\Xi$ with $N$ the number of Fourier grid points in each dimension, $\Xi$ the corresponding grid defined in and $L$ the domain size. First we change $J$ while keeping the other parameters fixed, and compare the result to a reference calculation with large $J = 72$. The relative error is calculated as $\sum_{j=0}^3 \lVert\mathcal{C}_{\mathrm{d}}[W]_j - \mathcal{C}_{\mathrm{d},\mathrm{ref}}[W]_j\rVert / \lVert \mathcal{C}_{\mathrm{d},\mathrm{ref}}[W]_j\rVert$ (and correspondingly for $\mathcal{C}_{\mathrm{c}}$), where $j$ denotes the Pauli matrix component as in and $\lVert \cdot \rVert$ is the $L^1$-norm of the representation as $N \times N$ matrices in Fourier grid space. One observes in Figure \[fig:conv-J\] that the relative error decreases exponentially fast as $J$ increases, confirming the previous discussion of exponential convergence with respect to $J$. Figure \[fig:conv-J\] also illustrates that $J = 32$ already achieves relative errors smaller than $10^{-8}$ (note that $J=32$ amounts to $64$ grid points in $\theta$ due to symmetry). We will fix this choice in the sequel. Next, we study the accuracy of the approximation for different choices of $M$. As discussed before, the integrand becomes more oscillatory as $N$ increases, and thus we expect that the number of quadrature nodes $M$ depends linearly on $N$. The relative error of the dissipative $\mathcal{C}_{\mathrm{d}}[W]$ collision term also seems to decay exponentially with $M$ before reaching the machine accuracy (see Figure \[fig:conv-M\]), thanks to the accuracy of the Gauss-Legendre quadrature. In our subsequent computations we fix $M = 32$, which is adequate to achieve $10^{-5}$ relative error for $N = 32$. Finally, the exponential convergence with respect to $N$ in Figure \[fig:conv-N\] verifies that our method achieves spectral accuracy in dealing with the collision operators. A large $M = 72$ is used in Figure \[fig:conv-N\] to ensure the accuracy of the approximation for each $N$. Spatially homogeneous equation ------------------------------ We now study the time evolution under the spatially *homogeneous* equation. Here we choose the initial condition to be a Fermi-Dirac state perturbed by $v$-dependent rotations: $$\label{eq:W0homogeneous} W(0,v) = \mathrm{e}^{-i X(v)} \cdot U_0 \cdot \mathrm{diag}\big( \mathrm{e}^{\left(\frac{1}{2} {\lvertv - v_0\rvert}^2 - \mu_s\right)/(k_{\mathrm{B}} T)} + 1 \big)^{-1}_s \cdot U_0^* \cdot \mathrm{e}^{i X(v)}$$ with $$U_0 = \begin{pmatrix} \cos(\pi/5) & -i \sin(\pi/5) \\ \sin(\pi/5) & i \cos(\pi/5) \end{pmatrix}, \quad X(v) = \sin\big(v_y^2\big)\,\mathbbm{1} + (v_x-1)\sigma_1 + 2 \big(v_x^2 + {\lvertv_y\rvert} + 1\big)^{-1} \sigma_2 + \cos\big(v_x + \tfrac{1}{2}v_y\big)\,\sigma_3$$ and the parameters $k_{\mathrm{B}} T = 5/4$, $\mu_{\uparrow} = 1$, $\mu_{\downarrow} = 3/2$ and $v_0 = (0.4, -0.1)$. In Figure \[fig:hom\_conservation\], the conserved quantities spin density , momentum and energy are plotted as a function of time. We use the matrix $L^2$-norm for the relative error of the $2 \times 2$ spin density $\rho(t)$. Excellent conservation is observed numerically (note that the scale of the y-axes are $10^{-12}$, $10^{-4}$ and $10^{-5}$ respectively). We have used the time step $\Delta t = 0.001$ for the simulation. ![Time-dependent range of the eigenvalues of the Wigner states, remaining between $0$ and $1$, as required (same simulation as in Figure \[fig:hom\_conservation\]).[]{data-label="fig:eigenvalue_range"}](eig_range){width="35.00000%"} As another consistency check, the eigenvalues of the Wigner spin-density matrices must stay between $0$ and $1$. This condition is satisfied by our numerical scheme: Figure \[fig:eigenvalue\_range\] visualizes the largest and smallest (with respect to $v$) of all eigenvalues. In accordance with the H-theorem, the entropy is monotonically increasing (see Figure \[fig:entropy\]). Physically, as $t \to \infty$ the Wigner state should converge to a thermal equilibrium Fermi-Dirac distribution $W_{{\mathrm{FD}}}$ with moments matching the conserved moments of the Wigner state (see Figure \[fig:hom\_conservation\]). The average velocity $u$ leads to a shift $W_{{\mathrm{FD}}}(v) \to W_{{\mathrm{FD}}}(v - u)$ in , and the eigenbasis of $W_{{\mathrm{FD}}}$ is equal to the eigenbasis of the spin density matrix $\rho$. We fit the temperature $T$ and chemical potentials $\mu_{\uparrow}$, $\mu_{\downarrow}$ in numerically to match the eigenvalues of $\rho$ and the energy. The expected convergence to $W_{{\mathrm{FD}}}$ is verified in Figure \[fig:relentropy\], showing the quantum relative entropy between the Wigner state and $W_{{\mathrm{FD}}}$. The quantum relative entropy is given by (recall that ${\tilde{W}} = \mathbbm{1} - W$) $$\label{eq:relentropy_hom} S(W \parallel W_{{\mathrm{FD}}}) = \int \operatorname{tr}\left[ W(v) \bigl(\log W(v) - \log W_{{\mathrm{FD}}}(v)\bigr) + {\tilde{W}}(v) \bigl(\log {\tilde{W}}(v) - \log {\tilde{W}}_{{\mathrm{FD}}}(v) \bigr) \right]{\,\mathrm{d}}v.$$ Using the H-theorem for the entropy and the conservation property of the fluid dynamic moments, the relative entropy is monotonically decaying to zero as $t \to \infty$. Finally, we demonstrate exponential convergence to $W_{{\mathrm{FD}}}$ in $L^1$-norm in Figure \[fig:convergenceFD\]. Spatially inhomogeneous equation with periodic boundary conditions ------------------------------------------------------------------ For *periodic* boundary conditions, the spin density, momentum, and energy are still conserved globally, i.e., after taking the integral over the spatial dimension. Figure \[fig:inhom\_periodic\_conservation\] shows the conservation in the numerical scheme, analogous to the spatially homogeneous case. The initial Wigner state is similar to but with $T$, $\mu_s$ and $v_0$ depending on the spatial location $x$. The quantum relative entropy now involves integration over the spatial domain: $$\label{eq:relentropy_inhom} S\big(W(t,\cdot,\cdot) \parallel W_{{\mathrm{FD}}}\big) = \int S\big(W(t,x,\cdot) \parallel W_{{\mathrm{FD}}}\big) {\,\mathrm{d}}x.$$ Note that the asymptotic Fermi-Dirac equilibrium state is independent of $t$ and $x$. For comparison, we define a quantum relative entropy with respect to locally fitted (at each $x$ and $t$) Fermi-Dirac states: $$\label{eq:relentropy_local} S_{\mathrm{loc}}\big(W(t,\cdot,\cdot) \parallel W_{{\mathrm{FD}},\mathrm{loc}}(t,\cdot,\cdot)\big) = \int S\big(W(t,x,\cdot) \parallel W_{{\mathrm{FD}},\mathrm{loc}}(t,x,\cdot)\big) {\,\mathrm{d}}x.$$ Here $W_{{\mathrm{FD}},\mathrm{loc}}(t,x,\cdot)$ is the Fermi-Dirac state with the same spin density, momentum and energy as $W(t,x,\cdot)$. Figure \[fig:inhom\_periodic\_entropy\] visualizes the monotonically increasing entropy, and Figure \[fig:inhom\_periodic\_relentropy\] the quantum relative entropy, both global and local (for the same simulation as in Figure \[fig:inhom\_periodic\_conservation\]). It is straightforward to verify that $$S\big(W(t) \parallel W_{{\mathrm{FD}}}\big) = S\big(W(t) \parallel W_{{\mathrm{FD}}, \mathrm{loc}}(t)\big) + S\big(W_{{\mathrm{FD}}, \mathrm{loc}}(t) \parallel W_{{\mathrm{FD}}}\big).$$ According to Figure \[fig:inhom\_periodic\_relentropy\], the quantum relative entropy is dominated by $S\big(W_{{\mathrm{FD}}, \mathrm{loc}}(t) \parallel W_{{\mathrm{FD}}}\big)$. In other words, the system quickly relaxes to a local Fermi-Dirac state, before converging to the global equilibrium. As next step, we investigate the effect of an external, $x$-dependent magnetic field, which enters the Boltzmann equation as the last term in Eq. . Specifically, we choose $$\label{eq:Bext_rot} \vec{B}(x) = (0, \cos(2\pi x), \sin(2\pi x))^T$$ for the simulation. Figure \[fig:density\_Bext\_rot\] visualizes the external magnetic field and shows the components of the spin density matrix at $t = 1$, with the Bloch vector part (scaled by $\frac{1}{2}$) defined as $\vec{\rho} = (\rho_1, \rho_2, \rho_3)$. Compared to the above simulation without magnetic field (density not shown), the Bloch vector components of $\rho(t,x)$ now change with $x$. Since the magnetic field acts as a unitary rotation of the Wigner state in the time evolution, the trace $\operatorname{tr}[W(t,x,v)]$ remains unaffected by the magnetic field and the momentum and energy conservation laws still hold. Spatially inhomogeneous equation with Dirichlet and Maxwell boundary conditions ------------------------------------------------------------------------------- First, we investigate a simulation with Dirichlet boundary conditions. The fixed states at the left and right boundary are Fermi-Dirac states with different temperatures and eigenbasis, as summarized in Table \[tab:Dirichlet\]. left right ------------------------ --------------- --------------------------------------------------------------------------------------- $1/(k_{\mathrm{B}} T)$ 0.8 1.2 $\mu_{\uparrow}$ 1.5 1.5 $\mu_{\downarrow}$ -1.5 -1.5 $U$ $\mathbbm{1}$ $\frac{1}{\sqrt{2}} \left(\begin{smallmatrix}1 & -1 \\ 1 & 1\end{smallmatrix}\right)$ : Temperature $T$, chemical potentials $\mu_{\uparrow}$, $\mu_{\downarrow}$ and eigenbasis $U$ of the incoming left and right Fermi-Dirac boundary states for a simulation with Dirichlet boundary conditions.[]{data-label="tab:Dirichlet"} Here, $U$ contains the spin-eigenbasis of the Fermi-Dirac state as column vectors. In Figure \[fig:Dirichlet\] we visualize the stationary density (after running the simulation until reaching stationarity) as well as the local temperature and entropy. The temperature is estimated by constructing a local Fermi-Dirac state with the same moments as the actual local Wigner state. Note that the stationary temperature at the boundary is not exactly equal to the values in Table \[tab:Dirichlet\] since the Dirichlet boundary condition fixes only the incoming parts of the Wigner states at the left and right boundary. Finally, we explore the effects of Maxwell boundary states with (i) different temperature and chemical potentials but common spin eigenbasis, and (ii) same temperature and chemical potentials but spin eigenvectors pointing to different directions, as summarized in Table \[tab:Maxwell\]. -------------- ------ ------- ------ ------- ------ ------- --------------- --------------------------------------------------------------------------------------- left right left right left right left right Maxwell (i) 0.6 1.1 -0.4 1.3 1.8 -0.9 $\mathbbm{1}$ $\mathbbm{1}$ Maxwell (ii) 1 1 1.5 1.5 -1.5 -1.5 $\mathbbm{1}$ $\frac{1}{\sqrt{2}} \left(\begin{smallmatrix}1 & -1 \\ 1 & 1\end{smallmatrix}\right)$ -------------- ------ ------- ------ ------- ------ ------- --------------- --------------------------------------------------------------------------------------- : Temperature $T$, chemical potentials $\mu_{\uparrow}$, $\mu_{\downarrow}$ and eigenbasis $U$ of the left and right diffusive reflection Fermi-Dirac states for two simulations with Maxwell boundary conditions.[]{data-label="tab:Maxwell"} \ The corresponding stationary states (after running the simulation up to $t = 4$) are shown in Figure \[fig:Maxwell\]. Somewhat surprisingly, the stationary density for case (i) with standard spin eigenbasis on the left and right contains nonzero off-diagonal entries, see Figure \[fig:Maxwell1\_density\]. This might result from the interaction of different spin components in the collision operator, such that the local eigenbasis of the stationary state changes with spatial location. Nonzero off-diagonal entries in the stationary spin density are also observed for case (ii) in Figure \[fig:Maxwell2\_density\], but here this effect is certainly attributable to the rotated eigenbasis of the right Maxwell boundary condition. The stationary temperature for case (i) smoothly interpolates between the left and right boundary condition (see Figure \[fig:Maxwell1\_temperature\]), as one might expect. The temperature for case (ii) remains constant, in accordance with the same temperature on the left and right (see Figure \[fig:Maxwell2\_temperature\]). Finally, the local entropy tends to decrease with temperature in case (i) and remains largely unaffected by the rotation of the local eigenbasis in case (ii). Recall that the entropy is invariant under a change of eigenbasis. In summary, the local temperature and entropy conform with reasonable expectations, but the local spin eigenbasis of the density $\rho(t, x)$ could have hardly been predicted from the Maxwell boundary conditions. Conclusions and outlook {#sec:conclusion} ======================= We have developed an efficient numerical algorithm based on spectral Fourier discretization for the matrix-valued quantum Boltzmann equation. The effective Hamiltonian appears only in the matrix-valued version (since a commutator of scalars vanishes) and consists of a principal value integral lacking microscopic energy conservation; we have introduced a shift in the numerical grid points to treat the singular part of the principal value. The resulting algorithm exhibit spectral accuracy as numerically confirmed in Figure \[fig:conv-N\]. Our numerical simulations support the picture of fast convergence to local equilibrium and slower global equilibration, see Figure \[fig:inhom\_periodic\_relentropy\]. This suggests that future work on effective hydrodynamic equations derived from the matrix-valued quantum Boltzmann equation might be a promising endeavor. A multiscale algorithm coupling the many-body Hubbard model, the kinetic description of the weakly interacting Hubbard system, and the hydrodynamic limit of the model would also be an interesting future direction to explore.

