Journal Pre-proof

Dominant eigenvalue and universal winners of digraphs

Sirshendu Pan, Sukanta Pati, Steve Kirkland

PII: S0024-3795(24)00196-4
DOI: https://doi.org/10.1016/j.laa.2024.04.033
Reference: LAA 16758

To appear in: Linear Algebra and its Applications

Received date: 5 January 2024

Revised date: 27 April 2024
Accepted date: 28 April 2024

Please cite this article as: S. Pan et al., Dominant eigenvalue and universal winners of digraphs, Linear Algebra Appl. (2024),
doi: https://doi.org/10.1016/j.laa.2024.04.033.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and
metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional
copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early
visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content,
and all legal disclaimers that apply to the journal pertain.

© 2024 Published by Elsevier.

 Dominant eigenvalue and universal winners of digraphs

Sirshendu Pana,∗, Sukanta Patia , Steve Kirklandb

a Department of Mathematics, Indian Institute of Technology Guwahati, Guwahati, India
b Department of Mathematics, University of Manitoba, Winnipeg, MB, Canada

Abstract

Let A be a nonnegative irreducible matrix of order n and Ai (t) be the matrix obtained by increasing its
ith diagonal entry by a positive number t. An index i ∈ [n] = {1, . . . , n} is called a universal winner for
A, if, letting ρ(·) denote the spectral radius, ρ(Ai (t)) ≥ ρ(Aj (t)) for j ∈ [n] and for t ∈ (0, ∞).
Let G be a strongly connected digraph with vertex set [n]. Let WG be the set of all nonnegative
matrices whose underlying graph structure is G. We say a vertex u structurally dominates another
vertex v in G, if ρ(Au (t)) ≥ ρ(Av (t)) for all t ∈ (0, ∞) and for all A ∈ WG . We characterize the class
of digraphs G that do not have a vertex that structurally dominates all other vertices. We say two
vertices u and v structurally tie if they structurally dominate each other in G. We supply an equivalent
graph theoretic condition for the structural tying of two vertices in G. Let S ⊆ [n] be nonempty. We
characterize the class of strongly connected digraphs G with vertex set [n] such that S is the set of
universal winners for each A ∈ WG . We also characterize the strongly connected digraphs whose vertex
set can be partitioned into subsets P1 , P2 , . . . , Pk such that vertices inside a part Pi structurally tie with
each other and vertices of Pi structurally dominate vertices of Pj strictly (without tying) for i < j.
Keywords: Weighted digraphs, Adjacency matrices, Spectral radius, Universal winners, Structural
domination.
MSC 05C50

1. Introduction

A real square matrix is said to be essentially nonnegative if all of its off-diagonal entries are non-
negative (see for example, [6]). Such a matrix A can be written as C + αI, where C is a nonnegative
matrix and α is a real number. So A has a real eigenvalue that is greater than or equal to the real part
of any other eigenvalue of A; equivalently, the spectral abscissa of A is an eigenvalue. This is called the
dominant eigenvalue of the essentially nonnegative matrix A. The dominant eigenvalue of an essentially
nonnegative matrix A need not have the largest modulus among its eigenvalues but when A is nonneg-
ative, by the Perron-Frobenius theorem, it is indeed the spectral radius ρ(A) of A. The behavior of the

∗ Corresponding author
Email addresses: [email protected] (Sirshendu Pan), [email protected] (Sukanta Pati),
[email protected] (Steve Kirkland)
dominant eigenvalue of an essentially nonnegative matrix, when some of its entries (not necessarily the
diagonal ones) are perturbed, has been studied extensively in the literature (see for example, [7, 8, 4]).
Given an essentially nonnegative matrix A, the problem of determining the minimum value of the
dominant eigenvalue of the sum of A and a real diagonal matrix D with trace tr(D) = 0 was introduced
by Johnson et al. [3] in 1994. They determined the minimum value of the dominant eigenvalue and
found all diagonal matrices for which the minimum is achieved. We state that result below. Here, e
denotes the column vector of all ones and λmax (A + D) denotes the dominant eigenvalue of A + D. By
AT we mean the transpose of the matrix A. A matrix A is line-sum-symmetric provided it satisfies
Ae = AT e, and it has a line-sum-symmetric diagonal similarity if there exists a diagonal matrix C with
positive diagonal such that CAC −1 is line-sum-symmetric. The matrix CAC −1 is said to be a line-
sum-symmetric diagonal similarity of A. It is well known that for an irreducible essentially nonnegative
matrix A, the matrix C such that CAC −1 is line-sum-symmetric is unique up to positive multiples (see
[3]).

Theorem 1.1. [3] Let A be an irreducible essentially nonnegative matrix of order n, and B be the
line-sum-symmetric diagonal similarity of A. Then

1 T
τ (A) := min{λmax (A + D) | D is diagonal and tr(D) = 0} = e Be
D n

1 T



and the minimum is only achieved by the matrix diag(τ (A)e − Be) = diag n e Be e − Be .

Looking at this beautiful result one would naturally be curious about whether

max{λmax (A + D) | D is diagonal and tr(D) = 0}

D

is bounded from above. However, it is easy to see that it is not. It is therefore natural to consider
the variant of the problem where the diagonal perturbations are constrained to be nonnegative with
trace t. Also, considering addition by a positive scalar matrix, the problem can be considered for an
irreducible nonnegative matrix. In that case, λmax (A) is just ρ(A). Indeed, Johnson, Loewy, Olesky,
van den Driessche [2] considered maximizing the spectral radius of an irreducible nonnegative matrix A
under a nonnegative, diagonal trace t perturbation. While that problem is not yet completely solved,
there are many interesting results on it, both matrix theoretic and combinatorial. It has been observed
that this problem has a strong connection with graph structures, and the results in this paper reinforce
that connection.
But first, let us describe some of the results given in [2]. We shall use some of that paper’s notation
and definitions. We shall use the notation A = [aij ] ≥ 0 for a nonnegative matrix. Let A ≥ 0 be an

2
irreducible matrix. Fix t > 0. Define

µ(A) := max{λmax (A + D) | D is diagonal and tr(D) = t}.

D

Let Eii denote the matrix whose only nonzero entry 1 is at the (i, i)-th position. Since the spectral
radius of a nonnegative matrix is a convex function of its diagonal entries (see [5]), and tr(D) = t, we
see that !
n n
X dii X dii
ρ(A + D) = ρ (A + tEii ) ≤ ρ(A + tEii ) ≤ max ρ(A + tEii ).
i=1
t i=1
t i

Hence µ(A) is attained by a matrix tEii , for some i. Thus, µ(A) = max ρ(A + tEii ).
1≤i≤n
The above characterization of µ(A) leads us to study the following functions of t ∈ [0, ∞),

µ(A; t) := max ρ(A + tEii )

1≤i≤n (1.1)
ρ(Ai (t)) := ρ(A + tEii ), for 1 ≤ i ≤ n.

Note that each of the above n functions is strictly increasing functions of t as A is irreducible. The indices
i at which µ(A; t) is attained may vary with t. To illustrate this, we supply the following example.

Example 1.2. Consider the following matrix A ≥ 0. It can be easily shown that it is an irreducible
matrix.  
4 1 0
 
1 0 6.
A= 
0 6 0

Here for t ∈ (0, 4.8), ρ(A2 (t)) = ρ(A + tE22 ) > ρ(A1 (t)) and ρ(A2 (t)) > ρ(A3 (t)). Thus µ(A; t) is
attained at the index 2 for all t ∈ (0, 4.8). But for large values of t, µ(A; t) occurs at the index 1. We
present the justifications in Example 2.2 below.

Example 1.2 motivates us to explore, for a given irreducible matrix A, whether µ(A; t) is attained
at the same index for all large values of t, µ(A; t) is attained at the same index for all small values of t.
This was studied in [2] from which we extract the following definition.

Definition 1.3. Let A ≥ 0 be irreducible. Assume that µ(A; t) and ρ(Ai (t)) are as defined in (1.1).

1. If ρ(Ai (t)) = ρ(Aj (t)) for all t ≥ 0, then we say the index i ties the index j.

2. If ρ(Ai (t)) > ρ(Aj (t)) for all t ∈ (t1 , t2 ), where 0 ≤ t1 < t2 then we say ‘i dominates j in (t1 , t2 )’.

3. Let t1 > 0. If i dominates or ties j in (0, t1 ) for each j, then i is called an initial winner.

4. Let t1 > 0. If i dominates or ties j in (t1 , ∞) for each j, then i is called a terminal winner.

5. Finally, if i dominates or ties j in (0, ∞) for each j, then i is called an universal winner.

3
 6. Let 0 < t0 < t1 < t2 . If j dominates i in (t0 , t1 ) and i dominates j in (t1 , t2 ), then we say the
indices i and j switch at t1 . Since the spectral radii ρ(Ai (t)) and ρ(Aj (t)) are continuous functions
of t, it follows that, if there is a switch between i and j at t1 > 0 then ρ(Ai (t1 )) = ρ(Aj (t1 )).

Notice that in Example 1.2 there is no universal winner. Instead, there is an initial winner (index 2)
and a terminal winner (index 1).
The following two theorems ensure that there is at least one initial winner and at least one terminal
(k)
winner. We use the notation aij to denote the (i, j)-th entry of the k-th power Ak of a matrix A. By
+(k)
aij we mean the (i, j)-th entry of the (A+ )k , where A+ is the Moore-Penrose inverse of A.

