Generalized gradient flows in Hadamard manifolds and convex optimization on entanglement polytopes
Extracted source text ยท arXiv 2511.12064
\documentclass[12pt,a4paper]{article}
\usepackage{amssymb}
\usepackage{amsmath, amsthm}
\usepackage{arydshln}
\usepackage{epsfig}
\usepackage{setspace}
\usepackage{comment}
\usepackage[margin=1in]{geometry}
%\usepackage{setspace}
%\geometry{left=27mm,right=27mm,top=27mm,bottom=27mm}
%\geometry{left=1in,right=1in,top=1in,bottom=1in}
%-------------- Macro ------------------------------
\def\RR{{\mathbb{R}}}
\def\ZZ{{\mathbb{Z}}}
\def\QQ{{\mathbb{Q}}}
\def\CC{{\mathbb{C}}}
\numberwithin{equation}{section}
%\newcommand{\spec}{\mathop{\rm spec} }
\newcommand{\proj}{\mathop{\rm proj} }
\newcommand{\pr}{\mathop{\rm Pr} }
\newcommand{\rank}{\mathop{\rm rank} }
\newcommand{\diam}{\mathop{\rm diam} }
\newcommand{\dom}{\mathop{\rm dom} }
\newcommand{\lv}{\mathop{\rm Lv} }
\newcommand{\interior}{\mathop{\rm int} }
\newcommand{\supp}{\mathop{\rm supp} }
\newcommand{\fr}{\mathop{\rm frac}}
\newcommand{\Conv}{\mathop{\rm Conv} }
\newcommand{\Cone}{\mathop{\rm Cone} }
\newcommand{\val}{\mathop{\rm val}}
\newcommand{\opt}{\mathop{\rm opt}}
\newcommand{\Argmin}{\mathop{\rm Argmin} }
\newcommand{\argmin}{\mathop{\rm argmin} }
\newcommand{\Det}{\mathop{\rm Det} }
\newcommand{\sgn}{\mathop{\rm sgn} }
\newcommand{\ncrank}{\mathop{\rm nc\mbox{-}rank} }
\newcommand{\ncrk}{\mathop{\rm ncrk }}
\newcommand{\ri}{\mathop{\rm ri} }
\newcommand{\diag}{\mathop{\rm diag} }
%\newcommand{\trace}{\mathop{\rm Tr} }
\newcommand{\capacity}{\mathop{\rm cap} }
\DeclareMathOperator{\trace}{Tr}
\DeclareMathOperator{\rk}{rk}
\DeclareMathOperator{\spec}{spec}
\newcommand{\ad}{\mathop{\rm ad} }
\newcommand{\Ad}{\mathop{\rm Ad} }
\newtheorem{Thm}{Theorem}[section]
\newtheorem{Prop}[Thm]{Proposition}
\newtheorem{Lem}[Thm]{Lemma}
\newtheorem{Cor}[Thm]{Corollary}
\newtheorem{Question}[Thm]{Question}
\newtheorem{Conj}[Thm]{Conjecture}
\theoremstyle{definition}
\newtheorem{Clm}{Claim}
\newtheorem{Def}[Thm]{Definition}
\newtheorem{Rem}[Thm]{Remark}
\newtheorem{Ex}[Thm]{Example}
\newtheorem{Ass}[Thm]{Assumption}
\title{Generalized gradient flows in Hadamard manifolds \\and\\
convex optimization on entanglement polytopes}
\author{Hiroshi Hirai\footnote{
Graduate School of Mathematics,
Nagoya University, Nagoya, 464-8602, Japan.
\texttt{\footnotesize hirai.hiroshi@math.nagoya-u.ac.jp}
}}
%\date{May 2025}
\begin{document}
\maketitle
\begin{abstract}
In this paper, we address the optimization problem of minimizing $Q(df_x)$ over a Hadamard manifold ${\cal M}$, where $f$ is a convex function on ${\cal M}$, $df_x$ is the differential of $f$ at $x \in {\cal M}$, and $Q$ is a function on the cotangent bundle of ${\cal M}$. This problem generalizes the problem of minimizing the gradient norm $\|\nabla f(x)\|$ over ${\cal M}$, studied by Hirai and Sakabe FOCS2024. We formulate a natural class of $Q$ in terms of convexity and invariance under parallel transports, and introduce a generalization of the gradient flow of $f$ that is expected to minimize $Q(df_x)$. For basic classes of manifolds, including the product of the manifolds of positive definite matrices, we prove that this gradient flow attains $\inf_{x\in {\cal M}} Q(df_x)$ in the limit, and yields a duality relation. This result is applied to the Kempf-Ness optimization for GL-actions on tensors, which is Euclidean convex optimization on the class of moment polytopes, known as the entanglement polytopes. This type of convex optimization arises from tensor-related subjects in theoretical computer science, such as quantum functional, $G$-stable rank, and noncommutative rank.
\end{abstract}
Keywords: Hadamard manifold, gradient flow, geodesic convexity, boundary at infinity,
positive definite matrices, moment map, moment polytope, entanglement polytope, tensor,
quantum functional, $G$-stable rank, noncommutative rank \\
MSC-classifications: 90C25, 53C35
\section{Introduction}
The present paper is a continuation of the work~\cite{Hirai2025,HiraiSakabe2024FOCS} studying ``divergent" gradient flows of lower-unbounded convex functions $f$ on Hadamard manifolds ${\cal M}$
and their applications to optimization problems hidden
in the boundary ${\cal M}^{\infty}$ at infinity of ${\cal M}$.
%
In~\cite{HiraiSakabe2024FOCS}, Hirai and Sakabe established
a duality relation of the infimum of the gradient-norm $\|\nabla f(x)\|$ over $x \in {\cal M}$ by the use of the gradient flow of $f$.
Specifically, if the infimum is positive,
then it is equal to
the supremum of the negative of the {\em recession function}
$f^{\infty}$~\cite{Hirai_Hadamard2022,KLM2009JDG} of $f$, which is a function defined on the boundary ${\cal M}^{\infty}$.
Further, the infimum and supremum are attained by
the limit of the solution $x(t)$ of the gradient-flow ODE $\dot x(t) = -\nabla f (x(t))$, $x(0) = x_0$ as follows:
\begin{equation}\label{eqn:min_gradient-norm}
\lim_{t \to \infty} \|\nabla f(x(t))\| = \inf_{x \in {\cal M}} \|\nabla f(x)\| =
\sup_{\xi \in {\cal M}^{\infty}} - f^{\infty}(\xi) = - f^{\infty} \left(\lim_{t\to \infty}x(t)\right),
\end{equation}
where the convergence $\lim_{t \to \infty} x(t) \in {\cal M}^{\infty}$ is with respect to the cone topology of
the visual compactification ${\cal M} \cup {\cal M}^{\infty}$.
Such an investigation originates from
{\em operator scaling}~\cite{Gurvits2004} and its generalizations that
have brought about remarkable developments and numerous applications in recent years; see e.g., \cite{BFGOWW_FOCS2018,BFGOWW_FOCS2019,GGOW} and the references therein. They involve optimization of the {\em Kempf-Ness functions}
associated with reductive group actions, which is formulated as
(possibly lower-unbounded) convex optimization on Hadamard manifolds.
%
Hirai and Sakabe~\cite{HiraiSakabe2024FOCS} utilized (\ref{eqn:min_gradient-norm}) to
reveal the asymptotic behavior of operator scaling
the gradient flow and its discretization ({\em gradient descent}) for operator scaling.
They also revealed the link with the theory of {\em moment-weight inequality}
(Georgoulas, Robbin, and Salamon~\cite{GRS_Book}), which provides
differential geometric understanding of stability in geometric invariant theory (GIT).
%
In a follow-up paper, Hirai~\cite{Hirai2025} showed a refinement of (\ref{eqn:min_gradient-norm})
for the Hadamard manifold ${\cal P}_n$ of positive definite matrices ({\em PD-manifold}),
in which the norm $\| \cdot \|$ is a general GL-invariant Finsler norm and $x(t)$ is a non-differentiable generalization~(a {\em curve of maximal slope}~\cite{Ambrosio_Book}) of gradient flow.
This leads to a new characterization of the {\em noncommutative rank}~\cite{IQS2017}
in terms of the trace norm of the scaling limit of operator scaling.
%
%a class of related nonpositively curved metric space
%Gromov hyperbolic space
In the light of these developments, we actively interpret (\ref{eqn:min_gradient-norm})
as an optimization framework
of minimizing the gradient-norm $\|\nabla f(x)\|$ over ${\cal M}$, and consider the following generalization.
{\em Given a function $Q$ on the cotangent bundle of ${\cal M}$,
minimize the $Q$-value of
the differential $df_x$ of $f$ over $x \in {\cal M}$:}
\begin{equation}\label{eqn:minimum_Q-gradient-norm_problem}
\mbox{Min.} \quad Q(df_x) \quad {\rm s.t.} \quad x \in {\cal M}.
\end{equation}
We call this problem the {\em minimum $Q$-gradient-norm problem}.
It is clear that
the above gradient-norm setting is obtained by taking the dual norm $Q := \|\cdot\|^*$. % since $\|\nabla f(x)\| = \|df_x\|^*$.
We then address the following questions:
\begin{itemize}
\item[1.] Which does $({\cal M},Q)$ admit a duality of form
$\displaystyle \inf_{x \in {\cal M}} Q(df_x) = \sup \ [\, \cdots]$?
\item[2.] Is there an analogue of gradient flow $x(t)$ that minimizes $Q(df_x)$ in its limit?
\item[3.] Can we explore further applications?
\end{itemize}
The main contribution of this paper is a partial but
satisfactory answer to these questions.
We formulate a natural class of functions $Q$ in terms of convexity and
invariance under parallel transports (or {\em holonomy group}),
and introduce a generalization of the gradient flow, called the {\em $Q$-gradient flow},
that is expected to minimize $Q(df_x)$ in the limit.
We establish in Theorem~\ref{thm:main}
an analogue of (\ref{eqn:min_gradient-norm}) for three basic classes of Hadamard manifolds
(Euclidean space, generic Hadamard manifold, the product of PD-manifolds),
although we could not prove it for the general case (Conjecture~\ref{conj}).
%
This limited result has promising applications.
Indeed, it is
applicable to the Kempf-Ness optimization for GL-actions on tensors, that widely appears in quantum information and theoretical computer science,
such as quantum marginal problem, tensor rank, and matrix multiplication; see~\cite{Burgisser_ACT_Book,BFGOWW_FOCS2018,Landsberg_tensors,Landsberg_tensors2019}.
In this setting,
the minimum $Q$-gradient-norm problem~(\ref{eqn:min_gradient-norm})
becomes Euclidean convex optimization on the {\em moment polytope} $\Delta(v)$~\cite{GuilleminSternberg1982-I,GuilleminSternberg1984-II,Kirwan1984}
for the GL-action on a tensor $v$. This class of moment polytopes is particularly called the {\em entanglement polytopes} for their importance of studying entanglements in quantum multi-particle systems~\cite{WDGC2013_Science}.
%
Convex optimization on the entanglement polytope $\Delta(v)$ has gained attention for
its connections with tensor-rank subjects.
However, there is no good inequality description for moment polytopes,
and optimizing on $\Delta(v)$ appears to be quite difficult; see the recent work~\cite{vdBCLNWZ_STOC2025} for this direction.
Our result implies an inf-sup duality relation (Theorem~\ref{thm:duality_entanglement})
for optimization problems on $\Delta(v)$ with convex objectives having a natural symmetry.
It is expected to lead to a good characterization
(NP $\cap$ co-NP characterization) for this problem.
%Also the dual problem is written as convex optimization
%on the {\em Euclidean building} arising as the boundary of the product of PD-manifolds.
Further, the convergence of the $Q$-gradient flow
gives rise to a sequence of points in $\Delta(v)$
accumulating to a minimizer (Theorem~\ref{thm:moment-limit}).
This generalizes the {\em moment-limit theorem}~\cite{GRS_Book},
and suggests a simple (sub)gradient-type algorithm solving this problem.
%
We demonstrate how our theory is applied to concrete examples ({\em quantum functional}~\cite{CristandlVramaZuiddam2023},
{\em $G$-stable rank}~\cite{Derksen2022ANT}, and noncommutative rank).
It is an important challenge to
discretize the $Q$-gradient flow and develop a unified algorithm for these problems.
The rest of the paper is organized as follows.
In Section~\ref{sec:preliminaries}, we provide the necessary backgrounds on Riemannian geometry, Hadamard manifold, convex analysis, and PD-manifold.
In Section~\ref{sec:Q-gradient_flow},
we introduce the $Q$-gradient flow, and establish the duality of
the minimum $Q$-gradient-norm problem for the above-mentioned classes of $({\cal M},Q)$.
In Section~\ref{sec:entanglement_polytope},
we present applications to convex optimization on entanglement polytopes.
\section{Preliminaries}\label{sec:preliminaries}
We will use the standard notions of Riemannian geometry; see e.g.,\cite{Sakai1996}. See also \cite{Boumal_Book} for optimization perspective.
We consider a complete Riemannian manifold ${\cal M}$
equipped with Riemannian metric $\langle, \rangle$ and Levi-Civita connection $\nabla$, where
$T_x = T_x{\cal M}$ and
$T_x^* = T_x^*{\cal M}$ denote the tangent and cotangent spaces at $x \in {\cal M}$, respectively.
%
Let $T{\cal M} = \bigsqcup_{x \in {\cal M}}T_x$ and $T{\cal M} = \bigsqcup_{x \in M} T_x^*$
denote the tangent and cotangent bundles of ${\cal M}$, repsectively.
%
The norm $\|\cdot\|$ on $T_x$ is defined by
$\|v\| = \sqrt{\langle v,v \rangle}$ $(v \in T_x)$, and
the dual norm $\|\cdot\|^*$ on $T_x^*$ is defined by
$\|p\|^* := \sup_{v \in T_x:\|v\|=1} p(v)$ $(p \in T_x^*)$.
%
For a curve
$\gamma: [0, \ell] \to {\cal M}$, let $\dot \gamma(t) \in T_{\gamma(t)}$ denote the tangent vector of $\gamma$ at $t$.
For a function $f$ on ${\cal M}$,
the differential $df_x \in T_x^*$ of $f$ at $x \in {\cal M}$ is given
by $df_x(v) := (d/dt)\mid_{t = 0} f(\gamma(t))$, where $\gamma$ is any curve with $\gamma(0) = x$ and $\dot \gamma(0) = v$.
The gradient $\nabla f(x) \in T_x$ is given by $\langle \nabla f(x),v\rangle = df_x(v)$ $(v \in T_x)$.
For $t \in [0,\ell]$,
let $\tau_{\gamma}^t:T_{\gamma(0)} \to T_{\gamma(t)}$ denote the parallel transport
along $\gamma$ with respect to $\nabla$.
The inverse map $\tau_{\gamma}^{-t}:= \tau_{\gamma^{-1}}^{\ell - t}$
is obtained by the reverse curve $t \mapsto \gamma^{-1}(t) := \gamma(\ell -t)$.
The parallel transport $\tau_{\gamma}^t:T^*_{\gamma(0)} \to T^*_{\gamma(t)}$ on the cotangent spaces
is defined by $\tau_{\gamma}^{t}(p)(\cdot) := p(\tau^{-t}_{\gamma} (\cdot))$
for $p \in T^{*}_{\gamma(0)}$.
For a vector/covector field $V = (V_x)_{x \in {\cal M}}$,
the covariant derivative $\nabla_{u} V$ by $u \in T_x$
is given
by $\nabla_{u}V :=
(d/dt)\mid_{t = 0} \tau_{\gamma}^{-t} V_{\gamma(t)}$, where $\gamma$ is any curve
with $\gamma(0) = x$ and $\dot\gamma(0) = u$.
The length of a curve $\gamma:[0,\ell] \to {\cal M}$ is given by $\int_{0}^{\ell} \|\dot \gamma(t)\|dt$.
The distance $d(x,y)$ of $x,y \in {\cal M}$ is defined as the infimum of the length of a curve connecting $x,y$.
By completeness, the infimum is attained
by a {\em geodesic}, which is a curve $\gamma$ with $\nabla_{\dot \gamma(t)}\dot\gamma = 0$.
%
\subsection{Hadamard manifold}
A {\em Hadamard manifold} is a simply-connected and complete Riemannian manifold
having nonpositive sectional curvature.
Examples of Hadamard manifolds include
Euclidean space, hyperbolic space,
and the manifold of positive definite matrices (Section~\ref{subsec:psd}).
%
In a Hadamard manifold ${\cal M}$,
a geodesic connecting two points is uniquely determined
(up to affine rescaling).
For $u \in T_x$, there is a unique geodesic ray
$\gamma:[0,\infty) \to {\cal M}$
such that $\gamma(0) = x$ and $\dot\gamma(0) = u$.
We denote $\gamma(t)$ particularly by $\exp_x tu$.
If $\|u\| =1$, then it is called a unit-speed geodesic ray.
The exponential map $u \mapsto \exp_x u$ is a diffeomorphism from $T_x$ to ${\cal M}$.
The parallel transport along the geodesic from $x$ to $y$
is simply denoted by $\tau_{x\to y}$.
A function $f: {\cal M} \to \RR$ is said to be {\em geodesically convex}
if $f \circ \gamma$ is convex for any geodesic $\gamma:[0,1] \to {\cal M}$:
\[
f(\gamma(t)) \leq (1-t) f(\gamma(0)) + t f(\gamma(1)) \quad (t \in [0,1]).
\]
If $f$ is twice differentiable,
then the geodesic convexity is equivalent to the property that $(d^2/dt^2)(f\circ \gamma)(t) \geq 0$ for any geodesic $\gamma$.
This condition is rephrased that
the {\em covariant Hessian} $(u,v) \mapsto (\nabla df)(u,v) := (\nabla_{u} df)(v)$
is positive semidefinite for every $x \in {\cal M}$; see \cite[Theorem 11.23]{Boumal_Book}.
Two unit-speed geodesic rays $\gamma,\gamma'$ are said to be {\em asymptotic}
if $\sup_{t \in \RR_{\geq 0}} d(\gamma(t),\gamma'(t)) < + \infty$.
The asymptotic relation is an equivalence relation.
The set ${\cal M}^{\infty}$ of equivalence classes of all unit-speed rays is called the {\em boundary of {\cal M} at infinity}; see \cite[Chapter II.8]{BridsonHaefliger1999}.
We consider its {\em Euclidean cone} $C{\cal M}^{\infty} := ({\cal M}^{\infty} \times \RR_{\geq 0})/\sim$,
where $(\xi,r) \sim (\xi',r')$ if and only if $r=r'=0$
or $(\xi,r) =(\xi',r')$.
Fix arbitrary $x_0 \in {\cal M}$.
Then, it is known that $u \mapsto (t \mapsto \exp_{x_0}t u) = ((t \mapsto \exp_{x_0}t u/\|u\|), \|u\|) $
is a bijection from $T_{x_0}$ to $C{\cal M}^{\infty}$.
This induces a topology on $C{\cal M}^{\infty} \simeq T_{x_0}$, called the {\em cone topology}, which is independent of $x_0$.
%
The boundary cone $C{\cal M}^{\infty}$ also has a geodesic metric determined
from the {\em Tits metric} on ${\cal M}^{\infty}$; see~\cite[Chapter II.9]{BridsonHaefliger1999}.