ArXiv

--- abstract: 'We exhibit geometric conditions on a family of toric hypersurfaces under which the value of a canonical normal function at a point of maximal unipotent monodromy is irrational.' author: - Matt Kerr title: Motivic irrationality proofs --- Introduction {#S1} ============ The limit of a generalized normal function at a point where the underlying variation of Hodge structure degenerates, as recently studied in [@7K], turns out to have an unexpected arithmetic application. R. Apéry’s famous proof (see [@vdP]) of irrationality of $\zeta(3):=\sum_{k\geq1}k^{-3}$ relies on the existence of rapidly divergent sequences $a_{m}\in\ZZ$, $b_{m}\in\QQ$ (the latter having denominators of bounded growth) with $2a_{m}\zeta(3)+b_{m}$ converging rapidly to zero. Beukers, Peters and Stienstra [@Be; @BP; @Pe; @PS] geometrically repackaged much of the proof, noting for instance that the generating function $\sum_{m\geq0}a_{m}\lambda^{m}=:A(\lambda)$ records periods of a holomorphic 2-form on a family of $K3$ surfaces $\{X^{\lambda}\}_{\lambda\in\PP^{1}}$, hence must satisfy a Picard-Fuchs differential equation $D_{\text{PF}}A(\lambda)=0$. Behind the remaining details of the irrationality proof lurks a family of cycles in (algebraic) $K_{3}$ of the $K3$. The associated higher normal function $\tilde{V}(\lambda)$ has special value $\tilde{V}(0)=-2\zeta(3)$, and satisfies the inhomogeneous equation $D_{\text{PF}}\tilde{V}(\lambda)=Y(\lambda)$, where $Y$ denotes the Yukawa coupling. Setting $\sum_{m\geq1}b_{m}\lambda^{m}:=A(\lambda)\tilde{V}(0)-\tilde{V}(\lambda)$, one deduces from this recurrence relations on the $\{b_{m}\}$ which (as presented here) give “half” of the required bounded denominator growth. The other “half” comes from the *Fermi family* of $K3$s studied, but not related to the Apéry proof, in [@PS]. Finally, the behavior of the cycles at singular members of the family $\{X^{\lambda}\}$ shows that $\tilde{V}(\lambda)$ has no mondromy about the conifold singular fiber closest to $\lambda=0$, implying the rapid convergence of $2a_{m}\zeta(3)+b_{m}\to0$. In this paper, we reveal a general criterion for the irrationality of special values of certain higher normal functions. Given a Laurent polynomial $\phi(x_{1},\ldots,x_{n})$ with reflexive Newton polytope $\Delta$, the equation $\phi(\underline{x})=\lambda$ defines a family of Calabi-Yau hypersurfaces $\{X^{\lambda}\}$ in the toric variety $\PP_{\Delta}$. Associated to this family is a pure irreducible variation $\V_{\phi}$ of weight $(n-1)$ (over a Zariski open $\mathcal{U}\subset \PP^1$), together with a canonical section $\{\tilde{\omega}^{\lambda}\}$ of the Hodge line bundle $\F^{n-1}\V_{\phi}$. We call $\phi$ *tempered* if the coordinate symbol $\{x_{1},\ldots,x_{n}\}$ lifts to a family of motivic cohomology classes $\Xi^{\lambda}$ on the family, producing an extension $$\label{ext} 0\to \V_{\phi}(n)\to \mathcal{E}_{\phi} \to \QQ(0)_{\mathcal{U}}\to 0$$of admissible variations of mixed Hodge structure over $\mathcal{U}$. (This temperedness typically holds, for example, for LG-models constructed from Minkowski polynomials [@dS].) Applying a variant of this hypothesis allows us to construct a canonical truncated higher normal function $\tilde{V}(\lambda)$ on $\PP^{1}\backslash\phi(\RR_{-}^{\times n})$ by pairing the regulator class of $\Xi^{\lambda}$ (i.e., the extension class of ) with $\tilde{\omega}^{\lambda}$ (see Theorem \[th3\]). To arrange $\tilde{V}(0)\notin\QQ$, we must impose several additional conditions on $\phi$, roughly as follows (see Theorem \[th2\]): - the local system of periods of $\tilde{\omega}^{\lambda}$ must be of rank $n$, admit an isomorphism to its pullback by $\lambda\mapsto C/\lambda$, and have two mild singularities apart from $0$ and $\infty$, one of which is very far from $0$; - $\phi(-\underline{x})$ has positive integer coefficients, and the Picard-Fuchs operator associated to $\tilde{\omega}^{\lambda}$ (suitably normalized) is integral; and - a finite ($r:1$) pullback of the family $X^{\lambda}$ can be presented as a family of toric hypersurfaces in $\PP_{\diamond}$, where $\diamond$ is a “facile” polytope (Definition \[def1\]). The role of the last condition is to produce a basis of periods whose power-series coefficients have the right denominator bounds (see Corollary \[cor1\]). This basis is closely tied to mirror symmetry [@HLY] and the Frobenius method [@Yo]; for $n\geq4$, Theorem \[th1\] uncovers a surprising arithmetic implication of the Hyperplane Conjecture [@HLY; @LZ]. We also remark that, assuming only temperedness, the higher normal function $\tilde{V}(\lambda)$ can always be written as one of the chain-integral solutions of [@HLYZ], while $\tilde{V}(0)$ may be interpreted as an Apéry constant as studied in [@Ga; @Go; @GGI; @GI; @GZ]. In the last section, we exhibit Laurent polynomials which satisfy all these conditions for $n=1$, $2$, and $3$, recovering irrationality of $\log(1+b^{-1})$ ($b\in\mathbb{N}$), $\zeta(2)$, and $\zeta(3)$. We also propose relaxations of some of the conditions, together with specific families of polynomials, for making contact with results on linear forms in more than one odd zeta value – for instance [@Va], [@Z1], and especially [@Br], whose basic cellular integrals on $\mathcal{M}_{0,n+3}$ are the power series coefficients of a $\tilde{V}(\lambda)$ as above. While the results on $\zeta(2)$ and $\zeta(3)$ complete, in a way, the story begun in [@BP], the reader familiar with those works will notice (perhaps deplore?) the complete lack of reference to modular forms in what follows. The omission is strategic, as a weight-$(n-1)$ VHS $\mathcal{V}_{\phi}$ with maximal unipotent and conifold monodromies *cannot* have a modular parametrization for $n\geq4$. This is, of course, precisely where we hope to stimulate the search for examples with Theorem \[th2\], starting with the increasingly sophisticated databases of polytopes, local systems, Calabi-Yau differential operators and their geometric realizations [@CYclass; @Fano1; @Fano; @DM]. The same reader may be puzzled by our reference to “splitting up the bound” on denominators of $b_m$, as in the Apéry story one simply shows that $2(L_m)^3 b_m \in \ZZ$, where $L_m := \text{lcm}\{1,2,\ldots,m\}$. What we are able to show under *general* hypotheses, using techniques from Hodge theory and mirror symmetry, is instead that (for some fixed $\varepsilon\in \mathbb{N}$) both $\varepsilon (L_{rm})^n b_m$ and $\varepsilon (m!)^n b_m$ are integers (with $r=2$ for Apéry), which *together* are enough for an irrationality proof. With that said, the results of this article are intended to be a narrow proof of concept for a Hodge-theoretic approach to irrationality proofs, rather than to be optimal as respects either methodology or hypotheses. We freely (though infrequently) use the notation and terminology of regulator currents and toric varieties throughout, as reviewed in $\S\S$1-2 of [@DK]. It is my pleasure to thank S. Bloch, F. Brown, C. Doran, V. Golyshev, B. Lian, V. Maillot, F. Rodriguez-Villegas, D. Zagier and W. Zudilin for helpful conversations and correspondence – some recent, and some further back. This work was partially supported by NSF FRG Grant DMS-1361147. Facile polytopes {#S2} ================ Let $\Delta,\Delta^{\circ}$ be a dual pair of reflexive polytopes in $\RR^{n}$, admitting regular projective triangulations $\mathcal{T},\mathcal{T}^{\circ}$. Take $\Sigma,\Sigma^{\circ}$ to be the fans on these triangulations, and $\PP_{\Delta},\PP_{\Delta^{\circ}}$ (respectively) the toric Fano $n$-folds determined by the fans. Write $$\mathcal{A}=\mathcal{A}_{\Delta}:=\Delta\cap\ZZ^{n}=\{\underline{v}^{(0)}=\underline{0},\,\underline{v}^{(1)},\ldots,\underline{v}^{(N)}\}$$ for the integer points, and $$\mathbb{L}:=\left\{ \left.\underline{\ell}=\left(\ell_{0}=\Sigma_{i=1}^{N}\ell_{i},\,\ell_{1},\ldots,\ell_{N}\right)\in\ZZ^{N+1}\right|\Sigma_{i=1}^{N}\ell_{i}\underline{v}^{(i)}=\underline{0}\right\}$$ for the lattice of integral relations on them. The irreducible components $\{D_{i}\}_{i=1}^{N}$ of $\PP_{\Delta^{\circ}}\backslash\mathbb{G}_{m}^{n}=\mathbb{D}_{\Delta^{\circ}}$ (with $\deg_{D_{i}}(x_{k})=v_{k}^{(i)}$) generate $H^{2}(\PP_{\Delta^{\circ}},\ZZ)$, and $K_{\PP_{\Delta^{\circ}}}=-\sum_{i=1}^{N}D_{i}=:D_{0}$. Assume that the Mori cone $\mathcal{M}=\mathcal{M}_{\Delta^{\circ}}\subset H_{2}(\PP_{\Delta^{\circ}},\RR)$ (of classes of effective curves) is regular simplicial, so that $\mathbb{M}:=\mathcal{M}\cap H_{2}(\PP_{\Delta^{\circ}},\ZZ)=\ZZ_{\geq0}\langle C_{1},\ldots,C_{M=N-n}\rangle.$ Write $C_{j}\mapsto\underline{\ell}^{(j)}=C_{j}\cdot\underline{D}$ for the images under $$H_{2}(\PP_{\Delta^{\circ}},\ZZ)\overset{\cong}{\to}\mathbb{L}\overset{\cong}{\to}\text{Hom}\left(H^{2}(\PP_{\Delta^{\circ}},\ZZ),\ZZ\right),$$ so that the dual (nef) cone $\overline{\mathcal{K}}\subset H^{2}(\PP_{\Delta^{\circ}},\RR)$ has $\mathbb{K}:=\overline{\mathcal{K}}\cap H^{2}(\PP_{\Delta^{\circ}},\ZZ)=\ZZ_{\geq0}\langle J_{1},\ldots,J_{M}\rangle$ with $J_{k}\cdot C_{j}=\delta_{kj}$ and $D_{i}=\sum_{k=1}^{M}\ell_{i}^{(k)}J_{k}$. (To compute the $\underline{\ell}^{(k)}$, one may use primitive collections as in [@LZ].) Let $f_{\underline{a}}(\underline{x}):=\sum_{i=0}^{N}a_{i}\underline{x}^{\underline{v}^{(i)}}$, and $X_{\underline{a}}^{\times}:=\{f_{\underline{a}}(\underline{x})=\underline{0}\}\subset\mathbb{G}_{m}^{n}$ with (CY $(n-1)$-fold) Zariski closure $X_{\underline{a}}\subset\PP_{\Delta}$; take $X^{\circ}\subset\PP_{\Delta^{\circ}}$ to be any smooth anticanonical hypersurface. We are interested in the $\mathbb{A}$-*periods* $\pi_{\gamma}(\underline{a})=\int_{\gamma}\omega_{\underline{a}}$ ($\gamma\in H_{n-1}(X_{\underline{a}},\mathbb{A})$, $\mathbb{A}=\ZZ,\QQ,\text{ or }\CC$) of $$\omega_{\underline{a}}=\text{Res}_{X_{\underline{a}}}\Omega_{\underline{a}}=\tfrac{a_{0}}{(2\pi{\bf i})^{n-1}}\text{Res}_{X_{\underline{a}}}\left(\tfrac{dx_{1}/x_{1}\wedge\cdots\wedge dx_{n}/x_{n}}{f_{\underline{a}}(\underline{x})}\right)\in\Omega^{n-1}(X_{\underline{a}}),$$ in the *large complex-structure limit* (LCSL) — i.e., where the $t_{k}:=\underline{a}^{\underline{\ell}^{(k)}}$ are sufficiently small. These are known to solve the GKZ system $$\tau_{\text{GKZ}}^{\Delta}:\;\;\begin{array}{c} \left\{ \left(\prod_{\ell_{i}>0}\partial_{a_{i}}^{\ell_{i}}-\prod_{\ell_{i}<0}\partial_{a_{i}}^{-\ell_{i}}\right)\tfrac{1}{a_{0}}\right\} _{\ell\in\mathbb{L}},\\ \left\{ \left(\Sigma_{i=0}^{N}v_{j}^{(i)}\delta_{a_{i}}\right)\tfrac{1}{a_{0}}\right\} _{j=1,\ldots,n},\;\left(\Sigma_{i=0}^{N}\delta_{a_{i}}+1\right)\tfrac{1}{a_{0}} \end{array}$$ whose remaining solutions are the other integrals of $\frac{1}{2\pi{\bf i}}\Omega_{\underline{a}}$ over *relative* cycles in $H_{n}\left(\PP_{\Delta}\backslash X_{\underline{a}},\DD_{\Delta}\backslash\DD_{\Delta}\cap X_{\underline{a}};\CC\right)$ [@HLYZ]. A formula for the solutions to $\tau_{\text{GKZ}}^{\Delta}$ in the LCSL was given by [@HLY]: writing $\tau_{i}:=\frac{\log(t_{i})}{2\pi{\bf i}}$, $\mathcal{O}:=\CC[[\underline{t}]][\underline{\tau}]$, they are precisely the functions $\psi(\mathcal{B}_{\Delta})\in\mathcal{O}$ where $\psi\in H^{\bullet}(\PP_{\Delta^{\circ}},\CC)^{\vee}$ and $$\mathcal{B}_{\Delta}:=\sum_{\underline{\ell}\in\mathbb{M}}\mathcal{B}_{\underline{\ell}}(\underline{D})\underline{a}^{\underline{\ell}+\underline{D}}=\sum_{\underline{n}\in\ZZ_{\geq0}^{M}}\mathbb{B}_{\underline{n}}(\underline{J})\underline{t}^{\underline{n}+\underline{J}}\in H^{\bullet}(\PP_{\Delta^{\circ}},\mathcal{O})$$ with $\mathcal{B}_{\underline{n}}(\underline{J}):=\mathcal{B}_{\sum n_{k}\underline{\ell}^{(k)}}\left(\sum\underline{\ell}^{(k)}J_{k}\right)$ and $$\label{3-1} \mathcal{B}_{\underline{\ell}}(\underline{D}) :=\frac{\prod_{i=1:\ell_i <0}^N D_i (D_i -1 )\cdots (D_i + \ell_i + 1)}{\prod_{i=1:\ell_i >0}^N (D_i + 1)\cdots (D_i + \ell_i )}\times (D_0 -1)\cdots (D_0 + \ell_0).$$According to the *Hyperplane Conjecture* [@HLY; @LZ], $\psi(\mathcal{B}_{\Delta})$ is a $\CC$-period (in the above sense) precisely when $\psi$ belongs to $$\text{im}\left\{ \imath_{*}^{\bullet}:\, H^{\bullet}(X^{\circ},\CC)^{\vee}\to H^{\bullet}(\PP_{\Delta^{\circ}},\CC)^{\vee}\right\} .$$ More precisely, for each $\underline{\kappa}:=(\kappa_{1},\ldots,\kappa_{M})\in\ZZ_{\geq0}^{M}$ with $|\underline{\kappa}|:=\sum\kappa_{j}\leq n$, we compute $$B^{\underline{\kappa}}(\underline{t},\underline{\tau}):=\left.\tfrac{1}{(2\pi{\bf i})^{|\underline{\kappa}|}}\left(\partial_{J_{1}}^{\kappa_{1}}\cdots\partial_{J_{M}}^{\kappa_{M}}\mathcal{B}_{\Delta}\right)\right|_{\underline{J}=\underline{0}}=\sum_{\underline{\kappa}'+\underline{\kappa}''=\underline{\kappa}}(2\pi{\bf i})^{-|\underline{\kappa}''|}\underline{\tau}^{\underline{\kappa}'}\sum_{\underline{n}\in\ZZ_{\geq0}^{M}}b_{\underline{n}}^{\underline{\kappa}''}\underline{t}^{\underline{n}}$$ where $b_{\underline{n}}^{\underline{\kappa}''}=\left(\partial_{J_{1}}^{\kappa_{1}''}\cdots\partial_{J_{M}}^{\kappa_{M}''}\mathbb{B}_{\underline{n}}\right)\left(\underline{0}\right).$ Given bases $\{\psi_{\ell}^{r}\}\subset H_{2r}(\PP_{\Delta^{\circ}},\ZZ)$ resp. $\{\hat{\psi}_{\ell}^{r}\}\subset\text{im}(\imath_{*}^{2r})_{\ZZ}$, we obtain $\CC$-bases for the solutions to $\tau_{\text{GKZ}}^{\Delta}$ resp. for the $\CC$-periods (assuming the Hyperplane Conjecture) which are $\ZZ$-linear combinations of the $\{B^{\underline{\kappa}}\}$. That is, writing $B^{\emptyset}(\underline{t})=\sum b_{\underline{n}}^{\emptyset}\underline{t}^{\underline{n}}=:\sum a_{\underline{n}}\underline{t}^{\underline{n}}=:A(\underline{t})$, we have$$\label{3-2} \overset{(\wedge)}{\psi}{}_{\ell}^r (\mathcal{B}_{\Delta}) = A(\underline{t})\overset{(\wedge)}{P}{}_{\ell}^r (\underline{\tau}) + \sum_{|\underline{\kappa}'|<r} (2\pi{\bf i})^{|\underline{\kappa}'|-r} \underline{\tau}^{\underline{\kappa}'} \sum_{\underline{n}} \left( \sum_{|\underline{\kappa}''|=r-|\underline{\kappa}'|} \overset{(\wedge)}{c}{}_{\underline{\kappa}',\underline{\kappa}''} b_{\underline{n}}^{\underline{\kappa}''} \right) \underline{t}^{\underline{n}}$$where $\overset{(\wedge)}{P}{}_{\ell}^{r}$ are $\ZZ$-homogeneous polynomials of degree $r$ and $\overset{(\wedge)}{c}{}_{\underline{\kappa}',\underline{\kappa}''}\in\ZZ.$ \(i) The full assertion for the $\CC$-periods holds without the Hyperplane Conjecture for $r\leq\min\left\{ \tfrac{n-1}{2},1\right\} .$ Writing $\mathcal{A}'\subset\mathcal{A}$ for the points (if any) interior to facets, the periods canot depend on the $\{\log(a_{i})\}_{\underline{v}^{(i)}\in\mathcal{A}'}$ because taking $a_{i}\to0$ does not make $X_{\underline{a}}$ singular. Moreover, $X^{\circ}$ avoids the corresponding (exceptional) $\{D_{i}\}$; so there are $M-|\mathcal{A}'|$ independent $\{\hat{\psi}_{\ell}^{1}\}$ (with leading terms $A(\underline{t})\,\times$ the $M-|\mathcal{A}'|$ independent linear combinations of the $\{\tau_{j}\}$ with no such $\log(a_{i})$’s). But these must all be periods since (by mirror symmetry) there are $h_{\text{tr}}^{1,1}(X_{\underline{a}})=h_{\text{alg}}^{1,1}(X^{\circ})=M-|\mathcal{A}'|$ independent periods. \(ii) Moreover, applying $r$ $\{N_{i}=\log(T_{i})\}$ to a $\ZZ$-period with “$\log^{r}(\underline{t})$” leading term must yield a $\ZZ$-multiple of $A(\underline{t})$. So (a fixed integer multiple of) this period must be a $\ZZ$-linear combination of the $\{\psi_{\ell}^{r}(\mathcal{B}_{\Delta})\}$ plus a $\CC$-linear combination of the $\{\psi_{\ell}^{r'}(\mathcal{B}_{\Delta})\}_{r'<r}$. If all of the $\CC$-linear combinations that can appear are themselves $\CC$-periods, they can be subtracted off. So one in fact has a basis of $\CC$-periods of the form for $r\leq\min\{\tfrac{n+1}{2},2\}.$ Call $\underline{\ell}(\underline{n}):=\sum n_{k}\underline{\ell}^{(k)}$ *effective* (resp. *quasi-effective*) if all $\ell_{i}(\underline{n})$ ($i>0$) are $\geq0$ (resp. at most one $<0$). Clearly the $$a_{\underline{n}}=\left\{ \begin{array}{cc} \frac{(-\ell_{0}(\underline{n}))!}{\prod_{i>0}\ell_{i}(\underline{n})!} & \underline{\ell}(\underline{n})\text{ effective}\\ 0 & \text{otherwise,} \end{array}\right.$$ being multinomial coefficients, are all integers. Now using : - If $\ul(\un)$ is effective, then the $ $$\left.\left(\d_{D_{0}}^{r_{0}}\cdots\d_{D_{N}}^{r_{N}}\mathcal{B}_{\ul(\un)}\right)(\underline{0})\right/a_{\un}$ are $\ZZ$-linear combinations of products $$\prod_{i=0}^{N}\prod_{\tiny\begin{array}{c} \underline{P}\in\ZZ_{\geq0}^{|\ell_{i}(\un)|}\\ |\underline{P}|=r_{i} \end{array}}\prod_{j=1}^{\ell_{i}(\un)}j^{-p_{j}}.$$ Since $b_{\un}^{\uk''}/a_{\un}$ is a $\ZZ$-linear combination of these with $|\underline{r}|=|\uk''|$, $b_{\un}^{\uk''}$ can be written as $\tfrac{P}{Q}$ ($P,Q\in\ZZ$) with $Q\mid L_{|\ell_{0}(\un)|}^{|\uk''|}$, where $L_{s}:=\text{lcm}\{1,\ldots,s\}$. - If $\ul(\un)$ is quasi-effective, with (say) $\ell_{1}(\un)<0$, then $\left(\d_{D_{1}}\mathcal{B}_{\ul(\un)}\right)(\uo)=$ $$\frac{(-\ell_{0}(\un))!\,(-\ell_{1}(\un)-1)!}{\prod_{i>1}\ell_{i}(\un)!}=\left.\frac{(\Sigma_{i>1}\ell_{i}(\un))!}{\Pi_{i>1}\ell_{i}(\un)!}\right/\frac{(\Sigma_{i>1}\ell_{i}(\un))!}{(-\ell_{0}(\un))!\,(-\ell_{1}(\un)-1)!}$$ is the quotient of a multinomial coefficient by an integer of the form $A!/(B-1)!(A-B)!$, which always divides $L_{A}$. Repeated differentiation as above now shows that $b_{\un}^{\uk}=P/Q$ with $Q\mid L_{|\ul(\un)|^{+}}^{|\uk|}$, where $|\ul|^{+}=\sum_{i:\,\ell_{i}>0}\ell_{i}\;(\geq-\ell_{0}).$ \[def1\] A reflexive polytope $\Delta\subset\RR^{n}$ is *facile* if:\ $\bullet$ $\Delta,\Delta^{\circ}$ admit regular projective triangulations;\ $\bullet$ the Mori cone $\mathcal{M}_{\Delta^{\circ}}$ is regular simplicial, with generators $\ul^{(k)}$;\ $\bullet$ $\ul(\un):=\sum n_{k}\ul^{(k)}$ is quasi-effective for each $\un\in\ZZ_{\geq0}^{M}$; and\ $\bullet$ $n\leq3$, or the Hyperplane Conjecture holds for $\Delta$. Now let $W_{\bullet}$ denote the monodromy weight filtration for the large complex structure limit, with $$h_{i}:=\text{rk}\left(\text{Gr}_{2i}^{W}H^{n-1}(X_{\underline{a}})\right)=\text{rk}\left(\text{Gr}_{-2i}^{W}H_{n-1}(X_{\underline{a}})\right)=\text{rk}\left(H^{i,n-i-1}(X_{\underline{a}})\right)$$ ($\underline{a}$ very general), which is $1$ for $i=0$ or $n-1$. \[th1\] If $\Delta$ is facile, there exists a basis of $H_{n-1}(X_{\underline{a}},\CC)$, of the form $\gamma_{\mu}^{r}\in W_{2(r-n+1)}H_{n-1}(X_{\underline{a}},\CC)$ $(r=0,\ldots,n-1;\,\mu=1,\ldots,h_{r})$, with periods $$\label{5-1} \pi^r_{\mu} := \int_{\gamma^r_{\mu}} \omega_{\underline{a}} = \sum_{|\uk|=r} c_{\mu,\uk}^r \utt^{\uk} \sum a_{\un} \ut^{\un} + \sum_{|\uk|<r} \tfrac{1}{(2\pi\mathbf{i})^{r-|\uk|}}\utt^{\uk}\sum_{\un\in\ZZ^M_{\geq 0}} \beta^{r,\uk}_{\mu,\un} \ut^{\un} ,$$where each $a_{\un}\in\ZZ$, each $c_{\mu,\uk}^{r}\in\ZZ$, $\beta_{\mu,\un}^{r,\uk}=\frac{P}{Q}$ $(P,Q\in\ZZ)$ with $Q\mid L_{|\ul(\un)|^{+}}^{r-|\uk|}$, and $($for each $r,\mu)$ some $c_{\mu,\uk}^{r}\neq0$. Moreover, this basis becomes $\QQ$-rational in the associated graded $\oplus\text{Gr}_{2i}^{W}$. For the irrationality proofs, we need to apply this to certain 1-parameter families. \[def2\] A *facile CY pencil* is a family of anticanonical hypersurfaces $\mathrm{X}_{\xi}=\overline{\{\phi(\xi)=0\}}\subset\PP_{\Delta}$ parametrized by $\xi\in\PP^{1}$, where:\ $\bullet$ $\Delta\subset\RR^{n}$ is a facile polytope;\ $\bullet$ $\phi(\xi,\ux):=\sum_{\um\in\mathcal{A}_{\Delta}}\xi^{o_{\um}}P_{\um}(\xi)\ux^{\um},$ with $o_{0}=0$ and $o_{\um}>0$ for $\um\neq\uo$, $P_{\um}(\xi)\in\ZZ[\xi]$ with $\gcd_{\um\in\mathcal{A}}\{P_{\um}(\xi)\}=1$, and (if $n>1$) $\prod_{i=0}^{N}P_{\uv^{(i)}}(0)^{\ell_{i}^{(k)}}=1$ $(\forall k)$, $P_{\uo}(0)=1$;\ $\bullet$ the VHS on $H^{n-1}(\mathrm{X}_{\xi})$ is pure, with a factor $H_{\text{tr}}^{n-1}(\mathrm{X}_{\xi})=:\mathcal{W}_{\phi}$ with Hodge numbers $h^{n-1,0}=h^{n-2,1}=\cdots=h^{0,n-1}=1$; and\ $\bullet$ $\sum_{i=1}^{N}o_{\uv^{(i)}}\ell_{i}(\un)\geq|\ul(\un)|^{+}$ $\forall\un\in\ZZ_{\geq0}^{M}$.[^1] In particular, note that $\mathcal{W}_{\phi}$ has maximal unipotent monodromy at $\xi=0$. Since the resulting $t_{k}=\xi^{\sum o_{i}\ell_{i}^{(k)}}g_{k}(\xi)$ with $g_{k}(\xi)\in1+\xi\ZZ[[\xi]]$, and $2\pi\mathbf{i}\tau_{k}=(\sum o_{i}\ell_{i}^{(k)})\log\xi+\sum_{m>0}\frac{h_{km}}{m}\xi^{m}$ ($h_{km}\in\ZZ$), while also $P_{\uo}(\xi)^{-1}\in1+\xi\ZZ[[\xi]]$, substituting into $\eqref{5-1}$ and normalizing yields at once the \[cor1\] Near $\xi=0$, any facile CY pencil admits a multivalued basis $\{\gamma_{j}\}_{j=0}^{n-1}$ of $\mathcal{W}_{\phi,\CC}^{\vee}\cong H_{n-1}^{\text{tr}}(\mathrm{X}_{\xi},\CC)$ $(\QQ$-rational in $\text{Gr}_{\bullet}^{W})$, holomorphic functions $f^{(j)}(\xi)=\sum_{m\geq0}\mathfrak{f}_{m}^{(j)}\xi^{m}$, and an integer $\varepsilon\in\mathbb{N}$, such that the periods $$\Pi_{\ell}:=\int_{\gamma_{\ell}}\omega_{\xi}:=\int_{\gamma_{r}}\tfrac{1}{(2\pi\mathbf{i})^{n-1}}\text{Res}_{\mathrm{X}_{\xi}}\left(\tfrac{dx_{1}/x_{1}\wedge\cdots\wedge dx_{n}/x_{n}}{\phi(\xi,\ux)}\right)$$ take the form $$\Pi_{\ell(\xi)}=\frac{1}{(2\pi\mathbf{i})^{\ell}}\sum_{j=0}^{\ell}\frac{1}{j!}\log^{j}(\xi)f^{(\ell-j)}(\xi),$$ with each $\varepsilon L_{m}^{j}\mathfrak{f}_{m}^{(j)}\in\ZZ$, and $\mathfrak{f}_{0}^{(0)}=1$. Note that in this scenario, $\gamma_{j}$ generates $\mathrm{Gr}_{2(j-n+1)}^{W}$, the monodromy logarithm $N$ sends $\gamma_{n-1}\mapsto\gamma_{n-2}\mapsto\cdots\mapsto\gamma_{1}\mapsto\gamma_{0}$, and $\gamma_{0}$ generates the local invariant cycles in $H_{n-1}^{\text{tr}}(\mathrm{X}_{\xi},\ZZ)$. Very special values {#S3} =================== Let $\phi\in\ZZ[x_{1}^{\pm1},\ldots,x_{n}^{\pm1}]$ ($n\geq1$) be an integral Laurent polynomial, with reflexive Newton polytope $\Delta=\Delta_{\phi},$ and $\PP_{\Delta}$ be the (possibly singular) toric variety associated to a maximal projective triangulation of $\Delta^{\circ}$ [@Ba]. We begin by defining several notions we shall require for the general irrationality statement. Denote by $\X_{\phi}\overset{\rho}{\to}\PP^{1}$ the Zariski closure in $\PP_{\Delta}\times\PP_{t}^{1}$ of $\mathcal{X}_{\phi}^{\times}:=\{1-t\phi(\ux)=0\}\subset\GG_{m}^{n}\times(\PP_{t}^{1}\backslash\{0\})$, and write $t=\lambda^{-1}$, $X_{\phi,t}:=X_{\phi}^{\lambda}:=\rho^{-1}(t)$. (Note that $\rho$ is given by $1/\phi$ resp. $\phi$ when working in $t$ resp. $\lambda$.) We shall call $\phi$ *involutive* if there exists a birational map $\mathcal{I}:\X_{\phi}\dashrightarrow\X_{\phi}$ over $t\mapsto\pm t^{-1}$, defined over $\QQ$. Further, $\phi$ is said to admit a *facile $r$-cover* ($r\in\NN$) if there is a facile CY pencil $\mathrm{X}=\{{}'\phi(\xi,\uy)=0\}\subset\PP_{\Delta'}\times\PP_{\xi}^{1}$ and a dominant (genrically $r:1$) rational map $\mathcal{J}:\mathrm{X}\dashrightarrow\X_{\phi}$ over $\xi\mapsto\xi^{r}=t$. We say that $\phi$ is of *conifold type* if[^2] the nonzero critical values of $\phi:(\CC^{\times})^{n}\to\CC$ underlie (isolated) ordinary double points in the fibers. While we will not assume that $1-t\phi$ is $\Delta$-regular for general $t$, or even that the generic fiber of $\rho$ is smooth, we shall impose the conditions that the variation $\H_{\phi}$ underlain by $R^{n-1}\rho_{\times}\QQ$ is pure (of weight $(n-1)$),[^3] and that the minimal sub-VHS $\V_{\phi}$ containing the class of $$\omega_{\phi,t}:=\omega_{\phi}^{\lambda}:=\tfrac{1}{(2\pi\mathbf{i})^{n-1}}\text{Res}_{X_{\phi}^{\lambda}}\left(\tfrac{dx_{1}/x_{1}\wedge\cdots\wedge dx_{n}/x_{n}}{1-t\phi(\ux)}\right)\in\Omega^{n-1}(X_{\phi}^{\lambda})$$ is of rank $n$, with Hodge numbers $\{h^{i,n-i-1}\}_{0\leq i\leq n-1}$ all $1$. In this case we call $\phi$ *principal* (since $\V_{\phi}$ is a “principal VHS” [@Ro]). Denote the singularity (discriminant) locus of $\V_{\phi}$ by $\D_{\phi}$; since $X_{\phi,0}=\DD_{\Delta}$, $\V_{\phi}$ has maximal unipotent monodromy at $t=0$, and so $0\in\D_{\phi}$. Write $\D_{\phi}^{\times}:=\D_{\phi}\cap\GG_{m}$, $P_{\phi}(t):=\prod_{t_{0}\in\D_{\phi}^{\times}}\left(\tfrac{t}{t_{0}}-1\right)$, $\delta_{\phi}:=|\D_{\phi}^{\times}|=\deg(P_{\phi})$, and $\mathfrak{r}_{\phi}=\min\{|\mathfrak{t}|\mid\mathfrak{t}\in\D_{\phi}^{\times}\}$. For $\phi$ of conifold type, the monodromy of $\V_{\phi}$ about each point of $\D_{\phi}^{*}$ is given by a single Picard-Lefschetz transformation. Since the local exponents are then given by $\{0,1,\ldots,n-2\}\cup\{\tfrac{n-2}{2}\}$, an easy calculation[^4] shows that the Picard-Fuchs operator for $\omega_{\phi}$ takes the form $$D_{\phi}(t):=P_{\phi}(t)\delta_{t}^{n}+\sum_{k=0}^{n-1}f_{\phi,k}(t)\delta_{t}^{k}\in\CC[t,\delta_{t}].$$ For $n>2$ we remark that $\rho$ is not semistable over $0$ (and possibly at other points of $\D_{\phi}$) without blowing up $\PP_{\Delta}$ along the singularities of the base locus, but this won’t be an issue for us. The isomorphism $\X_{\phi}^{\times}\to\GG_{m}^{n}$ provides $n$ coordinates $x_{i}\in\mathcal{O}^{*}(\X_{\phi}^{\times})\cong H_{\mathcal{M}}^{1}(\X_{\phi}^{*},\QQ(1))$ whose cup product produces the *coordinate symbol* $\{\ux\}=\{x_{1},\ldots,x_{n}\}\in H_{\M}^{n}(\X_{\phi}^{\times},\QQ(n))$. We shall call $\phi$ *tempered* if this lifts to a motivic cohomology class $\Xi\in H_{\M}^{n}(\X_{\phi}^{-},\QQ(n))$, where $\X_{\phi}^{-}:=\X_{\phi}\setminus X_{\phi}^{\infty}$. Writing $T_{x_{i}}$ [\[]{}resp. $T_{\{\ux\}}$[\]]{} for the analytic chain of $x_{i}^{-1}(\RR_{-})$ [\[]{}resp. $\cap_{i=1}^{n}x_{i}^{-1}(\RR_{-})$[\]]{} and $U_{\phi}:=\PP^{1}\setminus\rho(T_{\{\ux\}})$, we term $\phi$ *strongly tempered* if the *higher normal function (HNF)* defined by $$\nu_{\phi,t}:=\mathrm{AJ}_{X_{\phi,t}}^{n,n}(\Xi|_{X_{\phi,t}})\in H^{n-1}(X_{\phi,t},\CC/\QQ(n))$$ has a single-valued holomorphic family of lifts $$\tilde{\nu}_{\phi,t}\in H^{n-1}(X_{\phi,t},\CC)$$ over $U_{\phi}$.[^5] We shall only really care about the $\V_{\phi}$-component of $\nu_{\phi}$ below, so one might as well regard it as an element of $\mathrm{ANF}(\V_{\phi}(n))$. Note that any lift (such as $\tilde{\nu}_{\phi}$) of this component must have nontrivial monodromy on $\PP^{1}\setminus\D_{\phi}$ (as opposed to just $\V_{\phi}$): otherwise its topological invariant $[\nu_{\phi}]\in\mathrm{Hom}\left(\QQ(0),H^{n}\left(\rho^{-1}(\PP^{1}\setminus\D_{\phi}),\QQ(n)\right)\right)$ would vanish. Since the latter is computed by $\Omega_{\Xi},$ which restricts to the Haar form on $\X_{\phi}^{*}\cong\GG_{m}^{n}$, this is absurd. Now either $\omega_{\phi,t}$ or $\tilde{\omega}_{\phi,t}:=t\omega_{\phi,t}$ is a holomorphic section of the canonically extended Hodge bundle $\F_{e}:=\F^{n-1}\V_{\phi,e}$. As residue forms, they are really most naturally regarded as elements of $H_{n-1}(X_{\phi,t},\CC)$ for any $t\in\PP^{1}$. If $t\notin\D_{\phi}$, then purity of $\H_{\phi}$ makes $H^{n-1}\to H_{n-1}$ an isomorphism,[^6] allowing us to treat them as cohomology classes; but over all of $\PP^{1}$ or $U_{\phi}$, they only make sense as sections of $H_{n-1}$. Conveniently enough, this pairs with $H^{n-1}(X_{\phi,t},\CC)$, allowing us to define $$\tilde{V}_{\phi}(\lambda):=\left\langle \tilde{\nu}_{\phi}^{\lambda},\tilde{\omega}_{\phi}^{\lambda}\right\rangle \in\mathcal{O}(U_{\phi}).