Theorem 1.4. [2] Let A = [aij ] ≥ 0 be an irreducible matrix of order n. Define

n o
(k)
S0 := {1, 2, . . . , n} and Sk := s ∈ Sk−1 | a(k)
ss = max aii , for k = 1, 2, . . . , n − 1.
i∈Sk−1

Then the elements of Sn−1 are precisely the terminal winners. In particular, if Sk = {i} for some
k ∈ {1, . . . , n − 1} and i ∈ {1, . . . , n}, then i is the only terminal winner.

Theorem 1.5. [2] Let A = [aij ] ≥ 0 be an irreducible matrix of order n. Let x = [xi ] and y = [yi ],
respectively, be right and left Perron vectors of A with y T x = 1. Denote the matrix ρ(A)I − A by Q.
Define

T0 := {1, . . . , n},

T1 := {s ∈ T0 | xs ys = max xi yi }, and
i∈T0
n o
k +(k−1) +(k−1)
Tk := s ∈ Tk−1 | (−1) qss = max (−1)k qii , for k = 2, . . . , n.
i∈Tk−1

Then the elements of Tn are precisely the initial winners. In particular, if for some positive k, the
set Tk = {i} for some i, then i is the only initial winner. Moreover, for any two indices i and j, if
xi yi > xj yj , then ρ(Ai (t)) > ρ(Aj (t)) for all sufficiently small t > 0.

Let A ≥ 0 be an irreducible matrix of order n. Can the functions ρ(Ai (t)) and ρ(Aj (t)) be the
same over [0, ∞)? The following theorem provides an answer. For a matrix A, the characteristic
polynomial det(λIn − A) is denoted by φ(A, λ) or simply by φ(A). Here In is the identity matrix of
order n. We shall use A(i | i) to denote the principal submatrix of A obtained by deleting the i-th row
and column of A. The following theorem is taken from [2] with some notation changes.

Theorem 1.6. [2] Let A = [aij ] ≥ 0 be irreducible. Fix indices i and j. Then the following statements
are equivalent.

(i) The index i ties with index j, that is, ρ(Ai (t)) = ρ(Aj (t)), ∀t ≥ 0.

(ii) The characteristic polynomials φ(A(i | i), λ) and φ(A(j | j), λ) are equal.

4
 (k) (k)
(iii) The i-th and j-th diagonal entries of Ak are equal for all k ∈ N, that is, aii = ajj , k ∈ N.

(We note in passing that if A is the adjacency matrix of a weighted undirected graph and indices i, j
satisfy one (i)–(iii) of Theorem 1.6, then the corresponding vertices are called cospectral in the spectral
graph theory literature.)
Suppose that ρ(Ai (t)) and ρ(Aj (t)) are not identical. In this case, the following theorem facilitates
an upper bound on the number of values of t where the equality ρ(Ai (t)) = ρ(Aj (t)) holds.

Theorem 1.7. [2] Let A = [aij ] ≥ 0 be an irreducible matrix of order n, and let 1 ≤ i, j ≤ n be
fixed. If there exists an α > 0 such that ρ(Ai (α)) = ρ(Aj (α)), then ρ(Ai (α)) is a root of equation
φ(A(i | i), λ) − φ(A(j | j), λ) = 0. Conversely, if β > ρ(A) is a root of φ(A(i | i), λ) − φ(A(j | j), λ) = 0
then there exists a unique α > 0 such that β = ρ(Ai (α)) = ρ(Aj (α)).

As φ(A(i | i), λ) − φ(A(j | j), λ) is a polynomial of degree at most n − 2, using the previous theorem
we conclude that between the indices i and j at most n − 2 switches can happen.
Let us now limit our attention to irreducible matrices with an underlying graph structure. By a
graph we shall mean an undirected graph and by a digraph we shall mean a directed graph. We shall
consider only simple graphs and digraphs, that is, those without any loops or multiple edges. For a
digraph G with V (G) = [n], let WG be the set of all square nonnegative matrices A = [aij ] of order n
such that aij > 0 if and only if (i, j) is a directed edge in G. We define WG for an undirected graph
similarly, considering only symmetric matrices. Note that the adjacency matrix A = A(G) of G is the
matrix in WG that has all its entries from the set {0, 1}. A matrix from WG will be called a weight
matrix of G.
Suppose that G is a given graph or digraph with V (G) = [n]. Fix A ∈ WG and t > 0. The problem
of determining the vertices i at which µ(A; t) is achieved is an interesting combinatorial subject. It
was first studied by Fallat, Olesky and van den Driessche [1] in 1997 and it has since yielded several
interesting results. We note a few of them below.

Theorem 1.8. [1, Corollary 3.3] Let A be the adjacency matrix of a graph G. Let u, v be two vertices
such that u = f (v) for some automorphism f of G. Then u and v tie for A.

Theorem 1.9. [1, Theorem 2.2] Let G = Sn be the star graph with V (G) = [n]. Then for every A ∈ WG ,
only the central vertex is the universal winner.

The following result characterizes the directed cycle using the concept of universal winners.

Theorem 1.10. [1, Theorem 2.1] Let G be a digraph. For every A ∈ WG , all the vertices of G are
universal winners if and only if G is a directed cycle.

Motivated by the above, we plan to further explore the relationship between the universal winner
vertices and the graph structures. Let G be a strongly connected digraph and u, v be two vertices of G.

5
 Then we say u structurally dominates v, if ρ(Au (t)) ≥ ρ(Av (t)) for all t ∈ (0, ∞) and for each weight
matrix A of G. We say the vertices u and v structurally tie if they structurally dominate each other in
G, that is, u and v tie for every weight matrix A of G. Finally, if u structurally dominates v, without
tying, we say u structurally dominates v strictly.
In this paper we consider the following two questions.

Question 1. Let G be a digraph with V (G) = [n]. Suppose that there is a set S of k vertices
such that for every weight matrix A of G, the vertices of S are the only
universal winners. When k = n, a characterization of G is available (see [1,
Theorem 2.1]). Can we supply a characterization of G for other values of k?

Question 2. Given a natural number k, is it possible to find a strongly connected digraph G

whose vertex set can be partitioned into subsets P1 , . . . , Pk such that the vertices
inside Pi structurally tie with each other and the vertices of Pi structurally
dominate those of Pj strictly for i > j. Can we characterize all such digraphs?

We answer these questions in Theorems 4.5 and 4.7, respectively.

The article is organized as follows. In Section 2, we supply some new general results pertaining to
µ(A) of nonnegative matrices, which are used to answer questions in the later sections. In Section 3, we
consider strongly connected digraphs and prove an equivalent graph theoretic condition for structural
tying between two vertices. We also provide an equivalent graph theoretic condition for one vertex
to strictly structurally dominate another. We characterize the class of strongly connected digraphs
that do not have a vertex that structurally dominates the others. We also show that in any strongly
connected digraph with at least three vertices, a vertex u that has a unique neighbor (i.e., exactly one
in-neighbor and one out-neighbor both of which are the same vertex) is structurally dominated strictly
by its neighbor. We also develop a useful characteristic polynomial technique and use it to provide an
alternative proof. In Section 4, we answer Questions 1 and 2, and give constructive characterizations of
the respective families of digraphs. We also answer a variant of Question 1 for undirected graphs.

2. Domination and initial winners in a nonnegative irreducible matrix

In this section we provide a sufficient condition for one index to dominate another for all t > 0.
Prior to this, recall that Theorem 1.5 provides a way to identify the initial winners for a nonnegative
irreducible matrix. Below we discuss another useful technique that can be used in finding an initial
winner and for other purposes.

Lemma 2.1. Let A ≥ 0 be an irreducible matrix of order n. Fix indices i < j. Suppose that for some
δ > 0 we have φ(A(i | i), λ) > φ(A(j | j), λ) for all λ ∈ (ρ(A), ρ(A) + δ). Then i dominates j initially.

6
Proof. Suppose that the conclusion does not hold. Recall that ρ(Ai (t)) and ρ(Aj (t)) are equal for at most
n−2 values of t. Hence, we see that, for all t > 0 there exists xt in (0, t) such that ρ(Ai (xt )) < ρ(Aj (xt )).
Now the following two cases can occur.

a) There exists a t0 > 0 such that ρ(Ai (t)) < ρ(Aj (t)) for all t in (0, t0 ).

b) There exists a decreasing sequence {yk } of positive numbers converging to 0 such that ρ(Ai (yk )) <
ρ(Aj (yk )) and ρ(Ai (t)) ≥ ρ(Aj (t)) for all t in (yk+1 , yk ).

First suppose that case (b) occurs. Since the spectral radius is a continuous function of the matrix
entries, it follows that there will be more than n − 2 values of t where ρ(Ai (t)) = ρ(Aj (t)) holds, a
contradiction. Hence only case (a) can hold.
Let t > 0 and λ ∈ (ρ(A), ρ(A) + δ). By the hypothesis, we have φ(A(i | i), λ) > φ(A(j | j), λ). Hence
φ(A, λ) − tφ(A(i | i), λ) < φ(A, λ) − tφ(A(j | j), λ). That is,

φ(Ai (t), λ) < φ(Aj (t), λ) ∀t > 0, λ ∈ (ρ(A), ρ(A) + δ). (2.1)

Observe that for all sufficiently small t > 0 (assume t < t0 ) we have ρ(A) < ρ(Aj (t)) < ρ(A) + δ and
ρ(Ai (t)) < ρ(Aj (t)). Then for such a value of t, we have φ(Ai (t), ρ(Aj (t)) > 0, but φ(Aj (t), ρ(Aj (t)) = 0.
This contradicts Equation (2.1).
 
