American Journal of Combinatorics Research Article

Volume 2 (2023), Pages 59–71

Eigenvalues of multipart matrices and their applications

Ranjit Mehatari
(Communicated by Sudipta Mallik)

Abstract
A square matrix is called a multipart matrix if all its diagonal entries are zero
and all other entries in each column are constant. In this paper, we describe various
interesting spectral properties of multipart matrices. We provide suitable bounds for
the spectral radius of a multipart matrix. Later on, we show applications of multipart
matrices in spectral graph theory.

1 Introduction
A graph G = (V, E) is called a complete k-partite graph if its vertex set is partitioned as
V = V1 ∪ V2 ∪ · · · ∪ Vk such that two vertices u and v are adjacent if and only if u ∈ Vi
and u ∈ Vj where i 6= j, 1 ≤ i, j ≤ k. A complete k-partite graph with partition sizes
p1 , p2 , . . . , pk is denoted by Kp1 ,p2 ,··· ,pk . For a complete k-partite graph G with partition sizes
p1 , p2 , . . . , pk , the vertex set partition V1 ∪ V2 ∪ · · · ∪ Vk consisting of independent subsets of
G is an equitable partition. The quotient matrix Q for that equitable partition is given by
 
0 p2 p3 · · · pk
 p1 0 p3 · · · pk 
 
Q =  p1 p2 0 · · · pk  .
 
 .. .. .. . 
 . . . · · · .. 
p1 p2 p3 · · · 0

Let A denote the (0, 1)-adjacency matrix of Kp1 ,p2 ,··· ,pk . Then every eigenvalue of Q is
also an eigenvalue of A [5, Theorem 9.3.3]. Let G be a simple, connected, finite graph and
A be the (0, 1)-adjacency matrix of G. Let D be the diagonal matrix whose i-th diagonal
entry is the degree of the vertex i. The matrix A = D−1 A is called the normalized adjacency
1 1
matrix of G. The matrix A is similar to the matrix D− 2 AD− 2 , known as the Randić matrix
MSC2020: 05C50; Keywords: Multipart matrix, Complete multipartite graph, Equitable partition, Eigen-
value bound

©
Received May 15, 2023; Revised Aug 31, 2023; Accepted Sep 1, 2023; Published Sep 3, 2023
The author(s). Released under the CC BY 4.0 International License

59
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

[2]. We observe that the quotient matrix of A for the complete multipartite graph Kp1 ,p2 ,··· ,pk
is similar to
p2 p3 pk
0 · · · n−p
 
n−p2 n−p3 k
 p1 0 p3 pk
· · · n−p 
 n−p 1 n−p3 k 
 p1 p2
0 pk
· · · n−p
Q =  n−p1 n−p2 k .

 . .. .. .. 
 .. . . ··· . 
p1 p2 p3
n−p1 n−p2 n−p3
··· 0
Inspired by this special form of quotient matrices, we define a special class of matrices
and call them multipart matrices.
Let P = {p1 , p2 , . . . , pn } be a finite sequence of real numbers. The multipart matrix
A(P ) = [aij ] corresponding to the sequence P is an n × n matrix whose entries are given by
(
pj if i 6= j,
aij =
0 if i = j.

In other words, A(P ) is a n × n matrix whose j-th column is pj (e − e j ), where e is

