On matrix polynomials and the joint spectral radius over max-algebras
Extracted source text · arXiv 2605.27302
\documentclass[12pt, reqno]{amsart}
\usepackage{amsmath, amsthm, amscd, amsfonts, amssymb, mathtools, color, hyperref}
\usepackage{geometry}
\usepackage[dvipsnames]{xcolor}
\geometry{a4paper, left=3cm, right = 3cm, top=3.5cm, bottom=3.5cm}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{observation}[theorem]{Observation}
\newtheorem{fact}[theorem]{Fact}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\numberwithin{equation}{section}
\newcommand{\N}{\mathbb{N}}
\renewcommand{\baselinestretch}{1.5}
\title[Matrix polynomials and joint spectral radius] {On matrix polynomials and the joint spectral radius over max-algebras}
\author[Askar]{Askar Ali M$^{1}$}
\address{$^{1}$ Azim Premji University, Bhopal, India}
\email{askar.m@apu.edu.in}
\author[Sachindranath]{Sachindranath Jayaraman$^{2,\ast}$}
\address{$^{2}$ School of Mathematics, IISER Thiruvananthapuram, India}
\email{sachindranathj@iisertvm.ac.in, sachindranathj@gmail.com}
\thanks{$^\ast$Corresponding author}
\subjclass[2020]{Primary: 15A80 , 11B83, 15A18 Secondary: 37C25 }
\keywords{Max-algebras; matrix polynomial; joint spectral radius; matrix norm; periodic points; common eigenvectors}
\begin{document}
\noindent
\begin{abstract}
Our aim is to study matrix polynomials over max-algebras and their growth in terms of max-induced seminorms.
In particular, we compare the set growth of a bounded family $\Psi$ of matrix polynomials, measured in terms of the
seminorms $\eta_{\|\cdot\|}$ and $\hat{\eta}_{\|\cdot\|}$ with the induced joint spectral radius of the coefficient pool
$\Psi_0$ of the matrix polynomials. Dynamics of max-linear maps and convergence to periodic points under a single
joint spectral radius condition and the existence of common max-eigenvectors of the coefficient pool are also brought out.
\end{abstract}
\maketitle
\section{Introduction}
We work throughout over the field $\mathbb{R}$ of real numbers. By a max-algebra, we mean the triple $(\mathbb{R}_+,\oplus,\otimes)$, where $\mathbb{R}_+$ is the set of all nonnegative real numbers with the binary
operations $a \oplus b = \max\{a,b\}$ and $a \otimes b = ab$. Max-algebras provide an idempotent linear-algebraic
language for a broad class of nonlinear problems arising in discrete-event systems, scheduling, optimization and performance evaluation. This setting is isomorphic (through the exponential map) to the max-plus model commonly used in discrete-event
dynamics. Excellent references on this subject are the monographs \cite{Butkovic, Economics-note, Max@work}.
We shall denote by $M_n(\mathbb{R}_+)$ the collection of all $n \times n$ matrices with entries from the max-algebra
described in the previous paragraph. Given any two such matrices $A$ and $B$, their matrix product (denoted by $AB$)
is defined by $(A B)_{ij} = \displaystyle \max_{k} (a_{ik} . b_{kj})$ (this also includes the product $Ax$, when $x$ is a
vector). We shall also denote by $GL_n(\mathbb{R}_+)$ the set of all invertible matrices in a max algebra. It is easy to
prove that an element of $GL_n(\mathbb{R}_+)$ is necessarily a generalized permutation matrix - one that is a product
of a diagonal matrix and a permutation matrix.
Spectral analysis of max-algebra matrices plays a key role in several scenarios, such as discrete event dynamical systems,
stability analysis, scheduling problems and so on. We refer the readers to \cite{Butkovic, Gursoy-Mason, Neda-Ghasemi, DEDS-dynamic-product, Lur, Lur2005, Muller-Peperko}. Beyond single matrices, many problems lead naturally to matrix polynomials over max-algebras. Given matrices $A_0,\dots,A_{m-1} \in M_n(\mathbb{R}_+)$, a max-matrix polynomial
is an expression of the form $P(\lambda):= A_0 \oplus \lambda A_1 \oplus \cdots \oplus \lambda^{m-1}A_{m-1},\
\lambda\in \mathbb{R}_+$. Spectral properties for such polynomials have been investigated recently in
\cite{Gursoy-Mason, Muller-Peperko-0} and \cite{Neda-Ghasemi}. The asymptotic growth of a product of matrices is
measured by the joint spectral radius (JSR). Rota and Strang laid the foundation to this approach (see \cite{Rota-Strang})
and is now a standard reference in areas such as switching systems, wavelets and control \cite{Heil-Strang}. In \cite{Lur},
Lur initiated a study of max-algebraic version of the generalized or joint spectral radius theory. For a given norm
$||.||$ in $\mathbb{R}^n$ and $A\in M_n(\mathbb{R}_+)$, a seminorm, as introduced by Lur, is defined by
$$\eta_{\|\cdot\|}(A):= \displaystyle \sup \Bigl\{\frac{||Ax||}{||x||}: 0 \neq x \in \mathbb{R}_+ \Bigr\}.$$
For a bounded family of matrices $\Sigma$, one can then define (as done in \cite{Lur}) the joint spectral radius built
from $\eta$ by
$$\rho_\eta(\Sigma):= \displaystyle \limsup_{k\to\infty}\Big(\ \sup_{A_1,\dots,A_k \in \Sigma}
\eta_{\|\cdot\|}(A_k \dots A_1)\ \Big)^{1/k}.$$
In \cite{Lur2005}, the author introduced the $\hat{\eta}$-norm for a matrix $A \in M_n (\mathbb{R}_+)$ by
$$\hat{\eta}_{\|\cdot \|}(A) = \displaystyle \limsup_{k\to\infty} (\eta_{\|\cdot \|}(A^k))^{1/k}.$$
The set growth of a bounded family $\Sigma$ built from $\hat{\eta}$ is then given by
$$\hat{\eta}_{\|\cdot \|}(\Sigma) = \limsup_{k\to\infty} (\sup_{A \in \Sigma}\ \eta_{\|\cdot \|}(A^k))^{1/k}.$$
This notion of set growth accounts for asymptotic stability of matrices in max-algebras. Motivated by these approaches,
Muller and Peperko \cite{Muller-Peperko-0} developed a detailed study of the joint spectral radius in max-algebras.
Later on, the authors in \cite{Neda-Ghasemi}, extended these notions to matrix polynomials
with a view to study various spectral properties and norm inequalities in max-algebras. As Muller and Peperko point out,
there is an effective calculation of the joint spectral radius in max-algebras (for bounded families), in terms of the maximum
cycle geometric mean (see Proposition $2.1$ of \cite{Muller-Peperko-0}). We shall exploit this to illustrate our results
by means of examples.
The purpose of this article is to connect the above notions and define the joint-spectral-radius of a family of
max-matrix polynomials, in terms of the collection of coefficient matrices. In particular, given a bounded collection $\Psi$ of max-matrix polynomials (see Definition \ref{JSR-matrix-polynomials}),
let us associate to it the coefficient pool $\Psi_0\subset M_n(\mathbb{R}_+)$ obtained by collecting all coefficient matrices appearing in elements of $\Psi$. Then, using the above defined max-induced seminorm $\eta_{\|\cdot\|}$,
we define a set growth rate for $\Psi$ under polynomial multiplication and compare
it to the max-algebraic joint spectral radius of $\Psi_0$. This provides a quantitative bridge between polynomial
algebra on the one hand and switched max-linear dynamics on the other hand.
We summarize our results now. We begin with a single max-matrix polynomial \(P\) and show that the joint spectral radius
of the coefficient-set sits between the norms of the polynomial built from \(\hat{\eta}\) and \(\eta\) respectively (Theorem \ref{thm:etacap-rho-eta-inequality}). We then extend this idea for a bounded family \(\Psi\) of matrix polynomials, and we
prove two-sided bounds comparing the set growth of \(\Psi\) with the joint spectral radius of the coefficient pool \(\Psi_0\),
up to a factor depending only on the maximal degree in \(\Psi\) (Theorem \ref{thm:rho-eta-rho-inequality}). These inequalities sharpen the idea that the asymptotic growth of polynomial products is controlled by the growth of the matrices appearing as coefficients. When the coefficient pool \(\Psi_0\) is simultaneously triangularizable in the max-algebraic sense
(through a common permutation similarity), we derive an explicit formula for the joint spectral radius of \(\Psi_0\) as the maximum of the joint spectral radii of the induced one-dimensional diagonal pools (Theorem \ref{thm:SJR-triangular-family}). These results extend and complement the spectral descriptions of max-matrix polynomials in
\cite{Gursoy-Mason, Neda-Ghasemi} by linking them to the joint-growth quantities.
