Discrete Event Dynamic Systems (2023) 33:161–179

https://doi.org/10.1007/s10626-023-00380-0

Dynamics of products of matrices in max algebra

Shrihari Sridharan1 · Sachindranath Jayaraman1 · Yogesh Kumar Prajapaty1

Received: 6 July 2022 / Accepted: 30 June 2023 / Published online: 10 August 2023
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023

Abstract
The aim of this manuscript is to understand the dynamics of matrix products in a max algebra.
A consequence of the Perron-Fröbenius theorem on periodic points of a nonnegative matrix is
generalized to a max algebra setting. The same is then studied for a finite product associated
to a p-lettered word on N letters arising from a finite collection of nonnegative matrices,
with each member having its maximum circuit geometric mean at most 1.

Keywords Products of nonnegative matrices · Max-algebras · Boolean matrices · Fröbenius

normal form of nonnegative matrices · Circuit geometric mean · Periodic points

1 Introduction

We work throughout over the field R of real numbers and use the standard notations Rn and
Mn (R) to denote the n-dimensional real vector space of n-tuples of real numbers and the real
vector space of n × n square matrices with real entries respectively. We concern ourselves
with only those matrices whose entries are nonnegative real numbers. This set will be denoted
by Mn (R+ ). Other notations and terminologies used in this work will be introduced later.
Recall that given a self map f on a topological space X , an element x ∈ X is called a
periodic point of f if there exists a positive integer q such that f q (x) = x. In such a case,
the smallest such integer q that satisfies f q (x) = x is called the period of the periodic point
x. The starting point for this work is the following consequence of the Perron-Fröbenius
theorem.

Theorem 1.1 (Theorem B.4.7, Lemmens and Nussbaum (2012)) Let A ∈ Mn (R+ ) with
spectral radius less than or equal to 1. Then, there exists a positive integer q such that for

B Shrihari Sridharan
shrihari@iisertvm.ac.in
Sachindranath Jayaraman
sachindranathj@iisertvm.ac.in
Yogesh Kumar Prajapaty
prajapaty0916@iisertvm.ac.in
1 Indian Institute of Science Education and Research, Thiruvananthapuram (IISER-TVM), Vithura, India

123
162 Discrete Event Dynamic Systems (2023) 33:161–179

 
every x ∈ Rn with Ak x k ∈ Z bounded, we have
+

lim Akq x = ξx ,
k →∞
where ξx is a periodic point of A whose period divides q.
In an attempt to generalize Theorem 1.1, when the matrix A in the above theorem is replaced
by a product of the matrices
 Ar ’s, possibly an infinite one, drawn from the finite collection of
nonnegative matrices, A1 , A2 , . . . , A N , the following result was obtained by the authors.
The details may be found in Jayaraman et al. (2022).
 
Theorem 1.2 Jayaraman et al. (2022) Let A1 , A2 , . . . , A N , N < ∞, be a collection
of n × n matrices with nonnegative entries, each having spectral radius at most 1. Assume
that the collection has a nontrivial set of common eigenvectors, E. For any finite p, let Aω
denote the matrix product associated to a p-lettered word ω on the letters 1, 2, . . . , N .
Suppose LC (E) denotes the set of all real linear combinations of the vectors in E. Then, for
any x ∈ LC (E), there exists an integer q  1 (independent of ω) such that

ω x = ξx ,
lim Akq (1.1)
k →∞
where ξx (independent of ω) is a periodic point of Aω , whose period divides q.
The aim of this work is to explore the possibilities of extending Theorems 1.1 and 1.2 in
the setting of max algebras. By a max algebra, we mean the triple (R+ , ⊕, ⊗), where R+
denotes the set of nonnegative real numbers, ⊕ denotes the binary operation of taking the
maximum of two nonnegative numbers and ⊗ is the usual multiplication of two numbers.
There are several abstract examples of max algebras, as may be found in Butkovic (2010). The
one given above is more amenable to work with, while dealing with nonnegative matrices.
Another example is the set of real numbers, together with −∞, equipped with the binary
operations of maximization and addition, respectively. The latter system is isomorphic to the
former one via the exponential map. A good reference on max algebras is the monograph
by Butkovic (2010). For Perron-Fröbenius theorem in max algebras, one may refer to Bapat
(1998).

2 Preliminaries on max algebras

We introduce some preliminary notions from max algebras that we use, in this section. In
particular, we introduce the notion of a max eigenvalue, the corresponding max eigenvector
and a few properties of the same. We begin with a description of matrix product in max
algebra. Recall that for a, b ∈ R+

a ⊕ b = max{a, b} and a ⊗ b = ab.

Let A and B be two n × n nonnegative matrices. Then the matrix product under the
considered max algebra of A and B is denoted by AB = A ⊗ B and is defined by

[AB]i j = [A ⊗ B]i j = max {aik ⊗ bk j }.

k

Further, if A = B, then A⊗2 denotes the square of A. For simplicity and ease of writing,
we shall denote this by A2 . Further, Ak shall denote the k-th power of A in this algebra. By
convention, when k = 0, we assume A0 to be the identity matrix.

123
Discrete Event Dynamic Systems (2023) 33:161–179 163

Definition 2.1 An n × n matrix A is said to be reducible if there exists a permutation matrix

P such that  
B C
P APT = P ⊗ A ⊗ PT = , (2.1)
0 D

where B and D are square matrices. Otherwise A is irreducible.