the column vector of all ones and e j is the j-th standard basis element of Rn . It is easy
to verify that the eigenvalues of A(P ) do not depend on the arrangement of the entries in
the sequence P . Thus, without loss of generality, we consider the sequence P to be non-
decreasing and nonzero. Since the quotient matrices for both adjacency and normalized
adjacency matrices of complete multipartite graphs are special cases of multipart matrices,
the study of eigenvalues for multipart matrices seems significant and interesting.
There are several papers focused on adjacency eigenvalues of complete multipartite
graphs. Esser and Harary [4] discussed very interesting properties of the adjacency ma-
trix of complete multipartite graphs. In [1, 8] authors characterized complete multipartite
graphs having integral adjacency eigenvalues. Bapat and Karimi [1] provided examples of
non-isomorphic complete multipartite graphs with the same spectrum. Delorme obtained [3]
an explicit formula for the characteristic polynomial for complete multipartite graphs. Also,
he used optimization techniques to find the eigenvalues of a class of complete multipartite
graphs. Stevanović et al. [7] derived results related to the spectral radius and energy of
complete multipartite graphs. In this paper, we show that all these results can be easily
derived by considering the quotient matrix of a complete multipartite graph as a special case
of multipart matrices and taking P as a finite sequence of natural numbers. On the other
hand, the normalized adjacency eigenvalues of complete multipartite graphs were not much
studied.
In this paper, we first investigate the spectral properties of multipart matrices. Using
them, we show that some of the existing results of [3, 4] for the adjacency matrix of complete
multipartite graphs can be explained very easily. Later on, we establish spectral properties
for the normalized adjacency matrix of complete multipartite graphs. We also provide lower
and upper bounds for the largest adjacency eigenvalue of multipart matrices, that provides
suitable bounds for the spectral radius of complete multipartite graphs.

60
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

2 Eigenvalues of multipart matrices

In this section, we discuss various spectral properties of multipart matrices. For a finite
sequence P = {p1 , p2 , . . . , pn }, we show that some of the eigenvalues A(P ) can be determined
directly from the structure of P . We call such an eigenvalue a regular eigenvalue of A(P ),
otherwise we call it a non-regular eigenvalue of A(P ). We start with a determinant formula.
Theorem 2.1. Let P = {p1 , p2 , . . . , pn } be a finite sequence of real numbers. Then

det A(P ) = (−1)n−1 (n − 1)p1 p2 · · · pn .

Proof. We observe that,

det A(P ) = p1 p2 · · · pn det(Jn − In ),

where Jn is the all-ones square matrix of order n. Since Jn − In has eigenvalues −1 (with
multiplicity n − 1 and n − 1 (with multiplicity 1), we get the required result.
Corollary 2.2. If pi 6= 0 for all i = 1, 2, . . . , n, then A(P ) is nonsingular and consequently,
0 is not an eigenvalue of A(P ).
Theorem 2.3. Let P = {p1 , p2 , . . . , pn } be a finite sequence of real numbers, and let ΦP (x) =
xn + c1 xn−1 + · · · + cn be the characteristic polynomial of A(P ). Then
XY
ci = −(i − 1) pj ,
S j∈S

where the summation runs over all subsets S of {1, 2 . . . , n} of order i.

Proof. Since trace A(P ) is 0, the above equation holds for i = 1. For i ≥ 2, we consider the
following formula:
ci = (−1)i (sum of i × i principal minors).
Since an i × i principal minor of A(P ) is also a multipart matrix of order i, hence the
result follows by Theorem 2.1.
Theorem 2.4. If pi > 0 for all 1 ≤ i ≤ n, then the eigenvalues of A(P ) are all real.
1 1
Proof. Let D = diag[p1 , p2 , . . . , pn ]. Then the matrix D 2 A(P )D− 2 is a symmetric matrix
and similar to A(P ). Hence the eigenvalues of A(P ) are real.
Remark 2.5. If pi < 0 for some i, then the above result may not hold. Consider P = {1, −1}.
Then  
0 −1
A(P ) = ,
1 0
whose eigenvalues are ±i.
Theorem 2.6. Let P = {p1 , p2 , . . . , pn } be a finite sequence of non-decreasing real numbers.
If p1 = p2 = · · · = pl = p 6= pl+1 , then −p is an eigenvalue of A(P ) with multiplicity exactly
l − 1.

61
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

Proof. For all i ≥ 1, we define column vectors in Rn as follows:

i
X
Xi = e k − ie i+1 .
k=1

Then for i 6= j, XiT Xj = 0. So the set {X1 , X2 , . . . , Xl−1 } is orthogonal. Also note that,

A(P )Xi = −pXi for all 1 ≤ i ≤ l − 1.