4 1 0
 
1 0 6 and we stated that the index 1 is
Example 2.2. Recall that in Example 1.2, we had A =  
0 6 0
the terminal winner, whereas index 2 is the initial winner. Here we provide arguments to support those
statements. Using Theorem 1.4, we see that S0 = {1, 2, 3} and S1 = S2 = {1}. So, the index 1 is the
terminal winner. Now to find the initial winners, consider
f (λ) := φ(A(2 | 2), λ) − φ(A(1 | 1), λ) = λ(λ − 4) − (λ2 − 36) = −4λ + 36.
Observe that ρ(A) < 7. So for λ = ρ(A) we have f (λ) > 0. Since f (λ) is a continuous function of λ,
there exists δ > 0 such that f (λ) > 0 in (ρ(A), ρ(A) + δ). So by Lemma 2.1 we conclude that index
2 dominates index 1 initially. Similarly, we show that index 2 dominates index 3 initially. Therefore,
index 2 is the initial winner. Note that, we do not have a universal winner here.

To proceed further we need the following expression for the trace of Ai (t)k .

Lemma 2.3. Let A and Eii be of the order n. Then

k
X X
tr(Ai (t)k ) = tr(A + tEii )k = tr(Ak ) + tj tr(X),
j=1 X∈Arr(k−j,j)

where Arr(k − j, j) is the set of all arrangements of k − j copies of A and j copies of Eii .

7
 (k)
Proof. Notice that Eii Ak Eii = aii Eii , for k ∈ N. Also Eii
2
= Eii . Hence for m > 1 and k1 , . . . , km ∈ N,
we have
(k ) (k ) (k )
tr(Ak1 Eii Ak2 Eii · · · Eii Akm ) = aii 1 aii 2 · · · aii m
= tr(Eii Ak1 Eii Ak2 Eii · · · Eii Akm )
= tr(Ak1 Eii Ak2 Eii · · · Eii Akm Eii )
= tr(Eii Ak1 Eii Ak2 Eii · · · Eii Akm Eii ).

The identity follows.

The following result supplies a sufficient condition based on matrix powers, for an index to dominate
another for all t > 0.
(k) (k)
Lemma 2.4. Let A be a nonnegative matrix of order n. If aii ≥ ajj for all k ∈ N, then i dominates
j for all t > 0. That is, ρ(Ai (t)) ≥ ρ(Aj (t)) for all t > 0.

(k) (k)
Proof. Let t be any positive real number. Since aii ≥ ajj for all k ∈ N, using Lemma 2.3 we get
tr(Ai (t)k ) ≥ tr(Aj (t)k ). It is known that (see for example [4]), if X is a nonnegative matrix, then
lim sup(tr(X k ))1/k = ρ(X). Applying this to Ai (t) and Aj (t), we have
ρ(Ai (t)) = lim sup (tr(Ai (t)k ))1/k ≥ lim sup (tr(Aj (t)k ))1/k = ρ(Aj (t)).

(k) (k)
Remark 2.5. Consider a nonnegative irreducible matrix A of order n. Assume i < j and that aii ≥ ajj
holds for all k ∈ N. Let x and y T , respectively, be right and left Perron vectors of A. Then xi yi ≥ xj yj .
The proof follows by applying Lemma 2.4 and Theorem 1.5.

The converse of the above lemma does not hold, even for irreducible matrices. Here we provide an
example of that.

Example 2.6. Let A be the adjacency matrix of the digraph in Figure 1. Note that

f (λ) = φ(A(2 | 2), λ) − φ(A(3 | 3), λ) = (λ3 − 1) − (λ3 − λ) = λ − 1.

Observe that ρ(A) > 1. So, there is no root of f (λ) = 0 greater than 1. Therefore, by Theorem 1.7

1
2 4

3
(k) (k)
Figure 1: Vertex 2 dominates 3, but a22 ≥ a33 fails for k = 3.

there is no switch between the index 3 and the index 2. At λ = ρ(A), we have f (λ) > 0. By an argument
similar to that used in Example 2.2 we conclude that index 2 dominates index 3 initially. Therefore
(3) (3)
index 2 dominates index 3 for all values of t. However, a22 = 1 < 2 = a33 .

8
3. Some structural results for strongly connected digraphs

We know that for an irreducible nonnegative matrix A, there exists one or more initial winners and
one or more terminal winners. However, as Example 2.2 shows, there need not exist a universal winner.
The situation can be different if we consider some specific types of irreducible nonnegative matrices.
Recall that if G is a digraph, then WG stands for the set of all nonnegative matrices with aij > 0 if and
only if (i, j) ∈ G. We know that if G is a strongly connected digraph, then each weight matrix A of G is
irreducible. Note that the indices for A correspond to vertices of G. In this section we will explore how
the universal winners, terminal winners, and initial winners are related to the digraph’s structure.
Recall that when G is a digraph and A ∈ WG is fixed, then we can view A as the adjacency matrix
(k)
of a weighted digraph. Hence aii is the sum of the weights of all directed walks from i to i involving k
edges in that digraph, where the weight of a walk is the product of the weights of the edges present in
the walk. We refer to a nontrivial closed walk as a closed walk with at least one edge.
For a strongly connected digraph G, the following result gives us two useful graph theoretic conditions
helping to identify the winners. One is an equivalent condition for two vertices to structurally tie, and
the other is a sufficient condition under which a vertex u structurally dominates another vertex v strictly.

Theorem 3.1. Let G be a strongly connected digraph and u, v be two vertices of G. Then

(i) Vertices u and v, structurally tie if and only if every nontrivial closed walk passing through u passes
through v and every nontrivial closed walk passing through v passes through u.

(ii) Suppose that every nontrivial closed walk that passes through v also passes through u and there
exists a nontrivial closed walk passing through u that does not pass through v. Then u structurally
dominates v strictly.

Proof. (i) First suppose that every nontrivial closed walk passing through u also passes through v and
every nontrivial closed walk passing through v also passes through u. Then for every A ∈ WG , we have
(k) (k)
auu = avv for all k ∈ N. The proof follows by applying Theorem 1.6.
For the converse part, suppose that u and v structurally tie. We first show that every nontrivial
closed walk passing through u passes through v. If possible, assume that there exists a nontrivial closed
walk of length l passing through u that does not pass through v. Let  > 0 be a small number. Supply
a weight 1 to all the edges of G except those adjacent to v to which we assign weight . Let A = [aij ]
(l)
be the adjacency matrix of this weighted digraph G. Then avv = f (), where f () is a polynomial in .
(l)
So, for sufficiently small , we get avv < 1. As there exists a nontrivial closed walk of length l passing
(l) (l)
through u that does not pass through v, we have auu > 1 > avv . Since u and v tie for A , by Theorem
(k) (k)
1.6, auu = avv for all k ∈ N. This is a contradiction. Similarly, one can argue that every nontrivial
closed walk passing through v also passes through u.

9
 (ii) Let A ∈ WG . Note that every nontrivial closed walk that passes through v also passes through
u. Thus, by Lemma 2.4 we have

ρ(Au (t)) ≥ ρ(Av (t)) for all t > 0. (3.1)

As every nontrivial closed walk passing through v also passes through u and there is a nontrivial closed
(l) (l)
walk passing through u that does not pass through v, we get that auu > avv for some l ∈ N. By
Theorem 1.6 Item (i), we cannot have ρ(Au (t)) = ρ(Av (t)) for all t > 0. Hence, using (3.1), we see that
∃t > 0 such that ρ(Au (t)) > ρ(Av (t)). Therefore from (3.1) we conclude that u structurally dominates
v strictly.

In Item (ii) of the previous theorem, we have supplied a condition under which a vertex structurally
dominates another vertex strictly. A natural question is the following. ‘Suppose that G is a strongly
connected graph and u is a vertex. Can there exist a weight matrix A of G for which u is a universal
winner?’ The following result helps us to answer that.
Let G be a digraph and u be a vertex of G. By G − u, we denote the digraph obtained by deleting the
vertex u and all edges incident on it, from G. Fix X ∈ WG . Then for a sub-digraph H of G, by X(H)
we mean the square matrix indexed by the vertices of H (in the same order as in X), whose (u, v)-th
entry is the (u, v)-th entry of X if the edge (u, v) ∈ H. Thus X(G) = X.
Denote the n × n zero matrix by 0n . We use the notation [1, 2, . . . , n, 1] to represent a directed cycle
with vertices {1, 2, . . . , n} and edges {(i, i + 1) | i = 1, . . . , n − 1} ∪ {(n, 1)}.

Theorem 3.2. Let G be a strongly connected digraph and C be a cycle in G. Then there exists a weight
matrix A of G for which every vertex on C dominates every vertex outside C, initially. That is, the set
of initial winners of G for the weight matrix A is a subset of V (C).

