42 Basic Concepts of Graphs

1.10 Matrix Representation of Graphs

Definitions:
In this section, we introduce two kinds of matrix representations of a graph,
that is, the adjacency matrix and incidence matrix of the graph.
A graph G with the vertex-set V (G) = {x1 , x2 , · · · , vv } can be described by
means of matrices. The adjacency matrix of G is a v × v matrix

A(G) = (aij ), where aij = µ(xi , xj ) = |EG (xi , xj )|.

For example, for the digraph D and the undirected graph G shown in Figure 1.26,
their adjacency matrices A(D) and A(G) are as follows.
   
0 2 0 0 0 2 1 1
 0 0 0 0   2 0 1 0 
A(D) =  1 1 0 1
 , A(G) = 
 1 1
.
 0 1 
1 0 0 1 1 0 1 1

a1 e1
x1 x2 x1 x2
a2 e2

a6 e6
D: a5 a3 G: e5 e3
a7 e7
a4 e4
x4 x3 x4 x3

Figure 1.26: A digraph D and an undirected graph G

The incidence matrix of a loopless graph G is a v × ε matrix

M(G) = (mx (e)), x ∈ V (G) and e ∈ E(G),

where, if G is directed, then


 1, if x is the tail of e;
mx (e) = −1, if x is the head of e;

0, otherwise,

and if G is undirected, then


1, if e is incident with x;
mx (e) =
0, otherwise.

For example, for the digraph D and the undirected graph G shown in Figure 1.26,
the incidence matrix M(D − a7 ) and M(G) are as follows.
 
1 1 0 0 −1 −1
 −1 −1 −1 0 0 0 
M(D − a7 ) = 
 0
,
0 1 1 0 1 
0 0 0 −1 1 0
1.10. MATRIX REPRESENTATION OF GRAPHS 43
 
1 1 0 0 1 1 0
 1 1 1 0 0 0 0 
M(G) = 
 0
.
0 1 1 0 1 0 
0 0 0 1 1 0 2
The adjacency matrix or the incidence matrix of a graph is another representation
of the graph, and it is this form that a graph can be commonly stored in computers.
The matrix representation of a graph is often convenient if one intends to use a
computer to obtain some information or solve a problem concerning the graph. This
kind of representation of a graph is conducive to study properties of the graph by
means of algebraic methods.
Let
 
1 2 ··· n
σ=
i1 i2 · · · in
be a permutation of the set {1, 2, · · · , n}. Then we obtain an n × n permutation
matrix P = (pij ) defined by

1, if j = σ(i);
pij =
0, otherwise.

It is not difficult to see that the adjacency matrices of two isomorphic graphs
are permutedly similar. In other words, assume that A and B are the adjacency
matrices of two isomorphic graphs G and H, respectively, then there exists a v × v
permutation matrix P such that A = P−1 BP.
Similarly, the incidence matrices of two isomorphic graphs are permutedly equiv-
alent. In other words, assume that M and N are the incidence matrices of two iso-
morphic graphs G and H, respectively, then there exist a v × v permutation matrix
P and an ε × ε permutation matrix Q such that M = PNQ.

Relationship between Matrix and Graphical Representa-

tions:
It is these properties that makes us convenient to study structures of graphs
by using their matrix representations. We now present a very useful result on the
adjacency matrix of a graph as follows.

Theorem 1.11 Let A be the adjacency matrix of a digraph

G with the vertex set {x1 , x2 , · · · , xv }. Then the entry in position
(i, j) of Ak is the number of different (xi , xj )-walks of length k
in G.

Proof: The proof is by induction on k. The result is obvious for k = 1 since

there exist aij (xi , xj )-walks of length one if and only if there exist aij edges from
44 Basic Concepts of Graphs
 
(k−1) (k−1)
xi to xj in G. Let Ak−1 = aij and assume that aij
is the number of
 
(k)
different (xi , xj )-walks of length k − 1 in G; furthermore, let Ak = aij . Since
Ak = Ak−1 · A, we have
v
X
(k) (k−1)
aij = ail · alj . (1.11)
l=1

Every (xi , xj )-walk of length k in G consists of an (xi , xl )-walk of length k − 1, where

xl (1 ≤ l ≤ v) is adjacent to xj , followed by an edge from xl to xj . Thus by the
induction hypothesis and the equation (1.20), we have the desired result.