In Section \ref{Sec:com-eigenvector}, we examine common eigenvectors and periodic points of a bounded family of matrices
(or correspondingly, polynomials with coefficients from this bounded family), and its connection with the joint spectral radius. Periodic points arise naturally in the dynamics of cone-preserving maps, and in particular in the iteration of a nonnegative
matrix acting on $\mathbb{R}^n_+$. A classical consequence of nonlinear Perron--Frobenius theory (see for instance
Theorem $B.4.7$ of \cite{Lemmens-Nussbaum}) states that if a nonnegative matrix $A$ has spectral radius at most one,
then along a suitable arithmetic subsequence of iterates, every bounded orbit converges to a periodic point: there exists
$q\geq 1$ such that for each $x$ with $\{A^k x\}_{k\ge 1}$ bounded,
$\lim_{k\to\infty}A^{kq}x=\xi_x$, where $\xi_x$ is a periodic point of $A$ and the period of $\xi_x$ dividing $q$.
This viewpoint motivates studying convergence or limit sets through periodic points rather than
through pointwise convergence of $A^k x$. In applications, one is often led not to a single matrix, but to products drawn from
a finite collection $\{A_1,\dots,A_N\}$, indexed by a word $\omega$ (finite or infinite). In the Euclidean (standard-algebra) setting, S. Jayaraman {\it et. al.} \cite{Extractra-2022} prove that if each generator has spectral radius $1$ and the collection admits a nontrivial set of common eigenvectors $E$, then for each finite word $\omega$ and each initial condition
$x$ in the linear span $\mathcal{LC}(E)$, there exists an integer $q_\omega\ge 1$ such that
$\lim_{k\to\infty}A_\omega^{kq}x=\xi(x,\omega)$, where $\xi(x,\omega)$ is a periodic point
of the word product $A_\omega$ (with period dividing $q$). Moreover, once the word contains
all generators (in the sense that each symbol appears at least once), the integer $q_\omega$
and the limiting periodic point become independent of the particular choice of such a word. Recent work in the max-algebra
setting extends periodic-point phenomena from a single matrix to words under generator-wise cycle-mean bounds \cite{DEDS-dynamic-product}. In the present paper we strengthen this by replacing individual spectral constraints on the generators with a single global stability condition expressed in terms of the max-algebraic joint spectral radius. Assuming that a finite set \(\Sigma\subset M_n(\mathbb{R}_+)\) has max-algebraic joint spectral radius at most one and
admits common max-eigenvectors associated to its max-eigenvalues, we prove convergence of iterates along words containing
each generator at least once and the limit being a periodic point (Theorem \ref{thm:JSR-common-eig-periodic}). As a
consequence, we obtain fixed points for polynomial evaluations \(x\mapsto P(1)\otimes x\) when the coefficients of \(P\)
belong to \(\Sigma\) (Corollary \ref{cor:com-eigenvector-polynomial}) as well as asymptotic stability (convergence to \(0\))
when the joint spectral radius is strictly less than one (Corollary \ref{cor:JSR-strict-to-zero}). These statements connect
algebraic bounds of Section $3$ with the asymptotic product dynamics studied in \cite{DEDS-dynamic-product}.
\section{Preliminaries}
We shall collect some of the basic definitions and preliminary notions needed further. These may be found in
\cite{Butkovic} and \cite{Gursoy-Mason}. We begin with the notion of a digraph associated to $A \in M_n(\mathbb{R}_+)$.
\begin{definition}\label{def:digraph}
For $A=(a_{ij})\in M_n(\mathbb{R}_+)$, the directed graph $G_A$ associated to $A$ has vertices $\{1,\dots,n\}$ with
an edge $i \to j$ of weight $a_{ij}$ if and only if $a_{ij} > 0$.
\end{definition}
A (simple) circuit $C$ has weight $w(C) = \prod a_{i_\ell i_{\ell+1}}$; the cycle mean and the
maximum cycle mean of $A$ are then given by $\mu(C) = w(C)^{1/|C|}$ and
$\mu(A) = \max\{\mu(C):C\subseteq G_A\}$ respectively. The maximum cycle mean $\mu(A)$ (also known as the
Perron root) plays the role of the spectral radius in max-algebras. The A (right) max–eigenpair satisfies
$Ax = \lambda x$ with $\lambda \in \mathbb{R}_+$ being a max–eigenvalue. It is known that $\mu(A)$
is the largest max-eigenvalue of $A$. Moreover, if $A$ is irreducible, then $\mu(A)$ is the unique max-eigenvalue and every
max-eigenvector is positive (see for instance Theorem $2$ of \cite{Bapat}). Moreover, the max version of Gelfand formula
holds; that is, $\mu(A) = \displaystyle \lim_{m \to \infty} ||A^m||^{1/m}$ for any arbitrary matrix norm $||.||$ on
$M_n(\mathbb{R})$.
Let us now briefly point out the Frobenius normal form of a matrix. We will have an occasion to use it in a later section.
\begin{definition}\label{Frobenius-normal-form}
A matrix $A\in M_n(\mathbb{R}_+)$ is in \emph{Frobenius normal form} (FNF) if there exists $P \in GL_n(\mathbb{R}_+)$
such that
\[
P^{-1}AP=
\begin{bmatrix}
A_{11} & A_{12} & \cdots & A_{1t}\\
0 & A_{22} & \cdots & A_{2t}\\
\vdots & \ddots & \ddots & \vdots\\
0 & \cdots & 0 & A_{tt}
\end{bmatrix},
\]
where each diagonal block $A_{ii}$ is irreducible. The blocks correspond to the strongly connected components of $G_A$.
When every block is $1\times 1$, this reduces to an upper triangular form.
\end{definition}
Any reducible matrix can be written as a block upper triangular matrix via a permutation matrix, as above
\cite{Bapat-Raghavan}. The Frobenius normal form is a max-algebraic analogue of block triangularization in classical
nonnegative matrix theory. This decomposition plays a fundamental role in the spectral theory of max-algebraic matrices,
as the Perron roots and eigenvectors can be analyzed blockwise through the Frobenius structure (see for instance
Theorem $4.8$ of \cite{Katz-Schneider-Sergeev}).
\begin{definition}\label{def:triangular}
$A \in M_n(\mathbb{R}^n_+)$ is said to be triangularizable if $P^{-1} A P$ is upper triangular for some
$P \in GL_n(\mathbb{R}_+)$ (equivalently, for some permutation matrix $P$). A family $\{A_1, \dots, A_n\}$ is said
to be simultaneously triangularizable if there exists a $P \in GL_n(\mathbb{R}_+)$ such that $P^{-1}A_iP$ (once again,
for some permutation matrix $P$) is upper triangular for all $i = 1, \dots, n$.
\end{definition}
In a recent work \cite{Askar-Jayaraman-Himadri}, we have obtained several interesting results on simultaneous
triangularization of a family in the max-algebraic setting, some of which we shall make use of in a later section.
\begin{definition}
For a fixed $m\in\mathbb{N}$, a max–matrix polynomial of degree $m-1$ is an expression of the form
$P(\lambda) = A_0 \oplus \lambda A_1 \oplus \dots \oplus \lambda^{m-1}A_{m-1}, \ A_j\in M_n(\mathbb{R}_+), \ \lambda\in\mathbb{R}_+$.
\end{definition}
The (right) max–spectrum $\sigma_m[P]$ of $P(\lambda)$ consists of $k \in \mathbb{R}_+$ for which
$P(k) v = k^m v$ with $v \neq 0$. As in the classical set up, one can associate a block companion matrix $C_P$
of order $mn$ to $P(\lambda)$ as follows:
$C_P=\begin{bmatrix}
0&I&\cdots&0&0\\
&\ddots&\ddots&\vdots&\vdots\\
&&\ddots&I&0\\
A_0&A_1&\cdots&A_{m-2}&A_{m-1}
\end{bmatrix}$.