Note that the notion of irreducibility is the same in the max algebra and the Euclidean
algebra. If A is reducible and is in the form (2.1), and if a diagonal block is reducible, then
this block can be reduced further via a permutation similarity. Continuing this process, we
have a suitable permutation matrix P such that P A P T is in the block triangular form
⎡ ⎤
A11 A12 · · · A1m
⎢ 0 A22 · · · A2m ⎥
⎢ ⎥
P APT = ⎢ . .. . . . ⎥, (2.2)
⎣ .. . . .. ⎦
0 0 · · · Amm

where each block Aii is square and is either irreducible or a 1 × 1 null matrix. This block
triangular form is called the Fröbenius normal form of A.
Given an n ×n nonnegative matrix A, there is a natural way to associate a simple weighted
directed graph G (A) to the matrix as follows: G (A) has n vertices, say 1, . . . , n, such that
there is an edge from i to j with weight ai j if and only if ai j > 0. By a circuit, we always
mean a simple circuit. In contrast, our paths may include a vertex and/or an edge more than
once. By the product of a path, we mean the product of the weights of the edges in the path.
Let (i 1 , i 2 ), (i 2 , i 3 ), . . . , (i k , i 1 ) be a circuit in G (A). Then ai1 i2 ai2 i3 · · · aik i1 is the corre-
sponding circuit product and its k th root is the circuit geometric mean corresponding to the
circuit. The maximum among all possible circuit geometric means in G (A) is denoted by
μ(A). Observe that μ(A) > 0 or μ(A) = 0 indicates the presence or absence (respectively)
of a circuit in G (A). A circuit with circuit geometric mean μ(A) = μ(G (A)) is called a
critical circuit, and vertices on the critical circuits are critical vertices. The critical matrix of
A, denoted by AC = [aiCj ], is formed from a principal sub-matrix of A corresponding to the
critical vertices, by setting aiCj = ai j if (i, j) is in a critical circuit, and ai j = 0 otherwise.
Thus the critical graph G (AC ) has vertex set VC = {critical vertices}.
Note that the weighted directed graph associated with A, namely G (A) and the weighted
directed graph associated with P A P T , for a permutation matrix P, namely G (P A P T ) are
isomorphic. This implies that the corresponding circuit geometric means are equal. For more
on circuit geometric means, see Bapat (1998); Berman and Plemmons (1994); Elsner and
van den Driessche (1999).

Definition 2.2 Let A be a nonnegative matrix. We say that λ is a max eigenvalue of A if there
exists a nonzero, nonnegative vector x such that A ⊗ x = λ ⊗ x. Further, x is called a max
eigenvector associated to λ.

The following result is referred to as the max version of the Perron-Fröbenius theorem
and can be found in Bapat (1998).

Theorem 2.1 (Theorem 2, Bapat (1998)) Let A be an n × n nonnegative, irreducible matrix.

Then there exists a positive vector x such that A ⊗ x = μ(A) ⊗ x.

123
164 Discrete Event Dynamic Systems (2023) 33:161–179

It can be proved that an irreducible matrix has only one eigenvalue with a corresponding
eigenvector in max algebras. Moreover, unlike the eigenvalues in the Euclidean algebra, the
max eigenvalue has an explicit formula given by,
 1
μ(A) = μ(AC ) = max ai1 i2 ai2 i3 · · · ai i1  .


We now define a relation between two vertices in G (A).

Definition 2.3 Let A be a nonnegative square matrix of order n. For 1  i, j  n, we say

that i has access to j if there is a path from vertex i to vertex j in G (A), and that i and j
communicate if i has access to j and j has access to i.

Communication is an equivalence relation. Note that in the identity matrix, we assume

that every vertex i communicates with itself. The following result concerns the spectrum of
a nonnegative matrix in max algebra.

Theorem 2.2 Bapat (1998) Let A be an n × n nonnegative matrix in Fröbenius normal form,
⎡ ⎤
A11 A12 · · · A1m
⎢ 0 A22 · · · A2m ⎥
⎢ ⎥
⎢ .. .. . . .. ⎥ .
⎣ . . . . ⎦
0 0 · · · Amm
Then λ is an eigenvalue with a corresponding nonnegative eigenvector if and only if there
exists a positive integer i ∈ {1, 2, · · · , m} such that λ = μ(Aii ) and furthermore, class j
does not have access to class i whenever μ(A j j ) > μ(Aii ).

Writing the Fröbenius normal form of A as the sum of its diagonal blocks and the strict
upper triangular block, say
⎡ ⎤
A11 A12 · · · A1m
⎢ 0 A22 · · · A2m ⎥
⎢ ⎥
P APT = ⎢ . .. . . .. ⎥
⎣ .. . . . ⎦
0 0 · · · Amm
⎡ ⎤ ⎡ ⎤
A11 0 · · · 0 0 A12 · · · A1m
⎢ 0 A22 · · · 0 ⎥  ⎢0 0 · · · A2m ⎥
⎢ ⎥ ⎢ ⎥
= ⎢ . .. . . .. ⎥ ⎢ .. .. . . .. ⎥
⎣ .. . . . ⎦ ⎣ . . . . ⎦
0 0 · · · Amm 0 0 ··· 0
=: D A ⊕ N A , (2.3)
we observe that N A does not contribute in computing the circuit geometric mean of P A P T ,
and consequently has no role while determining μ(A). This follows from Theorem 2.2
and the communication relation defined above. Moreover, note that μ(A) = μ(D A ) =
max {μ(A11 ), μ(A22 ), . . . , μ(Amm )}.
We illustrate Theorem 2.2 below. Similar examples can be found in Bapat (1998).

Example 2.3 Let

    
40 42 52
A = ,B = and C = .
05 05 04

123
Discrete Event Dynamic Systems (2023) 33:161–179 165

Note that A has two max eigenvalues, namely, 4 and 5, with corresponding max eigenvec-
tors (1, 0) and (0, 1) respectively. Observe that B has two max eigenvalues namely 4 and
5 with corresponding max eigenvectors (1, 0) and (2, 5) respectively, whereas C has only
one max eigenvalue 5.

We now state two useful results due to Elsner. The first one is a D AD-type theorem that
says that a nonnegative irreducible matrix can be bounded by a matrix of all ones via a
D AD transform. The second one concerns the period of a nonnegative irreducible matrix
with μ(A) = 1. More about this will be discussed in the section on Boolean matrices.

Lemma 2.4 Elsner and van den Driessche (2001) Let A be an irreducible matrix with μ(A) 
1, x ∈ Rn+ , x  = 0 and z = A∗ x, where A∗ = I ⊕ A ⊕ A2 ⊕ . . . ⊕ An−1 . Then every
coordinate of z is strictly positive. If D = diag(z 1 , z 2 , . . . , z n ), then
   
D −1 AD ij
= D −1 ⊗ A ⊗ D i j  1.

In the hypothesis, suppose we start with μ(A) < 1, then the assertion of the Lemma is
also a strict inequality.