It should be noted that walks could not be replaced by paths in Theorem 1.11
in general.

It is easy to see that there is the unique (x, y)-walk of length n for any pair
(x, y) of vertices in B(d, n). We obtain from Theorem 1.11 immediately that if A
is the adjacency matrix of B(d, n), then An = J, where J is an n-square matrix
all of whose entries are 1. Similarly, if A is the adjacency matrix of K(d, n), then

An + An−1 = J.

Some Examples:

We will, in Section 1.11 this book, introduce an important application of the

adjacency matrix of a graph, specially Theorem 1.11, in matrix theory. We here
give three examples, which are important results in graph theory, to show that

adjacency and incidence matrices are very useful for studying graphs.

In Example 1.6.3, we show that if G is a strongly connected digraph of order v

and the maximum degree ∆, then

 k + 1, for ∆ = 1;
v ≤ 1 + ∆ + ∆2 + · · · + ∆k−1 + ∆k = ∆k+1 − 1
 , for ∆ > 1.
∆−1

Digraphs attaining this upper bound are called (∆, k)-Moore digraphs.

The following example is due to Plesnik and Znom (1974), and rediscovered by
Bridges and Toueg (1980).

Example 1.10.1 There is no (∆, k)-Moore digraph for ∆ ≥ 2

and k ≥ 2.

Proof: Assume that G is a (∆, k)-Moore digraph whose order n reaches the

Moore bound defined in (1.6), and let A be the adjacency matrix of G. By the
exercise 1.6.2, G is a ∆-regular and simple digraph. Furthermore, by Theorem 1.11,
1.10. MATRIX REPRESENTATION OF GRAPHS 45

we have

I + A + A2 + · · · + Ak = J, (1.12)

where I is an identity square matrix. The expression (1.12) implies that J is a

polynomial in A, and so the matrices A and J have a common set of eigenvectors.

It is not difficult to show that ∆ is an eigenvalue of A (see the exercise 1.10.6).

Let r be any eigenvalue other than ∆, and let X be an eigenvector corresponding to
r. Noting that the zero, as an eigenvalue of J, has the multiplicity n − 1, we have

AX = rX, JX = 0.

By (1.12), we obtain the relation

1 + r + r2 + · · · + rk = 0. (1.13)

The expression (1.13) shows that r has the multiplicity (k + 1) as the unite root, i.e.,
rk+1 = 1. Let r1 , r2 , · · · , rn−1 be n − 1 eigenvalues of A other than ∆. By Theorem
1.11, all the main diagonal entries of Ai (1 ≤ i ≤ k) are 0, that is,

TrAi = 0, i = 1, 2, · · · , k.

Thus the sum of the eigenvalues of Ai

n−1
X
∆i + rji = 0, i = 1, 2 · · · , k. (1.14)
j=1

Since rj r j = |rj |2 = 1 = rjk+1 , it follows that rj−1 = r j = rjk , where r j is the

conjugate complex number of rj . Setting i = 1 and i = k in (1.23), respectively, we
have
n−1
X n−1
X
k
−∆ = rj , −∆ = rjk .
j=1 j=1

Taking the conjugates of the above expressions and noting (1.14), we have that

n−1
X n−1
X
−∆ = rj−1 = rjk = −∆k ,
j=1 j=1

which holds if and only if either k = 1 or ∆ = 1. This contradicts to our assumption

and, thus, the conclusion follows.