This metric gives a different topology from the cone topology. but makes $C{\cal M}^{\infty}$ a {\em Hadamard space}, which is a metric-space generalization
of Hadamard manifolds and enjoys the unique geodesic property.
So we can consider geodesic convexity on the space $C{\cal M}^{\infty}$;
see e.g., \cite{Hirai_Hadamard2022}.
Let $f:{\cal M} \to \RR$ be a geodesically convex function.
To study the behavior of $f$ at the boundary $C{\cal M}^{\infty}$ of ${\cal M}$,
define
the {\em recession function} $f^{\infty} = f_{x_0}^{\infty}:C{\cal M}^{\infty} \to \overline{\RR}$ of $f$~\cite{Hirai_Hadamard2022,KLM2009JDG} by
\begin{equation}
f^{\infty}(\xi) := \lim_{t \to \infty}\frac{f(\exp_{x_0} t\xi)- f(x_0)}{t} \quad (\xi \in C{\cal M}^{\infty} \simeq T_{x_0}),
\end{equation}
where $f^{\infty}$ is independent of $x_0$. That is,
if $t \mapsto \exp_x t u$ and $t \mapsto \exp_y t v$ represent the same point in $C{\cal M}^{\infty}$,
then $f^{\infty}_{x}(u) = f^{\infty}_{y}(v)$.
%By this reason, we may omit subscript $x_0$.
It is known~\cite{Hirai_Hadamard2022} that $f^{\infty}$ is geodesically convex on the Hadamard space $C{\cal M}^{\infty}$.
To define norm-like functions $Q$ that behave well under parallel transports,
we will utilize the notion of
the {\em (restricted) Holonomy group}
\[
Hol_{x_0}({\cal M}) := \{\tau_\gamma^{1} \in SO(T_{x_0})
\mid \mbox{$\gamma:[0,1] \to {\cal M}:\gamma(0)=\gamma(1) =x_0$: piecewise loop at $x_0$}\},
\]
where $SO(T_{x_0})$ denotes the rotation group keeping the norm $\|\cdot\|$ invariant.
See e.g., \cite[Chapter III.6]{Sakai1996} for holonomy groups.
The holonomy group $Hol_{x_0}({\cal M})$ also acts on $T_{x_0}^*$
by $(\tau_{\gamma}^1,p) \mapsto \tau_{\gamma}^1 p (= p(\tau_\gamma^{-1}(\cdot)))$,
where $\gamma$ is a loop at $x_0$.
A Hadamard manifold ${\cal M}$ is said to be {\em generic} if
$Hol_{x_0}({\cal M}) = SO(T_{x_0})$.
For example, hyperbolic space is generic.
On the other hand, the holonomy group of Euclidean space is the trivial group.
\subsection{Convex analysis (on Hermitian matrices)}\label{subsec:convex-analysis}
%
We will utilize convex analysis on a finite-dimensional real vector space ${\cal V} (\simeq \RR^n)$;
see e.g.,\cite{Rockafellar}.
Let $\overline{\RR} := \RR \cup \{\infty\}$.
A function $F: {\cal V} \to \overline{\RR}$ is said to
be convex if $\lambda F(u) + (1- \lambda)F(v) \geq F(\lambda u + (1- \lambda) v)$
for $u,v \in {\cal V}$ and $\lambda \in [0,1]$.
Let $\dom F := \{u \in {\cal V}: F(u) < \infty\}$ denote the domain of $F$.
For $r \in \RR$, let $\lv(F,r) := \{u \in {\cal V}: F(u) \leq r\}$ denote the level set of $F$.
For a subset $C \subseteq {\cal V}$, let $\delta_C:{\cal V} \to \{0,\infty\}$
denote the indicator function defined by $\delta_C(u) := 0$ if $u \in C$ and $\infty$ otherwise.
For a norm $\|\cdot\|$ and $v \in {\cal V}$, let
$B_{\| \cdot \|}(v,r) := \{u \in {\cal V} \mid \|u-v\|\leq r\}$
denote the closed ball with respect to $\|\cdot\|$.
We (formally) consider the Legendre-Fenchel conjugate
$F^*:{\cal V}^* \to \overline{\RR}$ in the dual space ${\cal V}^*$
as
\[
F^*(p) := \sup_{u \in {\cal V}} p(u) - F(u) \quad (u \in {\cal V}^*).
\]
If $F$ is lower-semicontinuous convex, then ${F}^{**} = F$.
The subdifferential $\partial F(u)$ of $F$ at $u \in {\cal V}$ is given by
\begin{eqnarray}
\partial F(u) &:= & \{ p \in {\cal V}^* \mid F(v) \geq F(u) + p(v-u)\ (\forall v \in {\cal V})\} \nonumber \\
&=& \{ p \in {\cal V}^* \mid F(u) + F^*(p) = p(u)\}. \label{eqn:subdifferential}
\end{eqnarray}
An element $p$ in $\partial F(u)$ is called a subgradient of $F$ at $u$.
If $F$ is differentiable at $u$, then $\{dF_u\} = \partial F(u)$.
%Also, $p \in \partial F(u)$ if and only if $p(u) = F(u) + F^*(p)$.
In this paper, convex analysis on the space of Hermitian matrices (Lewis~\cite{Lewis1996})
will play important roles.
For a square complex matrix $A$, let $A^{\dagger}$ denote the conjugate transpose of $A$.
For $\lambda \in \RR^n$, let $\diag \lambda$ denote the diagonal matrix
having diagonals $\lambda_1,\lambda_2,\ldots,\lambda_n$ in order.
Let ${\cal H}_{n}$ denote the real vector space of $n \times n$ Hermitian matrices.
The dual space ${\cal H}_n^*$ is identified with ${\cal H}_n$ by $X(Y) := \trace XY$.
A convex function $F:{\cal H}_n \to \overline{\RR}$
is said to be {\em unitarily invariant}
if $F(k X k^{\dagger}) = F(X)$
for any $X \in {\cal H}_n$ and unitary matrix $k \in U_n$.
In this definition, the unitary group
$U_n$ may be replaced by the special unitary group $SU_n$.
A convex function $f:\RR^n \to \overline{\RR}$ is said to be {\em symmetric} if $f(x) = f(\sigma x)$
for any $x \in \RR^n$ and permutation matrix $\sigma$.
For a symmetric convex function $f$,
the function $F_f:{\cal H}_n \to \overline{\RR}$ defined by
$F_f(k \diag \lambda k^{\dagger}) := kf(\lambda)k^\dagger$ $(k \in SU_n,\lambda \in \RR^n)$ is unitarily invariant convex.
Conversely,
any unitarily invariant convex function is written in this way~\cite{Davis1957}.
The conjugate $F_f^*$ is also unitarily invariant, and is equal to $F_{f^*}$~\cite{Lewis1996}.
A subgradient of $F$ at $X$ can be computed by that of $f$ at $\lambda$;see \cite[Section 3]{Lewis1996} for details.
Specifically, if $F$ is differentiable,
then $dF_X = k^{\dagger} \diag df_{\lambda} k$ for $\lambda \in \RR^n$, $k \in U_n$ with
$X= k \diag \lambda k^\dagger$.
These arguments are naturally extended to convex functions on
the product of ${\cal H}_{n_1},{\cal H}_{n_2},\ldots {\cal H}_{n_k}$ that are invariant
under the action $X_i \mapsto k X_i k^{\dagger}$ of the unitary group $U_{n_i}$
for each component ${\cal H}_{n_i}$.
%
\begin{Ex}[Unitarily invariant norms $\|\cdot\|_{\rm op}$ and $\|\cdot \|_{\rm tr}$]\label{ex:norm}
The operator norm $\| \cdot \|_{\rm op}$ (the maximum of singular values) on ${\cal H}_n$
is unitarily invariant convex, and written as $F_{\|\cdot\|_{\infty}}$
for the $\ell_\infty$-norm $\|\cdot \|_{\infty}$ on $\RR^n$.
The trace norm $\|\cdot\|_{\rm tr}$ (the sum of singular values)
is unitarily invariant convex, and written as
$F_{\|\cdot\|_{1}}$ for the $\ell_1$-norm $\|\cdot \|_{1}$.
Their conjugates are given by
$(\| \cdot \|_{\rm op})^* = \delta_{B_{\|\cdot\|_{\rm tr}}(0,1)}$ and
$(\| \cdot \|_{\rm tr})^* = \delta_{B_{\|\cdot\|_{\rm op}}(0,1)}$.
The Frobenius norm $\|\cdot\|_{\rm F}$ is unitarily invariant convex,
is written as $F_{\|\cdot\|_2}$ for the $\ell_2$-norm $\|\cdot \|_{2}$, where $(\| \cdot \|_{\rm F})^* = \delta_{B_{\|\cdot\|_{\rm F}}(0,1)}$.
\end{Ex}
Let ${\cal H}_n^+ \subseteq {\cal H}_n$ denote the set of $n \times n$ positive semidefinite matrices, and
let ${\cal D}_n^+ \subseteq {\cal H}_n^+$ denote the set of positive semidefinite matrices
$X$ with $\trace X = 1$, i.e., {\em density matrices}.
%\begin{equation}
%{\cal D}^+_n = {\cal D}_n \cap {\cal H}_n^+.
%\end{equation}
%
\begin{Ex}[von Neumann entropy]\label{ex:entropy}
The {\em von Neumann entropy} $H: {\cal H}_n \to \RR \cup \{-\infty\}$ is defined by
\[
H(X) := - \trace X \log_2 X
\]
for $X \in {\cal D}_n^+$ and $-\infty$ for $X \not \in {\cal D}_n^+$,
where $\log_2 X$ is the matrix logarithm defined by $X = 2^{\log_2 X}$.
For $X = k \diag \lambda k^{\dagger}$ with $k \in U_n$, $\lambda \in \RR^n$, it holds $H(X) = - \sum_{i=1}^n \lambda_i \log_2 \lambda_i$ if $\lambda \in \RR^n_{\geq 0}$ with $\sum_{i=1}^n \lambda_i =1$ and $-\infty$ otherwise,
which is the {\em Shannon entropy} of the probability vector $\lambda$.
In particular, $-H$ is unitarily invariant convex.
The conjugate of the negative of the Shannon entropy is given by
$x \mapsto \log \sum_{i=1}^n 2^{x_i}$.
Therefore, the conjugate $(-H)^*$ of the negative of the von Neumann entropy is given by
\[
(-H)^*(Y) = \log \trace 2^{Y} \quad (Y \in {\cal H}^*_n = {\cal H}_n).
\]
\end{Ex}
\subsection{Manifold ${\cal P}_n$ of positive definite matrices}\label{subsec:psd}
Let ${\cal P}_n \subseteq {\cal H}_n$
denote the set of positive definite $n \times n$
Hermitian matrices. It is an open set of ${\cal H}_n$ and
has a natural manifold structure, where
the tangent space $T_x$ at each point $x \in {\cal P}_n$
is naturally identified with ${\cal H}_n$.
The general linear group $GL_n(\CC)$ acts transitively on ${\cal P}_n$
as $(g,x) \mapsto g x g^{\dagger}$, where the isotropy subgroup at $I$
is the unitary group $U_n$.
The Riemannian metric $\langle,\rangle$
on ${\cal P}_n$ is defined by $\langle X,Y \rangle := \trace x^{-1}X x^{-1}Y$ for $X,Y \in T_x$, which
which is invariant under this action.
The resulting Riemannian manifold ({\em PD-manifold}) is a well-studied and representative example
of Hadamard manifolds.
Various Riemannian notions are written explicitly; see e.g., \cite[Chapter 6]{Bhatia_PositiveDefiniteMatrices} and \cite[Chapter II.10]{BridsonHaefliger1999}.
For example,
any geodesic is written as $t \mapsto g e^{t H} g^\dagger$ for $g \in GL_n(\CC)$ and
$H \in {\cal H}_n$, where $e^{\bullet}$ denotes the matrix exponential.
In particular, a geodesic issuing at $x$ is written as
$t \mapsto x^{1/2} e^{t x^{-1/2} H x^{-1/2}} x^{1/2}$ for $H \in {\cal H}_n$.
The parallel transport $\tau_{x \to I}:T_x \to T_I$ is given by
$\tau_{x \to I}H = x^{-1/2}Hx^{-1/2}$.
The asymptotic relation is computed via eigenvalue decomposition and QR-factorization as follows:
For $H \in T_x$,
consider a geodesic ray $\gamma(t) = \exp_x t H
= x^{1/2} e^{tx^{-1/2}Hx^{-1/2}}x^{1/2} = g e^{t \diag \lambda} g^{\dagger}$ for $g \in GL_n(\CC)$ and $\lambda \in \RR^n$ with $\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_n$.
Observe\footnote{$d(b e^{t \diag \lambda}b^{\dagger}, e^{t \diag \lambda}) =
d(y(t)y(t)^{\dagger}, I)
$ for $y(t) := e^{-t\diag \lambda/2}b e^{t \diag \lambda/2}$ that is bounded over $t \in \RR_{\geq 0}$.
}
that $t \to b e^{t \diag \lambda} b^{\dagger}$ for upper-triangular $b$
is asymptotic to $t \mapsto e^{t \diag \lambda}$.
Via QR factorization $g = k b$ for unitary $k \in U_n$ and upper-triangular $b \in GL_n(\CC)$,
geodesic $\gamma$ is asymptotic to the geodesic ray $t \mapsto e^{tk\diag \lambda k^{\dagger}}$
issuing at $I$.
The submanifold ${\cal P}_n^1 := \{ x \in {\cal P}_n \mid \det x = 1\}$ of ${\cal P}_n$
is totally geodesic in the sense that any geodesic in ${\cal P}_n$ containing any two points in ${\cal P}_n^1$
belongs to ${\cal P}_n^1$.
The special linear group $SL_n(\CC)$ acts isometrically and transitively on ${\cal P}^1_n$
as $(g,x) \mapsto g x g^{\dagger}$, where the isotropy subgroup at $I$
is the special unitary group $SU_n$. So ${\cal P}_n^1$
is the irreducible symmetric space $SL_n(\CC) / SU_n$, whereas
${\cal P}_n = GL_n(\CC)/U_n$ is a reducible symmetric space
admitting de Rham decomposition ${\cal P}_n = \RR \times {\cal P}_n^1$.
%
%\[
%{\cal H}_n^0 := \{H \in {\cal H}_n \mid \trace H = 0\}.
%\]
%The cotangent space $T_I^*{\cal P}_n^1$ is identified with quotient space
%${\cal H}_n/\ker \trace$.
%We choose the representative
%as the space of Hermitian matrices with trace one:
%\[
%{\cal D}_n := \{H \in {\cal H}_n \mid \trace H = 1\}.
%\]
%That is, $T_I^*{\cal P}_n^1 = {\cal D}_n$,
%where the vector-space structure of ${\cal D}_n$ is induced from ${\cal H}_n/\ker \trace$.
The tangent space of ${\cal P}_n^1$ at $I$ is
given by $T_I{\cal P}_n^1 = \{ H \in {\cal H}_n \mid \trace H = 0\}$.
Consider the holonomy group $Hol_I({\cal P}_n^1)$ at $I$.
It is known (see e.g., \cite[Chapter IV, Theorem 6.9, Remark 6.14]{Sakai1996}) that the holonomy group
of a simply-connected irreducible symmetric space
is given by the adjoint action of the isotopy subgroup.
Therefore, $Hol_I({\cal P}_n^1)$ is equal to $SU_n$ acting on
$T_I {\cal P}_n^1$ as $(k,H) \mapsto k H k^{\dagger}$.
This is equal to the holonomy group $Hol_I({\cal P}_n)$ of ${\cal P}_n$ at $I$
since the holonomy group
is the product of that of irreducible components.
\paragraph{Euclidean building $C{\cal P}_n^{\infty}$.}
It is well-known \cite[Chapter II.10]{BridsonHaefliger1999} that the boundary $C{\cal P}_n^{\infty}$
has the structure of {\em Euclidean building}, which is a representative example of non-manifold Hadamard spaces.
We give a quick and informal introduction of this fact.
A point in our building ${\cal B}_n$ is represented by a pair of
a complete flag $\{0\} < {\cal Y}_1 < {\cal Y}_2 < \cdots < {\cal Y}_n = \CC^n$
of vector subspaces and nonincreasing weight
$\lambda \in \RR^n:\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_n$,
where $U \leq U'$ (resp. $U < U'$) means
that vector space $U$ is a (resp. proper) subspace of vector space $U'$.
Then ${\cal B}_n$ consists
of (equivalence classes of) all such pairs,
and is metrized so that for any ordered basis $u_1,u_2,\ldots,u_n$ of $\CC^n$, the map
\begin{equation}\label{eqn:expression}
\RR^n \ni \lambda \mapsto
%\left\{
\begin{array}{c}
\lambda_{i_1} \geq \lambda_{i_2}\geq \cdots \geq \lambda_{i_n}, \\
\langle u_{i_1}\rangle < \langle u_{i_1},u_{i_2} \rangle < \cdots < \langle u_{i_1},u_{i_2},\ldots,u_{i_n} \rangle
\end{array}
%\right.
\end{equation}
is an isometry from $\RR^n$ to ${\cal B}_n$,
where all $i_1,i_2,\ldots,i_n$ are different, and
$\langle u_{i_1},u_{i_2},\ldots u_{i_k} \rangle$ denotes the vector subspace spanned by $u_{i_1},u_{i_2},\ldots,u_{i_k}$.
An {\em apartment} is a subset of ${\cal B}_n$ that is the image of such an isometry.
The asymptotic class of a geodesic ray $t \mapsto g e^{t \diag \lambda}g^{\dagger}$
with nonincreasing $\lambda$ in ${\cal P}_n$
gives rise to such an expression
$\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_n$,
$\langle u_{1} \rangle < \langle u_{1},u_{2} \rangle < \cdots < \langle u_{1},u_{2},\ldots,u_{n}\rangle$,
where $u_i$ is the $i$-th column vector of $g$.
%
This expression is indeed independent of the choice of representatives,
and provides an isometry from $C{\cal P}_n^{\infty}$ to ${\cal B}_n$.
We can identify ${\cal H}_n (= T_I{\cal P}_n \simeq C{\cal P}_n^{\infty})$
with ${\cal B}_n$
by associating $H \in {\cal H}_n$
with geodesic ray $t \mapsto e^{tH} = k e^{t \diag \lambda} k^{\dagger}$ with $k \in U_n$ and $\lambda \in \RR^n$.
We note a link between unitarily invariant convexity
and geodesic convexity.
\begin{Lem}\label{lem:ui-convex=>g-convex}
In this identification ${\cal H}_n = {\cal B}_n$,
any unitarily invariant convex function $f:{\cal H}_n \to \overline{\RR}$
is geodesically convex on ${\cal B}_n$.
\end{Lem}
The proof is given in the appendix, since it only gives insights in this paper.
The building for $C({\cal P}_n^1)^{\infty}$ is obtained from the above expression by
adding $\sum_{i=1}^n \lambda_i = 0$,
where isometry (\ref{eqn:expression}) is from $\{ x\in \RR^n \mid \sum_{i=1}^n x_i = 0\}$.
It is an isometric subspace of ${\cal B}_n$.
\section{$Q$-gradient flow and asymptotic duality}\label{sec:Q-gradient_flow}
Let ${\cal M}$ be a Hadamard manifold, and
let $Q:T^*{\cal M} \to \overline{\RR}$ be a function
on the cotangent bundle of ${\cal M}$.