Theorem 2.5 Elsner and van den Driessche (1999); Butkovic (2010) Assume that A is a
nonnegative irreducible definite matrix, meaning μ(A) = 1. Then, there exist q and t0 in Z+
such that for all t  t0 , we have

At + q = At , where the powers are taken in the max algebra. (2.4)

It is interesting to note that the assertion in the above theorem is analogous to the concept
of sequences that are “ultimately geometric” and “ultimately periodic”, as introduced by
Gaubert (1994). de Schutter (2000) uses these concepts and obtains results pertaining to
powers of matrices in the max-plus algebraic setting. The notion of periodicity that we deal
with in this paper is the same as found in the assertion of Theorem 2.5; although we make a
precise definition of the same in Section 4.
A generalised version of Lemma 2.4 is available in the literature for what is known as the
scaling of a matrix A. If D is an invertible diagonal matrix, then D −1 AD is called a scaling
of A. For such scaling of A, we have the following theorem, as can be found in Butkovic
(2010).

Theorem 2.6 Butkovic (2010) Let A be a matrix with μ(A) = 1. Then there exists a positive
vector z ∈ Rn+ such that
   
D −1 AD ij
= D −1 ⊗ A ⊗ D i j  1 where D = diag(z 1 , z 2 , . . . , z n ).

Further, A and its scaling have the same dynamics.

We end this section with bounds and inequalities for μ(A).

Lemma 2.7 Bapat et al. (1995) Let A be an n × n irreducible nonnegative matrix.

1. Suppose there exist η1 > 0 and a vector z (1)  = 0 such that A ⊗ z (1)  η1 z (1) . Then
μ(A)  η1 .
2. Suppose there exist η2 > 0 and a vector z (2)  = 0 such that A ⊗ z (2)  η2 z (2) . Then
μ(A)  η2 .

123
166 Discrete Event Dynamic Systems (2023) 33:161–179

Consequently, we have:

Corollary 2.8 Bapat et al. (1995) Let A be an n × n be nonnegative matrix. Then

min max ai j  μ(A)  max ai j .
i j i, j

Let A and B be two commuting matrices. The following result gives us the circuit geo-
metric mean of the addition and product of the two matrices in max algebras.

Theorem 2.9 Katz et al. (2012) Let A and B commute. Then,

1. μ (A ⊗ B)  μ(A) ⊗ μ(B).
2. μ (A ⊕ B)  μ(A) ⊕ μ(B).
Moreover, equality holds in both the above inequalities if the matrices A and B are irreducible.

3 Statements of results

We state the main results of this paper, in this section. The first result includes an analogue
of Theorem 1.1, in the max algebra setting.

Theorem 3.1 Let A be a matrix with μ(A)  1. Then, there exists an integer q such that for
every 1 ≤ j ≤ q, we have
lim Akq+ j =  A( j) ,
k →∞

where 
A( j) is a periodic matrix, whose period divides q. Further, for every x ∈ Rn+ , we have
( j)
lim Akq+ j ⊗ x = ξx ,
k →∞
( j)
where ξx is a periodic point of the matrix A, whose period divides q.

The following example illustrates that the period of the periodic matrix A(j) and the
 period
( j) 0.5 1
of the periodic point ξx have no correlation between them. Consider A = . Then,
1 0.5
 
 (1)  (2) 1 0.5
we obtain q = 2. Further, A = A whose period is 2 whereas, A = has
0.5 1
( j)
period 1. For x = (1, 1)T , we find that the period of the periodic point ξx is 1, for j = 1, 2,
( j)
while for y = (1, 0)T , the period of the periodic point ξ y is 2, for j = 1, 2.
Experts working in the area of convergence of iterative schemes in max algebras may
be aware of the results due to Heidergott et al. (2006), however in the setting of max-plus
algebra. We combine two significant results from Heidergott et al. (2006) and present the
same in the max-product algebra as follows:

Theorem 3.2 Heidergott et al. (2006) Let A be a matrix with μ(A) ≤ 1. Then,

1. The infinite series Ak converges in max-algebra.
k ≥1
 1
k
2. If lim A ⊗ x0
k
exists, then the limit is independent of the starting point, provided
k →∞
we choose the starting point x0 from the interior of Rn+ .

123
Discrete Event Dynamic Systems (2023) 33:161–179 167

However, we urge the readers to observe through the following example that the two
limits namely, lim Akq+ j , as discussed in Theorem 3.1, and the infinite series Ak , as
k →∞
⎡ k ≥⎤1
0.2 1 4
discussed in Theorem 3.2 are quite different in nature. Consider A = ⎣ 1 0.5 6 ⎦. Then,
0 0 0.9
⎡ ⎤
 11 6
Ak = ⎣1 1 6 ⎦ while  A( j) ’s for this matrix are computed in Section 8. Further, taking
k ≥1 0 0 0.9
x0 = (1, 1, 1)T , we obtain
 1
k
lim Ak ⊗ x 0 = (1, 1, 0.9)T ,
k →∞

the limiting point not being a periodic point of any order.

The next result deals with a finite collection of matrices, say {A1 , A2 , · · · , A N }, with
each of them having μ(Ai ) ≤ 1. For any finite p ∈ Z+ , we denote the set of all p-lettered
words on the first N positive integers by
p  
 N := {ω = ω1 ω2 · · · ω p : 1  ωi  N },
p
and for any ω ∈  N , we define the matrix product

Aω := Aω p ⊗ Aω p−1 ⊗ · · · ⊗ Aω1 . (3.1)

We now state the second result of this paper, which is analogous to Theorem 1.2,
in the max algebra setting, however under the hypothesis that the collection of matrices
{A1 , A2 , · · · , A N } commute pairwise.

Theorem 3.3 Let {A1 , A2 , · · · , A N } be a finite collection of pairwise commuting nonneg-

p
ative matrices with μ(Ai ) ≤ 1, for every 1 ≤ i ≤ N . Suppose ω ∈  N is a p-lettered word
in which the letter i occurs pi many times. Then, there exists an integer q such that for every
1 ≤ j ≤ q, we have
N 
  pi
 ( j)
lim Akq+
ω
j
= Ai .
k →∞
i =1

Before stating our third result, let us assume that the collection {A1 , A2 , · · · , A N } of
matrices do not necessarily commute, nevertheless, has a set of common
 max eigenvectors,

(1) (n)
say, E = {v1 , v2 , · · · , vm }, (see Definition 2.2). Suppose vi = vi , . . . , vi . Define

 

m
LC (E) = (αi ⊗ vi ) : αi ∈ R+
i =1
     
( j) ( j)
= u = u (1) , . . . , u (n) ∈ Rn+ : u ( j) = max α1 v1 , · · · , αm vm ; αi ∈ R+ .