By Example 1.10.1, a digraph with the maximum degree ∆ and diameter 2 has
order at most ∆2 + ∆. We have known that the Kautz digraph K(∆, 2) had order

∆2 +∆. Therefore, K(∆, 2) is a maximum (∆, 2)-digraph, which is the unique known
maximum (∆, 2)-digraph up to now.
46 Basic Concepts of Graphs

Similarly, if G is a connected graph of order v and the maximum degree ∆, then

v ≤1 + ∆ + ∆(∆ − 1) + · · · + ∆(∆ − 1)2 + ∆(∆ − 1)k−1

 2k + 1, for ∆ = 2;
= ∆(∆ − 1)k − 2
 , for ∆ > 2.
∆−2

Graphs attaining this upper bound are called (∆, k)-Moore graphs.

Example 1.10.2 (Hoffman and Singleton, 1960) There is no

undirected (∆, 2)-Moore graph for ∆ 6= 2, 3, 7, 57.

Proof Assume that G is an undirected (∆, 2)-Moore graph of order n. Then G

is ∆-regular. The adjacency matrix A = A(G) is a real symmetric n-square matrix
with all main diagonal entries 0. Since G is ∆-regular, ∆ is an eigenvalue of A. Let
I be the identity n-square matrix, and J be an n-square matrix all of whose entries
are 1. Then n is an eigenvalue of J, otherwise are 0. Note d(G) = 2. By Theorem
1.11, all main diagonal entries of A2 are ∆, otherwise are 0 or 1, and the (i, j)-th
entry of A2 is 0 if and only if the corresponding two vertices are adjacent in G, that
is the (i, j)-th entry of A is 1. It follows that

A2 + A − (∆ − 1)I = J. (1.15)

This implies that J is a polynomial in A. So A and J have a common set of

eigenvectors. Suppose that one of these is X corresponding to the eigenvalue ∆.
Then

AX = ∆X, JX = nX.

For this eigenvector, the expression (1.15) supplies the relation ∆2 + 1 = n.

Let Y be any other eigenvector of A corresponding to an eigenvalue r. Then

AY = rY, JY = 0.

Using (1.15), we obtain the relation

r2 + r − (∆ − 1) = 0.

Hence A has two other distinct eigenvalues:

1 √ 1 √
r1 = (−1 + 4∆ − 3 ), r2 = (−1 − 4∆ − 3 ).
2 2

Since A is a real symmetric matrix, it has only real eigenvalues. Thus both r1 and
r2 are real numbers.
1.10. MATRIX REPRESENTATION OF GRAPHS 47

If ∆ is such that r1 and r2 are not rational, then each has multiplicity 12 (n − 1)
as an eigenvalue of A since A is rational. Since all main diagonal entries of A are 0,
by a result on the trace of a matrix, the sum of the eigenvalues of A is 0. Namely

1 1
∆ + (n − 1)(r1 + r2 ) = ∆ − ∆2 = 0. (1.16)
2 2

The value of ∆ satisfying (1.16) is only ∆ = 2, for which n = 5 and the corresponding
undirected (2, 2)-Moore graph is an undirected cycle C5 .
The values of ∆ for which r1 and r2 are rational are those for which there is an
integer s such that s2 = 4∆ − 3. Thus

1 1
r1 = (s − 1), r2 = − (s + 1).
2 2

Let t be the multiplicity of r1 . Then the sum of the eigenvalues of A is

s−1 −s − 1
∆+t + (n − 1 − t) = 0.
2 2

Using the relations n = 1 + ∆2 and s2 = 4∆ − 3, we have that

s5 + s4 + 6s3 − 2s2 + (9 − 32t)s = 15. (1.17)

Since the equation (1.17) requires solutions in integers, the only candidates for s are
the factors of 15. These possible solutions are:
s = ±1, t = 0, ∆ = 1, n = 2;
s = ±3, t = 5, ∆ = 3, n = 10;
s = ±5, t = 28, ∆ = 7, n = 50;
s = ±15, t = 1729, ∆ = 57, n = 3250.
There is no undirected graph of degree 1 and diameter 2. The theorem follows.

