Quotient graphs and stochastic matrices

arXiv:2411.09157v1 [math.CO] 14 Nov 2024

Frederico Cançado∗, Gabriel Coutinho†

November 15, 2024

Abstract
Whenever graphs admit equitable partitions, their quotient graphs
highlight the structure evidenced by the partition. It is therefore very
natural to ask what can be said about two graphs that have the same
quotient according to certain equitable partitions. This question has been
connected to the theory of fractional isomorphisms and covers of graphs
in well-known results that we briefly survey in this paper. We then depart
to develop theory of what happens when the two graphs have the same
symmetrized quotient, proving a structural result connecting this with the
existence of certain doubly stochastic matrices. We apply this theorem
to derive a new characterization of when two graphs have the same com-
binatorial quotient, and we also study graphs with weighted vertices and
the related concept of pseudo-equitable partitions. Our results connect
to known old and recent results, and are naturally applicable to study
quantum walks.

1 Introduction
Let G and H be graphs on the same number of vertices, with adjacency matrices
AG and AH . The linear programming relaxation of the graph isomorphism
problem consists of searching for a matrix M so that

AG M = M AH and M is doubly stochastic. (1)

Note that if M is required to have integer entries, then any such M is a permu-
tation matrix, and therefore describes an isomorphism between G and H. We
say the graphs G and H are fractionally isomorphic if there exists a matrix M
satisfying (1).
A well-known theorem [20, Chapter 6] provides equivalent characterizations
of fractionally isomorphic graphs.
Theorem 1 (Theorem 6.5.1 in [20]). For two graphs G and H on the same
number of vertices, the following are equivalent.
∗ [email protected].
† [email protected].

1
 (i) They are fractionally isomorphic, that is, there is a doubly-stochastic ma-
trix M such that AG M = M AH .
(ii) G and H have some common equitable partition.
(iii) G and H have in common the coarsest equitable partition.
(iv) D(G) = D(H).
The equivalence between (i) and (ii) was proved in [19], and we explain it
in details below. We overview the remaining items along with other details in
Section 1.1.
Given a partition π = {C1 , . . . , Ck } of the vertex set of graph, let P denote
its characteristic matrix, that is, the 01 matrix with rows indexed by vertices
and k columns, each equal to 1Ci , the indicator vector of the set Ci .
The partition π is called equitable if the colspace of P is AG -invariant.
Equivalently (see [12]), π is equitable if for any v ∈ Ci , the number of neighbours
of v in Cj depends only on i and j, thus denoted simply by pij . The directed
graph with the k cells of π as its vertices and an arc of weight pij from vertex
Ci to Cj is denoted by G/π, and is referred to as the quotient graph of G by π.
Note that if π is equitable, then (see [13, Chapter 9]).

AG P = P AG/π and AG/π = (P T P )−1 (P T AG P ).

A third equivalent definition comes from the notion of symmetrized quotient

graph, which is key to our paper. Given the characteristic matrix P , let Pe
denote the matrix obtained from P upon normalizing its columns, that is,

Pe = P (P T P )−1/2 .

g to be the weighted (unidirected) graph whose adjacency matrix

We define G/π
is given by
eT
g = P AG P ,
AG/π e and thus AG Pe = Pe A g , (2)
G/π

so it follows that π is equitable if and only if AG commutes with Pe Pe T (again,

see [12, Lemma 1.1]).

√
G= G/π = 1 2 g =
G/π 2

1 1

Figure 1: Example of a graph G with an equitable partition π (white and black

vertices), the quotient graph, and the symmetrized quotient graph.

2
 Equitable partitions and graph quotients were likely studied for the first
time in [21]. We briefly mention two recent applications of graph quotients to
completely different fields: the combinatorial quotient graphs, for instance, have
been applied to facilitate the solution of linear programs in [14]; and the sym-
metrized quotients have been used to study perfect state transfer — a quantum
walk phenomenon with applications in quantum computing [5, 16].
To say that two graphs G and H have a common equitable partition, in the
sense of Theorem 1, is to say that there are equitable partitions of G and H
with characteristic matrices P and Q respectively so that

P T P = QT Q, and P T AG P = QT AH Q.

(Note that we need not worry about using equality rather than some isomor-
phism sign, as the columns of each partition matrix may be freely reordered).
Naturally the first condition implies both graphs have the same number of ver-
tices, and together they imply that the quotient graphs are isomorphic.
Note that both conditions imply that AG/π = AH/σ , but the converse is not
necessarily true. A further observation is that

AG/π = AH/σ =⇒ AG/π

g = AH/σ
g,

as the procedure to symmetrize a matrix by diagonal similarity yields a unique

solution. We examine this latter property and we provide an equivalent con-
dition that goes in spirit of the equivalence between (i) and (ii) in Theorem 1.
Our first main result is the following:
Theorem 2. Let G and H be graphs with adjacency matrices AG and AH .
There are equitable partitions π in G and σ in H with equal symmetrized quotient
g = H/σ,
graphs, that is G/π g if, and only if, there is a nonnegative matrix M
satisfying:
(i) Both M M T and M T M are doubly stochastic.
(ii) AG M = M AH .
This Theorem is proved in Section 2. We also present here in the introduc-
tion, in Section 1.2, an application of this theorem to the study of quantum
walks in graphs.
In Section 3 we focus on the problem of what can be said about graphs
and equitable partitions so that AG/π = AH/σ . We prove that this condition
is equivalent to both graphs admitting the same quotient by their coarsest eq-
uitable partition. This is done upon examining how the structure of the set
of equitable partitions behaves when successive quotients are taken. While it
is plausible that the specialist in the topic of equitable partitions will find the
results of this section natural or familiar, we have not been able to locate them
in the literature, so we state them in this paper as original contributions. We
appreciate of course any indication to a suitable reference. We also verify that
techniques based on the work in [14] provide another equivalent condition to