We now state our third theorem in this article, which is analogous to Theorem 1.2, in the
max setting.

123
168 Discrete Event Dynamic Systems (2023) 33:161–179

Theorem 3.4 Let {A1 , A2 , · · · , A N } be a collection of nonnegative matrices with μ(Ai ) 

1. Assume that the collection has a non-trivial set of common eigenvectors, say E. Let Aω
p
be the matrix product associated with the word ω ∈  N . Then, for any x ∈ LC (E), we have

lim Akω ⊗ x = ξx ,
k →∞

where ξx is a fixed point for every matrix in the collection {A1 , A2 , · · · , A N }.

Note that Theorem 3.4 is also applicable in the case of commuting matrices. In the case
of commuting matrices, Theorem 3.4 however, does not give a complete picture as against
Theorem 3.3.

4 Boolean matrices

Recall that A is a Boolean matrix if all the entries of A are either 0 or 1. The following fact
on Boolean matrices can be found in Pang and Guu (2001); Rosenblatt (1957). Let A be a
Boolean matrix. Then there exist c0 and q such that
Ac + kq = Ac , for all k ∈ Z+ , c  c0  1. (4.1)
The minimal such q is called the period of the matrix A. Notice that this definition of the
period of a Boolean matrix coincides with the same notion introduced in Theorem 2.5 of
Section 2. We mean this same notion of periodicity in Theorem 3.1. We now introduce a few
more notions that will help us in understanding more about the period of a matrix. These are
taken from Rosenblatt (1957).

Definition 4.1 A Boolean matrix A is said to be

1. convergent if and only if there exists in the sequence {Ak : k  1} a power Am of A such
that Am = Am+1 .
2. oscillatory or periodic if and only if there exists in the sequence {Ak : k  1}, a power
Am of A such that Am = Am+q where q is the smallest integer for which this holds and
q > 1.

Theorem 4.1 If A is a Boolean matrix of order n with graph G (A), then the following hold:
1. Ak converges to the zero matrix if and only if G (A) contains no cyclic nets.
2. Ak converges to Jn (the matrix of order n with all entries being 1) if and only if G (A) is
a universal cyclic net.
3. A is oscillatory if and only if G (A) contains at least one maximal cyclic net which is not
universal. The period of a sub-matrix corresponding to a specified non-universal maximal
cyclic net in G (A) is given by the greatest common divisor of the order of all simple cyclic
nets contained in the maximal net.

Remark 4.2 Cyclic nets are the connected components of G (A) and the order of a cyclic net
is the number of vertices in the cyclic net. A cyclic net is said to be a universal cyclic net if
for some positive integer q every point of the cyclic net is attainable in q steps from some
fixed point in the cyclic net, see Rosenblatt (1957).

We now define asymptotic period of a sequence of matrices, as given in Pang and Guu
(2001).

123
Discrete Event Dynamic Systems (2023) 33:161–179 169

Definition 4.2 A sequence {Ak : k ∈ Z+ } of matrices in Mn (R) is called asymptotic q-

periodic if
lim A j+kq =  A( j) (4.2)
k →∞
exists for j = 1, 2, . . . , q. The minimal such q is called the asymptotic period of the sequence.

We note that the Definitions 4.1 and 4.2 coincide, for Boolean matrices.
Let J(m×n) be the m × n matrix of all ones. When m = n, we denote the matrix of all
ones by Jn . Suppose A is an n × n nonnegative matrix with A  Jn (entry-wise). Consider
the following decomposition:
A = B(A) ⊕ R(A), (4.3)
 
ai j if ai j = 1 0 if ai j = 1
where B(A) = and R(A) = .
0 if ai j < 1 ai j if ai j < 1
The following result is due to Pang and Guu.

Theorem 4.3 Pang and Guu (2001) Let A = B(A) ⊕ R(A). B(A) has period q if and only
if the sequence {Ak : k ∈ Z+ } has asymptotic period q.

Remark 4.4 From the above definitions, results and remarks, it follows that the period of a
Boolean matrix is the same as the greatest common divisor of all possible lengths of simple
circuits in G (AC ). Similar results on Max-Plus algebras and Max-Min algebras can also be
found in Gavalec (2000a, b) respectively.

5 Two technical lemmas

In this section, we state and prove two lemmas that will be useful in the sequel, while proving
the main theorems. We suppose that A is a reducible matrix with μ(A) ≤ 1, and can be
written in its Fröbenius normal form A = D A ⊕ N A (upto a permutation similarity, as stated
in Eq. (2.3)), where
⎡ ⎤ ⎡ ⎤
A11 0 · · · 0 0 A12 · · · A1m
⎢ 0 A22 · · · 0 ⎥ ⎢0 0 · · · A2m ⎥
⎢ ⎥ ⎢ ⎥
DA = ⎢ . . . . ⎥ and N A = ⎢ . . . . ⎥.
⎣ .. .. . . .. ⎦ ⎣ .. .. . . .. ⎦
0 0 · · · Amm 0 0 ··· 0
Define the matrix
⎡ ⎤
0 Jn 1 ×n 2 · · · Jn 1 ×n m−1 Jn 1 ×n m
 ⎢
⎢0 0 · · · Jn 2 ×n m−1 Jn 2 ×n m ⎥ ⎥
⎢ ⎥
N = max (N A )i j ⎢ ... ..
.
..
.
..
.
..
. ⎥, (5.1)
i, j ⎢ ⎥
⎣0 0 ··· 0 Jn m−1 ×n m ⎦
0 0 ··· 0 0
where Jn i ×n j is a matrix of size n i × n j , with all of its entries being 1. Suppose J =
 
diag Jn 1 , Jn 2 , · · · , Jn m , where Jn i is a square matrix of order n i , with all of its entries
 m
being 1. Observe that N = 0 and J k = J for any k ∈ Z+ . And, J ⊗ N = N ⊗ J = N .