One can then easily verify that $\sigma_m[P] = \sigma_m(C_P)$ (see for instance \cite{Neda-Ghasemi}). The following
remark is worth pointing out and we shall have an occasion to make use of it later on.
\begin{remark}\label{rem:irreducible-Cp}
For a matrix polynomial $P(\lambda)$, assume that each of the coefficient matrices $A_j$ are irreducible. Then, the
corresponding block companion matrix $C_P$ is also irreducible.
\end{remark}
The following definitions and discussions are from \cite{Neda-Ghasemi}. We have stated these here for the sake of
completeness.
\smallskip
A vector norm $||.||$ on $\mathbb{R}^n$ is monotone if $0 \leq x \leq y \Rightarrow \|x\| \leq \|y\|$.
Given a monotone norm $\|\cdot\|$ and $A\in M_n(\mathbb{R}_+)$, one can define a seminorm (see for instance
\cite{Lur}) on $M_n(\mathbb{R}_+)$ by
$$\eta_{||.||}(A):= \displaystyle \sup \biggl\{\frac{||Ax||}{||x||}: 0 \neq x \in \mathbb{R}^n_+ \biggr\}$$
This seminorm is subadditive with respect to $\oplus$ and submultiplicative with respect to $\otimes$; that is,
$\eta_{\|\cdot\|}(X\oplus Y) \leq \eta_{\|\cdot\|}(X) + \eta_{\|\cdot\|}(Y), \
\eta_{\|\cdot\|}(XY) \leq \eta_{\|\cdot\|}(X)\eta_{\|\cdot\|}(Y)$.
\begin{definition}\label{Eta-norm of a polynomial}
For a polynomial \(P(\lambda ) = \displaystyle \sum_{j=0}^{m-1}\lambda^j A_j\),
\[
\eta_{\|\cdot\|}[P]:= \displaystyle \max_{0\leq j\leq m-1}\eta_{\|\cdot\|}(A_j).
\]
\end{definition}
For a polynomial $P(\lambda) = A_0 \oplus \lambda A_1 \oplus \cdots \oplus \lambda^{m-1}A_{m-1}$, we denote by
$\Sigma_P = \{A_0, A_1, \cdots, A_{m-1}\}$, the collection of all coefficient matrices. With the max-algebra operations, there
is an obvious way to define the sum and product of two max-matrix polynomials. It is clear that the set $V$ of all max-matrix
polynomials is a real vector space under $\oplus$. Given any norm $||.||$ on $\mathbb{R}^n$ and any $P(\lambda) =
\displaystyle \sum_{j=0}^{m-1} \lambda^{j} A_j \in V$, one can define
$$||P(\lambda)||:= \max\{||A_0||, \cdots, ||A_{m-1}||\},$$ where $||A_j||$ is the induced matrix norm from the
vector norm $||.||$ on $\mathbb{R}^n$. It is easy to verify that $V$ is a real normed linear space with respect to the
above norm and that a subset $\Psi \subset V$ is bounded if and only if there exists an $M > 0$ such that
$||P(\lambda)|| \leq M$ for all $P(\lambda) \in \Psi$. It also turns out as a consequence of Proposition $2.7$ of
\cite{Neda-Ghasemi} that $\Psi$ is bounded if and only if there is an $M > 0$ such that $\eta_{\|\cdot\|}[P] \leq M$
for all $P(\lambda) \in \Psi$. Before proceeding further, let us mention that we shall denote by $\Psi^{k}$, the set of
all $k$-fold products of $P(\lambda)$ from the set $\Psi$. Similarly, we shall denote by $\Sigma_P^{k}$, the $k$-fold product
of coefficient matrices coming from the set $\Sigma_P$, corresponding to a matrix polynomial $P(\lambda)$.
\begin{definition}\label{JSR-matrix-polynomials}
Let \(\Psi\) be a bounded set of max–matrix polynomials.
Define the set growth of \(\Psi\) by
\begin{equation}\label{Eta-norm-Psi}
\eta_{\|\cdot\|}(\Psi):=\displaystyle \limsup_{k\to\infty}\Big(\ \sup_{P\in\Psi^{k}} \eta_{\|\cdot\|}[P]\ \Big)^{1/k}.
\end{equation}
If \(\Psi_0:= \displaystyle \bigcup_{P\in\Psi} \Sigma_P\) is the coefficient pool, the joint spectral radius built from
\(\eta\) is defined by
\begin{equation}\label{Rho-eta-Psi_0}
\rho_\eta(\Psi_0):= \displaystyle \limsup_{k\to\infty}\Big(\sup_{A_1,\dots,A_k\in\Psi_0} \eta_{\|\cdot\|}(A_1 \cdots A_k)\Big)^{1/k}.
\end{equation}
\end{definition}
The above definition of $\rho_\eta(.)$ is the same as the definition of the joint spectral radius of a collection of matrices,
except we have used the $\eta_{||.||}$.
\begin{definition}\label{def:eta-hat-norm}
For a matrix $A \in M_n (\mathbb{R}_+)$, the $\hat{\eta}$-norm is defined as
\begin{equation}\label{Eta-hat-norm-matrix}
\hat{\eta}_{\|\cdot \|}(A):= \displaystyle \limsup_{k\to\infty} (\eta_{\|\cdot \|}(A^k))^{1/k}.
\end{equation}
For a polynomial $P(\lambda) = A_0 \oplus \lambda A_1 \oplus \cdots \oplus \lambda^{m-1}A_{m-1}$,
\begin{equation}\label{Eta-hat-norm-polynomial}
\hat{\eta}_{\|\cdot \|}(P) = \max_j\ \hat{\eta}_{\|\cdot \|}(A_j).
\end{equation}
\end{definition}
Finally, the set growth of a bounded $\Psi$ built from \(\hat{\eta}\) is
\begin{equation}\label{Set-growth-from-eta-hat}
\hat{\eta}_{\|\cdot \|}(\Psi) = \limsup_{k\to\infty} (\sup_{P \in \Psi}\ \eta_{\|\cdot \|}(P^k))^{1/k}.
\end{equation}
Our objective is to compare the set growth of a bounded family $\Psi$ of matrix polynomials, measured in terms of
the seminorms $\eta_{\|\cdot\|}$ and $\hat{\eta}_{\|\cdot\|}$ as defined in Equations \eqref{Eta-norm-Psi} and \eqref{Set-growth-from-eta-hat} respectively, with the joint spectral radius of the coefficient pool $\Psi_0$. Since the set
growth is defined in terms of $\eta_{\|\cdot\|}$, it is natural to define the joint spectral radius of $\Psi_0$ via the same
seminorm, as in Equation \eqref{Rho-eta-Psi_0}, so that the two quantities are directly comparable. We note that this choice
of norm does not affect the value of the joint spectral radius, since all monotone norms on $\mathbb{R}^n$ are equivalent
in finite dimensions.
The following discussion is worth pointing out. Given a bounded subset $\Sigma$ of $M_n(\mathbb{R}_+)$,
the max-algebraic version of the generalized spectral radius $\mu(\Sigma$) of $\Sigma$, is defined as
$$\mu(\Sigma) = \displaystyle \limsup_{m \to \infty} \big[\sup_{A \in \Sigma^{m}} \mu(A) \big]^{1/m},$$ and is equal
to the joint spectral radius of $\Sigma$,
$$\rho(\Sigma) = \displaystyle \limsup_{m \to \infty} \big[\sup_{A \in \Sigma^{m}} ||A|| \big]^{1/m}.$$ We state this as
a theorem below and will have an occasion to use it later on.
\begin{theorem}\label{Determining the JSR}
Let $\Sigma \subset M_n(\mathbb{R}_+)$ be a bounded set. Then, the following hold.
\begin{enumerate}
\item (Proposition $2.1$, \cite{Muller-Peperko-0}) $\mu(\Sigma) = \mu(S(\Sigma))$, where $S(\Sigma) =
\displaystyle \bigoplus_{A \in \Sigma} A$.
\item (Corollary $2.3$, \cite{Muller-Peperko-0}) $\mu(\Sigma) = \rho(\Sigma)$.
\end{enumerate}
\end{theorem}
Statement $2$ in the above theorem is known as the Berger-Wang formula.
\section{Main Results}
The main results are presented in this section. We subdivide this section into three subsections for ease of reading. We begin
with a subsection, where we prove some norm inequalities connected to the joint spectral radius.