Which implies that the multiplicity of the eigenvalue −p is at least l − 1. Let the
multiplicity of −p be greater than l − 1. Then there exists a nonzero vector Y whose j-th
component is nonzero for some j > l and it satisfies the eigenvalue equation A(P )Y = −pY .
Then there exist real numbers c and d such that the vector cX1 + dY has the first and j-th
entries equal to 1. Now cX1 + dY is also an eigenvector corresponding to 1. Therefore,
comparing the first and j-th equations of A(cX1 + dY ) = −p(cX1 + dY ) we conclude that
p = pj . Which is a contradiction. Hence the multiplicity of −p is l − 1.

Theorem 2.7. If all pi ’s are distinct and positive, then A(P ) has n distinct eigenvalues.
Moreover, if λn < λn−1 < · · · < λ1 are eigenvalues of A(P ), then λ1 > 0 > λ2 .

Proof. Let λ be an eigenvalue of A(P ) with multiplicity > 1. By Theorem 2.4, A(P ) is
diagonalizable. Then there exist linearly independent eigenvectors X, Y ∈ Rn corresponding
to the eigenvalue λ. Since X and Y are linearly independent, there exists c, d ∈ R such that
cX + dY has two components equal to 1. Suppose the i-th and j-th components of cX + dY
are equal to 1. Since X and Y are eigenvectors corresponding to the same eigenvalue λ, we
have
A(cX + dY ) = λ(cX + dY ).
Now, comparing the i-th and j-th components, we get pi = pj . Which is a contradiction.
Therefore, all eigenvalues are distinct.
Now, since pi > 0 for all i = 1, 2, . . . , n, A(P ) is irreducible and non-negative. Hence
λ1 > 0 by Perron-Frobenius theorem ([6], Theorem 8.4.4). Again, A(P ) is similar to
1 1 1 1
D 2 A(P )D− 2 = D 2 (Jn − In )D 2 . Therefore, λ2 < 0 by Sylvester law of inertia ([6], Theorem
4.5.8). Therefore λ2 < 0 < λ1 . Which completes the proof.
Let P = {p1 , p2 , . . . , pn } be a finite sequence of real numbers. Suppose p1 = p2 = · · · =
pl = p. Let P 0 be the sequence {lp, pl+1 , . . . , pn }. We construct an (n − l + 1) × (n − l + 1)
matrix defined by
C(P 0 ) = A(P 0 ) + diag[(l − 1)p, 0, 0, . . . , 0].
Then, we have the following result:

Theorem 2.8. Let P = {p1 , p2 , . . . , pn } be a finite sequence of non-decreasing real numbers

and p1 = p2 = · · · = pl = p 6= pl+1 . If P 0 = {lp, pl+1 , . . . , pn }, then

(a) any eigenvalue of C(P 0 ) is also an eigenvalue of A(P ), and

62
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

(b) −p is not an eigenvalue of C(P 0 ).

Proof. (a) Let λ be an eigenvalue of C(P 0 ) and X = [x1 x2 · · · xn−l+1 ]T be a corresponding

eigenvector. Define Y = [y1 y2 · · · yn ]T by
(
x1 if i ≤ l,
yi =
xi−l+1 if l < i ≤ n.

Then Y is a nonzero vector in Rn and A(P )Y = λY. Therefore λ is also an eigenvalue of A(P ).