Lemma 5.1 For any k ∈ Z+ that is bigger than the index of nilpotency of N A , we have
Ak  D kA ⊕ N where N is some nilpotent matrix, whose index of nilpotency is at most m.

123
170 Discrete Event Dynamic Systems (2023) 33:161–179

Proof For a reducible matrix A = D A ⊕ N A , we observe that for any k ∈ Z+ with k ≥ m,

we have
Ak = (D A ⊕ N A )k
  
= D kA D k−1 k−2
A NA ⊕ DA NA DA ⊕ · · · ⊕ NA DA
k−1

   
D k−2 k−3 2 k−2
A NA ⊕ DA NA DA NA ⊕ · · · ⊕ NA DA
2
··· N Ak .(5.2)
 
k−r N r = J N r ≤ N r . In fact, for any product that involves
Further, D k−r
A NA ≤ J
r
A A
(k − r ) factors of D A and r factors of N A , we have
   r
k−r −1 r −1
D k−r
A N r
A ⊕ D A N A D A N A ⊕ · · · ⊕ N r k−r
D
A A ≤ N .

Thus,  2  m−1
Ak ≤ D kA ⊕ N ⊕ N ⊕ · · · ⊕ N .
 2  m−1
Define N = N ⊕ N ⊕ · · · ⊕ N , a nilpotent matrix whose index of nilpotency
is at most m. 

With the help of Lemma 5.1, we now find the limit of Ak when μ(A) < 1, in the following
lemma.

Lemma 5.2 Let A be a matrix with μ(A) < 1. Then, lim Ak = 0.

k →∞

Proof From the Fröbenius normal form of A, we know that we can express A = D A ⊕ N A
where the diagonal blocks are either irreducible or 1 × 1 null matrix. Then, applying Lemma
2.4 to Aii , there exists a diagonal matrix Di such that Di−1 Aii Di < Jn i , where n i is order
of Aii . However, if Aii is a 1 × 1 null matrix, we choose Di = [1]. Now, define D =
diag(D1 , D2 , · · · , Dm ). Then, D −1 D A D < J .
Defining  
D := max (D A )i j J,
i, j

observe that D and N (as defined in Eq. (5.1)) commute. We know from an analogous
analysis as in Eq. (5.2) in the proof of Lemma 5.1 that for any k ∈ Z+ with k ≥ m, we have
 k  k  k−1  k−m+1  m−1
Ak ≤ D ⊕ N = D ⊕ D N ⊕ ··· ⊕ D N .
 k−r
As k → ∞, D → 0 for every 0 ≤ r ≤ m − 1. Thus, lim Ak = 0. 
k →∞

Remark 5.3 μ(A) < 1 if one of the following conditions holds.

1. A is a nilpotent matrix.
2. B(A) = 0.
3. G (A) does not have any circuits.
Thus, Lemma 5.2 is relevant in all these cases.

6 Proof of Theorem 3.1

In this section, we prove Theorem 3.1. The easy cases are dealt with, in the beginning.

123
Discrete Event Dynamic Systems (2023) 33:161–179 171

6.1 Case - 1: Suppose A is irreducible

Here, we consider the matrix A to be irreducible. The reducible case will follow after the
proof of the irreducible case.

Proof Let A be a given irreducible matrix with μ(A) ≤ 1. By Lemma 2.4, without loss
of generality, we can assume A  Jn . By equation (4.3), we have A = B(A) ⊕ R(A).
Since B(A) is a Boolean matrix, there exists q ≥ 1 such that B(A) is q-periodic. Applying
Theorem 4.3 to A, the sequence {Ak : k ∈ Z+ } has asymptotic period q. That is, for each
j = 1, 2, . . . , q, lim A j+kq = 
A( j) , exists. 
k →∞

6.2 Case - 2: Suppose A is reducible

Recall from Eq. (2.3) that (upto a permutation similarity), we can express A = D A ⊕ N A .
Define = {i ∈ {1, 2, . . . , m} : μ(Aii ) = 1}. Then, for i ∈ , we apply Theorem
t+q (i)
2.5 to deduce that there exist q (i) and ti such that Aii = Aiit for all t  ti . However,
for i ∈
/ , we appeal to Lemma 5.2 to conclude that lim Aiik = 0, i.e., given > 0,
k →∞
there exists an integer ti such that Aiit < Jn i for all t  ti . In this case, q (i) = 1. Now,
defining q = lcm(q (1) , q (2) , . . . , q (m) ) and choosing t0 to be bigger than max{t1 , t2 , . . . , tm },
we obtain for all t ≥ t0 ,

t+q
Aii = Aiit for i ∈ and Aiit < Jn i for i ∈
/ . (6.1)

We present the arguments for the two cases, μ(A) < 1 and μ(A) = 1 separately. We urge
the readers to observe that the first case has already been dealt with in Lemma 5.2 with the
period of the limiting matrix being 1. Thus, we only need to prove the result in the second
case, when μ(A) = 1.
From Theorem 2.6, we know that there exists a positive vector z ∈ Rn+ such that D −1 AD 
Jn , where D = diag(z 1 , z 2 , · · · , z n ). Further, from Equation (4.3), we have that D −1 AD =
B(A) ⊕ R(A). Since B(A) is a Boolean matrix, there exist positive integers q and t0 such
that
(B(A))t+q = (B(A))t ∀t > t0 .

Observe that this positive integer q is the same as the one that we obtain in Eq. (6.1). Hence,
Theorem 4.3 implies that

lim D −1 Akq+ j D = D −1

A( j) D.
k →∞

Since by Theorem 2.6, any matrix and its scaling have the same dynamics, the proof of
the existence of the limiting matrix is complete.
We now prove that the limiting matrix  A( j) is periodic, the arguments involve a clever
usage of Elsner’s Theorem 2.5, a certain splitting of the max sum and comparison arguments.
We prove this by induction on the number of diagonal blocks in the Fröbenius normal form
of the given matrix A.