\subsection{Joint spectral radius and norm inequalities}\label{Sec:JSR-norm-inequalities}\hspace*{0.5cm}
We bring out norm-inequalities for a polynomial and a bounded collection of polynomials that connects
the joint spectral radius and the set growth built from $\eta$ and $\hat{\eta}$ in this section. Our first result is the following.
\begin{theorem}\label{thm:etacap-rho-eta-inequality}
Let $P(\lambda) = A_0 \oplus \lambda A_1 \oplus \cdots \oplus \lambda^{m-1}A_{m-1}$ be a matrix polynomial.
Then, $\hat{\eta}_{\| \cdot \|}(P) \leq \rho_\eta(\Sigma_P) \leq \eta_{\| \cdot \|} (P)$.
\end{theorem}
\begin{proof}
By Lemma $1 (V)$ of \cite{Lur}, we have, $\eta_{\| \cdot \|} (A B) \leq \eta_{\| \cdot \|} (A) \cdot \eta_{\| \cdot \|} (B)$.
Then,
\begin{equation*}
\eta_{\| \cdot \|} (A_{i_1} A_{i_2} \cdots A_{i_k}) \leq
\displaystyle \prod_{j=1}^{k} \eta_{\| \cdot \|} (A_{i_j}) \leq ( \max_j \eta_{\| \cdot \|} (A_{i_j})) =
(\eta_{\| \cdot \|} (P))^k .
\end{equation*}
By taking the $k^{th}$ root and then the supremum over $\Sigma_P^{k}$,
we get
\[\rho_\eta(\Sigma_P) \leq \eta_{\| \cdot \|} (P).\]
To prove the lower bound, observe that for each $j$,
\begin{equation*}
\begin{aligned}
\rho_\eta (\Sigma_P) & = \displaystyle \limsup_{k \to \infty} (\sup_{\Sigma_P}
( \eta_{\| \cdot \|} (A_{i_1} A_{i_2} \cdots A_{i_k})))^{1/k}\\
&\geq \limsup_{k \to \infty} ( \eta_{\| \cdot \|} (A_{i_j}^k )^{1/k}\\
& = \hat{\eta}_{\| \cdot \|}(A_{i_j}).
\end{aligned}
\end{equation*}
\end{proof}
We now move on to bounded families of matrix polynomials.
\begin{theorem}\label{thm:rho-eta-rho-inequality}
Let $\Psi$ be a bounded family of matrix polynomials. Then,
\[\rho_{\eta} (\Psi_0) \leq \eta_{\| \cdot \|} (\Psi) \leq m \rho_{\eta} (\Psi_0),\]
where $m = \displaystyle \max_{P \in \Psi} \deg(P)$.
\end{theorem}
\begin{proof}
Let $P_i(\lambda)= A_{0}^{(i)} \oplus \lambda A_{1}^{(i)} \oplus \cdots \oplus \lambda^{m_i-1}A_{{{m_i-1}}}^{(i)}$.
It is easy to verify that
\[P_1(\lambda) P_2(\lambda) \cdots P_k (\lambda) = \displaystyle
\bigoplus_{ s=0}^{m_1 +\cdots +m_k} \lambda^s C_s,\] where
$C_s = \displaystyle \bigoplus_{i_1+ \cdots + i_k=s, 0\leq i_j \leq m-1}
(A_{i_1}^{(1)} A_{i_2}^{(2)} \cdots A_{i_k}^{(k)})$. Then, one of the coefficients $C_s$ contains
the term $A_1 A_2 \cdots A_k$. For this coefficient,
\begin{equation*}
\begin{aligned}
C_s &\geq A_1 A_2 \cdots A_k.\\
\text{Thus, } \eta_{\|.\|}(C_s) &\geq \eta_{\|.\|}(A_1 A_2 \cdots A_k).
\end{aligned}
\end{equation*}
We also have,
\[\eta_{\|.\|} (P_1 P_2 \cdots P_k) = \max_s \eta_{\|.\|}(C_s) \geq \eta_{\|.\|} (A_1 A_2 \cdots A_k).\]
By taking the supremum over $P_i \in \Psi$, and $A_i \in \Psi_0$, the above inequality becomes,
\[\sup_{P \in \Psi^{k}} \eta_{\|.\|}(P) \geq
\sup_{\Psi_0^{k}} (A_1 A_2 \cdots A_k).\]
By taking the $k^{th}$ root and then taking the $\limsup$, we get
\[\eta_{\|.\|}(\Psi) \geq \rho_\eta (\Psi_0).\]
To get the upper bound, by sub-additivity of the $\eta-$semi norm, we have
\begin{equation*}
\begin{aligned}
\eta_{\|.\|}(C_s)
& \leq \displaystyle \bigoplus_{i_1+ \cdots + i_k=s, 0\leq i_j \leq m-1} \eta_{\|.\|}
(A_{i_1}^{(1)} A_{i_2}^{(2)} \cdots A_{i_k}^{(k)})\\
&\leq |\{ i_1+ \cdots + i_k=s|\ 0\leq i_j \leq m-1\}| \sup \eta_{\|.\|}
(A_{i_1}^{(1)} A_{i_2}^{(2)} \cdots A_{i_k}^{(k)})\\
&\leq m^k \sup \eta_{\|.\|} (A_{i_1}^{(1)} A_{i_2}^{(2)} \cdots A_{i_k}^{(k)}).
\end{aligned}
\end{equation*}
From the above, we have
$\eta_{\|.\|} (P_1 \otimes P_2 \cdots P_k) \leq m^k \sup \eta_{\|.\|}(A_{i_1}^{(1)} A_{i_2}^{(2)} \cdots A_{i_k}^{(k)})$.
By taking the $k^{th}$ root, the supremum over $\Psi^{k}$, and then
the $\limsup$, we get
\[\eta_{\|.\|}(\Psi) \leq m\ \rho_\eta (\Psi_0).\]
We thus have $\rho_{\eta} (\Psi_0) \leq \eta_{\| \cdot \|} (\Psi) \leq m \rho_{\eta} (\Psi_0)$.
\end{proof}
\subsection{Triangularizable coefficient matrices}\hspace*{0.5cm}
We now consider the case when the coefficients of the matrix polynomial are simultaneously triangularizable.
\begin{definition}\label{def:perm-invariant-norm}
A norm $\| \cdot \|$ on $\mathbb{R}_+^n$ is called permutation invariant if $\| Px \| = \| x \|$, for any
$x \in \mathbb{R}_+^n$, and for any permutation matrix $P$.
\end{definition}
\begin{lemma}\label{lem:perm-norm}
If $\| \cdot \|$ is a permutation invariant norm, then the $\eta_{\| \cdot \|}$ induced from $\| \cdot \|$ satisfies the
following property. For any $A \in M_n(\mathbb{R}_+)$ and any permutation matrix $P$,
$\eta_{\| \cdot \|}(P^{-1}AP) = \eta_{\| \cdot \|}(A)$.
\end{lemma}
\begin{proof}
Observe that $\eta_{\| \cdot \|}(P^{-1}AP) = \sup_{x \neq 0} \frac{\|P^{-1}APx\|}{\|x\|}$. For $y=Px$, we have the following.
\begin{equation*}
\begin{aligned}
\eta_{\| \cdot \|}(P^{-1}AP)
& = \sup_{x \neq 0} = \frac{\|P^{-1}Ay\|}{\|P^{-1}y\|}\\
& = \sup_{x \neq 0} = \frac{\|Ay\|}{\|y\|}\\
& = \eta_{\|\cdot\|} (A).
\end{aligned}
\end{equation*}
\end{proof}
Let us now consider a matrix polynomial where the coefficient matrices are simultaneously triangularizable.
\begin{proposition}\label{prop:triangular-spectrum}
Let $P(\lambda) = A_0 \oplus \lambda A_1 \oplus \cdots \oplus \lambda^{m-1}A_{m-1}$ be a matrix polynomial, with
$\Sigma_P = \{A_0,A_1, \cdots, A_{m-1}\}$ being a simultaneously triangularizable family through a permutation matrix $S$.
Taking $A_i' = S^{-1}AS$ to be upper triangular, define the scalar polynomials
\[p_i(\lambda) = (A'_0)_{ii} \oplus \lambda (A'_1)_{ii}\oplus \cdots \oplus \lambda^{m-1}(A'_{m-1})_{ii}. \]
Then, $\sigma_m[P] = \displaystyle \bigcup_{i=1}^{n} \sigma_m[p_i]$.