(b) Let −p be an eigenvalue of C(P 0 ) and x = [x1 x2 · · · xn−l+1 ]T an eigenvector corre-

sponding to −p. Then the nonzero vector Y = [y1 y2 · · · yn ]T defined by
(
x1 if i ≤ l,
yi =
xi−l+1 if l < i ≤ n,

satisfies A(P )Y = λY. Now since the first l components of Y are constants, XiT Y = 0 for all
1 ≤ i ≤ l −1. Which implies that, if −p is an eigenvalue of C(P 0 ), then the multiplicity of −p
as an eigenvalue of A(P ) is greater than l − 1. Which contradicts Theorem 2.6. Therefore,
−p is not an eigenvalue of C(P 0 ).
Theorem 2.6 gives us assurance that whenever l number of pi ’s are equal to a real number
p, −p is an eigenvalue of A(P ) with multiplicity l − 1. Let P = {p1 , p2 , . . . , pn } be a
non-decreasing sequence of positive real numbers. Suppose p1 , p2 , . . . , pn takes s distinct
values q1 , q2 , . . . , qs . To represent this situation, we rewrite P as {q1k1 , q2k2 , . . . , qsks }, where
k1 + k2 + · · · + ks = n. Define an s × s matrix C(P ) = [cij ] by
(
kj qj if i 6= j,
cij =
(kj − 1)qj if i = j.

Therefore, by using a similar approach to Theorem 2.8, we have the following result:

Theorem 2.9. Let P = {p1 , p2 , . . . , pn } = {q1k1 , q2k2 , . . . , qsks } be a finite sequence of nonzero
real numbers. Then the following are true:

(a) −qi is an eigenvalue of A(P ) with multiplicity ki − 1.

(b) Every eigenvalue of C(P ) is also an eigenvalue of A(P ).

(c) −qi , 1 ≤ i ≤ s is not an eigenvalue of C(P ).

An immediate consequence of Theorem 2.9 can be given as follows:

Corollary 2.10. Let P = {p1 , p2 , . . . , pn } = {q1k1 , q2k2 , . . . , qsks } be a finite sequence of nonzero
real numbers. Then
det C(P ) = (−1)s−1 (n − 1)q1 q2 · · · qs .

63
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

Proof. By Theorem 2.8, −qi is an eigenvalue of A(P ) with multiplicity ki − 1 and −qi ,
1 ≤ i ≤ s, is not an eigenvalue of C(P ). Therefore, the result follows from Theorem 2.1 as
det C(P ) equals to the product of eigenvalues of C(P ).

Theorem 2.11. Let P = {p1 , p2 , . . . , pn } = {q1k1 , q2k2 , . . . , qsks } be a finite sequence of nonzero
real numbers and let ΦC(P ) (x) = xn + c1 xn−1 + · · · + cn be the characteristic polynomial of
C(P ). Then
XX Y
cr = − ki − 1 qj ,
S i∈S j∈S

where the product runs over all subsets S of {1, 2 . . . , n} of order r.

Proof. The coefficient cr of ΦC(P ) (x) is given by the following formula

cr = (−1)r (Sum of r × r principal minors).

Let M (i1 , i2 . . . , ir ) be a principal minor obtained by taking ij -th, 1 ≤ j ≤ r rows and

k k k
corresponding columns. Then M (i1 , i2 . . . , ir ) = C(P 0 ) where P 0 = {qi1i1 , qi2i2 , . . . , qirir }.
Now

(ki1 − 1)qi1 ki2 qi2 ··· kir qir

ki1 qi1 (ki2 − 1)qi2 · · · kir qir
det C(P 0 ) = .. .. ..
. . ··· .
ki1 qi1 ki2 qi2 ··· (kir − 1)qir
ki1 − 1 ki2 ··· kir
ki1 ki2 − 1 ··· kir
= q i1 qi2 · · · qir .. .. ..
. . ··· .
ki1 ki2 ··· kir − 1
r−1
= (−1) (ki1 + ki2 + · · · + kir − 1)qi1 qi2 · · · qir .

Therefore, combining all possibilities, we get

XX Y
cr = − ki − 1 qj .
S i∈S j∈S

This completes the proof of the theorem.