%
We are going to study the minimum $Q$-gradient-norm problem~(\ref{eqn:minimum_Q-gradient-norm_problem}).
Since the setting is too general, we begin to formulate a meaningful special class of $({\cal M},Q)$.
For $x \in {\cal M}$, let $Q_x := Q|_{T_x^*}$
denote the restriction of $Q$ to $T_x^*$.
%
%
The Legendre-Fenchel conjugate $Q^*:T{\cal M} \to \overline{\RR}$ of $Q$ is defined
so that for each $x \in {\cal M}$, its restriction $(Q^*)_x$ to $T_{x}$ is equal to $Q_x^*$.
We then consider the following properties for $({\cal M},Q)$:
\begin{itemize}
\item[Q1:] For each $x \in {\cal M}$, the restriction $Q_x$ is
a $\RR_{\geq 0}$-valued convex function,
$Q_x^2$ is continuously differentiable,
and the level sets of $Q_x$ are bounded.
\item[Q2:] $Q$ is invariant under parallel transports; $Q(\tau_\gamma^1 p) = Q(p)$
for any $p \in T_x^*$, $x \in {\cal M}$ and curve $\gamma:[0,1] \to {\cal M}$ with $\gamma(0) = x$.
%Then the Legendre-Fenchel conjugate $Q^* = (Q_x^*)_{x \in M} :TM \to \overline{\RR}$ %defined by
%\[
%Q_x^*(u) := \sup_{p \in T_x^*} p(u) - Q_x(p) \quad (u \in T_x = T_x^{**})
%\]
%is also invariant under parallel transports.
%\footnote{}
%\begin{eqnarray*}
%G_{\gamma(t)}^*(\tau_\gamma^t u) &= & \sup_{p \in T_{\gamma(t)}^*} p(\tau_{\gamma}^tu) - G_{\gamma(t)}(p) \quad (u \in T_{\gamma(0)}) \\
%&=& \sup_{p \in T_{\gamma(t)}^*} (\tau_{\gamma}^{-t} p)(u) - G_{\gamma(0)}(\tau_\gamma^{-t}p) \\
%&=& G_{\gamma(0)}^*(u).
%\end{eqnarray*}}
\item[Q3:] For $x,y \in {\cal M}$ and $u \in T_x$, $v \in T_y$, if
$t \mapsto \exp_x tu$ and $t \mapsto \exp_y t v$ are asymptotic, then
$Q_x^*(-u) = Q_y^*(-v)$ holds. Define
$Q_{-}^*:C{\cal M}^{\infty} \to \overline{\RR}$ by
$Q_{-}^{*}(\xi) := Q^*_x(- u)$, where $t \mapsto \exp_{x}tu$ is the representative of $\xi \in C{\cal M}^{\infty}$ at $T_x$.
%\footnote{The negative $-\xi$ of $\xi \in C{\cal M}$ is not defined.
%Even if $u \in T_x$ and $v \in T_y$ are the same point in $C{\cal M}^{\infty}$, $-u$ and $-v$ are not.
%}
%The corresponding function on $CM^{\infty}$ is denoted by $G^*: CM^{\infty} \to \overline{\RR}$.
\end{itemize}
Note that
$Q^2_x$ is also convex in Q1.
%we will consider $Q^2/2:T^*{\cal M} \to \RR$ and its conjugate $(Q^2/2)^*: T{\cal M} \to \overline{\RR}$.
If Q2 holds, then $Q^*$ is also invariant under any parallel transport, since
$Q_{\gamma(1)}^*(\tau_\gamma^1 u) = \sup_{p \in T_{\gamma(1)}^*} p(\tau_{\gamma}^1u) - Q_{\gamma(1)}(p) = \sup_{p \in T_{\gamma(1)}^*} (\tau_{\gamma}^{-1} p)(u) - Q_{\gamma(0)}(\tau_\gamma^{-1}p) = Q_{\gamma(0)}^*(u)$.
%
It is important to note that Q2 can be formulated in terms of the holonomy group:
\begin{itemize}
\item[Q2$'$:] For some {\em $Hol_{x_0}({\cal M})$-invariant function} $Q_0:T_{x_0}^* \to \overline{\RR}$, it holds
\[
Q(p) = Q_0(\tau_{x \to x_0} p) \quad (p \in T_x^*, x \in {\cal M}),
\]
where $Hol_{x_0}({\cal M})$-invariance means that $Q_0$ satisfies
\[
Q_0(k\cdot p) = Q_0(p) \quad (\forall p \in T_{x_0}^*, \forall k \in Hol_{x_0}({\cal M})).
\]
\end{itemize}
It is clear that Q2 implies Q2$'$.
Conversely, If $Q$ satisfies Q2$'$, then,
for $\gamma:[0,1] \to {\cal M}$ with $\gamma(0) = x$ and $\gamma(1) = y$.
we have
\[
Q(\tau_\gamma^{1} p ) = Q_0(\tau_{y \to x_0} \circ \tau_\gamma^{1} p) = Q_0(\tau_{y \to x_0} \circ \tau_\gamma^{1} \circ \tau_{x_0 \to x} \circ \tau_{x \to x_0} p) = Q_0(\tau_{x \to x_0} p) =Q(p).
\]
\begin{Ex}[Euclidean space]
Suppose that ${\cal M} = \RR^n$.
Then $T_x$, $T_x^*$, and $C{\cal M}^{\infty}$ are identified with $\RR^n$,
where all parallel transports are the identity map.
Any convex function
$Q:\RR^n \to \RR_{\geq 0}$ for which it has bounded level sets and $Q^2$ is differentiable
gives rise to $Q:T^*{\cal M} \to \RR_{\geq 0}$ satisfying Q1-3.
\end{Ex}
\begin{Ex}[Dual norm]
The dual norm $Q := \|\cdot\|^*:T^*{\cal M} \to \RR_{\geq 0}$
satisfies Q1-3.
Indeed, the conjugate is given by $Q^*_x(u) = \delta_{B_{\|\cdot\|}(0;1)}$.
The norm $\|\cdot\|$ is invariant under parallel transports,
and gives the speed $\|H\|$ of the geodesic ray $t \mapsto \exp_x t H$ if
$Q^*$ is viewed as a function on $C{\cal M}^{\infty}$.
From this, we see Q2 and Q3.
More generally,
consider any nondecreasing convex function $h:\RR_{\geq 0} \to \RR_{\geq 0}$, and let $Q:= h(\|\cdot\|^*)$. Then it holds $Q^* = h^*(\|\cdot\|)$ for the conjugate $h^*(y) := \sup_{t \geq 0} yt - h(t)$ that is also nondecreasing.
From this, we see Q1-3.
If ${\cal M}$ is a generic Hadamard manifold,
then $Q$ is necessarily written in this way.
\end{Ex}
\begin{Ex}[PD-manifold and unitarily invariant convex function]\label{ex:PD-manifold}
Consider the PD-manifold ${\cal P}_n$.
Choose a unitarily invariant convex function
$Q_0:{\cal H}_n \to \RR_{\geq 0}$ such that $Q_0^2$ is continuously differentiable, and regard it as $T_I^* \to \RR_{\geq 0}$.
Then it is a $Hol_I({\cal P}_n)$-invariant function.
%
Define $Q:T^* {\cal P}_n \to \RR$ by
\[
Q_x(Y) := Q_0(\tau_{x \to I} Y) \quad (Y \in T^*_x{\cal P}_n).
\]
Then, $Q$ satisfies Q1 and Q2.
To see Q3, choose $H \in T_x$ and
suppose that $x^{-1/2}Hx^{-1/2} = u (\diag \lambda) u^{\dagger}$ for $u \in U_n$ and $\lambda \in \RR^n$.
Then $Q_x^*(- H) = Q_0^*(- \diag \lambda)$. On the other hand,
as explained in Section~\ref{subsec:psd},
geodesic $t \mapsto \exp_x t H$
is asymptotic to $t \mapsto \exp_I t k \diag \lambda k^{\dagger}$ for some $k \in U_n$.
Namely, $H \in T_x$ and $k \diag \lambda k^{\dagger} \in T_I$ represent the same point in $C{\cal M}^{\infty}$.
Then $Q^*_x(-H) = Q_0 (- \diag \lambda) = Q^*_I(- k \diag \lambda k^{\dagger})$.
This implies Q3.
%
This construction is generalized to the product of PD-manifolds,
more generally, symmetric spaces of nonpositive curvature.
%
\end{Ex}
Let $f: {\cal M} \to \RR$ be a
twice differentiable geodesically convex function.
Consider the minimum $Q$-gradient-norm problem (\ref{eqn:minimum_Q-gradient-norm_problem})
for ${\cal M},Q,f$.
%\begin{equation}\label{eqn:problem}
%{\rm Min.} \quad Q(df_x) \quad {\rm s.t.} \quad x \in {\cal M}.
%\end{equation}
Our goal is to establish an extension of (\ref{eqn:min_gradient-norm})
to this setting. We start with the weak duality that generalizes \cite[Lemma 2.2]{HiraiSakabe2024FOCS}.
\begin{Lem}[Weak duality]\label{lem:weakduality}
Suppose that $Q$ satisfies {\rm Q3}. Then it holds
\begin{equation}\label{eqn:weakduality}
\inf_{x \in {\cal M}} Q (df_x) \geq \sup_{\xi \in C{\cal M}^{\infty}} - f^{\infty}(\xi) - Q_{-}^*(\xi).
\end{equation}
\end{Lem}
\begin{proof} From the definition of $f^{\infty} = f_x^{\infty}$, it holds
\begin{eqnarray*}
f_x^{\infty}(\xi) &= &\lim_{t \to \infty} \frac{f(\exp_x t\xi) - f(x)}{t} \geq \lim_{t \to 0} \frac{f(\exp_x t\xi) - f(x)}{t} \\
& = & df_x(\xi) = - df_x(- \xi) \geq - Q_x(df_x) - Q_x^*(-\xi),
\end{eqnarray*}
where the first inequality follows from convexity of $f$ and the second inequality
follows from the Fenchel-Young inequality
$Q_x(p) + Q_x^*(- \xi) \geq p(-\xi)$
for $p= df_x \in T_x^*$ and $- \xi \in T_x$.
By Q3, we have the claim.
\end{proof}
%\begin{Question}
%Does the equality (strong duality) hold in (\ref{eqn:weakduality}) ?
%\end{Question}
In the case of Euclidean space, the strong duality holds.
\begin{Ex}
Consider ${\cal M} = \RR^n$ and a lower semicontinuous convex function $Q:\RR^n \to \overline{\RR}$.
Then $C{\cal M}^{\infty}$ is identified with $\RR^n$, and
$f^{\infty}$ is the support function of the closure $C$ of $\{ df_x \mid x \in \RR^n\} \subseteq (\RR^n)^*$
(see e.g.,\cite[Section 3.3]{HiraiSakabe2024FOCS}).
From the Fenchel duality~\cite[Section 31]{Rockafellar}, it holds
%\begin{eqnarray*}
\[
\inf_{x \in \RR^n} Q(df_x) = \inf_{p \in C} Q(p) = \inf_{p \in (\RR^n)^*} Q(p) + \delta_{C}(p) = \sup_{\xi \in \RR^n} - Q^*(- \xi) - f^{\infty}(\xi).
%\end{eqnarray*}
\]
\end{Ex}
%\begin{Ex}
%Consider the case $G=\|\cdot\|^*$; then $G^* = \delta_{B(0;1)}$ (indicator function of unit ball %$B(0;1)$).
%Hirai and Sakabe~\cite{HiraiSakabe2024FOCS} established
%the strong duality by gradient flow,
%which motivates the subsequent argument.
%\end{Ex}
We next introduce a generalization of the gradient flow $x(t)$ that is expected to
attain the equality in (\ref{eqn:weakduality}) by the limit.
From now on, we assume that $({\cal M}, Q)$ satisfies Q1, Q2, and Q3.
\begin{Def}[$Q$-gradient flow]\label{def:Q-gradient}
The {\em $Q$-gradient} $\nabla^{Q} f(x)$ at $x \in {\cal M}$ is defined by
\begin{equation*}
\nabla^{Q} f(x) := \frac{1}{2}d Q^{2}_{df_x}\mid_{T_x^*}\quad (\in T_x^{**} = T_x).
\end{equation*}
The {\em $Q$-gradient flow} is the solution $x(t)$ of the initial value ODE:
\begin{equation}\label{eqn:Q-gradient_flow}
\dot x(t) = - \nabla^{Q} f(x(t)), \quad x(0) = x_0.
\end{equation}
\end{Def}
\begin{Ex}
Suppose that $Q = \|\cdot\|^*$. From $Q(df_x) = \|df_x\|^* = \|\nabla f(x)\|$,
we see that the $Q$-gradient and the ordinary gradient are the same:
$\nabla^{Q} f(x) = \nabla f(x)$.
Accordingly, the $Q$-gradient flow is the ordinary gradient flow of $f$.
%More generally, suppose that $Q$ is a Minkowski norm in Example~\ref{ex:Finsler},
%Then, we see from $Q^2_x/2 = (N_x^2/2)^*$ that $\nabla^Q f(x)$ is equal to
%the gradient of $f(x)$ in this Finsler manifold $({\cal M}, N)$.
\end{Ex}
%\begin{eqnarray*}
%\frac{1}{2} dG^2_{df_x} \mid_{T^*_x} &=& \nabla f(x),\\
%G(df_x) &=& \|\nabla f(x)\|.
%\end{eqnarray*}
% \begin{proof}
% It is clear from convex analysis that $Q_x^* = (\|\cdot \|^*)^* = \delta_{B(0;1)}$.
% %For $v \in T_x \setminus \{0\}$,
% %choose $p^* \in T_x^*$ so that $\|p^*\|^* = 1$ and
% %$p^*v = \|v\|$.
% %Then $G^*(v) = \sup_{p \in T_x^*} vp - h(\|p\|) = \sup_{t\geq 0} t\|v\| - h(t) = h^*(\|v\|)$.
% %
% Define linear isomorphism $\varphi:T_x^* \to T_x$ via $p(v) = \langle \varphi(p),v \rangle$ $(v\in T_x)$.
% Then, $\varphi(df_x) = \nabla f(x)$. From $\|p\|^* = \|\varphi(p)\|$,
% we have $Q(df_x) = \|df_x\|^* = \|\nabla f(x)\|$.
% %From $\|p\|^* = \max_{v\in T_x:\|v\|=1} p(v) = \max_{v\in T_x:\|v\|=1} \langle \varphi(p),v %\rangle = \|\varphi(p)\|$.
% %Then we have $G(df_x) = h(\|df_x\|^*) = h(\|\varphi(df_x)\|) = h(\|\nabla f(x)\|)$, and
% For $p \in T_x^*$, we have
% \begin{eqnarray*}
% (\nabla^{Q}f(x))(p) & =& \frac{d}{dt}\mid_{t=0} \frac{(\|df_x+tp\|^*)^2}{2} = \frac{d}{dt}\mid_{t=0} \frac{\langle \varphi(df_x)+t\varphi(p),\varphi(df_x)+t\varphi(p)\rangle^2}{2} \\
% & = & \langle \varphi(df_x), \varphi(p) \rangle = (\nabla f(x)) (p).
% \end{eqnarray*}
% \end{proof}
Consider the $Q$-gradient flow $x(t)$ in (\ref{eqn:Q-gradient_flow}).
By Q1 and the Cauchy-Peano existence theorem for ODE,
such an $x(t)$ indeed exists and is defined on $[0,T)$ for some $T> 0$.
The following property generalizes the well-known monotonicity property of
$t \mapsto \|\nabla f(x(t))\|$ for the ordinary gradient flow $x(t)$.
\begin{Lem}\label{lem:nonincreasing}
$t \mapsto Q(df_{x(t)})$ is monotonically nonincreasing.
\end{Lem}
\begin{proof}
We show $(d/dt) Q^{2}(df_{x(t)}) \leq 0$. Let $\tau^{-{\delta}}$ denote the parallel transport along the curve $x(\cdot)$ from $t+\delta$ to $t$. Then it holds
\begin{eqnarray*}
&& \frac{d}{d\delta}\mid_{\delta = 0} Q^{2}(df_{x(t+ \delta)}) = \frac{d}{d\delta}\mid_{\delta = 0} Q^2(\tau^{- \delta } df_{x(t+\delta )}) \\
&& = d Q^2_{df_{x(t)}}( \nabla_{\dot x(t)} df) = - 2 \nabla df_{x(t)}(\dot x(t),\dot x(t)) \leq 0,
\end{eqnarray*}
where the first equality follows from Q2,
the second from the chain rule and the definition of covariant derivative,
the third from $d Q_{df_{x(t)}}^2 = - 2\dot x(t)$,
and the last inequality from convexity of~$f$.
\end{proof}
\begin{Lem}\label{lem:upperbound}
For a constant $\alpha \geq 0$ it holds
$
d(x(s),x(t)) \leq \alpha |s-t|.
$
\end{Lem}
\begin{proof}
The velocity of $x(t)$ is given by
\[
\|\dot x(t)\| = \| d(Q^2_{df_{x(t)}}/2)\mid_{T_x^*}\| = \| d(Q^2_{0}/2)_{\tau_{x(t) \to x_0}df_{x(t)}}\|.
\]
By the previous lemma, $t \mapsto Q(df_{x(t)}) = Q_0(\tau_{x(t) \to x_0} df_{x(t)})$ is monotonically nonincreasing.
Therefore, $\tau_{x(t) \to x_0} df_{x(t)}$ belongs to the level set
$\lv(Q_0, Q_0(df_{x_0}))$. Since it is compact by Q1,
the maximum $\alpha$ of $p \mapsto \| d(Q^2_0/2)_p\|$ over $p \in \lv(Q_0, Q_0(df_{x_0}))$ exists.
Then we have
$
\|\dot x(t)\| \leq \alpha$.
By the triangle inequality, we have
$d(x(s), x(t)) \leq \int_{s}^t \|\dot x(r)\|dr \leq \alpha |s-t|.$
\end{proof}
\begin{Lem}\label{lem:completeness}
The vector field $(- \nabla^{Q}f(x))_{x\in {\cal M}}$ is complete.
\end{Lem}
\begin{proof}
It suffices to show that
$\lim_{t \to T} x(t)$ exists in ${\cal M}$ if $T < \infty$.
By the previous lemma,
for every sequence $(t_i)_{i=1,2,\ldots}$ in $[0,T)$ converging to $T$,
the sequence $(x(t_i))_{i=1,2,\ldots}$ in ${\cal M}$ is Cauchy.
By the completeness of ${\cal M}$, the limit of the sequence exists, which must be $\lim_{t \to T} x(t) := x^*$.
Now the $Q$-gradient flow $x(t)$ $(t \in [0,T))$ is connected with the $Q$-gradient flow
$y(t)$ $(t \in [0, T'))$ with $y(0) = x^*$ and $T' > 0$.
%Indeed, by the previous lemma, $[0,T) \in t \mapsto x(t)$ is continuous, and $(x(t))_{t\in[0,T)}$ belongs
%to a closed ball with finite radius $\alpha T$.
%Otherwise, there is a sequence $(t_i)_{i=1,2,\ldots} \subseteq [0,T)$
%such that $t_i \to T$ and $(x(t_i)){i=1,2,\ldots}$ does not converge.
%By completeness of ${\cal M}$, $(x(t_i)){i=1,2,\ldots}$ is not Cauchy.