\end{proposition}
\begin{proof}
By definition, \(k \in \sigma_m[P]\) if and only if there exists \(v\geq 0\), \(v\neq 0\), such that
\[
P(k) v = k^{m} v.\]
Similarly, for the scalar polynomial \(p_i\), we have \(k\in\sigma_m[p_i]\) if and only if \(p_i(k)=k^m\).
Set \(P'(\lambda):=S^{-1}P(\lambda)S = \displaystyle \bigoplus_{r=0}^{m-1} \lambda^r A'_r\).
Since each \(A'_r\) is upper triangular, so is
\[
P'(k)=A'_0 \oplus kA'_1 \oplus \cdots \oplus k^{m-1}A'_{m-1}. \]
The $i^{th}$ diagonal entry of \(P'(k)\) is exactly \(p_i(k)\). An upper triangular matrix is already a block
upper triangular matrix with $1 \times 1$ irreducible diagonal blocks, and hence is in Frobenius normal form (FNF).
These blocks are precisely the scalars \(p_i(k)\) on the diagonal. Then, by Theorem $4.5.4$ of \cite{Butkovic}, it follows
that the eigenvalues of all irreducible blocks are eigenvalues of the matrix $P'$ and hence of \(P\) by similarity. We thus
have $\sigma_m[P] = \displaystyle \bigcup_{i=1}^{n} \sigma_m[p_i]$.
\end{proof}
\begin{definition}\label{def:SJR-diagonal-entry}
Given a bounded family $\Psi$ of matrix polynomials such that the set $\Psi_0$ of all coefficient matrices is a simultaneously triangularizable family through a permutation matrix $S$, let $\Delta\Psi_0:= \{S^{-1}AS\ \mid A \in \Psi_0\}$ and
$\Delta\Psi_0^{(i)} := \{ (S^{-1}AS)_{ii}\ \mid A \in \Psi_0\}$. For this scalar pool, we define the joint spectral radius by
\[\rho_\eta (\Delta\Psi_0^{(i)}) =
\displaystyle \limsup_{k \to \infty} \big(\sup_{a_1,a_2,\cdots,a_k \in \Delta\Psi_0^{(i)}} a_1a_2\cdots a_k \big)^{1/k}. \]
\end{definition}
\begin{lemma}\label{lem:limsup}
Let $W \subset \mathbb{R}_+$ be a bounded set. Then,
\[\displaystyle \limsup_{k \to \infty} (\sup_{a_1,a_2, \cdots, a_k \in \Delta\Psi_0^{(i)}} a_1a_2 \cdots a_k)^{1/k} =
\displaystyle \sup_{a \in W}\ a.\]
\end{lemma}
\begin{proof}
Since $W \subset \mathbb{R}_+$ is bounded and nonempty, we may set $M := \displaystyle \sup_{a \in W} a \in (0,\infty)$.
For each $k \in \mathbb{N}$, define $P_k := \displaystyle \sup_{a_1,\dots,a_k \in W} \bigl(a_1 a_2 \cdots a_k\bigr)^{1/k}$.
We shall show that $\displaystyle \limsup_{k \to \infty} P_k = M$. To prove this, let us fix $k \in \mathbb{N}$ and choose arbitrary $a_1,\dots, a_k \in W$. By the very definition of $M$, we have $a_j \leq M$ for each $j=1,\dots,k$, so that
\[
a_1 a_2 \cdots a_k \;\le\; M^k,
\qquad\text{hence}\qquad
\bigl(a_1 a_2 \cdots a_k\bigr)^{1/k} \leq M.
\]
Taking the supremum over all $a_1,\dots, a_k \in W$ we see that $P_k \leq M \quad \text{for all } k$. Therefore
$\displaystyle \limsup_{k \to \infty} P_k \leq M$. If $\varepsilon > 0$ is arbitrary, then by the definition of $M$ there exists $a_\varepsilon \in W$ such that $a_\varepsilon > M - \varepsilon$. For each $k \in \mathbb{N}$, consider the choice
$a_1 = a_2 = \cdots = a_k = a_\varepsilon$. Then $\displaystyle \bigl(a_1 a_2 \cdots a_k\bigr)^{1/k}
= \bigl(a_\varepsilon^k\bigr)^{1/k} = a_\varepsilon > M - \varepsilon$. Since $P_k$ is the supremum over all such products,
we obtain $P_k \geq a_\varepsilon > M - \varepsilon \quad \text{for all } k$.
Hence, $\displaystyle \limsup_{k \to \infty} P_k \geq M - \varepsilon$. $\varepsilon>0$ being arbitrary, we conclude that $\displaystyle \limsup_{k \to \infty} P_k \geq M$. Thus, $\displaystyle \limsup_{k \to \infty} P_k = M = \sup_{a \in W} a$.
\end{proof}
We are now in a position to prove the main result of this section.
\begin{theorem}\label{thm:SJR-triangular-family}
Let $\Psi$ be a bounded family of matrix polynomials with $\Psi_0$ being a simultaneously triangularizable family. Assume further that $\|.\|$ is a permutation invariant monotone norm and $\eta$ is induced by this norm. Then,
$\rho_\eta(\Psi) = \displaystyle \max_i \rho_\eta(\Delta\Psi_0^{(i)})$.
\end{theorem}
\begin{proof}
Let the matrices in $\Psi_0$ be simultaneously triangularizable through a permutation matrix $S$. For each $A_i \in \Psi_0$, assume $B_i = S^{-1}A_i S$ is upper triangular. Then, using Lemma \ref{lem:perm-norm}, we have $\eta_{\|\cdot \|}(A) = \eta_{\|\cdot \|}(B) $. It therefore suffices to verify $\rho_\eta(\Psi)$ for $B_i \in \Delta\Psi_0$. Let,
$s:= \displaystyle \max_i \sup_{B \in \Delta\Psi } B_{ii}$ and $e_i$ be the $i^{th}$ standard basis vector. Then,
$|(B_1 B_2 \cdots B_k) e_i \| \geq (B_1 B_2 \cdots B_k)_{ii} \cdot \|e_i\|$.
This gives $\eta_{\|\cdot \|}(B_1 B_2 \cdots B_k) \geq (B_1 B_2 \cdots B_k)_{ii} = \prod_{r=1}^k (B_r)_{ii}.$
We then have, $\rho_\eta (\Delta\Psi) = \rho_\eta (\Psi) \geq \displaystyle \sup_{B \in \Delta\Psi } B_{ii}$.
Thus, $\rho_\eta (\Psi) \geq \displaystyle \max_i \sup_{B \in \Delta\Psi } B_{ii}= s$. Since $\Psi$ is a bounded set, we have $\displaystyle \sup_{B \in \Delta\Psi } B_{ij} < M$, for some $M > 0$. Therefore, each contributing term in
$(B_k \cdots B_1)_{ij}$ contains at most \(n-1\) off-diagonal factors, each bounded above by \(M\), and at least \(k-(n-1)\) diagonal factors, each bounded above by \(s\). Hence,
$(B_1 B_2 \cdots B_k)_{ij} \leq M^{n-1}s^{k-(n-1)} \leq Cs^k, \ \text{ with } C:=\max\{1,M^{n-1}s^{-(n-1)}\}$.
Since all norms on $\mathbb{R}_+^n$ are equivalent, there exists a $c > 0$ such that
$\eta_{\|\cdot\|} (B_1 B_2 \cdots B_k) \leq \displaystyle c \max_{i,j} (B_1 B_2 \cdots B_k)_{ij} \leq cCs^k$.
Taking the $k^{th}$ root and the $\limsup$, the above inequality becomes, $\rho_\eta (\Psi) \leq s$. Combining these two inequalities, we have, $\rho_\eta(\Psi) = \displaystyle \max_i \rho_\eta(\Delta\Psi_0^{(i)})$.