Proof. Let us assume that V (G) = {1, 2, . . . , n} and C = [1, . . . , k, 1]. Let  > 0 be a small number.
Assign a weight  to all edges of G except those on the cycle, to which we assign weights 1. Let A be
the adjacency matrix of this weighted digraph G. Observe that as  approaches zero, A converges to
A(C) ⊕ 0n−k . Since spectral radius of a nonnegative matrix is a continuous function of its entries, ρ(A )
approaches to ρ(A(C)) = 1 as  → 0+ . Let v be a vertex of G such that v ∈
/ V (C). As φ(A (v | v), λ) is
a continuous function of , it follows that lim φ(A (v | v), λ) = λn−k−1 φ(A(C), λ). Therefore
→0

lim φ(A (v | v), ρ(A )) = ρ(A(C))n−k−1 φ(A(C), ρ(A(C))) = 0.

→0

Hence, ∃δv > 0 such that φ(A (v | v), ρ(A )) < 0.5 for all  < δv . One can continue the process for every
vertex v ∈ V (G) − V (C). Let δ0 = min{δv | v ∈ V (G) − V (C)}. Then we have

φ(A (v | v), ρ(A )) < 0.5 (∀ v ∈

/ V (C), ∀  < δ0 ). (3.2)

10
Let u be a vertex of C. Now observe that A (u | u) converges to A(C − u) ⊕ 0n−k−1 as  approaches zero.
Notice that ρ(A(C −u)⊕0n−k−1 ) = ρ(A(C −u)) = 0. Therefore we have φ(A(C −u)⊕0n−k−1 , λ) = λn−1 .
Since ρ(A ) approaches to ρ(A(C)) as  approaches to zero, it follows that lim φ(A (u | u), ρ(A )) =
→0
ρ(A(C))n−1 = 1. Hence ∃δ > 0, δ < δ0 , such that for all  < δ, we have φ(A (u | u), ρ(A )) > 0.5. It
follows that
φ(A (u | u), ρ(A )) > φ(A (v | v), ρ(A )) (∀ v ∈
/ V (C), ∀  < δ).

By using Lemma 2.1, it now follows that vertex u dominates every vertex v outside C initially. Hence
for the weight matrices A ( < δ) the set of initial winners is a subset of V (C).

Below we supply an alternative proof of Theorem 3.2.

Another proof of Theorem 3.2. Let us assume that V (G) = {1, 2, . . . , n} and C = [1, . . . , k, 1]. Let
 > 0 be a small number. Assign a weight  to all edges of G except those on the cycle, to which we
assign weights 1, and denote the adjacency matrix of this weighted digraph G by A . Consider the
matrix A(C) ⊕ 0n−k , which has spectral radius 1 as an algebraically simple eigenvalue. Observe that the
vector z with entries √1 on the vertices of C and 0 for vertices not on C is a vector of unit norm that
k
serves as both a right and left eigenvector for A(C) ⊕ 0n−k corresponding to the eigenvalue 1.
For each  > 0, let x() denote the right Perron vector of A having unit norm, and let y()T denote the
left Perron vector of A normalized so that y()T x() = 1. Since 1 is a simple eigenvalue of A(C) ⊕ 0n−k
we find that as  → 0+ , ρ(A(C) ⊕ 0n−k ) → 1, x() → z, and y() → z. We deduce that for all sufficiently
small  > 0, and vertices u, v with u ∈ V (C), v ∈ V (G) − V (C), we have x()u y()u > x()v y()v . The
conclusion now follows from Theorem 1.5.

As an application, we give an example of a digraph G that has some universal winners for one weight
matrix but does not have any universal winners for another weight matrix.

Example 3.3. Let G be the digraph in Figure 2. Let A ∈ WG . Since every nontrivial closed walk
passing through vertex 1 also passes through vertex 2, using Theorem 1.4 we see that S2 = {1, 2}
and S3 = {1}. Therefore for each A ∈ WG vertex 2 is the only terminal winner. Notice that every
(3) (3)
nontrivial closed walk that contains vertex 6 also contains vertex 5 and a55 > a66 . Thus by Item (ii)
of Theorem 3.1, vertex 5 structurally dominates vertex 6 strictly. Similarly, one can argue that vertex
5 also structurally dominates the vertices 7, 8 strictly and vertex 2 structurally dominates the vertices
1, 3, 4, strictly.
Consider the cycle C = [5, 6, 7, 5]. Then by Theorem 3.2, there exists a weight matrix B ∈ WG for which
the initial winners are a subset of V (C). As vertex 5 dominates the vertices 6, 7 for all A ∈ WG and for all
t > 0, it now follows that for B, vertex 5 is the initial winner. Note that we have already established that
for every A ∈ WG vertex 2 is the only terminal winner. So for the weight matrix B, there is no universal
winner. We now show that for the adjacency matrix A(G), vertex 2 is the only universal winner. Put

11
 3 4 8 7

1 2
5 6
Figure 2: A digraph G with A ∈ WG such that for A there is no universal winner.

A = A(G). It remains to show that ρ(A + tE22 ) = ρ(A2 (t)) ≥ ρ(A5 (t)) = ρ(A + tE55 ). Suppose that
this is not the case. Let there exist a t0 > 0 such that ρ(A2 (t0 )) < ρ(A5 (t0 )). Since vertex 2 is the
terminal winner and ρ(A2 (t0 )) < ρ(A5 (t0 )), there exists an α > 0 such that ρ(A2 (α)) = ρ(A5 (α)). Then
by Theorem 1.7, ρ(A2 (α)) is a root of the equation φ(A(2 | 2), λ) − φ(A(5 | 5), λ) = 0. But we have,

φ(A(2 | 2), λ) − φ(A(5 | 5), λ) = λ5 > 0 for all λ > 0.

This is a contradiction. Hence, vertex 2 is the only universal winner for A.

As another application of Theorem 3.2 we have the following result that characterizes the class of
strongly connected digraphs G that do not have a vertex that structurally dominates every other vertex
of G.

Theorem 3.4. Let G be a strongly connected digraph. Then G does not have a vertex that structurally
dominates the remaining vertices if and only if for every vertex u of G there exists a nontrivial closed
walk that does not pass through u.

Proof. First, suppose that for every vertex u of G, there exists a nontrivial closed walk that does not
pass through u. If possible, assume that v is a vertex of G that structurally dominates all other vertices.
Now, from the hypothesis, we know that there exists a nontrivial closed walk missing v. Then there
exists a cycle C that does not contain v. Using Theorem 3.2, we can find a weight matrix A for which
the vertices on C dominate the vertex v initially. This is a contradiction. Therefore, G does not have a
vertex that structurally dominates the remaining vertices.
For the converse part, suppose that the conclusion does not hold. In other words, there exists a
vertex v such that every nontrivial closed walk passes through v. Let u ∈ V (G) with u 6= v and A ∈ WG .
(k) (k)
Since every nontrivial closed walk contains v, we have avv ≥ auu , for all k ∈ N. Then, by Lemma 2.4,
ρ(Av (t)) ≥ ρ(Au (t)) for all t > 0. As A and u were arbitrary, we see that v structurally dominates all
other vertices. This contradicts the fact that G does not have a vertex that structurally dominates all
other vertices.

Recall that for a strongly connected digraph, Item (ii) of Theorem 3.1 provides a combinatorial
sufficient condition for strict structural dominance. As another application of Theorem 3.2, we now
show that the same condition is also necessary.

12
Corollary 3.5. Let u and v be two vertices of a strongly connected digraph G. Then the following
statements are equivalent.

(i) The vertex u structurally dominates the vertex v strictly.

(ii) Every nontrivial closed walk passing through v also passes through u, and there exists a nontrivial
closed walk passing through u that does not pass through v.

(iii) For every weight matrix A of G, with positive Perron vectors x and y, the inequality xu yu ≥ xv yv
holds and there exists a weight matrix A of G with positive Perron vectors x, y, such that xu yu >
xv yv .

Proof. In Item (ii) of Theorem 3.1 we have already shown that (ii) implies (i).
Proof of (i) implies (ii). Suppose that (i) holds. We first show that every nontrivial closed walk
passing through v also passes through u. On the contrary, suppose that there exist a nontrivial closed
walk passing through v that does not pass through u. Then there exists a cycle that contains v but does
not contain u. Using Theorem 3.2 we can find a matrix A of G for which vertex v dominates vertex u
initially. This contradicts (i).
So every nontrivial closed walk passing through v also passes through u. Now, if every nontrivial
closed walk passing through u also passes through v, then by Item (i) of Theorem 3.1, u and v will
structurally tie. This again contradicts (i).
Proof of (iii) implies (ii). We first show that every nontrivial closed walk passing through v also
passes through u. If possible, assume that there exists a nontrivial closed walk passing through v that
does not pass through u. Then there exists a cycle that contains v but misses u. In view of the alternate
proof for Theorem 3.2, we can find a weight matrix A, with Perron vectors x, y, such that xv yv > xu yu .
However, this contradicts condition (iii).
Thus, every nontrivial closed walk passing through v also passes through u. Now, if every nontrivial
closed walk passing through u also passes through v, then by Item (i) of Theorem 3.1, u and v will
structurally tie. Now, from the hypothesis, we have a matrix A ∈ WG with right and left Perron vectors
x and y T , respectively, such that xu yu > xv yv . Using Theorem 1.5, we see that for this weight matrix
A, vertex u dominates vertex v initially. This is a contradiction.
Proof of (ii) implies (iii). Suppose that (ii) is satisfied indicating that u structurally dominates v
strictly. From Theorem 1.5, it follows that for every weight matrix A of G, with right and left Perron
vectors x and y T , respectively, we have xu yu ≥ xv yv . Since, there exists a nontrivial closed walk passing
through u that does not pass through v, we can find a weight matrix A such that xu yu > xv yv .