$$ This extends to a multivalued-holomorphic function $$V_{\phi}(\lambda):=\left\langle \nu_{\phi}^{\lambda},\tilde{\omega}_{\phi}^{\lambda}\right\rangle \;\;\text{mod }\tilde{\mathscr{P}}_{\phi}^{\QQ(n)}$$ on $\PP^{1}\setminus\D_{\phi}$, defined up to $\QQ(n)$-periods of $\tilde{\omega}_{\phi}^{\lambda}$; we shall refer to both $V$ and $\tilde{V}$ as the *truncated higher normal function (THNF)* associated to $\Xi$ and $\tilde{\omega}_{\phi}$. Note that $\tilde{V}_{\phi}$ *cannot* extend to an entire function, since $\tilde{\nu}_{\phi}$ has nontrivial monodromy on $\PP^{1}\setminus\D_{\phi}$ and $\tilde{\omega}_{\phi}$ has $n$ independent periods (courtesy of the maximal unipotent monodromy). We can now state the main result of this section: \[th2\] Let $\phi(x_{1},\ldots,x_{n})$ be an integral Laurent polynomial, such that $\phi(-\ux)$ has all positive coefficients, which is reflexive, involutive, principal, strongly tempered, of conifold type, and admits a facile $r$-cover. Assume that $\delta_{\phi}=2$, $D_{\phi}\in\ZZ[t,\delta_{t}]$, and $\mathfrak{r}_{\phi}<e^{-n}$. Then $\tilde{V}_{\phi}(0)\notin\QQ^{\times}$. Proof of Theorem \[th2\] {#S3.proof} ------------------------ ### Step 1: The power series {#step-1-the-power-series .unnumbered} By involutivity, the four points of $\D_{\phi}$ have $\lambda$-values $0,\lambda_{0},\pm\lambda_{0}^{-1},\infty$, with $|\lambda_{0}|=\mathfrak{r}_{\phi}<1<|\lambda_{0}|^{-1}$. Moreover, the $\ZZ$-local system $\VV_{\phi}$ underlying $\V_{\phi}$ has maximal unipotent monodromy at $\lambda=0$, with (rank 1) invariant subsystem on the disk $D_{\mathfrak{r}_{\phi}}$ generated by a family of $(n-1)$-cycles $\vf_{0}^{\lambda}$. Indeed, we may assume that $\vf_{0}^{\lambda}=\mathcal{I}_{*}{}'\vf_{0}^{\mathcal{I}(\lambda)}$ where $\text{Tube}({}'\vf_{0})=\mathbb{T}_{n}:=\left\{ |x_{1}|=\cdots=|x_{n}|=1\right\} \subset\PP_{\Delta}\setminus X_{\phi}^{\mathcal{I}(\lambda)}$. As the Hodge bundle $\F_{e}$ has degree 1 [@GGK], and $\omega_{\phi}\in\Gamma(\PP^{1},\F_{e})$ has a zero at $\infty$, we have $(\mathcal{I}^{*}\omega)/\omega=M\lambda^{-1}$ for some $M\in\QQ^{\times}$. But then $(\mathcal{I}^{*})^{2}\omega=\mathcal{I}^{*}\frac{M}{\lambda}\omega=\frac{M^{2}}{\lambda\mathcal{I}(\lambda)}\omega=\pm M^{2}\omega,$ and since $(\mathcal{I}^{*})^{2}$ acts on $\VV_{\phi}$ (and $\VV_{\phi}^{\CC}$ is irreducible), we must have $M^{2}\in\ZZ^{\times}=\{\pm1\}$. This forces $M=\pm1$ hence $\mathcal{I}^{*}\omega=\pm\tilde{\omega}$. The holomorphic period at $\lambda=0$ is therefore A\_() &:= \_[\_0\^]{} \^\_ = \_[’\_0\^[()]{}]{} \^\* \^\_\ &= \_[’\_0\^[()]{}]{} \_\^[()]{} = \_[’\_[0,]{}]{} \_[,]{}\ &= \_[\_n]{} , and of course we may change the signs of $\vf_{0}$ and $\lambda$ if needed so that$$\label{13-1} A(\lambda)=\sum_{k\geq 0} [\phi^k]_{\uo} \lambda^k =: \sum_{k\geq 0}a_k \lambda^k$$(where $[\cdot]_{\uo}$ denotes “constant term”), with $a_{0}=1$. By the fundamental result of [@DvdK], we have $\mathfrak{r}_A := \left(\limsup_{k\to \infty} |a_k|^{1/k}\right)^{-1} \in \CC^{\times}$; and since $A$ is a period of $(\VV_{\phi},\tilde{\omega}_{\phi})$, $\mathfrak{r}_A$ must be the modulus of an element of $\D_{\phi}^{\times}$ (i.e., $\mathfrak{r}_{\phi}$ or $\mathfrak{r}_{\phi}^{-1}$). Since the $[\phi^k]_{\uo}$ are all (positive) integers, $\mathfrak{r}_A = \mathfrak{r}_{\phi}$. On the other hand, positivity of the coefficients of $\phi(-\ux)$ forces $\phi(T_{\ux})\subset [1,\infty]$. Now the global minimum of $\phi(T_{\ux})$ necessarily occurs at a critical point of $\phi$; more precisely, it may be regarded as the terminus of a Lefschetz thimble on the generator $\d T_{\Xi}$ of $\V_{\phi}^{0,\text{lim}}/N\V_{\phi}^{0,\text{lim}} \cong \text{Gr}^W_0 H_{n-1}(X_{\phi}^0)$, and thus is $\pm$ the $\lambda$-coordinate of a point of $\D_{\phi}^*$. Clearly this must be the larger one, so that $U_{\phi}$ contains the disk $D_{\mathfrak{r}_{\phi}^{-1}}$ about $\lambda=0$. Writing $$\tilde{V}_{\phi}(\lambda)=:\sum_{k\geq 0} v_k \lambda^k ,$$it follows that $\mathfrak{r}_V := \left(\limsup_{k\to\infty}|v_k|^{1/k}\right)^{-1}=\mathfrak{r}_{\phi}^{-1}$.[^7] Moreover, since $A(\lambda)$ is (up to scale) the *only* period of $\tilde{\omega}_{\phi}$ invariant about $\lambda =0$, and it is *not* invariant about $\lambda = \lambda_0$, we conclude that $\tilde{V}_{\phi}(\lambda)$ \[resp. $\tilde{\nu}_{\phi}^{\lambda}$\] is the unique analytic continuation of $V_{\phi}(\lambda)$ \[resp. holomorphic lift of $\nu_{\phi}^{\lambda}$\] which is well-defined on $U_{\phi}$. In particular, $\tilde{V}_{\phi}(0)=v_0\in \CC$ is well-defined. Assuming henceforth that $v_0 \neq 0$, we may consider $$B_{\phi}(\lambda):= -\tilde{V}_{\phi}(\lambda) + \tilde{V}_{\phi}(0)A_{\phi}(0) =: \sum_{k\geq 1} b_k \lambda^k.$$Obviously this has radius of convergence $\mathfrak{r}_B=\mathfrak{r}_A \,(<1)$, while $v_k = v_0 a_k - b_k \to 0$ as $k\to 0$. Restricting if necessary to a subsequence $a_{k_j} \to \infty$, we therefore have$$\lim_{j\to \infty}\frac{b_{k_j}}{a_{k_j}} = v_0.$$ ### Step 2: The inhomogeneous equation {#step-2-the-inhomogeneous-equation .unnumbered} Since the holomorphic period of $\tilde{\omega}$ about $\lambda = 0$ is obtained by substituting $\lambda$ for $t$ in the holomorphic period of $\omega$ about $t=0$, the same goes for the (inhomogeneous) Picard-Fuchs equations: that is,$$D_{\phi}(\lambda)\tilde{\mathscr{P}}_{\phi}^{\CC} = 0.$$Consequently, in $\V_{\phi}$ we have $$P_{\phi}(\lambda) \nabla_{\delta_{\lambda}}^n [\tilde{\omega}_{\phi}] = -\sum_{k=0}^{n-1} f_{\phi,k}(\lambda) \nabla^k_{\delta_k} [\tilde{\omega}_{\phi}];$$ and regarding $\tilde{\omega}_{\phi}$ as a section of $\V_{\phi,e}$, maximal unipotent monodromy at $\lambda = 0$ and $\text{Res}_0 \nabla = \tfrac{1}{2\pi \mathbf{i}} N$ imply the independence of the $\{ \nabla_{\delta_k}^k [\tilde{\omega}_{\phi}]\}_{k=0}^{n-1}$ in the fiber $\V_{\phi,e}^0$. As $N^n=0$, $\nabla_{\delta_{\lambda}}^n [\tilde{\omega}_{\phi}]$ vanishes in $\V_{\phi,e}^0$, and so all $f_{\phi,k}(0)=0$ $\implies$ $f_{\phi,k}(\lambda)=:\lambda g_{\phi,k}(\lambda)$. Moreover, since $D_{\phi}(\lambda)\circ \lambda^{-1}$ (like $D_{\phi}(t)$) kills $[\omega_{\phi}]$, we have $$F(t)\circ D_{\phi}(t) = D_{\phi}(\lambda)\circ \lambda^{-1} = D_{\phi}(\pm t^{-1})\circ t$$for some function $F(t)$. Since ($\delta_{\phi}=2\implies$) $P(\lambda)$ is quadratic, a short computation shows that $F(t)=\pm\frac{1}{t}$ and that the $t g_{\phi,k}(t^{-1})$ must be polynomials, forcing the $g_{\phi,k}$ to be linear. Next, define the *Yukawa coupling* by $$\label{15-1} Y(\lambda) :=Q\left( [\tilde{\omega}_{\phi}^{\lambda}],\nabla_{\delta_{\lambda}}^{n-1}[\tilde{\omega}^{\lambda}_{\phi}]\right) \in \CC(\lambda)^{\times}.$$ Let $w$ be a local coordinate at a point $p\in\D_{\phi}^{\times}$, about which monodromy is described by $C\mapsto C-\langle C,\sigma\rangle\sigma$ (since $\phi$ is of conifold type). Writing $[\tilde{\omega}_{\phi}^{\lambda}]$ with respect to a local basis $\{\sigma_0,\ldots,\sigma_{n-1}\}$ of $\VV_{\phi}$ with $\sigma_1,\ldots,\sigma_{n-1}$ invariant about $p$ and $\sigma = \sigma_0$ \[resp. $\sigma_1$\] for $n$ odd \[resp. even\] (with $\sigma_0\mapsto -\sigma_0$ resp. $\sigma_0 \mapsto \sigma_0 - \sigma_1$), the coefficients $u_i (w)$ of $\{\sigma_1,\ldots,\sigma_n\}$ are holomorphic and we have (near $p$) $u_0(w)\sim w^{\frac{n-2}{2}}$ \[resp. $u_0(w)\sim u_1(w)\log(w)\sim w^{\frac{n-2}{2}}\log(w)$\]. From this we compute $Y\sim w^{\frac{n-2}{2}-(n-1)}\times w^{\frac{n-2}{2}} = w^{-1}$, so that $Y$ has (at worst) simple poles at $\D_{\phi}^{\times}$. Moreover, the pairing makes sense in $\V_{\phi,e}^0$, so that $Y(0)\neq \infty$; while $\mathcal{I}^* \tilde{\omega}^{\lambda}_{\phi} = \pm \lambda \tilde{\omega}^{\lambda}_{\phi}$ $\implies$ $Y(\mathcal{I}(\lambda))=\pm \lambda^2 Y(\lambda)$ $ \implies$ $\text{ord}_{\infty}(Y)\geq 2$. Since $Y$ is rational with only 2 simple poles, it can have only the double zero at $\infty$, and thus[^8]$$Y(\lambda) = \frac{Y(0)}{P_{\phi}(\lambda)}.$$ To evaluate $Y(0)$, and also anticipating Step 3, we extend $\vf_0^{\lambda}$ to a basis $\{\vf_j^{\lambda}\}_{j=0}^{n-1}$ of $(\VV_{\phi}^{\lambda,\QQ})^* \cong H_{n-1}(X_\phi^{\lambda},\QQ)$ on an angular sector of the punctured disk $D_{\mathfrak{r}_{\phi}}^{\times}$ satisfying $\vf_{n-j-1} = N^j \vf_{n-1}$ hence $\vf_j \in W_{2(j-n+1)}$ ($N=\log(T)$ the monodromy logrithm at $0$, and $W_{\bullet} = W(N)[n-1]_{\bullet}$). From $NQ=-QN$ and $N^n = 0$, the $Q(\vf_j,\vf_k)$ are zero for $j+k<n-1$, and the $(-1)^j Q(\vf_j,\vf_{n-j-1})$ equal a common constant $Q_0^{-1}\in \QQ^{\times}$. We may then modify $\vf_j\mapsto \hat{\vf}_j = \vf_j + \sum_{i<j} \alpha_{ij}\vf_i \in W_{2(j-n+1)}$ ($\alpha_{ij}\in \QQ$) so that $Q(\hat{\vf}_i,\hat{\vf}_{n-j-1}) = (-1)^i Q_0^{-1} \delta_{ij}$, with $N\hat{\phi}_j = \hat{\phi}_{j-1} + \sum_{i<j-1} \eta_{ij} \hat{\vf}_i$ ($\eta_{ij}\in\QQ$) and dual basis $\{\check{\vf}_j\}$ of $\VV_{\phi}^{\QQ}$. (Note that $\check{\vf}_j \in W_{2(n-j-1)}\VV_{\phi}$, $-N\check{\vf}_j = \check{\vf}_{j+1} + \sum_{i>j+1} \eta_{ji} \check{\vf}_i$, while $Q(\check{\vf}_{n-i-1},\check{\vf}_j) = (-1)^i Q_0 \delta_{ij}$.) Writing locally $$[\tilde{\omega}_{\phi}^{\lambda}] = \sum_{j=0}^{n-1} \left( \int_{\hat{\vf}_j} \tilde{\omega}_{\phi}^{\lambda} \right) \check{\vf}_j =: \sum_{j=0}^{n-1} \hat{\pi}_j (\lambda) \check{\vf}_j$$in $\V_{\phi}$, and $\pi_j(\lambda):=\int_{\vf_j} \tilde{\omega}_{\phi}^{\lambda}$, we have $$\pi_0(\lambda) = \hat{\pi}_0(\lambda) = A_{\phi}(\lambda)=:A^{(0)}_{\phi}(\lambda)$$and $$\hat{\pi}_j = \pi_j(\lambda) + \sum_{i<j} \alpha_{ij} \pi_i (\lambda);$$while in accordance with the monodromy properties of $\vf_j$, $$\label{17-1} \pi_j(\lambda) = \sum_{k=0}^j \frac{\log^k (\lambda)}{(2\pi \mathbf{i})^k k!} A_{\phi}^{(j-k)}(\lambda)$$ for some functions $A_{\phi}^{(\ell)}(\lambda)$ holomorphic on $D_{\mathfrak{r}_{\phi}}$. Clearly, the limits $\lim_{\lambda\to 0} \log^{\ell}(\lambda) \left( \delta_{\lambda}^{n-1} \hat{\pi}_j \right) (\lambda)$ are zero for $j<n-1$, while we have $\lim_{\lambda\to 0} (2\pi\mathbf{i})^{n-1}\left( \delta_{\lambda}^{n-1}\hat{\pi}_{n-1}\right) (\lambda)=1=A^{(0)}_{\phi}(0)$; and so $Y(0)=(2\pi \mathbf{i})^{1-n}\times Q(\check{\vf}_0,\check{\vf}_{n-1})=\frac{\pm Q_0}{(2\pi\mathbf{i})^{n-1}}$. Now the fiberwise restrictions $\Xi^{\lambda}:=\Xi|_{X^{\lambda}_{\phi}}$ ($\lambda\neq \infty$) have trivial class in $\text{Hom}_{\text{MHS}}\left(\QQ(0),H^n(X^{\lambda}_{\phi},\QQ(n))\right)$, so that $T_{\Xi}$ is a coboundary over any sufficiently small $B\subset \mathbb{A}^1_{\lambda}$. Since the regulator current $R_{\Xi}$ has $dR_{\Xi} = \frac{dx_1}{x_1}\wedge \frac{dx_n}{x_n} - (2\pi\mathbf{i})^n \delta_{T_{\Xi}}$ on $\X_{\phi}^-$, writing $T_{\Xi}|_{\rho^{-1}(B)}\equiv \d \Gamma_B$ (mod $\d\rho^{-1}(B)$) yields a current $\tilde{R}^B_{\Xi}:=R_{\Xi}|_{\rho^{-1}(B)} + (2\pi\mathbf{i})^n \delta_{\Gamma_B}$ with closed fiberwise pullbacks. This yields $$\label{18-1} \nabla_{\delta_{\lambda}}\nu_{\phi}^{\lambda} = \nabla_{\delta_{\lambda}}[\tilde{R}_{\Xi^{\lambda}}^B ] = \left[ \left. \left( \tfrac{dx_1}{x_1}\wedge\cdots \wedge \tfrac{dx_n}{x_n} \,\lrcorner \,\widetilde{\delta_{\lambda}} \right)\right|_{X^{\lambda}_{\phi}}\right] = -(2\pi\mathbf{i})^{n-1}[\omega_{\phi}^{\lambda}].$$ Using this (and $Q(\omega,\nabla_{\delta_{\lambda}}^k \tilde{\omega})=0$ for $k<n-1$), one computes that $\delta_{\lambda}^k V(\lambda) = \delta_{\lambda}^k Q(\nu,\tilde{\omega})= Q(\nu,\nabla_{\delta_{\lambda}}^k \tilde{\omega})$ for $k<n$ and $-(2\pi\mathbf{i})^{n-1}Q(\omega,\nabla_{\delta_{\lambda}}^{n-1}\tilde{\omega})+Q(\nu,\nabla_{\delta_{\lambda}}^n \tilde{\omega})$ for $k=n$. Using , we have at once $$D_{\phi}(\lambda) V_{\phi}(\lambda)= \pm (2\pi \mathbf{i})^{n-1}\lambda Y(\lambda) P_{\phi}(\lambda) = \pm Q_0 \lambda.$$Notice that while $V_{\phi}(\lambda)$ is multivalued, the ambiguities are killed by $D_{\phi}(\lambda)$. Moreover, $\tilde{V}_{\phi}(\lambda)$ satisfies the same equation, so that $D_{\phi}(\lambda)A_{\phi}(\lambda)=0$ $\implies$ $$\label{18-2} D_{\phi}(\lambda)B_{\phi}(\lambda)=\mp Q_0 \lambda.$$ ### Step 3: Arithmetic of coefficients {#step-3-arithmetic-of-coefficients .unnumbered} By and the integrality of $\phi$, we have $a_m \in \ZZ$ ($\forall m$). Expressing $b_m$ as $\frac{p_m}{q_m}$ with $p_m \in \ZZ$ and $q_m \in \NN$ relatively prime, we claim that for some fixed $\varepsilon_B \in \NN$, $$\label{18-3} q_m \mid \varepsilon_B (m!)^n\;\;(\forall m).$$ Indeed, if we write $P_{\phi}(\lambda)=c_n ' \lambda^2 + c_n '' \lambda +1$ and $g_{\phi,j}(\lambda) = c_j '\lambda + c_j ''$ (al $c_j ', c_j '' \in \ZZ$; $c_n ' =\pm 1$), substituting $\sum_{m\geq 1}b_m \lambda^m = B_{\phi}(\lambda)$ in yields $b_1 = \pm Q_0 \in \QQ$ and the recurrence $$-m^n b_m = \left( \Sigma_{j=0}^n c_j ' (m-2)^j \right) b_{m-2} + \left( \Sigma_{j=0}^n c_j '' (m-1)^j \right) b_{m-1}.$$Taking $\varepsilon_B = q_1$, this establishes . We claim that, in addition (modifying $\varepsilon_B$ if necessary), $$\label{19-1} q_m \mid \varepsilon_B L_{rm}^n \;\; (\forall m).$$ To show this, we make use of the facile $r$-cover $\mathcal{J}$, which induces an isomorphism of VHS $\mathcal{J}^*\V_{\phi}\cong \W_{'\phi}$ hence of their extended Hodge bundles $\mathcal{O}_{\PP^1_{\xi}}(r)\cong \mathcal{J}^* \F_{e,\phi}^{n-1} \cong \F_{e,{}'\phi}^{n-1}$ over $\PP^1_{\xi}$, of which $\mathcal{J}^*\omega_{\phi,\xi^r}$ and $$'\omega_{\xi} = (2\pi\mathbf{i})^{1-n} \text{Res}_{\mathrm{X}_{\xi}}\left( \tfrac{dy_1/y_1\wedge \cdots \wedge dy_n/y_n}{'\phi(\xi,\uy)}\right)$$are sections. Since $\gcd_{\um\in {}'\mathcal{A}}\{ {}'P_{\um}(\xi)\}=1$, $'\omega_{\xi}$ has no zeroes over $\mathbb{A}^1_{\xi}$, so that both sections share the divisor $r[\infty]$. Since $(\mathcal{J}_{\xi}^{\pm 1})_*$ exchanges the generators $\mathcal{I}_*(\vf_0)$ and $\gamma_0$ of the integral invariant cycles about $\xi =0$ and $\t=0$, and $\lim_{t\to 0} \int_{\mathcal{I}_*(\vf_0)}\omega_{\phi,t} = 1 = \lim_{\xi \to 0}\int_{\gamma_0} {}'\omega_{\xi},$ we find that $'\omega_\xi = \mathcal{J}^* \omega_{\phi,\xi^r}$. Via $\mathcal{I}^*\tilde{\omega}_{\phi}^{\xi^r} = \omega_{\phi,\xi^r}$, Corollary \[cor1\], and , we therefore have ($0\leq \ell\leq n-1$) $$\hat{\pi}_{\ell} (\xi^r )= \int_{\hat{\vf}_{\ell}} \tilde{\omega}_{\phi}^{\xi^r} = \int_{\mathcal{J}_* \hat{\gamma}_{\ell}} \omega_{\phi,\xi^r} = \int_{\hat{\gamma}_{\ell}} {}'\omega_{\xi}$$for some $\hat{\gamma}_{\ell} = \sum_{k\leq \ell} (2\pi\mathbf{i})^{k-\ell}\beta_{k\ell}\gamma_k \in W_{2(\ell-n+1)} \W_{\phi,\QQ}^\vee$ ($\beta_{k\ell}\in \CC$; $\beta_{\ell\ell}=r^{\ell}$). That is, $$\label{20-1} \hat{\pi}_{\ell}(\xi^r) \;=\; \frac{1}{(2\pi \mathbf{i})^{\ell}} \sum_{\tiny\begin{array}{c} m\geq 0, \\ 0\leq j\leq k\leq \ell \end{array}} \frac{1}{j!}\,\mathfrak{f}_m^{(k-j)}\, \beta_{k\ell}\,\, \xi^m\, \log^j (\xi).$$ Returning to the $\lambda$-disk for a moment, we have $[\tilde{\omega}_{\phi}^0]\equiv Q_0 \hat{\vf}_{n-1}$ in $H_{n-1}(X_{\phi}^0)$ so that $\tilde{\nu}_{\phi}^0 = Q_0^{-1}v_0 \check{\vf}_{n-1}$ in $H^{n-1}(X_{\phi}^0 )$. This fixes the constant of integration, so that gives $$\label{20-2} \begin{split} \tilde{\nu}_{\phi}^{\lambda} - Q_0^{-1} v_0 \check{\vf}_{n-1} &= \int_0 \left( \nabla_{\delta_{\lambda}} \nu_{\phi}^{\lambda}\right) \frac{d\lambda}{\lambda} \\ &= -(2\pi\mathbf{i})^{n-1} \int_0 [\omega_{\phi}^{\lambda}]\frac{d\lambda}{\lambda}\\ &= (2\pi\mathbf{i})^{n-1}\sum_{\ell=0}^{n-1} \int_0 \hat{\pi}_{\ell}(\lambda)d\lambda \;\check{\vf}_{\ell} \end{split}$$ in $\V_{\phi}$. Since $Q(Q_0^{-1}v_0 \check{\vf}_{n-1},\tilde{\omega}_{\phi}^{\lambda})=v_0 A_{\phi}(\lambda)$ and $Q(\tilde{\nu}_{\phi}^{\lambda},\tilde{\omega}_{\phi}^{\lambda})=\tilde{V}_{\phi}(\lambda)$, pairing with $[\tilde{\omega}_{\phi}^{\lambda}]=\sum_{\ell'=0}^{n-1} \hat{\pi}_{\ell'}(\lambda) \check{\vf}_{\ell'}$ yields the key formula $$-B_{\phi}(\lambda) = (-2\pi\mathbf{i})^{n-1} Q_0 \sum_{\ell=0}^{n-1} (-1)^{\ell} \hat{\pi}_{n-\ell-1}(\lambda) \int_0 \hat{\pi}_{\ell}(\lambda)d\lambda.$$Substituting $\xi^r =\lambda$, this becomes $$\pm\sum_{\mu\geq 1}b_{\mu}\xi^{r\mu}=(2\pi\mathbf{i})^{n-1}Q_0 \sum_{\ell=0}^{n-1} (-1)^{\ell} \hat{\pi}_{n-\ell-1}(\xi^r)\int_0 \hat{\pi}_{\ell}(\xi^r)r\xi^{r-1}d\xi,$$and using $\int_0 \log^a(x)x^{b-1}dx=x^b \sum_{c=0}^a \frac{(-1)^{a-c}a!}{(a-c)! b^{c+1}}\log^{a-c}(x)$ together with , $$=Q_0 r\sum_{m\geq 0}\xi^m \sum_{(*)_m}\frac{(-1)^{\ell+j'-i'}}{(j'-i')!j''!}\frac{\mathfrak{f}_{m''}^{(k''-j'')}\mathfrak{f}_{m'}^{(k'-j')}\beta_{k'',n-\ell-1}\beta_{k',\ell}}{(m'+r)^{i'+1}}\log^{j''+j'-i'}(\xi),$$where the $\sum_{(*)_m}$ (finite for each $m$) is over $m'+m''=m-r$, $0\leq \ell\leq n-1$, $0\leq j''\leq k''\leq n-\ell-1$, and $0\leq i'\leq j'\leq k'\leq \ell$. Since this plainly has to be a power series in $\xi^r$, the $\log^* \xi$ terms must cancel out (forcing $j''=0=j'-i'$), leaving us with $$\label{21-1} =Q_0 r \sum_{m\geq 0}\xi^m \sum_{(*)_m} \frac{(-1)^{\ell}}{(m'+r)^{j'+1}}\mathfrak{f}^{(k'')}_{m''}\mathfrak{f}^{(k'-j')}_{m'} \beta_{k'',n-\ell-1}\beta_{k',\ell} .$$ Let $\{\sigma_1,\ldots,\sigma_d\}$ ($\sigma_1=1$) be a basis of the $\QQ$-vector space $\mathscr{B}$ generated by all the products $\{\beta_{ij}\beta_{k\ell}\}$, and write $\beta_{ij}\beta_{k\ell} = \sum_{s=1}^d \mathfrak{q}_s^{ijk\ell}\sigma_s$. Since $\eqref{21-1} = \pm \sum b_{\mu} \xi^{r\mu}$ with $b_{\mu}\in \QQ$ (and $\mathfrak{f}_b^{(a)}\in\QQ$), the resulting “$\sigma_s$-components” of vanish for $s>1$, while $$\label{21-2} \pm b_{\mu} = Q_0 r \sum_{(*)_{r\mu}} \frac{(-1)^{\ell}}{(m'+r)^{j'+1}} \mathfrak{f}^{(k'')}_{m''}\mathfrak{f}^{(k'-j')}_{m'} \mathfrak{q}_1^{k'',n-\ell-1,k',\ell}.$$ Choose $\epsilon\in\NN$ sufficiently large that all the $\{\epsilon Q_0 \mathfrak{q}_1^{ijk\ell}\}$ (a finite set) are integers. Now $m'+r\leq r\mu$ $\implies$ $\frac{L_{r\mu}^{j'+1}}{(m'+r)^{j'+1}}\in\ZZ$, while Corollary \[cor1\] $\implies$ $\varepsilon \mathfrak{f}_{m''}^{(k'')}L_{r\mu}^{k''}, \varepsilon \mathfrak{f}^{(k'-j')}_{m'} L_{r\mu}^{k'-j'}\in \ZZ$. Since in each term of RHS we have $k''+k'-j'+j'-1=k''+k'+1\leq (n-\ell-1)+\ell+1=n$, multiplying the original $\varepsilon_B$ by $\epsilon\varepsilon^2$ gives $\varepsilon_B b_{\mu} L_{r\mu}^n \in \ZZ$, as desired. ### Step 4: Irrationality of $v_0$ {#step-4-irrationality-of-v_0 .unnumbered} Let $\Lambda_m := \text{gcd}(m!,L_{rm})$, so that by Step 3 we have $\varepsilon_B \Lambda_m^n b_m =: B_m \in \ZZ$ ($\forall m$). Writing $\mathcal{P}_m :=\{ p \text{ prime}\mid p\leq m\}$, set $$\pi(m)=|\mathcal{P}_m | \;\; \text{and} \;\; \chi(m):=\sum_{p\in\mathcal{P}_m} \log(p).$$Evidently $$e^{\chi(m)}\leq \Lambda_m \leq \prod_{\tiny \begin{array}{c} p\leq m \\ p\text{ prime}\end{array}} p^{\lfloor \log_p (rm)\rfloor } \leq (rm)^{\pi(m)},$$hence $$\label{22-1} e^{\chi(m)/m} \leq \Lambda_m ^{1/m} \leq (rm)^{\pi(m)/m} .$$ By the Prime Number Theorem and its proof, $\frac{\pi(m)}{m} \sim \frac{1}{\log(m)}$ and $\lim_{m\to \infty} \frac{\chi(m)}{m}=1$. So the outer terms of limit to $e$, and so does $\Lambda_m^{1/m}$. Finally, suppose $v_0 = \frac{P}{Q}$ ($P\in\ZZ$, $Q\in \ZZ\setminus \{0\}$). Then \_[m]{} &| \_B \_m\^n a\_m P - B\_m Q|\^\ &= ( \_[m]{} \_m\^[1/m]{})\^n \_[m]{} |\_B Q|\^\_[m]{} |a\_m v\_0 - b\_m |\^\ &= e\^n 1 \_\ &&lt;1 , and for some sufficiently large $m$ we therefore have $$0<\left| \varepsilon_B \Lambda_m^n a_m P - B_m Q\right| <1 ,$$a contradiction. Q.E.D. Casting a wider net {#S3.last} ------------------- By [@7K Thm. 5.2], we can think of $\tilde{\nu}_{\phi}^0$ as computing the extension of $\QQ(0)$ by $W_{-2n}H^{n-1}(X_{\phi}^0,\QQ(n))\cong \QQ(n)$ associated to $\Xi|_{X_{\phi}^0} \in H^n_{\M}(X_{\phi}^0,\QQ(0))$. Since $(\mathcal{I}^*\omega)|_{X_{\phi}^0}$ generates the top graded piece $\text{Gr}^W_0 H_{n-1}(X_{\phi}^0,\QQ)$, pairing with it sends the generator of $\QQ(n)$ to $(2\pi\mathbf{i})^n$. So $\tilde{V}_{\phi}(0)$ realizes $\mathrm{AJ}\left( \Xi|_{X_{\phi}^0}\right)\in \CC/(2\pi\mathbf{i})^n\QQ$, which one interprets (arguing as in \[op. cit.\], at least for $n\leq 3$) as a Borel regulator value $r_{\text{Bor}}$ for $K_{2n-1}$ of the number field $k$ required to resolve $X_{\phi}^0$. Since this is just the field of definition of $\mathcal{I}$, the numbers $\tilde{V}_{\phi}(0)$ appearing in Theorem \[th2\] are limited (at best) to $\zeta(n)$. The first step in a generalization of this result would be to expand the notion of involutivity: - drop the requirement $k=\QQ$, - replace $t\mapsto \pm t^{-1}$ by $t\mapsto \frac{1}{Ct}$ ($C\in\ZZ$), and - allow $\mathcal{I}$ to be a correspondence inducing an isomorphism between the $\QQ$-VHS $\V_{\phi}$ and its pullback. Unfortunately, we have to pay for this expansion with a stronger bound: \[prop\] Let $\phi$ be as in Theorem \[th2\], but with the weaker involutivity just described. Assume in addition that $\mathfrak{r}_{\phi} < \frac{e^{-n}}{|C|}$. Then $\tilde{V}_{\phi}(0)\notin\QQ$. The proof is a straightforward, but tedious, generalization of the above. One nice formal consequence is that $\mathcal{I}^* \omega = \frac{c}{\lambda} \omega $ for $c\in\CC^{\times}$ with $\frac{c^2}{C}=\pm 1$, so that $\tilde{V}_{\phi}(0) = \langle \tilde{\nu}_{\phi}^0 , \tilde{\omega}_{\phi}^0 \rangle = c^{-1} \langle \tilde{\nu}_{\phi}^0 ,\mathcal{I}^* \omega_{\phi,0}\rangle = c^{-1} r_{\text{Bor}}.$ For example, if $n=2$ and $\mathcal{I}$ is defined over $\QQ(\mathbf{i})$, then one expects $r_{\text{Bor}}\underset{\QQ^{\times}}{\sim} \mathbf{i}{G}$, where $G = L(\chi_{-4},2)$ is Catalan’s constant. But then, one also expects $c\underset{\QQ^{\times}}{\sim}\mathbf{i}$, so that $\tilde{V}_{\phi}(0) \underset{\QQ^{\times}}{\sim} G$ — a not-insignificant “calibration”, as irrationality of $\mathbf{i}G$ and of $G$ are rather different things. Naturally, we don’t have a proof of the latter, but we will briefly discuss some higher normal functions with $G$ as a special value below. \[IJrem\] The function of $\mathcal{I}$ and $\mathcal{J}$ in the above proof is to match periods of a pullback of $\V_{\phi}$ with those of $\V_{\phi}$ resp. the facile family. If one has such a matching by other means, there is obviously no need for the maps of varieties. Another natural way to relax the hypotheses is to permit the Newton polytope of $\phi$ to be non-reflexive, as long as $\uo$ remains its unique interior integral point. This ensures that, while $X_{\phi}^{\lambda}$ may not be Calabi-Yau, $h^{n-1,0}(X_{\phi}^{\lambda})$ remains $1$. More significantly, one could abandon the “principality” constraint that Hodge numbers of $\V_{\phi}$ *all* equal $1$, in order to make contact with results (such as [@Z2]) involving linear forms in more than one zeta value; two likely sources of interesting examples will be discussed in $\S$\[S5.4\]. Low dimension {#S4} ============= In order to implement Theorem \[th2\] in any specific cases, we must be able to check strong temperedness and compute $\tilde{V}_{\phi}(0)$. To this end, we examine the boundary structure more closely. Let $\phi$ and $\PP_{\Delta}$ be as in the first paragraph of $\S$\[S3\], with associated family $\X_{\phi} \underset{\rho}{\to}\PP^1$ of anticanonical hypersurfaces. Consider the toric variety $\hat{\PP}_{\Delta}$ associated to $\Delta^{\circ}$ without the triangulation, with canonical blow-down morphisms $\PP_{\Delta} \underset{b}{\twoheadrightarrow} \hat{\PP}_{\Delta}$, $\DD_{\Delta} \underset{b}{\twoheadrightarrow} \hat{\DD}_{\Delta} = \hat{\PP}_{\Delta} \setminus \mathbb{G}_m^n$. Let $\sigma \subset \Delta$ be a codimension-$j$ face with ($(j-1)$-dimensional) dual $\sigma^{\circ}\subset \Delta^{\circ}$, and $\mathbb{G}_m^{n-j} \cong \hat{\DD}_{\sigma}^{\times} \subset \hat{\DD}_{\sigma}\subset \hat{\DD}_{\Delta}$ the corresponding (open resp. closed) codimension-$j$ stratum. The ($(n-j+1)$-dimensional) strata $\DD_{\sigma}^{\alpha} \subseteq b^{-1}(\hat{\DD}_{\sigma})$ correspond to the ($(j-i-1)$-dimensional) faces $\sigma_{\alpha}^{\circ} \subseteq \sigma^{\circ}$ in the triangulation of $\Delta^{\circ}$. A basis of the flag $\text{ann}(\sigma^{\circ})\subseteq \text{ann}(\sigma_{\alpha}^{\circ})\subseteq \ZZ^n$ produces $n-j+1$ toric coordinates $\{x_1^{\sigma},\ldots,x_{n-j}^{\sigma};y_1,\ldots,y_i\}$ on $\DD_{\sigma}^{\alpha,\times}$, with $b$ induced by forgetting the $\{y_i\}$. Writing $Y_{\phi}:=X_{\phi}^{\lambda} \cap \DD_{\Delta} \underset{b}{\twoheadrightarrow} \hat{Y}_{\phi}$, each[^9]$$Y_{\sigma}^{\alpha} := Y_{\phi} \cap \DD_{\sigma}^{\alpha} = \cup_{i=1}^{\mu_{\sigma}}Y_{\sigma}^{\alpha,i} \underset{b}{\twoheadrightarrow} \cup_{i=1}^{\mu_{\sigma}} \hat{Y}_{\sigma}^i = \hat{Y}_{\sigma} =:\hat{Y}_{\phi} \cap \DD_{\sigma}$$is cut out of $\DD_{\sigma}^{\alpha}$ resp. $\hat{\DD}_{\sigma}$ by a “face polynomial” $\phi_{\sigma}(x_1^{\sigma},\ldots ,x_{n-j}^{\sigma})$ given (up to a shift) by rewriting the terms $\{ c\ux^{\um} \mid \um\in\sigma \}$ of $\phi$ in terms of $\ux^{\sigma}$. A precondition for temperedness of $\phi$ (as defined above) is that the iterated residues of the coordinate symbol $\{\ux\}|_{(X_{\phi}^{\lambda})^{\times}}$ along strata $Y_{\sigma}^{\alpha}$ of $X_{\phi}^{\lambda}\setminus (X_{\phi}^{\lambda})^{\times}$ all vanish, which is equivalent to the vanishing of the $\{ \ux^{\sigma}\}$ on $\hat{Y}_{\sigma}^{\times}$: $\phi$ is *weakly tempered*[^10] if the symbols $$\label{26-1}\{ x_1^{\sigma},\ldots ,x_{n-j}^{\sigma}\}|_{\hat{Y}_{\sigma}^{\times}} = 0 \in K_{n-j}^M (\bar{\QQ}(\hat{Y}_{\sigma}^i))\otimes \QQ$$ for all $1\leq j\leq n-1$, $\sigma\in \Delta(j)$, and $i=1,\ldots,\mu_{\sigma}$. This is a condition on the face polynomials $\phi_{\sigma}=\prod\phi_{\sigma,i}$. For example, holds: - for $j=n-1$ if all edge polynomials are cyclotomic; and - for $j=n-2$ if (for all $2$-faces) the $\phi_{\sigma,i}$ are *Steinberg* (i.e. $\phi_{\sigma,i}(x,y)=0$ makes $\{ x,y\}=0$ in $K_2$). We will say that $\phi_{\sigma}$ is $\QQ$*-Steinberg* if its factors are Steinberg and defined over $\QQ$. For $n\geq 4$, it may not even be the case that weak temperedness implies the existence of pointwise lifts $\Xi^{\lambda} \in\mathrm{CH}^n(\widetilde{X_{\phi}^{\lambda}},n)\otimes \QQ$ of $\{\ux\}|_{X_{\phi}^{\lambda,\times}}$ to a smoothing. This is always true for $n\leq 3$; for $n=4,5$ it holds if (for instance) all boundary strata are rational and defined over a totally real number field. (See [@DK $\S$3] for other refinements.) For particular families one can certainly check temperedness in any dimension. However, in the remainder of this section, we prefer to restrict to the case $n\leq 3$. Let $\mathcal{K}_{\phi} :=\rho(T_{\{\ux\} })\subset \PP^1$ and $U_{\phi} := \PP^1 \setminus \mathcal{K}_{\phi}$ (as in $\S$\[S3\]). We are interested in conditions under which not only is $\phi$ strongly tempered, but where the (unique) single-valued lift $\tilde{\nu}_{\phi}^{\lambda}$ over $U_{\phi}$ is given by fiberwise restrictions of$$R_{\{\ux\} } = \sum_{j=1}^n (-1)^{(n-1)(j-1)} (2\pi \mathbf{i})^{j-1}\log(x_j) \frac{dx_{j+1}}{x_{j+1}} \wedge \cdots \wedge \frac{dx_n}{x_n} \delta_{T_{x_1} \cap \cdots \cap T_{x_{j-1}}}$$up to push-fowards of currents from $Y_{\phi}$ nd coboundary currents. The biggest nuisance turns out ot be the correction term $(2\pi\mathbf{i})^n\delta_{\Gamma^{\lambda}}$ added to $R_{\Xi^{\lambda}}$ to produce a closed current; part of this correction is (the current of integration over) a chain on $X_{\phi}^{\lambda}$ bounding on an unknown $(n-2)$-cycle on $Y_{\phi}$. In order that this not contribute to $\tilde{V}_{\phi}$, we have to assume that $H_{n-2}$ of $Y_{\phi}$ (or part of it) vanishes. To make these conditions relatively weak, we define some “bad” subsets of the boundary. Let $\mathscr{I}_{\phi}\subset Y_{\phi}$ be the generic $\Delta$-irregularity locus of $\lambda -\phi$ — that is, the closure of the union of all singularities and nonreduced components of all $Y_{\sigma}^{\times}$ (computable by taking partials of $\phi_{\sigma}$). Let $\mathscr{A}\subset \mathscr{I}$ be the locus of generic singularities of $X_{\phi}^{\lambda}$.