%However this contradicts the previous lemma $d(x(t_i), x(t_{j})) \leq \alpha |t_i - t_j|$.
% and the curve $x(t) (t \in [0,T))$ belongs to the closed ball $B$ with center $x_0$ and radius $\alpha T$.
% Since $B$ is compact,
% $x(t)$ $(t \in [0,T))$ has accumulation points in $B$.
% Suppose indirectly that distinct accumulation points $y,y'$ exist.
% For $\epsilon := d(y,y')/3 > 0$, consider disjoint balls $B_1 = B(y;\epsilon)$, $B_2 = B(y';\epsilon)$.
% The curve $x(t)$ $(t \in [0,T))$ meets them infinitely many times.
% The distance between $B_1$ and $B_2$ is $\epsilon >0$.
% This implies that the length of curve $x(t)$ $(t \in [0,T))$ must be $\infty$,
% which contradicts $\int_{0}^T \|\dot x(s)\|ds \leq Q_0(df_{x_0}) \alpha T$.
% Thus an accumulation point is unique and must be the limit $\lim_{t \to T} x(t)$.
\end{proof}
Thus, $x(t)$ is defined on $\RR_{\geq 0}$.
Next, we show the energy identity of the $Q$-gradient flow.
\begin{Lem}[Energy identity]\label{lem:energy_identity}
It holds
\begin{equation}\label{eqn:energy_identity}
f(x(t)) - f(x_0) = -\int_{0}^t \frac{1}{2} Q^2(df_{x(s)}) + \left(\frac{1}{2}Q^2\right)^* (- \dot x(s)) ds.
%&=& -\int_{0}^T G^2_x(- df_x) + G_x(-df_x) G_x^* \left(\frac{\dot x}{G_x(-df_x)}\right) dt
\end{equation}
If $Q(df_x) > 0$ for every $x \in {\cal M}$, then
\begin{equation}\label{eqn:energy_identity_modified}
f(x(t)) - f(x_0) + \int_0^{t} Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right) Q(df_{x(s)}) ds = - \int_0^t Q^2(df_{x(s)})ds.
\end{equation}
%\end{eqnarray*}
\end{Lem}
\begin{proof}
If $v =d(Q_x^2/2)_p$ for $p \in T_x^*,v \in T_x$, then by (\ref{eqn:subdifferential}) it holds
\begin{equation}\label{eqn:p(v)}
p(v) = \frac{1}{2}Q_x^2(p) + \left( \frac{1}{2}Q^2_x\right)^*(v).
\end{equation}
Therefore, for the setting $v := - \dot x(t)$ and $p := df_{x(t)}$,
it holds $v = d(Q_x^2/2)_{p}$ by Definition~\ref{def:Q-gradient}.
From (\ref{eqn:p(v)}), we have
\begin{equation*}
\frac{d}{dt}f(x(t)) = df_{x(t)}(\dot x(t)) = -\frac{1}{2} Q^2(df_{x(t)}) - \left(\frac{1}{2}Q^2\right)^* (-\dot x(t)).
\end{equation*}
By integrating it from $0$ to $t$, we have
(\ref{eqn:energy_identity}).
We show the latter part.
If $v = d(Q_x^2/2)_p \in Q_x(p) \cdot \partial (Q_x)(p)$ and $Q_p(p) > 0$,
%\footnote{$Q_x (= \sqrt{Q_x^2})$ may be nondifferentiable at the origin},
then
it holds $v/Q_x(p) \in \partial (Q_x) (p)$, and
$
p(v/Q_x(p)) = Q_x(p) + Q_x^*(v/Q_x(p))$ by (\ref{eqn:subdifferential}),
that is,
\[
p(v) = Q^2_x(p) + Q_x(p) Q_x^{*}\left(\frac{v}{Q_x(p)}\right).
\]
Considering the setting $p := df_x$
and $v := -\dot x(t)$, we have (\ref{eqn:energy_identity_modified}) as above.
\end{proof}
% \begin{Lem}
% If $\displaystyle \inf_{x \in {\cal M}} Q(df_x) > 0$,
% then $\displaystyle \lim_{t \to \infty}f(x(t)) = - \infty$ and $\displaystyle \lim_{t \to \infty} d(x_0,x(t)) = \infty$.
% \end{Lem}
% \begin{proof}
% Note that $(Q^2/2)^{*}$ is nonnegative-valued
% since $(Q^2_x/2)^{*}(u) \geq u(0) - (Q^2/2)(0) = 0$.
% Let $v^* := \inf_{x \in {\cal M}} Q(df_{x(t)}) > 0$,
% By Lemmas~\ref{lem:nonincreasing} and \ref{lem:energy_identity}, it holds
% $f(x(T)) - f(x_0) \leq - v^*T$.
% Thus $\lim_{t \to \infty}f(x(t)) = - \infty$.
% Suppose for contradiction that $d(x_0,x(t))$ has an accumulation point
% $r < \infty$. Then
% $x(t)$ has an accumulation point
% in the closed ball $B$ with center $x_0$ and radius $r$.
% $\inf_{x \in B} f(x) = \lim_{t \to \infty} f(x(t))= -\infty$, contradicting $f: M \to \RR$.
% \end{proof}
% We note that if $Q(0) = 0$,
% then the $Q$-gradient flow also does not increase $f$,
% though we do not use this fact.
% \begin{Lem}
% If $Q_x(0) = 0$ for $x \in {\cal M}$,
% then $t \mapsto f(x(t))$ is nonincreasing.
% %In addition, if $\inf_{x \in {\cal M}}Q(df_x) > 0$,
% %then $\lim_{t \to \infty} f(x(t)) = -\infty$.
% \end{Lem}
% \begin{proof}
% This follows from energy identity and nonnegativity of $(Q^2/2)^*$
% by $(Q_x^2/2)^*(v) \geq v(0) - (Q_x/2)(0) = 0$.
% \end{proof}
For $t \in \RR_{\geq 0}$, define $R(t) \in \RR_{\geq 0}$ and $u(t) \in T_{x_0}$ by
\begin{equation*}
R(t) := \int_0^t Q(df_{x(s)})ds, \quad x(t) = \exp_{x_0} R(t) u(t).
\end{equation*}
\begin{Lem}
Suppose that $\inf_{x\in {\cal M}} Q(df_{x}) > 0$. Then the following hold:
\begin{itemize}
\item[(1)] $(u(t))_{t \in \RR_{\geq 0}}$ is bounded in $T_{x_0}$.
\item[(2)] For $t \in \RR_{>0}$ and $r \in [0, R(t)]$, it holds
\begin{equation}\label{eqn:important}
\frac{f(\exp_{x_0} r u(t)) - f(x_0)}{r} + \frac{1}{R(t)}
\int_0^{t} Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right) Q(df_{x(s)}) ds
\leq - \lim_{t \to \infty} Q(df_{x(t)}).
\end{equation}
\end{itemize}
\end{Lem}
The essence of the inequality (\ref{eqn:important}) is given in the proof of \cite[Theorem 3.1]{HiraiSakabe2024FOCS}.
\begin{proof}
(1).
By $d(x_0,x(t)) = \|R(t) u(t)\| = R(t)\|u(t)\|$ and
Lemmas~\ref{lem:nonincreasing} and \ref{lem:upperbound}, it holds
\[
\|u(t)\| = \frac{d(x_0,x(t))}{R(t)} = \frac{1}{\frac{1}{t}\int_{0}^t Q(df_{x(s)})ds}\frac{d(x_0,x(t))}{t}
\leq \frac{\alpha}{\lim_{t \to \infty} Q(df_{x(t)})} < \infty.
\]
%Thus $u(t)$ is bounded.
(2). Consider the geodesic $[0,R(t)] \ni r \mapsto \exp_{x_0} r u(t)$ from $x_0$ to $x(t)$.
By convexity of $f$ along this geodesic, it holds
\begin{equation*}
f(\exp_{x_0} r u(t)) - f(x_0) \leq \frac{r}{R(t)}(f(x(t))- f(x_{0})) \quad (r \in [0,R(t)]).
\end{equation*}
By the energy identity (\ref{eqn:energy_identity_modified}), we have
\begin{equation}\label{eqn:important0}
\frac{f(\exp_{x_0} r u(t)) - f(x_0)}{r} \leq \frac{-1}{R(t)} \left(
\int_0^{t} Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right) Q(df_x(s)) ds + \int_0^t Q^2(df_{x(s)})ds
\right).
\end{equation}
By the Cauchy-Schwarz inequality and Lemma~\ref{lem:nonincreasing}, it holds
\[
- \frac{\int_0^t Q^2(df_{x(s)})ds}{\int_0^t Q(df_{x(s)})ds} \leq -\frac{1}{t}\int_0^t Q(df_{x(s)})ds \leq - \lim_{t \to \infty} Q(df_{x(t)}).
\]
By substituting it for the second term of the RHS in (\ref{eqn:important0}),
we have (\ref{eqn:important}).
\end{proof}
The main result in this section is an extension of (\ref{eqn:min_gradient-norm})
to basic cases of $({\cal M}, Q)$:
\begin{Thm}\label{thm:main}
Assume that one of the following holds:
\begin{itemize}
\item[(i)] ${\cal M} = \RR^n$.
\item[(ii)] ${\cal M}$ is generic, or more generally, $Q$ is written as $Q = h(\|\cdot \|^*)$.
\item[(iii)] ${\cal M} = {\cal P}^1_n$.
\item[(iv)] ${\cal M} = \RR^n \times {\cal P}^1_{n_1} \times \cdots \times {\cal P}^1_{n_d}$.
\end{itemize}
Suppose $\inf_{x \in {\cal M}} Q(df_x) >0$. Then it holds
\begin{equation}\label{eqn:main}
\lim_{t \to \infty} Q(df_{x(t)}) = \inf_{x \in {\cal M}} Q(df_x) = \sup_{\xi \in C{\cal M}^{\infty}} - f^{\infty}(\xi) - Q^*_{-}(\xi)= - f^{\infty}(u^*) - Q^*_{-}(u^*),
\end{equation}
where $u^*$ is any accumulation point of $(u(t))_{t \in \RR_{\geq 0}}$.
%If the minimizer $\xi^*$
%of $\xi \mapsto f^{\infty}(\xi)+ Q^*_{-}(\xi)$
%is unique, then $u(t)$ converges to $\xi^*$.
\end{Thm}
The proof follows the approach developed in \cite{Hirai2025,HiraiSakabe2024FOCS},
which is a combination of the weak duality (Lemma~\ref{lem:weakduality}) and the inequality (\ref{eqn:important}).
In the case of $Q = \|\cdot\|^*$ (with $Q^*_{-} = \delta_{B_{\|\cdot\|}(0,1)}$),
the second term of the LHS of (\ref{eqn:important}) vanishes, and
$u(t)$ belongs to $B_{\|\cdot\|}(0,1)$ by the triangle inequality $d(x_0,x(t)) \leq R(t)$.
From $t \to \infty$ and $r \to \infty$, we obtain $u^* \in B_{\|\cdot\|}(0,1)$
with the desired inequality $f^{\infty}(u^*) \leq - \inf_{x \in {\cal M}} Q(df_x)$
(the reverse of the weak duality (Lemma~\ref{lem:weakduality})).
For the general case, however,
the second term in (\ref{eqn:important}) makes the argument more involved, and
we were only able to prove the above limited cases.
%
Fortunately, the cases (iii) and (iv) are expected to have enough applications,
as we will see in Section~\ref{sec:entanglement_polytope}.
\begin{proof}[Proof of Theorem~\ref{thm:main}]
% Case (i). $Q_x^* = \delta_{C_x}$ for a closed convex set $C_x$ in $T_x$.
% Necessarily $- \dot x(s)/Q(df_{x(s)}) \in C_x$ and
% the second term of LHS of (\ref{eqn:important}) is zero.
Case (i). From $x(t) -x_0 = R(t)u(t)$, we have
\[
u(t) = \frac{1}{R(t)} \int_{0}^t \dot x(s) ds.
\]
Also, by convexity (Jensen's inequality) of $Q^*$, we have
\[
\frac{1}{R(t)}
\int_0^{t} Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right) Q(df_{x(s)}) ds
\geq Q^* \left( \frac{1}{R(t)}\int_{0}^t - \dot x(s) ds \right) = Q^*(-u(t)).
\]
Therefore, from (\ref{eqn:important}), we have
\begin{equation}\label{eqn:becomes}
\frac{f(\exp_{x_0} r u(t)) - f(x_0)}{r} + Q^*(-u(t))
\leq - \lim_{t \to \infty} Q(df_{x(t)}).
\end{equation}
Choose any subsequence $(t_i)_{i=1,2,\ldots}$ such that
$t_i \to \infty$ and $u(t_i)$ converges to some $u^* \in T_{x_0}$.
Since $\inf_{x \in {\cal M}} Q(df_x) >0$, it holds $R(t_i) \to \infty$.
By (\ref{eqn:becomes}), for any $r \in \RR_{\geq 0}$ it holds
\[
\frac{f(\exp_{x_0} r u^*) - f(x_0)}{r} + Q^*(-u^*)
\leq - \lim_{t \to \infty} Q(df_{x(t)}).
\]
For $r \to \infty$, we obtain
\begin{equation}
f^{\infty}(u^*) + Q^*(-u^*) \leq - \lim_{t \to \infty}Q(df_{x(t)}).
\end{equation}
Combining it with the weak duality (Lemma~\ref{lem:weakduality}), we have
\begin{eqnarray*}
\lim_{t \to \infty}Q(df_{x(t)}) &\leq& - f^{\infty}(u^*) - Q^* (-u^*) \\
& \leq& \sup_{\xi \in C{\cal M}^{\infty}} - f^{\infty}(\xi) - Q^*_{-}(\xi)
\leq \inf_{x \in {\cal M}} Q(df_x) \leq \lim_{t \to \infty}Q(df_{x(t)}).
\end{eqnarray*}
Thus, the equality (\ref{eqn:main}) holds.
%Further, if the minimizer $\xi^*$ is unique, then $u^*$ must be $\xi^*$.
%\end{proof}
% We study a more detailed expression of (\ref{eqn:u(t)}).
% \begin{Lem}[{See e.g.,\cite[p. 35]{Sakai1996}}]
% For $u \in S_{x_0}$, $t \in \RR_{++}$, and $H \in T_{tu}T_{x_0} = T_{x_0}$, it holds
% \[
% d (\exp_{x_0})_{t u} H = \frac{X(t)}{t},
% \]
% where $X(t)$ is the Jacobi field along geodesic $t \mapsto \exp_{x_0}tu$
% with $X(0) = 0$ and $\nabla X(0) = H$.
% \end{Lem}
%\begin{proof}
Case (ii). Let $H(t) := R(t)u(t)$.
Then
\[
u(t) = \frac{1}{R(t)}\int_{0}^t \dot H(s) ds.
\]
By
$x(s) = \exp_{x_0} H(s)$, it holds
\[
\dot x(s) = d(\exp_{x_0})_{H(s)} \dot H(s).
\]
It is well-known that the differential of the exponential map $\exp_{x_0}$
is given by a Jacobi field; see e.g.,~\cite[p. 35]{Sakai1996}.
Specifically, it holds
\[
d (\exp_{x_0})_{r H(s)} \dot H(s) = \frac{J_s(r)}{r},
\]
for the Jacobi field $J_s(r)$ along geodesic $r \mapsto \gamma(r) = \exp_{x_0}r H(s)$
with $J_s(0) = 0$ and $\nabla_{\dot \gamma(0)} J_s = \dot H(s)$.
Therefore,
$\dot x(s) = J_s(1).$
By the Rauch comparison theorem~\cite[Theorem IV. 2.3]{Sakai1996} between $\RR^n$ and ${\cal M}$, it holds
$
\|J_s(r)\| \geq \|r \dot H(s)\|
$, and hence
\[
\|\dot x(s)\| \geq \|\dot H(s)\|.
\]
Since $Q^*$ is written as $Q^* = h^*(\|\cdot\|)$
for a nondecreasing convex function $h^*: \RR_{\geq 0} \to \RR$, we have
\begin{eqnarray*}
&& \frac{1}{R(t)}
\int_0^{t} Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right) Q(df_{x(s)}) ds
\geq \frac{1}{R(t)}
\int_0^{t} Q^* \left( -\frac{\dot H(s)}{Q(df_{x(s)})} \right) Q(df_{x(s)}) ds \\
&& \geq Q^* \left( \frac{1}{R(t)}\int_{0}^t - \dot H(s) ds \right) = Q^*(- u(t)),
\end{eqnarray*}
where the second inequality is Jensen's inequality as above.
Therefore, we have (\ref{eqn:becomes}).
The rest is the same as in the case (i).
Case (iii). Since
the $SL_n(\CC)$-action on ${\cal P}_n^1$ is transitive and naturally extends to $C({\cal P}_n^1)^{\infty}$,
we can assume that $x_0 = I$.
Specifically, choose $g\in SL_n(\CC)$ with $gg^{\dagger}=x_0$,
and consider $\tilde f(x) := f(g x g^{\dagger})$
and $\tilde Q(Y) := Q((g^{\dagger})^{-1} Y g^{-1})$ instead of $f,Q$.
%
As in (ii), let $H(t) := R(t)u(t)$. Then $x(t) = e^{H(t)}$ and
\[
u(t) = \frac{1}{R(t)} \int_{0}^t \dot H(s) ds.
\]
\begin{Clm}\label{clm:show}
There are $k_1,k_2,\ldots,k_m \in SU_n$ and $p_1,p_2,\ldots,p_m \in \RR_{>0}$
with $\sum_{j=1}^m p_j =1$ such that
\begin{equation*}
\dot H(s) = \sum_{j=1}^m p_j k_j^{\dagger} (\tau_{x(s) \to x_0} \dot x(s)) k_j.
\end{equation*}
%
%\begin{equation}\label{eqn:show}
%Q^* \left( -\frac{x(s)^{-1/2} \dot x(s) x(s)^{-1/2}}{Q(df_{x(s)})} \right) \geq Q^*\left( -%\frac{\dot H(s)}{Q(df_{x(s)})} \right).
%\end{equation}
\end{Clm}
We show this claim later.
By this claim, we obtain
\begin{equation}\label{eqn:IEMI}
Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right) \geq Q^*\left( -\frac{\dot H(s)}{Q(df_{x(s)})} \right)
\end{equation}
since
\begin{eqnarray}
&& Q^*\left( -\frac{\dot H(s)}{Q(df_{x(s)})} \right) =
Q^*\left( \sum_{j=1}^m - p_jk_j^{\dagger}\frac{\tau_{x(s) \to x_0} \dot x(s)}{Q(df_{x(s)})}k_j\right) \leq
\sum_{j=1}^m p_j Q^*\left( - k_j^{\dagger}\frac{\tau_{x(s) \to x_0} \dot x(s)}{Q(df_{x(s)})}k_j\right) \nonumber
\\
&& = \sum_{j=1}^m p_j Q^*\left( - \frac{\tau_{x(s) \to x_0} \dot x(s)}{Q(df_{x(s)})}\right)
= Q^* \left( -\frac{\tau_{x(s) \to x_0} \dot x(s)}{Q(df_{x(s)})} \right) = Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right), \label{eqn:as_in}
\end{eqnarray}
where the inequality follows from convexity and the next equality follows from unitary invariance.