As another application of Theorem 3.2, we have the following result about the set of vertices that
are universal winners for every weight matrix. Let G be a digraph and u be a vertex of G. We say that
the vertex u is a cut vertex of G if it is a cut vertex of the underlying undirected graph of G.

13
Corollary 3.6. Let G be a strongly connected digraph. Suppose that G has two cut vertices, u and v.
Let S be the set of universal winners for every weight matrix A of G. Then S = ∅.

Proof. Note that there exists a cycle Cu containing u that does not contain v. Let Gv be the component
of G − u containing the vertex v. As v was also a cut vertex of G, there is a cycle Cv containing v
inside Gv . So the cycles Cu and Cv are disjoint. Therefore by Theorem 3.2 there exist weight matrices
A and B for which the set of initial winners of G are subsets of V (Cu ) and V (Cv ), respectively. Hence
we cannot have a set of universal winners common for all weight matrices.

Recall that for the star graph the central vertex is the universal winner for every weight matrix.
This means that no switch happens between a pendent vertex and the center. In [1], it is shown that
for a connected undirected graph G with at least 3 vertices, a pendent vertex v is never an initial or a
terminal winner for the adjacency matrix A(G).
As an application of Corollary 3.5, we supply a more general result for a strongly connected digraph.

Theorem 3.7. Let G be a strongly connected digraph on n ≥ 3 vertices. Suppose that v is a vertex with
a unique neighbor u (i.e., u is the only in-neighbor and out-neighbor). Then u structurally dominates v
strictly.

Proof. First observe that since u is the only neighbor of v, every nontrivial closed walk from v to v
passes through u. As G is strongly connected and has at least 3 vertices, u has a neighbor other than v.
So there exist a nontrivial closed walk from u to u that does not pass through v. Therefore by Corollary
3.5, we conclude that u structurally dominates v strictly.

As a further application of Corollary 3.5, we now prove a result for a connected undirected graph
(with at least 3 vertices) stating that for any weight matrix, a pendent vertex v is neither an initial nor
a terminal winner.

Theorem 3.8. Let G be a connected graph on n ≥ 3 vertices. Suppose that v is a pendent vertex. Then
for each weight matrix A of G, the vertex v is never an initial or a terminal winner.

Proof. It follows directly from Theorem 3.7 by replacing each undirected edge with a pair of oppositely
directed edges of the same weight.
We now end this section with an alternate proof of Theorem 3.7. This alternate proof is crucial as
it exhibits a new technique. Prior to this we supply a result about the characteristic polynomial that
is needed for the alternate proof. By a connected digraph we mean the underlying undirected graph is
connected.

Lemma 3.9. Let G be a connected digraph on vertices 1, 2, . . . , n. Let H be a spanning sub-digraph of

G. Let X = X(G) ∈ WG be fixed. Then φ(X(H), λ) ≥ φ(X(G), λ) for all λ ≥ ρ(X(G)).

14
Proof. We will prove the result by applying induction on the number of vertices of G. For n = 2, it can
be easily verified. Assume that the statement holds true for all connected digraphs on k vertices. Let
G be a connected digraph on k + 1 vertices and X = X(G) ∈ WG . Then the derivative of φ(X(G), λ) is
k+1
φ0 (X(G), λ) = tr(adj(λIk+1 − A)) =
P
φ(X(G − i), λ). Let H be a spanning sub-digraph of G. Then,
i=1

k+1
X
0 0
 
φ (X(H), λ) − φ (X(G), λ) = φ(X(H − i), λ) − φ(X(G − i), λ) .
i=1

Since H is a spanning sub-digraph of G, H − i is also a spanning sub-digraph of G − i. So by the

induction hypothesis φ(X(H − i), λ) − φ(X(G − i), λ) ≥ 0, ∀λ ≥ ρ(X(G − i). Using the above equation,
we have
φ0 (X(H), λ) − φ0 (X(G), λ) ≥ 0, ∀λ ≥ max ρ(X(G − i)).
1≤i≤k+1

Therefore, φ(X(H), λ) − φ(X(G), λ) is monotonically increasing in [ρ(X(G)), ∞). Note that ρ(X(G)) ≥
ρ(X(G − i)), as G − i is a sub-digraph of G. So at λ = ρ(X(G)), we have φ(X(H), λ) − φ(X(G), λ) ≥ 0.
Hence φ(X(H), λ) − φ(X(G), λ) ≥ 0 in [ρ(X(G)), ∞).

Now we are ready to give the alternative proof of Theorem 3.7.

Another proof of Theorem 3.7. Fix A ∈ WG . Observe that the matrix A(u | u) (respectively,
A(v | v)) represents the graph G − u (respectively, G − v). Since G is strongly connected and v has only
one neighbor u, it follows that G − v is also strongly connected. Thus A(v | v) is irreducible. Note that
" #
A(G − u − v) 0
A(G − u) = , where the last row and column correspond to the vertex v. Hence
0 0
ρ(A(u | u)) = ρ(A(G − u − v)) < ρ(A(v | v)). Consider the digraph H obtained by adding an isolated
vertex u in G − u − v. Then H is a proper spanning sub-digraph of G − v. Note that A(H) = A(u | u).
Now using Lemma 3.9, we have that

φ(A(u | u), λ) ≥ φ(A(v | v), λ) for all λ ≥ ρ(A(v | v)).

In particular, this is true for all λ ≥ ρ(A).

Now we show that ρ(Au (t)) ≥ ρ(Av (t)). Suppose that this is not the case. Let there exist a t0 > 0
such that ρ(Au (t0 )) < ρ(Av (t0 )). Put λ0 = ρ(Av (t0 )). Hence φ(Au (t), λ0 ) > 0. Note that φ(Au (t), λ) =
1
φ(A + tEuu , λ) = φ(A, λ) − tφ(A(u | u), λ). Therefore we have t0 φ(A, λ0 ) > φ(A(u | u), λ0 ). Again, as
1
φ(Av (t0 ), λ0 ) = 0, we see that φ(A(v | v), λ0 ) = t0 φ(A, λ0 ). Hence, φ(A(v | v), λ0 ) > φ(A(u | u), λ0 ).
This is a contradiction.

We remark here that another interesting combinatorial proof of Theorem 3.7 is given in the Appendix.
It simply relies on the fact that, if X ≥ 0 is an irreducible matrix with a nonzero diagonal entry, then
 1/k
lim X k (i, i) = ρ(X).
k→∞

15
4. Digraphs with a fixed subset of vertices as universal winners for every weight matrix

Let G be a strongly connected digraph. Let S be the set of vertices that are universal winners for
every weight matrix A of G. Recall that from Corollary 3.6, we know that if G has two cut vertices,
then S = ∅. So a natural question is whether we have some strongly connected digraphs such that the
same nonempty subset S ⊆ V (G) becomes the set of universal winners for every weight matrix. This
question, for the case S = V (G) has been answered in [1, Theorem 2.1]. Notice that if G is the directed
cycle on two vertices, then by [1, Theorem 2.1] every vertex of G is a universal winner. Therefore in this
section, we shall consider strongly connected digraphs on n ≥ 3 vertices. Given S = {1, 2, . . . , k}, we
want to characterize all strongly connected digraphs G with S ( V (G) such that S is the set of universal
winners of G for every weight matrix. In order to supply the characterization let us first introduce some
terminology.

Definition 4.1. Let Γ = [u1 , u2 , . . . , un , u1 ] be a directed cycle on n vertices. We say that the vertices
x1 , x2 , . . . xk ∈ V (Γ) are cyclically placed on Γ, if for each i = 1, . . . , k − 1, the directed path from xi to
xi+1 does not contain any other xj , j 6= i, i + 1. (Notice that, in this case, the directed path from xk to
x1 does not contain any of x2 , . . . , xk−1 .)

Definition 4.2. Let G and H be two vertex disjoint strongly connected digraphs. Let x1 , x2 , . . . , xk be
k cyclically placed vertices on some cycle in G and y1 , y2 , . . . , yk be k cyclically placed vertices on some
cycle in H. Let G ⊕ H be the digraph obtained by identifying the vertices xi with yi , i = 1, . . . , k.

Example 4.3. The digraph G in Figure 3 is obtained by identifying the vertex xi on Γ1 with the vertex
yi on Γ2 , i = 1, . . . , 5. We have used pi to label the vertex obtained by identifying xi with yi , i = 1, . . . , 5.

Let S = {1, 2, . . . , k}. We now supply a recursive construction of the family FS of strongly connected
digraphs in which S is the set of universal winners for every weight matrix.
Construction.

A) First we construct a family G in the following way.

Step-1. Take two vertex disjoint directed cycles Γ1 and Γ2 containing k or more vertices. Let
x1 , . . . , xk1 be some cyclically placed vertices on Γ1 and y1 , . . . , yk1 be some cyclically placed
vertices on Γ2 , where k1 ≥ k. Obtain the digraph Γ1 ⊕ Γ2 by identifying xi with yi , i =
1, . . . , k1 . Let us use pi to label the vertex obtained by identifying xi with yi , i = 1 . . . , k1 .
Call p1 , . . . , pk1 the principal vertices of Γ1 ⊕ Γ2 . Put all such digraphs Γ1 ⊕ Γ2 in G. Let
H ∈ G. Notice that the principal vertices of H are cyclically placed on every directed cycle
in H.