[^11] Denoting by $\mathbb{I}_{\Delta}$ the intersection with $\DD_{\Delta}$ of the closure of the locus $\cup_{i=1}^n \{ x_i = 1\}$ in $\PP_{\Delta}$, we write $\mathscr{J}_{\phi}$ for the union of all components of $Y_{\phi}$ not contained in $\mathbb{I}_{\Delta}$, and (for $n=3$) not of “Steinberg type” $\{ x_{i_1}+ x_{i_2} = 1,\, x_{i_3} = 0 \text{ or }\infty\}$. \[th3\] Let $\phi \in \ZZ[x_1^{\pm1},\ldots ,x_n^{\pm1}]$ be a reflexive Laurent polynomial for $n=1$, $2$, or $3$. 1. Assume that edge polynomials \[resp. 2-face polynomials\] of $\phi$ are cyclotomic \[resp. $\QQ$-Steinberg\], and that $\mathscr{I}\subset \mathbb{I}_{\Delta}$ and $\mathscr{I}\cap \mathscr{J}\subset \mathscr{A}$; if $n=3$, assume $\mathscr{A}$ consists only of $A_1$ singularities and lies in $\mathbb{I}_{\Delta}\cap \text{sing}(\DD_{\Delta})$. Then $\phi$ is tempered. 2. If $n=2$, assume in addition that $\mathscr{J}$ is one point. If $n=3$, assume in addition that $H_1(\mathscr{J}\setminus \mathscr{J}\cap \mathscr{A})=\{0\}$. Then $\phi$ is strongly tempered, with single-valued THNF $$\label{28-1} \tilde{V}_{\phi}(\lambda) = \int_{X_{\phi}^{\lambda}} R\{\ux\} \wedge \tilde{\omega}_{\phi}^{\lambda}$$ over $U_{\phi}$. \(a) is a restatement of part of [@DK Thm. 3.8]. The strong-temperedness part of (b) recapitulates [@DK Prop. 4.12]. We now show that is consistent with the lift constructed in the proof of \[loc. cit.\]. Assuming first that $\mathscr{A}=\emptyset$, we can extend $\{\ux\}$ to a closed precycle on $Z^n(\X_{\phi}^-\setminus \mathscr{J}\times \mathbb{A}^1,n)$ by taking Zariski closure and (for $n=3$) adding precycles of the form $\left( t,1-t,1-\tfrac{x_{i_1}}{t}\right)_{t\in\PP^1}$ over the “Steinberg type” locus $\mathscr{S}$. As in \[loc. cit.\], one then has $\gamma\in Z^n(\X_{\phi}^- \setminus\mathscr{J}\times \mathbb{A}^1,n+1)$ and (closed) $\Xi\in Z^n(\X_{\phi}^-,n)$ with $\Xi|_{\X_{\phi}^-\setminus \mathscr{J}\times\mathbb{A}^1} = \{\ux\}+\imath^{\mathscr{S}}_*Z + \d\gamma$ $\implies$ $R_{\Xi}|_{\X_{\phi}^-\setminus \mathscr{J}\times\mathbb{A}^1} = R_{\{\ux\}} + \tfrac{1}{2\pi\mathbf{i}}dR_{\gamma} +\imath^{\mathscr{S}}_* R_Z - (2\pi \mathbf{i})^n\delta_{T_{\gamma}}$ and $T_{\Xi} \equiv \d \overline{T_{\gamma}} + \tau$ mod $\rho^{-1}(\mathcal{K})$, with $\tau$ supported on $\mathscr{J}\times (\PP^1,\mathcal{K})$. Arguing as in \[loc. cit.\], our hypotheses give that $\tau \underset{\text{hom}}{\equiv} 0$ on $\mathscr{J}\times(\PP^1,\mathcal{K})$. Writing this as $\tau\equiv\d C_{\mathscr{J}}$, the “lift” of $R_{\Xi}$ given there is $R_{\Xi}' = R_{\Xi} + (2\pi\mathbf{i})^n \delta_{\overline{T_{\gamma}}+C_{\mathscr{J}}}$, so that $$\left. R_{\Xi} ' \right|_{\X_{\phi}^- \setminus \mathscr{J}\times\mathbb{A}^1} = R_{\{\ux\}} + \tfrac{1}{2\pi\mathbf{i}} dR_{\gamma}.$$Writing $X_{\phi,\epsilon}^{\lambda}$ for $X_{\phi}^{\lambda}$ minus a small tubular neighborhood of $\mathscr{J}$, this leads to $$\tilde{V}_{\phi}(\lambda) = \int_{X_{\phi}^{\lambda}} R_{\Xi}'\wedge \tilde{\omega}_{\phi}^{\lambda} = \lim_{\epsilon \to 0} \left\{ \int_{X_{\phi,\epsilon}^{\lambda}} R_{\{\ux\}}\wedge \tilde{\omega} + \tfrac{1}{2\pi\mathbf{i}} \int_{\d X_{\phi,\epsilon}^{\lambda}} R_{\gamma}\wedge\tilde{\omega} \right\}$$hence , provided $\lim_{\epsilon\to 0}\int_{\d X_{\phi,\epsilon}^{\lambda}} R_{\gamma}\wedge \tilde{\omega}=0.$ Viewing $\gamma$ as a family of curves in $\square^{n+1}$ over $\X_{\phi}$, the $(n-2)$-current $R_{\gamma}$ is the push-forward of $R_{\{\uz\}}$ ($\uz = $ coordinates on $\square^{n+1}$). If $n=2$, and $n$ is a local holomorphic coordinate about $\mathscr{J}$, one checks that $R_{\gamma}$ is locally $\mathcal{O}(\log^2 u)$; clearly $\oint_{|u|=\epsilon}\log^2 (u)du\to 0$. For $n=3$ (with $|(u)_0|=\mathscr{J}$ locally), $R_{\gamma}$ is the current of integration over a 3-chain times a locally-$\mathcal{O}(\log^2(u))$ function, with the same result. Finally, in the event that $\mathscr{A}$ is nonempty, and $\widetilde{\X_{\phi}^-}$ the blowup along $\mathscr{A}\times\PP^1$ (with exceptional divisor $\mathcal{E}$, one replaces all complexes (of higher Chow cycles and currents) on $\X_{\phi}^-$ by cone (double-)complexes for the morphism $\mathcal{E}\longrightarrow \mathscr{A}\amalg \widetilde{\X_{\phi}^-}$. The assumption that $H_1(\mathscr{J}\setminus\mathscr{J}\cap\mathscr{A})=\{0\}$ allows $C_{\mathscr{J}}$ to be drawn so as to avoid $\mathcal{E}$. This result is likely far from optimal, but suffices for the applications in the next section. \[cor3\] For $\phi$ as in Theorem \[th3\](b), if $\phi(-\ux)$ has all positive coefficients, then$$\tilde{V}_{\phi}(0)=\int_{\psi} R_{\{\ux\}}|_{X_{\lambda}^0}$$for any $(n-1)$-cycle $\psi\subset X_{\lambda}^0 \setminus \mathscr{J}$ representing $Q_0 \hat{\vf}_{n-1} \in H_{n-1}(X_{\lambda}^0)$.[^12] The additional hypothesis on $\phi$ ensures that $\lambda=0$ belongs to $U_{\phi}$. Now $Z^0 = \text{Res}_{X^0}(\mathcal{I}^*\Xi)=\mathcal{I}^*\text{Res}_{X_0}(\Xi) = \mathcal{I}^* Z_0$ has $\Omega_{Z^0} = \mathcal{I}^* \Omega_{Z_0} = (2\pi\mathbf{i})^{n-1} \mathcal{I}^* \omega_0 = (2\pi\mathbf{i})^{n-1}\tilde{\omega}^0$, while $T_{Z^0}$ is an $(n-1)$-cyle with $dR_{Z^0} = \Omega_{Z^0} - (2\pi\mathbf{i})^{n-1}\delta_{T_{Z^0}}$ $\implies$ $[T_{Z^0}]=[\tilde{\omega}^0]=Q_0\hat{\vf}_{n-1}$ in homology. So for $\psi$ as above, there exists an $(n-2)$-current $\mathcal{R}$ on $X^0$ with $d\mathcal{R} = \omega^0 - \delta_{\psi}$, with closed hence (by hypothesis on $\mathscr{J}$) exact restriction to $\mathscr{J}$. We may therefore assume that $\mathcal{R}$ (is of intersection type with respect to $\mathscr{J}$ and) pulls back to $0$ on $\mathscr{J}$, so that $\lim_{\epsilon\to 0}\int_{\d X_{\phi,\epsilon}^0} R_{\{\ux\}}\wedge \mathcal{R} = 0$. It now follows that $\tilde{V}_{\phi}(0) = \lim_{\epsilon \to 0} \int_{X_{\phi,\epsilon}^0} R_{\{\ux\}}\wedge \tilde{\omega}_{\phi}^0 = \lim_{\epsilon\to 0} \int_{X_{\phi,\epsilon}^0} R_{\{\ux\}} \wedge \delta_{\psi}$ as claimed. Examples and near-examples {#S5} ========================== Here we record some Laurent polynomials that satisfy the conditions of Theorems \[th2\] and \[th3\], as well as a few which stray close enough to warrant attention. $n=1$ ----- Let $b\in \ZZ_+$, $a=2b+1$, and set $$\phi(x):=-x+a+\frac{1-a^2}{4x}.$$$X_{\phi}^{\lambda}$ is a pair of points $\{ p_+^{\lambda},p_-^{\lambda}\}$ which are distinct unless $\lambda$ is a root of $P_a(\lambda):=\lambda^2 -2a\lambda + 1$, and $\X_{\phi}$ has involution[^13]$$(x,\lambda)\;\longmapsto \; \left( \lambda^{-1}x + \tfrac{1}{2}(1-\lambda^{-1})(a+1),\lambda^{-1}\right).$$We have $p_{\pm}^{\lambda}=\tfrac{1}{2}(a-\lambda \pm P_a(\lambda)^{1/2})$, and the 0-form$$\tilde{\omega}_{\phi}^{\lambda} = \lambda^{-1}\omega^{\lambda}_{\phi} = \frac{\pm 1}{2\sqrt{P_a(\lambda)}}\;\;\;\text{on}\;\;\;p_{\pm}^{\lambda}$$has period $$A(\lambda)=\int_{p_+^{\lambda}-p_-^{\lambda}}\tilde{\omega}_{\phi}^{\lambda} = \frac{1}{\sqrt{P_a(\lambda)}}.$$The regulator 0-current $R_{\lambda}=\log(p_{\pm}^{\lambda})$ on $p_{\pm}^{\l}$, and so the HNF is$$\tilde{V}_{\phi}(\l) = \frac{1}{\sqrt{P_a(\l)}}\log (p_+^{\l}/p_-^{\l}).$$Since $\mathfrak{r}_{\phi}=a-\sqrt{a^2 - 1}<e^{-1}$ for all $b\geq 1$, we conclude that $$\tilde{V}_{\phi}(0)=\log\left( \frac{b+1}{b}\right)\notin\QQ.$$ $n=2$ ----- Let $$\phi(x_1,x_2):= x_1^{-1}x_2^{-1} (1-x_1)(1-x_2)(1-x_1-x_2);$$the picture$$\includegraphics[scale=0.5]{apery1.pdf}$$indicates the (reflexive) Newton polytope $\Delta$ and $X^{\lambda}_{\phi}\subset \PP_{\Delta}$. The green \[resp. red\] dots represent $Y_{\phi}\setminus \mathscr{J}$ \[resp. $\mathscr{J}$\]; edge polynomials are $x^{\sigma}-1$ or $(x^{\sigma}-1)^2$, so $\phi$ is strongly tempered. The singular fibers of $\X_{\phi}$ are over $0$, $\infty$, and $t_{\pm} = \tfrac{-11\pm 5\sqrt{5}}{2}$, with $\mathfrak{r}_{\phi} = t_+ <e^{-2}$ (and $\X_{\phi,t_{\pm}}$ of type $I_1$), while$$D_{\phi}(t) = (t^2 + 11t-1)\delta_t^2 + t(2t+11)\delta_t + t(t+3)$$is integral.[^14] Involutivity is ensured by the (order 4) automorphism$$\mathcal{I}:\;(x_1,x_2,t)\longmapsto \left( \tfrac{x_1}{x_1 -1},\tfrac{1-x_2}{1-x_1-x_2},-\tfrac{1}{t}\right) ,$$while the facile 2-cover given by$$(1-\xi^2) + \xi y_1 + \xi^2 y_2 -\xi y_1 y_2 - \xi y_1^{-1} y_2^{-1} = 0$$maps down by$$\mathcal{J}:\; (y_1,y_2,\xi)\longmapsto \left( \tfrac{y_1}{\xi},\tfrac{y_1 - \xi}{y_1 y_2 -\xi},\xi^2 \right) .$$To compute the special value, note that $X^0 = \{ x_1 =1 \} \cup \{ x_2 =1 \} \cup \{ x_1 + x_2 = 1\}$, with $\psi$ going “once around” the figure. Furthermore, $R_{\{\ux\}} = \log(x_1 )\tfrac{dx_1}{x_1} - 2\pi\mathbf{i} \log (x_2)\delta_{T_{x_1}}$ vanishes on the first two components. Parametrizing the remaining component of $-\psi$ by $[0,1]\ni s \mapsto (1-s,s)$,$$\tilde{V}_{\phi}(0) = -\int_0^1 \log(1-s)\tfrac{ds}{s} = \mathrm{Li}_2 (1) = \tfrac{\pi^2}{6} \notin \QQ.$$ In view of the results of Zagier’s search for recurrencies of Apéry type [@Za], it seems likely that this $\phi$ is the unique example for $n=2$ that satisfies the conditions of Proposition \[prop\]. One can match tempered Laurent polynomials (hence higher normal functions) to the sporadic examples of \[op. cit.\], but outside case “D” (just treated), both the bound on $\mathfrak{r}_{\phi}$ and involutivity fail. For instance, case “E” of \[op. cit.\] is $\phi(\ux) = (x_1 + x_1^{-1})(x_2 + x_2^{-1}) + 4$; this has $\tilde{V}_{\phi}(0) = 2G$ (Catalan), but $\mathfrak{r}_{\phi} C = \tfrac{1}{8}\cdot 32 = 4$ (too big), and the Kodaira fiber types at $\lambda_0$ and $C/\lambda_0$ (or $0$ and $\infty$) don’t match, so that $\V_{\phi} \ncong \mathcal{I}^* \V_{\phi}$. This non-involutivity is not a problem for the approach via modular forms, which gives a different means for obtaining period expansions about any cusp; we are trading off this advantage for (at least in principle) the ability to treat non-modular families in higher dimension. One other “near-example” related to Catalan’s constant arises from work of Zudilin [@Z3], who found an Apéry-like recurrence with rational solutions $a_m, b_m$ whose ratios $b_m/a_m$ converge rapidly to $G$. With some work, one can write the generating series $\sum_{m\geq 0} (a_m G - b_m) \lambda^m$ as a normal function associated to a higher cycle on a family of open[^15] genus-9 curve, which are branched 4:1 covers of the “baby Apéry” family of elliptic curves above! By construction, this has $V(0)=G$. $n=3$ ----- The Newton polytope $\Delta$ of $$\begin{gathered} \label{35-1} \phi(x_1,x_2,x_3)=\\ x_1^{-1}x_2^{-1}x_3^{-1}(x_1 - 1)(x_2 - 1)(x_3 - 1)(1-x_1 -x_2 +x_1 x_2 -x_1 x_2 x_3)\end{gathered}$$ and its dual are $$\includegraphics[scale=0.5]{apery2.pdf}$$showing reflexivity; maximal triangulation adds the green edge. The edge and facet polynomials are products of $(x^{\sigma}_i - 1)$, $(x_i^{\sigma}-1)^2$, and $(1-x_1^{\sigma} \pm x_2^{\sigma})$, and $\mathscr{A}$ (red), $\mathscr{J}$ (blue), and $\mathscr{I}$ (green) are as depicted:$$\includegraphics[scale=0.5]{apery3.pdf}$$In particular, $\mathscr{J}\setminus\mathscr{J}\cap\mathscr{A}$ is two copies of $\mathbb{A}^1$ attached at a point, and we conclude that $\phi$ is strongly tempered. Singular fibers are at $0$, $\infty$, $t_{\pm}=(\sqrt{2}\pm 1)^4$, and $1$; the last of these does not contribute to monodromy of $\VV_{\phi}$, and so $\delta_{\phi} = 2$, while $\mathfrak{r}_{\phi}=t_- < e^{-3}$. To see that the generic Picard rank is $19$, one can use a torically-induced elliptic fibration (cf. [@Ke $\S$2]). The Picard-Fuchs operator is$$D_{\phi}(t) = (t^2 -34t+1)\delta_t^3 + 3t(t-17)\delta_t^2 + 3t(t-9)\delta_t + t(t-5),$$and the $\{a_m\}$ the famous Apéry sequence $a_m = \sum_{k=0}^m {\binom{m}{k}}^2 {\binom{m+k}{k}}^2 = 1,5,73,1445,\ldots .$ Changing coordinates by $\sx_i = \tfrac{x_i}{x_i - 1}$ brings $1-t\phi(\ux)=0$ into the form studied by Beukers and Peters [@BP]. By the results of Peters and Stienstra [@PS], $\X_{\phi,t}$ thus has a (facile) 2-cover by the *Fermi family*$$\rx_{\xi} := \left\{ \xi \sum_{i=1}^3 (y_i + y_i^{-1}) + 1 + \xi^2 =0 \right\}.$$It also has an involution, by $$\mathcal{I}:\; (x_1,x_2,x_3,t)\longmapsto \left( \tfrac{x_3}{x_3 -1},\tfrac{(x_1 - 1)(x_2 -1 )}{1-x_1 -x_2 + x_1 x_2 -x_1 x_2 x_3 },\tfrac{x_1}{x_1 - 1},\tfrac{1}{t} \right) .$$ The 2-current $$R_{\{\ux\}} = \log(x_1)\tfrac{dx_2}{x_2} \wedge \tfrac{dx_3}{x_3} + (2\pi \mathbf{i})\log(x_2) \tfrac{dx_3}{x_3} \delta_{T_{x_1}} + (2\pi \mathbf{i})^2 \log(x_3) \delta_{T_{x_1}\cap T_{x_2}}$$vanishes on the components $\{ x_i = 1\}$ of $X_{\phi}^0$. The piece of $\psi$ on the remaining component $x_3 = \tfrac{(1-x_1)(1-x_2)}{x_1x_2}$ is parametrized by $0\leq r\leq s\leq 1\; \longmapsto \;(1-r,s,\tfrac{r(1-s)}{s(1-r)})$, and so \_(0) &= \_[0rs1]{} (1-r)(s)()\ &= -\_0\^1 (1-r)(r)()\ &= -2\_0\^1 (1-r)(r) (r)\ &= 2 \_[k1]{} \_0\^1 r\^[k-1]{} (r) dr = -2(3) . There are at least three “near-examples” for $n=3$, identified in [@dS] (and closely related to [@Go]), which satisfy all the criteria in Proposition \[prop\] that we have checked, except for the bound: writing $\phi_{\rm{I}}$ for , these are \_[[II]{}]{} &= (1-x\_1 - x\_2 -x\_3)(1-x\_1\^[-1]{})(1-x\_2\^[-1]{})(1-x\_3\^[-1]{})\ \_[[III]{}]{} &= (x\_1 + x\_2 +x\_3)(-1+x\_1\^[\^[-1]{}]{} + x\_2\^[\^[-1]{}]{} + x\_3\^[\^[-1]{}]{} - x\_1\^[\^[-1]{}]{}x\_2\^[\^[-1]{}]{} - x\_1\^[\^[-1]{}]{}x\_3\^[\^[-1]{}]{} - x\_2\^[\^[-1]{}]{}x\_3\^[\^[-1]{}]{})\ \_[[IV]{}]{} &= (1-x\_1 -x\_2 -x\_3)(1-x\_1\^[-1]{}-x\_2\^[-1]{}-x\_3\^[-1]{}) . They are reflexive, tempered, and involutive, with $C_{\rm{I}}=1$, $C_{\rm{II}}=16$, $C_{\rm{III}}=-27$, and $C_{\rm{IV}}=64$; while $\tilde{V}(0) \underset{\QQ^{\times}}{\sim} \zeta(3)$ except for $\phi_{\rm{III}}$, where $\tilde{V}(0) \underset{\QQ^{\times}}{\sim} L(\chi_{-3},3)\underset{\QQ^{\times}}{\sim} \pi^3 /\sqrt{3}$. For $\mathfrak{r}_{\phi}|C|e^3$ we obtain $\approx 0.59$, $13.78$, $27.97$, resp. $80.34$, which satisfies the required bound ($<1$) only in the first case. Higher dimension? {#S5.4} ----------------- Here we propose two sources for examples with $n\geq 4$, if one is prepared to weaken the hypotheses as in the last paragraph of $\S$\[S3.last\]. In both cases, the Laurent polynomials considered, while not in general reflexive, all have Newton polytope $\Delta \subset [-1,1]^n$ having the origin as unique integer interior point. Details, proofs, and further developments will appear elsewhere. Define the *VZ polynomials* $\{\phi_n\}$ inductively by $\psi_1 = 1$,$$\psi_n (x_1,\ldots,x_n) :=x_1\cdots x_n + (1-x_n) \psi_{n-1}(x_1,\ldots,x_{n-1}),$$$$\phi_n(\ux) := (1-x_1^{-1})\cdots (1-x_n^{-1}) \psi_n (\ux) ;$$they are obtained (by substituting $\sx_i :=\tfrac{x_i}{x_i -1}$) from denominators of integrals first considered by Vasilyev [@Va] and Zudilin [@Z2] in their works on linear forms in zeta values. For $n=2$ and $3$, this recovers (up to inversion and permutation of coordinates) the Apéry polynomials above. For $n=5$, we expect that $\phi_5$ is strongly tempered, and conjecture that Hodge numbers of $\V_{\phi}$ are $(1,1,2,1,1)$. This would result in *two* invariant periods $A(\lambda) = 1+\sum_{m\geq 1} a_m\lambda^m$, $B(\lambda) = \lambda +\sum_{m\geq 2} b_m \lambda^m$ about the maximal unipotent monodromy point, as $N$ has two primitive classes. Writing $V(\lambda)$ for the HNF, one then expects $\mathfrak{r}_{\phi}<e^{-5}$, and C() &:= -V(ł) + A(ł)V(0) + B(ł)( -V(0) A’(0) + V’(0))\ &= \_[m2]{} c\_m ł\^m to satisfy an inhomogeneous equation as above. Combining this with Vasilyev’s results, one would conclude that$$V(\l)=\sum_{m\geq 0} \left( 2a_m \zeta(5) + b_m \zeta(3) - c_m \right) ,$$where $a_m,L_m^2 b_m,L_m^5 c_m \in \ZZ$, with the innocuous consequence that “at least one of $\zeta(3)$ and $\zeta(5)$ is irrational.” Recent work of F. Brown [@Br] provides another expected source of interesting Laurent polynomials. Given a permutation $\pi \in \mathfrak{S}_{n+3}$, write formally $$\theta_{\pi}(z_1,\ldots,z_{n+3}):= \prod_{i\in \ZZ/(n+3)\ZZ} (z_{\pi(i)}-z_{\pi(i+1)}),$$$$x_j^{\pi}:=-\text{CR}\left( z_{\pi(1)},z_{\pi(n+2-j)},z_{\pi(n+3-j)},z_{\pi(n+4-j)}\right),$$where $j=1,\ldots,n$, and $$\Omega_{\pi} :=\frac{dx_1^{\pi}}{x_1^{\pi}}\wedge \cdots \wedge \frac{dx_n^{\pi}}{x_n^{\pi}} ;$$if $\pi=\text{Id}$ then we drop the sub- and superscripts. Now let $\sigma \in \mathfrak{S}_{n+3}$ be a *convergent* permutation in the sense of \[op. cit.\]; namely, we assume that for any $i\in \ZZ/(n+3)\ZZ$ and $2\leq k \leq n+1$, $\{ \sigma(i),\ldots,\sigma(i+k)\}$ is not a consecutive sequence of integers mod $(n+3)$. It turns out that $\theta_{\sigma}(\uz)/\theta(\uz)$ can be written as a Laurent polynomial $\phi_{\sigma}(x_1,\ldots,x_n)$ (with Newton polytope $\Delta_{\sigma}$), and the *basic cellular integrals* on $\mathcal{M}_{0,n}$ of \[op. cit.\] become the integrals$$I_{\sigma}^{(k)} := \int_{T_{\{\ux\}}}\phi_{\sigma}(\ux)^{-k} \Omega_{\sigma}$$on $\PP_{\Delta_{\sigma}}$. Defining $X_{\sigma}^{\l}\subset \PP_{\Delta_{\sigma}}$ by $\l=\phi_{\sigma}(\ux)$, the generating series V\_(ł)&:= (2)\^[1-n]{}\_[k0]{} I\_\^[(k)]{}\ &=(2)\^[1-n]{}\_[T\_[{}]{}]{} may be rewritten (using integration by parts) in the form $$=\int_{X_{\sigma}^{\l}} \left. R_{\{\ux\}} \right|_{X_{\sigma}^{\l}} \wedge \tilde{\omega}^{\l} ,$$which is a truncated HNF under a strong temperedness hypothesis. Finally, involutivity may be arranged via the additional hypothesis that $\sigma^{-1} = \pi_1 \circ \sigma \circ \pi_2$, with $\pi_1,\pi_2$ belonging to the dihedral group $D_{n+3}$. [?????]{} G. Almkvist, C. van Enckevort, D. van Straten, and W. Zudilin, *Tables of Calabi-Yau equations*, arXiv:math/0507430; database at http://www2.mathematik.uni-mainz.de/CYequations/db. V. Batyrev, *Dual polyhedra and mirror symmetry for Calabi-Yau hypersurfaces in toric varieties*, J. Algebraic Geom. 3 (1994), no. 3, 493-535. F. Beukers, *Irrationality of $\pi^2$, periods of an elliptic curve and $\Gamma_1(5)$*, in “Diophantine approximations and transcendental numbers (Luminy, 1982)”, 47-66, Prog. Math. 31, Birkhäuser, Boston, 1983. F. Beukers and C. Peters, *A family of $K3$ surfaces and $\zeta(3)$*, J. Reine Angew. Math. 351 (1984), 42-54. S. Bloch, M. Kerr and P. Vanhove, *A Feynman integral via higher normal functions*, Compos. Math. 151 (2015), no. 12, 2329-2375. F. Brown, *Irrationality proofs for zeta values, moduli spaces and dinner parties*, Mosc. J. Comb. Number Theory 6 (2016), no. 2-3, 102-165. G. Brown, A Kasprzyk et al, “Graded Ring Database”, at http://www.grdb.co.uk. T. Coates, A. Corti, S. Galkin, V. Golyshev and A. Kasprzyk, “Fano varieties and extremal Laurent polynomials”, research blog and database, http://www.fanosearch.net. G. da Silva Jr., *On the arithmetic of Landau-Ginzburg model of a certain class of threefolds*, preprint, arXiv:1601.00990v2. P. del Angel, C. Doran, J. Iyer, M. Kerr, J. Lewis, S. Müller-Stach, and D. Patel, *Specialization of cycles and the $K$-theory elevator*, preprint, arXiv:1704.04779. C. Doran and M. Kerr, *Algebraic $K$-theory of toric hypersurfaces*, Commun. Number Theory Phys. 5 (2011), no. 2, 397-600. C. Doran and A. Malmendier, *Calabi-Yau manifolds realizing symplectically rigid monodromy tuples*, preprint, arXiv:1503.07500. J. Duistermaat and W. van der Kallen, *Constant terms in powers of a Laurent polynomial*, Indag. Math. N.S. 9 (1998), 221-231. S. Galkin, *Apéry constants of homogeneous varieties*, arXiv:1604.04652. S. Galkin, V. Golyshev and H. Iritani, *Gamma classes and quantum cohomology of Fano manifolds: gamma conjectures*, Duke Math. J. 165 (2016), no. 11, 2005-2077. S. Galkin and H. Iritani, *Gamma conjecture via mirror symmetry*, preprint, arXiv:1508.00719. M. Green, P. Griffiths, and M. Kerr, *Some enumerative global properties of varietions of Hodge structures*, Mosc. Math. J. 9 (2009), no. 3, 469-530. V. Golyshev, *Deresonating a Tate period*, arXiv:0908.1458. V. Golyshev and D. Zagier, *Proof of the gamma conjecture for Fano 3-folds of Picard rank 1*, Izv. Math. 80 (2016), no. 1, 24-49. S. Hosono, B. Lian and S.-T. Yau, *GKZ-generalized hypergeometric systems in mirror symmetry of Calabi-Yau hypersurfaces*, Comm. Math. Phys. 182 (1996), no. 3, 535-577. A. Huang, B. Lian, S.-T. Yau and X. Zhu, *Chain integral solutions to tautological systems*, Math. Res. Lett. 23 (2016), no. 6, 1721-1736. K. Iwasaki, H. Kimura, S. Shinemura, and M. Yoshida, “From Gauss to Painlevé: A modern theory of special functions”, Aspects of Mathematics, E16, Friedr. Vieweg & Sohn, Braunschweig, 1991. M. Kerr, *Indecomposable $K_1$ of elliptically fibered $K3$ surfaces: a tale of two cycles*, in “Arithmetic and geometry of $K3$ surfaces and C-Y threefolds (Laza, Schütt, Yui Eds.)”, 387-409, Fields Inst. Comm. 67, Springer, New York, 2013. M. Kerr and J. Lewis, *The Abel-Jacobi map for higher Chow groups, II*, Invent. Math. 170 (2007), 355-420. M. Kerr, J. Lewis, and S. Müller-Stach, *The Abel-Jacobi map for higher Chow groups*, Compos. Math. 142 (2006), 374-396. B. Lian and M. Zhu, *On the hyperplane conjecture for periods of Calabi-Yau hypersurfaces in $\PP^n$*, preprint, arXiv:1610.07125. C. Peters, *Monodromy and Picard-Fuchs equations for families of $K3$-surfaces and elliptic curves*, Ann. Sci. École Norm. Sup. (4) 19 (1986), no. 4, 583-607. C. Peters and J. Stienstra, *A pencil of $K3$ surfaces related to Apéry’s recurrence for $\zeta(3)$ and Fermi surfaces for potential zero*, in “Arithmetic of complex manifolds (Erlangen, 1988)”, 110-127, Lecture Notes in Math. 1399, Springer, Berlin, 1989. C. Robles, *Principal Hodge representations*, in “Hodge theory, complex geometry, and representation theory”, 259-283, Contemp. Math. 608, AMS, Providence, 2014. A. van der Poorten, *A proof that Euler missed . . . Apéry’s proof of the irrationality of $\zeta(3)$. An informal report*, Math. Intellegencer 1 (1978/79), no. 4, 195-203. D. Vasilyev, *On small linear forms for the values of the Riemann zeta-function at odd points*, Preprint no. 1 (558), Nat. Acad. Sci. Belarus, Institute Math., Minsk, 2001. D. Zagier, *Integral solutions of Apéry-like recurrence equations*, Groups and symmetries, 349-366, CRM Proc. Lecture Notes, 47, Amer. Math. Soc., Providence, RI, 2009. W. Zudilin, *One of the numbers $\zeta(5),\zeta(7),\zeta(9),\zeta(11)$ is irrational* (Russian), Uspekhi Mat. Nauk 56 (2001), no. 4 (340), 149-150; translation in Russian Math. Surveys 56 (2001), no. 4, 774-776. W. Zudilin, *Well-poised hypergeometric service for Diophantine problems of zeta values*, J. Théor. Nombres Bordeaux 15 (2003), no. 2, 593-626. W. Zudilin, *An Apéry-like difference equation for Catalan’s constant*, Electron. J. Combin. 10 (2003), Research Paper 14, 10 pp. [^1]: this is automatic if $\ul(\un)$ is effective. [^2]: When $\phi$ is not $\Delta$-regular, we also assume that no *non-generic* singularities of $X_{\phi}^{\lambda}$ along the base locus $X_{\phi}^{\lambda}\cap\DD_{\Delta}$ occur for values $\lambda\in\D_{\phi}^{*}$ (hence don’t affect the local system $\VV_{\phi}$), see below. [^3]: For $n=1$, the generic $X_{\phi,t}$ is a pair of points, and $H^{0}$[\[]{}resp. $H_{0}$[\]]{} means everywhere the augmentation cokernel [\[]{}resp. kernel[\]]{}, i.e. reduced (co)homology. Involutivity and reflexivity imply that $x\phi(x)$ is a quadratic polynomial with two distinct roots. [^4]: With no information on exponents, one would have $P_{\phi}(t)\delta_{t}^{n}+\sum_{k=0}^{n-1}\frac{F_{k}(t)}{P_{\phi}(t)^{n-k-1}}\delta_{t}^{k}$ ($F_{k}\in\CC[t]$); the existence of holomorphic solutions with orders $0$ thru $n-2$ about each root $\mathfrak{t}$ of $P_{\phi}$ forces $(t-\mathfrak{t})^{n-k-1}\mid F_{k}$. Note that if $n$ is even, the exponent $\frac{n-2}{2}$ is repeated. [^5]: Note that if we assume this only over $U_{\phi}\setminus U_{\phi}\cap\D_{\phi}$, the lift extends to $U_{\phi}$ anyway: the single-valuedness of $\tilde{\nu}$ on a punctured disk about $t_{0}$ means $\nu$ has no singularity at the center, so that $\tilde{\nu}$ uniquely extends to the whole disk. The value $\tilde{\nu}(t_{0})$ lies in $\ker(T-I)$ in the canonical extension $\H_{\phi}^{\text{lim}}$, which contains the image of $H^{n-1}(X_{\phi,t_{0}})$; see [@7K $\S$5]. [^6]: Here and below, we use the polarization $Q$ to make this identification (up to twist). [^7]: Recall that $\tilde{V}_{\phi}$ cannot be entire. [^8]: One may also show that $g_{\phi,n-1}(\lambda)=\frac{n}{2} P_{\phi} ' (\lambda)$, but we won’t need this. [^9]: The $\{ \hat{Y}_{\sigma},Y_{\sigma}^{\alpha}\}$ need not be irreducible, as $\lambda - \phi$ is not assumed $\Delta$-regular for generic $\lambda$. [^10]: Note that this is the definition of *temperedness* in [@DK $\S$3]. Our use of “tempered” here correlates to the property of “$\{\ux\}$ completes to a family of motivic cohomology classes” in \[op. cit.\]. [^11]: $\mathscr{A}$ is empty for $n=1$ or $2$. [^12]: $Q_0 \hat{\vf}_{n-1}$ is the class with intersection number $(\pm)1$ against $\hat{\vf}_0$; one should think of a membrane “stretched once around” $X_{\lambda}^0$. [^13]: of course, it is its own facile 1-cover! [^14]: The coefficients of the holomorphic solution are the “baby Apéry” sequence $a_m = \sum_{k=0}^m {\binom{m}{k}}^2 \binom{m+k}{k} = 1,\,3,\,19,\,147,\,\ldots .$ [^15]: The corresponding Laurent polynomial is neither reflexive nor tempered; 2 points are removed from each fiber.