3
AG/π = AH/σ , along the lines of condition (i) in Theorem 1, thus we enrich the
known Theorem 5 with two extra equivalent characterizations.
Lastly, and on the topic of weighted graphs, we discuss what has been called
pseudo-equitable partitions (concept first introduced in [10] and recently studied
in [1,2] in connection with eigenvalue bounds, though in a sense more restricted
than ours). Given G, we say that a partition π with matrix P is pseudo-equitable
if there is a diagonal matrix D of positive vertex weights u so that the colspace of
P is (AD)-invariant. We denote the quotient graph by G/(u, π). This has been
particularly useful to generalize the fact that the trivial partition with only one
class is equitable if and only if the graph is regular. By putting D = Diag(u),
where u is the Perron-eigenvector of G, it follows that the trivial partition with
only one class is pseudo-equitable for all graphs. Going further, note now that
any two graphs with the same largest eigenvalue will be pseudo-fractionally-
isomorphic. More details come in Section 5.
The main results in Section 5 come from analysing the well known results
that transform certain matrices into doubly stochastic matrices by diagonal
conjugation (see [4, 22]) in conjunction to our analysis of equitable partitions.
There we show the following theorem.
Theorem 3. Suppose that G and H are positive-weighted graphs with adjacency
matrices AG and AH respectively. The following are equivalent:
(i) There is a nonnegative matrix M such that the connected components of
the graph ZM with adjacency matrix
 
M
MT

are complete bipartite graphs, and AG M = M AH ;

(ii) there are pseudo-equitable partitions (u, π) and (w, σ) of G and H respec-
tively such that AG/(u,π)
^ = AH/(w,σ)^ .

If the conditions hold true, the connected components of M are in bijection

with pairs of cells of the equitable partitions of each graph.
As a consequence of the previous result, we have a fairly useful description
of all pairs of pseudo-equitable partitions from two distinct graphs that admit
the same symmetrized quotient.
We point out that if AG and AH have the same largest eigenvalue, then both
conditions of the previous theorem are always satisfied for the partition with
only one part.

1.1 Equivalent characterizations of fractional isomorphism

and other notions of similar equitable partitions
Let π, σ be two partitions of a graph. We say that σ is coarser than π, and
denote by π ≤ σ, when each cell of π is contained in some cell of σ. In this case,

4
we can also say that π is finer than σ. Given two partitions, π, σ, there is the
finest partition coarser than both, which is said to be π join σ and is denoted
by π ∨ σ. Likewise, π ∧ σ denotes the meet, that is, the coarsest partition finer
than both. Equitable partitions form a lattice:
Proposition 4 (see [19]). Suppose π, σ are equitable partitions of X. Then
π ∨ σ and π ∧ σ are equitable partitions.
Moreover, all graphs have the finest equitable partition, which is the trivial
one (each cell is a singleton), and joining all equitable partitions will give us the
coarsest equitable partition, which is unique. The equivalence between (i) and
(iii) in Theorem 1 was proved in [23]. Symbols D(G) and D(H) in item (iv)
in Theorem 1 refer to encodings of all partial outputs of the well-known color
refinement algorithm (equivalent to the 1-dimensional Weisfeiler-Leman [24]),
and its equivalence to (iii) is quite straightforward (see [17]).
Assume G and H are graphs and π and σ are their coarsest equitable par-
titions, respectively, so that G/π ≃ H/σ. In [17, Section 3], three equivalent
conditions to this were presented, all based on the notion of a graph cover.
A (possibly infinite) graph Γ covers G via maps γV : V (Γ) → V (G) and
γE : E(Γ) → E(G) if
• u ∈ e in Γ implies γV (u) ∈ γE (e) in G,
• both maps γV and γE are onto,
• γE is locally one-to-one, that is, γE is a bijection between the sets of edges
{e ∈ E(Γ) : u ∈ e} and {e ∈ E(G) : γV (u) ∈ e}.
The graph Γ is called a cover of G. It is straightforward to notice that a
connected graph G has a unique tree that covers it, called the universal cover,
and unless G is itself a tree, this universal cover is an infinite graph (for instance:
if G is a cycle, the universal cover is the infinite path in which all vertices have
degree 2).
Theorem 5 (see [17], Section 3). If G and H are graphs, then the following
are equivalent:
(i) G and H share a common finite cover.
(ii) G and H have the same universal cover.
(iii) G and H share a common (possibly infinite) cover.
(iv) If π and σ are the coarsest equitable partitions of G and H respectively,
then G/π ≃ H/σ.
Our work in Section 3 provides an equivalent condition to these:
Theorem 6. Two graphs G and H admit coarsest equitable partitions π and σ
so that G/π ≃ H/σ if and only if there is a nonnegative nonzero matrix M with
constant row sum and constant column sum such that AG M = M AH .

5
1.2 Application to quantum walks
A continuous-time quantum walk on a graph G is the map U that takes non-
negative real numbers t and returns a symmetric unitary matrix, as

U (t) = exp(itAG ).

There is a very large literature on the topic, and as this is not the main theme of
this paper, we will simply refer to the survey [8] and references therein. A major
problem in this field is to find graphs G so that for some t, an off diagonal entry
U (t) has absolute value 1. In this case, G is said to admit perfect state transfer.
The connection between this and equitable partitions was first studied in [11],
and we summarize what is of our interest as follows. Say π is an equitable
partition of G with partition matrix P , and denote A = AG and B = AG/π g.
Recall that B is symmetric as in (2). Then, as exp(itA) is a polynomial in A,

APe = Pe B =⇒ exp(itA)Pe = Pe exp(itB),

whence if two classes of π are singletons, corresponding to vertices u and v, the

equation above says that | exp(itA)uv | = 1 if, and only if, | exp(itB){u}{v} | = 1.
Thus, if any two graphs admit the same symmetrized quotient B with the two
vertices as singletons, as above, then perfect state transfer in one implies it in
the other. Bottom line: understanding the combinatorics of pairs of graphs that
admit the same symmetrized quotient might lead to new ways of constructing
examples of perfect state transfer, perhaps allowing for optimizing the size of the
graph, as was shown in limited settings in [7, 16], and discussed more lengthily
in [18].