Example 2.12. Consider the finite sequence P = {2, 2, 2, 3, 3, 3}. Then

 
0 2 2 3 3 3
 2 0 2 3 3 3 
 
 2 2 0 3 3 3 
A(P ) =   2 2 2 0 3 3 ,

 
 2 2 2 3 0 3 
2 2 2 3 3 0

64
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

and  
4 9
C(P ) = .
6 6
In this case regular eigenvalues of A(P ) are −2 (multiplicity 2) and −3 (multiplicity 2). The
non-regular eigenvalues are roots of the quadratic equation

x2 − 10x − 30 = 0.
√
Therefore, eigenvalues of A(P ) are −22 , −32 , 5 ± 55.

2.1 Inverse and eigenvalue bounds

In this section, we first provide the inverse of a nonsingular multipart matrix. By corollary
2.2, if pi 6= 0 for all i = 1, 2, . . . , n, then A(P ) is nonsingular. For that case, we can find
A(P )−1 explicitly.

Theorem 2.13. Let P = {p1 , p2 , . . . , pn } be a finite sequence of nonzero real numbers. Then
n−2 h1 1 1i
A(P )−1 = A(P 0 )T − diag , ,..., ,
n−1 p1 p2 pn

where P 0 = { p11 , p12 , . . . , p1n }.

Proof. Let D = diag[p1 , p2 , . . . , pn ]. Observe that, A(P ) = (Jn − In )D, where Jn is the
all-ones square matrix of order n. Then, we have
1    1  2 
Jn − (n − 1)In Jn − In = Jn − (n − 1)Jn − Jn + (n − 1)In = In
n−1 n−1
and
1    1  2 
Jn − In Jn − (n − 1)In = Jn − (n − 1)Jn − Jn + (n − 1)In = In .
n−1 n−1
 −1  
1
Therefore, Jn − In = n−1 Jn − (n − 1)In . Hence,
 −1 1  
A(P )−1 = D−1 Jn − In = D−1 (Jn − In ) − (n − 2)In
n−1
1 0 T n−2 h1 1 1i
= A(P ) − diag , ,..., .
n−1 n−1 p1 p2 pn
This completes the proof of the theorem.
Now we provide bounds for the eigenvalues for A(P ). For an n × n matrix with real
eigenvalues, we arrange its eigenvalues as follows:

λn ≤ λn−1 ≤ · · · λ2 ≤ λ1 .

65
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

Theorem 2.14. Let P = {p1 , p2 , . . . , pn } = {q1k1 , q2k2 , . . . , qsks } be a finite sequence of nonzero
real numbers. Then the negative eigenvalues of C(P ) satisfy
−qs < λs < −qs−1 < λs−1 < · · · < −q2 < λ2 < −q1 .
√ √ √ 1 1
Proof. Let D = diag[ k1 q1 , k2 q2 , . . . , ks qs ]. Then D 2 C(P )D− 2 is symmetric, and it has
the same eigenvalues as C(P ). Let B be the s×s matrix whose each entry in the i-th column
is ki qi . Then

1 1 1 1
D 2 C(P )D− 2 = D 2 BD− 2 + diag[−q1 , −q2 , . . . , −qs ].
1 1
Since the matrix D 2 BD− 2 has eigenvalues 0 (with multiplicity n − 1) and
P
pi (with
multiplicity 1). Therefore, by Weyl’s inequality [6, Theorem 4.3.1], we conclude that
−qs < λs < −qs−1 < λs−1 < · · · < −q2 < λ2 < −q1 .
Now the eigenvalues of B are n (with multiplicity 1) and 0 (with multiplicity s − 1).
Hence the required inequalities follow, as −qi is not an eigenvalue of C(P ).
By Theorem 2.14, we conclude that no two non-regular eigenvalues of a multipart matrix
are equal. Let P = {p1 , p2 , . . . , pn } = {q1k1 , q2k2 , . . . , qsks } be a finite sequence of non-decreasing
positive real numbers, and let µ1 > µ2 > · · · > µs be the non-regular eigenvalues of A(P ).
Then we can write
n
X
λ1 = µ1 = − λi
i=2
s
X s
X
= (ki − 1)qi + −µi .
i=1 i=2

Now, by Theorem 2.14, qi < −µi < qi−1 for 2 ≤ i ≤ s. Therefore,

X X
pi − pn < λ1 < pi − p1 .