123
172 Discrete Event Dynamic Systems (2023) 33:161–179

For 1 ≤ j ≤ q, we know from the Fröbenius normal form of A that there exists a
permutation matrix P such that

Here, for i ∈ ,

( j) kq+ j t + j+(kq−t0 )
Aii = lim Aii = lim Aii0 .
k →∞ k →∞

Suppose t0 = rq + s where 0 ≤ s ≤ q − 1, observe that

( j) t + j+(kq−rq−s) t + j+q(k−r −1)+q−s t + j+q(k−r −1)+s

Aii = lim Aii0 = lim Aii0 = lim Aii0
k →∞ k →∞ k →∞
t + j+s
= Aii0 ,

where 0 ≤ s ≤ q − 1. Assuming t0 to be a multiple of q, as assured by Eq. (6.1), we have


t +j
( j) Aii0 when μ(Aii ) = 1
Aii = (6.2)
0 when μ(Aii ) < 1,

where the latter case is given by Lemma 5.2. We continue with the proof, however, only for
the case when j = q. For any other 1 ≤ j < q, we note from Eq. (6.2) that we merely
j
multiply with a factor of Aii to find the appropriate limit. For notational convenience, we
supress the superscript in the notation   the components of the matrix
A(q) and merely write A,
also following suit.
We start with the base case of induction when m = 2 and when μ(A11 ) = 1 and μ(A22 ) <
1. Then, we have
⎡ t (t+q) t (t+q−1) ⎤ ⎡ t t t (t−1) ⎤
 t+q
0
A11 0
A11 
A12 0
A11 0
A11 
A12  t

A = ⎣ ⎦ = ⎣ ⎦ = A ∀t > t = 1.
0 0 0 0

A similar exercise holds when μ(A11 ) < 1 and μ(A22 ) = 1 and when μ(A11 ) < 1 and
μ(A22 ) < 1. Finally, in the case when μ(A11 ) = μ(A22 ) = 1, we have
⎡ ⎤
t0 (t+q) 
t+q−1
0   t0 (t+q−−1)
 t+q ⎢ A11 At11 A12 A22 ⎥

A = ⎢
⎣
=0 ⎥.
⎦
t (t+q)
0
0 A22

123
Discrete Event Dynamic Systems (2023) 33:161–179 173

Consider


t+q−1
t (t+q−−1)

t−2
t (t+q−−1)

t+q−2
t (t+q− −1)
 t (t+q−1)
0  0 t0  0 0  0
At11 A12 A22 = A11 A12 A22 At11 A12 A22 0
A11 A12
=0 =0  = t−1


t−2
t0  0 t (t−−1)+t0 q
 t (t−1)+t0 q

t+q−2
t (t+q− −1)
0  0
= A11 A12 A22 0
A11 A12 At11 A12 A22
=0  = t−1


t−1
t0  t0 (t−−1)

t+q−2
t (t+q− −1)
0  0
= A11 A12 A22 At11 A12 A22 .
=0  = t−1

Whenever t +1 ≥ q, observe that for every t −1 ≤  ≤ t +q −2, there exists a 0 ≤  ≤ t −1

such that
t0   t0 (t−−1) 0   t0 (t+q− −1)
A11 A12 A22 = At11 A12 A22 .
Thus,
⎡ ⎤
⎡ t (t+q) ⎤ 
t+q−1
0  t (t+q−−1)
 t+q
0
A11 0  ⎢0 At11 0
A12 A22 ⎥

A =⎣ ⎦ ⎢ =0 ⎥
t (t+q)
⎣ ⎦
0
0 A22
0 0
⎡ t−1 ⎤
⎡ ⎤  t0  t0 (t−−1)
At11
0t
0  ⎢0 A11 
A12 A22 ⎥
=⎣ ⎦ ⎢ =0 ⎥
t0 t ⎣ ⎦
0 A22
0 0
 t

= A ∀t > t ≥ q − 1,


proving the periodicity of A.
For the general case, we consider the following partition


At11
0
when μ(A11 ) = 1
where 
B=
A11 =
0 when μ(A11 ) < 1.
 a block upper triangular matrix of order
Further, owing to the induction hypothesis, D,
m − 1 is a q-periodic matrix. Thus, there exists a t1 (multiple of q) such that for all t 
(t+q) = D
t1 , D t . Now, for t ≥ t1 + q + 1, consider


t+q−1 
q 
t+q−1
 D
B C D
(t+q−−1) = C (t+q−1)  D
B C (t+q−−1)  D
B C (t+q− −1)
=0 =1  = q+1


q 
t−1
D
=C (t−1)  D
B C (t−−1)+q  D
B  +q C (t+q− −q−1)
=1  =1

123
174 Discrete Event Dynamic Systems (2023) 33:161–179


q 
t−1
=  D
B C (t−−1)  D
B C (t− −1)
=1  =0

t−1
=  D
B C (t− −1) .
 =0

Thus,
⎡ ⎤
⎡ ⎤ 
t+q−1

B (t+q) 0⎢ 0 
B 
C 
D (t+q−−1)
⎥
 (t+q)

A =⎣ ⎦ ⎢ =0 ⎥
⎣ ⎦
0 D (t+q)
0 0
⎡ ⎤
⎡ ⎤  
t−1
t
B 0 ⎢0 
B C D (t+q−−1)
⎥  t
=⎣ ⎦ ⎢ =0 ⎥ = A  for all t  t1 + q + 1
⎣ ⎦
0 D t
0 0


proving the periodicity of A.
Finally, for any x ∈ R+ , define ξx := 
( j) ( j)
n A( j) ⊗ x. Now we show that ξx is a periodic
point of A with its period dividing q. Consider
( j) ( j)
Aq ⊗ξx = Aq ⊗ lim Akq+ j ⊗x = lim Akq+q+ j ⊗x = lim A(k+1)q+ j ⊗x = ξx .
k →∞ k →∞ k →∞
( j)
Hence ξx is a periodic point of A with its period dividing q.

7 Proofs of Theorems 3.3 and 3.4

We start this section with the hypothesis that {A1 , A2 , · · · , A N } is a collection of pairwise
p
commuting nonnegative matrices with μ(Ai )  1. Consider a p-lettered word ω ∈  N ,
possibly with all letters present at least once. Let Aω be the matrix associated with the
p-lettered word ω, as defined in Eq. (3.1). From Theorem 2.9, we know that
μ(Aω )  μ(Aω p ) × μ(Aω p−1 ) × . . . × μ(Aω1 )  1.