\end{proof}
We illustrate the above theorem by means of an example.
\begin{example}\label{eg-1}
Let $\Psi = \{P_1, P_2\}$ where $P_i(\lambda) = A_0^{(i)} \oplus \lambda A_1^{(i)}$ are max-matrix polynomials over $M_3(\mathbb{R}_+)$, with coefficient matrices
\[
A_0^{(1)} =
\begin{pmatrix}
3 & 0 & 2 \\
0 & 1 & 0 \\
0 & 2 & 4
\end{pmatrix},
\
A_1^{(1)} =
\begin{pmatrix}
2 & 3 & 1 \\
0 & 2 & 0 \\
0 & 1 & 3
\end{pmatrix},
\
A_0^{(2)} =
\begin{pmatrix}
1 & 2 & 3 \\
0 & 0 & 0 \\
0 & 1 & 2
\end{pmatrix},
\ \text{and } \
A_1^{(2)} =
\begin{pmatrix}
4 & 1 & 0 \\
0 & 1 & 0 \\
0 & 3 & 2
\end{pmatrix}.
\]
The coefficient pool is $\Psi_0 = \bigl\{A_0^{(1)},\, A_1^{(1)},\, A_0^{(2)},\,A_1^{(2)}\bigr\}$. Let $S$ be the permutation
matrix corresponding to the transposition $(2\;3)$; that is,
\[
S =
\begin{pmatrix}
1 & 0 & 0 \\
0 & 0 & 1 \\
0 & 1 & 0
\end{pmatrix} = S^{-1}.
\]
For any $A \in \Psi_0$, a direct computation gives $S^{-1}AS$ is upper triangular. Hence $\Psi_0$ is simultaneously triangularizable through $S$. Applying this to each element of $\Psi_0$, we obtain
\begin{align*}
S^{-1}A_0^{(1)}S &
= \begin{pmatrix}
3 & 2 & 0 \\
0 & 4 & 2 \\
0 & 0 & 1
\end{pmatrix},
&
S^{-1}A_1^{(1)}S & =
\begin{pmatrix}
2 & 1 & 3 \\
0 & 3 & 1 \\
0 & 0 & 2
\end{pmatrix},
\\[6pt]
S^{-1}A_0^{(2)}S & =
\begin{pmatrix}
1 & 3 & 2 \\
0 & 2 & 1 \\
0 & 0 & 0
\end{pmatrix},
&
S^{-1}A_1^{(2)}S & =
\begin{pmatrix}
4 & 0 & 1 \\
0 & 2 & 3 \\
0 & 0 & 1
\end{pmatrix}.
\end{align*}
Set $\Delta\Psi_0 = \{S^{-1}AS \mid A \in \Psi_0\}$. Reading off the diagonal entries at each position $i = 1,2,3$, the scalar
pools are,
\[
\Delta\Psi_0^{(1)} = \{3,\,2,\,1,\,4\},
\qquad
\Delta\Psi_0^{(2)} = \{4,\,3,\,2,\,2\},
\qquad
\Delta\Psi_0^{(3)} = \{1,\,2,\,0,\,1\}.
\]
By Lemma \ref{lem:limsup}, $\rho_\eta\!\left(\Delta\Psi_0^{(1)}\right) = \sup \Delta\Psi_0^{(1)} = 4,
\ \rho_\eta\!\left(\Delta\Psi_0^{(2)}\right) = \sup \Delta\Psi_0^{(2)} = 4,$ \\
$\rho_\eta\!\left(\Delta\Psi_0^{(3)}\right) = \sup \Delta\Psi_0^{(3)} = 2$.
We then infer from Theorem~\ref{thm:SJR-triangular-family} that,
\[
\rho_\eta(\Psi_0)
= \max_{i=1,2,3}\,\rho_\eta\!\left(\Delta\Psi_0^{(i)}\right)
= \max\{4,\,4,\,2\} = 4.
\]
One can verify this directly from statements $(1)$ and $(2)$ of Theorem \ref{Determining the JSR}, which actually gives
$\rho_\eta(\Psi_0) = \mu (A_0^{(1)}\oplus A_1^{(1)} \oplus A_0^{(2)}\oplus A_1^{(2)})$. We also have,
\[ A_0^{(1)}\oplus A_1^{(1)} \oplus A_0^{(2)}\oplus
A_1^{(2)} = \begin{pmatrix}
4&3&3\\
0&2&0\\
0&3&4
\end{pmatrix}.\]
Thus, $\mu (A_0^{(1)} \oplus A_1^{(1)} \oplus A_0^{(2)} \oplus A_1^{(2)}) = 4$, confirming $\rho_\eta(\Psi_0) = 4$.
\end{example}
\subsection{Common eigenvectors and periodic points}\label{Sec:com-eigenvector}\hspace*{0.5cm}
This is the last section of the paper where we investigate self maps on $\mathbb{R}_+^n$ of the form $Ax$, for some
$A\in M_n(\mathbb{R}_+)$ and $x \in \mathbb{R}_+^n$. For a bounded family of matrices $\Sigma$, and matrix
polynomials with coefficients from this family $\Sigma$, we examine the relation between common eigenvectors and
periodic points of the above maps with the joint spectral radius. More precisely, given $A\in M_n(\mathbb{R}_+)$, we
consider the max--linear self-map $T_A:\mathbb{R}_+^n\to\mathbb{R}_+^n,\ T_A(x)= Ax$. A vector $x\in\mathbb{R}_+^n$ is
called a fixed point of $A$ (or of $T_A$) if $Ax = x$. More generally, $x$ is called a periodic point of $A$ if there exists
$q\in\mathbb{N}$ such that $A^{q} x = x \ (\text{equivalently } T_A^{\,q}(x) = x)$.
The least such $q$ (when it exists) is called the period of $x$. Before proceeding further with our results, the following
is worth pointing out. In \cite{Butkovic-CuninghameGreen}, P. Butkovic and Cuninghame-Green discussed the periodic
behaviour of max-algebraic matrices. In particular, for an irreducible matrix $A$, there exist integers $p\geq 1$ and
$k_0$ such that $A^{k+p}=\lambda(A)^{p} A^k$ for all $k \geq k_0$, where $\lambda(A)$ is the max-algebraic eigenvalue (maximum cycle mean). The smallest such $p$ is called the period of $A$, and is denoted by $\mathrm{per}(A)$. A matrix
is called robust if every orbit eventually reaches an eigenvector of $A$; equivalently, $\mathrm{per}(A)=1$. This
single-matrix robustness condition can be seen as a precursor to our setting: where Butkovič and Cuninghame-Green require periodicity one for a single matrix to guarantee orbit convergence, we replace this matrix-wise condition with a single
global bound in terms of the joint spectral radius of the entire collection $\Sigma$, and characterize convergence to
periodic points for all word products simultaneously.
\begin{proposition}\label{prop:com-eigenvector-polynomial-eigenvalue}
Let $P(\lambda) = A_0 \oplus \lambda A_1 \oplus \cdots \oplus \lambda^{m-1}A_{m-1}$ be a matrix polynomial.
Assume that there exists a nontrivial vector $v \in \mathbb{R}^n_+$ such that $A_j v = \alpha_jv, \text{ for }
j=0,1,\cdots,m-1$. Define $q(\lambda) = \alpha_0 \oplus \lambda \alpha_1 \oplus \cdots \oplus \lambda^{m-1}\alpha_{m-1}$.
If $q(k) = k^m$ for some $k \geq 0$, then, $P(k)v = k^m v, \ \text{ equivalently, }\ k \in \sigma_m[P]$.
\end{proposition}
\begin{proof}
Using max–linearity and the hypothesis, we have
$P(k) v = \displaystyle \bigoplus_{j=0}^{m-1} k^j (A_j v)
= \displaystyle \bigoplus_{j=0}^{m-1} (k^j \alpha_j)\, v
= \displaystyle \big(q(k)\big)\, v
= k^m v$.
\end{proof}
The following proposition is worth pointing out.
\begin{proposition}\label{prop:eigen-Cp-1}
Let $P(\lambda) = A_0\oplus \lambda A_1\oplus\cdots\oplus \lambda^{m-1}A_{m-1}$ be a matrix polynomial with all the
$A_j$'s irreducible. Let $C_P$ be the associated companion matrix. If there exists a nonzero periodic point of
$y \mapsto C_P y$, then $\mu(C_P) = 1$.
\end{proposition}
\begin{proof}
Assume that $C_P^{q} x = x$ for some $x \neq 0$ and $q \geq 1$. Define
$y:= x \oplus\ (C_P x) \oplus\ (C_P^{2} x)\ \oplus \cdots \oplus\ (C_P^{q-1} x)$.