ArXiv

--- abstract: 'We present new measurements of the flux power-spectrum $P(k)$ of the $z<0.5$ [H$\,$[i]{}]{} Lyman-$\alpha$ forest spanning scales $k \sim 0.001-0.1\, \mathrm{s \, km}^{-1}$. These results were derived from 65 far ultraviolet quasar spectra (resolution $R \sim 18000$) observed with the Cosmic Origin Spectrograph (COS) on board the Hubble Space Telescope. The analysis required careful masking of all contaminating, coincident absorption from [H$\,$[i]{}]{} and metal-line transitions of the Galactic interstellar medium and intervening absorbers as well as proper treatment of the complex COS line-spread function. From the $P(k)$ measurements, we estimate the [H$\,$[i]{}]{} photoionization rate ($\Gamma_{\rm HI}$) in the $z<0.5$ intergalactic medium. Our results confirm most of the previous $\Gamma_{\rm HI}$ estimates. We conclude that previous concerns of a photon underproduction crisis are now resolved by demonstrating that the measured $\Gamma_{\rm HI}$ can be accounted for by ultraviolet emission from quasars alone. In a companion paper, we will present constraints on the thermal state of the $z<0.5$ intergalactic medium from the $P(k)$ measurements presented here.' author: - | \ $^{1}$Physics Department, Broida Hall, University of California Santa Barbara, Santa Barbara, CA 93106-9530, USA\ $^{2}$Max-Planck-Institut für Astronomie, Königstuhl 17, 69117 Heidelberg, Germany\ $^{3}$International Max Planck Research School for Astronomy & Cosmic Physics at the University of Heidelberg, Germany\ $^{4}$Royal Observatories, Blackford Hill, Edinburgh EH9 3HJ, UK\ $^{5}$Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, CA 94720, USA\ $^{6}$Astronomy and Astrophysics, UC Santa Cruz, 1156 High St., Santa Cruz, CA 95064, USA\ $^{7}$Department of Astronomy, University of Massachusetts - Amherst, 710 North Pleasant Street, Amherst, MA 01003-9305, USA\ $^{8}$UCO / Lick Observatory, University of California, Santa Cruz, CA 95064, USA\ $^{9}$Department of Physics and Astronomy, University of California, Irvine, CA 92697, USA bibliography: - 'vikrambib.bib' date: 'Accepted 2019 January 29. Received 2019 January 14; in original form 2018 September 4' title: 'The Power Spectrum of the Lyman-$\alpha$ Forest at $z < 0.5$ ' --- \[firstpage\] [Intergalactic medium, UV background, Lyman-$\alpha$ forest, quasars]{} Introduction {#sec:intro} ============ The intergalactic medium (IGM), being the largest reservoir of the baryons in the Universe, plays an important role in the formation of cosmic structures. The ultraviolet (UV) radiation emanating from this cosmic structure photoionizes and heats the IGM. The trace amount of neutral hydrogen in the highly-ionized IGM imprints a swath of absorption lines on the spectra of background quasars known as the Lyman-$\alpha$ forest. Observations of the Lyman-$\alpha$ forest in a large sample of background quasar sightlines can probe the underlying density fluctuations in the IGM, measure its thermal state, and determine the amplitude of the UV ionizing background. The temperature and density of the photoionized IGM follow a tight power-law relation over two decades in the density, $T(\Delta)=T_0\,\Delta^{\gamma -1}$, where $\Delta=\rho/\bar{\rho}$ is the overdensity, $T_0$ is the temperature at mean density $\Delta=1$, and $\gamma$ is the power-law index. This power-law relation quantifies the thermal state of the IGM [@Hui97; @Theuns98; @McQuinn16]. While a wide variety of statistics have been applied to Lyman-$\alpha$ forest spectra with the goal of measuring its thermal state [@Haehnelt98; @Schaye99; @Theuns00; @Zaldarriaga01; @McDonald06; @Lidz10; @Becker11t; @Bolton12; @Rorai17; @Hiss18], the power spectrum of the transmitted flux is appealing for several reasons: 1) it is sensitive to a broad range of scales, in particular, the small-scales that encode information about the IGM thermal state, 2) it is thus capable of breaking strong parameter degeneracies, 3) systematics due to noise, metal-line contamination, resolution effects, and continuum errors impact it in well-understood ways, and 4) it can be described by a simple multivariate Gaussian likelihood enabling straightforward principled statistical analysis and parameter inference [@Irsic17; @Walther18a; @Walther19]. For these reasons, the power-spectrum has been used to constrain parameters such as the UV ionizing background intensity [@Gaikwad17a], alternate cosmology models with warm and fuzzy dark matter [@Viel08; @Viel13; @Garzilli17; @Irsic17], and cosmological parameters including neutrino masses [@McDonald06; @Palanque13ps; @PD15Neutrino; @Yeche17Neutrino; @Irsic17t]. There are many measurements of the flux power-spectrum at high redshifts [e.g @McDonald00; @Croft02; @Kim04ps; @Palanque13ps; @Irsic17; @Yeche17Neutrino; @Walther18a] where ground-based telescopes with medium or high resolution spectrographs were used to observe the Lyman-$\alpha$ forest redshifted to optical wavelengths. However to date, there are no measurements at low-redshifts $z < 1.6$ [but see @Gaikwad17a] where space-based observations are required because the redshifted Lyman-$\alpha$ transition lies in the UV below the atmospheric cutoff. Recently, large surveys [e.g, @Tumlinson13; @Danforth14; @Burchett15; @Borthakur15] have gathered a significant amount of Lyman-$\alpha$ forest spectra using the Cosmic Origin Spectrograph (COS) on-board the Hubble Space Telescope (HST) that can be used to measure the Lyman-$\alpha$ forest power-spectrum at low redshifts. The power spectrum at low redshifts is of particular interest since it provides another method for measuring the UV ionizing background, whereas previous work based on fitting the distribution of column densities argued for a ‘photon underproduction crisis’ [@Kollmeier14; @Wakker15]. Also, it can measure the thermal state of the low redshift IGM where long after the impulsive photoheating from reionization events is complete, theory robustly predicts the IGM should have cooled down to temperatures of $T_0\simeq 5000$ K at $z = 0$ [see e.g @McQuinn16; @Pheobe16t0]. Constraints on the low-redshift thermal state would thus provide an important check on our theoretical understanding of the IGM and shed light on the degree to which any other processes such as blazar heating, feedback from galaxy formation, or any other exotic physics can inject heat into the IGM. In an earlier study using the Space Telescope Imaging Spectrograph, @Dave01 obtained preliminary evidence that the $T_0$ is indeed about $5000$ K at $z\sim 0$, but they also found that the observed low-$z$ Lyman-$\alpha$ lines are not consistent with pure thermal broadening and may therefore also be broadened by some additional processes such as some type of feedback. Similar issues with line broadening are also reported by @Gaikwad17b [@Viel17] and @Nasir17. We can now revisit these issues with much larger sample. In this paper, we present new measurements of the Lyman-$\alpha$ forest flux power spectrum at $z<0.5$ in five redshift bins. We use high quality Lyman-$\alpha$ forest spectra (S/N per pixel $>10$) observed in 65 background quasars from the sample of @Danforth14. Combining these power spectrum measurements with state-of-the-art cosmological hydrodynamical simulations run with the Nyx code [@Almgren13; @Lukic15], we constrain the intensity of UV background $\Gamma_\mathrm{HI}$ at $z<0.5$. Our UV background measurements are consistent with recent studies [@Shull15; @Gaikwad17a; @Gaikwad17b; @Fumagalli17] which confirm that there is no crisis with UV photon production at $z<0.5$ and that the primary contributors to the UV background are quasars. In a companion paper (Walther et al. in prep.), we will use these power-spectrum measurements to constrain the thermal state of the IGM at $z<0.5$. The paper is organized as follows. In Section 2, we discuss the Lyman-$\alpha$ forest data. In Section 3, we describe our method to compute the power spectrum and present the resulting measurements. In Section 4, we discuss the implications of our power spectrum regarding the UV ionizing background and compare with previous work. In Section 5, we present our conclusions and discuss future directions. Throughout the paper, we adopt a flat $\Lambda$CDM cosmology with parameters $\Omega_m=0.319181$, $\Omega_b h^2 = 0.022312$, $h = 0.670386$, $n_s = 0.96$, and $\sigma_8 = 0.8288$ consistent with @Planck18. This cosmology is used both for our power spectrum measurements as well as for our cosmological hydrodynamical simulations. All of the distances quoted are comoving. {width="\textwidth" height="\textheight"} Data and Masking {#sec:data} ================ {width="\textwidth" height="\textheight"} We use high-quality medium-resolution ($R\sim 18000$, $\Delta v \sim 17$ km s$^{-1}$) quasar spectra obtained from HST/COS as a part of the large low-$z$ IGM survey by @Danforth14. This survey contains 82 quasar spectra at $z_{\rm em}<0.72$ observed in the wavelength range from $1130-1800$ Å, which covers the Lyman-$\alpha$ forest at $z<0.48$. The observations were obtained with the G130M and G160M gratings between year 2009 and 2013. @Danforth14 co-added individual spectra (combining both gratings whenever available), fitted continua, and identified nearly all individual absorption and emission lines. We use these continuum-fitted spectra along with the line catalog made publicly available by @Danforth14 as a high-level science product at Mikulski Archive for Space Telescopes[^1]. To calculate the Lyman-$\alpha$ forest flux power spectrum, P$(k)$, we mask all absorption lines arising from higher Lyman series transitions than Lyman-$\alpha$, the metal lines arising from intervening systems and the interstellar medium (ISM) of the Milky Way, all emission lines including geocoronal airglow emission, low quality data having S/N $<5$ per pixel[^2], and all gaps in the spectral coverage. To illustrate our masking procedure, we show five random chunks of spectra with different S/N in Fig. \[fig1\]. The shaded regions in Fig. \[fig1\] show our masks. In the Lyman-$\alpha$ forest, masking is critical to removing metal contamination and estimating the power spectrum correctly [see, e.g, @Walther18a]. We discuss the effect of not masking metals on the power spectrum in Appendix \[A\]. After masking, we restrict our further analysis to the rest-frame wavelength range between 1050 and 1180 Å in each quasar spectrum to avoid proximity zones of quasars ($\lambda>1180$ Å) and the excessive masking of data due to the presence of higher Lyman series forest lines ($\lambda<1050$Å). The redshift path covered by the unmasked Lyman-$\alpha$ forest data is shown in Fig. \[fig2\] where the quasars are ordered vertically by increasing redshift. Gaps in the horizontal lines show our masking. Vertically aligned gaps indicate masking due to strong Milky Way metal lines. A large gap at 1305 Å ($z\sim 0.07$) is due to a combination of an O [i]{} geocoronal airglow emission line and strong Si [ii]{} and O [i]{} absorption lines from the Milky Way (see e.g the middle panel of Fig. \[fig1\]). After masking and choosing the relevant wavelength range, we calculate the median S/N per pixel in the unmasked regions and impose a median S/N $>10$ (per pixel) cut to the 82 quasar spectra. Although in our power spectrum calculation we subtract the noise (see Section 3), we choose this S/N cut to ensure that we are not sensitive to systematic errors associated with how well we know the properties of the noise. Note that our S/N estimate is different than the default values provided in @Danforth14 who computed the S/N per resolution element over the entire spectrum. After applying our S/N cut, we are left with the Lyman-$\alpha$ forest of 66 quasars out of the initial 82. Individual values of this S/N per pixel are indicated in Fig. \[fig2\] via different colors. We then split the total redshift path covered by these 66 high S/N Lyman-$\alpha$ forest spectra into five redshift bins as illustrated by the vertical dashed lines in Fig. \[fig2\] and summarized in Table \[tab1\]. The first bin is chosen from $z=0.005$ to $0.06$ to remove any systematics arising from the extended wings of geocoronal Lyman-$\alpha$ emission line, which sets the lower limit of this redshift bin. The next three bins ($z=0.06-0.16$, $z=0.16-0.26$ and $z=0.26-0.36$) are chosen to have the same width ($\Delta z =0.1$) and also because the mean redshifts of the Lyman-$\alpha$ forests in these bins are nicely centered at $z=0.1, \,0.2$, and $0.3$ where we can compare our UV background measurements with previous studies [e.g @Shull15; @Gaikwad17a; @Gaikwad17b]. The last redshift bin ($z=0.36-0.48$) encloses the remaining redshift-path covered by the data. Finally, in each redshift bin we removed short spectra that span less than 10% of the redshift bin-width. This criterion removes one more spectrum from the last bin and we are left with a total of 65 quasar spectra shown in Fig. \[fig2\]. -------------- -------------- ------------ ------------ ------------------------ Redshift ${\bar z}^a$ Number Simulation Simulation bin of quasars redshift T$_0$ (K) and $\gamma$ 0.005 - 0.06 0.03 39 0.03 5033 1.73 0.06 - 0.16 0.10 34 0.10 5288 1.72 0.16 - 0.26 0.20 19 0.20 5652 1.71 0.26 - 0.36 0.30 13 0.30 6010 1.69 0.36 - 0.48 0.41 9 0.40 6368 1.68 -------------- -------------- ------------ ------------ ------------------------ : Details of the data and comparison simulation \ \[tab1\] Power Spectrum {#sec:ps} ============== In this section we discuss our method for measuring the power spectrum and present the measurement. Method {#sec3.1} ------ Once the data are prepared, we calculate the P($k$) following the method presented in @Walther18a. A brief description of the method is as follows. We first calculate the flux contrast of each spectrum in a redshift bin $\delta_F = (F- \bar{F})/\bar{F} $ where $\bar{F}$ is the mean flux of that spectral chunk in the bin. Then we use a Lomb-Scargle periodogram [@Lomb76; @Scargle82] to calculate the raw power spectrum $P_{\rm raw}(k)$. We subtract off the noise power $P_{\rm noise}(k)$ from $P_{\rm raw}(k)$ and divide the difference by the square of the window function $W(k, R)$ corresponding to the appropriate COS line spread function (LSF) to correct for finite resolution and pixelization [for more details see @Palanque13ps; @Walther18a]. Therefore, our final power spectrum is $$\label{eq1} P(k)=\bigg \langle \, \frac{P_{\rm raw}(k)-P_{\rm noise}(k) }{W^2(k, R)} \, \bigg \rangle .$$ We use the same logarithmic binning used in @Walther18a, and the average is performed over individual Lomb-Scargle periodograms of all Lyman-$\alpha$ forest chunks in each bin. We follow the standard normalization of the power spectrum, i.e. the variance in the flux contrast is $\sigma^2_{\delta_F}=\int_{-\infty}^{\infty}{dk \, P(k)/2\pi}$. The noise power within a bin is calculated using many realizations of Gaussian random noise[^3] generated from the error vector of each spectrum of the bin. For estimating the window function, we used the COS LSFs corresponding to the different gratings and lifetime positions[^4] depending on the observational parameters for each spectrum. In particular, spectra at $\bar{z}=0.2$ bin have overlapping contribution from both the G130M and the 160M grating. In this redshift bin, motivated by the co-addition routine used by @Danforth14 we take 1460 Å as the wavelength where the transition between the gratings happen.[^5] The COS LSF is quite different from the typical Gaussian LSFs that govern ground based spectroscopic observations and exhibits broad wings. In Appendix \[C\], we discuss the effect of incorrectly assuming a Gaussian LSF on the obtained $P(k)$ instead of using the correct non-Gaussian COS LSF. Finally, we calculate the uncertainties in P$(k)$, the diagonal elements of covariance matrix $C_{ii}\equiv\sigma_i^2$, by bootstrap resampling using $10^4$ random realizations of the dataset. Note that we do not have enough data to estimate the full covariance matrix $C_{ij}$. We recommend that researchers attempting to fit our power spectrum calculate the full covariance matrix by computing the correlation matrix $R_{ij} = C_{ij}\slash \sigma_i \sigma_j$ from their models, which can then be scaled by our diagonal elements to determine the full covariance matrix [see e.g. @Walther18a for details]. Our P($k$) measurements and the diagonal elements of covariance matrix are given in Table \[tab2\]. Results ------- {width="99.50000%" height="\textheight"} In Fig \[fig3\], we show our power spectrum measurement in different redshift bins, which reliably probes the power over the range $k\sim 5\times 10^{-4}$ to 0.15 s km$^{-1}$ (see Table \[tab2\]). In contrast to $z>2$, the relative ease of identifying and masking metal absorption lines at these low redshifts allows us to probe the small-scale (high-$k$) power-spectrum with negligible systematics due to non-Lyman-$\alpha$ absorption. We discuss the effect of not masking metals on the power spectrum in Appendix \[A\]. At all redshifts, our measured power spectrum shows a clear small scale cut-off at $k>0.03$ s km$^{-1}$. This cut-off is a signature of pressure smoothing and Doppler broadening of the Lyman-$\alpha$ forest. At small scales, the Lyman-$\alpha$ forest is supported by the thermal-pressure and does not follow the dark-matter density fluctuations [@Hui97; @Kulkarni15; @Onorbe17; @Rorai17; @Nasir17]. This pressure support, in addition to the Doppler broadening, smooths out the fluctuations in the Lyman-$\alpha$ forest flux and gives rise to the small-scale cut-off seen in the power spectrum [see e.g, @Peeples10; @Rorai13]. This cut-off is an important feature that probes the thermal state of the IGM [@Zaldarriaga01; @Walther19]. The amplitude of the power spectra at all $k$ is significantly smaller than that obtained at high-$z$ [see, e.g, @Walther18a]. This is because the density evolution of the Universe results in lower overall opacity in the low-z IGM giving rise to a thinner low-z Lyman-$\alpha$ forest. This reduced opacity also reduces the power at all scales. Nevertheless, we have obtained high precision measurements (15% at $z<0.2$, 25% at $z=0.3$ and 30% at $z=0.41$ over $0.001<k<0.1$ s km$^{-1}$ scale) of the power spectrum because of the large sample size. In contrast to high-$z$, the redshift evolution of the amplitude of the power appears quite shallow. This evolution is so weak that it is challenging to identify; for example, our large scale power ($k<0.02$ s km$^{-1}$) measurement at $\bar z =0.2$ is only slightly lower than that at $\bar z=0.1$. Such non-monotonic redshift evolution is unexpected but likely results from a combination of noise fluctuations and very shallow redshift evolution. Indeed, extremely shallow evolution in the power spectrum amplitude is not unexpected at low-$z$. The power spectrum $P(k,z)\propto \langle FF^\ast\rangle \propto e^{-2\tau(z)}$ scales with the evolution of the optical depth $\tau(z)$. Following the fluctuating Gunn-Peterson approximation (FGPA), we can write [@Gunn65; @Croft98] $$\label{eq2} \tau(z) \propto \, \Gamma_{\rm HI} ^{-1} \, T_0^{-0.7} \, n_{\rm H}^2,$$ where $\Gamma_{\rm HI}$ is the photoionization rate of [H$\,$[i]{}]{} from the UV background (UVB) and $n_{\rm H}$ is the hydrogen density. Using $\Gamma_{\rm HI}(z) \propto (1+z)^5$ [from @Shull15; @Gaikwad17a], $n_{\rm H}(z) \propto (1+z)^3$ from cosmological density evolution, and assuming a power law redshift evolution for $T_0(z) \propto (1+z)^{\beta}$, the opacity of the IGM should scale as $\tau \propto (1+z)^{1-0.7{\beta}}$. In the absence of any non-standard heating processes, theory predicts a cool-down of the IGM at low redshift suggesting $\beta>0$ [see e.g, @Pheobe16t0]. This suggests that $\tau(z)$ and the resulting amplitude of the power spectrum evolve slowly at low redshifts. However, note that the FGPA is not a good approximation for the low-$z$ IGM, at least for the power spectrum at large scales (small $k$), because it does not include the effects of shock-heated gas, as explained in Section 4 and Fig. \[fig4\]. Therefore, the redshift evolution of the amplitude of the large scale power is likely to be more complicated than the simple picture presented here. The flux power spectrum at redshifts $0.1\le z \le 0.4$ was also presented by @Gaikwad17a, however they split observed spectra into chunks of size 50 cMpc/h to compare with their simulation box size for the specific purpose of only evaluating the UV background. Also their method of calculating the power spectrum and the normalization is quite different from ours. For example, they fill the masks with added continuum and random noise and estimate the power spectrum of the flux ($F$) rather than the flux-contrast ($\delta F$). For these reasons, we are unable to compare with the @Gaikwad17a power spectrum measurements directly. However, we can compare with their UV background measurements, which use not only the power-spectrum but also the flux probability density function (PDF) and column-density distribution function [CDDF; @Gaikwad17b]. [1.2]{}[0.8]{} ${\bar z}$ $k (s/km)$ $kP(k)/\pi$ $\sigma_{kP(k)/\pi}$ ${\bar z}$ $k (s/km)$ $kP(k)/\pi$ $\sigma_{kP(k)/\pi}$ ------------ ----------------------- ----------------------- ----------------------- ------------ ----------------------- ----------------------- ----------------------- 0.03 3.937$\times 10^{-4}$ 1.874$\times 10^{-4}$ 7.550$\times 10^{-5}$ 0.20 2.794$\times 10^{-3}$ 1.298$\times 10^{-3}$ 2.129$\times 10^{-4}$ 0.03 4.390$\times 10^{-4}$ 1.228$\times 10^{-4}$ 5.164$\times 10^{-5}$ 0.20 3.551$\times 10^{-3}$ 1.892$\times 10^{-3}$ 2.949$\times 10^{-4}$ 0.03 7.825$\times 10^{-4}$ 3.137$\times 10^{-4}$ 9.136$\times 10^{-5}$ 0.20 4.466$\times 10^{-3}$ 2.219$\times 10^{-3}$ 4.231$\times 10^{-4}$ 0.03 8.805$\times 10^{-4}$ 2.267$\times 10^{-4}$ 6.759$\times 10^{-5}$ 0.20 5.636$\times 10^{-3}$ 2.348$\times 10^{-3}$ 3.451$\times 10^{-4}$ 0.03 1.170$\times 10^{-3}$ 5.812$\times 10^{-4}$ 1.833$\times 10^{-4}$ 0.20 7.117$\times 10^{-3}$ 2.612$\times 10^{-3}$ 5.248$\times 10^{-4}$ 0.03 1.528$\times 10^{-3}$ 5.083$\times 10^{-4}$ 1.726$\times 10^{-4}$ 0.20 8.969$\times 10^{-3}$ 3.162$\times 10^{-3}$ 5.229$\times 10^{-4}$ 0.03 1.874$\times 10^{-3}$ 5.534$\times 10^{-4}$ 1.232$\times 10^{-4}$ 0.20 1.130$\times 10^{-2}$ 3.278$\times 10^{-3}$ 5.957$\times 10^{-4}$ 0.03 2.322$\times 10^{-3}$ 8.393$\times 10^{-4}$ 1.814$\times 10^{-4}$ 0.20 1.420$\times 10^{-2}$ 3.584$\times 10^{-3}$ 6.451$\times 10^{-4}$ 0.03 2.915$\times 10^{-3}$ 9.651$\times 10^{-4}$ 1.717$\times 10^{-4}$ 0.20 1.788$\times 10^{-2}$ 3.943$\times 10^{-3}$ 5.999$\times 10^{-4}$ 0.03 3.670$\times 10^{-3}$ 1.265$\times 10^{-3}$ 2.399$\times 10^{-4}$ 0.20 2.255$\times 10^{-2}$ 3.809$\times 10^{-3}$ 5.501$\times 10^{-4}$ 0.03 4.523$\times 10^{-3}$ 1.161$\times 10^{-3}$ 2.277$\times 10^{-4}$ 0.20 2.838$\times 10^{-2}$ 3.695$\times 10^{-3}$ 5.647$\times 10^{-4}$ 0.03 5.691$\times 10^{-3}$ 1.576$\times 10^{-3}$ 2.837$\times 10^{-4}$ 0.20 3.576$\times 10^{-2}$ 2.974$\times 10^{-3}$ 4.456$\times 10^{-4}$ 0.03 7.217$\times 10^{-3}$ 2.721$\times 10^{-3}$ 6.823$\times 10^{-4}$ 0.20 4.503$\times 10^{-2}$ 2.129$\times 10^{-3}$ 2.573$\times 10^{-4}$ 0.03 9.028$\times 10^{-3}$ 2.958$\times 10^{-3}$ 5.895$\times 10^{-4}$ 0.20 5.667$\times 10^{-2}$ 1.648$\times 10^{-3}$ 2.354$\times 10^{-4}$ 0.03 1.137$\times 10^{-2}$ 3.248$\times 10^{-3}$ 5.704$\times 10^{-4}$ 0.20 7.131$\times 10^{-2}$ 1.138$\times 10^{-3}$ 2.070$\times 10^{-4}$ 0.03 1.431$\times 10^{-2}$ 3.343$\times 10^{-3}$ 6.273$\times 10^{-4}$ 0.20 8.975$\times 10^{-2}$ 7.136$\times 10^{-4}$ 1.244$\times 10^{-4}$ 0.03 1.792$\times 10^{-2}$ 3.142$\times 10^{-3}$ 4.820$\times 10^{-4}$ 0.20 1.130$\times 10^{-1}$ 3.709$\times 10^{-4}$ 2.719$\times 10^{-4}$ 0.03 2.250$\times 10^{-2}$ 2.907$\times 10^{-3}$ 4.776$\times 10^{-4}$ 0.30 5.576$\times 10^{-4}$ 4.136$\times 10^{-4}$ 2.045$\times 10^{-4}$ 0.03 2.837$\times 10^{-2}$ 2.966$\times 10^{-3}$ 3.903$\times 10^{-4}$ 0.30 8.690$\times 10^{-4}$ 5.103$\times 10^{-4}$ 1.624$\times 10^{-4}$ 0.03 3.578$\times 10^{-2}$ 2.438$\times 10^{-3}$ 3.273$\times 10^{-4}$ 0.30 1.126$\times 10^{-3}$ 6.460$\times 10^{-4}$ 2.556$\times 10^{-4}$ 0.03 4.503$\times 10^{-2}$ 2.045$\times 10^{-3}$ 2.542$\times 10^{-4}$ 0.30 1.406$\times 10^{-3}$ 8.403$\times 10^{-4}$ 3.366$\times 10^{-4}$ 0.03 5.666$\times 10^{-2}$ 1.349$\times 10^{-3}$ 1.670$\times 10^{-4}$ 0.30 1.786$\times 10^{-3}$ 1.175$\times 10^{-3}$ 3.081$\times 10^{-4}$ 0.03 7.133$\times 10^{-2}$ 1.137$\times 10^{-3}$ 1.502$\times 10^{-4}$ 0.30 2.330$\times 10^{-3}$ 1.396$\times 10^{-3}$ 5.073$\times 10^{-4}$ 0.03 8.982$\times 10^{-2}$ 1.082$\times 10^{-4}$ 1.681$\times 10^{-4}$ 0.30 2.846$\times 10^{-3}$ 1.315$\times 10^{-3}$ 4.777$\times 10^{-4}$ 0.03 1.131$\times 10^{-1}$ 3.235$\times 10^{-4}$ 2.101$\times 10^{-4}$ 0.30 3.575$\times 10^{-3}$ 2.593$\times 10^{-3}$ 9.301$\times 10^{-4}$ 0.10 2.397$\times 10^{-4}$ 2.721$\times 10^{-4}$ 1.521$\times 10^{-4}$ 0.30 4.537$\times 10^{-3}$ 2.771$\times 10^{-3}$ 9.136$\times 10^{-4}$ 0.10 2.861$\times 10^{-4}$ 1.436$\times 10^{-4}$ 4.891$\times 10^{-5}$ 0.30 5.706$\times 10^{-3}$ 2.767$\times 10^{-3}$ 7.532$\times 10^{-4}$ 0.10 3.521$\times 10^{-4}$ 1.093$\times 10^{-4}$ 5.191$\times 10^{-5}$ 0.30 7.137$\times 10^{-3}$ 2.966$\times 10^{-3}$ 7.580$\times 10^{-4}$ 0.10 4.724$\times 10^{-4}$ 5.567$\times 10^{-4}$ 4.046$\times 10^{-4}$ 0.30 8.992$\times 10^{-3}$ 4.268$\times 10^{-3}$ 1.529$\times 10^{-3}$ 0.10 5.742$\times 10^{-4}$ 3.166$\times 10^{-4}$ 8.959$\times 10^{-5}$ 0.30 1.131$\times 10^{-2}$ 3.695$\times 10^{-3}$ 9.241$\times 10^{-4}$ 0.10 7.142$\times 10^{-4}$ 3.239$\times 10^{-4}$ 1.059$\times 10^{-4}$ 0.30 1.424$\times 10^{-2}$ 3.936$\times 10^{-3}$ 8.243$\times 10^{-4}$ 0.10 9.321$\times 10^{-4}$ 7.483$\times 10^{-4}$ 2.526$\times 10^{-4}$ 0.30 1.793$\times 10^{-2}$ 5.021$\times 10^{-3}$ 1.199$\times 10^{-3}$ 0.10 1.139$\times 10^{-3}$ 8.075$\times 10^{-4}$ 2.949$\times 10^{-4}$ 0.30 2.256$\times 10^{-2}$ 4.804$\times 10^{-3}$ 7.301$\times 10^{-4}$ 0.10 1.417$\times 10^{-3}$ 6.968$\times 10^{-4}$ 1.449$\times 10^{-4}$ 0.30 2.842$\times 10^{-2}$ 3.819$\times 10^{-3}$ 6.866$\times 10^{-4}$ 0.10 1.812$\times 10^{-3}$ 9.472$\times 10^{-4}$ 2.221$\times 10^{-4}$ 0.30 3.576$\times 10^{-2}$ 2.999$\times 10^{-3}$ 4.241$\times 10^{-4}$ 0.10 2.265$\times 10^{-3}$ 1.291$\times 10^{-3}$ 3.512$\times 10^{-4}$ 0.30 4.499$\times 10^{-2}$ 2.994$\times 10^{-3}$ 3.940$\times 10^{-4}$ 0.10 2.841$\times 10^{-3}$ 1.423$\times 10^{-3}$ 2.675$\times 10^{-4}$ 0.30 5.664$\times 10^{-2}$ 2.110$\times 10^{-3}$ 2.131$\times 10^{-4}$ 0.10 3.590$\times 10^{-3}$ 1.917$\times 10^{-3}$ 3.613$\times 10^{-4}$ 0.30 7.130$\times 10^{-2}$ 1.101$\times 10^{-3}$ 2.666$\times 10^{-4}$ 0.10 4.511$\times 10^{-3}$ 2.127$\times 10^{-3}$ 3.669$\times 10^{-4}$ 0.30 8.975$\times 10^{-2}$ 7.149$\times 10^{-4}$ 2.101$\times 10^{-4}$ 0.10 5.677$\times 10^{-3}$ 2.735$\times 10^{-3}$ 4.798$\times 10^{-4}$ 0.41 1.118$\times 10^{-3}$ 1.717$\times 10^{-3}$ 5.549$\times 10^{-4}$ 0.10 7.151$\times 10^{-3}$ 3.390$\times 10^{-3}$ 5.174$\times 10^{-4}$ 0.41 1.395$\times 10^{-3}$ 1.287$\times 10^{-3}$ 5.887$\times 10^{-4}$ 0.10 8.990$\times 10^{-3}$ 4.109$\times 10^{-3}$ 7.895$\times 10^{-4}$ 0.41 1.801$\times 10^{-3}$ 3.991$\times 10^{-3}$ 2.017$\times 10^{-3}$ 0.10 1.130$\times 10^{-2}$ 4.269$\times 10^{-3}$ 6.702$\times 10^{-4}$ 0.41 2.259$\times 10^{-3}$ 2.333$\times 10^{-3}$ 1.102$\times 10^{-3}$ 0.10 1.423$\times 10^{-2}$ 4.864$\times 10^{-3}$ 1.069$\times 10^{-3}$ 0.41 2.870$\times 10^{-3}$ 3.762$\times 10^{-3}$ 1.463$\times 10^{-3}$ 0.10 1.792$\times 10^{-2}$ 4.292$\times 10^{-3}$ 5.291$\times 10^{-4}$ 0.41 3.616$\times 10^{-3}$ 2.445$\times 10^{-3}$ 8.784$\times 10^{-4}$ 0.10 2.255$\times 10^{-2}$ 4.104$\times 10^{-3}$ 7.505$\times 10^{-4}$ 0.41 4.516$\times 10^{-3}$ 5.841$\times 10^{-3}$ 2.415$\times 10^{-3}$ 0.10 2.839$\times 10^{-2}$ 3.489$\times 10^{-3}$ 3.827$\times 10^{-4}$ 0.41 5.687$\times 10^{-3}$ 4.887$\times 10^{-3}$ 2.170$\times 10^{-3}$ 0.10 3.573$\times 10^{-2}$ 2.375$\times 10^{-3}$ 3.648$\times 10^{-4}$ 0.41 7.172$\times 10^{-3}$ 8.052$\times 10^{-3}$ 3.469$\times 10^{-3}$ 0.10 4.497$\times 10^{-2}$ 2.027$\times 10^{-3}$ 2.117$\times 10^{-4}$ 0.41 9.012$\times 10^{-3}$ 7.250$\times 10^{-3}$ 2.318$\times 10^{-3}$ 0.10 5.663$\times 10^{-2}$ 1.736$\times 10^{-3}$ 1.238$\times 10^{-4}$ 0.41 1.133$\times 10^{-2}$ 7.995$\times 10^{-3}$ 1.968$\times 10^{-3}$ 0.10 7.131$\times 10^{-2}$ 9.309$\times 10^{-4}$ 1.217$\times 10^{-4}$ 0.41 1.427$\times 10^{-2}$ 8.106$\times 10^{-3}$ 2.493$\times 10^{-3}$ 0.10 8.977$\times 10^{-2}$ 2.450$\times 10^{-4}$ 1.744$\times 10^{-4}$ 0.41 1.793$\times 10^{-2}$ 8.183$\times 10^{-3}$ 2.148$\times 10^{-3}$ 0.20 2.666$\times 10^{-4}$ 9.025$\times 10^{-5}$ 2.044$\times 10^{-5}$ 0.41 2.252$\times 10^{-2}$ 7.539$\times 10^{-3}$ 2.298$\times 10^{-3}$ 0.20 5.359$\times 10^{-4}$ 2.741$\times 10^{-4}$ 8.676$\times 10^{-5}$ 0.41 2.835$\times 10^{-2}$ 6.100$\times 10^{-3}$ 1.462$\times 10^{-3}$ 0.20 7.364$\times 10^{-4}$ 1.893$\times 10^{-4}$ 5.990$\times 10^{-5}$ 0.41 3.576$\times 10^{-2}$ 4.826$\times 10^{-3}$ 9.621$\times 10^{-4}$ 0.20 8.586$\times 10^{-4}$ 3.430$\times 10^{-4}$ 1.778$\times 10^{-4}$ 0.41 4.502$\times 10^{-2}$ 3.285$\times 10^{-3}$ 6.987$\times 10^{-4}$ 0.20 1.078$\times 10^{-3}$ 4.733$\times 10^{-4}$ 1.528$\times 10^{-4}$ 0.41 5.666$\times 10^{-2}$ 2.789$\times 10^{-3}$ 2.951$\times 10^{-4}$ 0.20 1.394$\times 10^{-3}$ 8.696$\times 10^{-4}$ 1.969$\times 10^{-4}$ 0.41 7.131$\times 10^{-2}$ 2.367$\times 10^{-3}$ 3.749$\times 10^{-4}$ 0.20 1.767$\times 10^{-3}$ 9.598$\times 10^{-4}$ 2.439$\times 10^{-4}$ 0.41 8.977$\times 10^{-2}$ 1.319$\times 10^{-3}$ 3.876$\times 10^{-4}$ 0.20 2.200$\times 10^{-3}$ 1.607$\times 10^{-3}$ 4.047$\times 10^{-4}$ 0.41 1.131$\times 10^{-1}$ 4.726$\times 10^{-4}$ 3.530$\times 10^{-4}$ \ \[tab2\] Implications for the UV background {#sec:uvb} ================================== The amplitude of the Lyman-$\alpha$ forest power spectrum is sensitive to the UV background; therefore, it can be used to measure the UV background quantified by the [H$\,$[i]{}]{} photoionization rate $\Gamma_{{{\mbox{\scriptsize {H$\,${\sc i}}}}}}$. The basic idea behind the measurement is to compare the power-spectrum with cosmological hydrodynamical simulations of the IGM where $\Gamma_{{{\mbox{\scriptsize {H$\,${\sc i}}}}}}$ is one of the free parameters. In this section, we first discuss our IGM simulations and then the $\Gamma_{{{\mbox{\scriptsize {H$\,${\sc i}}}}}}$ measurements obtained from them. Simulations ----------- For comparing our power spectrum measurements with the simulated IGM, we ran an Nyx cosmological hydrodynamic simulation [@Almgren13; @Lukic15] from $z=159$ to $z=0.03$. This is an Eulerian hydrodynamical simulation of box size 20 cMpc/h and 1024$^3$ cells. The initial conditions were generated using MUSIC code [@Hahn11] along with the transfer function from CAMB [@Lewis00; @Howlett12]. We evolve baryon hydrodynamics in Eulerian approach on a fixed Cartesian grid with 1024$^3$ cells, and follow the evolution of dark matter using 1024$^3$ Lagrangian (N-body) particles. This simulation resolves 19.5 ckpc/h scales (i.e, $\Delta v <2$ km s$^{-1}$ at $z<0.5$). In this simulation, we used the photoheating rates from the @Puchwein19 non-equilibrium models (the equivalent-equilibrium rates since our code assumes ionization equilibrium). We stored simulation outputs at different redshifts corresponding to our $\bar z$ at the measured P($k$). We determined T$_0$ and $\gamma$ by fitting the distribution of densities and temperatures in the simulation following the linear least squares method described in @Lukic15. The simulation redshifts, T$_0$, and $\gamma$ are provided in Table \[tab1\]. The values of T$_0$ and $\gamma$ are also consistent with the theoretical models presented in @McQuinn16 and obtained in the simulations presented by @Shull15 and @Gaikwad17a. For our current purposes, the only free parameter in this simulation is $\Gamma_{\rm H\, I}$. We vary $\Gamma_{\rm H\, I}$ and generate the simulated Lyman-$\alpha$ forest as follows. First, we calculate the ionization fractions of hydrogen and helium under the assumption of ionization equilibrium including both photoionization and collisional ionization. For this we have used updated cross-sections and recombination rates from @Lukic15. Then, we extract a large number ($5\times 10^4$) of random lines-of-sight (skewers) parallel to the (arbitrarily chosen) $z-$axis of the simulation cube. Along these lines-of-sight, we store the ionization densities, temperatures and $z$-component of the velocities. As our procedure does not include radiative transfer, we model the self-shielding of dense cells using the prescription described in @Rahmati13. We generate the simulated Lyman-$\alpha$ optical depth for each cell along the line of sight, which we will refer to as $\tau$-skewers, by summing all the real space contributions to the redshift space optical depth using the full Voigt profile resulting from each (real-space) cell following the approximations used in [@Tepper06]. The flux $F=e^{-\tau}$ gives us the continuum normalized Lyman-$\alpha$ forest flux along these skewers. These constitute our ‘perfect skewers’ from the simulation. We calculate and store 5$\times$10$^4$ simulated skewers from each box for different values of $\Gamma_{\rm H\, I}$. Note that every-time we change the $\Gamma_{\rm H\, I}$ we recalculate the skewers following the procedure described above. We do not simply rescale the $\tau$ values along the skewers when we change the $\Gamma_{\rm H\, I}$, as is typically applied to simulations of the Lyman-$\alpha$ forest at higher redshifts following the FGPA. According to the FGPA, $\tau \propto \Gamma_{\rm H\, I}^{-1}$ (from Eq. \[eq2\]), which is a good approximation at high-$z$ when most of the gas in the IGM is photoionized. However, a large amount of gas in the low-$z$ Universe is collisionally ionized because it is heated to $T> 10^5\,{\rm K}$ by structure formation shocks [see @Dave10; @Shull12]. In this regime, the simple rescaling of $\tau$ following the FGPA leads to erroneous results, because the contribution of collisional ionization implies that the true optical depth is no longer linearly proportional to $\Gamma_{\rm H\, I}$ [see also @Lukic15]. This is illustrated in Fig. \[fig4\] where we show the power-spectrum for $\tau$-skewers generated with a fiducial value of $\Gamma_{\rm H\,I \, Fid}=1.75 \times 10^{-13}$ s$^{-1}$ (black-curve) and three more power-spectra where the $\tau$-skewers were initially calculated for $\Gamma_{\rm H\, I \, Fid} \times 4$, $\Gamma_{\rm H\, I\, Fid} /3.5$ and $\Gamma_{\rm H\, I, Fid} /2$ values and then rescaled to get the $\tau$-skewers corresponding to $\Gamma_{\rm H\, I \, Fid}$ following FGPA. The latter three deviate significantly from the fiducial power-spectrum on large scales (low $k <0.04$ s km$^{-1}$values). The bottom panel of Fig. \[fig4\] shows the percentage differences in the correct calculation versus the ones obtained by using the FGPA. The differences are large (of the order of 10 to 25%) when the FGPA is applied for larger difference in $\Gamma_{\rm H\, I}$ (of factor 2 to 4). Given that the reported low-$z$ $\Gamma_{\rm H\, I}$ measurements vary over factors of $2-5$, the results obtained by incorrectly using FGPA can give large systematic differences in the derived $\Gamma_{\rm H\, I}$ values. ![The effect of using the FGPA on power-spectra. Top panel shows $z=0.2$ power-spectrum estimated for $\Gamma_{\rm H\, I\, Fid}=1.75 \times 10^{-13}$ s$^{-1}$ (black-curve) by correctly calculating the $\tau$-skewers. The green, blue and red curves show power-spectra estimated when $\tau$-skewers, which were initially calculated for $\Gamma_{\rm H\, I\, Fid} \times 4$, $\Gamma_{\rm H\, I} /2$ and $\Gamma_{\rm H\, I\, Fid} /3.5$ values, were rescaled to $\Gamma_{\rm H\, I\, Fid}$ values by following FGPA. At $k<0.04$ s km$^{-1}$ the power-spectra obtained using FGPA deviates from the correct power-spectrum (black-curve) and the percentage deviation are shown in the bottom panel with the same colors. The deviation is large for large differences in the initial and final $\Gamma_{\rm H\, I}$ values.[]{data-label="fig4"}](rescale.pdf){width="48.00000%" height="\textheight"} Once the perfect-skewers are constructed, we adopt a forward modeling approach to make them look like realistic spectra. To this end, we follow the procedure developed in @Walther18a but for the HST COS data. We first stitch randomly drawn skewers together to cover the Lyman-$\alpha$ forest redshift path of each quasar. Then, we convolve these skewers with the finite COS LSFs corresponding to the same gratings and lifetime positions as the data for each quasar. Then, we rebin these to match the pixels of the actual data, and we add Gaussian random noise at each pixel generated with the standard deviation of the error-vector from the observed spectra. Finally, we mask these spectra in exactly same manner as the data and follow the same procedure to calculate the power-spectrum (see Section 3.1). We calculate $P(k)$ from these forward models created for a large number of $\Gamma_{\rm H\, I}$ values and estimate the $\Gamma_{\rm H\, I}$ by comparing with the $P(k)$ measurements as explained in the next sub-section. Constraints on the UV background {#sec4.2} -------------------------------- The power spectrum is not only sensitive to $\Gamma_{\rm H\, I}$ but also to the thermal state of the IGM quantified by $T_0$ and $\gamma$. Therefore, to correctly measure the $\Gamma_{\rm H\, I}$ using our power spectrum, we need a large ensemble of simulations of different IGM thermal state models stored at the redshifts where we have measured the power spectrum. We have the THERMAL grid[^6] [@Hiss18; @Walther19] available with $>70$ Nyx simulations (with box size 20 cMpc/h and 1024$^3$ particles) of different IGM thermal models but only a single redshift $z=0.2$ overlaps with our dataset presented here. We use 50 simulations from this suite and follow the Bayesian inference approach presented in @Walther19, where the likelihood of the model is given by $$\begin{aligned} \mathcal{L}\equiv&P(\mathrm{data}|\mathrm{model})\\ \propto&\prod_\mathrm{datasets}\frac{1}{\sqrt{\det(C)}}\exp\left(-\frac{\mathbf{\Delta}^\mathrm{T} C^{-1} \mathbf{\Delta}}{2}\right) \nonumber\\ \mathbf{\Delta}=&\mathbf{P}_\mathrm{data}-\mathbf{P}_\mathrm{model} \nonumber \label{eq:likelihood}\end{aligned}$$ Here $C$ is the covariance matrix of the measurements where the diagonal elements are taken from the uncertainties obtained here (see Table \[tab2\]), and the off-diagonal elements are estimated from the forward model which is closest in the parameter space to the model in question. Using this $\mathcal{L}$, we perform a Bayesian inference at $z=0.2$ using Markov Chain Monte Carlo methods (MCMC) and jointly estimate the $T_0$, $\gamma$, and $\Gamma_{\rm H\, I}$ at $z=0.2$ (Walther et al. in prep.). We then marginalize the joint posterior distribution over parameters $T_0$ and $\gamma$ to determine $\Gamma_{\rm H\, I}$ and its 68% confidence interval. At other redshifts, we do not yet have such a large simulation grid. However the analysis at $z=0.2$ provides a clear indication of how much the degeneracies in the thermal state of the low-z IGM propagate into uncertainties on the $\Gamma_{\rm H\, I}$ measurements. To estimate the $\Gamma_{\rm H\, I}$ at other redshifts but using only one thermal model rather than a large grid of thermal models, we simply assume that the scaling between uncertainties at these redshifts behave similarly to our $z=0.2$ $\Gamma_{\rm H\, I}$ measurement. To quantify the amount by which the uncertainties on the $\Gamma_{\rm H\, I}$ can be underestimated if one uses only a single simulation instead of the full grid, we repeat the analysis mentioned above at $z=0.2$ but using only one simulation with $\Gamma_{\rm H\, I}$ being the only free parameter. In this analysis, we are using the same simulation for which we have stored outputs at other redshifts as well (as described in Section 3 and Table \[tab1\]). Also, for this $\Gamma_{\rm HI}$ estimate, we opt to use only the diagonal elements of the covariance matrix because the shape of power spectrum does not vary widely in this single parameter model where the IGM thermal state is fixed. We found that fitting with the full covariance led to spurious bad fits resulting from the rigidity of the model of the power spectrum shape if the thermal state is fixed. In this case, where we use only one simulation (and hence one thermal model) and the diagonal elements of covariances, we calculate the $\Gamma_{\rm H\, I}$ and its uncertainties represented by 68% confidence intervals by performing a maximum likelihood analysis. We find that these uncertainties are underestimated by a factor of 5.75 as compared to the one obtained using full simulation grid and covariances. We repeat this calculation at other redshifts and obtain the $\Gamma_{\rm H\, I}$ and the 68% confidence interval using only the diagonal elements. Next, we multiply this confidence interval by factor of 5.75 so that it correctly represents the uncertainties arising from degeneracies in the thermal state of the low-$z$ IGM. We believe that this approach of obtaining uncertainties at redshifts other than $z=0.2$, although approximate, is nevertheless a significant improvement over previous estimates of $\Gamma_{\rm H\,I}$ that are based on IGM simulation outputs which effectively assume perfect knowledge of the thermal state of the IGM. Nevertheless, in a companion paper (Walther et al. in prep.), we will present joint constraints on $T_0$, $\gamma$, and $\Gamma_{\rm H\, I}$. For all $\Gamma_{\rm H\, I}$ estimates discussed here, we have fit our power spectrum measurements over the range $0.03<k<0.1$ s km$^{-1}$, where the smallest $k$ is chosen to minimize the box size effects and largest $k$ corresponds to scales where we have reliable uncertainties on the power spectrum measurements. However, note that our results are not sensitive to the choice of the smallest $k$ value ($k_{\rm min}$) as long as $0.03>k_{\rm min}>0.01$ s km$^{-1}$. ![Our estimated $\Gamma_{\rm H\, I}$ values (black circles) compared with previous measurements by @Gaikwad17a [green diamonds], @Kollmeier14 [red star], a fitting form by @Shull15 [orange curve] and upper limits from @Fumagalli17 [downward arrow]. Prediction from a new UV background model of @KS19 [blue-dash curve] is consistent with our measurements. Also, the UV background predictions using updated QLF from @Kulkarni18 [red dotted curve] give consistent results with the measurements (see Section \[sec4.2\] for more details). The UV background from @HM12 [magenta dot-dash curve] is factor of $\sim 2$ smaller than the measurements.[]{data-label="fig5"}](Gamma_zoom_new.pdf){width="48.00000%" height="\textheight"} ${\bar z}$ $\Gamma_{\rm H\ I}$ (10$^{-13}$s$^{-1}$) ------------ ------------------------------------------ -- -- 0.03 0.585 $^{+0.17}_{-0.18}$ 0.10 0.756 $^{+0.17}_{-0.18}$ 0.20 1.135 $^{+0.32}_{-0.32}$ 0.30 1.479 $^{+0.44}_{-0.52}$ 0.41 1.418 $^{+0.53}_{-0.57}$ : $\Gamma_{\rm H\ I}$ measurements \[tab3\] Our $\Gamma_{\rm H\, I}$ measurements and the 68% confidence interval on them obtained following the method mentioned above are provided in Table \[tab3\] and shown in Fig. \[fig5\]. We also show the previous measurements reported in the literature at $z<0.5$ for comparison [however, see @Dave01 for an early attempt]. The upper limit at $z\sim 0$ by @Fumagalli17 was obtained from the H$\alpha$ and 21-cm observations of a nearby galaxy. All other measurements have used the Lyman-$\alpha$ forest from @Danforth14. The measurement by @Kollmeier14 at $z\sim 0.1$ and the fit resulting from the analysis by @Shull15 at $z<0.5$ were obtained by modeling the CDDF, whereas measurements of @Gaikwad17a were obtained using the flux PDF and power-spectrum (and later confirmed with the CDDF in @Gaikwad17b). Our measurements are consistent with the previous studies except for @Kollmeier14, which is $\sim 2.5$ times higher. As compared to previous studies, we have independent measurements obtained only from the Lyman-$\alpha$ forest power spectrum extending down to $z= 0.03$. In Fig. \[fig5\], we also show predictions from the UV background models by @HM12 and @KS19[^7]. Ours as well as the other $\Gamma_{\rm H\, I}$ measurements [except for @Kollmeier14] are consistent with the @KS19 UV background and are a factor of $\sim 2$ higher than @HM12. The @KS19 UV background at $z<0.5$ is mostly contributed by quasars since it has been obtained with a negligible contribution from galaxies (at $z<2$; see their Eq. 13). Therefore, we argue that quasar emission is sufficient to produce the low-z UV background [@Khaire15puc see also]. The factor of five discrepancy between the measurement from @Kollmeier14 and the prediction from @HM12 was argued to represent a photon-underproduction crisis at low-$z$. Soon after the @Kollmeier14 study, @Khaire15puc showed that the UV background models that include an updated quasar emissivity predict a factor of two higher $\Gamma_{\rm H\, I}$, without requiring any additional contribution from galaxies. This model later turned out to be consistent with many new $\Gamma_{\rm H\, I}$ measurements [@Shull15; @Gaikwad17a; @Gaikwad17b; @Gurvich17; @Viel17; @Fumagalli17]. As shown in Fig \[fig5\], this is also consistent with our new measurements. Other recent updates to UV background models [@Madau15; @KS19; @Puchwein19] also came to the same conclusion using more recent estimates for the quasar emissivity similar to @Khaire15puc. Recently, @Kulkarni18 updated the fits to the quasar luminosity functions (QLF) and emissivity across a large redshift range and claim that the quasar contribution to the low-$z$ UV background is factor of $\sim 2$ smaller than the predictions by these recent UV background models [@Madau15; @KS19; @Puchwein19] and most of the measurements. However, the @Kulkarni18 calculation uses the CDDF of the IGM from tabulated fits by @HM12 and an ionizing spectral slope of quasars $f_{\nu} \propto \nu^{-1.7}$ [@Lusso15]. Instead, using the @InoueAK14 CDDF results in a 40-60% higher UV background at $z<0.5$ [see, for e.g., Figure 16 of @Gaikwad17a]. This along with including the recombination emissivities in the UV background calculations should easily increase the $\Gamma_{\rm H\, I}$ estimates of @Kulkarni18 by factor of $\sim 1.8$ as shown in Fig. \[fig5\]. For this calculation we used the UV background code presented in @KS19 but with the updated quasar emissivity determined by @Kulkarni18, with the same limiting magnitude at 1450 Å ($M_{\rm 1450, lim}=-18$) and the same spectral slope $f_{\nu} \propto \nu^{-1.7}$ that they adopted, but using the @InoueAK14 CDDF instead of the @HM12 fits used by @Kulkarni18. We note that adopting a flatter spectral slope for quasars can further increase the $\Gamma_{\rm H\, I}$. For example, redoing the calculation mentioned above using the @Kulkarni18 emissivity and the @KS19 UV background model, but changing the quasar spectra slope to $f_{\nu} \propto \nu^{-1.4}$ consistent with the harder slope measured by @Shull12 and @Stevans14, results in a further increase of the $\Gamma_{\rm H\, I}$ by a factor of 1.3. Finally, note that even if the $\Gamma_{\rm H\, I}$ is as high as that purported by @Kollmeier14, there is no crisis associated with photon production since a negligible contribution from galaxies, along with updated cosmic star formation histories [@Madau14; @Khaire15ebl], can easily reproduce such a large $\Gamma_{\rm H\, I}$ [see for more details @Khaire15puc]. Summary and conclusions ======================= We present a new high precision high resolution power spectrum measurement in five different redshift bins at $z<0.5$. For this measurement, we have used high-quality medium resolution Lyman-$\alpha$ forest data from HST/COS from the largest low-$z$ IGM survey published by @Danforth14. We applied the procedure developed in @Walther18a, which takes into account masked metal-line absorptions, noise in the data and the finite resolution of the instrument. The data allow us to reliably probe the power spectrum up to small scales $k<0.1$ s km$^{-1}$ and our measurements show the expected thermal cut-off in the power at small scales $k>0.03$ s km$^{-1}$, resulting from pressure smoothing of IGM gas and thermal Doppler broadening of absorption lines. Our power spectrum measurements are provided in Table \[tab2\]. We compare these with cosmological hydrodynamical simulations and obtain constraints of the UV background at $z<0.5$. Our measured hydrogen photoionization rates (Table \[tab3\]) are consistent with the previous estimates [@Shull15; @Gaikwad17a; @Gaikwad17b; @Fumagalli17] and recent UV background models [@KS19; @Puchwein19]. This suggests that the low-$z$ UV background is dominated by ionizing photons emitted by quasars without requiring any significant contribution from galaxies. The power-spectrum measurements presented here, in principle, can probe the thermal state of the low-$z$ IGM. At low-$z$, theoretical calculations using standard heating and cooling rates show that the IGM loses memory of the previous heating episodes caused by the hydrogen and helium reionization. This makes understanding and predicting the structure of the low-$z$ IGM relatively simple, as it is independent of the physics associated with hydrogen and helium reionization heating, which complicate modeling at higher redshifts ($z \gtrsim 2$). In the absence of any other heating processes, theory predicts that the diffuse low-density photoionized IGM cools down after $z \sim 2$ and asymptotes toward a single temperature-density relation with $\gamma$ close to 1.6 and T$_0 \sim 5000$ K [@McQuinn16] at $z=0$. Such a predicted cool-down of the IGM at low-$z$ has not yet been observationally confirmed. Indeed, there are no reliable measurements of the thermal state of the IGM at $z<1.6$ [but see @Ricotti00] where the atmospheric cut-off does not allow us to observe Lyman-$\alpha$ forest from ground based telescopes. In a companion paper (Walther et al. in prep.) we will use the power spectrum measurements presented here to jointly constrain the IGM thermal state (T$_0$, $\gamma$) and the UV background ($\Gamma_{\rm H\, I}$). These measurements will help us understand the physics of IGM and address important questions such as whether feedback processes associated with galaxy formation modify the thermal state of the IGM at low-$z$ [@Viel17; @Nasir17] and/or if there is any room for the existence of non-standard heating processes powered by TeV Blazars [@Puchwein12; @Lamberts15] or decaying dark matter [@Furlanetto06DM_decay; @Araya14]. In contrast with the high-$z$ Universe, the much lower opacity of Lyman series absorption arising in the low-$z$ IGM results in dramatically reduced line blanketing, making it relatively straightforward to identify all lines as either resulting from the Lyman-series or metal absorption. This results in a large redshift path length where Lyman series absorption can be studied, data analysis and preparation are simplified, and important systematics from metal-line contamination arising at small-scales [high-$k$; see section 4.1 of @Walther18a] are mitigated. Furthermore, our measurements demonstrate that COS resolution is sufficient to obtain high-quality power-spectrum measurements even at the small-scales (high-$k$, $k\sim 0.1\,{\rm s\,km^{-1}}$) required for probing the thermal state of the IGM and demonstrate the important role that HST/UV spectroscopy can play in our understanding of the low-$z$ IGM. We conclude by noting that, owing to the paucity of archival near-UV spectra covering the Lyman-$\alpha$ transition at $0.5<z<1.6$, there are essentially no constraints on the physical state of IGM gas in this redshift interval, representing 5 Gyr of the Universe’s history. It is critical that HST UV spectroscopy fill this gap in our understanding of the Universe before HST’s mission is complete, otherwise we could remain in the dark for decades. acknowledgement {#acknowledgement .unnumbered} =============== VK thanks R. Srianand, T. R. Choudhury and P. Gaikwad for insightful discussion on the power spectrum normalization. We thank all members of the ENIGMA group[^8] at University of California Santa Barbara for useful discussions and suggestions. Financial support for this work was provided to VK by NASA through grant number HST-AR-15032.002-A from the Space Telescope Science Institute, which is operated by Associated Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. Calculations presented in this paper used the draco and hydra clusters of the Max Planck Computing and Data Facility, a center of the Max Planck Society in Garching (Germany). We have also used the National Energy Research Scientific Computing Center (NERSC) supported by the U.S. Department of Energy (DoE) under Contract No. DE-AC02-05CH11231. ZL was in part supported by the Scientific Discovery through Advanced Computing (SciDAC) program funded by the DoE, the Office of High Energy Physics and the Office of Advanced Scientific Computing Research. Power spectrum without masking metal absorption lines {#A} ===================================================== {width="99.50000%" height="\textheight"} While preparing data for the power spectrum calculation, we have properly masked the metal lines originating from Milky-way and intervening metals as well as other contaminations (see Section 2). This is an important procedure to correctly calculate the power spectrum. However, to estimate the relevance of different metal contamination, in Fig. \[figA1\] we show the power spectrum calculated without masking any metal absorption lines (green points) and masking only Milky-way metal lines (red points) in the Lyman-$\alpha$ forest. However, note that we mask the spectral gaps and all other emission lines, which appear mainly in our lowest redshift bin ($\bar z = 0.03$). As expected, when metal lines are not masked the power is systematically higher because of the extra fluctuations in the flux introduced by metal lines. However the difference in power is larger at lower redshift mostly because of strong metal contamination arising from the ISM of the Milky-way and many of the absorption lines are correlated as they originate from the same source and with same wavelength separation. The intervening metals have insignificant impact on the Lyman-$\alpha$ power spectrum at low-$z$ (the difference between red and black points in Fig \[figA1\]). On the other hand, Milky-way metal lines are easy to identify which makes any incompleteness in intervening metal identification unimportant for calculating the power spectrum. Effect of masking on the power-spectrum {#B} ======================================= To check if the spectral masking is introducing any significant contamination in the power-spectrum we calculate the power spectrum from our forward models (using $\Gamma_{\rm H \,I}$ values consistent with our measurements) but without using any masking on the Lyman-$\alpha$ forest. In these forward models for the purpose of only studying the effect of masking, we are using infinite resolution noiseless spectra. In Fig. \[figA2\] we compare the resulting power spectrum with and without masking the Lyman-$\alpha$ forest. The bottom panel shows a fractional difference between these. The maximum change in relevant $k<0.1$ s km$^{-1}$ values is no more than 5% at any redshift. For comparing with the actual uncertainties in the power spectrum, we also plot the fractional errors in the bottom panel. The figure clearly shows that the fractional change in the power-spectrum because of masking is less than the errors on the actual power spectrum measurements. This motivates us not to perform the masking corrections on the measurements as done in @Walther18a. {width="99.50000%" height="\textheight"} Effect of the COS LSF on the power-spectrum {#C} =========================================== The COS LSF is quite different from a Gaussian as it exhibits strong non-Gaussian wings. In the power spectrum calculation the LSF appears in the window function correction (see Eq. 1) and it is also important to generate the forward models. To see how much the power spectrum can be affected if one uses a Gaussian LSF with a similar resolution to COS we perform the following analysis. We use mock spectra generated from our forward models (at $\Gamma_{\rm H \,I}$ values consistent with our measurements) generated using the correct COS LSF but while calculating the power spectrum we use a Gaussian LSF. The resulting power spectrum (red curves) is shown in the Fig \[figA3\] along with that obtained by using the correct COS LSF (black curves). The amplitude of the power spectrum in the former is smaller at large $k$ values and the differences, as shown in the bottom panels, can be as large as 80%. This is much too large as compared to the actual uncertainties from our measurements. Although the differences decrease at higher redshifts, because of small sample sizes, the discrepancy is too significant to ignore. This exercise shows that it is never a sound choice to approximate the COS LSF as a Gaussian. {width="99.50000%" height="\textheight"} \[lastpage\] [^1]: Link: http://archive.stsci.edu/prepds/igm/ [^2]: A pixel in this dataset corresponds to $\Delta v = 6.67$ km s$^{-1}$. [^3]: Since our S/N cut is $\ge 10$, we always record sufficient photons to be in the Gaussian regime. [^4]: We use the python package [linetools]{} (<https://linetools.readthedocs.io/en/latest/api.html>) which can interpolate the COS LSFs from <http://www.stsci.edu/hst/cos/performance/spectral_resolution/> to any central wavelength. [^5]: Using a different wavelength in the range $1400-1500$ Å for the transition from G130M to G160M grating has negligible effect on the calculated power spectrum at $\bar{z}=0.2$. [^6]: Link: http://thermal.joseonorbe.com/ [^7]: For @KS19 UV background, we use their fiducial Q18 model which uses quasar spectral slope $f_{\nu} \propto \nu^{-1.8}$ at hydrogen ionizing energies [following @Lusso15; @Khaire17sed] in the UV background calculations. [^8]: <http://enigma.physics.ucsb.edu/>