2 Symmetrized quotient
Our goal in this section is to prove Theorem 2. Again, this has been motivated
by the equivalence between (i) and (ii) in Theorem 1. We start with definitions,
and standard lemma (see [20, Chapter 6].
A square matrix M is called decomposable if, for some permutation matrices
P and Q, we have  
A 0
PMQ = ,
0 B
where A, B are square matrices. If M is not decomposable, then it is called
indecomposable. It is quite straightforward to realize that for any square matrix
M , there is a permutation matrix P so that P T M P is a block matrix, each of
its blocks indecomposable.
A square matrix M is called reducible if, for some permutation matrix P ,
we have  
A C
PMPT = ,
0 B

6
where A, B are square matrices, and C is of appropriate dimension. If M is
not reducible, then it is called irreducible, and if P M is irreducible for all
permutation matrices P , we say that M is strongly irreducible.
Lemma 7 (see [20], Proposition 6.2.1). If M is doubly stochastic and indecom-
posable, then M is strongly irreducible.
This will be useful to us when applied together with the celebrated Perron-
Frobenius theorem.
Theorem 8 (see for instance [15], Chapter 8). Let M be an irreducible, non-
negative matrix. Then among all eigenvalues of M with maximum absolute
value, there is one which is positive and its multiplicity is one. This eigenvalue
has an associated eigenvector that is entry-wise positive. Further, eigenvectors
associated with other eigenvalues are not nonnegative.
Recall from Section 1 that if π is an equitable partition and S denotes its
g denotes the symmetric weighted graph
normalized partition matrix, then G/π
whose adjacency matrix is given by
T
AG/π
g = S AG S.

We restate Theorem 2 for convenience.

Theorem 2. Let G and H be graphs with adjacency matrices A = AG and B =
AH . There are equitable partitions π in G and σ in H with equal symmetrized
g ≃ H/σ,
quotient graphs, that is G/π g if, and only if, there is a nonnegative
matrix M satisfying:
(i) Both M M T and M T M are doubly stochastic.
(ii) AM = M B.
Proof. First we prove that if G and H have equal symmetrized quotients, then
conditions (i) and (ii) hold. This is quite straightforward.
Let S and T be normalized matrices related to equitable partitions of G and
H, respectively, satisfying the equation

S T AS = T T BT.

Recalling that S T S = T T T = I, and that A commutes with SS T and B com-

mutes with T T T (as stated right after (2)), it follows that

A(ST T ) = (ST T )B

The matrix M = ST T is nonnegative, as both S and T are nonnegative, and

ST T (ST T )T = S(T T T )S T = SS T ,

(ST T )T ST T = T (S T S)T T = T T T,

7
which are readily seen to be both doubly-stochastic.
Moving on to the converse, suppose a matrix M satisfying conditions (i) and
(ii) exists. In this case, M M T commutes with A since

AM M T = M BM T = (M BM T )T = M M T A.

From Lemma 7, there exists some permutation P such that

 
S1
 S2 
 
T T  S 
P (M M )P =  3 ,
 . .. 
 
Sk

where Sr is irreducible for all r. Let us denote

 
A11 A12 . . . A1k
A21 A22 . . . A2k 
 
P T AP =  .. ,
 . 
Ak1 Ak2 . . . Akk
with the same block sizes as P T (M M T )P . The equation AM M T = M M T A is
equivalent to
Sr Ars = Ars Ss ∀r, s.
Multiplying by 1 we get
Sr Ars 1 = Ars Ss 1,
Sr Ars 1 = Ars 1,
and because Sr is irreducible, by the Perron-Frobenius Theorem we have that
Ars 1 = c1, which means that these blocks correspond to a partition π of V (G)
which is equitable (this previous argument is essentially equivalent to the impli-
cation (i) to (ii) from Theorem 1, and we reference [20, Section 6] for a detailed
exposition).
Let us denote π = {C1 , . . . , Ck }, each Ci ⊆ V (G). By an analogous argu-
ment, the partition of M T M into indecomposable blocks induces an equitable
partition σ of H, say σ = {D1 , . . . , Dm }, Di ⊆ V (H).
In π, we will have that a and b are in the same cell if, and only if, they are in
some irreducible block of M M T , that is, (M M T )ℓa,b 6= 0 for some ℓ. Let us say
this cell is Cr , and denote by 1r its indicator vector. We will show that M T 1r
corresponds in some way to a cell of σ.

• M T 1r is supported in one cell of σ.

We suppose that a and b are in the support of M T 1r and prove that
(M T M )ℓa,b 6= 0 for some ℓ. The support is non-empty, for M M T 1r 6= 0.

8
 Using its decomposition in blocks, each 1-eigenvector v satisfies
 
S1
 S2 
 
 S  T
P 3  P v = v,
 . . 
 . 
Sk

from where each block of P T v is an eigenvector of Ss . As Ss is inde-

composable (and doubly-stochastic), by Perron-Frobenius, we must have
that v is constant in these blocks. Hence, {1s }s forms aPbasis for the
1-eigenspace of M M T and the related eigenprojector is αs 1s 1T
s , for
αs = k1s k > 0.
−2
P P
Let f (x) = cℓ xℓ be such that f (M M T ) = αs 1s 1T
s . If a and b are
not in the same indecomposable block of M M , then ((M T M )ℓ )a,b = 0
T

for any ℓ > 0, whence

X 
0= cℓ (M T M )ℓ+1
a,b
 X  
T T ℓ
= M cℓ (M M ) M
a,b
 X  
= M T
αs 1s 1s M
T
a,b