Proof [of Theorem 3.3] Applying Theorem 3.1 to Ai , we obtain an integer qi such that for
every 1 ≤ j ≤ qi , we have

kqi + j ( j)
lim Ai = Ai .
k →∞

Define q := lcm (q1 , q2 , · · · , q N ). As the matrices commute, we see from the associa-
tivity of ⊗ that
p p p
Aω = A1 1 ⊗ A2 2 ⊗ · · · ⊗ A NN ,
where pi is the number of occurrences of i in the p-lettered word ω with p = p1 + p2 +
kp kp kp
· · · + p N . Observe that Akω = A1 1 ⊗ A2 2 ⊗ · · · ⊗ A N N from which, it follows that
N 
  pi
 ( j)
lim Akq+
ω
j
= Ai .
k →∞
i =1

123
Discrete Event Dynamic Systems (2023) 33:161–179 175

Finally, by our definition of q, we observe that it is independent of the choice of the word
ω as well as its length. For any word ω that may not contain all the letters, one makes an
analogous argument with a set that has fewer letters. 
The following is an obvious corollary to the above theorem.
Corollary 7.1 Let A1 and A2 be two commuting matrices with all the max eigenvalues being
p
either 0 or 1. Suppose ω ∈ 2 is such that the letter i occurs pi many times and has the
associated matrix product Aω . Then, there exist q and t0 such that for any 1 ≤ j ≤ q, we
have
lim Akq+
ω
j
= L (ωj) ,
k →∞
 
( j) p t p t p (t +1) p (t0 +1) p (t0 +q−1) p (t0 +q−1)
where Lω ∈ A1 1 0 ⊗ A2 2 0 , A1 1 0 ⊗ A2 2 , · · · , A1 1 ⊗ A2 2 .
We now prove Theorem 3.4 with the supposition that the collection {A1 , A2 , · · · , A N }
is not necessarily pairwise commuting. A key hypothesis in the proof of Theorem 3.4 is the
existence of a set of common max eigenvectors for the collection {A1 , A2 , · · · , A N }.
Proof [of Theorem 3.4] Let E = {v1 , v2 , · · · , vm } be the set of common max eigenvectors
for the collection {A1 , A2 , · · · , A N }. We rearrange the eigenvectors (and rename them, if
necessary) as
{v1 , v2 , · · · , vκ , vκ+1 , vκ+2 , · · · , vm }
p
so that Ai ⊗ v j = v j for all 1  i  N and 1  j  κ. Assuming that the word ω ∈  N
contains all the letters at least once, it follows that for every κ +1  j  m, lim Akω ⊗v j =
k →∞
0. Now, for any x = α1 v1 ⊕ α2 v2 ⊕ · · · ⊕ αm vm ∈ LC (E), consider
        
lim Akω ⊗ x = α1 lim Akω ⊗ v1 ⊕ α2 lim Akω ⊗ v2 ⊕ · · · ⊕ αm lim Akω ⊗ vm
k →∞ k →∞ k →∞ k →∞
= α1 v1 ⊕ α2 v2 ⊕ · · · ⊕ ακ vκ
=: ξx ,

with ξx being a fixed point for every matrix in the collection {A1 , A2 , · · · , A N }. 

8 Examples

In this section, we illustrate the theorems in this article with a few examples. We begin with
an example that concerns Theorem 3.1.
⎡ ⎤
0.2 1 4
Example 8.1 Let A = ⎣ 1 0.5 6 ⎦ with corresponding graph given by
0 0 0.9

123
176 Discrete Event Dynamic Systems (2023) 33:161–179

Recall that the critical circuit is the one where the maximum circuit geometric mean is
attained in G (A); in this case, it is (2, 1)(1, 2), that yields μ(A) = 1 and using Remark 4.4,
we also obtain q = 2. Then, for any k ∈ Z+ , we have
⎡ ⎤ ⎡ ⎤
0.5 1 5.4 1 0.5 6
A2k+1 = ⎣ 1 0.5 6 ⎦ while A2k = ⎣0.5 1 5.4 ⎦ .
0 0 (0.9) 2k+1 0 0 (0.9)2k
Thus, ⎡ ⎤ ⎡ ⎤
0.5 1 5.4 1 0.5 6

A(1) = ⎣ 1 0.5 6 ⎦ while 
A(2) = ⎣0.5 1 5.4⎦ .
0 0 0 0 0 0
Hence, for any x = (x1 , x2 , x3 )t ∈ R3+ , it is easy to verify that
⎛ ⎞ ⎛ ⎞
0.5x1 ⊕ x2 ⊕ 5.4x3 x1 ⊕ 0.5x2 ⊕ 6x3
lim A2k+1 ⊗x = ⎝ x1 ⊕ 0.5x2 ⊕ 6x3 ⎠ whereas lim A2k ⊗x =⎝0.5x1 ⊕ x2 ⊕ 5.4x3 ⎠.
k →∞ k →∞
0 0
The following example illustrates Theorem 3.3 and Corollary 7.1.
Example 8.2 Consider the collection {A1 , A2 , A3 } of pairwise commuting 5 × 5 nonnegative
matrices given by
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
010 8 5 00199 100 8 9
⎢0 0 1 5 8 ⎥ ⎢ 1 0 0 9 9⎥ ⎢0 1 0 8 9 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
A1 = ⎢1 0 0 8 5 ⎥ , A2 = ⎢0 1 0 9 9⎥ and A3 = ⎢
⎢ ⎥ ⎢ ⎥
⎢0 0 1 8 9 ⎥
⎥
⎣0 0 0 1 0.5⎦ ⎣ 0 0 0 1 1⎦ ⎣0 0 0 0.8 1 ⎦
0 0 0 0.5 1 00011 0 0 0 1 0.8
with μ(Ai ) = 1 for all i. Observe that all the max eigenvalues of the three matrices are
equal to 1. Then, owing to the Theorem 3.1, we find that the matrices A1 , A2 and A3 have
kq + j 
( j)
asymptotic periods q1 , q2 and q3 given by 3, 3 and 2 respectively, i.e., lim Ai i = Ai .
k →∞
We now discuss all possible cases of ω ∈ 32 , with distinct letters.
kq + j
(i) Let Aω = A1 ⊗ A2 . In this case, we find that lim Aω ω = Aω (irrespective of j)
k →∞
is a matrix with period 1, while qω = 3.
kq + j ( j) ( j)
(ii) Let Aω = A2 ⊗ A3 . Then, lim Aω ω = L ω , where L ω has period 3 while qω = 6.
k →∞
( j)  
Further, L ω ∈ A2 ⊗ A3 , A22 ⊗ A23 , A32 ⊗ A33 .
kq + j ( j) ( j)
(iii) Let Aω = A1 ⊗ A3 . In this case, lim Aω ω = L ω , where L ω has period qω = 6.
k →∞
( j)  
Further, L ω ∈ A1 ⊗ A3 , A21 ⊗ A23 , · · · , A51 ⊗ A53 .
For a general p-lettered word ω that contains each of the letters at least once, suppose pi
kq+ j ( j)
counts the number of occurrences of the letter i in ω, we find that lim Aω = L ω . One
k →∞
( j)
obtains the period of L ω to be either 1 or 3 while q = 6, here. This is a special occasion
since the second diagonal block of the matrices A2 and A3 , expressed in their Fröbenius
normal form, when multiplied yields J2 . Thus, the period of A3 makes no contribution in the
( j)
computation of the period of L ω . Hence,
 