Step-2. Take any digraph G ∈ G and let Γ be a directed cycle in G. Let z1 , . . . , zt be t cyclically
placed vertices on Γ such that {z1 , . . . , zt } contains at least k principal vertices of G. Take

16
 a new directed cycle Γ0 that is vertex disjoint from G. Let y1 , . . . , yt be t cyclically placed
vertices on Γ0 . Obtain the digraph G ⊕ Γ0 by identifying zi with yi , i = 1, . . . , t. For the
digraph G ⊕ Γ0 , define the principal vertices to be those principal vertices of G that are
selected in z1 , . . . , zt , written in the cyclical order. Put all these digraphs G ⊕ Γ0 in G.

Step-3. Repeat Step-2 a few more times.

B) Let FS be the set of digraphs in G that have exactly k principal vertices.

We will prove that this is precisely the family we are looking for. Before that let us illustrate the
previous construction with an example.

Example 4.4. In this example, we create a digraph in which three vertices are the universal winners
for all weight matrices.

a) We start with a directed cycle Γ1 on 6 vertices from which we have selected 5 vertices (those are
surrounded by circles) and a directed cycle Γ2 on 5 vertices from which we have selected all the
vertices. Then we obtain G = Γ1 ⊕ Γ2 by identifying the corresponding cyclically placed selected
vertices. The result is shown below. The five principal vertices are shown with a square around
each.

x1 y1 p1

x2 x5 y2 y5 p2 p5
⊕ =
x3 x4 y3 y4 p3 p4

Γ1 Γ2 Figure 3: G = Γ1 ⊕ Γ2 ;

b) Now we take G from the previous picture and select 5 of its vertices from a cycle. Out of these
5 vertices we must select at least 3 principal vertices. Here we have selected 4 principal vertices
(these have both a square and a circle around them). Now we take a new directed cycle Γ of length
6 that is vertex disjoint from G and select 5 vertices. The digraph H = G ⊕ Γ is shown below.
For the new digraph the principal vertices are those principal vertices of the old digraph that are
selected. We continue to use the square to represent them.

x1 y1 p1

x5 y5 p4
⊕ =
x2 x4 y2 y4 p2 p3

x3 y3
G Γ Figure 4: H = G ⊕ Γ

17
 c) Now we take the digraph H from the previous picture and select 4 vertices (from a cycle) including
3 principal ones. Next we take a directed cycle Γ0 of length 5 that is vertex disjoint from H and
select 4 vertices. The digraph Z = H ⊕ Γ0 is shown below. The new principal vertices are marked
with a square.

x1 y1 p1

x2 y2 y4
⊕ =
x3 x4 y3
p2 p3

H Γ0 Figure 5: Z = H ⊕ Γ0

d) The digraph Z is a digraph in which has the same set of three universal winners for every weight
matrix. In fact this set is precisely the set of principal vertices of Z.

We are now ready to supply the complete characterization of strongly connected digraphs in which
the same subset of vertices remains as the set of universal winners for all weight matrices.

Theorem 4.5. Let G be a strongly connected digraph and S ( V (G) be nonempty. Then the following
statements are equivalent.

(i) For every weight matrix A of G, the digraph G has S as the set of universal winners.

(ii) Each nontrivial closed walk in G passes through every vertex of S. For any other vertex v ∈
/ S,
there exists a nontrivial closed walk that does not pass through v (which pass through all vertices
of S).

(iii) There is a digraph Ĝ in FS that is isomorphic to G for which the set of principal vertices is the
isomorphic image of S.

Proof. Proof of (i) implies (ii). Let C be a nontrivial closed walk in G and u ∈ S. If possible, assume
that C does not contain any vertex of S. Let v be a vertex on C. So v ∈
/ S. Since u is a universal winner
for each weight matrix A of G and v is not, by definition, u structurally dominates vertex v strictly.
Then by Corollary 3.5, every nontrivial closed walk passing through v should also pass through u. In
particular C also passes through u. This is a contradiction. So C must pass through some vertices of
S. As the vertices in S are the universal winners for all weight matrices, these vertices structurally tie
with each other. From Item (i) of Theorem 3.1, we see that C contains every vertex of S.
Next, let u ∈ S and v ∈
/ S. Then u structurally dominates vertex v strictly. So by Corollary 3.5,
there exist a nontrivial closed walk passing through u that does not pass through v. This completes our
first assertion.

18
 Proof of (ii) implies (iii). Suppose that (ii) is satisfied. We shall create a digraph Ĝ in FS that is
isomorphic to G. The construction of the digraph Ĝ proceeds as follows.

Step-1. As G is strongly connected, it contains a cycle, say, C. By Item b), C contains every vertex
of S. If C is the only cycle in G, then G = C. Then by [1, Theorem 2.1] every vertex in G is a
universal winner, that is, S = V (G). This is a contradiction to our hypothesis that S ( V (G).

So G contains at least two cycles, say, C1 and C2 , each containing all the vertices of S. Let
{z1 , . . . , zk1 } be the set of common vertices of C1 and C2 and assume that z1 , . . . , zk1 are cyclically
placed on C1 (hence on C2 ). Let Γ1 be a copy of C1 and Γ2 be a copy of C2 , on disjoint set
of vertices. Let x1 , . . . , xk1 be the vertices on Γ1 mirroring z1 , . . . , zk1 on C1 and y1 , . . . , yk1 be
the vertices on Γ2 mirroring z1 , . . . , zk1 on C2 . Define H2 to be the digraph Γ1 ⊕ Γ2 obtained by
identifying xi with yi (to form the principal vertex pi ), i = 1, . . . , k1 . Let P2 := {p1 , p2 , . . . , pk1 }
be the set of principal vertices of H2 . By our construction H2 is isomorphic to the sub-digraph
G2 := C1 ∪ C2 of G. Let f be an isomorphism from G2 to H2 . Since S ⊆ {z1 , . . . , zk1 }, we see
that f (S) = {f (s) | s ∈ S} ⊆ P2 = f ({z1 , . . . , zk1 }).

If G = G2 , then every cycle in G contains the common vertices z1 , . . . , zk1 , implying that S =
{z1 , . . . , zk1 }. In this case, H2 is the required digraph. The proof is complete for this case.

Step-2. Assume that we have obtained Hm and Gm for some m ≥ 2, f from Gm to Hm is an

isomorphism, Pm be the set of principal vertices of Hm , f (S) ⊆ Pm , and that each cycle in Hm
contains Pm .

Suppose that G 6= Gm . So there exists a cycle C 0 in G that is not present in Gm . Now take any
cycle C in Gm . Let w1 , . . . , wt be the common vertices of C and C 0 . Note that S ⊆ {w1 , . . . , wt }.
Assume that w1 , . . . , wt are cyclically placed on C (and, consequently on C 0 ). Let Γ = f (C),
the image of C under f . Take a copy Γ0 of the cycle C 0 such that V (Hm ) ∩ V (Γ0 ) = ∅. Let
x1 , . . . , xt be the vertices on Γ corresponding to w1 , . . . , wt on C and y1 , . . . , yt be the vertices on
Γ0 corresponding to w1 , . . . , wt on C 0 . Put Hm+1 := Hm ⊕ Γ0 , the digraph obtained by identifying
xi with yi , i = 1, . . . , t. Call the set Pm+1 := Pm ∩ f ({w1 , . . . , wt }) the set of principal vertices of
Hm+1 . From the construction we see that Hm+1 is isomorphic to the sub-digraph Gm+1 := Gm ∪C 0
of G. Let f be an isomorphism from Gm+1 to Hm+1 . Since f (S) ⊆ Pm and f (S) ⊆ f ({w1 , . . . , wt }),
we see that f (S) ⊆ Pm+1 . Also, it is not difficult to argue that as each each cycle in Hm+1 contains
Pm+1 .

Step-3. We repeat Steps 2 until G = Gm0 for some m0 . Then put Ĝ = Hm0 . By our construction, it is
isomorphic to G. Let f be an isomorphism between Gm0 and Hm0 .

We claim that Pm0 = f (S). If possible assume that v0 ∈ Pm0 \ f (S). Let u0 ∈ V (Gm0 ) with
f (u0 ) = v0 . Then by Item b), there is a directed cycle C in G = Gm0 that does not contain u0 .

19
 Then the cycle Γ = f (C) in Hm0 will not contain v0 . As v0 ∈ Pm0 , this contradicts the fact that
every cycle in Hm contains all vertices of Pm , for all m ≥ 2. Hence our claim is justified.

So we have constructed a digraph Ĝ ∈ FS isomorphic to G for which the set of principal vertices
is the isomorphic image of S.

Proof of (iii) implies (i). Let Ĝ be a strongly connected digraph in FS isomorphic to G for which
the set of principal vertices is the isomorphic image of S. Note that, by construction, the principal
vertices of Ĝ are precisely those vertices that are contained in every cycle. Then due to isomorphism,
the elements of S are precisely those vertices that are contained in every cycle in G. So by Corollary
3.5, S is the set of universal winners in G for every weight matrix.