≥ αr M T 1r 1T r M a,b

= αr (M T 1r )a (M T 1r )b
> 0,

a contradiction. Hence, we may conclude that both a and b belong to the

same cell within the equitable partition of B associated with M T M .
• M T 1r is constant on each cell of σ.
Let η : {1, . . . , k} → {1, . . . , m} defined by having Dη(r) be the cell of σ
that contains the support of M T 1r . We denote the indicator vector of
Dη(r) , which is in RV (H) , by 1η(r) .
Considering the expression (M T M )(M T 1r ) = M T (M M T 1r ) = M T 1r , we
observe that M T 1r is an eigenvector of M T M , just like 1η(r) , and both
are nonzero only on the entries corresponding to the submatrix of M T M
with entries in Dη(r) , which is irreducible. Therefore M T 1r is parallel to
1η(r) by the Perron-Frobenius theorem.
We shall now demonstrate that the function η establishes an isomorphism
g and H/σ.
between G/π g
P
Since 1r = 1 and M T 1 is supported on all entries (if M T has a row
equal to zero, then M T M would not be doubly-stochastic), and since M T is

9
nonnegative, it follows that all cells of the equitable partitions are covered by
{1η(r) }r , or, in others words, η is onto.
Also, η is one-to-one, since:

1η(r) = 1η(s) =⇒ M T 1r k M T 1s =⇒ M M T1r k M M T1s =⇒ r = s.

In particular k = m.
At last, we shall prove that η is a graph isomorphism. We begin by noting
that
kM T 1r k2 = 1Tr (M M 1r ) = 1r 1r ,
T T

from where
M T 1r 1η(r)
p =q .
1 1
T
r r 1Tη(r) 1η(r)
Thus, a direct calculation gives us

1Tη(r) 1η(s) 1T 1s
q Bq = p r M BM T p
1Tη(r) 1η(r) 1η(s) 1η(s)
T 1r 1r
T 1Ts 1s
1Tr 1s
= p AM M T p
1r 1r
T 1Ts 1s
1T 1s
= p r Ap
1Tr 1r 1Ts 1s
or, likewise
(AG/π g )η(r)η(s) .
g )rs = (AH/σ

which completes the proof that η is an isomorphism.

Corollary 9. Let G and H be graphs with adjacency matrices A = AG and
B = AH . Assume there is a nonnegative matrix M satisfying:
(i) Both M M T and M T M are doubly stochastic.

(ii) AM = M B.
Consider the graph ZM whose adjacency matrix is the support of
 
M
.
MT

Let π = {C1 , . . . , Ck } and σ = {D1 , . . . , Dk } be the equitable partitions as in

Theorem 2. Then ZM is a bipartite graph with k connected components, each
equal to Cr ∪ Dr , for r ∈ {1, . . . , k}.

10
Proof. We observe that two vertices a, b ∈ V (G) are in the same cell if and only
if there is some ℓ for which (M M T )ℓa,b 6= 0. This is equivalent to having some
(even) path between a and b in the graph ZM . Likewise for if these are vertices
in V (H).
To see that vertices in Cr are connected to vertices in Dr , let 1r be the
indicator vector of Cr in RV (G) . Note that M T 1r is parallel to the indicator
vector of Dr , both nonzero, thus some entry M T whose row corresponds to a
vertex of Dr and column corresponds to a vertex of Cr is nonzero. In other
words, Cr ∪ Dr belong to the same connected component.

3 Different notions of common quotients

Theorems 1,5 and 2 describe equivalent conditions to three relations between
graphs G and H, namely
(A) There are equitable partitions π and σ of G and H respectively, with par-
tition matrices P and S, so that
P TP = S TS and P T AG P = S T AH S.
(This is the condition used in Theorem 1.)
(B) There are equitable partitions π and σ of G and H respectively, with par-
tition matrices P and S, so that
(P T P )−1 P T AG P = (S T S)−1 S T AH S.
(This is the condition used in Theorem 5.)
(C) There are equitable partitions π and σ of G and H respectively, with nor-
malized partition matrices Pe and S,
e so that
fT AH S.
PfT AG Pe = S e

(This is the condition used in Theorem 2.)

The attentive reader will have noticed that while Theorem 5 speaks about the
coarsest equitable partitions, we have stated relation (B) with no such restric-
tion. This is fine because of the following result, which should not be surprising
but we could not find a reference to it.
Recall the lattice of equitable partitions introduced in Section 1.1, which
we denote by Π(G). If σ ∈ Π(G), we denote its characteristic matrix by Pσ ,
and Dσ = (PσT Pσ )−1 . Because a partition corresponds to a partition matrix, we
might abuse the notation and say that P ∈ Π(G) if P = Pπ , when π ∈ Π(G).
Proposition 10. Let G be a graph on n vertices. Let σ ∈ Π(G). Define the
function
ϕ : {π ∈ Π(G) : σ ≤ π} → Π(G/σ),
π 7→ Dσ PσT Pπ .

11
Then ϕ is an order-preserving bijection. If S ∈ Π(G/σ), its inverse is given by

ϕ−1 (S) = Pσ S,

and we also have

(G/σ)/ϕ(π) ≃ G/π.
Proof. Fix some π ∈ Π(G). The fact that Dσ PσT Pπ is a partition matrix of
V (G/σ) follows immediately — it is the partition induced by π on the classes
of σ. To see that it is an equitable partition, we proceed as follows:

AG/σ (Dσ PσT Pπ ) = (Dσ PσT Pσ )AG/σ (Dσ PσT Pπ ) (3)

= (Dσ PσT )APσ (Dσ PσT Pπ ) (4)
= (Dσ PσT )APπ (5)
= (Dσ PσT Pπ )AG/π (6)

where (3) follows from Dσ PσT Pσ = I; (4) from the fact that σ is an equitable
partition, thus APσ = Pσ AG/σ ; (5) because Pσ Dσ PσT is an orthogonal projection
and the columns of Pπ are in its range; and (6) because π is an equitable partition
of G.
To see that ϕ is one-to-one, a simple calculation gives us

Dσ PσT Pπ = Dσ PσT Pν =⇒ Pσ Dσ PσT Pπ = Pσ Dσ PσT Pν =⇒ Pπ = Pν .