By max--linearity, we have $C_P y = (C_P x) \oplus (C_P^{2} x) \oplus \cdots \oplus (C_P^{q} x)
= (C_P x) \oplus (C_P^{2} x) \oplus \cdots \oplus x = y$,
where the last equality comes from the fact that $C_P^{q} x = x$. Hence $y \neq 0$ is a max--eigenvector of $C_P$
with eigenvalue $1$. Since the $A_j$'s are irreducible, by Remark \ref{rem:irreducible-Cp}, $C_P$ is irreducible.
For an irreducible max--algebraic matrix, the (unique) eigenvalue admitting a nonzero nonnegative
eigenvector equals the maximum cycle geometric mean $\mu(C_P)$. Therefore $\mu(C_P) = 1$.
\end{proof}
For a non-empty subset $S$ of $\mathbb{R}^n_+$, we denote by $\mathcal{LC}(S)$ the max--cone generated by $S$. For a bounded family of matrices $\Sigma = \{A_1,\dots,A_N\} \subset M_n(\mathbb{R}_+)$, we denote the collection of common eigenvectors of $\Sigma$ as $E$. Note that $\mathcal{LC}(E)$ is invariant for the collection $\Sigma$, as well
as for any matrix polynomial $P(\lambda)$ with coefficients from the set $\Sigma$. For a $p-$letter word
$\omega = \omega_1 \omega_2 \cdots \omega_p$ on
$\{1,\dots,N\}$, we define the associated product by
$A_\omega:= A_{\omega_p} A_{\omega_{p-1}} \cdots A_{\omega_1}$. The following theorem replaces the individual spectral constraints $\mu(A_i)\leq 1$ in Theorem $3.4$ of \cite{DEDS-dynamic-product} by the single joint spectral radius
condition $\rho_\eta(\Sigma)\leq 1$, which controls the growth of all products uniformly. This global bound is essential in our setting, as it allows one to deduce full convergence of $A_\omega^k x$ for every word $\omega$ using a single stability
parameter, rather than relying on matrix–wise spectral estimates. Our main result is the following.
\begin{theorem}\label{thm:JSR-common-eig-periodic}
Let $\Sigma = \{A_1,\dots,A_N\}\subset M_n(\mathbb{R}_+)$. Assume that $\rho_\eta(\Sigma)\leq 1$ and that $\Sigma$
admits a nontrivial set $E$ of common max-eigenvectors. Then for every $x \in \mathcal{LC}(E)$, and a word
$\omega = \omega_1 \omega_2 \cdots \omega_p$ on $\{1,\dots,N\}$, the limit
$\displaystyle \lim_{k\to\infty} A_\omega^k\otimes x := \xi_{x,\omega}$ exists, and $\xi_{x,\omega}$ is a periodic point of $A_\omega$.
\end{theorem}
\begin{proof}
For $A_i \in \Sigma$ and $v_j \in E$, let $A_i v_j = \alpha_{ij}v_j$, for some $\alpha_{ij} \in \mathbb{R}_+$. Then
max--linearity gives
$A_i^k v_j = \alpha_{ij}^k v_j$ for all $k \geq 1$. Hence $\eta(A_i^k)\;\geq\;\frac{\|A_i^kv_j\|}{\|v_j\|}
= \frac{\|\alpha_{ij}^k v_j\|}{\|v_j\|} = \alpha_{ij}^k$. Therefore, $\widehat{\eta}(A_i) \geq \alpha_{ij}$. It then
follows from Theorem \ref{thm:etacap-rho-eta-inequality} that $\rho_\eta(\Sigma) \geq \widehat{\eta}(A_i)$ and so
$\alpha_{ij} \leq \rho_\eta(\Sigma) \leq 1$. Now fix a word $\omega = \omega_1 \cdots \omega_p$ and set
$\beta_j(\omega):= \displaystyle \prod_{t=1}^{p} \alpha_{\omega_t,j} \in [0,1]$. Then
$A_\omega v_j = \beta_j(\omega)v_j$ and $A_\omega^k v_j = \beta_j(\omega)^k v_j$.
Take $x \in \mathcal{LC}(E)$ and choose scalars $\gamma_1, \dots, \gamma_r \in \mathbb{R}_+$ such that
$x = \displaystyle \bigoplus_{j=1}^{r} \gamma_j v_j$. Using max--linearity, we have
\[
A_\omega^k x
= A_\omega^k (\bigoplus_{j=1}^{r} \gamma_j v_j)
= \displaystyle \bigoplus_{j=1}^{r} \gamma_j\,(A_\omega^k v_j)
= \displaystyle \bigoplus_{j=1}^{r} \gamma_j\,\beta_j(\omega)^k\,v_j.
\]
Since each $\beta_j(\omega) \in [0,1]$, the scalar limit $\beta_j(\omega)^k \to 0$ if
$\beta_j(\omega) < 1$, while $\beta_j(\omega)^k = 1$ for all $k$ if $\beta_j(\omega) = 1$.
Thus $A_\omega^k x$ converges, and its limit $\xi_{x,\omega}$ satisfies
$A_\omega \xi_{x,\omega} = \xi_{x,\omega}$, thereby proving the periodicity conclusion as well.
\end{proof}
Note that in the above theorem, if the word $\omega$ contains the letter $i$ at least once, then
$A_i \xi_{x,\omega} = \xi_{x,\omega}$. In particular, if $\omega$ contains every letter $1,\dots,N$ at least once, then
$\xi_{x,\omega}$ is a common fixed point for the entire collection $\Sigma$, and the limit
$\xi_{x,\omega}$ is independent of the choice of such a word $\omega$. This gives an immediate consequence to a matrix polynomial with coefficients from $\Sigma$. We state this as a corollary and skip its proof.
\begin{corollary}\label{cor:com-eigenvector-polynomial}
Let $\Sigma = \{A_1,\dots,A_N\}\subset M_n(\mathbb{R}_+)$ satisfy the hypotheses of Theorem \ref{thm:JSR-common-eig-periodic}. Also, let $P(\lambda) = B_0 \oplus \lambda B_1\oplus \cdots \oplus
\lambda^{m-1}B_{m-1}$ be a max--matrix polynomial whose coefficients satisfy $B_j \in \Sigma$ for all
$j = 0,1, \dots, m-1$. Then, any fixed point $\xi_x$ for $A_\omega$, as described in Theorem
\ref{thm:JSR-common-eig-periodic} is a fixed point of the map $x \mapsto P(1)\otimes x$. In other words,
$P(1) \xi_x = \xi_x$.
\end{corollary}
We end the paper with the following corollary.
\begin{corollary}\label{cor:JSR-strict-to-zero}
Let $\Sigma = \{A_1,\dots, A_N\} \subset M_n(\mathbb{R}_+)$ be finite and suppose that
$\rho_\eta(\Sigma) < 1$. Then for every word $\omega$ and every $x \in \mathbb{R}^n_+$,
$\displaystyle \lim_{k\to\infty} A_\omega^k x = 0$. In particular, the only periodic point of the map
$x \mapsto A_\omega x$ is $0$.
\end{corollary}
\begin{proof}
Fix a word $\omega$ of length $p$ and choose $r$ such that $\rho_\eta(\Sigma) < r <1$.
By the definition of $\rho_\eta(\Sigma)$, there exists $K \in \mathbb{N}$ such that for all $t \geq K,
\displaystyle \sup_{B_1,\dots, B_t \in \Sigma}\eta(B_t \cdots B_1) \leq r^t$. For $k$ large enough so that $kp \geq K$,
the matrix $A_\omega^k$ is a product of length $kp$
with factors in $\Sigma$ and hence $\eta(A_\omega^k) \leq r^{kp}$. Therefore
$\|A_\omega^k\otimes x\| \leq \eta(A_\omega^k) \|x\| \leq r^{kp} \|x\| \rightarrow 0$, which implies
$A_\omega^k x\to 0$ in $\mathbb{R}^n_+$. If $y \neq 0$ were a periodic
point, say $A_\omega^q y = y$, then $A_\omega^{kq} y = y$ for all $k$, contradicting
the convergence to $0$. Hence $0$ is the only periodic point.