Thus we have
\begin{eqnarray*}
&& \frac{1}{R(t)}
\int_0^{t} Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right) Q(df_{x(s)}) ds
\geq \frac{1}{R(t)}
\int_0^{t} Q^* \left( -\frac{\dot H(s)}{Q(df_{x(s)})} \right) Q(df_{x(s)}) ds \\
&& \geq Q^* \left( \frac{1}{R(t)}\int_{0}^t - \dot H(s) ds \right) = Q^*(- u(t)),
\end{eqnarray*}
where the second inequality is Jensen's inequality as before.
Then we have (\ref{eqn:becomes}), and complete the proof as before.
Case (iv). Suppose ${\cal M} =\RR^n \times {\cal P}_{n_1}^1 \times {\cal P}_{n_2}^1 \times \cdots \times {\cal P}_{n_d}^1$.
We write points $x \in {\cal M}$ as $x = (y, z_1,z_2,\ldots, z_d)$, where $y \in \RR^n$ and $z_i \in {\cal P}_{n_i}^1$.
Also $T_x{\cal M} = \RR^n \oplus T_{z_1}{\cal P}_{n_1}^1 \oplus \cdots \oplus T_{z_d}{\cal P}_{n_d}^1$.
We can assume that $x_0 := (0,I,I,\ldots,I)$ as in case (iii).
The holonomy group $Hol_I({\cal M})$ is given by
$SU_{n_1} \times SU_{n_2} \times \cdots \times SU_{n_d}$
acting $T_{x_0}M$ as $(k_1,k_2,\ldots,k_d) \mapsto (v,k_1 H_1 k_1^{\dagger}, k_2H_2k_2^{\dagger},\ldots, k_dH_dk_d^{\dagger})$ for $(v,H_1,H_2,\ldots,H_d) \in T_x{\cal M}$.
Represent $x(t)$ as $x(t) = \exp_{x_0} H(t)$ as above.
Then $\exp_{x_0} H(t) = (y(t), e^{H_1(t)},\ldots, e^{H_d(t)})$
for $H_i(t) \in TP_{n_i}^1$.
For each $\dot H_i(s)$, Claim~\ref{clm:show} holds.
Thus, as in (\ref{eqn:as_in}), we obtain~(\ref{eqn:becomes}):
\begin{eqnarray*}
&& Q^*\left( -\frac{\dot H(s)}{Q(df_{x(s)})} \right)
= Q^*\left( -\frac{(\dot y(s), \dot H_1(s), \dot H_2(s),\ldots,\dot H_d(s))} {Q(df_{x(s)})} \right) \\
&& \leq Q^*\left( -\frac{(\dot y(s), \tau_{z_1(s) \to I}\dot z_1(s), \dot H_2(s),\ldots,\dot H_d (s))} {Q(df_{x(s)})} \right) \\
&& \leq \cdots \leq Q^*\left( -\frac{(\dot y(s), \tau_{z_1(s) \to I}\dot z_1(s), \tau_{z_2(s) \to I}\dot z_2(s),\ldots, \tau_{z_d(s) \to I}\dot z_d(s))} {Q(df_{x(s)})} \right) \\
&& = Q^* \left( -\frac{\tau_{x(s) \to x_0} \dot x(s)}{Q(df_{x(s)})} \right) = Q^* \left( -\frac{\dot x(s)}{Q(df_{x(s)})} \right).
\end{eqnarray*}
The rest is the same as above.
\end{proof}
We are going to prove Claim~\ref{clm:show}.
We first provide an explicit formula of $\dot H(s)$.
For $h \in \RR^n$, let $S^h = (S^h_{ij})$ be the $n \times n$ matrix defined by
\begin{equation*}
S^h_{ij} := \frac{(h_i-h_j)/2}{\sinh ((h_i-h_j)/2)} \quad (1\leq i,j \leq n).
\end{equation*}
For two matrices $A =(a_{ij})$, $B=(b_{ij})$ of the same size,
the {\em Hadamard product (Schur product)} $A \odot B$ is defined by
$(A \odot B)_{ij} := a_{ij}b_{ij}$.
\begin{Lem}[{See \cite[Part II, Lemma 10.40]{BridsonHaefliger1999}}]\label{lem:dotH(s)}
Suppose that $x(s) = e^{H(s)}$, where
$H(s) = k \diag h k^\dagger$ for $u \in SU_n$ and $h \in \RR^n$.
Then it holds
\begin{equation}\label{eqn:dotH(s)}
\dot H(s) = k (S^h \odot (k^\dagger \tau_{x(s) \to x_0} \dot x(s) k)) k^{\dagger}.
\end{equation}
\end{Lem}
\begin{proof}
Recall the parallel transport on ${\cal P}_n$ is given by
\[
\tau_{x(s) \to x_0} \dot x(s) = x(s)^{-1/2} \dot x(s) x(s)^{-1/2}.
\]
Then, \cite[Part II, Lemma 10.40]{BridsonHaefliger1999} says that
\[
x(s)^{-1/2} \dot x(s) x(s)^{-1/2} = \frac{\sinh (\ad H(s)/2)}{\ad H(s)/2} \dot H(s),
\]
where $(\ad H)(X) = [H,X] = HX - XH$ and $\frac{\sinh \ad H}{\ad H} (X):= \sum_{i=0}^\infty (-1)^k \frac{(\ad H)^{2k} (X)}{(2k+1)!}$.
This formula can also be deduced by Jacobi field calculation
with curvature tensor $R(X,Y)Z = - [X,Y]Z/4$ at $T_I {\cal P}_n$.
From $(\ad H(s)) (\dot H(s)) = k ((\ad \diag h) (k^{\dagger}\dot H(s) k))k^\dagger$
and $(\ad \diag h) E_{ij} = (h_i- h_j)E_{ij}$, where $E_{ij}$ is the matrix unit,
we obtain
\[
(k^{\dagger}\tau_{x(s) \to x_0} \dot x(s)k)_{ij}
= \frac{\sinh ((h_i-h_j)/2)}{(h_i-h_j)/2} (k^{\dagger} \dot H(s) k)_{ij},
\]
which implies (\ref{eqn:dotH(s)}).
\end{proof}
We combine the above formula (\ref{eqn:dotH(s)}) with the methods
of {\em matrix analysis}~\cite{Bhatia_PositiveDefiniteMatrices}.
Recall that a linear operator $\Psi:\CC^{n\times n} \to \CC^{n \times n}$ is said to be {\em doubly stochastic} if
\begin{itemize}
\item[(1)] $\Psi$ is {\em positive}, i.e., $\Psi(A) \in {\cal H}_n^+$ whenever $A \in {\cal H}_n^+$.
\item[(2)] $\Psi$ is {\em unital}, i.e., $\Psi(I) = I$.
\item[(3)] $\Psi^*$ is also unital, i.e., $\Psi^*(I) = I$,
\end{itemize}
where $\Psi^*$ is defined by $\trace \Psi(A)^{\dagger}B = \trace A^{\dagger} \Psi^*(B)$.
\begin{Lem}\label{lem:doublystochastic1}
For $k \in U_n$ and $h \in \RR^n$,
the linear operator $\Psi$ defined by $A \mapsto k (S^h \odot (k^{\dagger} Ak)) k^{\dagger}$
is doubly stochastic.
\end{Lem}
\begin{proof}
(1). It is clear that $A \mapsto u A u^{\dagger}$ for any $u \in U_n$ is positive.
Since $x \mapsto x/\sinh x$ is a positive definite function~\cite[5.2.9]{Bhatia_PositiveDefiniteMatrices},
it holds $S(h) \in {\cal H}_n^+$.
Therefore $A \mapsto S^h \odot A$
is positive~\cite[Example 2.2.1 (ix)]{Bhatia_PositiveDefiniteMatrices}.
Consequently, the composition $\Psi$ is also positive.
(2) follows from $S^h \odot I = I$.
(3). This follows from the observation that $\Psi^*$ is given by $B \mapsto k^{\dagger} (S^h \odot (k Bk^{\dagger})) k$; see \cite[Exercise 2.7.11 (ii)]{Bhatia_PositiveDefiniteMatrices}.
\end{proof}
Now Claim~\ref{clm:show} is obtained from the next lemma in the setting of
$A = \dot H(s)$, $B = \tau_{x(s) \to x_0} \dot x(s)$, $\Psi = k (S^h\odot (k^\dagger (\cdot) k)) k^{\dagger}$.
%
\begin{Lem}[{\cite[Exercise 2.7.13]{Bhatia_PositiveDefiniteMatrices}}]\label{lem:doublystochastic2}
Let $A,B \in {\cal H}_n$. Suppose that $A = \Psi(B)$ for some doubly stochastic operator $\Psi$.
Then there are $k_1,k_2,\ldots,k_m \in U_n$
and $p_1,p_2,\ldots,p_m \in \RR_{>0}$ with $\sum_{j=1}^mp_j = 1$ such that
\[
A = \sum_{j=1}^m p_j k_j^{\dagger} B k_j.
\]
\end{Lem}
\begin{proof}
Suppose that $A = u \diag \lambda u^{\dagger}$ and $B = k \diag \mu k^{\dagger}$
for $u,k \in U_n$ and $\lambda,\mu \in \RR^n$.
Define a linear map $\varphi:\RR^n \to \RR^n$
by $\varphi(z)_i:= (u^{\dagger} \Psi(k \diag z k^{\dagger}) u)_{ii}$.
Since $\Phi$ is positive,
$\varphi$ maps $\RR^n_{\geq 0}$ to $\RR^n_{\geq 0}$.
Thus $\varphi$ is written as $\varphi(z) = M z$ for a nonnegative matrix $M$.
Since $\Psi$ and $\Psi^*$ are unital, $M$ is a doubly stochastic matrix.
By the Birkhoff-von Neumann theorem, $M$ is a convex combination $\sum_{j=1}^m p_j \sigma_j$ for permutation matrices $\sigma_1,\sigma_2,\ldots,\sigma_m$.
Thus $\diag \varphi(z) = \sum_{j=1}^m p_j \sigma_j (\diag z) \sigma_j^{\dagger}$.
By the assumption $A = \Psi(B)$, it holds $\varphi(\mu) = \lambda$, and
\[
A = u (\diag \varphi(\mu)) u^{\dagger} =
\sum_{j=1}^m p_j u \sigma_j (\diag \mu) \sigma_j^{\dagger}u^{\dagger} = \sum_{j=1}^m p_j
(u\sigma_j k^{\dagger}) B (k \sigma_j^{\dagger} u^{\dagger}). \qedhere
\]
\end{proof}
\begin{Conj}\label{conj}
The conclusion of Theorem~\ref{thm:main} (and subsequent Theorem~\ref{thm:main2})
holds for any Hadamard manifold.
\end{Conj}
\begin{Rem}
%We conjecture that Theorem~\ref{thm:main} holds for any Hadamard manifolds.
For proving this conjecture,
it suffices to show that
\begin{equation}\label{eqn:ggIEMI}
F(\tau_{\exp_{x_0}H \to x_0} d(\exp_{x_0})_{H} K) \geq F(K) \quad (H,K \in T_{x_0})
\end{equation}
holds for any $Hol_{x_0}({\cal M})$-invariant convex function $F$ on $T_{x_0}$.
Indeed, it implies (\ref{eqn:IEMI}) in the setting of $F = Q^*_0$,
$H = H(s)$, and $K=-\dot H(s)/Q(df_{x(s)})$.
The property (\ref{eqn:ggIEMI}) may be interesting in its own right.
For the ${\cal P}_n^1$ case above, we have shown in (\ref{eqn:as_in}) that
\begin{equation*}
F ( e^{-H/2} (d (e^{(\cdot)})_H K) e^{-H/2} ) \geq F(H) \quad (H,K \in {\cal H}_n).
\end{equation*}
holds for any unitarily invariant convex function
$F:{\cal H}_n \to \RR$.
This property is viewed as a convex function version of
the {\em generalized infinitesimal exponential metric
increasing property (generalized IEMI)}~\cite[Proposition 6.4.1]{Bhatia_PositiveDefiniteMatrices}.
\end{Rem}
Next, by using Theorem~\ref{thm:main},
we show that the strong duality itself holds
for a more general class of $Q$.
Relaxing Q1, consider
\begin{itemize}
\item[$\tilde {\rm Q}$1:] For each $x \in {\cal M}$,
the restriction $Q_x :T_x^* \to \overline{\RR}$ is
a lower-semicontinuous convex function having bounded level sets.
%such that
%$Q_x(p)= 0$ if and only if $p=0$.
\end{itemize}
Now suppose that $Q$ satisfies $\tilde {\rm Q}$1, Q2, and Q3,
and that $f$ satisfies
\begin{equation*}
df_x \in \dom Q_x \quad (x \in {\cal M}).
\end{equation*}
\begin{Thm}\label{thm:main2} Under this setting,
assume that one of the conditions (i)-(iv) in Theorem~\ref{thm:main} holds. Then, it holds
\begin{equation}\label{eqn:strong_duality}
\inf_{x \in {\cal M}} Q(df_x) = \sup_{\xi \in C{\cal M}^{\infty}} - f^{\infty}(\xi) - Q^*_{-}(\xi).
\end{equation}
\end{Thm}
\begin{proof}
We can assume that $Q$ is $\RR_{\geq 0}$-valued and $\inf_{x \in {\cal M}}Q(df_x)>0$
(by adding a positive constant to $Q$).
%
%The weak duality (Lemma~\ref{lem:weakduality}) holds in this setting:
%\[
%\inf_{x \in {\cal M}} Q_x(df_x) \geq \sup_{\xi \in C{\cal M}^{\infty}} - f^{\infty}(\xi) - Q^*(\xi).
%\]
We utilize the {\em Moreau envelope} of $Q$; see e.g., \cite[Chapter 1.G]{RockafellarWets}.
Specifically, for $\lambda > 0$, define Moreau envelope $e_{\lambda} Q:T^*{\cal M} \to \RR$ by
\[
(e_\lambda Q)_x(p) := \inf_{q \in T_x^*} Q_x(q) + \frac{1}{2\lambda} (\|p-q\|^*)^2 \quad (p \in T_x^*, x\in {\cal M}).
\]
Then $(e_\lambda Q)_x$ is continuously differentiable convex (\cite[Theorem 2.26]{RockafellarWets}), and satisfies Q1.
The conjugate $(e_\lambda Q)^*$ is given by
\[
(e_\lambda Q)^*_x = Q_x^* + \frac{\lambda}{2} \|\cdot\|^2.
\]
See \cite[Example 11.26]{RockafellarWets}.
From this, we see that $e_\lambda Q (= (e_\lambda Q)^{**})$ satisfies Q2 and Q3.
By applying Theorem~\ref{thm:main} to $e_\lambda Q$, we have
\begin{equation}\label{eqn:strong_duality_e_lambdaQ}
\alpha_{\lambda} := \inf_{x\in {\cal M}} e_{\lambda}Q(df_{x}) = \sup_{\xi \in C{\cal M}^{\infty}} - f^{\infty}(\xi) - Q^*_{-}(\xi) - \frac{\lambda}{2} \|\xi\|^2 \quad (\lambda > 0).
\end{equation}
From $Q \geq e_{\lambda}Q \geq e_{\lambda'}Q$ if $0 < \lambda \leq \lambda'$,
we see that $\lambda \mapsto \alpha_{\lambda}$ is nonincreasing.
Since it has an upper bound $\alpha^* := \inf_{x \in {\cal M}} Q(df_{x})$,
the limit $\alpha_0 := \lim_{\lambda \to +0} \alpha_\lambda (= \sup_{\lambda >0} \alpha_\lambda)$ exists.
We show
\[
\alpha_0 = \alpha^*
\]
Define $D_f \subseteq T_{x_0}^*$ by
\[
D_f := \{ \tau_{x \to x_0} df_{x} \mid x \in {\cal M} \}.
\]
Then
\[
\alpha_{\lambda} = \inf_{p \in D_f} e_{\lambda}Q(p).
\]
Since $e_{\lambda}Q$ has compact level sets,
the infimum is attained by some point in $\overline{D_f}$ (closure of $D_f$).
Since $\lv(e_{\lambda}Q,r) \subseteq \lv(e_{\lambda'}Q,r)$ for $\lambda \leq \lambda'$ and $r \geq 0$,
we can choose convergent sequences $(\lambda_i)_{i=1,2,\ldots} \subseteq \RR_{>0}$ and
$(p_i)_{i=1,2,\ldots} \subseteq \overline{D_f}$ such that
$\lambda_i \to 0$,
$e_{\lambda_i} Q(p_i) = \alpha_{\lambda_i}$, and $p_i \to p^*\in \overline{D_f}$.
%We show that $Q(p^*) = \alpha_0$.
%which implies $\alpha_0 = \alpha^*$
%(since $Q(p) = \alpha_0 \leq \alpha^* \leq Q(p)$)
Here $(\lambda,p) \mapsto e_{\lambda} Q(p)$ is lower semicontinuous,
since it is written as
the supremum of affine functions
$(p,\lambda) \mapsto p(v) - Q^{*}(v)- (\lambda/2)\|v\|^2$ over $v \in T_x$.
Therefore
$\alpha_0 = \lim_{i \to \infty}e_{\lambda_i} Q(p_i) \geq Q(p^*) \geq \inf_{p \in D_f} Q(p) = \alpha^* \geq \alpha_0$,
which implies $\alpha_0 = \alpha^*$.
By the weak duality (Lemma~\ref{lem:weakduality}) and (\ref{eqn:strong_duality_e_lambdaQ}), we have
\begin{eqnarray*}
&& \inf_{x \in {\cal M}} Q(df_x) \geq \sup_{\xi \in C{\cal M}^{\infty}} - f^{\infty}(\xi) - Q^*_{-}(\xi) \geq \sup_{\xi \in C{\cal M}^{\infty}} - f^{\infty}(\xi) - Q^*_{-}(\xi) - \frac{\lambda_i}{2} \|\xi\|^2 \\
&& = \inf_{x\in {\cal M}} e_{\lambda_i}Q(df_{x}) = \alpha_{\lambda_i} \quad \underset{i \to \infty}{\longrightarrow} \quad \alpha_0 = \alpha^* = \inf_{x \in {\cal M}} Q(df_x).
\end{eqnarray*}
Thus, the strong duality holds.
\end{proof}
We provide a geodesic convexity interpretation of the RHS in (\ref{eqn:strong_duality}).
\begin{Prop}\label{prop:g-convexity}
Assume that one of the conditions (i)-(iv) in Theorem~\ref{thm:main} holds.
Then, $f^{\infty} + Q_{-}^*$ is geodesically convex on $C{\cal M}^{\infty}$.
\end{Prop}
\begin{proof}
The case (i) is obvious.
It is shown \cite{Hirai_Hadamard2022}
that the recession function $f^{\infty}$ is geodesically convex on $C{\cal M}^{\infty}$.
Then the cases (iii) and (iv) follow from Lemma~\ref{lem:ui-convex=>g-convex}.
Consider the case (ii).
In Hadamard space $C{\cal M}^{\infty}$,
the norm $\|\cdot\|$ represents the distance from $0$
and is geodesically convex~\cite[II.2.5]{BridsonHaefliger1999}.
With the fact that $h^*$ is nondecreasing convex,
we see geodesic convexity of $Q_{-}^* = h^{*}(\|\cdot\|)$.
\end{proof}
\paragraph{$Q$-subgradient method.}
Here we consider to solve (\ref{eqn:min_gradient-norm}) by discretization of the $Q$-gradient flow.