For each S that is a partition matrix of an equitable partition of G/σ, we

have AG/σ S = SC for some C. Applying Pσ we get

Pσ AG/σ S = Pσ SC =⇒ AG Pσ S = Pσ SC.

As each line of Pσ has only one nonzero entry, we can also conclude that Pσ S is a
matrix of 0’s and 1’s. Thus, it is an equitable partition of G. As Dm uPσT Pσ S =
S, the image of Pσ S is indeed S, as we wanted.
Finally, we note that π ≤ ν if and only if rng Pπ ⊇ rng Pν , and as the
composition of functions preserves the inclusion of their range, we conclude
that ϕ preserves order.
As a consequence, it follows that the quotient graph by the coarsest equi-
table partition of G coincides with the quotient graph by the coarsest equitable
partition of any of its quotients graphs by other equitable partitions. We may
therefore add one extra equivalent condition to the formulation of Theorem 5,
justifying the phrasing of Condition (B). Motivated partly by the work in [14],
we provide a third characterization after the lemma.
Lemma 11. Let G, H be graphs with equitable partitions π and σ and ϕ : Cr 7→
Dr a map between the cells of π and σ. Then ϕ induces an isomorphism from
G/π to H/σ if and only if ϕ induces an isomorphism from G/πg to H/σg and
there is some constant λ for which |Cr | = λ|Dr |.

12
Proof. First off, recall that if G/π and H/σ are isomorphic, then we may assume
AG/π = AH/σ , and because there is only one way to symmetrize these matrices
g and
by diagonal similarity, it follows that A g = A g , so it follows that G/π
G/π H/σ
g are isomorphic.
H/σ
g ≃ H/σ
We observe that G/π g is equivalent to having a convenient indexing
which satisfies
N (Cr , Cs ) N (Dr , Ds )
p = p ∀r, s;
|Cr ||Cs | |Dr ||Ds |
which is also equivalent to
p p !
N (Cr , Cs ) |Dr | |Cs | N (Dr , Ds )
= p ·p ∀r, s.
|Cr | |Cr | |Ds | |Dr |

Recalling that
N (Cr , Cs )
(AG/π )r,s = ,
|Cr |
we have an isomorphism in the non-symmetrized quotient if and only if
p p
|Dr | |Cs |
p ·p =1 ∀r, s;
|Cr | |Ds |

which is equivalent to the desired condition.

Now, we can enunciate and prove the following:
Theorem 12. If G and H are graphs, then the following are equivalent:
(i) If π and σ are the coarsest equitable partitions of G and H respectively,
then G/π ≃ H/σ.
(ii) There are equitable partitions π and σ of G and H respectively such that
G/π ≃ H/σ.
(iii) There is a nonnegative and nonzero matrix M with constant row sum and
constant column sum such that AG M = M AH .
Proof. As shown in Proposition 10, the first two conditions are equivalent.
We start by proving (iii) =⇒ (ii). Normalizing M , we may assume that
M M T and M T M are doubly stochastic. Applying Theorem 2, it follows that
g ≃ H/σ,
G/π g for the equitable partitions π, σ induced by M .
Let 1Cr be the indicator of Cr . By the argument in the proof of Theorem
2, we have M 1Cr parallel to the indicator vector of the cell isomorphic to Cr
in the partition σ of H, which we will denote by Dr . As M 1 is a multiple of 1,
and the vectors M 1Cr have disjoint support, then for each r, M 1Cr must be a
multiple of 1Dr by a factor that is independent of r. Hence, using the previous
lemma we obtain the result.

13
 For the converse, recall that AG PeSeT = Pe SeT AH , where Pe, Se are the normal-
ized matrices of π, σ. A simple calculation gives us that
X
(Pe SeT 1)k = Pek,r Ser,s
r,s
1 X 1
=p p , where k ∈ Cr ,
|Cr | s∈Dr |Dr |
s
|Dr |
=
|Cr |
= γ.

The existence of a value of γ independent of r follows from the lemma above.

An analogous argument shows that Pe SeT is column stochastic.
Coming back to the conditions in the beginning of this section, note the
obvious implication (A) =⇒ (B) =⇒ (C). None of them is an equivalence,
and we show below a large class of examples satisfying only condition (C).

In general, it is easy to see that Kr,s and Kr′ ,s′ have a common symmetrized
quotient if rs = r′ s′ , and naturally if r 6= r′ it cannot be that they admit the
same combinatorial quotient.
While Conditions (A) and (B) are clearly equivalence relations, Condition
(C) is not, as the following example shows:

√
2

1 1

In which the first graph is related to the second, and the second to the third
but the first and last one are not related.
Lemma 13 (see for instance Theorem 6.2.4 in [20]). If S and T are doubly
stochastic matrices and v = Su and u = T v, then v = P u, for some permutation
matrix P .

14
4 Pseudo-equitable partitions — a linear alge-
bra characterization
The concept of a pseudo-equitable partition was likely introduced in [10], and
has been recently studied in [2] (see references therein for other works on the
topic). Essentially a partition is pseudo-equitable if there is a positive vertex
weighting that makes it equitable. In most cases, it turns out that this vertex
weighting is given by the eigenvector corresponding to the largest eigenvalue
(the Perron-eigenvector, as it is sometimes called), but in this section we allow
for any positive weighting.
Let w ∈ RV+ be a vector with all positive entries, and let Dw = Diag(w)
be the diagonal matrix whose diagonal entries correspond to w. We say that a
partition π of V (G) is pseudo-equitable with respect to w if the column space
of Dw Pπ is AG -invariant, or better yet, if there is a matrix B such that

AG (Dw Pπ ) = (Dw Pπ )B.

It is immediate to note that an equivalent way of seeing this is to consider that

−1
π is equitable for the weighted graph Dw AG Dw . We say that the matrix B is
the adjacency matrix of the quotient graph G/(w, π), that is, B = AG/(w,π) .
We are particularly invested in scenarios where two graphs admit the same
quotient graph, but perhaps one of them by means of a pseudo-equitable parti-
tion.