In the next theorem, we provide a lower bound for the positive eigenvalue of a multipart
matrix.
Theorem 2.15. Let P = {p1 , p2 , . . . , pn } = {q1k1 , q2k2 , . . . , qsks } be a finite sequence of positive
real numbers. Then
s
1X sp
λ1 ≥ (ki − 1)qi + s (n − 1)q1 q2 · · · qs .
2 i=1 2

Proof. Since λ2 < 0 < λ1 and trace of A(P ) is 0, so λ1 = −(λ2 + λ3 + · · · + λn ). Let

λ1 = µ1 > µ2 > · · · > µs be the non-regular eigenvalues of A(P ).
n
X s
X s
X
2|λ1 | = |λi | = (ki − 1)qi + |µi |.
i=1 i=1 i=1

66
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

Now, applying A.M.≥ G.M. on |µ1 |, |µ2 |, . . . , |µs | and using Corollary 2.10, we get
s
1X sp
|λ1 | ≥ (ki − 1)qi + s (n − 1)q1 q2 · · · qs .
2 i=1 2

This completes the proof.

Remark 2.16. The bound provided in Theorem 2.14 is sharp. The equality holds if P =
{p, p, p, . . . , p} or P = {p, q}.
In the next theorem, we provide an upper bound for the largest eigenvalue (spectral
radius) of multipart matrix.
Theorem 2.17. Let P = {p1 , p2 , . . . , pn } = {q1k1 , q2k2 , . . . , qsks } be a finite sequence of nonzero
real numbers. Then
n−1
σ(A(P )) = λ1 ≤ (n − 1)(p1 p2 . . . pn ) n .
Proof. By Theorem 2.14, λ1 is the only positive eigenvalue A(P ). Therefore λ11 is the largest
eigenvalue of A(P )−1 . Then for any nonzero vector X = [x1 x2 · · · xn ]T ∈ Rn , we have
1 X T A(P )−1 X
≥ .
λ1 XT X
We take X to be the column vector with all component equal to 1. Then we have
1 1h i
≥ sum of entries of of A(P )−1
λ1 n
1 X 1
=
n(n − 1) pi
p 1 + p2 + · · · + pn
= .
n(n − 1)p1 p2 . . . pn
Applying A.M.≥G.M. on p1 , p2 , . . . , pn we get
1 1
≥ n−1 .
λ1 (n − 1)(p1 p2 . . . pn ) n
n−1
Therefore, λ1 ≤ (n − 1)(p1 p2 . . . pn ) n .
Remark 2.18. The bound provided in the above theorem is sharp. The equality holds if
P = {p, p, p, . . . , p} or P = {p, q}.

3 Applications in spectral graph theory

Now we provide some applications of the multipart matrices. We have already observed that
the quotient matrix of the adjacency matrix of a complete multipartite graph is a multipart
matrix. Here we consider two connectivity matrices of a complete multipartite graph, namely,
the adjacency matrix and the normalized adjacency matrix.

67
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

3.1 Adjacency eigenvalues of complete multipartite graphs

In this subsection, we try to analyze the spectral properties of the adjacency matrix of
complete multipartite graphs. The complete multipartite graph Kp1 ,p2 ,...,pk on n vertices has
nullity n − k. An orthogonal set of eigenvectors corresponding to 0 can be constructed as
follows:
T n
P any nonzero vector X = [x1 x2 · · · xn ] ∈ R , whose entries satisfy the
Note that
equation j∼i xj = 0, for all 1 ≤ i ≤ n, is an eigenvector corresponding to 0. Let Or denote
the r-component zero row-vector, and for i > 1 define the row-vectors Xji by

Xji = e 1 (i)T + e 2 (i)T + · · · + e j (i)T − je j+1 (i)T , for all 1 ≤ j ≤ i − 1,

where e j (i) is the j-th standard basis element of Ri . Now, for Pi ≥ 2, we define

Yi (j) = [Op1 Op2 · · · Opi−1 Xjpi Opi+1 · · · Opk ]T , for all 1 ≤ i ≤ k, 1 ≤ j ≤ pi − 1.