A natural discretization scheme of the $Q$-gradient flow is the following:
\begin{equation}\label{eqn:Q-subgradient_descent}
x_{i+1} := \exp_{x_i} - \delta_i Z_i,\quad Z_i \in \partial (Q_{x_i}^2/2)(df_{x_i}) \quad (i=0,1,2,\ldots),
\end{equation}
where $\delta_i >0$ is a step size.
By $\partial Q_{x_i}(df_{x_i}) = \tau_{I \to x_i}\partial Q_0(\tau_{x_i \to I}df_{x_i})$,
the $Q$-subgradient method is implemented by subgradient computation of $Q_0$.
\begin{Question}
Under which conditions of $f,Q,\delta_i$ does the sequence $Q(df_{x_i})$ converge to $\inf_{x\in {\cal M}}Q(df_{x})$ ?
\end{Question}
Hirai and Sakabe~\cite{HiraiSakabe2024FOCS} showed that
the required convergence holds if $f$ is $L$-smooth convex, $Q = \|\cdot\|^*$, and $\delta_i=1/L$.
This important problem goes beyond the scope of this paper
and has to be left to subsequent research.
\section{Convex optimization on entanglement polytopes}\label{sec:entanglement_polytope}
In this section, we apply the results in Section~\ref{sec:Q-gradient_flow}
to the Kempf-Ness optimization for GL-action on tensors.
% which is a geodesically convex function on the product of PD-manifolds.
% The corresponding minimum $Q$-gradient-norm problem (\ref{eqn:minimum_Q-gradient-norm_problem})
% is then convex optimization
% on a class of moment polytopes, known as {\em entanglement polytopes}~\cite{WDGC2013_Science}.
For a positive integer $n$, let $[n] := \{1,2,\ldots,n\}$.
Let $Mat_n(\CC)$ denote the set of $n \times n$ matrices over $\CC$.
We will regard $Mat_{n_1}(\CC) \times Mat_{n_2}(\CC) \times \cdots \times Mat_{n_d}(\CC)$
as a block-diagonal subspace of $Mat_{n_1 + n_2 + \cdots + n_d}(\CC)$ by
\begin{equation}\label{eqn:block-diagonal}
(A_1,A_2,\ldots,A_d) \ \equiv \
\left(
\begin{array}{cccc}
A_1 & & & \\
& A_2 & & \\
& & \ddots & \\
& & & A_d
\end{array}
\right),
\end{equation}
where multiplication and other operations are considered in $Mat_{n_1 + n_2 + \cdots + n_d}(\CC)$.
%as $(A_1,A_2,\ldots,A_d) \cdot (B_1,B_2,\ldots,B_d) = (A_1B_1,A_2B_2,\ldots,A_dB_d)$.
\subsection{GL-action on tensors and entanglement polytopes}
Let $d$ be a positive integer.
For each $i \in [d]$, let ${\cal V}_i := \CC^{n_i}$ be
an $n_i$-dimensional $\CC$-vector space with standard Hermitian inner product
$\langle u,v\rangle := u^{\dagger}v$.
Consider the tensor product space ${\cal V}:= {\cal V}_1 \otimes {\cal V}_2 \otimes \cdots \otimes {\cal V}_d$.
An element in ${\cal V}$ is called a ($d$-)tensor.
A tensor $v \in {\cal V}$ is uniquely written as
\begin{equation*}%\label{eqn:v_j1j2...jd}
v = \sum_{j_1,j_2,\ldots,j_d} v_{j_1j_2\cdots j_d} e_{j_1} \otimes e_{j_2} \otimes \cdots \otimes e_{j_d},
\end{equation*}
where $e_i$ denotes the $i$-th unit vector and the index $j_k$ ranges over $[n_k]$.
The Hermitian inner product $\langle, \rangle$ on ${\cal V}$ is determined by extending
$\langle u_1 \otimes \cdots \otimes u_d, v_1 \otimes \cdots \otimes v_d \rangle := \langle u_1,v_1\rangle \cdots \langle u_d,v_d\rangle$ (bi)linearly. The norm $\|\cdot\|$ on ${\cal V}$
is defined by $\|v\| := \sqrt{\langle v, v\rangle}$ accordingly.
%
The product of matrices $Mat_{n_1}(\CC) \times Mat_{n_2}(\CC) \times \cdots \times Mat_{n_d}(\CC)$
acts on ${\cal V}$ by
\[
(a_1,a_2,\ldots,a_d) \cdot (v_1 \otimes v_2 \otimes \cdots \otimes v_d)
= (a_1 v_1) \otimes (a_2 v_2) \otimes \cdots \otimes (a_d v_d).
\]
Let $G := GL_{n_1}(\CC) \times GL_{n_2}(\CC) \times \cdots \times GL_{n_d}(\CC)$.
%and $K := SU_{n_1} \times SU_{n_2} \times \cdots \times SU_{n_d}$.
We consider the $G$-action on ${\cal V}$.
The associated {\em moment map} $\mu:{\cal V} \to {\cal H} := {\cal H}_{n_1} \times {\cal H}_{n_2} \times \cdots \times {\cal H}_{n_d}$ is given by
\begin{equation*}
\mu(v) = \left( \mu_1(v), \mu_2(v),\ldots,\mu_d(v)\right),
\end{equation*}
where each $\mu_k(v) \in {\cal H}_{n_k}$ is given by
\begin{equation*}
\trace \mu_k(v) H = \frac{\langle v, (I,\ldots, I, \overset{k}{H}, I,\ldots,I)\cdot v\rangle}{\|v \|^2}
\quad (H \in {\cal H}_{n_k}).
\end{equation*}
% \begin{equation}
% \mu(v) =
% \left(
% \begin{array}{cccc}
% \mu_1(v) & & & \\
% & \mu_2(v) & & \\
% & & \ddots & \\
% & & & \mu_d(v)
% \end{array}
% \right),
%\end{equation}
%where each $\mu_i(v) \in V_i^{*} \times V_i$ is given by
%\begin{equation}
%\mu_i(v) = \frac{\trace^{-i}v^{\dagger} \otimes v}{\langle v, v \rangle}.
%\end{equation}
The matrix $\mu_k(v)$ is explicitly written as
\begin{equation}\label{eqn:mu_k(v)_ij}
\mu_k(v)_{ij}= \frac{1}{\| v \|^2}\sum_{j_1,\ldots, j_{k-1},j_{k+1}, \ldots,j_d}
v_{j_1 \cdots i \cdots j_d} \overline{v_{j_1 \cdots j \cdots j_d}} \quad (i,j \in [n_k]),
\end{equation}
where the indices $i,j$ place on the $k$-th index and the index $j_\ell$ ranges over $[n_\ell]$ for $\ell \in [d]\setminus \{k\}$.
%and $\overline{(\cdot)}$ means the complex conjugate.
% The matrix $\mu_k(v)$ is written by the {\em $k$-flatting} of $v$.
% Specifically, regard $v$ as a linear map $v:({\cal V}_1 \otimes \cdots \otimes {\cal V}_{k-1} \otimes {\cal V}_{k+1} \otimes \cdots \otimes {\cal V}_d)^* \to {\cal V}_k$,
% and consider its matrix representation by the basis of tensor product of unit vectors.
% For $i \in [n_k]$, let $v_i[k]$ be
% the $\prod_{j \in [d]\setminus \{k\}} n_j$-dimensional vector
% consisting of $v_{j_1j_2\cdots j_{k-1}ij_{k+1}\cdots j_{d}}$,
% where we fix an arbitrary ordering on indices $j_{1},j_2,\ldots,j_{k-1},j_{k+1},\ldots,j_{d}$.
% Then $\mu_k(v)$ is written as
% \begin{equation}
% \mu_k(v) = \left( \begin{array}{c} v_1[k]^{\dagger} \\ v_2[k]^{\dagger} \\ \vdots \\ v_{n_k}[k]^{\dagger} \end{array}\right)
% \end{equation}
From this, we see that $\mu_i(v) \in {\cal D}^+_{n_i}$,
and $\mu(v) \in {\cal D}^+:= {\cal D}^+_{n_1} \times {\cal D}^+_{n_2} \times \cdots \times {\cal D}^+_{n_d}$.
% \begin{eqnarray}
% \trace \mu_i(v) &=& 1, \\
% \mu_i(v) & \succeq & 0.
% \end{eqnarray}
Now we are given a convex function $S:{\cal H} \to \overline{\RR}$
that is invariant under the adjoint action of $K := U_{n_1} \times U_{n_2} \times \cdots \times U_{n_d}$:
\begin{itemize}
\item[S1:] $S$ is a lower-semicontinuous convex function with
${\cal D}^+ \subseteq \dom S$.
\item[S2:] $S$ is {\em $K$-invariant}, that is, $S(Y) = S(k Y k^{\dagger})$
holds for any $Y \in {\cal H}$ and $k \in K$.
\end{itemize}
In S2, we use the convention (\ref{eqn:block-diagonal}).
Observe that $S$ is a function of the eigenvalues of $Y_i$.
For $(p^{(1)},p^{(2)},\ldots, p^{(d)}) \in \RR^{n_1} \times \RR^{n_2} \times \cdots \times \RR^{n_d}$,
we denote
$S(\diag p^{(1)},\diag p^{(2)},\ldots, \diag p^{(d)})$
simply by $S(p^{(1)},p^{(2)},\ldots, p^{(d)})$.
Consider the following optimization problem on the group $G$:
\begin{equation}\label{eqn:S(mu(gv))}
{\rm Min.} \quad S(\mu(g\cdot v)) \quad {\rm s.t.} \quad g\in G.
\end{equation}
This problem is viewed as a convex optimization problem on the Euclidean space $\RR^{n_1+\cdots + n_d}$ as well
as the minimum $Q$-gradient-norm problem (\ref{eqn:minimum_Q-gradient-norm_problem}).
For $i \in [d]$,
let $\spec \mu_i(v) \in \RR^{n_i}$ denote the nonincreasing
vector $p^{(i)} \in \RR^{n_i}$ with $p_1^{(i)} \geq p_2^{(i)} \geq \cdots \geq p_{n_i}^{(i)}$
of the eigenvalues of $\mu_i(v)$ counting multiplicity.
Let $\spec \mu(v) := (\spec \mu_1(v),\spec \mu_2(v),\ldots,\spec \mu_d(v))$.
Define $\Delta(v) \subseteq \RR^{n_1} \times \RR^{n_2} \times \cdots \times \RR^{n_d}$ by
\begin{equation}
\Delta(v) := \overline{\{ \spec \mu(g \cdot v) \mid g \in G \}}.
\end{equation}
Then the problem (\ref{eqn:S(mu(gv))}) is rewritten as
%
\begin{equation}\label{eqn:min_S(p)}
{\rm Min.} \quad S(p) \quad {\rm s.t.} \quad p = (p^{(1)}, p^{(2)},\ldots,p^{(d)}) \in \Delta(v).
\end{equation}
%
The set $\Delta(v)$ is known as the {\em moment polytope} of the Hamiltonian action $G \curvearrowright \mathbb{P}({\cal V})$.
Then, the classical result in GIT implies:
\begin{Thm}[{Convexity theorem~\cite{GuilleminSternberg1982-I,GuilleminSternberg1984-II,Kirwan1984}}]
$\Delta(v)$ is a convex polytope.
\end{Thm}
%
Although this theorem holds for more general actions,
it was explored by Walter, Doran, Gross, and Christandl~\cite{WDGC2013_Science} that
this moment polytope $\Delta(v)$ is an effective tool for studying the SLOCC entanglement class of a quantum multi-particle system,
and is particularly called the {\em entanglement polytope}.
Note also that in the literature of K-stability,
Lee, Sturm, and Wang~\cite{Lee2022momentmapconvexfunction}
studied other aspects of invariant convex functions
on moment polytopes (for general actions).
The problem (\ref{eqn:S(mu(gv))}) is also written as
the minimum $Q$-gradient-norm problem (\ref{eqn:minimum_Q-gradient-norm_problem}) on
the product of PD-manifolds ${\cal M} := {\cal P}_{n_1} \times {\cal P}_{n_2} \times \cdots \times {\cal P}_{n_d}
\subseteq {\cal P}_{n_1 + \cdots + n_d}$.
Here $T_I = T_I{\cal M}$ is given by ${\cal H}$,
and $T_I^* = T_I^*{\cal M}$ is also identified with ${\cal H}$
by $Y(X) := \sum_{i=1}^d \trace Y_iX_i$ $(X \in T_I = {\cal H})$.
%
Now we can regard ${\cal D}^{+} \subseteq T_I^*$ and $S:T_I^* \to \overline{\RR}$.
%
Define the {\em Kempf-Ness function} $\Phi_v: {\cal M} \to \RR$ by
\begin{equation}
\Phi_v(x) := \log \langle v, x \cdot v\rangle \quad (x \in {\cal M}).
\end{equation}
It is well-known that the Kempf-Ness function is a geodesically convex function on the symmetric space ${\cal M} \simeq G/K$.
\begin{Lem}[see e.g., \cite{BFGOWW_FOCS2018,BFGOWW_FOCS2019,GRS_Book}]
$\Phi_v$ is geodesically convex.
\end{Lem}
The moment map $\mu$ is interpreted as transported differentials of $\Phi$.
\begin{Lem}[see \cite{BFGOWW_FOCS2018,HiraiSakabe2024FOCS}]\label{lem:dPhi_x}
$\tau_{x \to I}d(\Phi_v)_x = \mu( x^{1/2} \cdot v)$.
\end{Lem}
\begin{proof}
For $X = (X_1,X_2,\ldots,X_d) \in T_x$, we have
\begin{eqnarray*}
d(\Phi_v)_x(X) &= & \frac{d}{dt}\mid_{t = 0} \log \langle v , (x_1+tX_1,\ldots, x_d + tX_d) \cdot v \rangle \\
&=& \frac{1}{\| x^{1/2}\cdot v\|^2}\sum_{i=1}^d \langle x^{1/2} \cdot v , (I,\ldots,I,x_i^{-1/2} X_ix_i^{-1/2},I,\ldots,I) \cdot x^{1/2} \cdot v \rangle \\
&=& \sum_{i=1}^d \trace \mu_i(x^{1/2}\cdot v) (x_i^{-1/2}X_ix_i^{-1/2}).
\end{eqnarray*}
Hence $d(\Phi_v)_x(X) = \mu(x^{1/2}\cdot v)(\tau_{x \to I}X)$.
%Projecting $\mu(v)$ to $T_I^*$, we obtain the claim.
\end{proof}
Extend $S$ to $T^*{\cal M} \to \overline{\RR}$ by
\begin{equation*}
S_x(Y) := S\left(\tau_{x \to I} Y \right) \quad (x \in {\cal M}, Y \in T^*_x{\cal M}).
\end{equation*}
If $x = g^{\dagger}g$ for $g \in G$,
then $x^{1/2} = k g$ for $k \in K$ and
$S(d(\Phi_v)_x) = S(\tau_{x \to I} d (\Phi_v)_x) = S(\mu(x^{1/2}\cdot v)) = S(\mu(kg\cdot v)) = S(k^{\dagger} \mu(g\cdot v)k) = S(\mu(g \cdot v))$.
Hence, the problem (\ref{eqn:S(mu(gv))}) becomes
\begin{equation*}
{\rm Min.} \quad S(d(\Phi_v)_x) \quad {\rm s.t.}\quad x\in {\cal M} = {\cal P}_{n_1} \times {\cal P}_{n_2} \times \cdots \times {\cal P}_{n_d}
\end{equation*}
for which the results in Section~\ref{sec:Q-gradient_flow} are applicable.
Since ${\cal D}^+$ is compact, we can assume that level sets of $S$ are bounded.
Then, $({\cal M},S)$ satisfies $\tilde {\rm Q}$1, Q2, and Q3 (see Example~\ref{ex:PD-manifold}).
By Theorem~\ref{thm:main2}, we have:
\begin{Thm}[Duality theorem for convex optimization on entanglement polytope] \label{thm:duality_entanglement}
\begin{equation}\label{eqn:duality_entanglement}
\min_{p \in \Delta(v)} S(p) = \sup_{\xi \in C{\cal M}^{\infty}} - \Phi_v^{\infty}(\xi) - S^*_{-}(\xi).
\end{equation}
\end{Thm}
We also give an explicit formula for the recession function $\Phi_v^{\infty}$.
Note that the recession function of the Kempf-Ness function is essentially Mumford's numerical invariant {\em $\mu$-weight $w_{\mu}$}~\cite[Definition 5.1]{GRS_Book}.
\begin{Lem}[{Formula of $\Phi_v^{\infty}$; see \cite[Lemma 3.9 (1)]{Hirai_Hadamard2022}}]
\begin{eqnarray*}
&& \Phi_v^{\infty}(Y) = \max \{\lambda_{j_1}^{(1)}+ \lambda_{j_2}^{(2)} + \cdots + \lambda_{j_d}^{(d)} \mid j_1, j_2,\ldots, j_d: (k\cdot v)_{j_1j_2\cdots j_d} \neq 0\},
\end{eqnarray*}
where $k \in K$ such that
$Y_i = k_i^{\dagger} \diag \lambda^{(i)} k_i$ and $\lambda^{(i)} \in \RR^{n_i}$ for $i\in [d]$.
\end{Lem}
\begin{proof}
By definition of $\Phi_v$ and $\Phi_v^{\infty}$, we have
\begin{eqnarray}
&& \Phi_v^{\infty}(Y) = \lim_{t \to \infty} \frac{\Phi_v(e^{tY})}{t}
= \lim_{t \to \infty} \frac{1}{t}\log \| (e^{t \diag \lambda^{(1)}/2},\ldots,e^{t \diag \lambda^{(d)}/2}) k \cdot v \|^2 \nonumber \\
&& = \lim_{t \to \infty} \frac{1}{t} \log \sum_{j_1,j_2,\ldots,j_d} | ( k \cdot v)_{j_1j_2\cdots j_d}|^2 e^{t\left(\lambda^{(1)}_{j_1}+ \lambda^{(2)}_{j_2} + \cdots +\lambda^{(d)}_{j_d}\right)}. \label{eqn:Phi^infty}
\end{eqnarray}
Apply $\lim_{t \to \infty}(1/t)\log \sum_{i} e^{ta_i+b_i} = \max_{i} a_i$.
\end{proof}
%This is true for
%moment polytopes for general $SL$-actions.
%If Conjecture~\ref{conj} is true for general symmetric spaces,
%it is true for moment polytopes of reductive group actions.
%
By identifying $C{\cal M}^{\infty}$ with the Euclidean building
${\cal B} := {\cal B}_{n_1} \times {\cal B}_{n_2} \times \cdots \times {\cal B}_{n_d}$,
the negative of the RHS of (\ref{eqn:duality_entanglement})
is formulated as a geodesically convex optimization on the building~${\cal B}$ (see also Proposition~\ref{prop:g-convexity}):
\begin{eqnarray}
\mbox{Max.} && -S^*(-(\lambda^{(1)}, \lambda^{(2)},\ldots, \lambda^{(d)})) \nonumber \\
&& \quad - \max \left\{\sum_{i=1}^d \lambda_{j_i}^{(i)} \mid j_1,j_2,\ldots,j_d:
v({\cal Y}^{(1)}_{j_1},{\cal Y}_{j_2}^{(2)},\ldots,{\cal Y}_{j_d}^{(d)}) \neq \{0\} \right\} \nonumber \\
%&& \quad \quad \quad+ \sum_{i=1}^d \theta(i) \log \sum_{j=1}^{n_i} e^{- \lambda_j^{(i)}}, \nonumber\\
\mbox{s.t.} && \{0\} < {\cal Y}_1^{(i)} < {\cal Y}_2^{(i)} < \cdots < {\cal Y}_{n_i}^{(i)} = \CC^{n_i},\lambda_1^{(i)} \geq \lambda_2^{(i)} \geq \cdots \geq \lambda_{n_i}^{(i)} \ (i \in [d]), \label{eqn:S(v)_building}
\end{eqnarray}
where $v$
is viewed as a multilinear map $\CC^{n_1} \times \CC^{n_2} \times \cdots \times \CC^{n_d} \to \CC$.