Consider the digraph G given below. For this digraph the set {1, 2, 3, 4} is the set of universal
winners for every weight matrix. But something more happens here. Note that by definition, the
universal winners for every weight matrix structurally tie with each other. What about the remaining
vertices? In general, we may not have the remaining vertices structurally tying with each other. But
in this graph they do. In other words, the vertex set of this graph can be partitioned in two subsets
of structurally tying vertices such that vertices from one subset structurally dominates from the other
strictly.

7 6 5

1 4

2 3

So let us put the following question.

Given a natural number k, is it possible to find a strongly connected digraph G
such that V (G) can be partitioned into k subsets P1 , . . . , Pk such that we have the
following.

i) Vertices inside Pi structurally tie with each other for i = 1, . . . , k.

ii) Vertices of Pi structurally dominates vertices of Pj strictly for i < j.

The answer is in the affirmative. Let G be such a strongly connected digraph. Then we call the sets
P1 , . . . , Pk as layers of G. Thus the vertices of P1 are the universal winners of G, the vertices of P2 can
be seen as the next set of winners and so on. Let us rephrase the previous question a little differently.
Given any partition P = {P1 , P2 , . . . , Pk } of {1, 2, . . . , n}, does there exist a strongly
connected digraph G with V (G) = {1, 2, . . . , n} whose layers are given by P. Can we
characterize them?
The following lemma answers both of these questions in the affirmative.

20
Lemma 4.6. Let P = {P1 = {1, . . . , l1 }, P2 = {l1 + 1, . . . , l2 }, . . . , Pk = {lk−1 + 1, . . . , n}} be a partition
of {1, 2, . . . , n}. Consider the digraph G obtained from the directed cycle Γ = [1, 2, . . . , n, 1], by adding
the directed edges (l1 , 1), (l2 , 1), . . . , (lk−1 , 1). Then P1 , . . . , Pk are the layers of G.

1 n

2 n−1
.. ..
. .
l1 − 1 lk−1 + 1

l1 ...... ......
l1 + 1 l2 l2 + 1 lk−1

Proof. Fix i and j such that 1 ≤ i < j ≤ k. Notice that if C is a nontrivial closed walk that passes
through one vertex of Pi , then C passes through every vertex of Pi . Hence by Item (i) of Theorem 3.1
all the vertices of Pi structurally tie with each other.
Now, let u ∈ Pi and v ∈ Pj . Observe that every nontrivial closed walk passing through the vertex v
passes through the vertex u and there is a cycle C = [1, 2, . . . , li , 1] that passes through u but not through
v. So by Item (ii) of Theorem 3.1, vertex u structurally dominates the vertex v strictly. Therefore vertices
of Pi structurally dominates vertices of Pj strictly.

Thus we have shown that there are digraphs G such that the vertex set can be partitioned into k
subsets P1 , . . . , Pk such that the vertices of Pi structurally tie with each other and every vertex of Pi
structurally dominates every vertex of Pj strictly, for 1 ≤ i < j ≤ k. In the following theorem we
characterize all such digraphs.

Theorem 4.7. Let G be a strongly connected digraph on vertices 1, 2, . . . , n. Then G has k > 1 layers
if and only if G can be seen as a digraph constructed as described in Lemma 4.6.

Proof. Assume that G has k layers P1 , . . . , Pk . Fix v ∈ Pk . As G is strongly connected, it contains

a cycle Γv passing through v. Since the vertices of Pk structurally tie with each other, by Item (i) of
Theorem 3.1, Γv contains every vertex of Pk . Select any Pi , i 6= k and any u ∈ Pi . Since u structurally
dominates v strictly, by Item (ii) of Theorem 3.1, Γv also passes through u. As Pi and u were arbitrary,
we see that Γv contains every vertex of G.
Let u ∈ P1 and w ∈ P2 . Since u structurally dominates w strictly, by Corollary 3.5, there is a cycle
Γ1 containing u that does not contain w. As the vertices of P2 structurally tie with each other and the
vertices of P1 structurally tie with each other, by Item (i) of Theorem 3.1, Γ1 does not contain any vertex
of P2 and Γ1 contains every vertex of P1 . Similarly one can argue that there is a cycle, say Γ2 , that
does not contain any vertex of P3 but contains every vertex of P2 . Since the vertices of P1 structurally
dominate those of P2 strictly, by Corollary 3.5, Γ2 also contains every vertex of P1 . Using a similar

21
argument, we can conclude that there are cycles Γi (2 < i ≤ k − 1) that does not contain any vertex of
Pi+1 but contains every vertex of Pj for all j ≤ i.
We claim that V (Γ1 ) = P1 . Suppose that Γ1 contains a vertex v0 from Pj for some j > 2. Let
u0 ∈ P2 . Note that u0 structurally dominates v0 strictly. So every cycle that contains v0 must contain
u0 , by Item (ii) of Theorem 3.1. But we already know that Γ1 does not contain any vertex of P2 .
Therefore, V (Γ1 ) = P1 .
i
S
Similarly one can argue that V (Γi ) = Pj .
j=1
Assume that Γ1 = [i1 , i2 , . . . , im , i1 ]. Note that in G the edges (i1 , i3 ), . . . , (i1 , im ) are absent,
otherwise we would have a cycle that does not contain all of P1 . With a similar argument we see
that P1 induces the cycle Γ1 . Also observe that, the vertices i1 , i2 , . . . , im are cyclically present on
Γv . Assuming otherwise there exist a smallest integer l > 2 such that Γv [i1 , . . . , il ] contains il0 with
l0 > l. Here Γv [i1 , . . . , il ] refers to the directed path from i1 to il that is on Γv . Then the cycle
C := Γv [i1 , i2 , . . . , il−1 ] ∪ (il−1 , il ) ∪ Γv [il , . . . , i1 ] does not contain il0 . Since the vertices of P1 struc-
turally tie with each other, by Item (i) of Theorem 3.1 we already know that C contains every vertex of
P1 . This is a contradiction. So, the vertices i1 , i2 , . . . , im are cyclically present on Γv . Note that all the
edges of Γ1 cannot be present on Γv , otherwise Γ1 = Γv , which is not possible.
Claim : All edges of Γ1 except exactly one are present on Γv . Suppose that two edges (i1 , i2 ) and
(il , il+1 ) (l > 1) of Γ1 are not on Γv . (In the case that l = m, assume il+1 to be i1 .) As i1 is on Γv , let
x1 be the vertex next to i1 on Γv , that is, (i1 , x1 ) is on Γv . Note that x1 ∈
/ V (Γ1 ) as (i1 , i3 ), . . . , (i1 , im )
are absent in G. Similarly, let x2 , x3 , x4 ∈
/ V (Γ1 ) such that the directed edges (x2 , i2 ), (il , x3 ), (x4 , il+1 )
are in Γv . Here we allow that x1 = x2 and x3 = x4 .
Notice that the cycle Γv [il , il+1 ] ∪ Γ1 [il+1 , . . . , il ] does not contain x1 . Similarly the cycle Γv [i1 , i2 ] ∪
Γ1 [i2 , . . . , i1 ] does not contain x3 . So, by Theorem 3.2, there exists a matrix A ∈ WG for which x3
dominates x1 initially and there exists a matrix B for which x1 dominates x3 initially. Hence x1 and
x3 cannot be in the same layer. They cannot be in two different layers, as a vertex from a layer Pj
structurally dominates a vertex from Pi strictly if j > i. This is a contradiction. So our claim is
justified.
Therefore Γ1 has only one edge outside Γv . Cyclically renaming the vertices i1 , . . . , im of Γ1 we
assume that the edge (im , i1 ) is not present on Γv . Now we claim that Γ1 [i1 , . . . , im ] is present in
Γ2 . Suppose that this is not the case. Then there exist vertices ij with j 6= m and w such that the
directed edge (ij , w) is present in Γ2 . Then the cycle (ij , w) ∪ Γv [w, . . . , ij ] does not contain ij+1 . This
is a contradiction to the fact that every cycle that contains a vertex of P1 contains all of P1 . Now
assume that Γ2 = [i1 , . . . , im , j1 , . . . , jq ]. By an argument similar to that used to show that i1 , . . . , im
are cyclically placed on Γv , we can conclude that the vertices j1 , . . . , jq are also cyclically placed on Γv .
By an argument similar to that used to show Γ1 has exactly one edge outside Γv , we can conclude that
Γ2 also has exactly one edge outside Γv . Similarly, we see that each Γi contains only one edge outside

22
Γv , i = 1, . . . , k − 1.
Now it remains to show that G does not contain any cycle other than Γ1 , . . . , Γk−1 , Γk := Γv . Let
0
C be any cycle in G. Let j be the maximum index such that Pj contains a vertex of C. Let v ∈ Pj be
on C. As the vertices inside Pj structurally tie with each other, we see that all vertices of Pj lie on C.
As v is structurally dominated by every vertex of Pl strictly for l < j, we see that C must contain every
vertex of Pl . Hence C = Γj . Therefore, G does not contain any other cycle. This completes the proof
for this case.
Note that we have already shown the converse part in Lemma 4.6

Remark 4.8. Recall that Theorem 4.5 characterizes all strongly connected digraphs in which the same
subset of vertices remains as the set of universal winners for all weight matrices. The following question
arises naturally.
Let G be a strongly connected digraph. Suppose we have a set S ( V (G) such
that the vertices of S are the only initial winners for every A ∈ WG . Can we
characterize all such digraphs? The answer is in the positive as we shall show. In fact the
characterization is same that of the universal winners.