ArXiv

--- author: - 'A. Nebot Gómez-Morán, B. T. Gänsicke, M. R. Schreiber, A. Rebassa-Mansergas, A. D. Schwope, J. Southworth, A. Aungwerojwit, M. Bothe, P. J. Davis, U. Kolb, M. Müller, C. Papadaki, S. Pyrzas, A. Rabitz, P. Rodríguez-Gil, L. Schmidtobreick, R. Schwarz, C. Tappert, O. Toloza, J. Vogel, M. Zorotovic' bibliography: - 'aamnem99.bib' - 'references.bib' date: 'Received / Accepted ' title: | Post common envelope binaries from SDSS. XII:\ The orbital period distribution --- [The complexity of the common envelope phase and of magnetic stellar wind braking currently limits our understanding of close binary evolution. Because of their intrinsically simple structure, observational population studies of white dwarf plus main sequence (WDMS) binaries hold the potential to test theoretical models and constrain their parameters.]{} [The Sloan Digital Sky Survey (SDSS) has provided a large and homogeneously selected sample of WDMS binaries, which we are characterising in terms of orbital and stellar parameters.]{} [We have obtained radial velocity information for 385 WDMS binaries from follow-up spectroscopy, and for an additional 861 systems from the SDSS sub-spectra. Radial velocity variations identify 191 of these WDMS binaries as post common envelope binaries (PCEBs). Orbital periods of 58 PCEBs were subsequently measured, predominantly from time-resolved spectroscopy, the total number of SDSS PCEBs with orbital parameters to 79. Observational biases inherent to this PCEB sample were evaluated through extensive Monte Carlo simulations.]{} [We find that $21-24$% of all SDSS WDMS binaries have undergone common envelope evolution, which is in good agreement with published binary population models and high-resolution *HST* imaging of WDMS binaries unresolved from the ground. The bias corrected orbital period distribution of PCEBs ranges from 1.9h to 4.3d and follows approximately a normal distribution in $\log({\mbox{$P_\mathrm{orb}$}})$, peaking at $\sim10.3$h. There is no observational evidence for a significant population of PCEBs with periods in the range of days to weeks. ]{} [The large and homogeneous sample of SDSS WDMS binaries provides the means to test fundamental predictions of binary population models, and hence to observationally constrain the evolution of all close compact binaries. ]{} Introduction ============ It is well known that all types of close compact binaries, including X-ray binaries, cataclysmic variables, and double degenerates, form through common envelope evolution [@paczynski76-1]. This phase is required to explain the observed small binary separations in compact binaries. The main concept is that, for a suitable range of orbital separations, the massive star (i.e. the primary) will fill its Roche lobe as it evolves up the giant (or asymptotic giant) branch, and will then in most cases become susceptible to dynamically unstable mass transfer. The transferred material cannot cool as fast as it is being transferred and the core of the primary star and the companion star (also referred to as secondary star) are both engulfed by the envelope of the former. The envelope is not co-rotating with the binary, creating drag forces which transport orbital angular momentum and energy from the binary to the common envelope, leading to a shrinkage of the orbit [@ricker+taam08-1]. Eventually, the common envelope is ejected, leaving behind a post common envelope binary (PCEB). Although the basic concept of the common envelope phase has been outlined 30 years ago, it is still the most poorly understood phase of compact binary evolution. While we have learnt that this phase must be very short ($\le 10^3$ yrs) [@hjellming+taam91-1; @webbink08-1], current hydrodynamical simulations of common envelope evolution are still unable to unambiguously link the initial parameters of the binary with the outcome of the common envelope, and numerically too expensive to be run on large ranges of initial binary parameters [@ricker+taam08-1]. Therefore theoretical binary population synthesis studies (BPS) usually adopt a simple parametrisation: a certain fraction ($\alpha_\mathrm{CE}$) of the binary’s binding energy which is released in the spiralling-in process is used to unbind the common envelope [@webbink84-1]. However, @nelemansetal00-1 and @nelemans+tout05-1 find difficulties even with this scaled energy relation when trying to reproduce the evolution of double white dwarf binaries, and propose a prescription based on angular momentum conservation. A fundamental consequence of this approach is that clear observational constraints on these parameters are necessary to make any meaningful predictions on the Galactic population of compact binaries. PCEBs[^1] composed of a white dwarf and a main-sequence (WDMS) companion represent the most promising population to derive such observational constraints, as they are the intrinsically most abundant population of PCEBs and hence easily accessible to intense studies using the current suite of 2–8m telescopes, they underwent only one common envelope phase, are not altered by ongoing accretion, and their stellar components are structurally simple. In general, WDMS binaries comprise two fundamentally different types of systems: wide binaries that evolved like single stars, i.e. without interaction, and PCEBs. The major shortcoming of using PCEBs for studying compact binary evolution has been their small number. The comprehensive study of @schreiber+gaensicke03-1 contained only 30 systems with accurate binary parameters, and highlighted that the sample of PCEBs known at that time was not only small but also heavily biased towards hot white dwarfs and late secondary star spectral types. This bias was identified by @schreiber+gaensicke03-1 as a natural consequence of the fact that most of these PCEBs were discovered in blue-excess surveys. Over the past few years, the Sloan Digital Sky Survey (SDSS, @yorketal00-1 [@abazajianetal09-1; @yannyetal09-1]) has totally changed this situation. Thanks to its broad $ugriz$ colour space, and massive spectroscopic follow-up, SDSS has identified over 1900 WDMS binaries [@smolcicetal04-1; @silvestrietal06-1; @silvestrietal07-1; @schreiberetal07-1; @helleretal09-1; @rebassa-mansergasetal07-1; @rebassa-mansergasetal10-1]. We have initiated an intensive follow-up study of the SDSS WDMS binaries, with the ultimate aim to substantially improve our understanding of compact binary evolution. First results on individual systems and small sub-sets of this sample have been published [@schreiberetal08-1; @rebassa-mansergasetal08-1; @nebotgomez-moranetal09-1; @pyrzasetal09-1; @schwopeetal09-1], as well as several studies that analyse the total sample in the context of close binary evolution [@schreiberetal10-1; @zorotovicetal10-1; @rebassa-mansergasetal11-1]. Here we present one of the fundamental results of this project: the orbital period distribution of 79 SDSS WDMS binaries, which we discuss in the context of current theories of compact binary evolution. Overall strategy and input sample {#sec:sample} ================================= Our aim of characterising a large and homogeneous sample of PCEBs requires a multi-stage approach. The first step is to identify all WDMS binaries within the SDSS spectroscopic data release. The latest catalogues of WDMS binaries from SDSS list 1903 WDMS systems, $\sim1600$ from @rebassa-mansergasetal10-1 based on the SDSS Data Release (DR) 6, plus another $\sim300$ that have been discovered by us [@schreiberetal07-1] as part of the Sloan Extension for Galactic Understanding and Exploration (SEGUE, @yannyetal09-1). These catalogues include white dwarf temperatures and masses, as well as the spectral types of the companion stars. In brief, stellar parameters are derived from fitting the SDSS spectrum to template spectra made of white dwarf plus dM templates. The spectral type of the secondary star is derived from the best fit and the mass is calculated using a spectral type-mass relation [see @rebassa-mansergasetal07-1 for details]. The mass and the temperature of the white dwarf are derived from fitting both lines and continuum (after the contribution from the companion has been subtracted) to stellar models from [@koesteretal05-1]. The second step is to identify the PCEBs among the entire sample of WDMS binaries. The comprehensive population synthesis by @willems+kolb04-1 suggested that $\sim75\%$ of all Galactic WDMS binaries are wide systems, and that the vast majority of the ${\raisebox{-0.4ex}{${\stackrel{<}{\scriptstyle \sim}}$}}\,25\%$ underwent a common envelope phase, hence several hundred PCEBs are to be expected among the SDSS WDMS sample. A small fraction of WDMS binaries appear as blended white dwarf plus M-dwarf pairs in the SDSS imaging, and were flagged as wide systems by @rebassa-mansergasetal10-1. However, even those WDMS binaries that are genuinely unresolved on the SDSS images can still have orbital separations of many tens of astronomical units. The key to identify the close WDMS binaries is hence to carry out a radial velocity survey of the spatially unresolved SDSS WDMS binaries, obtaining at least two spectra, separated by at least one night, of as many WDMS binaries as possible. In a final step, intense follow-up spectroscopy of the PCEBs, i.e. WDMS binaries found to be radial velocity variable, is needed to measure their orbital periods. Observations, reduction, and radial velocity measurements {#sec:observations} ========================================================= We obtained spectroscopic follow-up observations for 385 systems (Sect.\[sec:spec-follow\]), and used exclusively SDSS spectroscopic data for another 861 systems, bringing the number of SDSS WDMS binaries with at least two reliable radial velocity measurements to 1246. In addition, we obtained photometric time-series for seven systems (Sect.\[sec:phot-follow\]). A log of the observations is given in the online version of this article (Table \[t:log\]), listing the name of the object, the *ugriz* magnitudes, the telescopes, the number of spectra taken and the duration of the observations. Spectroscopy {#sec:spec-follow} ------------ Throughout the period January 2006 to July 2010, we have used the 3.5m telescope at Calar Alto (CA), the 3.5 New Technology Telescope (NTT), the 4.2m William Herschel Telescope (WHT), the two 6.5m Magellan telescopes (Baade/Clay), the two 8m Gemini telescopes (GN/GS), and the 8m Very Large Telescope (VLT) to obtain spectroscopy of 385 SDSS WDMS binaries. Details on the instrumentation and setup are given in Table \[t:res\]. The data obtained at Calar Alto were reduced within MIDAS[^2] and the spectra were extracted using the optimal algorithm [@horne86-1]. The data from all other telescopes was reduced using the STARLINK[^3] packages FIGARO and KAPPA, spectra were optimally extracted using PAMELA [@marsh89-1] and wavelength-calibrated within MOLLY[^4]. For all the spectra, the wavelength calibration was checked, and if necessary corrected, for telescope/instrument flexure using the night sky lines. The spectra were flux calibrated and corrected for telluric absorption using observations of the stars BD+28$^\circ$4211, BD+33+$^\circ$2642, BD+25$^\circ$4655, Feige66, Feige110, GD108, and LTT3218. For additional details on the reduction and calibration of the data, see [@schreiberetal08-1; @rebassa-mansergasetal08-1; @nebotgomez-moranetal09-1]. In addition to our own follow-up spectroscopy, we made use of the spectroscopic data from SDSS. Every SDSS spectrum is the average of (typically) three 15min exposures. These *sub-spectra* are publically available since DR6 [@adelman-mccarthyetal08-1]. As shown by [@rebassa-mansergasetal08-1] and [@schwopeetal09-1] it is possible to use many of these sub-spectra to measure radial velocities and identify short-period WDMS binaries. We retrieved the subspectra for all 1903 systems in our input catalogue. Discarding spectra of insufficient signal-to-noise ratio, we were able to measure at least one velocity for 1147 WDMS binaries, and we measured a total of 5171 radial velocities using the SDSS data. The wavelength calibration of the SDSS spectra can be accurate to $\simeq2$ in dark nights, while in bright nights there can be shifts up to 20. A systematic error of 7 has been reported by @stoughtonetal02-1 and @yannyetal09-1. Although the SEGUE spectroscopic data has been corrected for that offset, this shift has only been applied to the combined spectra and not the individual subspectra. We take this into account by adding in quadrature 10 to the statistical errors on the radial velocities measured from Gaussian fits (see Sect. \[sec:rv\_measur\]). Radial velocity measurements {#sec:rv_measur} ---------------------------- We measured the radial velocities of the companion stars by a second-order polynomial plus two Gaussians to the observed $8183/8194$Å NaI absorption doublet. The separation of the two Gaussians was fixed to the laboratory value of the NaI doublet. We also measured the radial velocities from the line (where present) by fitting a second-order polynomial to the underlying continuum, plus a Gaussian for the line [see @rebassa-mansergasetal07-1; @rebassa-mansergasetal08-1 for further details on the method]. Radial velocities, errors, time of the observations and the telescope used for the identification of the system are available via the CDS; an excerpt is presented in Table\[t:rvs\_iden\]. Photometric follow-up observations {#sec:phot-follow} ---------------------------------- We carried out time-series photometry for 7 WDMS binaries to probe for orbital variability (Table\[t:log\]) where the available spectroscopy suggested a short orbital period. These data were obtained between September 2006 and August 2009 using the 2.5m DuPont telescope, the 2.2m telescope at Calar Alto, and the 0.8m IAC80 telescope. All observations were carried out with red filters (Bessel $R$, Sloan $r$ or $i$) to maximise the sensitivity to changes in the brightness of the secondary star (ellipsoidal modulation and/or reflection effect). The IAC80 data were reduced using standard packages in IRAF[^5]. The DuPont and Calar Alto data were de-biased and flat-fielded in a standard fashion within MIDAS, and instrumental magnitudes were derived SExtractor [@bertin+arnouts96-1]. For a full description of the photometric reduction pipeline see @gaensickeetal04-2. The results of the photometric observations are discussed in Sect.\[sec:porb\_photo\] and \[sec:incl\], and details on individual systems are given in the online version (Appendix\[sec:app2\]). Observational results ===================== Post common envelope binary identification {#sec:subspectra_ana} ------------------------------------------ Our observational program started back in 2005. The original strategy for identification of PCEBs was to take two to three spectra for each WDMS binary at random times, with the only constraint that they should be separated by at least one night. Any system displaying significant radial velocity variations was then flagged for intense follow-up spectroscopy to measure its orbital period. With the wealth of SDSS increasing with every data release, our strategy has evolved. Within DR5, [@rebassa-mansergasetal07-1] found that $\sim100$ WDMS had multiple SDSS spectra, and identified a certain fraction of PCEBs among them. As mentioned in Sect. \[sec:spec-follow\] each SDSS spectrum is typically the of three individual exposures. Occasionally, the subspectra were obtained on different nights, and for a small number of objects more than three subspectra are available. These subspectra were made publically available from DR6 onwards, and we made use of them for the identification of PCEBs. For $\sim100$ systems, the signal-to-noise ratio of the subspectra was insufficient for an accurate radial velocity measurement. In these cases, we combined the radial velocity determined from the co-added SDSS spectra with our own measurements. For those WDMS binaries that have SDSS data obtained on different nights, we usually took only one additional spectrum to have an independent measurement. ![Upper panel: frequency of the number of spectra taken per system from our spectroscopic campaign (dashed), from the SDSS (solid) and total combined number of spectra (gray). Bottom panel: distribution of time span between the first and the last spectrum taken for identification of close and wide binaries. See text for a detailed explanation.[]{data-label="g:num_spec"}](fig1.ps){width="50.00000%"} Figure\[g:num\_spec\] shows the distribution of the number of spectra per system used to separate PCEBs and wide binary candidates (upper panel), and the time elapsed between the first and last spectrum for a given object (bottom panel), with the SDSS data (solid line), our follow-up data (dashed line), and the combined data (gray). Most of the WDMS binaries have 3 SDSS subspectra which were taken in the same night, and are hence sensitive only to short orbital periods. A second peak around a few days indicates SDSS repeat observations of the same spectroscopic plate. Our own spectroscopy was typically carried out with a separation of $\sim1-3$days in visitor mode observing runs, with a long tail coming from service mode observations on the VLT and on Gemini. We make the reader note that the combined distributions are not the direct sum of the other two, there are systems that were observed only once by us but had several SDSS subspectra, or the other way round. For the identification of radial-velocity variable stars, i. e. PCEBs, we analysed the radial velocities measured from the NaI doublet and the H$\alpha$ emission line separately, since these two velocities can differ by more than the statistical errors [@rebassa-mansergasetal07-1], e.g. if H$\alpha$ originates predominantly in the inner hemisphere of the companion. We followed the approach of [@maxtedetal00-1] to evaluate the significance of the measured radial velocity differences. We calculated for each system the $\chi^2$ statistic with respect to the mean radial velocity, which represents the best fit for a non-variable system. For the hypothesis of a constant radial velocity, the $Q$ of obtaining a given $\chi^2$ is high when radial velocity are dominated by random noise, and low for intrinsically radial-velocity variable systems. The probability of a system showing high radial velocity variations is hence given by $P(\chi^2)=1-Q$. We consider strong PCEB candidates those systems showing radial velocity variations at a $>3\sigma$ significance, or $P(\chi^2)>99.73$%. The number of spectra used for each system, time span covered by these spectra, and the probabilities $P(\chi^2)$ are available via the CDS; for an extract see Table \[t:stats\]. ![Distribution of the probability of measuring strong radial velocity variations among the WDMS binaries in the sample. Those above a $3\sigma$ ($P(\chi^2)>99.73$) threshold are considered to be PCEBs. Line coding as in Fig. \[g:num\_spec\]. A zoom in the region between 0.995 and 1 is shown in the box, note that the bin size is smaller. For a detailed explanation see text.[]{data-label="g:probab"}](fig2.ps){width="50.00000%"} The distribution of $P(\chi^2)$ is shown in Fig. \[g:probab\], again for the SDSS data (solid line), our follow-up data (dashed line), and the combined data (gray). The combined probability distribution presents a pronounced peak at $P(\chi^2)=1$, corresponding to systems exhibiting a statistically significant radial velocity variation, i. e. PCEBs, and a less pronounced one close to $P(\chi^2)=0$, corresponding mainly to very wide systems, and a small contribution of PCEBs with very low inclinations resulting in projected radial velocities that are too low to be picked up by the spectroscopic data. In total we found 191 PCEBs among the 1246 SDSS WDMS binaries for which at least two radial velocity measurements are available. These numbers can be used to set a strict lower limit on the fraction of PCEBs, see Sect.\[sec:pcebfraction\] for a more detailed analysis. For clarity, as for Fig. \[g:num\_spec\], we mention that the combined distribution of $P(\chi^2)$ is not the direct sum of the individual sets and illustrate this with an example: SDSS0848+0058 has three SDSS subspectra taken within 1.4 hours, and we took two spectra separated by five nights. We obtained $P(\chi^2)=0.4573$ and $P(\chi^2)=0.4014$ respectively. Combining the data we have 5 spectra, separated by 7 years and obtain $P(\chi^2)=0.7140$. It is clear that any radial velocity survey will be subject to observational biases related to the spectral resolution employed, the number of spectra obtained for each object, and the temporal spacing of these spectra. These biases are evaluated in Sect.\[sec:biases\]. They are also responsible for the differences between the probability distribution from the SDSS data and from our own follow-up observations, which are discussed in Sect.\[sec:porb\_dis\]. Orbital period measurements --------------------------- As mentioned above, 191 WDMS binaries exhibited significant radial velocity variations and were flagged for detailed follow-up observations. 15 orbital periods were already published by [@pyrzasetal11-1; @schreiberetal08-1; @rebassa-mansergasetal08-1; @nebotgomez-moranetal09-1] and [@pyrzasetal09-1]. Here we report period determinations for 58 additional systems, four of which were measured from follow-up photometry, the rest from follow-up spectroscopy. ### Orbital periods from photometric observations {#sec:porb_photo} Photometric variability in PCEBs can be caused by four different mechanisms: ellipsoidal modulation due to the distorted shape of the secondary star, reflection effect on the surface of the secondary star due to irradiation from the hot primary, eclipses, and rotational modulation due to starspots (BY Dra phenomenon). The first three types of variability are, by definition, modulated on the orbital period. For most PCEBs, also variability due to star spots will be modulated on the orbital period, as the companion will by tidally synchronised. However, for wide binaries, star spots on the companion star may mimic variability of a short-period PCEB. Four out of the seven WDMS binaries for which we obtained time-series photometry (Table\[t:log\]) show ellipsoidal modulation, and we determined their orbital periods by fitting sine curves to the observations. Periodograms and phase folded light curves are presented and discussed in more detail in Appendix \[sec:app2\]. ### Orbital periods from spectroscopic observations {#sec:Porb} We obtained further spectroscopy for 68 of the 191 PCEBs identified by our radial velocity survey (Table\[t:log\]). A period search was performed by computing periodograms adopting @schwarzenberg-czerny96-1’s (1996) ORT/TSA method within MIDAS, which fits periodic orthogonal polynomials to the observations, combined with an analysis of variance statistic. In a second step, we fitted sine curves of the form $$v_r = \gamma_{\mathrm{sec}} + K_{\mathrm{sec}} \sin\left[\frac{2\pi(t-t_\mathrm{0})}{{\mbox{$P_\mathrm{orb}$}}}\right],$$ to the radial velocity data, with $\gamma_{\mathrm{sec}}$ the systemic velocity of the secondary star, $t_\mathrm{0}$ the zero point defined by the inferior conjunction of the secondary star, and $K_{\mathrm{sec}}$ the radial velocity amplitude of the secondary star. Initial values for ${\mbox{$P_\mathrm{orb}$}}$ were selected from the highest peaks in the periodograms. We adopt as orbital period the values of the best-fit solution, which was always coinciding with the highest peak in the periodogram. We could measure orbital periods for 58 of the 68 PCEBs that were spectroscopically followed up. For the remaining 10 systems the radial velocity data was insufficient to unambiguously identify the true orbital periods. The radial velocity variations measured for these 10 systems range from 50 to 490 so a priori there is no reason to think that their orbital period distribution differs significantly from the one drawn from the rest of the population of PCEBs. The 58 orbital periods range from $1.97$ to $104.56$ hours (0.08 to 4.35 days) (Table \[t:allparam\]). The spectra of these PCEBs, and their phase-folded radial velocity curves are shown in Fig. \[g:spectra1\], \[g:spectra2\] and Fig. \[g:rv1\], \[g:rv2\], respectively, and their stellar and binary parameters given in Table \[t:allparam\], which also lists $K_{\mathrm{sec}}$ and $\gamma_{\mathrm{sec}}$. The orbital period distribution of the SDSS PCEBs is discussed in more detail in Sect.\[sec:porb\_dis\]. ### Orbital inclinations and candidate eclipsing systems {#sec:incl} We estimated the orbital inclinations and binary separations of those PCEBs for which masses of the white dwarf and the companion star (Sect. \[sec:sample\]), $K_2$ and ${\mbox{$P_\mathrm{orb}$}}$ were available using Kepler’s third law, and propagated the errors in all quantities accordingly. We further calculated $K_\mathrm{WD}$, as well as the Roche-lobe size of the companion star, $R_\mathrm{lob}$ using @eggleton83-1’s (1983) approximation. Finally, using the empirical radius-spectral type relation of [@rebassa-mansergasetal07-1] we estimate the Roche-lobe filling factors of the secondary stars. All this ancillary information is given in Table \[t:allparam\]. In case the effective temperature of the primary is smaller than $12~000$ K, the derived spectroscopic mass is inherently uncertain [@koesteretal09-1; @tremblayetal10-1] and the derived parameters should be used with caution. For six systems, the estimated orbital inclinations are close to 90 degrees, making them good candidates for being eclipsing binaries: SDSS0238–0005, SDSS1348+1834, SDSS1434+5335, SDSS1506–0120, SDSS2132+0031 and SDSS2318–0935. Eclipsing systems offer the best opportunity to derive accurate fundamental stellar parameters, i.e. masses and radii, and we have begun to exploit the potential offered by the eclipsing PCEBs [@steinfadtetal08-1; @nebotgomez-moranetal09-1; @pyrzasetal09-1]. SDSS1348+1834, aka CSS21357, was recently identified in the Catalina Real Time Transient Survey [@drakeetal09-1] as an eclipsing PCEB [@drakeetal10-1], confirming the predictive power of our estimates. ### Observed orbital period distribution {#sec:obspdist} We have measured here the orbital periods of 58 PCEBs. Together with the 15 periods already published by our group [@pyrzasetal11-1; @rebassa-mansergasetal08-1; @schreiberetal08-1; @pyrzasetal09-1; @nebotgomez-moranetal09-1], two more by [@drakeetal10-1], and four additional PCEBs that are subject to forthcoming individual studies (Pyrzas et al in prep.; Parsons et al in prep.), the total number of PCEBs from SDSS with orbital period measurements stands at 79, which is more than double the corresponding number of pre-SDSS PCEBs. The orbital period distribution of the SDSS PCEBs is shown in Fig. \[g:porbs58\] on a logarithmic scale, in light and dark gray those systems that were identified as PCEBs based on spectroscopy (separating systems with PCEB identification spectra taken in the same night, light gray) and photometrically identified PCEBs are shown in medium gray. The 58 periods determined in this paper are indicated by vertical lines. Taking this observed period distribution at face value, PCEBs are found to have orbital periods ranging from 1.97h to 4.35d, they follow approximately a normal distribution in $\log({\mbox{$P_\mathrm{orb}$}})$ with a peak centred on $\sim8.1$h and a width of $0.41$ in $\log({\mbox{$P_\mathrm{orb}$}}[d])$ as determined from a Gaussian fit. Analysis {#sec:porb_dis} ======== We have performed a radial velocity survey of 1246 WDMS binaries. The majority of the systems have three radial velocity measurements, spread over hours to hundreds of days (see Fig.\[g:num\_spec\]). We have identified 191 of these systems as PCEBs based on the detection of significant radial velocity variations. While we consider the remaining 1055 systems to be candidate wide binaries, it is clear that there will still be a number of PCEBs among them which our radial velocity survey failed to identify because of observational selection effects. Our ultimate aim is to establish the intrinsic period distribution of the SDSS PCEBs. We will first investigate in the following sections what information might be gleaned from the distribution of probabilities of our initial radial velocity survey (Fig.\[g:probab\]). Subsequently, we will analyse the observational biases within the orbital period distribution of the SDSS PCEBs (Fig.\[g:porbs58\]), and derive a bias-corrected period distribution, that will then be discussed in Sect.\[sec:discussion\] in the context of compact binary evolution. Does our radial velocity survey constrain the orbital period distribution of WDMS binaries? ------------------------------------------------------------------------------------------- The detailed shape of the probability distribution used to identify PCEBs among SDSS WDMS binaries will depend on the technical details of our radial velocity survey (spectral resolution, temporal sampling, number of spectra) and the intrinsic period distribution of the WDMS binaries. To test how sensitive the observed probability distribution is to differences in the underlying orbital period distribution, we performed extensive Monte Carlo simulations. As input, we use the measured stellar masses (Table\[t:allparam\]), the times of observation, and the radial velocities and associated errors (Table\[t:rvs\_iden\]) for all 1246 WDMS binaries within our radial velocity survey. We then adopted four different orbital period distributions: (1) short-period binaries only, with $2\,\mathrm{h}<{\mbox{$P_\mathrm{orb}$}}<100$d, and uniformly distributed in $\log({\mbox{$P_\mathrm{orb}$}})$; (2) wide binaries only, with $3.5<\log({\mbox{$P_\mathrm{orb}$}}[\mathrm{d}])<6.5$, uniformly distributed in $\log({\mbox{$P_\mathrm{orb}$}})$; (3) a bi-modal distribution with a short-period binary population peaking at 7.5h, normally distributed in $\log({\mbox{$P_\mathrm{orb}$}})$ and with a minimum period of $\sim 1.5$h, plus a second population of wide binaries with $3.5<\log({\mbox{$P_\mathrm{orb}$}}[\mathrm{d}])<7.5$, uniformly distributed in $\log({\mbox{$P_\mathrm{orb}$}})$, and finally (4) a uniform period distribution in $\log({\mbox{$P_\mathrm{orb}$}})$ with $\log({\mbox{$P_\mathrm{orb}$}}[\mathrm{d}])<8.5$. Obviously (1), (2) and (4) are very unlikely to be physically relevant, but were adopted to run the test with several radically different period distributions. We then used $10^4$ random inclinations per object, uniformly distributed in $\cos i$, and calculated the probability distributions of the two adopted period distributions. We assumed circular orbits for simplicity (and note that the orbits of PCEBs will be, in most cases, circularised). The results of this simulation are illustrated in Fig.\[g:SDSS\_Porbs\], showing the adopted period distributions on the left hand side, and the resulting probability distributions on the right hand side, with the observed probability distribution shown for comparison. The top two panels show that a peak at $P(\chi^2)=1$ requires close binaries, whereas a broad accumulation towards $P(\chi^2)=0$ is dominated by wide binaries. In other words, the observed probability distribution (Fig.\[g:probab\]) can only be explained if the underlying orbital period distribution of the observed WDMS binaries contains both wide binaries (${\mbox{$P_\mathrm{orb}$}}{\raisebox{-0.4ex}{${\stackrel{>}{\scriptstyle \sim}}$}}10^3$d) and PCEBs with shorter periods (${\mbox{$P_\mathrm{orb}$}}{\raisebox{-0.4ex}{${\stackrel{<}{\scriptstyle \sim}}$}}\,100$d). However, the bottom two panels show that the two period distributions which contain both short-period PCEBs and wide binaries, but radically differ in shape, both lead to probability distributions that are very similar to the observed one. We conclude that our radial velocity survey carried out to identify PCEBs alone cannot constrain the detailed shape of the underlying WDMS binary orbital period distribution, and that it is hence necessary to accurately measure the orbital periods for a large and homogeneous sample of PCEBs. Observational biases and selection effects {#sec:biases} ------------------------------------------ We have presented in Sect.\[sec:obspdist\] the orbital period distribution of 79 PCEBs that were spectroscopically identified by SDSS (Fig.\[g:porbs58\]). The resulting orbital period distribution peaks at $\sim8$hrs and very few systems with orbital periods longer than a day have been found. This is not only the largest but also the most homogeneous sample of PCEBs available so far, which has a large potential to provide crucial constraints on the theories of close compact binary formation and evolution [e.g. @schreiberetal10-1; @zorotovicetal10-1; @rebassa-mansergasetal11-1]. It is, however, of fundamental importance to understand any observational selection effects that may affect the properties of the SDSS PCEB sample. In the context of the orbital period distribution, there are two particular points that need to be investigated. Firstly, our radial velocity PCEB identification method is obviously more sensitive to short orbital period and high inclination systems and secondly, we only measured the orbital periods of a sub-sample of the identified PCEBs which is not necessarily representative for the entire sample. ### PCEB identification {#sec:pceb_iden} While we have compiled the largest and most homogeneous sample of PCEBs to date, not all the systems in this sample have been identified in the same way. Ten systems were initially classified as PCEBs because of photometric variability [^6] and four because of radial velocity variations seen over the course of a single night[^7]. Both methods favour the identification of short orbital period PCEBs. This is illustrated in Fig. \[g:porbs58\], where all those 14 systems have periods $\la9$h (0.4d), and we discard them in the following analysis. The remaining 65 PCEBs were identified by radial velocity variations from spectra that have been obtained on different nights, and we will concentrate our analysis of the observational biases on this “core” sample. The strength of the observational bias introduced in a radial velocity search for binaries depends on the intrinsic orbital period and the sampling frequency. We have performed Monte-Carlo simulations similar to those presented in @schreiberetal08-1 [@rebassa-mansergasetal08-1; @schreiberetal10-1] to evaluate if the observed orbital period distribution (Sect. \[sec:Porb\]) is significantly biased by the design of our radial velocity survey, and we calculated the detection probability as a function of orbital period for all 65 PCEBs. Given the times of observation and the error of the radial velocity measurements, we randomly select $10^4$ phases and inclinations for a given orbital period, and determine the corresponding $\chi^2$ probability. The PCEB detection probability for each system as a function of the orbital period is then simply given by the number of cases with $P(\chi^2)>0.9973$ ($>3\sigma$) divided by $10^4$. The resulting detection probabilities depend on the number of spectra taken, the temporal sampling of the measurements, and the accuracy of the radial velocity measurement (i.e. basically the spectral resolution of the telescope/instrument combination that has been used). Figure \[g:biases2\] shows the obtained detection probabilities for three rather extreme examples: SDSS0807+0724 (green, ${\mbox{$P_\mathrm{orb}$}}=11.45$h, identified as a PCEB from two VLT/FORS spectra), SDSS0152–0058 (blue, ${\mbox{$P_\mathrm{orb}$}}=2.15$h, $3\times$WHT, 2$\times$Magellan, 1$\times$SDSS), and SDSS0225+0055 (red, ${\mbox{$P_\mathrm{orb}$}}=21.86$h, 16 SDSS subspectra). We have chosen these three cases to illustrate that our PCEB identification method was sensitive to PCEBs with orbital periods in the range of $1-10$days (detection probabilities $P(\chi^2){\raisebox{-0.4ex}{${\stackrel{>}{\scriptstyle \sim}}$}}0.35$) as long as the time baseline exceeded one night. In addition, Fig. \[g:biases2\] illustrates the following: (1) the lower the resolution, the lower is the detection probability, even in the most favourable case, i.e. a large number of spectra. (2) Variations in the detection probability with orbital period are dominated by phase coverage and are therefore strongest when only few spectra are available. Orbital periods near one day and their first subharmonics (12h, 8h) have substantially depressed detection probabilities, as this increases the chances to sample the same orbital phase twice on two subsequent nights. (3) Superimposed on the fine structure related to the details of the sampling is a decreasing envelope, which is caused by a larger range of (low) inclinations leading to radial velocity variations that are too low to be resolved. The mean detection probability for the whole sample, calculated from the 65 individual detection probabilities, is higher than $35\%$ up to 10 days orbital periods (Fig.\[g:biases\], upper panel), demonstrating that our radial velocity survey is well capable of identifying PCEBs with long orbital periods. We will use this mean detection probability to bias-correct the observed orbital period distribution in Sect. \[sec:porbbiascor\]. ### PCEBs selected for orbital period measurements {#sec:biasporb} Given that we have measured orbital periods only for a fraction of the PCEBs that we identified, the next question to address is whether the subset of systems is representative of the whole sample. For instance, our follow-up strategy might have been biased towards the “easier” targets, i.e. those which exhibited large variations in our radial velocity survey. For this purpose, we compared the cumulative radial velocity variations of the 65 PCEBs identified in our radial velocity survey (see Sect \[sec:subspectra\_ana\]), and of those PCEBs for which we obtained no additional follow-up observations (Fig.\[g:cdfs\_rv\]). A Kolmogorov-Smirnov test between both sets gives a probability of 0.97 of both samples being drawn from the same parent population. We can conclude that both distributions are very similar, and that we have measured the orbital period of a representative sub-sample of the SDSS PCEB population. ### Bias corrected orbital period distribution {#sec:porbbiascor} We have demonstrated in Sect.\[sec:biasporb\] that the 65 PCEBs that were identified from multiple-night radial velocity snapshots and have measured orbital periods are representative of the entire SDSS PCEB sample. Here, we apply the average detection probability (Sect.\[sec:pceb\_iden\]) to correct their observed period distribution (dark gray histogram in Fig. \[g:biases\]). The upper panel of Fig.\[g:biases\] shows the detection probability averaged across all 65 systems, and its binned values as red dots. The bias-corrected period distribution is shown in Fig.\[g:biases\] (white histogram). It is evident that the observational bias of our radial velocity survey does not dramatically affect the conclusions drawn in Sect.\[sec:obspdist\], i.e. the period distribution of the SDSS PCEB sample follows broadly a normal distribution centred at $\sim10.3$h (0.43d), a width of $0.44$ in $\log({\mbox{$P_\mathrm{orb}$}}[d])$, a short-period cut-off at $\sim2$h and a long-period cut-off at $\sim5$d, where the position of the peak and the width have been determined from a Gaussian fit (dashed line). The bias-corrected period distribution is discussed in the context of published binary population models in Sect.\[sec:comp\_bps\]. Discussion {#sec:discussion} ========== The PCEB fraction among WDMS binaries {#sec:pcebfraction} ------------------------------------- Analysing the radial velocity variations of 1246 SDSS WDMS binaries we identified 191 PCEBs (Sect. \[sec:subspectra\_ana\]), which provides a lower limit of $\sim15\pm1$% to the fraction of PCEBs among the SDSS WDMS binaries. However, for a substantial number of these 1246 systems the only available source of radial velocities are the SDSS subspectra, which, as they are usually taken during a single night (see Fig.\[g:num\_spec\]), are sensitive only to very short orbital periods. If we only consider the sub-sample of 718 WDMS binaries that were observed on different nights, we find 164 PCEBs, i.e. a PCEB fraction of $\sim 23\pm2$%. This number is similar to that found by [@schreiberetal08-1], based on our VLT/FORS radial velocity survey of 26 SDSS WDMS binaries. This empirical determination of the PCEB fraction is broadly consistent with the results of the binary population synthesis carried out by [@willems+kolb04-1]. Two observational selection effects will affect this empirical estimate of the PCEB fraction. On the one hand, our analysis in Sect.\[sec:pceb\_iden\] showed that a number of PCEBs will escape identification because of their low orbital inclination. From the mean detection probability and the bias corrected orbital period distribution we learn that $\sim16\%$ of the PCEBs are not detected due to a combination of inclination and phase coverage (see Fig. \[g:biases\]). Assuming that the 164 detected PCEBs (observed in different nights) follow the same orbital period distribution as the observed one (drawn from 65 of them) the number of expected PCEBs is 195, implying a fraction of PCEB to WDMS binaries of $27\pm2\%$ ($195/718$). On the other hand, WDMS binaries that are sufficiently wide will appear as blended objects, or even fully resolved objects, and not have been followed up spectroscopically, as they have unreliable colours or colours of single white dwarfs and M-dwarfs, respectively. Hence, our estimate of $27\pm2\%$ PCEBs among the spatially unresolved SDSS WDMS binaries can be taken as an upper limit on the intrinsic fraction of PCEBs among all WDMS. @schreiberetal10-1 estimated the fraction of SDSS resolved WDMS to be 12-23%. Taking this into account we finally estimate the intrinsic fraction of PCEBs to be $\sim21-24$%. @farihietal06-1 [@farihietal10-1] pursued a complementary approach to investigate the make-up of the known WDMS binary population, obtaining high spatial resolution imaging with *HST*/ACS. They observed a total of 72 WDMS binaries unresolved from the ground, and found a bimodal distribution in projected binary separations: all spatially resolved systems had binary separations $\ga10$AU, clearly above the detection threshold of $1-2$AU. Accounting for the frequency of WDMS binaries that can be resolved from the ground @farihietal05-1 [@farihietal10-1] conclude that the fraction of short-period binaries among all WDMS binaries is $\simeq25$%, which is in excellent agreement with our estimate presented here. 19 of the WDMS binaries imaged with HST were also spectroscopically re-discovered by SDSS, and are hence in our WDMS binary sample (Table\[t:farihi\]). @farihietal10-1 predict that all the WDMS binaries unresolved by *HST* (Table5 in their paper) should be short-period PCEBs, which is confirmed for all five systems for which we have obtained follow-up observation, spanning more than one night. For an additional two of Farihi’s unresolved systems (SDSS1341+6026, 1500+1659) we have only the radial velocity information of the SDSS subspectra spanning less than one hour, and we classified both systems as wide binary candidates. This underlines what we said above, i.e. that the single-night SDSS subspectra allow us to identify only the shortest-period WDMS binaries (which we fully take into account by our Monte Carlo analysis of the selection biases). For completeness, we note that one additional unresolved binary from @farihietal10-1 was recently found to be an eclipsing WDMS binary with an orbital period of 2.93h @drakeetal10-1. Vice versa, based on our radial velocity information for nine of the WDMS binaries spatially resolved by *HST* (Table4 of @farihietal10-1), we classified all of them as wide binary candidates (labeled as “wide” in the bottom portion of Table \[t:farihi\]). PCEBs and detached CVs in the orbital period gap ------------------------------------------------ The observed orbital period distribution of CVs presents a sharp decrease of systems between 2–3h [@warner76-1; @whyte+eggleton80-1; @knigge06-1], referred to as the orbital period gap of CVs, which is interpreted as a key to understanding their long-term evolution. The mass ratio of CVs implies that they are undergoing stable mass transfer, and it is angular momentum loss that drives them towards shorter orbital periods, keeping the companion stars in Roche-lobe contact. Above the period gap, the secondary stars have a radiative core, and are subject to magnetic wind braking that acts as a strong agent of angular momentum loss. The corresponding mass loss rates are sufficient to drive the companion stars of these long-period CVs out of thermal equilibrium. At $\sim3$h, the companion stars become fully convective, and magnetic wind braking is thought to stop. From there on, the evolution towards shorter periods is driven by gravitational wave radiation, resulting in much lower mass loss rates. Consequently, the secondary stars have time to relax to their thermal equilibrium radii, shrink within their Roche lobes, and turn into detached WDMS binaries with no mass transfer. Gravitational wave radiation keeps driving these detached systems towards shorter periods. At $\sim2$h the secondary stars re-establish Roche-lobe contact, mass transfer stars again, and the systems re-appear as CVs. This “disrupted magnetic braking” scenario is currently the standard paradigm of CV evolution [@rappaportetal83-1]. Observationally, there is very little difference between genuine pre-CVs that will start mass transfer within, or shortly below, the CV period gap, and those detached CVs (dCVs) that are crossing the gap. The common hallmark of both populations is that they are WDMS binaries with secondary spectral types $\sim$M6 to $\sim$M4. A very simple, but so far untested, prediction of the disrupted magnetic braking scenario is that there should be an *excess* of systems with periods of 2–3h in the period distribution of WDMS binaries, due to dCVs that are crossing the orbital period gap. [@davisetal08-1] carried out a number of binary population studies to estimate the strength of this effect, and predict that within the CV period gap the number of dCVs should exceed that of genuine pre-CVs with secondary stars in the mass range $0.17-0.36$ by a factor $\sim4-13$, which implies an increase of systems (PCEBs plus dCVs) of 1.5–2 in the gap in comparison with PCEBs at longer periods (see their Figs.4 &10). The large sample of SDSS PCEBs provides a first opportunity to test this prediction. Observationally, the secondary stars in the SDSS WDMS binaries are characterised in terms of their spectral type. CVs just above and below the period gap have $\sim$M3–4, and $\sim$M5 secondary stars, respectively. We show in Fig.\[g:porb\_gap\] the period distributions of all SDSS WDMS binaries with M3–4 companions (top panel), and a somewhat broader range of M2–5 (bottom panel) which accounts, somewhat conservatively, for the uncertainty in the spectral types of the companion stars. The expected increase towards shorter periods by a factor of 1.5–2 is not present in our data. In contrast the number of dCVs plus PCEBs is slightly decreasing (M3–4) or constant (M2–5), although the number of systems in this sub-sample is still rather small. Comparing the number of dCVs plus PCEBs in the gap with the number of PCEBs between 3–4 hours, we obtain a fraction of dCVs plus PCEBs (in the gap) to PCEBs (above the gap) of $0.25\pm0.30$ (M3–4) or $0.43\pm0.30$ (M2–5), which excludes the predicted increase of 1.5–2 on a $3\sigma$ (M3–4) or a $2\sigma$ (M2–5) level. Three SDSS PCEBs that match the expected characteristics of dCVs are SDSS0052–0053, SDSS1611+4640 and, SDSS2243+3122, with orbital periods 2.74, 1.98 and 2.87h respectively. The Roche-lobe filling factors of SDSS0052–0053 and SDSS1611+4640 are $1.0 \pm 0.6$ and $1.52 \pm 0.50$ respectively. For SDSS2243+3122 we have no estimate of the (DC) white dwarf mass, and hence can not determine the filling factor. ![Orbital period distribution of PCEBs with secondary spectral types in the range M3–M4 (upper panel) and M2–M5 (bottom panel). Magnetic braking disruption predicts a peak at around 2–3 hours corresponding to an increase in systems due to detached CVs, we highlighted the region in gray.[]{data-label="g:porb_gap"}](fig12.ps){width="50.00000%"} Comparison with non SDSS PCEBs ------------------------------ In this section we compare the orbital period distribution of PCEBs discovered within the SDSS, with that of PCEBs that were discovered by other surveys. We compiled a list of 48 non-SDSS PCEBs containing a white dwarf plus a main sequence star from @ritter+kolb03-1 ([-@ritter+kolb03-1], version v7.14) and [@brownetal11-1]. The orbital period distribution of non-SDSS PCEBs (Figure \[g:Porbs\], gray histogram) is overall similar to that of the SDSS PCEBs (white hatched histogram), presenting both a peak at around $\sim5$h and decreasing number of systems towards long orbital periods. A Kolmogorov-Smirnov test between the cumulative period distributions (bottom panel in Figure \[g:Porbs\]) gives a probability of 35%, i.e. consistent with the two samples being drawn from the same parent population. We note that this result is in agreement with the conclusions of [@miszalskietal09-1], who first noticed the resemblance between the orbital period distribution of binary central stars in planetary nebulae (PNe) and that one of PCEBs as presented by @rebassa-mansergasetal08-1. This was later confirmed by [@hillwig11-1], who selected 27 WDMS binaries that are the central stars of planetary nebulae (PNe) and the 35 SDSS WDMS binaries (as published by @zorotovicetal10-1). Surveys dedicated to measure the orbital period of the binary central stars of PNe have been mostly photometric [see @bond00-1 and references therein], implying a bias towards short orbital periods. [@demarcoetal08-1] studied the possible biases introduced in the photometric sample, and concluded that the decrease of the number of PNe towards long periods is unlikely to be due to observational biases. Comparison with predictions from BPS models {#sec:comp_bps} ------------------------------------------- In this section we discuss the SDSS WDMS binary sample in the context of published binary population synthesis models. However, it is necesary to keep in mind that the SDSS PCEB sample is subject to selection effects, i.e. systems with companion spectral types earlier than $\sim$M0 are severely underrepresented, and a comparison of the predicted orbital period distribution of the different BPS studies for the corresponding mass domain is in most cases not possible. Common to all BPS models is the prediction that the population of WDMS binaries should be made up of wide binaries, with orbital periods longer than one year, in which the components never interacted, and close binaries which formed from common envelope evolution. The two most important parameters determining the outcome of the CE phase are the initial mass ratio distribution of the progenitor binaries, and the efficiency of the ejection of the common envelope, $\alpha_\mathrm{CE}$, which will affect the position of the peak in the orbital period distribution of the PCEBs, and the slope towards longer orbital periods. One of the first BPS for WDMS binaries was carried out by [@dekool92-1] and [@dekool+ritter93-1], predicting that most PCEBs should have periods shorter than $\sim$ one day. An extended tail towards longer orbital periods up to 1000 days is expected for $\alpha_\mathrm{CE}=1$, while no tail is observed for $\alpha_\mathrm{CE}=0.3$. The authors also simulated the expected properties of a volume limited blue sample of present-day WDMS binaries, however, comparison with the very small number of PCEBs at that time did not lead to any significant observational constraints. Our orbital period distribution (Fig. \[g:biases\]) is very similar to their predicted one for $\alpha_\mathrm{CE}=0.3$ and the mass of the secondary star picked randomly from the IMF, already pointed out by [@miszalskietal09-1] in the context of PN, or with $dN\propto dq$ (see their Fig.12b). However, this comparison should be regarded with some caution, as [@dekool+ritter93-1] simulated their “observational” sample for the selection criteria of the Palomar-Green Survey [@greenetal86-1] ($U<16$ and $U-B<-0.5$), whereas the SDSS WDMS sample has a much larger limiting magnitude $i\simeq19.1$ and is drawn from a much larger $ugriz$ colour volume. [@willems+kolb04-1] carried out an extensive BPS study of WDMS binaries, using $\alpha_\mathrm{CE}=1$, and for different initial mass ratio distributions. Their BPS predicts that the orbital period distribution of PCEBs peaks around 1 day, with a substantial tail towards longer orbital periods of up to 100 days, which is in contrast to the period distribution of the SDSS PCEBs (Fig. \[g:biases\]). However, [@willems+kolb04-1] did not convolve their model WDMS binary population with detailed observational selection effects, but just discussed their results in terms of luminosity ratio distributions, $L_\mathrm{wd}/L_\mathrm{MS}$, which prevents any more quantitative comparison with our SDSS WDMS binary sample. [@politano+weiler07-1] investigated the effects on the population of PCEBs for $\alpha_\mathrm{CE}$ depending on the mass of the secondary star. They predict an orbital period distribution peaking at around 3 days, which is incompatible with the observed period distribution of the SDSS WDMS binaries. However, also @politano+weiler07-1 did not simulate a realistic “observational” WDMS population, again preventing a quantitative comparison with our sample. The most recent BPS of WDMS binaries was performed by [@davisetal10-1], who, in contrast with previous works, calculate the binding energy parameter $\lambda$ from stellar evolution models incorporating the internal energy of the envelope. While this is a significant step forward compared to previous PBS that worked with a constant value of $\lambda$, [@davisetal10-1] obtained the values of $\lambda$ from the tables of [@dewi+tauris00-1], which cover only a very small part of the relevant parameter space. The orbital period distribution of [@davisetal10-1] predicts a long tail of systems with periods $>1$d, inconsistent with the period distribution of the PCEBs known at the time of publication (their Fig.10), even when taking into account the detection probabilities of [@rebassa-mansergasetal08-1]. Our more detailed observed (Fig.\[g:porbs58\]) and bias-corrected (Fig.\[g:biases\]) orbital period distributions are in better agreement with their predicted orbital period distribution limited to PCEBs with companion masses $\le0.5$, i.e. spectral types later than $\sim$M0 (see their Fig.10 middle panel). In summary, most published BPS models predict a substantial number of PCEBs with orbital periods $>1$d, which is currently incompatible with the orbital period distribution of the SDSS PCEBs. The observed dearth of long-period PCEBs could be reconciled with a low value of $\alpha_\mathrm{CE}$. [@zorotovicetal10-1] used a sub-set of 35 of the PCEBs presented here, along with some other 25 non-SDSS PCEBs, to constrain the value of $\alpha_\mathrm{CE}$ by reconstructing their evolution. They found that there seems to be no dependence of $\alpha_\mathrm{CE}$ on the mass of the secondary star, and that if the value is universal only a small value of $\alpha_\mathrm{CE}$ can explain the short orbital periods observed among the known PCEBs. However, it is also clear that even the SDSS PCEB sample is subject to selection effect, i.e. systems with companion spectral types earlier than $\sim$M0, i.e. masses $\ga0.5$ are severely underrepresented, which are expected to have longer orbital periods than the currently dominant population of PCEBs with low-mass (mid M-type) companions. Conclusions =========== Nearly 2000 WDMS binaries have been spectroscopically identified within the SDSS survey. A detailed characterisation and analysis of this large and homogeneously selected sample promises to improve substantially our understanding of close binary evolution, in particular of the common envelope phase. Using follow-up observations of 385 WDMS binaries, combined with the SDSS subspectra for an additional 864 systems, we identified 191 PCEBs. We determine orbital periods for 58 of these systems, increasing the total number of SDSS PCEBs with known orbital periods to 79. We analyse the observational biases inherent to the SDSS PCEB sample, and find that $21-24$% of all SDSS WDMS binaries with M dwarf companions must have undergone common envelope evolution. The bias-corrected orbital period distribution of the SDSS PCEBs shows a peak near $\sim10.3$h, and drops fairly steeply to both shorter and longer periods, with no systems at ${\mbox{$P_\mathrm{orb}$}}<1.9$h, and few systems at $P>1$d. Comparing the properties of the SDSS PCEB population to the results of published binary population, we conclude that the observed dearth of long-period systems can be explained by low values of $\alpha_\mathrm{CE}$. Observational selection effects against earlier-type companions can also play a role. Given a Roche lobe-filling giant or asymptotic giant (the progenitor of the white dwarf component) in a binary system, the more massive the companion star, the greater the orbital energy available to eject the common envelope, and therefore the longer the orbital period of the remnant PCEB, other things being equal. Unlocking the diagnostic potential of the SDSS WDMS binary population now requires the computation of a state-of-the art binary population model that includes a realistic simulation of the observational selection effects inherent to this sample. Based on observations collected at: the European Organisation for Astronomical Research in the Southern Hemisphere, Chile (078.D-0719, 079.D-0531, 080.D-0407, 082.D-0507, 083.D-0862, 085.D-0974); the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the Science and Technology Facilities Council (United Kingdom), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), Ministério da Ciência e Tecnologia (Brazil) and Ministerio de Ciencia, Tecnología e Innovación Productiva (Argentina) (GS-2007B-Q-27, GS-2008A-Q-31, GS-2008B-Q-40, GS-2009A-Q-56); the Magellan Telescopes Baade and Clay, and the Iréné Dupont Telescope located at Las Campanas Observatory, Chile; the William Herschel Telescope, operated on the island of La Palma by the Isaac Newton Group in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofísica de Canarias; the IAC80 operated on the island of Tenerife by the Instituto de Astrofísica de Canarias in the Spanish Observatorio del Teide; and at the Centro Astronómico Hispano Alemán (CAHA) at Calar Alto, operated jointly by the Max-Planck Institut für Astronomie and the Instituto de Astrofísica de Andalucá. Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the U.S. Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, the Max Planck Society, and the Higher Education Funding Council for England. The SDSS Web Site is http://www.sdss.org/. The authors would like to thank the referee, R. Webbink, for carefully reading the manuscript and for providing useful remarks which helped improving the quality of the paper. We would like to thank T. Forveille and B. Miszalski for useful comments. ANGM acknowledges support by the Centre National d’Etudes Spatial (CNES, ref. 60015). BTG was supported by a rolling grant from STFC. ANGM and MRS acknowledge support by the Deutsches Zentrum für Luft und Raumfahrt (DLR) GmbH under contract No. FKZ 50 OR 0404. MRS was also supported by FONDECYT (grant 1100782), and the Center of Astrophysics at the Universidad de Valparaiso. ARM acknowledges financial support from FONDECYT (grant 3110049). This work was supported by the Deutsche Forschungsgemeinschaft (DFG) under contract numbers Schw 536/26-1, 29-1, 30-1, 31-1, 32-1, 33-1, and 34-1. JS acknowledges financial support from STFC in the form of an Advanced Fellowship. Optical light curves {#sec:app2} ==================== Photometric follow-up observations were carried out for a total of seven WDMS binaries where the available spectroscopy suggested a short orbital period. Telescope, filter and duration of the observations are listed in Table \[t:log\]. Four out of the seven observed systems showed ellipsoidal modulation: SDSS0853+0720, SDSS1611+4640, SDSS2112+1014, and SDSS2243+3122. We determined their orbital periods by computing periodograms and fitting sine curves to the phase folded light curves (see Fig. \[g:sdss0853\_2243\]). We estimated ${\mbox{$P_\mathrm{orb}$}}=3.6, 1.9, 2.2,$ and $2.8$ hours respectively. Spectroscopic follow-up observations were carried out for SDSS0853+0720, SDSS1611+4640, and SDSS2243+3122 covering a full cycle, and confirming the measured periods. We could connect the spectroscopic and photometric runs for SDSS1611+4640, and SDSS2243+3122, without cycle count aliases, which allowed us to refine their orbital periods. The light curves of SDSS0853+0720 and SDSS2243+3122 show both two uneven maxima at phases $0.25$ and $0.75$. For SDSS0853+0720 the first maximum is brighter than the second one, while for SDSS2243+3122 is the other way round. In both systems equal minima at phases $0$ and $0.5$ are observed and the maximum variation is almost $0.2$ magnitudes in the *r* and the *i* bands respectively. Ellipsoidal modulation and spots in one of the hemispheres of the secondary star could explain the uneven maxima (O’Connell effect, see [@liuandyang03-1]). Photometry carried out on the 15th of August 2009 for SDSS2243+3122 revealed a flare with $\sim25$ minutes length in the decay, and rising by $0.35$ magnitudes in the *i* band (see Figure \[g:flare\]). This would strengthen the argument of magnetic activity being the cause of the observed disparity in the maxima (see Figure \[g:sdss0853\_2243\]). Although no radial velocity curve is available for SDSS2112+1014, the five available spectra show radial velocity variations of more than 300 in less than 40 minutes, indicating a rather short orbital period, consistent with the measured photometric orbital period of $2.2$ hours. Notes on individual systems =========================== **SDSS0225+0054**, the primary star was classified as a DC by [@kleinmanetal04-1]. Although it is true that no absorption lines are visible in the primary star we consider that the signal to noise of the residual white dwarf after the subtraction of the best fitted secondary template is insufficient for giving any other classification than white dwarf [see @rebassa-mansergasetal07-1 for details on the spectral analysis]. SDSS0225+0054 together with **SDSS0853+0720** are WDMS binaries consisting of cold DC white dwarf and an $\mathrm{M4.5}$ and $\mathrm{M3}$ companion star, clear examples of the old PCEBs predicted by @schreiber+gaensicke03-1 that would not have been found in any search for blue objects. **SDSS0307+3848** (SDSSJ030716.44+384822.8) has a red star nearby - within 5- (SDSSJ030716.16+384821.4) and is slightly brighter in the r band (16.70 versus 15.17). To be sure that there was no contamination in our spectra by this source we rotated the slit so as to take spectra simultaneously and compared the radial velocities with those of this system. No variations in the radial velocity of this visual companion were observed. The optical spectrum of SDSS0307+3848 is dominated by the secondary star and it presents strong emission lines. We measured the radial velocities from the line in the SDSS-subspectra. The and the lines are in phase and the amplitudes were consistent with each other within the errors. **SDSS1006+0044**, **SDSS1523+4604** and **SDSS1844+4120** have very cold WDs, ${\mbox{$T_\mathrm{eff}$}}= 7\,800, 8\,400, 7\,600$K, late secondary star spectral types $\mathrm{M9}$, $\mathrm{M7}$, $\mathrm{M6}$, and are among the closest PCEBs from the sample, ${\mbox{$d_{\mathrm{wd}}$}}= 60, 76, 73$ pc respectively. SDSS1844+4120 is the system with the highest systemic velocity and it has CaII H&K in absorption, i. e. the white dwarf is a DAZ. This photospheric absorption could be originated by some wind from the secondary star that is being accreted by the WD (see e.g. LTT560 [@tappertetal07-1]). Looking at the SDSS images of SDSS1523+4604, aka CSO749 [@sanduleak+pesch89-1], the components of this binary seem to be slightly separated, also noted by [@helleretal09-1], and there seems to be a third red object towards the North-East. We have measured an orbital period of $9.93$ hours, so it could be that SDSS1523+4604 is a triple system formed by a close binary and a wide companion. **SDSS1143+0009**, aka WD1140+004, is an X-ray emitting source [@aguerosetal09-1] from the Rosat All Sky Survey [@vogesetal99-1]. The X-ray to optical flux ratio ($\log\mathrm{f_X/f_g}$) is 0.22, a value too high to be explained by the coronal emission of the $\mathrm{dMe}$ star [see @maccacaroetal88-1; @motchetal98-1 for typical values], and the SDSS spectrum presents some marginal emission (although this can be partially masked by the flux of the white dwarf). The high flux ratio could be explained if some wind from the secondary star is being accreted by the white dwarf. If so, some CaII H&K absorption would then be expected, which is not visible in the spectrum. A small fraction of isolated white dwarfs show hard (higher than 0.5 keV) X-ray emission [@o'dwyeretal03-1; @chuetal04-1; @bilikovaetal10-1], nevertheless the effective temperature of the white dwarf in SDSS1143+0009 ($\sim 17~000$K) is too cold to explain it’s hard X-ray emission. @briggsetal07-1 list other six PCEBs X-ray emitters: BPM 71214, RR Cae, UZ Sex, EG UMa, Feige 24 and WD 1541-381 . We investigated the X-ray fluxes of these sources within the XMM-Newton making use of xcat-DB[^8] [@micheletal04-1]. None of them show an X-ray spectrum as hard as that of SDSS1143+0009. It has a counterpart the Galaxy Evolution Explorer (GALEX) database [@martinetal05-1], with detection only in the FUV flux, $17.97 \pm 0.08$ mag, consistent with the effective temperature of the white dwarf. We conclude that there is no straight-forward explanation of the hard X-ray emission of SDSS1143+0009. **SDSS1352+0910** contains a hot WD, ${\mbox{$T_\mathrm{eff}$}}= 36~200$K and no companion visible in SDSS spectrum. Nevertheless emission in the line was observed in the SDSS spectrum, pointing to a binary nature of the system, and triggering further spectroscopy which revealed it to be a close binary. We investigated infrared surveys to learn more about the secondary star. We found a counterpart in the UKIDSS catalogue, ULAS J135228.14+091039.0. For details on the UKIDSS project, camera, photometric system, calibration, pipeline processing and science archive see [@lawrenceetal07-1; @casalietal07-1; @hewettetal06-1; @hodgkinetal09-1; @hamblyetal08-1]. Magnitudes found in the 7th data release are Y $= 18.24 \pm 0.03$, J $= 18.04 \pm 0.03$, H = $+17.68 \pm 0.04$, K = $17.49 \pm 0.06$ mag, yielding infrared colours J-H = 0.36 and H-K = 0.19 magnitudes. We subtracted the infrared contribution of the white dwarf and, using the empirical luminosity-mass relation from [@delfosseetal00-1] we estimated a mass of 0.175 for the secondary star. Using the spectral type mass radius relation from [@rebassa-mansergasetal07-1] this is consistent with a spectral type of $\mathrm{M6.5}$. **SDSS1434+5335** (SDSSJ143443.24+533521.2) has a background or nearby red object (SDSSJ143443.19+533525.0), classified as a galaxy by the SDSS. The background galaxy is too faint, and our object is sufficiently bright so that integration time was relatively short and a possible contamination of our spectra is negligible. **SDSS1436+5741** and **SDSS2156–0002** present a slight blue excess in the SDSS spectra with respect to single dM stars. Spectroscopic observations carried out for these two systems (see Table \[t:log\]) showed significant radial velocity variation, confirming their binary nature. We explored GALEX with the aim of learning about the primary stars and found that SDSS2156–0002 has GALEXJ215614.5-000237 as counterpart. Only flux in the NUV was detected, with a magnitude of $21.99 \pm 0.34$, consistent with a cold white dwarf. **SDSS1559+0356** has a hot white dwarf, ${\mbox{$T_\mathrm{eff}$}}=48~212$K, and no companion is visible in the SDSS spectrum. It has a strong emission line from which we were able to measure the radial velocities. SDSS1559+0356 was listed as a CV candidate by [@szkodyetal07-1], but there is no obvious sign of mass transfer in the spectrum. With its orbital period of $2.2$ hours this system is located in the CV orbital period gap. If it would be a detached CV one would expect a cooler white dwarf [@townsley+gaensicke09-1]. We investigated infrared surveys to have more information for the secondary star, but no counterpart was found for the object. **SDSS1611+4640** has the shortest period among the PCEBs, with ${\mbox{$P_\mathrm{orb}$}}= 1.97$ hours, and has a cold white dwarf, ${\mbox{$T_\mathrm{eff}$}}\sim 10~300$K and a late secondary star spectral type, $\mathrm{M5}$. **SDSS1731+6233** has a red nearby star, care was taken so as not to contaminate our spectra with any contribution from it. Unbeknownst to us at the start of our spectroscopic campaign, this system had been previously classified a probable binary by [@pourbaixetal05-1] on the basis of radial velocity variations of $\sim 40$ . **SDSS2149–0717** has SDSS photometry but it was not an SDSS spectroscopic target. It was identified as a WDMS by [@southworthetal07-1] during some pilot follow-up studies of WDMS candidates in the SDSS footprint at the UK Schmidt telescope with the 6dF spectrograph. The spectral type of the secondary star is less certain than for SDSS spectra since the 6dF only covers until 7 500Å. **SDSS2208+1221** contains the hottest white dwarf in our sample, ${\mbox{$T_\mathrm{eff}$}}\sim 100~000$K and the earliest spectral type secondary star, a $\mathrm{K7}$ star [@silvestrietal06-1]. **SDSS2243+3122** has a $\mathrm{M5}\pm1$ so that it could be a detached CV (dCV) or hibernating CV, that has relaxed to its equilibrium radius. The SDSS spectrum presents the Balmer lines in emission. Radial velocities from the emission line showed no difference in phasing from the absorption doublet, i.e. both sets of lines must originate from the secondary which argues against a CV nature of the system. [^1]: For the remainder of the paper, we will use *PCEB* as a synonym for a short-period WDMS binary that underwent a common envelope phase. [^2]: http://www.eso.org/sci/software/esomidas [^3]: starlink.jach.hawaii.edu/starlink [^4]: www.warwick.ac.uk/go/trmarsh/software [^5]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under contract with the National Science Foundation, http://iraf.noao.edu [^6]: SDSS0106–0014, 0110+1326, 0303+0054, 0908+0604, 1212–0123, 1423+2409, 1548+4057, 1648+2811, 1724+5620 and 2112+1014 [^7]: SDSS0307+3848, 1300+1908, 1348+1834 and 1625+6400 [^8]: http://xcatdb.u-strasbg.fr/