L (ωj) ∈ A1 1 ⊗ A2 2 ⊗ A3 3 , A1 1 ⊗ A2 2 ⊗ A3 3 , A1 1 ⊗ A2 2 ⊗ A3 3 .
p p p 2p 2p 2p 3p 3p 3p

123
Discrete Event Dynamic Systems (2023) 33:161–179 177

We conclude this section and the article with an example in the non-commuting case, that
explains Theorem 3.4.

Example 8.3 Consider the matrices A1 and A2 , where

⎡ ⎤ ⎡ ⎤
0.9 0.45 5 6 27 1 1 27 6 2
⎢0.45 0.9 1 23 8 ⎥ ⎢0.5 1 17 3 23 ⎥
⎢ ⎥ ⎢ ⎥
A1 = ⎢ ⎢ 0 0 0.9 1 0 ⎥ ⎢
⎥ and A2 = ⎢ 0 0 0.4 1 0⎥ ⎥.
⎣ 0 0 0 0.2 1 ⎦ ⎣0 0 1 0.8 1⎦
0 0 1 0 0 0 0 0.9 1 0.2

The max eigenvalues for the matrix A1 are 1 and 0.9, while those of A2 are 1 and 1, with
the corresponding common max eigenvectors for the matrices being u = (27, 23, 1, 1, 1)T
p
and v = (2, 1, 0, 0, 0)T respectively. Suppose ω ∈ 2 with the presence of both the letters
at least once. For x ∈ LC (E) with x = αu ⊕ βv for some α, β  0, we have

lim Akω ⊗ x = ξx = αu,

k →∞

where ξx is a common fixed point of A1 and A2 .

Acknowledgements The authors are thankful to all the referees, whose helpful comments have shortened the
proof of Theorem 3.1 considerably.

Declaration
The authors declare that they have no conflict of interests.

References
Bapat RB (1998) A max version of the Perron-Fröbenius theorem. Linear Algebra Appl 275(376):3–18
Bapat RB, Stanford DP, van den Driessche P (1995) Pattern properties and spectral inequalities in max algebra.
SIAM J Matrix Anal Appl 16:964–976
Berman A, Plemmons RJ (1994) Nonnegative matrices in the mathematical sciences 9, Society for Industrial
and Applied Mathematics (SIAM), Philadelphia, PA
Butkovic P (2010) Max-linear Systems: Theory and Algorithms Springer Monographs in Mathematics 27
Elsner L, van den Driessche P (1999) On the power method in max algebra. Linear Algebra Appl 302(303):
17–32
Elsner L, van den Driessche P (2001) Modifying the power method in max algebra. Linear Algebra Appl
332(334):3–13
Gaubert S (1994) Rational series over dioids and discrete event systems. In: G. Cohen, J.P. Quadrat (Eds.), Pro-
ceedings of the 11th International conference on analysis and optimization of systems, Sophia- Antipolis,
France, June 1994, Lecture Notes in Control and Information Sciences, 199, Springer, London, 1994, pp
247–256
Gavalec M (2000) Linear matrix period in max-plus algebra. Linear Algebra Appl 307:167–182
Gavalec M (2000) Computing orbit period in max-min algebra. Discrete Appl Math 100:49–65
Heidergott B, Oldser GJ, van der Woude J (2006) Max plus at work Princeton Series in Applied Mathematics.
Princeton University Press
Jayaraman S, Prajapaty YK, Sridharan S (2022) Dynamics of products of nonnegative matrices. Extracta Math
37:223–242
Lemmens B, Nussbaum RD (2012) Nonlinear perron-frobenius theory. Cambridge Tracts in Mathematics 189,
Cambridge University Press
Katz RD, Schneider H, Sergeev S (2012) On commuting matrices in max algebra and in classical nonnegative
algebra. Linear Algebra Appl 436:276–292
Pang CT, Guu SM (2001) A note on the sequence of consecutive powers of a nonnegative matrix in max
algebra. Linear Algebra Appl 330:209–213

123
178 Discrete Event Dynamic Systems (2023) 33:161–179

Rosenblatt D (1957) On the graphs and asymptotic forms of finite Boolean relation matrices and stochastic
matrices. Naval Res Logist Quart 4:151–167
de Schutter B (2000) On the ultimate behavior of the sequence of consecutive powers of a matrix in the
max-plus algebra. Linear Algebra Appl 307:103–117

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.

Shrihari Sridharan is a faculty member in the School of Mathematics

at IISER-TVM.

Sachindranath Jayaraman is a faculty member in the School of Math-

ematics at IISER-TVM.

123
Discrete Event Dynamic Systems (2023) 33:161–179 179

Yogesh Kumar Prajapaty is a graduate student in the School of Mathe-

matics at IISER-TVM, working under the joint supervision of Sachin-
dranath Jayaraman and Shrihari Sridharan.

123