Theorem 4.9. Let G be a strongly connected digraph and S ( V (G) be nonempty. Then the following
statements are equivalent.

(i) For every weight matrix A of G, the digraph G has S as the set of initial winners.

(ii) Each nontrivial closed walk in G passes through every vertex of S. For any other vertex v ∈
/ S,
there exists a nontrivial closed walk that does not pass through v (which pass through all vertices
of S).

(iii) There is a digraph Ĝ in FS that is isomorphic to G for which the set of principal vertices is the
isomorphic image of S.

Proof. Proof of (i) implies (ii). Let C be a nontrivial closed walk in G and u ∈ S. If possible, assume
that C does not contain any vertex of S. Let v be a vertex on C. So v ∈
/ S. Notice that there exists
a cycle Γ in C containing v. Then by Theorem 3.2, we can find a matrix A for which the set of initial
winners is a subset of V (Γ) ⊆ V (C). This is a contradiction. Therefore, C must pass through some
vertices of S. Observe that if two vertices are initial winners for a weight matrix A of G, then these
vertices tie for A. As the vertices of S are the initial winners for every weight matrices, by definition,
these vertices structurally tie with each other. By Item (i) of Theorem 3.1, C contains every vertex of
S.
Next, let u ∈ S and v ∈
/ S. Suppose that the conclusion does not hold. That is every nontrivial
closed walk passing through u also passes through v. Since every nontrivial closed walk in G contains

23
every vertex of S, by Item (ii) of Theorem 3.1, vertices u and v tie. Therefore, vertex v is also an initial
winner. This is a contradiction to the fact that v ∈
/ S. This completes the proof for this case.
Proof of (ii) implies (iii). The proof is identical to the Proof of (ii) implies (iii) in Theorem 4.5.
Proof of (iii) implies (i). Let Ĝ be a strongly connected digraph in FS that is isomorphic to G
with the property that the set of principal vertices in Ĝ corresponds to the isomorphic image of S. By
construction, the principal vertices of Ĝ are precisely those vertices that are present in every cycle. Then
due to isomorphism, the elements of S correspond to the vertices contained in every cycle of G. By Item
(i) of Theorem 3.1, the vertices in S structurally tie with each other. By Item (ii) of same theorem the
vertices of S structurally dominates every other vertex strictly. Therefore the vertices of S are precisely
the initial winners for every weight matrix A of G.

We end this section with a result that answers a similar question for undirected graphs. Recall that
if we take an undirected graph G, then WG contains the symmetric nonnegative matrices A with aij > 0
if and only if [i, j] ∈ G.
Can we characterize all connected (undirected) graphs such that for each matrix
A ∈ WG , we have the same subset of vertices as universal winners?
The following result answers the above question by showing that the stars are the only such graphs.

Theorem 4.10. Let G be a connected graph on n ≥ 3 vertices and S ( V (G) be nonempty. Suppose
that G has S as the set of universal winners for all weight matrix A ∈ WG . Then G must be a star.

Proof. Suppose that the conclusion does not hold. Let G be a connected graph that is not a star.
Suppose that G has S as set of universal winners for every A ∈ WG . Take u ∈ S. Since G is not a
star, there exists an edge [v, w] with v, w 6= u. Suppose that u is adjacent to l vertices. Assign weight
l + 1 to the edge [v, w], while all other edges in G bear the weight 1. Let A be the adjacency matrix
(2) (2)
of this weighted graph G. Consequently, auu = l < l + 1 ≤ avv . Using Theorem 1.4, we see that u is
not a terminal winner. This contradicts the assumption that u is a universal winner for every A ∈ WG .
Hence, G must be a star.

5. Appendix

Here we provide an alternative proof of Theorem 3.7

Another proof of Theorem 3.7. Fix A ∈ WG . We now show that ρ(A + tEuu ) = ρ(Au (t)) ≥
ρ(Av (t)) = ρ(A + tEvv ). Let us use Aku to denote the matrix (A + tEuu )k . So Akv (i, j) denotes the
(i, j)-th element of Akv . We claim that Auk+2 (v, v) ≥ Akv (v, v). Note that the matrix Av represents the
digraph Gv that is obtained from G assigning edge weights according to A and adding a loop of weight
t at the vertex v. Similarly, the digraph Gu is defined using Au .

24
 Note also that Akv (i, j) is the sum of the weights of all directed walks of length k from i to j, where
the weight of a walk is the product of the weights of the edges present in the walk.
Let P = [v = v1 , v2 , . . . , vk+1 = v] a walk of length k in Gv . By a maximal stretch of consecutive v’s in
P , we mean a sequence vl = vl+1 = · · · = vl+m = v, where either l = 1 or vl−1 = u and either l+m = k+1
or vl+m+1 = u, l, m ∈ N. Corresponding to P , we generate a walk P 0 = [v = v10 , v20 , . . . , vk+3
0
= v] in Gu ,
in the following way.

a) Find all the maximal stretches of v’s in P . Replace each of such stretch vl = vl+1 = · · · = vl+m = v
with vl0 = v, vl+1
0 0
= · · · = vl+m = u.

0
b) Assign vk+2 = u.

c) For other vertices vi in P we keep vi0 = vi .

For example, assuming 1, 2, 3, u, v are distinct vertices and the necessary edges in G, the walk P shown
below will correspond to the walk P 0 .

P = [v, v, u, 1, 2, 3, u, v, v, u, 1, 2, 3, u, v, v]

P0 = [v, u, u, 1, 2, 3, u, v, u, u, 1, 2, 3, u, v, u, u, v].

Note that P 0 does not have two consecutive v’s, by construction. So, one can obtain P from P 0 simply
replacing the maximal stretches of the form v, u, u, . . . , u, u with v, v, v, . . . , v, u and then by deleting the
last two entries. Hence,

i) the number of loops in P and P 0 are the same,

ii) P 0 has one more [u, v] and [v, u] edge than of P and

iii) the remaining edges in P and P 0 are the same.

Then the weight of the walk P 0 equals the product of the weight of P with auv avu . If auv avu ≥ 1, our
claim is justified. Otherwise, we can consider a positive scalar multiple of A. Thus our claim is justified.
It is known that (see for example [4]), if X ≥ 0 is an irreducible matrix with a nonzero diagonal entry,
 1/k
then lim X k (i, i) = ρ(X). Applying this to Av and Au , we get that
k→∞

ρ(Au (t)) = lim (Ak+2

u (v, v))
1/k+2
= lim (Ak+2
u (v, v))
1/k
≥ lim (Akv (v, v))1/k = ρ(Av (t)).
k→∞ k→∞ k→∞

Observe that any walk of length k from v to v in G has the form [v = v1 , v2 = u, v3 , . . . , vk = u, vk+1 = v].
Then [v2 = u, v3 , . . . , vk = u, vk+1 = v, u] is a corresponding walk from u to u of the same length in G.
It is easy to see that distinct walks of length k from v to v in G give distinct walks from u to u of the
same length in G.

25
 Also, as G is a strongly connected digraph with at least 3 vertices, u has a neighbor other than v. So
there exists a walk from u to u of some length k that does not pass through v. Then, for that k we have
(k) (k)
auu > avv . Therefore by Theorem 1.6, we conclude that ρ(Au (t)) = ρ(Av (t)), ∀t ≥ 0 cannot happen.
Therefore the vertex u dominates the vertex v for all t > 0.

Acknowledgement: S.K.’s research is supported by the Natural Sciences and Engineering Research
Council of Canada, grant number RGPIN-2019-05408. The research of S.Pan is funded through the
fellowship of Indian Institute of Technology Guwahati (IIT Guwahati).

References

[1] S. Fallat, D.D. Olesky, P. van den Driessche, Graph theoretic aspects of maximizing the spectral
radius of nonnegative matrices, Linear Algebra Appl. 253 (1997) 61–77.

[2] C.R. Johnson, R. Loewy, D.D. Olesky, and P. van den Driessche, Maximizing the spectral radius of
fixed trace diagonal perturbations of nonnegative matrices, Linear Algebra Appl. 241-243 (1996)
635–654.

[3] C.R. Johnson, D. Stanford, D.D. Olesky, P. van den Driessche, Dominant eigenvalues under trace-
preserving diagonal perturbations, Linear Algebra Appl. 212/213 (1994) 415–435.

[4] A. Berman, R.J. lemmons, Nonnegative Matrices in the Mathematical Sciences, SIAM Publications,
Philadelphia, 1994.

[5] J.E. Cohen, Convexity of the dominant eigenvalue of an essentially nonnegative matrix, Proc.
Amer. Math. Soc. 81 (4) (1981) 657–658.

[6] C.R. Johnson, J.A. Dias da Silva, Symmetric matrices associated with a nonnegative matrix, Circuits
Systems Signal Process 9 (2) (1990) 171–180.

[7] J.E. Cohen, Derivatives of the spectral radius as a function of a non-negative matrix, Math. Proc.
Cambridge Philos. 83 (1978) 183–190.

[8] E. Deutsch, M. Neumann, Derivatives of the Perron root at an essentially nonnegative matrix and
the group inverse of an M-matrix, J. Math. Anal. Appl. 102 (1984) 1–29.

26