Figure 2: Example of two graphs that admit the same quotient graph with
respect to the classes with the same horizontal coordinate. However for the
graph on the right hand side it is necessary to use a pseudo-equitable partition
whose vector of weights assigns 1 to all white vertices, and 2 to the black vertex.

One application comes from quantum walks, as discussed in Section 1.1. The
graph on the left hand side in Figure 2 admits perfect state transfer because it
is a cartesian power of P2 , and therefore its symmetrized quotient graph also
does. Because the graph on the right hand side admits the same quotient, it
also admits perfect state transfer (see [6, Section 5]).
In the next section we will characterize when two graphs admit the same
symmetrized quotient with respect to pseudo-equitable partitions, but let us
first show a theorem that associates to pseudo-equitable partitions very natural
objects.
We need the following lemma, whose proof can be found in [3].

15
Lemma 14. An orthogonal projection matrix P has nonnegative entries if and
only if rng P has an orthonormal basis with only nonnegative vectors.
For the next theorem, it will be convenient to assume all pseudo-equitable
partitions are presented with weight vectors that have been normalized in each
class of the partition (this is without loss of any generality). We denote by Π̂(G)
the set of all pairs of the form (w, π) where π is a pseudo-equitable partitions
of G with weight vector w, and w is normalized in each class of π.
Let P(G) denote the set of all orthogonal projection matrices with nonneg-
ative entries that satisfy the following two properties
(A) If M ∈ P(G), then there is a positive vector u in its range.
(B) M commutes with AG .
Note that if M ∈ P(G), we may decompose it into irreducible blocks. Each block
has rank 1 and contains a unique positive eigenvector to the eigenvalue 1, as a
consequence of the Perron-Frobenius theorem (Theorem 8). The vector which
is the sum of the normalized Perron eigenvectors of each irreducible block is
positive and belongs to the range of M . We call it the special positive eigenvector
of M .
Given (w, π) ∈ Π̂(G), we let Pπ denote its (unweighted) characteristic ma-
trix, and P(w,π) = Dw Pπ its weighted characteristic matrix. Note that its
columns are already normalized. Then we denote by S(w,π) the projection onto
the column space of P(w,π) , that is
T
S(w,π) = P(w,π) P(w,π) = Dw Pπ PπT Dw . (7)

With that, we can relate projectors and pseudo-equitable partitions by the fol-
lowing theorem.
Theorem 15. Let G be a graph. The following function is a bijection

ϕ : Π̂(G) → P(G),
(w, π) 7→ S(w,π) .

Then, if π and σ are pseudo-equitable partitions with respect to the same weight
vector w, now dropped from the subscripts for clarity, it follows that:
(i) π ≤ σ ⇐⇒ Sπ  Sσ ⇐⇒ rng Sπ ⊇ rng Sσ ,
(ii) Sπ∨σ is the orthogonal projection onto rng Sπ ∩ rng Sσ .
Proof. Given (w, π), the matrix S(w,π) is a nonnegative orthogonal projection
by definition, and it commutes with AG due to an argument analogous to [12,
Lemma 1.1] and shown in [1, Lemma 3.1] for when w is the Perron-eigenvector.
Equation (7) gives an immediate way to verify that w is an eigenvector of S(w,π) ,
and in fact it is going to be the special positive eigenvector of S(w,π) .
Given two elements of Π̂(G), if they correspond to different partitions, then
clearly their images under ϕ are different. If the partitions are equal but the

16
weight vectors are different, the unicity of the special positive eigenvector implies
that their images will be different. Thus ϕ is injective.
Now, let P ∈ P(G) be arbitrary. Using Lemma 14, we have an orthonormal
basis ζr of nonnegative vectors for rng P . As ζr are nonnegative, we must have
for r 6= s that supp ζr ∩ supp ζs = ∅ to ensure orthogonality, and because there
is at least one positive eigenvector of P , the supports of the ζr correspond to
a decomposition of P into irreducible blocks. The sum of the ζr is the special
positive eigenvector of P , which we call u. Thus we conclude that P = S(u,π) ,
where π is the partition into {supp ζr }r . The fact that π is pseudo-equitable with
weighting u follows from the fact that P and A commute (again, see [12, Lemma
1.1] or [1, Lemma 3.1]).
Fix w and assume (w, π) and (w, σ) are pseudo-equitable. We drop w from
the subscripts for clarity.
Condition (i) follows from the fact that π ≤ σ if and only if col Pπ ⊇ col Pσ ,
combined with the property rng Sπ = col Pπ , and that Sπ and Sσ are projections.
Based on (i), we can conclude that rng Sπ∨σ is the largest vector space which
is contained in both rng Sπ and rng Sσ , and whose projector is also in P(G).
The projector onto rng Sπ ∩ rng Sσ is indeed in P(G) since it is the limit of
nonnegative operators that commute with A: let S = limn (Sπ Sσ Sπ )n — it
indeed exists and is a projector, as all entries of the Jordan matrix of Sπ Sσ Sπ
converge to zero, except the ones in the diagonal. It is self-adjoint as it is the
limit of self-adjoint operators, and rng S = rng Sπ ∩ rng Sσ because as Sπ Sσ Sπ
is a contraction, the fix points of the convergent are the same as the fixed points
of Sπ Sσ Sπ , which are rng Sπ ∩ rng Sσ .

5 Characterizing pseudo-equitable partitions

In this section, we present an analysis of when it is possible to use the technology
from Theorem 2 to understand pseudo-equitable partitions.
Given graphs G and H, suppose that for some nonzero nonnegative matrix
M we have AG M = M AH . Is there a weighting of these graphs that result in
pseudo-equitable partitions whose quotients are the same? It turns out that this
is the case provided there are diagonal matrices D, E such that DM E 2 M T D
and EM T D2 M E are doubly stochastic. Are these always available?
We will use a well-known theorem (see for instance [22]) in our result. Be-
forehand, a definition. We say that an n × n matrix M has total support if for
each nonzero entry Ma,b we have some permutation σ ∈ Sn such that σa = b
and for each r, Mr,σr 6= 0.
Theorem 16 (from [22]). Given a nonnegative square matrix M , there are
diagonal matrices D, E with positive nonzero entries such that DM E is doubly
stochastic if and only if M has total support. Moreover, the matrix DM E is
unique; and D, E are also unique up to scalar multiplication if and only if M is
fully indecomposable.