ArXiv

--- abstract: 'In this paper, we focus on maximum principles of a time-space fractional diffusion equation. Maximum principles for classical solution and weak solution are all obtained by using properties of the time fractional derivative operator and the fractional Laplace operator. We deduce maximum principles for a full fractional diffusion equation, other than time-fractional and spatial-integer order diffusion equations.' address: - 'Department of Mathematics, Xi’an Jiaotong University, Xi’an 710049, China; ' - 'Department of Mathematics, Xi’an Jiaotong University, Xi’an 710049, China;' author: - Junxiong Jia - Kexue Li bibliography: - 'reference.bib' title: 'Maximum principles for a time-space fractional diffusion equation' --- Introduction ============ In this paper, we focus on the following time-space fractional diffusion equation $$\begin{aligned} \label{TimeSpaceEquation1} \left\{\begin{aligned} \partial_{t}^{\alpha}(u(x,t)-u_{0}(x)) + (-\Delta)^{\beta}u(x,t) & = f(x,t) \quad \text{in }\Omega\times[0,\infty), \\ u(x,t) & = 0 \quad\quad\quad\, \text{in }\mathbb{R}^{N}\backslash\Omega, \, t\geq 0, \\ u(x,0) & = u_{0}(x) \quad\,\, \text{in }\Omega, \end{aligned}\right.\end{aligned}$$ where $\Omega \subset\mathbb{R}^{N}$($N \geq 1$) is a bounded domain in $N$-dimensional space, $\alpha, \beta \in (0,1)$ and $\partial_{t}^{\alpha}\cdot$ represents the Riemann-Liouville time-fractional derivative defined as follow $$\begin{aligned} \label{CaputoDefinition1} \partial_{t}^{\alpha}v(t) := \frac{d}{dt}(g_{1-\alpha}*v(\cdot))(t),\end{aligned}$$ with $g_{\gamma}(t) = \frac{t^{\gamma - 1}}{\Gamma(\gamma)}$ and “$*$” represents usual convolution operator. The fractional Laplace operator could be defined as follow $$\begin{aligned} \label{DefinitionFractionalLaplace} (-\Delta)^{\beta}v(x) = c_{N,\beta}\int_{\mathbb{R}^{n}} \frac{v(x) - v(y)}{|x - y|^{N+2\beta}}dy,\end{aligned}$$ with $c_{N,\beta} = \frac{\beta 2^{2\beta}\Gamma(\frac{N+2\beta}{2})}{\pi^{N/2}\Gamma(1-\beta)}$ and $\Gamma(\cdot)$ represents the usual Gamma function. For more properties about fractional Laplace operator, we refer to [@CPA:CPA20153]. There are much research about maximum principles for equation (\[TimeSpaceEquation1\]) when $\beta = 1$ [@Luchko2009218; @al2014maximum], which is a time fractional diffusion equation. In the fractional elliptic partial differential equation field, there are also lots of research about maximum principles e.g. [@grecohopf]. Recently, some maximum principles for the time fractional diffusion equations have been applied to inverse source problems in [@Luchko2013Rundell]. Although maximum principles are important tools, to the best of our knowledge, there are few results about maximum principles for equation (\[TimeSpaceEquation1\]) when $\alpha$, $\beta$ are both non-integers. In this paper, we prove weak maximum principles for classical and weak solutions of full fractional diffusion equation (\[TimeSpaceEquation1\]) which may provide important tools for other research. **Notations**: In the sequel, $W^{k,p}$ denotes the usual Sobolev spaces with derivative $k$ and Lebesgue exponent $p$; $C^{k}$ denotes $k$ times differentiable function spaces. Fundamental Identity of the Time Fractional Derivative ====================================================== In the following proof, we need an important formula which could be found in [@Zacher2008137] that is for a sufficiently smooth function $u$ on $(0,T)$ one has for a.e. $t\in (0,T)$, $$\begin{aligned} \label{FundamentalIdentity} \begin{split} & H'(u(t))\frac{d}{dt}(k*u)(t) = \frac{d}{dt}(k*H(u))(t) + (-H(u(t))+H'(u(t))u(t))k(t) \\ & \quad + \int_{0}^{t}(H(u(t-s))-H(u(t))-H'(u(t))[u(t-s)-u(t)])\left(-\frac{dk(s)}{ds}\right)ds, \end{split}\end{aligned}$$ where $H\in C^{1}(\mathbb{R})$ and $k\in W^{1,1}([0,T])$. Denote $y^{+} = \max\{y,0\}$ and $y^{-}=\max\{-y,0\}$. Now, taking $H(y)=\frac{1}{2}(y^{+})^{2}$, for any function $u \in L^{2}([0,T])$, there will be a direct corollary of the above formula $$\begin{aligned} \label{AppendexTimeFractional1} u(t)^{+}\frac{d}{dt}(k*u)(t) \geq \frac{1}{2}\frac{d}{dt}(k*(u^{+})^{2}), \quad \text{a.e. }t\in(0,T).\end{aligned}$$ Denote $v = -u$ and for $v$, we could also obtain $$\begin{aligned} \label{AppendexTimeFractional2} v(t)^{+}\frac{d}{dt}(k*v)(t) \geq \frac{1}{2}\frac{d}{dt}(k*(v^{+})^{2}), \quad \text{a.e. }t\in(0,T).\end{aligned}$$ Now replacing $u$ back into (\[AppendexTimeFractional2\]), we find that $$\begin{aligned} \label{AppendexTimeFractional3} u(t)^{-}\frac{d}{dt}(k*u)(t) \leq -\frac{1}{2}\frac{d}{dt}(k*(u^{-})^{2}), \quad \text{a.e. }t\in(0,T).\end{aligned}$$ Maximum Principle for Classical Solution ======================================== In this section, firstly, let us introduce a lemma which could easily be obtained by using Theorem 1 in [@Luchko2009218] and formula (1.20) in [@bajlekova2001fractional]. \[CaputoMaximumLemma\] Let a function $f \in W^{1,1}((0,T)) \cap C([0,T])$ attain its maximum (minimum) over the interval $[0,T]$ at the point $\tau = t_{0}$, $t_{0}\in (0,T]$. Then the Riemann-Liouville fractional derivative of the function $f(\cdot) - f(0)$ is non-negative (non-positive) at the point $t_{0}$ for any $\alpha$, $0<\alpha<1$, $$\begin{aligned} \partial_{t}^{\alpha}(f(t_{0})-f(0)) \geq 0, \quad (\partial_{t}^{\alpha}(f(t_{0})-f(0)) \leq 0), \quad 0<\alpha < 1.\end{aligned}$$ \[BoundaryDefinition1\] Define the following concepts regarding the domain of the solution: 1. $Q_{T} := \Omega\times(0,T) \subset \mathbb{R}^{N+1}$. 2. Lateral boundary of $Q_{T}$: $\partial_{L}Q_{T}:=\partial\Omega\times[0,T]$. 3. Parabolic boundary of $Q_{T}$: $\partial_{p}Q_{T} := (\Omega\times\{0\}) \cup \partial_{L}Q_{T}$. \[MaximumPrincipleClassicalSol1\] Let $\Omega \subset \mathbb{R}^{N}$ to be a bounded domain, and let $u(x,t)$ be a function that is $C^{2}$ in $x$ and $C^{1}$ in $t$ for $(x,t)\in \Omega\times(0,T)$, and continuous in both $x$ and $t$ for $(x,t) \in \bar{\Omega}\times[0,T]$; and $u$ is a solution of equation (\[TimeSpaceEquation1\]) with $f \geq 0$ in $\bar{Q}_{T}$, and $u_{0} \geq 0$ in $\Omega$. Then $u \geq 0$ in $\bar{Q}_{T}$. Consider $0 < T' < T$, and $\bar{Q}_{T'}$, and let us argue by contradiction. Assume $u < 0$ somewhere in $\bar{Q}_{T'}$. Because $u \in C(\bar{Q}_{T'})$, and $\bar{Q}_{T'}$ compact, there exist $(x_{0},t_{0}) \in \bar{Q}_{T'}$ such that $u(x_{0},t_{0}) = \min_{\bar{Q}_{T'}}u < 0$. Since $u \geq 0$ in $\partial_{p}\bar{Q}_{T'} \subset \partial_{p}\bar{Q}_{T}$, we have $(x_{0},t_{0}) \notin \partial_{p}\bar{Q}_{T'}$. No matter $(x_{0},t_{0}) \in Q_{T'}$ is a minimum or $(x_{0},t_{0}) \in \Omega\times\{T'\}$ is a minimum, we know that $\partial_{t}^{\alpha}(u(x_{0},t_{0}) - u(x_{0},0)) \leq 0$ from Lemma \[CaputoMaximumLemma\]. Because $u(\cdot,t_{0}) \in C^{2}(\Omega)\cap C(\bar{\Omega})$ and is zero outside the domain and $u$ attains minimum at $(x_{0},t_{0})$, we have $$\begin{aligned} \label{ClassicalMaximumProof1} (-\Delta)^{\beta}u(x_{0},t_{0}) = c_{n,\beta}\int_{\mathbb{R}^{n}} \frac{u(x_{0},t_{0}) - u(x,t_{0})}{|x_{0} - x|^{N+2\beta}}dx \leq 0.\end{aligned}$$ If $(-\Delta)^{\beta}u(x_{0},t_{0}) = 0$, then $u(\cdot,t_{0}) = 0$, which is a contradiction with $u(x_{0},t_{0}) < 0$, therefore $(-\Delta)^{\beta}u(x_{0},t_{0}) < 0$. But, we have $0 \leq f(x,t) = \partial_{t}^{\alpha}(u(x_{0},t_{0}) - u(x_{0},0)) + (-\Delta)^{\beta}u(x_{0},t_{0}) < 0$. It is a contradiction. Therefore, $u\geq 0$ in $Q_{T'}$. Now we obtain $u \geq 0$ in $\bar{Q}_{T'}$ for all $T' < T$. By continuity, $u \geq 0$ in $\bar{Q}_{T}$. \[MaximumPrincipleClassicalSol2\] Let $\Omega \subset \mathbb{R}^{n}$ be a bounded domain, $T > 0$ and let $u$ be a function with the same regularity as in Theorem \[MaximumPrincipleClassicalSol1\] and Dirichlet (zero) exterior conditions. Then we have the following two assertions 1. If $\partial_{t}^{\alpha}(u-u_{0}) + (-\Delta)^{\beta}u \leq 0$ in $\Omega$, $t \in [0,T]$, then $\max_{\bar{Q}_{T}}u = \max_{\partial_{p}Q_{T}}u$. 2. If $\partial_{t}^{\alpha}(u-u_{0}) + (-\Delta)^{\beta}u \geq 0$ in $\Omega$, $t \in [0,T]$, then $\min_{\bar{Q}_{T}}u = \min_{\partial_{p}Q_{T}}u$. We only prove the second result, the first one could be proved similarly. If $u(x,0) \geq 0$, then we use Theorem \[MaximumPrincipleClassicalSol1\] to see $u \geq 0$ in $\bar{Q}_{T}$, and since $\partial_{p}Q_{T}\subset \bar{Q}_{T}$ and $u|_{\partial_{p}Q_{T}} = 0$, $\min_{\bar{Q}_{T}}u = \min_{\partial_{p}Q_{T}}u = 0$. Otherwise, we assume that $u \geq 0$ not hold everywhere in $Q_{T}$, so there exists $(x_{0},t_{0}) \in \bar{Q}_{T}$ such that $\min_{\bar{Q}_{T}}u = u(x_{0},t_{0}) < 0$. By the proof of Theorem \[MaximumPrincipleClassicalSol1\], it is not possible that there exists a negative minimum in $Q_{T}\cup (\Omega\times\{T\})$, therefore, the minimum in $\bar{Q}_{T}$ must be in $\partial_{p}Q_{T}$. Maximum Principle for Weak Supersolution ======================================== For convenience, denote $H_{e}^{s}(\Omega)$ ($s\in \mathbb{R}$) as follow $$\begin{aligned} \label{DefineHExteriorSobolev1} H_{e}^{s}(\Omega) := \left\{ u \in W^{s,2}(\mathbb{R}^{N}) \, : \, u = 0 \text{ in }\mathbb{R}^{N}\backslash\Omega \right\},\end{aligned}$$ and $L_{e}^{p}(\Omega)$ ($1 \leq p \leq \infty$) as $$\begin{aligned} \label{DefineLExteroriLegesgue1} L_{e}^{p}(\Omega) := \left\{ u \in L_{e}^{p}(\mathbb{R}^{N}) \, : \, u = 0 \text{ in }\mathbb{R}^{N}\backslash\Omega \right\}.\end{aligned}$$ Denote $$\begin{aligned} \label{SecondForm1} a(u,v) := \frac{c_{N,\beta}}{2}\int_{\mathbb{R}^{N}}\int_{\mathbb{R}^{N}}\frac{(u(x,t)-u(y,t))(\eta(x,t)-\eta(y,t))}{|x-y|^{N+2\beta}}dxdy.\end{aligned}$$ We say that a function $u$ is a weak supersolution of (\[TimeSpaceEquation1\]) in $Q_{T}$ with $f \in L^{\infty}(Q_{T})$ and $u_{0} \in L_{e}^{2}(\Omega)$, if $u$ belongs to the space $$\begin{aligned} \begin{split} V_{p} := & \Big\{ u \in L^{2p}([0,T];L_{e}^{2}(\Omega)) \cap L^{2}([0,T];H_{e}^{\beta}(\Omega)) \\ & \text{ such that }g_{1-\alpha}*(u-u_{0}) \in C([0,T];L_{e}^{2}(\Omega)), \text{ and }(g_{1-\alpha}*(u-u_{0}))|_{t=0} = 0 \Big\}, \end{split}\end{aligned}$$ and for any nonnegative test function $$\begin{aligned} \label{TestFun1} \eta \in H_{e}^{1,\beta}(Q_{T}) := W^{1,2}([0,T];L_{e}^{2}(\Omega)) \cap L^{2}([0,T];H_{e}^{\beta}(\Omega))\end{aligned}$$ with $\eta|_{t=T} = 0$ there holds $$\begin{aligned} \label{WeakFormulation1} \begin{split} \int_{0}^{T}\int_{\Omega}-\eta_{t}\left[ g_{1-\alpha}*(u-u_{0}) \right]dxdt + \int_{0}^{T}a(u,\eta)dt \geq \int_{0}^{T}\int_{\Omega}f\eta dxdt. \end{split}\end{aligned}$$ We could provide an equivalent weak formulation of (\[TimeSpaceEquation1\]) where kernel $g_{1-\alpha}$ is replaced by a more regular kernel $g_{1-\alpha,m}$($m\in\mathbb{N}$). For the detailed definition of $g_{1-\alpha,m}$, we refer to Section 2 in [@Zacher2008137]. We could also introduce a function $h_{m}$ which satisfy $g_{1-\alpha,m} = g_{1-\alpha}*h_{m}$ with “$*$” represents the convolution operator. For concisely, we only provide some important properties of functions $g_{1-\alpha,m}$ and $h_{m}$ as follows $$\begin{aligned} \label{propertiesApp} \begin{split} & g_{1-\alpha,m} \in W^{1,1}([0,T]), \quad g_{1-\alpha,m} \rightarrow g_{1-\alpha} \text{ in }L^{1}([0,T]) \text{ as }m\rightarrow \infty, \\ & g_{1-\alpha,m} \text{ and }h_{m} \text{ are all nonnegative functions for every }m \in \mathbb{N}, \\ & \text{If }f\in L^{p}([0,T],X), 1\leq p<\infty, \text{ there holds }h_{m}*f \rightarrow f \text{ in }L^{p}([0,T],X), \end{split}\end{aligned}$$ where $X$ represents a Banach space. Now we could show another definition of weak solution which is equivalent to equation (\[WeakFormulation1\]). \[WeakEquivalent1\] Let $u \in V_{p}$ is a weak supersolution of equation (\[TimeSpaceEquation1\]) if and only if for any nonnegative function $\psi \in H_{e}^{\beta}(\Omega)$ one has $$\begin{aligned} \label{WeakFormulation2} \begin{split} &\int_{\Omega}\psi \partial_{t}\left[ g_{1-\alpha,m}*(u-u_{0}) \right]dx + a(h_{m}*u,\psi) \\ &\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\,\,\,\, \geq \int_{\Omega}(h_{m}*f)\psi dx \, \text{ a.e.}\,\,t\in (0,T), \, m\in\mathbb{N}. \end{split}\end{aligned}$$ The ‘if’ part is readily seen as follows. Given an arbitrary nonnegative $\eta \in H_{e}^{1,\beta}(Q_{T})$ satisfying $\eta|_{t=T} = 0$, we take in (\[WeakFormulation2\]) $\psi(x) = \eta(t,x)$ for any fixed $t\in (0,T)$, integrate from $t = 0$ to $t = T$, and integrate by parts with respect to the time variable. Then by using the approximating properties of the kernels $h_{m}$, we obtain (\[WeakFormulation1\]). To show the ‘only-if’ part, we choose the test function $$\begin{aligned} \label{TestFun2} \eta(x,t) = \int_{t}^{T}h_{m}(\sigma - t)\varphi(\sigma,x)d\sigma = \int_{0}^{T-t}h_{m}(\sigma)\varphi(\sigma+t,x)d\sigma,\end{aligned}$$ with arbitrary $m\in\mathbb{N}$ and nonnegative $\varphi\in H_{e}^{1,\beta}(Q_{T})$ satisfying $\varphi|_{t=T} = 0$; $\eta$ is a nonnegative since $\varphi$ and $h_{m}$ are both nonnegative functions. For the first term in (\[WeakFormulation1\]), it can be transformed to $$\begin{aligned} \label{EquivalentTran1} \int_{0}^{T}\int_{\Omega}-\varphi_{t}\left[ g_{1-\alpha,m}*(u-u_{0}) \right] dxdt,\end{aligned}$$ where we used $g_{1-\alpha,m} = g_{1-\alpha}*h_{m}$ and the Fubini’s theorem. For the term $\int_{0}^{T}a(u,\eta)dt$, we have $$\begin{aligned} \begin{split} &\int_{0}^{T}a(u,\eta)dt \\ = & \frac{c_{N,\beta}}{2}\int_{0}^{T}\int_{\mathbb{R}^{N}\times\mathbb{R}^{N}}\int_{t}^{T} h_{m}(\sigma-t)\frac{(u(x,t)-u(y,t))(\varphi(x,\sigma)-\varphi(y,\sigma))}{|x-y|^{N+2\beta}}d\sigma dx dy dt \\ = & \frac{c_{N,\beta}}{2}\int_{0}^{T}\int_{\mathbb{R}^{N}}\int_{\mathbb{R}^{N}} \frac{((h_{m}*u)(x,t)-(h_{m}*u)(y,t))(\varphi(x,t)-\varphi(y,t))}{|x-y|^{N+2\beta}}dxdydt \\ = & \int_{0}^{T}a(h_{m}*u,\varphi)dt. \end{split}\end{aligned}$$ Observe that $g_{1-\alpha,m}*(u-u_{0}) \in {_{0}}W{^{1,2}}([0,T];L_{e}^{2}(\Omega))$ where $0$ means vanishing at $t=0$. Therefore, combining (\[EquivalentTran1\]) and the above equation, then integrating by parts and using $\varphi|_{t = T} = 0$ yields $$\begin{aligned} \label{WeakFormWithT} \int_{0}^{T}\int_{\Omega}\varphi \partial_{t}\left[ g_{1-\alpha,m}*(u-u_{0}) \right]dx + a(h_{m}*u,\varphi)dt \geq \int_{0}^{T}\int_{\Omega}(h_{m}*f)\varphi dxdt,\end{aligned}$$ for all $m \in \mathbb{N}$ and $\varphi \in H_{e}^{1,\beta}(Q_{T})$ with $\varphi|_{t=T} = 0$. By means of a simple approximation argument, we obtain that (\[WeakFormWithT\]) holds true for any $\varphi$ of the form $\varphi(x,t) = \chi_{(t_{1},t_{2})}\psi(x)$ where $\chi_{(t_{1},t_{2})}$ denotes the characteristic function of the time-interval $(t_{1},t_{2})$, $0<t_{1}<t_{2}<T$ and $\psi \in H_{e}^{\beta}(\Omega)$ is nonnegative. Appealing to the Lebesgue’s differentiation theorem [@grafakos2014classical], the proof is complete. Now, we prove the maximum principle for the weak supersolution of (\[TimeSpaceEquation1\]). \[WeakMaximumParabolic1\] Let $\Omega \subset \mathbb{R}^{N}$ be a bounded domain, $T > 0$, and $u$ a weak supersolution of problem (\[TimeSpaceEquation1\]) with $u_{0} \geq 0$ a.e. in $\Omega$ and $f \geq 0$ a.e. in $\Omega \times [0,T]$. Then $u \geq 0$ a.e. in $\mathbb{R}^{N}\times[0,T]$. We proceed by a contradiction argument. Taking $\varphi$ in (\[WeakFormWithT\]) to be $u^{-}$, the negative part of $u$. Suppose $u^{-}$ is nonzero in a set of positive measure. We know that $$\begin{aligned} \label{WeakMaxProof1} \begin{split} \int_{0}^{T}\int_{\Omega}u^{-}\partial_{t}\left[ k_{m}*(u-u_{0}) \right]dx + & a(h_{m}*u,u^{-})\, dt \\ & \geq \int_{0}^{T}\int_{\Omega}(h_{m}*f)u^{-} dxdt. \end{split}\end{aligned}$$ Let us first analyze the second term on the left hand side of (\[WeakMaxProof1\]). Because $h_{m}*u \rightarrow u$ in $L^{2}([0,T];L_{e}^{2}(\Omega))$ as $m\rightarrow \infty$, we could deduce that $\int_{0}^{T}a(h_{m}*u,u^{-})\,dt \rightarrow \int_{0}^{T}a(u,u^{-})\,dt \quad \text{as }m\rightarrow \infty.$ From $$\begin{aligned} &\quad\quad\quad\quad \int_{0}^{T}a(u,u^{-})dt = \int_{0}^{T}a(u^{+},u^{-})dt - \int_{0}^{T}a(u^{-},u^{-})dt, \\ &\int_{0}^{T}a(u^{-},u^{-})dt = \frac{c_{N,\beta}}{2}\int_{0}^{T}\int_{\mathbb{R}^{N}}\int_{\mathbb{R}^{N}} \frac{(u^{-}(x,t) - u^{-}(y,t))^{2}}{|x-y|^{N+2\beta}}dxdydt > 0,\end{aligned}$$ we find that $$\begin{aligned} \int_{0}^{T}a(u,u^{-})dt < \int_{0}^{T}a(u^{+},u^{-})dt.\end{aligned}$$ Noticing that $(u^{+}(x,t) - u^{+}(y,t))(u^{-}(x,t) - u^{-}(y,t)) \leq 0$, we obtain $$\begin{aligned} \label{WeakMaxProof2} \int_{0}^{T}a(u,u^{-})dt < \int_{0}^{T}a(u^{+},u^{-})dt \leq 0.\end{aligned}$$ Hence, there exists a large positive number $M > 0$ such that if $m \geq M$, we have $$\begin{aligned} \label{WeakMaxProof3} \int_{0}^{T}a(h_{m}*u,u^{-})dt < 0.\end{aligned}$$ For the first term on the left hand side of (\[WeakMaxProof1\]), we have $$\begin{aligned} \int_{0}^{T}\int_{\Omega}u^{-}\partial_{t} & \left[ g_{1-\alpha,m}*(u-u_{0}) \right]dxdt \\ & = \int_{0}^{T}\int_{\Omega}u^{-}\partial_{t}\left[ g_{1-\alpha,m}*u\right]dxdt - \int_{0}^{T}\int_{\Omega}u^{-}g_{1-\alpha,m}u_{0}dxdt.\end{aligned}$$ Noticing that the second term on the righthand side is bigger than or equal to zero, we infer that $$\begin{aligned} \label{WeakMaxProof4} \int_{0}^{T}\int_{\Omega}u^{-}\partial_{t}\left[ g_{1-\alpha,m}*(u-u_{0}) \right]dxdt \leq \int_{0}^{T}\int_{\Omega}u^{-}\partial_{t}\left[ g_{1-\alpha,m}*u\right]dxdt.\end{aligned}$$ Using formula (\[AppendexTimeFractional3\]), we obtain $$\begin{aligned} \label{WeakMaxProof5} \int_{0}^{T}\int_{\Omega}u^{-}\partial_{t}\left[ g_{1-\alpha,m}*u\right]dxdt \leq -\frac{1}{2}\int_{\Omega}(g_{1-\alpha,m}*(u^{-})^{2})(x,T)dx \leq 0.\end{aligned}$$ From (\[WeakMaxProof4\]) and (\[WeakMaxProof5\]), we conclude that $$\begin{aligned} \label{WeakMaxProof6} \int_{0}^{T}\int_{\Omega}u^{-}\partial_{t}\left[ g_{1-\alpha,m}*(u-u_{0}) \right]dxdt \leq 0 \quad \text{for }m\in \mathbb{N}.\end{aligned}$$ Considering (\[WeakMaxProof3\]) and (\[WeakMaxProof6\]), for sufficiently large $m$, we deduce that $$\begin{aligned} \label{WeakMaxProof7} \begin{split} \int_{0}^{T}\int_{\Omega}u^{-}\partial_{t}\left[ g_{1-\alpha,m}*(u-u_{0}) \right]dx + a(h_{m}*u,u^{-})\, dt < 0 \end{split}\end{aligned}$$ Since $f \geq 0$ a.e. on $Q_{T}$, $u^{-} \geq 0$ a.e. on $Q_{T}$ and $g_{1-\alpha,m} \geq 0$ on $(0,T)$, we obtain $$\begin{aligned} \int_{0}^{T}\int_{\Omega}(h_{m}*f)u^{-} dxdt \geq 0,\end{aligned}$$ which contradicts to (\[WeakMaxProof1\]) and (\[WeakMaxProof7\]). Therefore, $u \geq 0$ a.e. in $\mathbb{R}^{N}\times[0,T]$. Acknowledgements ================ This work was partially supported by the National Natural Science Foundation of China under grant no. 11501439 and the postdoctoral science foundation project of China under grant no. 2015M580826.

ArXiv

--- abstract: 'Recently, it has been observed that a quantum field theory need not be Hermitian to have a real, positive spectrum. What seems to be required is symmetry under combined parity and time-reversal transformations. This idea is extended to massless electrodynamics, in which the photon couples to the axial-vector current with an imaginary coupling constant. The eigenvalue condition necessary for the finiteness of the theory can now be solved; the value for the charge appears to be stable order-by-order. Similarly, the semiclassical Casimir model for the fine-structure constant yields a positive value.' address: - 'Department of Physics, Washington University, St. Louis, MO 63130, USA' - 'Department of Physics and Astronomy, University of Oklahoma, Norman, OK 73019, USA' author: - 'Carl M. Bender[^1]' - 'Kimball A. Milton[^2]' title: Massless quantum electrodynamics with a critical point --- Recently, there have been investigations of quantum theories whose Hamiltonians are non-Hermitian. It has been found that the energy spectra are real and positive when these theories respect ${\cal PT}$ invariance, where ${\cal P}$ and ${\cal T}$ represent parity and time reversal. A class of quantum-mechanical theories having this property is defined by the Hamiltonian [@QM] $$\begin{aligned} H=p^2-(ix)^N\quad(N~{\rm real}). \label{e1}\end{aligned}$$ For all $N\geq2$ the spectrum of $H$ is discrete, real, and positive [@BBM]. Note that this theory does not respect parity invariance; thus for all $N$ (including $N=4$) the expectation value of $x$ is nonvanishing [@QM]. This surprising result is a consequence of the boundary conditions [@ROT]. Quantum field theories having this property have also been studied. A generalization of Eq. (\[e1\]) to scalar quantum field theory is described by the Lagrangian $$\begin{aligned} {\cal L}={1\over2}(\partial\phi)^2+{1\over2}m^2\phi^2-g(i\phi)^N\quad(N\geq2). \label{e2}\end{aligned}$$ This theory is not symmetric under $\cal P$ or $\cal T$ separately, but it is invariant under the product ${\cal PT}$. The Hamiltonian for this theory is not Hermitian and thus the theory is not unitary in the conventional sense. However, there is strong evidence that the energy spectrum is real and bounded below [@PARITY]. One can understand positivity heuristically in the context of the weak-coupling expansion for the case $N=3$. The Lagrangian for this theory is $$\begin{aligned} {\cal L}={1\over2}(\partial\phi)^2+{1\over2}m^2\phi^2+gi\phi^3. \label{e3}\end{aligned}$$ In a conventional $g\phi^3$ theory the weak-coupling expansion is real, and (apart from a possible overall factor of $g$) the Green’s functions are formal power series in $g^2$. These series are not Borel summable because they do not alternate in sign. Nonsummability reflects the fact that the spectrum of the underlying theory is not bounded below. However, when we replace $g$ by $ig$, the perturbation series remains real but now alternates in sign. The perturbation series is now summable and this suggests that the underlying theory has a real positive spectrum. We emphasize that replacing $g$ by $ig$ in a $\phi^3$ field theory or $g$ by $-g$ in a $\phi^4$ field theory gives a non-Hermitian Hamiltonian. However, the ${\cal PT}$ invariance of the resulting theory is crucial and appears to guarantee that the energy spectrum is positive. The purpose of this note is extend these notions to quantum electrodynamics [@REST]. In particular, we wish to discuss the case of massless quantum electrodynamics and to re-examine the program of Johnson, Baker, and Willey [@JBW]. In brief, the objective of their program is to find a critical value of the coupling constant $e$ in the Lagrangian describing massless quantum electrodynamics $$\begin{aligned} {\cal L}=-{1\over4}F^{\mu\nu}F_{\mu\nu}-{\bar\psi}\gamma^\mu{1\over i} \partial_\mu\psi+e{\bar\psi}\gamma^\mu A_\mu\psi. \label{e4}\end{aligned}$$ The coupling constant $e$ is determined by the condition that the theory be entirely finite. The mass shift in this theory is finite because the unrenormalized masses are zero. Thus, the only possible infinite quantities are associated with the three renormalization constants $Z_1$, $Z_2$, and $Z_3$. However, to any order in powers of $e$, it is possible to find a gauge in which $Z_1=Z_2$ is finite. Thus, the only remaining divergent quantity is $1/Z_3$. Demanding that this be finite translates into an eigenvalue condition on the fine structure constant $$\begin{aligned} \alpha={e^2\over4\pi}. \label{e5}\end{aligned}$$ This eigenvalue condition takes the form $$\begin{aligned} F_1(\alpha)=0. \label{e6}\end{aligned}$$ The function $F_1(\alpha)$ has been calculated to three loops (four terms) in weak-coupling perturbation theory: $$\begin{aligned} F_1(\alpha)={4\over3}\left({\alpha\over4\pi}\right)+4\left({\alpha\over4\pi} \right)^2-2\left({\alpha\over4\pi}\right)^3-46\left({\alpha\over4\pi}\right)^4 +\cdots. \label{e7}\end{aligned}$$ The first two terms in this series were calculated by Jost and Luttinger [@JL]. Unfortunately, with just two terms the only nontrivial solution of Eq. (\[e6\]) is negative, which gives an unphysical imaginary value for $e$. In a dramatic development, Rosner [@ROSNER] calculated the third term in the series. The negative sign of his result is significant because now there is a positive root to the cubic polynomial equation obtained by truncating $F_1(\alpha)$ after three terms: $$\begin{aligned} \alpha=13.872. \label{e8}\end{aligned}$$ Rosner’s two-loop result is surprising because it is rational. His work suggests the conjecture that all of the coefficients in the expansion of $F_1(\alpha)$ might be rational, possibly reflecting a deep symmetry of the underlying massless theory [@BKZ]. This conjecture has recently gained support through the stunning calculation of the three-loop coefficient by Gorishny and coworkers [@GKLS; @BK; @REVIEW]. The fourth-degree equation gives one positive nontrivial root for $\alpha$: $$\begin{aligned} \alpha=3.969. \label{e9}\end{aligned}$$ This value differs from the result in Eq. (\[e8\]) by a factor of $3.5$, which suggests that this nontrivial root is unstable. One might wonder if a stable positive root can be found by first converting the expansion of $F_1(\alpha)$ to Padé form. The $(1,1)$ Padé of the Rosner result gives no positive root at all. The $(1,2)$ Padé of the four-term series gives one positive root, $\alpha=0.814$, and the $(2,1)$ Padé gives $\alpha=0.545$. There seems to be no sensible pattern to these numerical results. The results regarding ${\cal PT}$-symmetric non-Hermitian quantum field theories are intriguing because they suggest that it is possible to formulate a new kind of electrodynamics. Instead of coupling the $A$ field to a vector current, why not couple this field to an axial-vector current? Of course, this coupling breaks parity symmetry. Therefore, we also replace $e$ by $ie$, thereby breaking time-reversal invariance as well! The resulting ${\cal PT}$-symmetric, massless Lagrangian is $$\begin{aligned} {\cal L}=-{1\over4}F^{\mu\nu}F_{\mu\nu}-{1\over2}\psi\gamma^0\gamma^\mu{1\over i}\partial_\mu\psi+e{1\over2}\psi\gamma^0\gamma^5\gamma^\mu A_\mu\psi. \label{e10}\end{aligned}$$ We conjecture on the basis of our experience with scalar theories that the spectrum of this theory is physically acceptable in that it is bounded below. Our conventions in Eq. (\[e10\]) are as follows: $\gamma^0$ is antisymmetric and pure imaginary, $\gamma^0\gamma^\mu$ is symmetric and real, $\gamma^5= \gamma^0\gamma^1\gamma^2\gamma^3$ is antisymmetric and real, and $(\gamma^5)^2= -1$. The fermion field $\psi$ is expected to be complex, as are the operators $x$ and $p$ in Eq. (\[e1\]) and $\phi$ in Eqs. (\[e2\]) and (\[e3\]). Like the conventional electrodynamics described by the Lagrangian (\[e4\]), ${\cal L}$ in (\[e10\]) possesses gauge invariance; ${\cal L}$ is invariant under the replacements $$\begin{aligned} A^\mu\to A^\mu+\partial^\mu\Lambda,\qquad\psi\to e^{-ie\gamma^5\Lambda}\psi. \label{e11}\end{aligned}$$ Note that this gauge transformation on the fermion field is not a phase transformation when $e$ is real; rather it changes the scale of $\psi$. However, the bilinear forms in the fermion field in the Lagrangian and in the energy-momentum tensor are all invariant. Apart from a possible overall factor of $e\gamma^5$ in some of the Green’s functions, the Feynman rules for the Lagrangian (\[e10\]) give precisely the same weak-coupling expansion as in the conventional massless quantum electrodynamics (\[e4\]), except that $\alpha$ is now replaced by $-\alpha$. Thus, in this new and peculiar theory of quantum electrodynamics, the expansion of $F_1(\alpha)$ becomes $$\begin{aligned} F_1(\alpha)=-{4\over3}\left({\alpha\over4\pi}\right)+4\left({\alpha\over4\pi} \right)^2+2\left({\alpha\over4\pi}\right)^3-46\left({\alpha\over4\pi}\right)^4 +\cdots. \label{e12}\end{aligned}$$ Now, we find that there [*is*]{} a nontrivial positive value $\alpha_2$ satisfying the condition (\[e6\]) when only the first two terms are retained: $$\begin{aligned} \alpha_2=4.189. \label{e13}\end{aligned}$$ If the first three terms are retained, the (unique) positive root is $$\begin{aligned} \alpha_3=3.657, \label{e14}\end{aligned}$$ which differs from $\alpha_2$ by 12%. We feel that since $\alpha_2$ is determined in effect from a $(2,0)$ Padé it is more reasonable to convert the three-term series to a $(2,1)$ Padé and then to find the root. The slightly different result is now $$\begin{aligned} \alpha_3=3.590. \label{e15}\end{aligned}$$ The natural continuation of this process is to calculate the $(3,1)$ Padé of the four-term series. The result is $$\begin{aligned} \alpha_4=4.110. \label{e16}\end{aligned}$$ (This is the only Padé that gives a stable positive root.) Note that the sequence of roots $\alpha_2$, $\alpha_3$, $\alpha_4$ is remarkably stable. It would be extremely interesting to calculate the roots of the $(3,2)$, $(4,2)$, $(4,3)$, $\ldots$, Padés. We conclude this note with a related observation. Recall that Casimir proposed a model for determining the charge of the electron. In this model the Coulomb repulsion of a compact charge distribution is balanced by an attractive zero-point energy [@CASIMIR]. Unfortunately, although the Casimir force for parallel plates is attractive, in a landmark paper Boyer showed that it is repulsive for a perfectly conducting spherical shell [@BOYER; @OTHERS], and thus no balance of forces is possible. However, with ${\cal PT}$-symmetric quantum electrodynamics such a balance is achievable. In the absence of radiative corrections, the Casimir or zero-point energy of a perfectly conducting spherical shell of radius $a$ is $$E_{\rm Casimir}={0.09235\over2a}\hbar c. \label{e17}$$ This energy results from fluctuations of the electromagnetic field inside and outside the shell. This result will be unchanged in the new theory if the boundary conditions are unaltered because the energy is independent of the coupling to the fermion. But now, rather than a Coulomb repulsion, we have an attraction because in effect $e\to ie$. If a charge $e$ is uniformly distributed over a spherical shell of radius $a$, that attractive energy is $$E_{\rm Coulomb}=-{1\over8\pi}{e^2\over a}. \label{e18}$$ Thus, stability is achieved if the two energies cancel: $$E_{\rm Casimir}+E_{\rm Coulomb}=0. \label{e19}$$ This implies a real, positive value for the fine structure constant: $$\alpha={e^2\over4\pi\hbar c}=0.09235.$$ This is an order of magnitude larger than the physical value $1\over137$, and 40 times smaller than the value found above for a finite quantum electrodynamics. But what is significant here is that a positive solution for $\alpha$ actually exists. ACKNOWLEDGMENTS {#s1 .unnumbered} =============== We thank A. Kataev and D. Broadhurst for helpful communications and we are grateful to the U.S. Department of Energy for financial support. C. M. Bender and S. Boettcher, submitted to Phys. Rev. Lett. When $N<2$ some eigenvalues are real and others are complex because ${\cal PT}$ symmetry is spontaneously broken. See C. M. Bender, S. Boettcher, and P. N. Meisinger, in preparation. For a discussion of the effect of boundary conditions see C. M. Bender and A. Turbiner, Phys. Lett. A [**173**]{}, 442 (1993). C. M. Bender and K. A. Milton, Phys. Rev. D [**55**]{}, R3255 (1997). This idea has been extended to quasi-exactly solvable potentials in quantum mechanics in C. M. Bender and S. Boettcher, J. Phys. A, in press. It has been extended to supersymmetric quantum field theory in C. M. Bender and K. A. Milton, Phys. Rev. D, in press. K. Johnson, M. Baker, R. Willey, Phys. Rev. Lett. [**11**]{}, 518 (1963) and Phys.  Rev. [**136**]{}, B1111 (1964); K. Johnson, R. Willey, M. Baker, Phys. Rev. [**163**]{}, 1699 (1967). R. Jost and J. M. Luttinger, Helv. Phys. Acta [**23**]{}, 201 (1950). J. Rosner, Phys. Rev. Lett. [**17**]{}, 1190 (1966) and Ann. Phys. (N.Y.) [**44**]{}, 11 (1967). C. M. Bender, R. W. Keener, R. E. Zippel, Phys. Rev. D [**15**]{}, 1572 (1977). S. G. Gorishny, A. L. Kataev, S. A. Larin, and L. R. Surguladze, Phys. Lett. B [**256**]{}, 81 (1991). D. J. Broadhurst and A. L. Kataev, Phys. Lett. B [**315**]{}, 179 (1993). A. L. Kataev, in [*Proceedings of the Workshop on Continuous Advances in QCD*]{}, Minneapolis, MN, ed. M. I. Polikarpov (World Scientific, Singapore, 1996), p. 107 (hep-ph/9607426). H. B. G. Casimir, Physica [**19**]{}, 846 (1956). T. H. Boyer, Phys. Rev. [**174**]{}, 1764 (1968). B. Davies, J. Math. Phys. [**13**]{}, 1324 (1972); R. Balian and B. Duplantier, Ann. Phys. (N.Y.) [**112**]{}, 165 (1978); K. A. Milton, L. L. DeRaad, Jr., and J. Schwinger, Ann. Phys. (N.Y.) [**115**]{}, 388 (1978). [^1]: Electronic address: [email protected] [^2]: Electronic address: [email protected]