\end{proof}
We end the paper with an example that illustrates the above Theorem and Corollary.
\begin{example}\label{eg-2}
We construct a family $\Sigma = \{A_1, A_2\} \subset M_4(\mathbb{R}_+)$ illustrating Theorem \ref{thm:JSR-common-eig-periodic}. Let
$A_1 =
\begin{pmatrix}
1 & 0.5 & 0 & 0 \\
0.3 & 1 & 0 & 0 \\
0 & 0 & 0.9 & 0.7 \\
0 & 0 & 0.5 & 0.9
\end{pmatrix},
\
A_2 =
\begin{pmatrix}
0.8 & 1 & 0 & 0 \\
1 & 0.4 & 0 & 0 \\
0 & 0 & 0.8 & 0.6 \\
0 & 0 & 0.4 & 0.8
\end{pmatrix}, \ v_1 = (1,1,0,0)^T$ and $v_2 = (0,0,1,1)^T$. A direct computation gives
\[
A_1 \otimes v_1 = v_1, \quad A_2 \otimes v_1 = v_1,
\]
and therefore $v_1$ is a common max-eigenvector of $\Sigma$ with corresponding eigenvalue $1$. Similarly,
\[
A_1 \otimes v_2 = 0.9\, v_2, \quad A_2 \otimes v_2 = 0.8\, v_2,
\]
making $v_2$ a common max-eigenvector of $A_1$ with eigenvalue $0.9$, and of $A_2$ with
eigenvalue $0.8$. Thus $E = \{v_1, v_2\}$ is a nontrivial common max-eigenvector set for $\Sigma$. By
statements $(1)$ and $(2)$ of Theorem \ref{Determining the JSR}, $\rho_\eta(\Sigma) = \mu(A_1 \oplus A_2)$. From the above, we have,
\[A_1 \oplus A_2 = \begin{pmatrix}
1 & 1 & 0 & 0 \\
1 & 1& 0 & 0 \\
0 & 0 & 0.9 & 0.7 \\
0 & 0 & 0.5 & 0.9
\end{pmatrix}.\]
The spectral radius of $A_1 \oplus A_2$ is $\mu(A_1 \oplus A_2) =1$. Thus, $\rho_\eta(\Sigma) = 1$. Let
$x = \alpha\, v_1\oplus \beta\, v_2 = (\alpha,\alpha,\beta,\beta)^T \in \mathcal{LC}(E)$, for
$\alpha, \beta \in \mathbb{R}_+$. Consider the word $\omega = 12$, so that $A_\omega = A_2\otimes A_1$. Then
$A_\omega \otimes v_1 = v_1$ and $A_\omega \otimes v_2 = (0.9 \times 0.8)\,v_2 = 0.72\,v_2$. By max-linearity,
\[
A_\omega^k \otimes x
\;=\;
\alpha\, v_1 \oplus \beta\,(0.72)^k\, v_2
\;\xrightarrow{k\to\infty}\;
\alpha\, v_1
\;=\;
(\alpha,\,\alpha,\,0,\,0)^T
\;:=\;
\xi_{x,\omega}.
\]
Since $A_\omega \otimes \xi_{x,\omega} = \xi_{x,\omega}$, the limit $\xi_{x,\omega}$ is a fixed point of $A_\omega$, confirming Theorem \ref{thm:JSR-common-eig-periodic}. The component along $v_2$ decays geometrically because both eigenvalues ($0.9$ and $0.8$) are strictly less than $1$, while the component along $v_1$ stabilizes because the
corresponding eigenvalue equals $1$. Moreover, since the word $\omega=12$ contains every letter in $\{1,2\}$, the limit $\xi_{x,\omega}$ is a common fixed point of $\Sigma$ and is independent of the particular choice of such a word.
\end{example}
\bibliographystyle{amsplain}
\begin{thebibliography}{99}
\bibitem{Askar-Jayaraman-Himadri}
Askar Ali M, S. Jayaraman and H. Mukherjee, \emph{Simultaneous triangularization over max-algebras}, J. Algebra
Appl., (Online Ready) doi:10.1142/S0219498827502410.
\bibitem{Bapat}
R. B. Bapat, \emph{A max version of the Perron–Frobenius theorem}, Linear Algebra Appl. \textbf{275–276} (1998) 3 -- 18.
\bibitem{Bapat-Raghavan}
R. B. Bapat and T. E. S. Raghavan, \emph{Nonnegative Matrices and Applications}, Encyclopedia of Mathematics and
Applications, Vol. 64, Cambridge University Press, Cambridge, 1997.
\bibitem{Butkovic-0}
P. Butkovic, \emph{Max-algebra: the linear algebra of combinatorics ?}, Linear Algebra Appl., \textbf{367} (2003),
313 -- 335.
\bibitem{Butkovic}
P. Butkovič, \emph{Max-linear Systems: Theory and Algorithms}, Springer Monographs in Mathematics,
Springer Verlag, London, 2010.
\bibitem{Butkovic-CuninghameGreen}
P. Butkovič and R. A. Cuninghame-Green, \emph{On matrix powers in max-algebra}, Linear Algebra Appl., \textbf{421}
(2007), 370 -- 381.
\bibitem{Economics-note}
R. A. Cuninghame-Green, \emph{Minimax Algebra}, Lecture Notes in Economics and Mathematical Systems, Vol. 166,
Springer, Berlin, 1979.
\bibitem{Gursoy-Mason}
B. B. Gursoy and O. Mason, \emph{Spectral properties of matrix polynomials in the max algebra}, Linear Algebra Appl., \textbf{435} (2011), 1626 --1636.
\bibitem{Neda-Ghasemi}
N. Ghasemizadeh and Gh. Aghamollaei, \emph{Some results on matrix polynomials in the
max algebra}, Banach J. Math. Anal., \textbf{9} (2015), 17 -- 26.
\bibitem{Max@work}
B. Heidergott, G. J. Olsder and J. W. van der Woude, \emph{Max Plus at Work: Modeling and Analysis of Synchronized
Systems: a Course on Max-Plus Algebra and Its Applications}, Princeton Series in Applied Mathematics,
Princeton University Press, Princeton, New Jersey, 2006.
\bibitem{Heil-Strang}
C. Heil and G. Strang, \emph{Continuity of the joint spectral radius: application to wavelets}, \emph{Linear Algebra for Signal Processing}, A. Bojanczyk and G. Cybenko (eds.), IMA Vol. Math. Appl., \textbf{69}, Springer--Verlag, New York,
1995, pp. 51 -- 61.
\bibitem{DEDS-dynamic-product}
S. Jayaraman, S. Sridharan, and Y. K. Prajapaty, \emph{Dynamics of products of matrices
in max algebra}, Discrete Event Dynamic Systems, \textbf{33} (2023), 473 -- 522.
\bibitem{Extractra-2022}
S. Jayaraman, Y. K. Prajapaty and S. Sridharan,
\emph{Dynamics of products of nonnegative matrices}, Extracta Math., \textbf{37} (2022), 223 -- 242.
\bibitem{Katz-Schneider-Sergeev}
R. D. Katz, H. Schneider and S. Sergeev, \emph{On commuting matrices in max algebra and in classical nonnegative algebra},
Linear Algebra Appl., \textbf{436}(2) (2012), 276 -- 292.
\bibitem{Lemmens-Nussbaum}
B. Lemmens and R. D. Nussbaum, \emph{Nonlinear Perron--Frobenius Theory}, Cambridge Tracts in Mathematics,
\textbf{189}, Cambridge University Press, Cambridge, 2012.
\bibitem{Lur}
Y.-Y. Lur, \emph{A max version of the generalized spectral radius theorem}, Linear Algebra Appl., \textbf{418} (2006),
336 -- 346.
\bibitem{Lur2005}
Y.-Y. Lur, \emph{On the asymptotic stability of nonnegative matrices in max algebra}, Linear Algebra Appl., \textbf{407} (2005),
149 -- 161.
\bibitem{Muller-Peperko-0}
V. Muller and A. Peperko, \emph{Generalized spectral radius and its max algebra version}, Linear Algebra Appl., \textbf{439} (2013), 1006 -- 1016.
\bibitem{Muller-Peperko}
V. Muller and A. Peperko, \emph{On the spectrum in max algebra}, Linear Algebra Appl., \textbf{485} (2015), 250 -- 266.
\bibitem{Rota-Strang}
G.-C. Rota and W. G. Strang, \emph{A note on the joint spectral radius}, Indag. Math., \textbf{22} (1960), 379 -- 381.
\end{thebibliography}
\end{document}