We expect that the duality (\ref{eqn:duality_entanglement}) would lead to
a good characterization
(NP $\cap$ co-NP characterization) for
the decision problem of deciding $\min_{p \in \Delta(v)} S(p) \leq \alpha$ or not.
Indeed, if $\min_{p \in \Delta(v)} S(p) > \alpha$, then
a feasible solution $(\{{\cal Y}_j^{(i)}\}, \{\lambda_j^{(i)}\})$ of (\ref{eqn:S(v)_building}) with the objective value greater than $\alpha$
is a NO-certificate.
%
To accomplish this for specific problems,
the bit-complexity of $(\{{\cal Y}_j^{(i)}\}, \{\lambda_j^{(i)}\})$ must be polynomially bounded,
which is an important future research topic.
%By Proposition~\ref{prop:g-convexity}, the objective function is actually geodesically convex.
Next, consider the case where the $S$-gradient flow $x(t)$ of $\Phi_v$ is definable.
By Theorem~\ref{thm:main},
$S(d(\Phi_v)_{x(t)}) = S(\mu(x^{1/2}(t)\cdot v))$
converges to the minimum $S$-gradient norm over ${\cal M}$,
which implies that $\spec \mu(x^{1/2}(t) \cdot v)$
accumulates to minimizers of
the problem (\ref{eqn:min_S(p)}).
%%
\begin{Thm}[{$S$-moment-limit theorem}]\label{thm:moment-limit}
Suppose that $S$ is $\RR_{\geq 0}$-valued and $S^2$ is continuously differentiable.
Let $x(t)$ be the $S$-gradient flow of $\phi_v$.
Then, any accumulation point of $\spec \mu(x^{1/2}(t) \cdot v)$ is a minimizer of $S$ over $\Delta (v)$.
%If a minimizer is unique, then $\spec \mu(x^{1/2}(t) \cdot v)$ converges to the unique minimizer.
\end{Thm}
From the view of the theory of {\em moment-weight inequality} (Georgoulas, Robbin, and Salamon\cite{GRS_Book}),
this theorem is a generalization of the {\em moment-limit theorem}~\cite[Theorem 6.4]{GRS_Book}
from $\|\cdot\|_{\rm F}$ to $S$ (in our setting of the GL-action on tensors).
We have verified that several parts of the theory
are directly generalized to the $S$-setting.
For example, the weak duality $\geq$ in (\ref{eqn:duality_entanglement})
is the $S$-version of the moment-weight inequality.
Also, the $S$-gradient flow is related to a generalization
of the {\em gradient flow of moment-map squared}~\cite[Chapter 3]{GRS_Book}
from $\|\cdot\|_{\rm F}^2/2$ to $S^2/2$.
However, we shall not pursue this issue further; the full generalization needs to solve Conjecture~\ref{conj}
for arbitrary symmetric spaces.
We expect that the $S$-moment-limit theorem will lead to an algorithm
for convex optimization~(\ref{eqn:min_S(p)}).
Indeed, it strongly suggests
the convergence of the $S$-subgradient method with an appropriate step-size.
We here only give a formulation of the $S$-subgradient method
by updating group element $g_i \in G$ in the spirit
of {\em noncommutative ``group'' optimization}~\cite{BFGOWW_FOCS2019}:
\begin{equation*}
g_{i+1} := e^{-\delta_i Z_i/2} g_i,\quad Z_i \in \partial (S^2/2) (\mu(g_i\cdot v)) \quad (i=1,2,\ldots).
\end{equation*}
Then $x_i := g_i^\dagger g_i$ satisfies (\ref{eqn:Q-subgradient_descent}).
This can be seen from Lemma~\ref{lem:dPhi_x} and
$\partial (S^2/2) (\mu(k x_i^{1/2}\cdot v)) = \partial (S^2/2) (k\mu( x_i^{1/2}\cdot v)k^{\dagger})
= k^{\dagger} \partial (S^2/2) (\mu(x_i^{1/2}\cdot v)) k$
for $g_i = k x_i^{1/2}$ with $k \in K$.
%
In the subsequent sections, we consider three concrete examples
(quantum functional, $G$-stable rank, noncommutative rank)
for which the developed theory is applicable.
\subsection{Quantum functional}
A $d$-tensor in $\bigotimes_{i=1}^d \CC^{n_i} = \CC^{n_1} \otimes \cdots \otimes \CC^{n_d}$ is viewed as a multilinear map $\CC^{n_1} \times \cdots \times \CC^{n_d} \to \CC$.
For two $d$-tensors $u \in \bigotimes_{i=1}^d \CC^{n_i}$ and
$v \in \bigotimes_{i=1}^d \CC^{m_i}$,
the tensor product $u \otimes v \in \bigotimes_{i=1}^d \CC^{n_i} \otimes \CC^{m_i}$
and the direct sum $u \oplus v \in
\bigotimes_{i=1}^d \CC^{n_i} \oplus \CC^{m_i}$ are defined
by $(u \otimes v) (x_1\otimes y_1,\ldots,x_d \otimes y_d) := u(x_1,\ldots,x_d)v(y_1,\ldots,y_d)$ and $(u \oplus v)((x_1,y_1),\ldots,(x_d,y_d)) := u(x_1,\ldots,x_d)+ v(y_1,\ldots,y_d)$, respectively.
We say
that $u$ {\em restricts to} $v$, denoted by $v \leq u$, if there are linear maps $A_k:\CC^{m_k} \to \CC^{n_k}$
such that $v = u \circ (A_1,A_2,\ldots,A_d)$.
A function $\xi: \{ \mbox{$d$-tensors} \} \to \RR_{\geq 0}$ is
called a (universal) {\em spectral point}
if it is (a) monotone under $\leq$, (b) multiplicative under $\otimes$, (c) additive under $\oplus$,
and (d) normalized so that $\xi(\langle n\rangle) = n$ for any unit tensor
$\langle n \rangle := \sum_{i=1}^n e_i \otimes \cdots \otimes e_i \in (\CC^{n})^{\otimes d}$.
Spectral points of tensors play significant roles in Strassen's theory~\cite{Strassen1988_asymptotic_spectrum}
on {\em asymptotic spectra}
and matrix multiplication complexity; see \cite[part 3]{Landsberg_tensors2019} and \cite{WigdersonZuiddam2023} for survey.
%
Nontrivial spectral points (other than the flattening rank) were not known until
the recent discovery of {\em quantum functionals} by
Christandl, Vrana, and Zuiddam~\cite{CristandlVramaZuiddam2023}.
%
For a probability vector $\theta \in \RR_{>0}^d$ with $\sum_{i=1}^d \theta_i = 1$,
the quantum functional $F_{\theta}: \{ \mbox{$d$-tensors} \} \to \RR_{\geq 0}$ is defined by
\begin{equation}\label{eqn:quantum_functional}
F_{\theta}(v) := 2^{E_{\theta}(v)},\quad E_{\theta}(v) := \sup_{g \in G} \sum_{i=1}^d \theta_i H (\mu_i(g \cdot v)),
\end{equation}
where $H$ is the von Neumann entropy (Example~\ref{ex:entropy}).
It is clear that
the logarithmic quantum functional $E_{\theta}$ is obtained by
entropy maximization of the entanglement polytope:
\begin{equation*}
{\rm Max.} \quad \sum_{i=1}^d \theta_i H (p^{(i)}) \quad {\rm s.t.} \quad p=(p^{(1)},p^{(2)},\ldots,p^{(d)})\in \Delta(v).
\end{equation*}
By applying Theorem~\ref{thm:duality_entanglement}
for the setting $S(Y) := - \sum_{i=1}^d \theta_i H(Y_i)$, we obtain
a new expression of the logarithmic quantum functional $E_{\theta}$.
%
\begin{Thm}
\begin{equation}\label{eqn:E_theta(v)}
E_\theta(v) = \inf_{X \in {\cal H}} \Phi_v^{\infty}(X) + \sum_{i=1}^d \theta_i \log_2 \trace 2^{- X_i}.
\end{equation}
\end{Thm}
In this formula, we identify $C{\cal M}^{\infty}$ with $T_I = {\cal H}$.
See Example~\ref{ex:entropy} for the conjugate of $-H$.
\paragraph{Seeing that quantum functionals are spectral points.}
It may be worth seeing the properties (a),(b),(c),(d) of quantum functionals from our approach.
As in \cite[Section 3.1]{CristandlVramaZuiddam2023},
the properties (a),(d), super-multiplicativity $F_{\theta}(v \otimes v') \geq F_{\theta}(v) F_{\theta}(v')$ in (b), and
super-additivity $F_{\theta}(v \oplus v') \geq F_{\theta}(v) + F_{\theta}(v')$ in (c)
are directly shown from the definition (\ref{eqn:quantum_functional}).
%
The difficult part (for us) is to show the remaining properties $F_{\theta}(v \otimes v') \leq F_{\theta}(v) F_{\theta}(v')$
and $F_{\theta}(v \oplus v') \leq F_{\theta}(v) + F_{\theta}(v')$, which requires
representation-theoretic methods~\cite[Section 3.2]{CristandlVramaZuiddam2023} \cite[Sections 7.3.1, 7.3.2]{Landsberg_tensors2019}.
The above formula (\ref{eqn:E_theta(v)}) directly proves sub-multiplicativity of $F_{\theta}$ (sub-additivity of $E_{\theta}$) as follows.
Let $X = (X_i)$ and $X' = (X_i')$ be any solutions of the optimization problems in (\ref{eqn:E_theta(v)}) for
$E_\theta(v)$ and $E_\theta(v')$, respectively.
Consider $Z = (Z_i) := (X_i \otimes I + I \otimes X_i')$,
which is a solution of the problem for $E_\theta(v \otimes v')$.
Then we observe $\Phi_v^{\infty}(Z) = \Phi_v^{\infty}(X) + \Phi_v^{\infty}(X')$
and $\log_2 \trace 2^{-Z_i} = \log_2 \trace 2^{-X_i} + \log_2 \trace 2^{-X_i'}$.
From this, we have sub-additivity
$E_{\theta}(v \otimes v') \leq \inf_{Z = (X_i \otimes I + I \otimes X_i')} \Phi_v^{\infty}(Z) +
\sum_{i} \theta_i \log_2 \trace 2^{Z_i} = E_{\theta}(v)+ E_{\theta}(v')$.
The additivity (c) can be seen from the convergence of the $S$-gradient flow
(Theorem~\ref{thm:moment-limit}), where we can choose $S$
as the Moreau envelop
of $- E_{\theta} + C > 0$
for a constant $C$ and an arbitrary small parameter $\lambda >0$; see the proof of Theorem~\ref{thm:main2}.
Let ${\cal V}'$, $G'$, and ${\cal M}'$ denote the tensor space, group, and manifold, respectively, for the second tensor $v'$.
Let $\bar{G}$ and $\bar{\cal M}$ denote the group and manifold for the sum $v \oplus v'$, respectively.
By direct calculation $2^{\sup_{g \in G,g'\in G'} \sum_{i} \theta(i) H(\mu_i( (g\cdot v) \oplus (g' \cdot v')))}
= F_\theta(v) + F_\theta(v')$~\cite[Lemmas 3.8--3.10]{CristandlVramaZuiddam2023},
it suffices to show
\begin{equation}\label{eqn:sup_h=sup_g,g'}
(E_{\theta}(v \oplus v') :=)\ \sup_{h \in \bar G} \sum_{i=1}^d \theta_i H(\mu_i( h\cdot (v \oplus v'))) = \sup_{g \in G,g'\in G'} \sum_{i=1}^d \theta_i H(\mu_i( (g\cdot v) \oplus (g' \cdot v'))).
\end{equation}
Here, the moment map $\mu( (g\cdot v) \oplus (g'\cdot v')) \in T_{I}^*\bar{\cal M}$ is written as
\[
\mu((g\cdot v) \oplus (g'\cdot v')) = \left(
\begin{array}{cc}
\frac{\|g\cdot v\|^2}{\|g\cdot v\|^2+\|g'\cdot v'\|^2}\mu(g\cdot v) & O \\
O & \frac{\|g'\cdot v'\|^2}{\|g\cdot v\|^2+\|g'\cdot v'\|^2}\mu(g'\cdot v')
\end{array}
\right) \quad \in T_{I,I}^*({\cal M} \times {\cal M}').
\]
From this matrix structure and the formula of $dF$ of a unitarily invariant convex function $F$ (Section~\ref{subsec:convex-analysis}),
we see that
the $S$-gradient $\nabla^{S} \Phi_{v \oplus v'}(x,x') = \tau_{(I,I) \to (x,x')} d (S^2/2)_{\mu((x^{1/2}\cdot v) \oplus ((x')^{1/2}\cdot v')))}$ at the submanifold ${{\cal M} \times {\cal M}'} \subseteq \bar{\cal M}$ belong to $T({\cal M} \times {\cal M}')$.
Consequently, the $S$-gradient flow of $\Phi_{v \oplus v'}$
with initial point $x_0 = I \in {\cal M} \times {\cal M}'$
remains in ${\cal M} \times {\cal M}'$ but attains the minimum $S$-gradient-norm over $\bar{\cal M}$ in the limit.
This implies (\ref{eqn:sup_h=sup_g,g'}) by $\sup_{h \in \bar{G}} = \sup_{\bar x \in \bar {\cal M}}
= \sup_{(x,x') \in {\cal M} \times {\cal M}'} = \sup_{g\in G, g' \in G'}$.
%van den Berg et al.\cite{vdBCLNWZ_STOC2025}
\subsection{$G$-stable rank}
Derksen~\cite{Derksen2022ANT} introduced a new rank concept of tensors,
called the {\em $G$-stable rank},
and revealed interesting properties and applications.
As a consequence of~\cite[Theorem 5.2]{Derksen2022ANT},
the $G$-stable rank for a complex $d$-tensor $v \in {\cal V}$
is obtained by a convex optimization problem on the entanglement polytope.
Here we introduce the $G$-stable rank via this optimization problem.
For each $i\in [d]$,
the tensor $v$ is viewed as a linear map
$v:({\cal V}_1 \otimes \cdots \otimes {\cal V}_{i-1} \otimes {\cal V}_{i+1} \otimes \cdots \otimes {\cal V}_d)^* \to {\cal V}_i$, which is called the {\em $i$-flattening} of $v$.
The corresponding operator norm is denoted by $\|v\|_{i, {\rm op}}$.
For a positive vector $\alpha \in \RR^d_{>0}$,
the $G$-stable rank $\rk^G_\alpha (v)$ of $v$ is defined by
\begin{equation*}
\rk^G_{\alpha}(v) := \sup_{g \in G} \min_{i \in [d]} \frac{\alpha_i \| g\cdot v \|^2}{\|g \cdot v\|^2_{i,{\rm op}}}.
\end{equation*}
Consider the matrix representation $A_{v,i} = (v_{j_1\cdots j_i \cdots j_d})$
of the $i$-flattening of $v$, where
$v_{j_1\cdots j_i \cdots j_d}$ is the element of $A_{v,i}$ with the row index $j_i$
and column index $(j_1,\ldots, j_{i-1},j_{i+1},\ldots, j_d)$.
Then we see from (\ref{eqn:mu_k(v)_ij}) that $\mu_i(v) = A_{v,i} A^{\dagger}_{v,i}/ \langle v,v \rangle$, and
\[
\frac{\|v\|^2_{i,\rm{op}}}{\|v\|^2} = \|\mu_i(v)\|_{\rm op}.
\]
% Therefore
% $\rk^G (v)$ is written as
% \begin{equation}
% \rk^G(v) = \sup_{g \in G} \min_{i \in [d]} \frac{\alpha_i}{\|\mu_i(g \cdot v)\|_{\rm op}}
% \end{equation}
Therefore, $1/ \rk_{\alpha}^G(v)$ is given by convex optimization~(\ref{eqn:S(mu(gv))})
with the setting $S(Y) := \max_{i \in [d]}\|Y_i\|_{\rm op}/\alpha_i$:
\begin{equation*}
{\rm Min.} \quad \max_{i\in [d]} \frac{1}{\alpha_i} \|\mu_i(g \cdot v)\|_{\rm op} \quad {\rm s.t.} \quad g\in G.
\end{equation*}
It is a weighted $\ell_{\infty}$-norm minimization problem
on the entanglement polytope:
\begin{equation*}
{\rm Min.} \quad \max_{i\in [d]} \frac{1}{\alpha_i} \|p^{(i)}\|_{\rm \infty} \quad {\rm s.t.} \quad p=(p^{(1)},p^{(2)},\ldots,p^{(d)})\in \Delta(v).
\end{equation*}
By applying Theorem~\ref{thm:duality_entanglement},
we obtain a dual expression of $\rk^G_{\alpha}$:
\begin{Thm}%[{see \cite[Theorem 5.2]{Derksen2022ANT}}]
\begin{equation}\label{eqn:1/rk^G}
\frac{1}{\rk^G_{\alpha}(v)} = \sup \left\{ - \Phi_v^{\infty}(X) \mid X \in {\cal H}:\sum_{i=1}^d \alpha_i \| X_i \|_{\rm tr} \leq 1 \right\}.
\end{equation}
\end{Thm}
\begin{proof}
Since $S$ is a norm, the conjugate $S^*$ is given by $\delta_B$ for
the unit ball $B$ with respect to the dual norm of $S$, which
is given by $X \mapsto
\sup_{Y_i: \|Y_i/\alpha_i\|_{\rm op}\leq 1} \sum_{i=1}^d \trace X_i Y_i = \sum_{i=1}^d \alpha_i \|X_i\|_{\rm tr}$ (see Example~\ref{ex:norm}).
%Then the LHS of (\ref{eqn:duality_entanglement})
%is $\sup_{X \in {\cal H}}- \Phi^{\infty}(X) - \delta_{B}$.
\end{proof}
%
This formula matches the original definition of
$G$-stable rank~\cite[Definition 1.3, Theorem 2.4]{Derksen2022ANT}.\footnote{
By using notation $\val_{s}$ in $\cite{Derksen2022ANT}$,
the correspondence is given by $\Phi^{\infty}(Y) = - \val_{s} (s^{-Y} \cdot v)$ (seen from (\ref{eqn:Phi^infty}))
and $\|Y_k\|_{\rm tr} = - \val_{s} \det s^{-Y_i}$,
where the eigenvalues $\lambda^{(i)}_j$ of $Y_i$ in (\ref{eqn:1/rk^G}) can be assumed to be nonpositive}
\subsection{Noncommutative rank}\label{subsec:ncrank}
Finally, we explain the noncommutative rank from our viewpoint.
Consider a linear symbolic matrix
\[
A = \sum_{k=1}^m A_k x_k,
\]
where $A_k \in Mat_n(\CC)$ and $x_k$ is an indeterminate for $k \in [m]$.