ArXiv

--- abstract: 'Abraham Lorentz (AL) formula of Radiation Reaction and its relativistic generalization, Abraham Lorentz Dirac (ALD) formula, are valid only for periodic (accelerated) motion of a charged particle, where the particle returns back to its original state. Thus, they both represent time averaged solutions for radiation reaction force. In this paper, another expression has been derived for radiation reaction following a new approach, starting from Larmor formula, considering instantaneous change (rather than periodic change) in velocity, which is a more realistic situation. Further, it has been also shown that the new expression for Radiation Reaction is free of pathological solutions; which were unpleasant parts of AL as well as ALD equations; and remained unresolved for about 100 years.' author: - | Nikhil D. Hadap\ Bhabha Atomic Research Center\ Trombay, Mumbai- 400085, India title: 'Radiation Reaction of an accelerating Point Charge: A new Approach' --- Introduction: ============= One of the most important results of classical electrodynamics is the fact that accelerated (or decelerated) charge radiates away energy in the form of electromagnetic waves. The formula for total power radiated was first derived by J. J. Larmor in 1897; known as the Larmor formula [@grif]: $$\label{1} P\ =\ \frac{\mu _0q^2a^2}{6\pi c}$$ The Larmor formula is applicable for non-relativistic particles; where $v << c$. Relativistic generalization of Larmor formula leads to: $$\label{2} P\ =\ \frac{\mu _0q^2\gamma ^6}{6\pi c}\left[a^2-\frac 1{c^2}({\vec v}\times {\vec a})^2 \right]$$ Which is Liénard’s result; which was first obtained in 1898 [@grif]. As radiation carries energy and momentum; in order to satisfy energy and momentum conservation, the charged particle must experience a recoil at the time of emission. The radiation must exert an additional force ${\vec F}_{rad}$ back on the charged particle. This force is known as Radiation Reaction Force or simply, Radiation Reaction. In 1902 Max Abraham derived a non- relativistic formula for radiation- reaction by using Larmor formula (\[1\]), assuming cyclic change in particle’s velocity [@grif], by evaluating: $$\label{3} \int _{t1}^{t2} {\vec F}_{rad} \cdot {\vec v}\ dt\ =\ -\frac{\mu _0q^2 \gamma ^6}{6\pi c}\int _{t1}^{t2}a^2\ dt$$ Which led to the well known Abraham- Lorentz (AL) formula: $$\label{4} {\vec F}_{rad}\ =\ \frac{\mu _0q^2}{6\pi c}\frac{d{\vec a}} {dt}$$ In 1916, H. A. Lorentz showed that the AL force is actually the ”Self Force” [@lor] which is exerted by the accelerated (extended) charge on itself, in order to resist change in state of motion. In 1938, P. A. M. Dirac derived another formula for radiation reaction; by making use of advanced as well as retarded fields and evaluating stress-energy- tensor for point particles [@dir]: $$\label{5} F_{{rad}\ \mu }\ =\ \frac{\mu _0q^2}{6\pi mc} \left[ \frac{d^2 p_{\mu }}{d \tau ^2}\ +\frac{p_{\mu }}{m^2 c^2} \left( \frac{d p_{\nu }}{d\tau } \frac{d p^{\nu }}{d\tau } \right) \right]$$ Which is, actually, the relativistic generalization of AL formula [@roh]; and is known as Abraham- Lorentz- Dirac (ALD) formula. The 3- force (time- space part) of \[5\] gives: $$\label{5-a} {\vec F}_{rad}\ =\ \frac{\mu _0q^2}{6\pi c} \left[ \gamma ^3 \frac{d{\vec a}}{dt}\ + \frac{3 \gamma ^5} c ( {\vec a} \cdot {\vec{\beta }} ) {\vec a} \ +\gamma ^5 ( {\vec{\beta }} \cdot \frac{d{\vec a}}{dt} ) {\vec{\beta }} \ + \frac{3 \gamma ^7} c ( {\vec{\beta }} \cdot {\vec a} ) ^2 {\vec{\beta }} \right]$$ The main problem with AL (and also ALD) formula is its run-away solution [@grif] [@jak]: $a\ =\ a_0e^{t/\tau }$ ; where $\tau \ =\ (\mu _0q^2)/(6\pi c)$ ; Which says that a charged particle initially at rest, would spontaneously accelerate without bounds. Further, when equation of motion is solved assuming presence of non-zero external force, the solution [@jak]: $$\label{6} m \frac{d{\vec v}}{dt}\ =\ \frac 1{\tau }\ e^{\frac t{\tau }}\int _t^{\infty} e^{\frac{-t}{\tau}}\ {\vec F}(t')\ {dt'}$$ predicts acausal pre-acceleration; i.e. the particle accelerates in advance before the force would act. The AL force is the result of the most fundamental calculation of the effect of self-generated fields. However, why it gives unpleasant pathological solutions, is still an unsolved mystery in classical electrodynamics. It looks the problem may be due to positive sign at RHS; in both (\[4\]) and (\[5\]). Dirac also pointed out possible problem of sign in his expression (\[5\]), in his 1938 paper [@dir]. Further, The AL force is the average force that an accelerating charged particle feels in the recoil from the emission of radiation, between the two identical states. Thus, it represents an special case, which requires the particle to return to its initial state of motion, at the end of a cycle. However, a charged particle does radiate while in non periodic accelerated motion; i.e. when it accelerates and finally settles at different velocity. The purpose of this paper to present a new formula for radiation reaction, starting from Larmor formula while considering general case of instantaneously accelerated charge. In section- \[derivation\], general expression for radiation- reaction has been derived. In section- \[analysis\], a qualitative picture of radiation reaction has been presented. In section- \[direction\_of\_Frad\], direction of ${\vec F}_{rad}$ has been determined to complete final expression for radiation- reaction. In section- \[eqn\_of\_motion\], equation of motion has been solved, to test validity of new expression for ${\vec F}_{rad}$. Section- \[conclusion\] is the conclusion, followed by acknowledgement and references. Derivation of a new expression for Radiation Reaction Force: {#derivation} ============================================================ Starting with Larmor- Liénard formula (\[2\]) for electromagnetic power radiated by an accelerating charge; and modifying cross product term by applying scalar triple product rule: ${\vec A} \cdot ({\vec B} \times {\vec C})\ ={\vec B} \cdot ({\vec C} \times {\vec A})$, yields: $$\label{7} ({\vec v} \times {\vec a})^2\ =({\vec v} \times {\vec a}) \cdot ({\vec v} \times {\vec a})\ =\ v^2a^2\ -({\vec v} \cdot {\vec a})^2$$ Putting the same in Larmor- Liénard formula \[2\] yields: $$\label{8} P\ =\ \frac{\mu _0 q^2}{6 \pi c}\ \left[ \gamma^4 a^2\ -\frac{\gamma^6}{c^2}\ ({\vec v} \cdot {\vec a})^2 \right]$$ Whenever there is change in charge’s velocity $d{\vec v}$ there would be radiation. Thus, any radiation received from a point charge would indicate that there was a change in its velocity at earlier times. Thus, according to conservation of energy, the radiation received (at present time) must equal the negative work- done by Radiation Reaction Force (at retarded time): $$\label{9} P\ =\ -\int _{v1}^{v2} {\vec F}_{rad} \cdot {\vec{dv}}\ =\ -\int _{t1}^{t2} {\vec F}_{rad} \cdot {\vec a}\ {dt}$$ which implies: $$\label{10} {\vec F}_{rad} \cdot {\vec a}\ =\ -\frac{dP}{dt}\ =\frac{-\mu_0 q^2}{6 \pi c}\ \frac{d}{dt} \left[ \gamma^4 a^2\ -\frac{\gamma ^6}{c^2} ( {\vec v} \cdot {\vec a})^2 \right]$$ Where $t$ is retarded coordinate time. Carrying out differentiation by parts yields: $$\label{10-a} {\vec F}_{rad}\ \cdot {\vec a}\ =\ -\frac{\mu _0 q^2}{3 \pi c}\ \left[ \gamma^4 ( {\vec a} \cdot \frac{d{\vec a}}{dt})\ +\frac{3 \gamma^6}{c^2}( {\vec v} \cdot {\vec a})\ a^2\ +\frac{\gamma^6}{c^2}( {\vec v} \cdot {\vec a})( {\vec v} \cdot \frac{d{\vec a}}{dt} )\ +\frac{3 \gamma^6}{c^4}( {\vec v} \cdot {\vec a} )^3 \right]$$ Or: $$\label{10-b} {\vec F}_{rad}\ \cdot {\vec a}\ =\ -\frac{\mu _0 q^2}{3\pi c}\ \left[ \gamma^4 ( {\vec a} \cdot \frac{d{\vec a}}{dt} )\ +\frac{3 \gamma ^6}{c} ( {\vec{\beta}} \cdot {\vec a} )a^2\ +\gamma ^6 ( {\vec{\beta }} \cdot {\vec a} ) ( {\vec{\beta }} \cdot \frac{d{\vec a}}{dt} )\ +\frac{3 \gamma ^8}{c} ( {\vec{\beta }} \cdot {\vec a})^3 \right]$$ As equation (\[10-b\]) is an scalar expression, it doesn’t represent unique solution for ${\vec F}_{rad}$. However, it does have a general solution: $$\label{11} \begin{split} {\vec F}_{rad}\ =\ -\frac{\mu_0 q^2}{3 \pi c} \left[ \gamma^4 \frac{d{\vec a}}{dt}\ +\frac{3 \gamma^6}{c} x({\vec a} \cdot {\vec{\beta}}) {\vec a}\ +\frac{3 \gamma^6}{c} (1-x)a^2 {\vec{\beta}}\ +\gamma^6 ( {\vec{\beta}} \cdot {\vec{\frac{d{\vec a}}{dt}}} ) {\vec{\beta}} \right. \\ \left. +\frac{3\gamma^8}{c} ( {\vec{\beta}} \cdot {\vec a} )^2 {\vec{\beta }} \right] \end{split}$$ Where, $x$ is a number; and 0 ${\leq}$ x ${\leq}$ 1. Contrary to AL equation (\[4\]) and ALD equation (\[5\]), here in (\[11\]), there is negative sign at RHS. Thus, the expression is expected to be free from runaway solutions. However, (\[11\]) is not a complete expression; because the value of $x$ is yet to be determined; which shall be evaluated later on, in this paper. However, before proceeding further, a qualitative analysis of radiation reaction has been presented in the next section, which would provide valuable information about ${\vec F}_{rad}$. Qualitative Picture of Radiation Reaction: {#analysis} ========================================== Rewriting the obtained general expression of radiation reaction (\[11\]) in vector component form: $$\label{11-a} \begin{split} {\vec F}_{rad}\ =\ -\frac{\mu_0 q^2}{3 \pi c} \left[ \gamma^4 \frac{d{\vec a}}{dt}\ +\frac{3 \gamma^6}{c} x({\vec a} \cdot {\vec{\beta}}) {\vec a}\ + \left\{\ \frac{3 \gamma^6}{c} (1-x)a^2\ +\gamma^6 ( {\vec{\beta}} \cdot {\frac{d{\vec a}}{dt}} ) \right. \right. \\ \left. \left. +\frac{3\gamma^8}{c} ( {\vec{\beta}} \cdot {\vec a} )^2 \right\} {\vec{\beta }}\ \right] \end{split}$$ The first term of (\[11-a\]) represents component of ${\vec F}_{rad}$ that points opposite to the instantaneous direction of jerk (direction of change in acceleration). This term is like Schott term of ALD equation (\[5\]); the only term which survives in particle’s rest frame ($\beta=0$). Effectively, this term appears to resist any change in acceleration. The second term of (\[11-a\]) represents component of ${\vec F}_{rad}$ that points opposite to the instantaneous direction of acceleration. Effectively, this term appears to resist acceleration; i.e. the change in velocity. Remaining terms of (\[11-a\]) comprise Radiation Recoil term, which points opposite to the instantaneous direction of velocity. Magnitude of which depends upon instantaneous values of jerk and acceleration, coupled with instantaneous velocity. The derived expression (\[11-a\]) represents instantaneous values of ${\vec F}_{rad}$. However, ${\vec F}_{rad}$ is still a ”space-averaged” quantity; because it has been derived from Larmor- Liénard formula (\[2\]), which is ”Total Power radiated in All Directions”. Thus, at any instance any ”Radiation Emitted in all Space” is a cause; and radiation reaction is its effect; its ”Mirror Image”. Therefore, valuable information about direction of ${\vec F}_{rad}$ can be inferred from examination of shape of radiation patterns; which must always point opposite to the ”Net Direction of radiation”. Figure \[fig:1\] and Figure \[fig:2\] are two specific cases of radiation profiles from accelerated charge particles: ![(Color Online) Radiation Pattern emitted by a Linearly accelerated charge[]{data-label="fig:1"}](fig1.eps) ![(Color Online) Radiation Pattern emitted by a charge under Circular motion[]{data-label="fig:2"}](fig2.eps) Figure \[fig:1\] represents radiation patterns from a linearly accelerated charge (${\vec a}\parallel{\vec v}$), as in case of Bremsstrrahlung. Whereas, figure \[fig:2\] represents radiation from a charge under circular motion (${\vec a}\perp {\vec v}$ ), as in case of Synchrotron radiation. In both the cases, radiation patterns are symmetrical about velocity vector. Thus, they provide an idea that ${\vec F}_{rad}$ must point in $- \vec v$ direction; and therefore suggest $x = 0$ in equation (\[11\]). However, in general case when ${\vec a}$ and ${\vec v}$ were at certain angle $\alpha$, and $ 0< \alpha< \pi $, shape of radiation patterns Figure \[fig:3\] provide no help in deducing direction of ${\vec F}_{rad}$ directly. ![(Color Online) Snapshots of Radiation Patterns of a charged particle with acceleration vectors pulling its trajectory at various angles.[]{data-label="fig:3"}](fig3.eps){width="12cm"} Therefore, direction of ${\vec F}_{rad}$ needs to be found analytically, which has been presented in the next section. Direction of Radiation Reaction Vector: {#direction_of_Frad} ======================================= Starting with the Poyenting vector for radiation fields, given by [@grif]: $$\label{12} {\vec S}_{{RAD}}\ =\ \left( \frac{ {\vec R} \cdot {\vec U} }{Rc} \right) \frac{1}{\mu_0 c} E_{RAD}^2 {\hat n}\ =\ ( 1-{\hat n} \cdot {\vec{\beta }} ) \frac{1}{\mu_0 c} E_{RAD}^2 {\hat n}$$ Where: ${\hat n}=\ {\vec R}/R$, the unit vector in the direction of observer, at large distance R, ${\vec U}=\ c{\hat n}\ -{\vec v}$, the Retarded relative vector, & ${\vec E}_{RAD}$ is Radiation Electric Field; given by [@grif]: $$\label{13} \begin{split} {\vec E}_{RAD}\ =\ \frac{q}{4 \pi \epsilon_0} \frac{R}{ ({\vec R} \cdot {\vec U})^3 } \left\{ {\vec R} \times ({\vec U} \times {\vec a}) \right\}\ =\ \frac{q}{4 \pi \epsilon_0 R c^2} \left\{ \frac{ ({\hat n} \cdot {\vec a}) }{(1-{\hat n} \cdot {\vec{\beta }})^3} \right. \\ \left. -\frac{{\vec a}}{(1- {\hat n} \cdot {\vec{\beta }})^2} \right\} \end{split}$$ Thus, using (\[13\]) in (\[12\]) the expression for Radiation Poyenting Vector becomes: $$\label{12-a} {\vec S}_{RAD}\ =\ \frac{q^2 {\hat n}}{16 \pi^2 \epsilon_0 R^2 c^3} \left\{ \frac{a^2}{ (1-{\hat n} \cdot {\vec{\beta}})^3 }\ +\frac{2 ({\hat n} \cdot {\vec a})({\vec a} \cdot {\vec{\beta }}) }{(1-{\hat n} \cdot {\vec{\beta }})^4}\ -\frac{(1-\beta^2)({\hat n} \cdot {\vec a})^2}{(1-{\hat n} \cdot {\vec{\beta}})^5} \right\}$$ Using Cartesian coordinate vectors: ${\vec{\beta}}\ =\ \beta{\hat z}$, ${\vec a}\ =\ a(\sin \alpha {\hat x}\ +cos \alpha {\hat z})$, & ${\hat n}\ =\ \sin \theta \cos \phi {\hat x}\ +sin \theta \sin \phi {\hat y}\ +cos \theta {\hat z}$ where, $\alpha$ is the angle between particle’s instantaneous acceleration and instantaneous velocity; the equation (\[12-a\]) becomes: $$\label{12-b} \begin{split} {\vec S}_{RAD}\ =\ \frac{q^2 a^2}{16 \pi^2 \epsilon_0 R^2 c^3} \left[ \frac{1}{(1-\beta \cos \theta )^3}\ +\frac{2 \beta \sin \alpha \cos \alpha \sin \theta \cos \phi }{(1-\beta \cos \theta)^4}\ +\frac{2 \beta \cos^2 \alpha \cos \theta}{(1-\beta \cos \theta )^4} \right. \\ \left. -\frac{(1-\beta^2) \sin^2 \alpha \sin^2 \theta \cos^2 \phi }{(1-\beta \cos \theta)^5}\ -\frac{(1-\beta^2) \cos^2 \alpha \cos^2 \theta}{(1-\beta \cos \theta )^5} \right.\\ \left. -\frac{2 (1-\beta^2) \sin \alpha \cos \alpha \sin \theta \cos \theta \cos \phi}{(1-\beta \cos \theta)^5} \right] (\sin \theta \cos \phi {\hat x}\ +\sin \theta \sin \phi {\hat y}\ +\cos \theta {\hat z}) \end{split}$$ However, $\alpha=0$ cannot remain constant (unless $\alpha=0$ or $\pi/ 2$); because, component of $\vec a$ perpendicular to $\vec v$ would continuously pull particle’s trajectory more and more towards $\vec a$ (\[fig:4\]). Thus $\alpha$ would continuously change in time. ![(Color Online) Particle’s Velocity & Acceleration vectors in space, along its trajectory.[]{data-label="fig:4"}](fig4.eps) Even if, the acceleration (or applied force) were constant in space; in particle’s frame it would seem to be changing in the direction of changing $\alpha$. The direction of change in $\alpha$ would be the same as the the direction of $d{\vec a}/{dt}$ (\[fig:5\]). ![(Color Online) Velocity & Acceleration vectors in particle’s own Rest Frame.[]{data-label="fig:5"}](fig5.eps) This is why $d{\vec a}/{dt}$ term comes into the expression of ${\vec F}_{rad}$; which is a higher order term, and becomes significant when history of particle’s motion (over a certain small period) were also taken into account. However, purpose of this analysis is to know components of ${\vec F}_{rad}$ along $\vec a$ and $\vec v$ (to know value of $x$ in equation (\[11\])). So only ”instantaneous snapshot” of particle’s state of motion is being considered. At any instance radiation reaction acts like ”Mirror Image” to the ”Radiation Emitted in all Space”. Therefore, if negative of all radiation poyenting vectors (\[12-b\]) were algebraically added up together, over all space, the resultant vector ${\vec W}$ would point in the direction of ${\vec F}_{rad}$: $$\label{15} {\vec W}\ =\ -\oint_v {\vec S}_{RAD}\ {da}\ =\ -\int_{\theta=0}^{\pi} \int_{\phi=0}^{2 \pi} {\vec S}_{RAD}\ R^2 \sin \theta\ d\theta d\phi$$ Substituting expression of ${\vec S}_{RAD}$ from (\[12-b\]) and writing only non- vanishing integrals in $\phi$ yields: $$\label{15-a} \begin{split} {\vec W}\ =\ -\frac{q^2 a^2}{16 \pi^2 \epsilon_0 c^3} \left[ {\hat x} \sin \alpha \cos \alpha \left\{ 2 \beta \int_{\theta=0}^{\pi} \int_{\phi=0}^{2 \pi} \frac{\sin^3 \theta \cos^2 \phi \ d\theta d\phi}{(1-\beta \cos \theta )^4} \right. \right. \\ \left. \left. -2(1-\beta^2) \int_{\theta=0}^{\pi}\int_{\phi=0}^{2 \pi} \frac{\sin^3 \theta \cos \theta \cos^2 \phi \ d\theta d\phi }{(1-\beta \cos \theta)^5} \right\} \right. \\ \left. +{\hat z} \left\{ \int_{\theta=0}^{\pi} \int_{\phi=0}^{2 \pi} \frac{\sin \theta \cos \theta \ d\theta d\phi }{(1-\beta \cos \theta )^3}\ +2\beta \cos^2 \alpha \int_{\theta=0}^{\pi} \int_{\phi=0}^{2 \pi} \frac{\sin \theta \cos^2 \theta \ d\theta d\phi}{(1-\beta \cos \theta)^4} \right. \right. \\ \left. \left. -(1-\beta^2) \sin^2 \alpha \int_{\theta=0}^{\pi} \int_{\phi=0}^{2 \pi} \frac{\sin^3 \theta \cos \theta \cos^2 \phi \ d\theta d\phi}{(1-\beta \cos \theta)^5} \right. \right. \\ \left. \left. -(1-\beta^2) \cos^2 \alpha \int_{\theta=0}^{\pi} \int_{\phi=0}^{2 \pi} \frac{\sin \theta \cos^3 \theta \ d\theta d\phi}{(1-\beta \cos \theta)^5} \right\} \right] \end{split}$$ Substituting suitable $\phi$ integral values yields: $$\label{15-b} \begin{split} {\vec W}\ =\ \frac{-q^2 a^2}{16 \pi^2 \epsilon_0 c^3} \left[ {\hat x} \sin \alpha \cos \alpha \left\{ 2 \pi \beta \int_{0}^{\pi} \frac{\sin^3 \theta \ d\theta}{(1-\beta \cos \theta )^4}\ -2 \pi (1-\beta ^2) \int_{0}^{\pi} \frac{\sin^3 \theta \cos \theta \ d\theta }{(1-\beta \cos \theta)^5} \right\} \right. \\ \left. +{\hat z} \left\{ 2 \pi \int_{0}^{\pi} \frac{\sin \theta \cos \theta \ d\theta}{(1-\beta \cos \theta )^3}\ +4 \pi \beta \cos^2 \alpha \int_{0}^{\pi} \frac{\sin \theta \cos^2 \theta \ d\theta}{(1-\beta \cos \theta)^4} \right. \right. \\ \left. \left. -\pi (1-\beta ^2) \sin^2 \alpha \int_{0}^{\pi} \frac{\sin^3 \theta \cos \theta \ d\theta}{(1-\beta \cos \theta)^5}\ -2 \pi (1-\beta^2) \cos^2 \alpha \int _{0}^{\pi} \frac{\sin \theta \cos^3 \theta \ d\theta}{(1-\beta \cos \theta)^5} \right\} \right] \end{split}$$ Rearranging the $\sin \theta$ and $\cos \theta$ terms; (\[15-b\]) simplifies to: $$\label{15-c} \begin{split} {\vec W}\ =\ -\frac{q^2 a^2}{16 \pi \epsilon_0 c^3} \left[ 2 {\hat x} \sin \alpha \cos \alpha \left\{ \beta \int_{0}^{\pi} \frac{\sin \theta \ d\theta}{(1-\beta \cos \theta)^4}\ -\beta \int_{0}^{\pi} \frac{\sin \theta \cos^2 \theta \ d\theta}{(1-\beta \cos \theta)^4} \right. \right. \\ \left. \left. -(1-\beta^2) \int_{0}^{\pi} \frac{\sin \theta \cos \theta \ d\theta}{(1-\beta \cos \theta)^5}\ +(1-\beta^2) \int_{0}^{\pi} \frac{\sin \theta \cos^3 \theta \ d\theta }{(1-\beta \cos \theta)^5} \right\} \right. \\ \left. +{\hat z} \left\{ 2 \int_{0}^{\pi} \frac{\sin \theta \cos \theta \ d\theta}{(1-\beta \cos \theta)^3}\ +4\beta \cos ^2\alpha \int_{0}^{\pi} \frac{\sin \theta \cos^2 \theta \ d\theta}{(1-\beta \cos \theta)^4} \right. \right. \\ \left. \left. -(1-\beta^2) \sin^2 \alpha \int_{0}^{\pi} \frac{\sin \theta \cos \theta \ d\theta}{(1-\beta \cos \theta)^5}\ +(1-\beta^2) \sin^2 \alpha \int_{0}^{\pi} \frac{\sin \theta \cos^3 \theta \ d\theta}{(1-\beta \cos \theta)^5} \right. \right. \\ \left. \left. -2(1-\beta^2) \cos^2 \alpha \int_{0}^{\pi} \frac{\sin \theta \cos^3 \theta \ d\theta}{(1-\beta \cos \theta)^5} \right\} \right] \end{split}$$ Solving the integrals by substitution $\beta \cos \theta= t$; $\sin \theta d\theta= -dt/ \beta$ and changing the limits ($\theta= 0$ to $\pi$) to ($t= \beta$ to $-\beta$ ) yields: $$\label{15-d} \begin{split} {\vec W}\ =\ \frac{q^2 a^2}{16 \pi \epsilon_0 c^3} \left[ 2 {\hat x} \sin \alpha \cos \alpha \left\{ \int_{\beta }^{-\beta } \frac{dt}{(1-t)^4}\ -\frac{1}{\beta^2} \int_{\beta }^{-\beta} \frac{t^2\ dt}{(1-t)^4}\ -\frac{(1-\beta ^2)}{\beta ^2} \int_{\beta}^{-\beta} \frac{t\ dt}{(1-t)^5} \right. \right. \\ \left. \left. +\frac{(1-\beta^2)}{\beta^4} \int_{\beta}^{-\beta} \frac{t^3\ dt}{(1-t)^5} \right\}\ + {\hat z} \left\{ \frac{2}{\beta^2} \int_{\beta}^{-\beta} \frac{t\ dt}{(1-t)^3}\ +\frac{4}{\beta^2} \cos^2 \alpha \int_{\beta}^{-\beta} \frac{t^2\ dt}{(1-t)^4} \right. \right. \\ \left. \left. -\frac{(1-\beta^2)}{\beta^2} \sin^2 \alpha \int_{\beta}^{-\beta} \frac{t\ dt}{(1-t)^5}\ +\frac{(1-\beta^2)}{\beta^4} \sin^2 \alpha \int_{\beta}^{-\beta} \frac{t^3\ dt}{(1-t)^5} \right. \right. \\ \left. \left. -2 \frac{(1-\beta^2)}{\beta^4} \cos^2 \alpha \int_{\beta}^{-\beta} \frac{t^3\ dt}{(1-t)^5} \right\} \right] \end{split}$$ The integrals can be solved easily; either by parts or by partial fraction. Following are the solutions: $$\label{16} \begin{split} \int_{\beta}^{-\beta} \frac{t\ dt}{(1-t)^3}\ =\ -\frac{2 \beta^3}{(1-\beta^2)^2};\\ \int_{\beta}^{-\beta} \frac{dt}{(1-t)^4}\ =\ -\frac{2 \beta}{3} \frac{(3+ \beta^2)}{(1-\beta^2)^3};\\ \int_{\beta}^{-\beta} \frac{t^2\ dt}{(1-t)^4}\ =\ -\frac{2\beta^3}{3} \frac{(1+3\beta^2)}{(1-\beta ^2)^3};\\ \int_{\beta}^{-\beta} \frac{t\ dt}{(1-t)^5}\ =\ -\frac{2 \beta^3}{3} \frac{(5+\beta^2)}{(1-\beta^2)^4}\\ \int_{\beta}^{-\beta} \frac{t^3\ dt}{(1-t)^5}\ =\ -2 \beta^5 \frac{(1+\beta ^2)}{(1-\beta ^2)^4} \end{split}$$ Putting the values of various integrals (\[16\]) into (\[15-d\]) yields: $$\label{15-e} \begin{split} {\vec W}\ =\ \frac{q^2 a^2 \beta}{16 \pi \epsilon_0 c^3\ (1-\beta^2)^3} \left[ 2 {\hat x} \sin \alpha \cos \alpha \left\{ \frac{-2}{3}(3+\beta^2)\ +\frac{2}{3}(1+3\beta^2)\ +\frac{2}{3}(5+\beta^2) \right. \right. \\ \left. \left. -2(1+\beta^2) \right\}\ +{\hat z} \left\{ -4 (1-\beta ^2)\ -\frac{8}{3} (1+3\beta^2) \cos^2 \alpha \right. \right. \\ \left. \left. +\frac{2}{3} (5+\beta^2) \sin^2 \alpha \ -2 (1+\beta^2) \sin^2 \alpha \ +4 (1+\beta^2) \cos^2 \alpha \right\} \right] \end{split}$$ Which gets simplified to: $$\label{15-f} {\vec W}\ =\ -\frac{q^2 a^2 \beta \gamma^6}{6 \pi \epsilon_0 c^3}\ {\hat z} \cos^2 \alpha$$ which points in $-{\hat z}$ direction, i.e. opposite to the direction of instantaneous velocity (like friction force), for any value of $\alpha$. However, if ${\vec F}_{rad}$ indeed points in $-{\vec v}$ direction, its expression must not contain any term going with ${\vec a}$. Thus, the correct expression for ${\vec F}_{rad}$ is equation (\[11-a\]) with $x= 0$. Rewriting the final expression for Radiation Reaction vector: $$\label{11-d} {\vec F}_{rad}\ =\ -\frac{\mu_0 q^2}{3 \pi c} \left[ \gamma^4 \frac{\vec da}{dt}\ +\left\{ \frac{3 \gamma^6}{c} a^2\ +\gamma^6 (\vec{\beta} \cdot \frac{\vec da}{dt})\ +\frac{3 \gamma^8}{c} ({\vec{\beta}} \cdot {\vec a})^2 \right\} {\vec{\beta }} \right]$$ Further, equation (\[11-d\]) can be expressed in terms of relativistic mechanical force (or rate of change of momentum) which is experienced by the charged particle: $$\label{17} {\vec F}_{MECH}\ =\ \frac{\vec dp}{dt}\ =\ m \frac{d(\gamma {\vec v})}{dt}\ =\ \gamma m {\vec a}\ +\frac{\gamma^3 m}{c^2}({\vec v} \cdot {\vec a}) {\vec v}$$ Differentiating (\[17\]), once again w.r.t. (retarded) time, yields: $$\label{18} \begin{split} \frac{{\vec dF}_{MECH}}{dt}\ =\ \frac{d^2 {\vec p}}{{dt}^2}\ =\ m \gamma \frac{\vec da}{dt}\ +m \frac{d \gamma}{dt} {\vec a}\ +\frac{m \gamma^3}{c^2}({\vec v} \cdot {\vec a}) {\vec a}\ +\frac{m \gamma^3}{c^2}({\vec v} \cdot \frac{\vec da}{dt}) {\vec v} \\ +\frac{m \gamma^3}{c^2} a^2 {\vec v}\ +\frac{3 m \gamma^2}{c^2}({\vec v} \cdot {\vec a}) \frac{d \gamma}{dt} {\vec v} \end{split}$$ Meanwhile: $$\label{19} \frac{d\gamma}{dt}\ =\ \frac{\gamma^3}{c^2}({\vec v} \cdot {\vec a})$$ Thus, it yields from equation (\[18\]): $$\label{20} \begin{split} \frac{1}{m} \frac{{\vec dF}_{MECH}}{dt}\ =\ \frac{1}{m} \frac{d^2 {\vec p}}{{dt}^2}\ =\ \gamma \frac{\vec da}{dt}\ +\frac{2 \gamma^3}{c}({\vec{\beta}} \cdot {\vec a}) {\vec a}\ +\left\{ \gamma^3 \frac{a^2}{c}\ +\frac{3 \gamma^5}{c} ({\vec{\beta}} \cdot {\vec a})^2 \right. \\ \left. +\gamma^3 \left( {\vec{\beta}} \cdot \frac{\vec da}{dt} \right) {\vec{\beta }} \right\} {\vec{\beta }} \end{split}$$ Or: $$\label{20-a} \begin{split} \frac{1}{m} \frac{{\vec dF}_{MECH}}{dt}\ =\ \frac{1}{m} \frac{d^2{\vec p}}{{dt}^2}\ =\ \gamma \frac{\vec da}{dt}\ +\left\{\frac{3 \gamma ^3 a^2}{c}\ +\frac{3 \gamma^5}{c} ({\vec{\beta}} \cdot {\vec a})^2\ +\gamma^3 \left( {\vec{\beta}} \cdot \frac{\vec da}{dt} \right) {\vec{\beta}} \right\} {\vec{\beta}} \\ +\frac{2 \gamma^3}{c} ({\vec{\beta }} \cdot {\vec a}) {\vec a}\ -\frac{2 \gamma^3}{c} a^2 {\vec{\beta }} \end{split}$$ Comparing equation (\[20-a\]) with (\[11-d\]); there arrives another expression for radiation reaction force: $$\label{11-e} {\vec F}_{rad}\ =\ -\frac{\mu_0 q^2}{3 \pi c} \left[ \frac{\gamma^3}{m} \frac{{\vec dF}_{MECH}}{dt}\ -\frac{2 \gamma^6}{c} \{ {\vec a} \times ({\vec a} \times {\vec{\beta}}) \} \right]$$ Or: $$\label{11-f} {\vec F}_{rad}\ =\ -\frac{\mu_0 q^2}{3 \pi c} \left[ \frac{\gamma^3}{m} \frac{d^2 {\vec p}}{{dt}^2}\ -\frac{2 \gamma^6}{c} \{ {\vec a} \times ({\vec a} \times {\vec{\beta}}) \} \right]$$ Equation of Motion: {#eqn_of_motion} =================== Considering non-relativistic case where $\gamma \approx 1$. From Newton’s second law, the equation of motion for a charged particle is: $$\label{21} {\vec F}\ +{\vec F}_{rad}\ =\ m {\vec a}$$ Further, considering motion in straight line (${\vec a} \parallel {\vec{\beta }}$); the second term in equation (\[11-f\]) drops and the equation of radiation reaction reduces to: $$\label{22} {\vec F}_{rad}\ =\ -\frac{\mu_0 q^2 \gamma^3}{3 \pi m c} \frac{d^2 p}{{dt}^2} {\hat x}\ \approx \ -\frac{\mu_0 q^2}{3 \pi c} \frac{da}{dt} {\hat x}$$ Using expression of ${\vec F}_{rad}$ from equation (\[22\]) in equation of motion (\[21\]) yields: $$\label{22-a} F\ {\hat x}\ -\frac{\mu _0 q^2}{3 \pi c} \frac{da}{dt} {\hat x}\ =\ m a\ {\hat x}$$ Or: $$\label{22-b} \frac{F}{m}\ =\ a\ +\frac{\mu_0 q^2}{3 \pi m c} \frac{da}{dt}\ =\ a\ +\tau \frac{da}{dt}$$ Or: $$\label{22-c} \frac{F}{\tau\ m}\ =\ \frac{a}{\tau}\ +\frac{da}{dt}$$ Where; $\tau =\ \frac{\mu_0 q^2}{3 \pi m c}$; the time constant of a moving charged particle. Using integrating factor $e^{t/\tau}$, the solution of equation (\[22-c\]) comes out to be: $$\label{23} a\ =\ \frac{1}{\tau\ m}\ e^{-t/\tau} \int_{T_0}^t F(t')\ e^{t'/\tau}\ {dt'}$$ Which is similar to equation (\[6\]) but with opposite signs in time variable. The constant of integration $T_0$ needs to be found from physical ground. Here, unlike usual mechanical systems, the acceleration doesn’t depend upon instantaneous value of force; but on its weighted time average. Presence of $t'/\tau$ indicate that only small interval of the order of $\tau$ is involved. Integrating by parts in equation (\[23\]) yields: $$\label{23-a} a\ =\ \frac{1}{m} \left\{ F(t)\ -F(T_0)\ e^{(T_0\ -t)/\tau } \right\} \ -\frac{1}{m}\ e^{-t/\tau } \int_{T_0}^t \left\{ \frac{dF}{dt}\ e^{t'/\tau } \right\} {dt}'$$ Integrating further, the second term of (\[23-a\]) by parts yields: $$\label{23-b} \begin{split} a\ =\ \frac{1}{m} \left\{ F(t)\ -F(T_0)\ e^{(T_0\ -t)/\tau } \right\} -\frac{\tau}{m} \left\{ \frac{dF}{dt}(t)\ -\frac{dF}{dt}(T_0)\ e^{(T_0\ -t)/\tau} \right\} \\ +\frac{\tau}{m}\ e^{-t/\tau } \int_{T_0}^t \left\{ \frac{d^2 F}{{dt}^2}\ e^{t'/\tau} \right\} {dt}' \end{split}$$ As $q \to 0$ (for neutral particle); $\tau \to 0$; the acceleration must reduce to $F(t)/ m$, as in case of usual mechanical systems. Thus, (\[23-b\]) demands $T_0 \leq 0$. Therefore, setting time origin $T_0=0$; the successive integration by parts in (\[23-b\]) yields: $$\label{23-c} \begin{split} a\ =\ \frac{1}{m} \left\{ F(t)\ -\tau \frac{dF}{dt}(t)\ +\tau^2 \frac{d^2 F}{{dt}^2}\ -\tau^3 \frac{d^3 F}{{dt}^3}\ +.... \right\} \\ -\frac{e^{-t/\tau }}{m} \left\{ F(0)\ -\tau \frac{dF}{dt}(0)\ +\tau^2 \frac{d^2 F}{{dt}^2}(0)\ -\tau^3 \frac{d^3 F}{{dt}^3}(0)\ +.... \right\} \end{split}$$ Which (using Taylor Series) reduces to: $$\label{23-d} a\ =\ \frac{1}{m} \left[ F(t\ -\tau)\ -F(-\tau)\ e^{-t/\tau } \right]$$ Thus, there is no pre-acceleration seen here. The second term in (\[23-d\]) would vanish as $t \to \infty$; i.e. if motion were observed for a long enough time. The first term simply indicates, “Delay in Response”; i.e. the present acceleration would depend upon force which was received by the particle a little earlier. Thus, causality is not violated here. However, as $\tau $ is too small (of the order of $10^{-24}s$ for fundamental particles), it doesn’t come into picture in macroscopic systems. For constant force, the acceleration becomes: $$\label{23-e} a\ =\ \frac{F}{m} \left(1\ -e^{-t/\tau } \right)$$ Which reminds of equations of current/ voltage growths in LR/ RC circuits respectively; or any first order equation of motion with damping force. Again going back to equation (\[22-c\]) and considering absence of any external force, $F=0$; the solution comes out to be: $$a\ =\ a_0\ e^{-t/\tau }$$ Which says that any perturbation in motion of charged particle would die out quickly and the charge would acquire another steady state. Thus here, there is no runaway seen here as well. Conclusion: =========== Equations of motion (\[23-d\]) and (\[23-e\]), based on newly derived equation of Radiation Reaction, give clear impression that the Radiation Reaction is Electromagnetic Friction force that provides damping effect to the motion of charged particle. The new expressions of Radiation Reaction, (\[11-d\]) and (\[11-f\]), successfully eliminate pathological solutions of runaway and acausal pre-accelaration, which were disturbing parts of AL equation (\[4\]) and ALD equation (\[5\]) so far. Further, they are found to be in same form and containing the similar terms as in ALD formula (\[5-a\]). Therefore, they can be believed upon. However, here particle’s time constant has been found to be $\mu _0 q^2/(3 \pi mc)$, which is twice that of AL formula (\[4\]) as well as ALD formula (\[5\]). The source of this discrepancy might be due to the fact that AL equation (\[4\]) was derived considering cyclic change in velocity; the process which might have actually averaged out contributions from rising and returning accelerations. The value of time constant $\tau$ mismatches with Lorentz’s theory of self- force [@lor] as well. Thus, self force phenomenon needs to be revisited and to be studied further to check this discrepancy. Acknowledgement =============== I am thankful to my institution Bhabha Atomic Research Center that inspired me to learn; and also for providing me facilities and resources that helped me in conceptualizing and writing this paper. [9]{} Griffiths D. ”Introduction to Electrodynamics”, 3rd edn (New Delhi: PHI) pp 460–473 (2011). Jackson J. D. ”Classical Electrodynamics”, Sect. 17.7, pp 597-599 (1962). Dirac P. A. M., ”Classical theory of radiating electrons”, Proc. Royal Society A 167, 148 (1938). Lorentz H. A., ”Theory of Electrons”, 2nd edn, Note 18, p 252 (1916) Rohrlich F., ”The dynamics of a charged sphere and the electron”, American Journal of Physics, 65, 1051 (1997)

ArXiv

--- abstract: 'We introduce a new notion of solution, which we call weak\* solutions, for systems of conservation laws. These solutions can be used to handle singular situations that standard weak solutions cannot, such as vacuums in Lagrangian gas dynamics or cavities in elasticity. Our framework allows us to treat the systems as ODEs in Banach space. Starting with the observation that solutions act linearly on test functions $\alpha\in X$, we require solutions to take values in the dual space $X^*$ of $X$. Moreover, we weaken the usual requirement of measurability of solutions. In order to do this, we develop the calculus of the Gelfand integral, which is appropriate for weak\* measurable functions. We then use the Gelfand calculus to define weak\* solutions, and show that they are stronger than the usual notion of weak solution, although for $BV$ solutions the notions are equivalent. It is expected that these solutions will also shed light on vexing issues of ill-posedness for multi-dimensional systems.' author: - 'Alexey Miroshnikov[^1], Robin Young [^2]' bibliography: - 'Refs.bib' title: | Weak\* solutions I: A new perspective on solutions\ to systems of conservation laws --- intro prelims gelfand hcl spaces [^1]: Department of Biostatistics and Epidemiology, University of Massachusetts Amherst, [email protected] [^2]: Department of Mathematics and Statistics, University of Massachusetts Amherst, [email protected]

ArXiv

--- abstract: 'Generative Adversarial Networks (GANs) have shown great promise recently in image generation. Training GANs for language generation has proven to be more difficult, because of the non-differentiable nature of generating text with recurrent neural networks. Consequently, past work has either resorted to pre-training with maximum-likelihood or used convolutional networks for generation. In this work, we show that recurrent neural networks can be trained to generate text with GANs from scratch using curriculum learning, by slowly teaching the model to generate sequences of increasing and variable length. We empirically show that our approach vastly improves the quality of generated sequences compared to a convolutional baseline. [^1]' author: - | **Ofir Press**[^2] $^{1}$, **Amir Bar**$^{*1,2}$, **Ben Bogin**$^{*1,3}$\ **Jonathan Berant**$^{1}$, **Lior Wolf**$^{1,4}$\ $^1$ School of Computer Science, Tel-Aviv University\ $^2$ Zebra Medical Vision     $^3$ IBM Research     $^4$ Facebook AI Research\ [[email protected]]{}\ bibliography: - 'emnlp2017.bib' title: 'Language Generation with Recurrent Generative Adversarial Networks without Pre-training' --- Introduction ============ Generative adversarial networks [@GAN] have achieved state-of-the-art results in image generation [@GAN; @DCGAN; @WGAN; @improvedWGAN]. For text generation, training GANs with recurrent neural networks (RNNs) has been more challenging, mostly due to the non-differentiable nature of generating discrete symbols. Consequently, past work on using GANs for text generation has been based on pre-training [@seqgan; @advDialogue; @nmtGAN; @nmtGAN2; @rttGAN; @zhangGenerating; @captionGAN] or joint training [@profF; @mladGAN] of the generator and discriminator with a supervised maximum-likelihood loss. Recently, two initial attempts to generate text using purely generative adversarial training were conducted by @improvedWGAN and @boundarySeekingGAN. In these works, a convolutional neural network (CNN) was trained to produce sequences of 32 characters. This CNN architecture is fully differentiable, and the authors demonstrated that it generates text at a reasonable level. However, the generated text was still filled with spelling errors and had little coherence. RNNs are a more natural architecture for language generation, since they condition each generated character on the entire history, and are not constrained to generating a fixed number of characters. In this paper, we extend the setup of @improvedWGAN and present a method for generating text with GANs. Our main contribution is a model that employs an RNN for both the generator and discriminator, similar to current state-of-the-art approaches for language generation [@sutskever2011generating; @mikolov2012statistical; @jozefowicz2016exploring]. We succeed in training the model by using curriculum learning [@elman1993learning; @bengio2009curriculum; @ranzato2015sequence]: At each stage we increase the maximal length of generated sequences, and train over sequences of variable length that are shorter than that maximal length. In addition, we aid the model by feeding it with ground truth characters before generation. We show that these methods vastly improve the quality of generated sequences. Sequences contain substantially more n-grams from a development set compared to those generated by a CNN, and generation generalizes to sequences that are longer than the sequences the model was trained on. Motivation ========== While models trained with a maximum-likelihood objective (ML) have shown success in language generation [@sutskever2011generating; @mikolov2012statistical; @jozefowicz2016exploring], there are drawbacks to using ML, that suggest training with GANs. First, using ML suffers from “exposure bias”, that is, at training time the model is exposed to gold data only, but at test time it observes its own predictions, and thus wrong predictions quickly accumulate, resulting in bad text generation. Secondly, the ML loss function is very stringent. When training with ML, the model aims to allocate all probability mass to the $i$-th character of the training set given the previous $i-1$ characters, and considers any deviation from the gold sequence as incorrect, although there are many possible sequences given a certain prefix. GANs suffer less from this problem, because the objective is to fool the discriminator, and thus the objective evolves dynamically as training unfolds. While at the beginning the generator might only generate sequences of random letters with spaces, as the discriminator learns to better discriminate, the generator will evolve to generate words and after that it may advance to longer, more coherent sequences of text. This interplay between the discriminator and generator helps incremental learning of text generation. Preliminaries {#sec:preliminaries} ============= @improvedWGAN and @boundarySeekingGAN trained a purely generative adversarial model (without pre-training) for character-level sentence generation. We briefly review the setup of @improvedWGAN, who use the Improved Wasserstein GAN objective [@WGAN; @improvedWGAN], which we employ as well. @boundarySeekingGAN have a similar setup, but employ the Boundary-Seeking GAN objective. The generator $G$ in @improvedWGAN is a CNN that transforms a noise vector $z \sim N(0,1)$ into a matrix $M \in \mathbb{R}^{32 \times V}$, where $V$ is the size of the character vocabulary, and 32 is the length of the generated text. In this matrix the $i$-th row is a probability distribution over characters that represents a prediction for the $i$-th output in the character sequence. To decode a sequence, they choose the highest probability character in each row. The discriminator $D$ is another CNN that receives a matrix as input and needs to determine if this matrix is the output of the generator $G$ or sampled from the real data (where each row in the matrix now is a one-hot vector). The loss of the Improved WGAN generator is: $$L_G = -{\mathbb{E}}_{\widetilde{x} \sim {\mathbb{P}}_g}[D(\widetilde{x})],$$ and the loss of the discriminator is: $$\begin{gathered} L_D = {\mathbb{E}}_{\widetilde{x} \sim {\mathbb{P}}_g}[D(\widetilde{x})] - {\mathbb{E}}_{x \sim {\mathbb{P}}_r} [D(x)] \\ + \lambda {\mathbb{E}}_{\widehat{x} \sim {\mathbb{P}}_{\widehat{x}}} [(\| \nabla_{\widehat{x}} D(\widehat{x}) \|_2 - 1)^2],\end{gathered}$$ Where ${\mathbb{P}}_r$ is the data distribution and ${\mathbb{P}}_g$ is the generator distribution implicitly defined by $\widetilde{x} = G(z)$. The last term of the objective controls the complexity of the discriminator function and penalizes functions that have high gradient norm, that is, change too rapidly. ${\mathbb{P}}_{\widehat{x}}$ is defined by sampling uniformly along a straight line between a point sampled from the data distribution and a point sampled from the generator distribution. A disadvantage of the generators in @improvedWGAN and @boundarySeekingGAN is that they use CNNs for generation, and thus the $i$-th generated character is not directly conditioned on the entire history of $i-1$ generated characters. This might be a factor in the frequent spelling mistakes and lack of coherence in the output of these models. We now present a model for language generation with GANs that utilizes RNNs, which are state-of-the-art in language generation. --------- ------ ------ ----- ----- Samples 1 2 3 4 64.4 25.9 5.1 0.4 --------- ------ ------ ----- ----- Recurrent Models ================ We employ a GRU [@gru] based RNN for our generator and discriminator. The generator is initialized by feeding it with a noise vector $z$ as the hidden state, and an embedded start-of-sequence symbol as input. The generator then generates a sequence of distributions over characters, using a softmax layer over the hidden state at each time step. Because we want to have a fully-differentiable generator, the input to the RNN generator at each time step is not the most probable character from the previous time step. Instead we employ a continuous relaxation, and provide at time step $i$ the weighted average representation given by the output distribution of step $i-1$. More formally, let $\alpha_{i-1}^c$ be the probability of generating the character $c$ computed at time step $i-1$, and let $\phi(c)$ be the embedding of the character $c$, then the input to the GRU at time step $i$ is $\sum_c \alpha_{i-1}^c \phi(c)$. This is fully differentiable compared to $\arg\max_{\phi(c)} \alpha_{i-1}^c$. We empirically observe that the RNN quickly learns to output very skewed distributions. The **discriminator** is another GRU that receives a sequence of character distributions as input, either one-hot vectors (for real data) or softer distributions (for generated data). Character embeddings are computed from the distributions and fed into the GRU. The discriminator then takes the final hidden state and feeds it into a fully connected layer which outputs a single number, representing the score that the discriminator assigns to the input. The models are trained with the aforementioned Improved WGAN objective (Section \[sec:preliminaries\]). An advantage of a recurrent generator compared to the convolutional generator of @improvedWGAN and @boundarySeekingGAN is that can output sequences of varying lengths, as we empirically show in Section \[sec:results\]. Our baseline model trains the generator and discriminator over sequences of length $32$, similar to how CNNs were trained in @improvedWGAN. We found that training this baseline was difficult and resulted in nonsensical text. We now present three extensions that stabilize the training process. \[tab:table2\] ---- ---- ---- -------------------------------------------------- ---------- ---------- ---------- --------- CL VL TH Samples 1 2 3 4 28.8 3.7 0.0 0.0 80.6 8.6 0.0 0.0 27.0 7.9 2.0 0.0 68.1 24.5 4.4 0.5 79.4 44.6 11.5 0.7 **87.7** **54.1** **19.2** **3.8** **Sequences of length 64. Examples in Table 3.** 87.5 51.3 15.1 1.7 ---- ---- ---- -------------------------------------------------- ---------- ---------- ---------- --------- \[tab:table3\] [l]{}\ #### Curriculum Learning (CL) In this extension, we start by training on short sequences and then slowly increase sequence length. In the first training stage, the generator $G$ generates sequences of length $1$, and the discriminator $D$ receives real and generated sequences of length $1$ as input. Then, the generator generates sequences of length $2$ and the discriminator receives sequences of length $2$. We increase sequence length in this manner until the maximum length of 32 characters. #### Variable Length (VL) Here, we define a maximum length $l$, and generate during training sequences of every length $\leq l$ in every batch. Without curriculum learning, this amounts to training $G$ and $D$ in every batch with sequences of length $i$, $1 \leq i \leq 32$. With curriculum learning, we generate at each step sequences of length $i$, $1 \leq i \leq l$, and slowly increase $l$ throughout training. #### Teacher Helping (TH) Finally, we propose a procedure where we help the generator learn to generate long sequences by conditioning on shorter ground truth sequences. Recall that in our baseline, the generator generates an entire sequence of characters that are fed as input to the discriminator. Here, when generating sequences of length $i$, we feed the generator a sequence of $i-1$ characters, sampled from the real data. Then, the generator generates a distribution over characters for the final character, which we concatenate to the real characters and feed as input to the discriminator. The discriminator observes a sequence of length $i$ composed of $i-1$ real characters and one character that is either real or generated. This could be viewed as a conditional GAN [@CGAN], where the first $i-1$ characters are the input and the final character is the output. Note that this extension may suffer from exposure bias, similar to the ML objective, and we plan to address this problem in future work. Results {#sec:results} ======= To directly compare to @improvedWGAN, we follow their setup and train our models on the Billion Word dataset [@1billion]. We evaluate by generating 640 sequences from each model and measuring %-IN-TEST-$n$, that is, the proportion of word $n$-grams from generated sequences that also appear in a held-out test set. We evaluate these metrics for $n \in \{1, 2, 3,4\}$. Our goal is to measure the extent to which the generator is able to generate real words with local coherence. In contrast to @WGAN and @improvedWGAN, where the generator is trained once for every $10$ training iterations of the discriminator, we found that training the generator for $50$ iterations every $10$ training iterations of the discriminator resulted in superior performance. In addition, instead of using noise vectors sampled from the $N(0,1)$ distribution as in @improvedWGAN, we sample noise vectors from the $N(0,10)$ distribution, since we found this leads to a greater variance in the generated samples when using RNNs. In all our experiments, we used single layer GRUs for both the discriminator and generator. The embedding dimension and hidden state dimension are both of size $512$. Following @improvedWGAN, we train all our models on sequences whose maximum length is $32$ characters. Table 1 shows results of the baseline model of @improvedWGAN, and Table 2 presents results of our models with various combinations of extensions (Curriculum Learning, Variable Length, and Teacher Helping). Our best model combines all of the extensions and outperforms the baseline by a wide margin on all metrics. The samples show that models that used both the Variable Length and Teacher Helping extensions performed better than those that did not. This is also backed by the empirical evaluation, which shows that 3.8% of the word 4-grams generated by the CL+VL+TH model also appear in the held-out test set. The weak performance of the curriculum learning model without the other extensions shows that curriculum learning by itself does not lead to better performance, and that training on variable lengths and with Teacher Helping is important. We note that curriculum learning did not perform well at generating sequences of length $32$, but did perform well at generating sequences of shorter lengths earlier in the training process. For example, the model that used only curriculum learning had a %-IN-TEST-1 of $79.9$ when it was trained on sequences of length $5$. This decreased to $59.7$ when the model reached sequences of length $10$, and continued decreasing until training stopped. This also shows the importance of Variable Length and Teacher Helping. Finally, to check the ability of our models to generalize to longer sequences, we generated sequences of length $64$ with our CL+VL+TH model, which was trained on sequences of up to $32$ characters (Table 3). We then evaluated the generated text, and this evaluation shows that there is a small degradation in performance (Table 2). Conclusion ========== We show for the first time an RNN trained with a GAN objective that learns to generate natural language from scratch. Moreover, we demonstrate that our model generalizes to sequences longer than the ones seen during training. In future work, we plan to apply these models to tasks such as image captioning and translation, comparing them to models trained with maximum likelihood. [^1]: Code for our models and evaluation methods is available at <https://github.com/amirbar/rnn.wgan> [^2]:   Denotes equal contribution. Author ordering determined by coin flip.

ArXiv