17
 The following result can be found in [9].
Corollary 17. Given a symmetric matrix M with total support, we have a
unique diagonal matrix D with positive nonzero entries such that DM D is dou-
bly stochastic.
We now start by describing which conditions on M we need in order to apply
the theorem. We will from now on suppose that M is k × l and call Rr the rth
row-vector of M and Cr the rth column-vector.
Lemma 18. Given a matrix M we have that M M T has total support if and
only if M has no null rows. Moreover, in this case, M DM T also has total
support for any positive diagonal matrix D.
Proof. We start by observing that

(M M T )a,b 6= 0 ⇐⇒ Ra 6⊥ Rb .

If M M T has total support, then it is immediate from the definition that no row
of M M T is zero, thus M has no null rows.
For the converse, we have that if (M M T )a,b 6= 0 then for σ = (ab) we have
Ra 6⊥ Rb by hypothesis and Rx 6⊥ Rx since Rx 6= 0.
Putting D in the product M DM T does not change the orthogonality rela-
tions, whence the result is still valid.
Theorem 19. Let M be a k × l nonnegative matrix which possesses no null row
or column, and such that for permutations matrices P, Q we have
 
M1
 M2 
 
M =P .  Q,
 .. 
Mp

where Mr has positive entries.

Then, there is a k × k positive diagonal D and a l × l positive diagonal E
such that for N := DM E, both N N T and N T N are doubly stochastic.
Proof. We will use Corollary 17 to recursively find matrices that satisfy partially
what we wanted. Then, we will examine the sub-sequences of these matrices
and conclude that there is one of them that converges, by Bolzano-Weierstrass.
The limit will be the matrices we wanted.
By Lemma 18 and Corollary 17, there is a k × k diagonal D1 , for which
D1 M M T D1 is doubly stochastic. Applying this again, there is a l × l diagonal
E1 for which E1 M T D12 M E1 is doubly stochastic. An induction argument gives
us Di , Ei for which
Di+1 M Ei2 M T Di+1 1 = 1; (8)
Ei M T Di2 M Ei 1 = 1. (9)

18
 For each indecomposable block Mr Equation (8) and Equation (9) still hold,
with Mr replacing M and the appropriate diagonal block of Ds and Es replacing
the whole diagonal. So we can prove the result for each block independently
and conclude that it is valid for the whole matrix M . For the sake of simplicity,
we will maintain the notation and suppose that M has only positive entries.
Denoting di,x := (Di )x,x and ei,y := (Ei )y,y , Equation 9 is equivalent to
m X
X n
ei,b Mx,b d2i,x Mx,y ei,y = 1; b = 1, 2, . . . , l.
x=1 y=1

We will now prove that each sequence di,x , ei,y is bounded above and below
by nonzero constants, which implies that Di , Ei converge for some sub-sequence
to a positive diagonal.
Equation (9) gives us that
1 1
ei,b = P 2 ≤ 2 2 e ,
x,y Mx,b di,x Mx,y ei,y di,a Ma,b i,b

from where, calling M := maxx,y {Mx,y } and m := minx,y {Mx,y }, we have:

1
di,x ei,y ≤ , ∀x, y. (10)
m
Equation (9) and inequality (11) imply
1 ml
ei,b = P 2 M ≥ 2P . (11)
x,y M x,b di,x x,y e i,y M x di,x

Taking sub-sequences we may suppose that all di,x , ei,y either converge (pos-
sibly to zero) or go to infinity. Inequality (10) assures us that if some di,x goes
to infinity then all ei,y go to zero; and if some ei,y goes to infinity then all di,x
go to zero.
Inequality (11) implies that if some ei,y goes to zero then some di,x goes to
infinity.
Thus, if no di,x or ei,y goes to zero, then no sequence goes to infinity and
we are done. Suppose otherwise. We have now two cases: for some x, di,x → 0
or there is y for which ei,y → 0. We will consider the latter. Equation (9), will
imply, by the above argument, that ei,y → 0 for all y.
Taking sub-sequences we can suppose that each of the following limits exists:
ei,y′
→ Ly′ ,y ∈ [0, ∞].
ei,y
We will denote by y ≪ y ′ when Ly′ ,y = ∞. This relation is transitive and
anti-symmetrical, hence is a partial order on the values of y, and we have some
maximal element, which we will call c.
Now, call Ẽi := e−1
i,c Ei and D̃i := ei,c Di . As all ei,y → 0, by taking some
sub-sequence, we can suppose that
ei+1,c
→ 0.
ei,c

19
 By Equation (8) we conclude that
 2
ei+1,c
D̃i+1 M Ẽi2 M T D̃i+1 1 = 1;
ei,c
and
Ẽi M T D̃i2 M Ẽi 1 = 1.
As Equation (9) is still valid, and Ẽc,c = 1, we will have that no entry of
Ẽi goes to zero. Also, since y 6≪ c for all y, we have (Ẽi )y,y 6→ ∞. Whence,
Ẽi → Ẽ, for some sub-sequence. As Ẽi M T D̃i converges (for some sub-sequence)
due to Inequality (10) and the fact that Ec,c = 1, we have D̃i → D̃. This implies
D̃M Ẽ 2 M T D̃1 = 0, which contradicts ẼM T D̃M 2 Ẽ 1 = 1.
The case di,x → 0 for some x is analogous. We change the role of di,x and
ei,y in Inequality (11) and use Equation (8) instead of Equation (9), and the
argument follows similarly.
The theorem below follows immediately.
Theorem 20. Suppose that G and H are positive-weighted graphs. The follow-
ing are equivalent
(i) There is a nonzero nonnegative matrix M such that the graph ZM with
adjacency matrix  
M
MT
has as connected components complete bipartite graphs, and AG M = M AH ;
(ii) there are pseudo-equitable partitions (u, π) and (w, σ) of G and H respec-
tively such that AG/(u,π
^ ) = AH/(w,σ)
^ .
1

In the case (i) =⇒ (ii), the classes and isomorphism are defined by the con-
nected components of ZM .