The undirected (7, 2)-Moore bound is m(7, 2) = 50, the corresponding undi-
rected (7, 2)-Moore graph is determined by Hoffman-Singleton. The uniqueness of
the undirected (3, 2)- and (7, 2)-Moore graphs have been shown by Hoffman and
Singleton.
The undirected (57, 2)-Moore has order 3250, the corresponding undirected (57, 2)-
Moore graph might exist, but no one has been able to construct (or prove the nonex-

istence of) such a graph so far.

Example 1.10.3 Let M be the incidence matrix of a digraph

G without loops, and let Mi be the matrix obtained from M
by deleting the ith row. Then the algebraic cofactor of any
T T
entry in MM is equal to the determinant det (Mi Mi ), where
T
M denotes the transpose of M.
48 Basic Concepts of Graphs

T
Proof: Suppose that the vertex-set of G is {x1 , x2 , · · · , xv }, and let N = MM .
Suppose that the entry in position (i, j) of N is nij . Then N is symmetric and

dG (xi ) = d+ −
G (xi ) + dG (xi ), for j = i;
nij =
−µ(xi , xj ) − µ(xj , xi ), 6 i.
for j =

Thus the sum of any row and the sum of any column of N all are 0. It is a routine
T
algebraic exercise to show that the algebraic cofactors of all entries in MM have
the same value (the exercise 1.10.5). Let Nij be the algebraic cofactors of nij in
T
MM , and, without loss of generality, suppose

Nij = N11 , ∀ 1 ≤ i, j ≤ v.

Let α1 be the first row vector of M. Then

   T T
!

T α1 T T α1 α1 α1 M1
N = MM = · α1 M1 = T T .
M1 M1 α1 M1 M1

Thus, we have
T
Nij = N11 = det (M1 M1 ), ∀ 1 ≤ i, j ≤ v

as desired.

We conclude this section with some remarks. Let A be the adjacency matrix of
an undirected graph G with the vertex-set {x1 , x2 , · · · , xv }, M the incidence matrix
of any oriented graph D of G, and let B be the v × v diagonal matrix with the
main diagonal entries bii = dG (xi ). It is easily shown that (the exercise 1.10.5)
T
MM = B − A, which is called Laplace matrix in the literature and textbook on
graph theory.
We have known from Example 1.10.3 that all the algebraic cofactors of entries
T T
in MM have the same value, the determinant det (Mi Mi ). This value is a very
important invariant of isomorphic graphs. We will, in Section 2.3, know what this
invariant is.

Exercises: 1.10.4, 1.10.5, 1.10.6

1.10. MATRIX REPRESENTATION OF GRAPHS 49

Exercises
1.10.6 Let A be the adjacency matrix of a graph G (undirected or directed). The
eigenvalues of A is referred to as the eigenvalues of G; the characteristic
polynomial det (λI − A) is referred to as the characteristic polynomial of
G. Suppose that characteristic polynomial of G is

PG (λ) = det(λI − A) = λv + c1 λv−1 + · · · + cv−1 λ + cv .

(a) Count the characteristic polynomials of the following two graphs.

(Exercise 1.10.6)
(b) Prove
X
ck = (−1)ω(H) , k = 1, 2, · · · , v,
H∈Hk

where Hk is the set of (1-) 2-regular subgraphs with order k of (di)graphs of

G. (M.Milic (1964), H. Sachs (1964), L.Spialter (1964))

(c) Prove that c1 = 0; −c2 = ε; and −c3 is equal to twice the number of
triangles in G.

(d) Prove that if λ1 , λ2 , · · · , λv are all eigenvalues of G, then

(i) λ1 + λ2 + · · · + λv = −c1 ;

(ii) the number of different directed closed walks of length k in G

is (λk1 + λk2 + · · · + λkv ).

(e) Let λ be the maximum eigenvalue of G. Prove that

(i) δ + ≤ λ ≤ ∆+ , and δ − ≤ λ ≤ ∆− (or δ ≤ λ ≤ ∆),

and the equalities hold if and only if G is regular;

(ii) if G is strongly connected and regular, then λ has the

multiplicity 1.

(f) Prove that a strongly connected digraph of diameter d has at leat d + 1

distinct eigenvalues.