The {\em noncommutative rank} $\ncrk A$~(Ivanyos, Qiao, and Subrahmanyam~\cite{IQS2017})
is the rank of $A$ considered in the {\em free skew field}~\cite{Cohn}
generated by $x_1,x_2,\ldots,x_m$ being pairwise noncommutative $x_ix_j \neq x_jx_i$.
%of the noncommutative polynomial ring $\CC\langle x_1,x_2,\ldots,x_m\rangle$.
%under the setting where $x_i$ are noncommutative $x_{i}x_j \neq x_jx_i$.
%which is defined in the {\em free skew field}---the most generic skew field of
%fractions of the noncommutative polynomial ring $\CC\langle x_1,x_2,\ldots,x_m\rangle$~\cite{IQS2017}.
%
The polynomial-time computation of $\ncrk A$ by \cite{GGOW,HamadaHirai2021,IQS2018}
is the starting point of what we have developed here.
%
It is shown by Fortin and Reutenauer~\cite{FortinReutenauer04} that $\ncrk A$ is given by a vector-space optimization problem:
%
\begin{Thm}[\cite{FortinReutenauer04}]\label{thm:FR}
\begin{equation}\label{eqn:ncrank}
\ncrk A = 2n - \max \{ \dim {\cal X} + \dim {\cal Y} \mid {\cal X},{\cal Y} \leq \CC^n: A_k({\cal X},{\cal Y}) = \{0\}\ (k \in [m])\},
\end{equation}
where each $A_k$ is viewed as a bilinear form $\CC^n \times \CC^n \to \CC$ by $(u,v) \mapsto u^{\top}A_kv$.
\end{Thm}
%
We regard the matrix $A$ as a 3-tensor $A \in \CC^n \otimes \CC^n \otimes \CC^m$ by
\begin{equation*}
A = \sum_{i,j \in [n],k \in [m]} (A_k)_{ij} e_i \otimes e_j \otimes e_k.
\end{equation*}
Consider the $G$-action for $G = GL_n(\CC) \times GL_n(\CC) \times GL_m(\CC)$ on $A$, as above.
Although it is shown \cite[Proposition 2.9]{Derksen2022ANT} that
the noncommutative rank
$\ncrk A$ coincides with the $G$-stable rank $\rk_{\alpha}^G A$ for $\alpha = (1,1,n)$,
we explain a different approach \cite{Hirai2025}
relating $\ncrk A$ with optimization on a moment polytope.
We assume for simplicity that the common left or right kernel of $A_k$ is trivial:
\begin{equation}\label{eqn:commonkernel}
\bigcap_{k=1}^m \ker A_k = \{0\} \quad \mbox{or} \quad \bigcap_{k=1}^m \ker A_k^{\dagger} = \{0\}.
\end{equation}
Otherwise, $\ncrk A$ reduces to smaller matrices.
We restrict the $G$-action
to $GL_n(\CC) \times GL_n(\CC)$, which is known as the {\em left-right action}.
The first two components of the moment map $\mu(A) = (\mu_1(A),\mu_2(A),\mu_3(A))$
are given by
\begin{equation*}
\mu_1(A) = \frac{1}{\|A\|^2} \sum_{k=1}^m A_k A_k^{\dagger},\quad \mu_2(A)^{\top} = \frac{1}{\| A\|^2} \sum_{k=1}^m A_k^{\dagger} A_k.
\end{equation*}
Then $\ncrk A$ is written as:
\begin{Thm}[\cite{Hirai2025}]
\begin{equation*}
\ncrk A = n - \frac{n}{2} \inf_{g_1,g_2 \in GL_n(\CC)} \left\| \mu_1((g_1,g_2,I) \cdot A) - \frac{1}{n}I \right\|_{\rm tr} + \left\| \mu_2((g_1,g_2,I) \cdot A) - \frac{1}{n}I \right\|_{\rm tr}.
\end{equation*}
% \begin{eqnarray*}
% n - ncrk A &=& \frac{n}{2} \inf_{g_1,g_2 \in GL_n(\CC)} \left\| \mu_1((g_1,g_2,I) \cdot A) - \frac{1}{n}I \right\|_{\rm tr} + \left\| \mu_2((g_1,g_2,I) \cdot A) - \frac{1}{n}I \right\|_{\rm tr} \\
% &=& \frac{n}{2} \inf_{(p_1,p_2) \in \Delta(A)} \|p_1 - {\bf 1}/n \|_1 + \|p_2 - {\bf 1}/n \|_1.
% \end{eqnarray*}
\end{Thm}
Although the formula in~\cite{Hirai2025}
does not have the denominator $\|(g_1,g_2,I) \cdot A\|^2$ of $\mu_i((g_1,g_2,I) \cdot A)$,
by the assumption~(\ref{eqn:commonkernel}),
we may restrict $g_1,g_2$ to satisfy $\|(g_1,g_2,I) \cdot A\|^2 = n$ in the infimum;
see \cite[(4.8)]{Hirai2025}.
%
Consider the following variant of the entanglement polytope
\[
\Delta'(A) := \overline{\{ (\mu_1( (g_1,g_2,I)\cdot A),\mu_2((g_1,g_2,I)\cdot A) \mid g_1,g_2\in GL_n(\CC)\}},
\]
which is also a convex polytope; it is the moment polytope for the left-right action.
Therefore, $\ncrk A$ is obtained from $\ell_1$-norm minimization on $\Delta'(A)$:
\begin{equation*}
\mbox{Min.} \quad \left\|p^{(1)} - \frac{\bf 1}{n} \right\|_1
+ \left\|p^{(2)} - \frac{\bf 1}{n} \right\|_1 \quad \mbox{s.t.} \quad (p^{(1)},p^{(2)})
\in \Delta'(A).
\end{equation*}
The nc-rank computation algorithm by
Hamada and Hirai~\cite{HamadaHirai2021} solves (a variant of)
the corresponding dual problem (\ref{eqn:S(v)_building}),
which is a piecewise linear extension ({\em Lov\'asz extension}~\cite{Hirai_Hadamard2022}) of
the optimization problem in (\ref{eqn:ncrank}); see also~\cite{Hirai2025}.
\section*{Acknowledgments}
The author thanks Keiya Sakabe, Asuka Takatsu, Shinichi Ohta for discussion, Cheng Zhuohao for comments, and Michael Walter for suggesting paper~\cite{Lee2022momentmapconvexfunction}.
The author was supported by JSPS KAKENHI Grant Number JP24K21315.
\bibliographystyle{plain}
%\small
\bibliography{G-gradientflow}
\appendix
\section{Proof of Lemma~\ref{lem:ui-convex=>g-convex}}
Suppose that unitarily invariant convex function $F:{\cal H}_n \to \overline{\RR}$ is written as $F = F_f$
for a symmetric convex function $f:\RR^n \to \overline{\RR}$.
We first observe ($*$1) if $\xi \in {\cal B}_n$ is
represented by $\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_n$,
${\cal Y}_1 < {\cal Y}_2 <\cdots < {\cal Y}_n$,
then $F(\xi) = f(\lambda)$.
Indeed, by the Gram-Schmidt orthonormalization, we choose an orthonormal basis $u_1,u_2,\ldots,u_n$ such that $\langle u_1,u_2,\ldots,u_j\rangle = {\cal Y}_j$ for $j \in [n]$. Then $\xi$ corresponds Hermitian matrix $k \diag \lambda k^{\dagger}$
for $k= (u_1\ u_2\ \cdots \ u_n) \in U_n$. Thus $F(\xi) = f(\lambda)$.
Since any two points of ${\cal B}_n$ belong to a common apartment,
it suffices to show that $F$ is convex on each apartment.
For any unitary $k \in U_n$,
the set ${\cal E}_k:= \{ k \diag x k^{\dagger} \mid x \in \RR^n \}$
is one of apartments, in which $\RR^n \ni x \mapsto k \diag x k^{\dagger}$ is an isometry.
Therefore, $F$ is convex on such an apartment ${\cal E}_k$.
%
Now consider a general apartment ${\cal A}$ for
(non orthonormal) basis $v_1,v_2,\ldots,v_n$.
Choose an orthonormal basis $u_1,u_2,\ldots,u_n$ such that $\langle u_1,u_2,\ldots,u_j\rangle = \langle v_1,v_2,\ldots,v_j\rangle$ for $j \in [n]$. Let $k:= (u_1\ u_2\ \cdots \ u_n) \in U_n$.
Let $\rho:{\cal A} \to {\cal E}_{k}$ be
defined by
\begin{equation*}
\begin{array}{c}
y_{i_1} \geq y_{i_2}\geq \cdots \geq y_{i_n}, \\
\langle v_{i_1}\rangle < \langle v_{i_1},v_{i_2} \rangle < \cdots < \langle v_{i_1},v_{i_2},\ldots,v_{i_n} \rangle
\end{array} \mapsto
\begin{array}{c}
y_{i_1} \geq y_{i_2}\geq \cdots \geq y_{i_n}, \\
\langle u_{i_1}\rangle < \langle u_{i_1},u_{i_2} \rangle < \cdots < \langle u_{i_1},u_{i_2},\ldots,u_{i_n} \rangle
\end{array}.
\end{equation*}
It is clear that $\rho:{\cal A} \to {\cal E}_{k}$ is an isometry.
By ($*$1), it holds ($*$2) $F(\xi) = F \circ \rho(\xi) = f(y)$.
Therefore, for $\xi,\xi' \in {\cal A}$ and $t \in [0,1]$,
denoting by $(1-t) \xi + t \xi'$ the point
$\gamma(t)$ on the unique geodesic
$\gamma:[0,1] \to {\cal B}_n$ with $\gamma(0) = \xi$ and $\gamma(1) = \xi'$,
we have
$F((1-t) \xi + t\xi') = F \circ \rho ((1-t)\xi + t \xi') = F((1-t) \rho(\xi) + t\rho (\xi'))
\leq (1-t) F \circ \rho(\xi) + t F \circ \rho(\xi') = (1-t) F(\xi) + t F(\xi')$,
where the first and last equalities follow from ($*$2), the second from the fact
that $\rho$ is an isometry, and the third follows from the convexity of $F$ on ${\cal E}_k$.
\end{document}
\begin{thebibliography}{10}
\bibitem{Ambrosio_Book}
N.~Ambrosio, L.~Gigli and G.~Savar\'{e}.
\newblock {\em Gradient Flows: In Metric Spaces and in the Space of Probability
Measures}.
\newblock Birkh\"{a}user, Basel, {S}econd edition, 2008.
\bibitem{Bhatia_PositiveDefiniteMatrices}
R.~Bhatia.
\newblock {\em Positive Definite Matrices}.
\newblock Princeton University Press, Princeton, NJ, 2007.
\bibitem{Boumal_Book}
N.~Boumal.
\newblock {\em An Introduction to Optimization on Smooth Manifolds}.
\newblock Cambridge University Press, Cambridge, 2023.
\bibitem{BridsonHaefliger1999}
M.~R. Bridson and A.~Haefliger.
\newblock {\em Metric Spaces of Non-Positive Curvature}.
\newblock Springer-Verlag, Berlin, 1999.
\bibitem{Burgisser_ACT_Book}
P.~B\"urgisser, M.~Clausen, and M.~A. Shokrollahi.
\newblock {\em Algebraic Complexity Theory}.
\newblock Springer-Verlag, Berlin, 1997.
\bibitem{BFGOWW_FOCS2018}
P.~B{\"{u}}rgisser, C.~Franks, A.~Garg, R.~Oliveira, M.~Walter, and
A.~Wigderson.
\newblock Efficient algorithms for tensor scaling, quantum marginals, and
moment polytopes.
\newblock In {\em 59th {IEEE} Annual Symposium on Foundations of Computer
Science, {FOCS} 2018}, pages 883--897, 2018.
\bibitem{BFGOWW_FOCS2019}
P.~B\"{u}rgisser, C.~Franks, A.~Garg, R.~Oliveira, M.~Walter, and A.~Wigderson.
\newblock Towards a theory of non-commutative optimization: geodesic 1st and
2nd order methods for moment maps and polytopes.
\newblock In {\em 60th {IEEE} {A}nnual {S}ymposium on {F}oundations of
{C}omputer {S}cience, FOCS 2019}, pages 845--861, 2019.
\bibitem{CristandlVramaZuiddam2023}
M.~Christandl, P.~Vrana, and J.~Zuiddam.
\newblock Universal points in the asymptotic spectrum of tensors.
\newblock {\em J. Amer. Math. Soc.}, 36(1):31--79, 2023.
\bibitem{Cohn}
P.~M. Cohn.
\newblock {\em Skew Fields: Theory of General Division Rings}.
\newblock Cambridge University Press, Cambridge, 1995.
\bibitem{Davis1957}
C.~Davis.
\newblock All convex invariant functions of hermitian matrices.
\newblock {\em Arch. Math.}, 8:276--278, 1957.
\bibitem{Derksen2022ANT}
H.~Derksen.
\newblock The {$G$}-stable rank for tensors and the cap set problem.
\newblock {\em Algebra Number Theory}, 16(5):1071--1097, 2022.
\bibitem{FortinReutenauer04}
M.~Fortin and C.~Reutenauer.
\newblock Commutative/non-commutative rank of linear matrices and subspaces of
matrices of low rank.
\newblock {\em S\'em. Lothar. Combin.}, 52:B52f, 2004.
\bibitem{GGOW}
A.~Garg, L.~Gurvits, R.~Oliveira, and A.~Wigderson.
\newblock Operator scaling: theory and applications.
\newblock {\em Found. Comput. Math.}, 20(2):223--290, 2020.
\bibitem{GRS_Book}
V.~Georgoulas, J.~W. Robbin, and D.~A. Salamon.
\newblock {\em The Moment-Weight Inequality and the {H}ilbert-{M}umford
Criterion---{GIT} from the Differential Geometric Viewpoint}, volume 2297 of
{\em Lecture Notes in Mathematics}.
\newblock Springer, Cham, 2021.
\bibitem{GuilleminSternberg1982-I}
V.~Guillemin and S.~Sternberg.
\newblock Convexity properties of the moment mapping.
\newblock {\em Invent. Math.}, 67(3):491--513, 1982.
\bibitem{GuilleminSternberg1984-II}
V.~Guillemin and S.~Sternberg.
\newblock Convexity properties of the moment mapping. {II}.
\newblock {\em Invent. Math.}, 77(3):533--546, 1984.
\bibitem{Gurvits2004}
L.~Gurvits.
\newblock Classical complexity and quantum entanglement.
\newblock {\em J. Comput. System Sci.}, 69(3):448--484, 2004.
\bibitem{HamadaHirai2021}
M.~Hamada and H.~Hirai.
\newblock Computing the nc-rank via discrete convex optimization on {${\rm
CAT}(0)$} spaces.
\newblock {\em SIAM J. Appl. Algebra Geom.}, 5(3):455--478, 2021.
\bibitem{Hirai_Hadamard2022}
H.~Hirai.
\newblock Convex analysis on {H}adamard spaces and scaling problems.
\newblock {\em Found. Comput. Math.}, 24:1979--2016, 2024.
\bibitem{Hirai2025}
H.~Hirai.
\newblock A scaling characterization of nc-rank via unbounded gradient flow.
\newblock {\em Linear Algebra Appl.}, 730:525--545, 2026.
\bibitem{HiraiSakabe2024FOCS}
H.~Hirai and K.~Sakabe.
\newblock Gradient descent for unbounded convex functions on {H}adamard
manifolds and its applications to scaling problems.
\newblock In {\em 65th {IEEE} Annual Symposium on Foundations of Computer
Science, {FOCS} 2024}, pages 2387--2402. {IEEE}, 2024.
\newblock {\tt arXiv:2404.09746}.
\bibitem{IQS2017}
G.~Ivanyos, Y.~Qiao, and K.~V. Subrahmanyam.
\newblock Non-commutative {E}dmonds' problem and matrix semi-invariants.
\newblock {\em Comput. Complex.}, 26(3):717--763, 2017.
\bibitem{IQS2018}
G.~Ivanyos, Y.~Qiao, and K.~V. Subrahmanyam.
\newblock Constructive non-commutative rank computation is in deterministic
polynomial time.
\newblock {\em Comput. Complex.}, 27(4):561--593, 2018.
\bibitem{KLM2009JDG}
M.~Kapovich, B.~Leeb, and J.~Millson.
\newblock Convex functions on symmetric spaces, side lengths of polygons and
the stability inequalities for weighted configurations at infinity.
\newblock {\em J. Differential Geom.}, 81(2):297--354, 2009.
\bibitem{Kirwan1984}
F.~Kirwan.
\newblock Convexity properties of the moment mapping. {III}.
\newblock {\em Invent. Math.}, 77(3):547--552, 1984.
\bibitem{Landsberg_tensors}
J.~M. Landsberg.
\newblock {\em Tensors: {G}eometry and {A}pplications}.
\newblock American Mathematical Society, Providence, RI, 2012.
\bibitem{Landsberg_tensors2019}
J.~M. Landsberg.
\newblock {\em Tensors: {A}symptotic {G}eometry and {D}evelopments 2016--2018}.
\newblock American Mathematical Society, Providence, RI, 2019.
\bibitem{Lee2022momentmapconvexfunction}
K.~L. Lee, J.~Sturm, and X.~Wang.
\newblock Moment map, convex function and extremal point, 2022.
\newblock arXiv:2208.03724.
\bibitem{Lewis1996}
A.~S. Lewis.
\newblock Convex analysis on the {H}ermitian matrices.
\newblock {\em SIAM J. Optim.}, 6:164--177, 1996.
\bibitem{Rockafellar}
R.~T. Rockafellar.
\newblock {\em Convex Analysis}.
\newblock Princeton University Press, Princeton, NJ, 1970.
\bibitem{RockafellarWets}
R.~T. Rockafellar and R.~J-B. Wets.
\newblock {\em Variational Analysis}.
\newblock Springer-Verlag Berlin, Heidelberg, 1998.
\bibitem{Sakai1996}
T.~Sakai.
\newblock {\em Riemannian Geometry}.
\newblock American Mathematical Society, Providence, RI, 1996.
\bibitem{Strassen1988_asymptotic_spectrum}
V.~Strassen.
\newblock The asymptotic spectrum of tensors.
\newblock {\em J. Reine Angew. Math.}, 384:102--152, 1988.
\bibitem{vdBCLNWZ_STOC2025}
M.~van~den Berg, M.~Christandl, V.~Lysikov, H.~Nieuwboer, M.~Walter, and
J.~Zuiddam.
\newblock Computing moment polytopes of tensors, with applications in algebraic
complexity and quantum information.
\newblock In {\em S{TOC}'25---{P}roceedings of the 57th {A}nnual {ACM}
{S}ymposium on {T}heory of {C}omputing}, pages 756--765. ACM, 2025.
\newblock arXiv:2510.08336.
\bibitem{WDGC2013_Science}
M.~Walter, B.~Doran, D.~Gross, and M.~Christandl.
\newblock Entanglement polytopes: Multiparticle entanglement from
single-particle information.
\newblock {\em Science}, 340(6137):1205--1208, 2013.
\bibitem{WigdersonZuiddam2023}
A.~Wigderson and J.~Zuiddam.
\newblock Asymtotic spectra: {T}heory, applications, and extensions, 2023.
\newblock {\tt https://staff.fnwi.uva.nl/j.zuiddam/papers/convexity.pdf}.
\end{thebibliography}