6 Open problems
There exist polynomial time algorithms to decide whether two graphs are frac-
tionally isomorphic and whether two graphs admit isomorphic quotients with
respect to their coarsest equitable partitions. An obvious question is whether
there is an efficient algorithm to decide with two graphs admit isomorphic sym-
metrized quotients.
Our original motivation to look at symmetrized quotients is the connection
with quantum walks shown in Section 1.2. The main problem we would like to
solve is that of finding the smallest graph whose symmetrized quotient matrix is
a given one. We warn this is a hard problem, as evidenced by the efforts in [16].
Fractional isomorphism can be characterized by means of tree homomor-
phism count. Is there an analogous theory for isomorphic quotients (combina-
torial or symmetrized)?

20
Acknowledgements
Authors thank Thomás J. Spier, Chris Godsil and Aida Abiad for conversations
about the topic of this paper.
Frederico Cançado acknowledges support from CAPES. Gabriel Coutinho
acknowledges support from CNPq and FAPEMIG.

References
[1] Aida Abiad. A characterization of weight-regular partitions of graphs. Elec-
tronic Notes in Discrete Mathematics, 68:293–298, 2018.
[2] Aida Abiad, Christopher Hojny, and Sjanne Zeijlemaker. Characterizing
and computing weight-equitable partitions of graphs. Linear Algebra and
its Applications, 645:30–51, 2022.

[3] Daniel Brosch and Etienne de Klerk. Jordan symmetry reduction for conic
optimization over the doubly nonnegative cone: theory and software. Op-
timization Methods and Software, 37(6):2001–2020, 2022.
[4] Richard A Brualdi, Seymour V Parter, and Hans Schneider. The diagonal
equivalence of a nonnegative matrix to a stochastic matrix. Journal of
Mathematical Analysis and Applications, 16(1):31–50, 1966.
[5] Matthias Christandl, Nilanjana Datta, Tony C. Dorlas, Artur Ekert, Alas-
tair Kay, and Andrew J. Landahl. Perfect transfer of arbitrary states in
quantum spin networks. Phys. Rev. A, 71:032312, Mar 2005.
[6] Gabriel Coutinho. Spectrally extremal vertices, strong cospectrality, and
state transfer. The Electronic Journal of Combinatorics, pages P1–46, 2016.
[7] Gabriel Coutinho. Quantum walks and the size of the graph. Discrete
Mathematics, 342(10):2765–2769, 2019.
[8] Gabriel Coutinho and Chris Godsil. Continuous-time quantum walks in
graphs. The Bulletin of the International Linear Algebra Society, page 5,
2018.
[9] Judit Csima and Biswa Nath Datta. The DAD theorem for symmetric non-
negative matrices. Journal of Combinatorial Theory, Series A, 12(1):147–
152, 1972.
[10] M.A. Fiol, E. Garriga, and J.L.A. Yebra. Locally pseudo-distance-regular
graphs. Journal of Combinatorial Theory, Series B, 68(2):179–205, 1996.
[11] Yang Ge, Benjamin Greenberg, Oscar Perez, and Christino Tamon. Per-
fect state transfer, graph products and equitable partitions. International
Journal of Quantum Information, 9(03):823–842, 2011.

21
[12] C.D. Godsil. Compact graphs and equitable partitions. Linear Algebra and
its Applications, 255(1):259–266, 1997.
[13] Chris Godsil and Gordon F Royle. Algebraic graph theory, volume 207.
Springer Science & Business Media, 2001.
[14] Martin Grohe, Kristian Kersting, Martin Mladenov, and Erkal Selman.
Dimension reduction via colour refinement. In Algorithms-ESA 2014: 22th
Annual European Symposium, Wroclaw, Poland, September 8-10, 2014.
Proceedings 21, pages 505–516. Springer, 2014.
[15] Roger A Horn and Charles R Johnson. Matrix analysis. Cambridge uni-
versity press, 2012.
[16] Alastair Kay. The perfect state transfer graph limbo. arXiv preprint
arXiv:1808.00696, 2018.
[17] Frank Thomson Leighton. Finite common coverings of graphs. Journal of
Combinatorial Theory, Series B, 33(3):231–238, 1982.
[18] Frederico Cançado Pereira. Exploring quantum walks: weighted paths and
quotient graphs unveiled. Master’s thesis, Universidade Federal de Minas
Gerais, 2023.
[19] Motakuri V Ramana, Edward R Scheinerman, and Daniel Ullman. Frac-
tional isomorphism of graphs. Discrete Mathematics, 132(1-3):247–265,
1994.
[20] Edward R Scheinerman and Daniel H Ullman. Fractional graph theory: a
rational approach to the theory of graphs. Courier Corporation, 2013.
[21] Allen J Schwenk. Computing the characteristic polynomial of a graph.
In Graphs and Combinatorics: Proceedings of the Capital Conference on
Graph Theory and Combinatorics at the George Washington University
June 18–22, 1973, pages 153–162. Springer, 2006.

[22] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and
doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343–348,
1967.
[23] Gottfried Tinhofer. Graph isomorphism and theorems of Birkhoff-type.
Computing, 36(4):285–300, 1986.
[24] Boris Weisfeiler and Andrei Leman. The reduction of a graph to canonical
form and the algebra which appears therein. nti, Series, 2(9):12–16, 1968.

22
Gabriel Coutinho
Dept. of Computer Science
Universidade Federal de Minas Gerais, Brazil
E-mail address: [email protected]

Frederico Cançado
Dept. of Computer Science
Universidade Federal de Minas Gerais, Brazil
E-mail address: [email protected]

23