Then, for each pi > 1, the set {Yi (j)|1 ≤ i ≤ k, 1 ≤ j ≤ pi − 1 and pi > 1} contains
n − k orthogonal eigenvectors corresponding to 0. Therefore, a complete multipartite graph
Kp1 ,p2 ,...,pk on n vertices has at most k distinct nonzero eigenvalues. Again, the nonzero
eigenvalues of the adjacency matrix are eigenvalues of the multipart matrix A(P ), where
P = {p1 , p2 , · · · , pk }.
Let P = {p1 , p2 , . . . , pk } = {q1k1 , q2k2 , . . . , qsks } with
P P P
ki = k and ki qi = pi = n.
Then Theorem 2.6, Theorem 2.9, and Theorem 2.14 provide alternative proofs for some of
the existing results [3, 4] related to adjacency eigenvalues of Kp1 ,p2 ,...,pk .

Theorem 3.1. [4] Let P = {p1 , p2 , . . . , pk } = {q1k1 , q2k2 , . . . , qsks } be non-decreasing sequence
of natural numbers with p1 + p2 + · · · + pk = n. If λ1 is the positive adjacency eigenvalue of
Kp1 ,p2 ,...,pk , then the following are true:

(a) 0 is an eigenvalue with multiplicity n − k.

(b) −qi is an eigenvalue with multiplicity ki − 1.

(c) If µs < µs−1 < · · · < µ1 are the non-regular eigenvalues, then

−qs < µs < −qs−1 < µs−1 < · · · ≤ −q2 < µ2 ≤ −q1 .

Theorem 3.2. [3] The characteristics polynomial of Kp1 ,p2 ,··· ,pk is given by

xn−k (xk − c2 xk−2 − c3 xk−3 − c4 xk−4 − · · · − ck ),

P Q
where ci = (i − 1) S j∈S pj , 2 ≤ i ≤ k and the summation runs over all subsets S of
{1, 2 . . . , k} of order i.

Now since all non-regular eigenvalues of A(P ) are also eigenvalues of C(P ), we have the
following characteristic polynomial formula:

68
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

Theorem 3.3. Let P = {p1 , p2 , . . . , pk } = {q1k1 , q2k2 , . . . , qsks } be non-decreasing sequence of

natural numbers with p1 +p2 +· · ·+pk = n. Then the characteristics polynomial of Kp1 ,p2 ,··· ,pk
is given by
s
Y 
x n−k
(x + qj )kj −1 (xs − c1 xs−1 − c2 xs−2 − c3 xs−3 − · · · − cs ),
j=1

P P Q
where cr = S i∈S ki − 1 j∈S qj , 1 ≤ r ≤ s and the summation runs over all subsets
S of {1, 2 . . . , k} of order r.

Proof. The result follows from Theorem 3.1 and Theorem 2.11.

Example 3.4. Consider the complete multipartite graph G = K2,2,2,3,3,3,4,4,4 . The coefficients
in the Theorem 3.3 are c1 = 18, c2 = 130, c3 = 192. Consequently the characteristic
polynomial of G is

ΦA (x) = x18 (x + 2)2 (x + 3)2 (x + 4)2 (x3 − 18x2 − 130x − 192).

Therefore, the adjacency eigenvalues of K2,2,2,3,3,3,4,4,4 are −42 , −3.4884, −32 , −2.3125, −22
018 , 23.8009.

The next result gives a lower bound and an upper bound for the positive adjacency
eigenvalue of complete multipartite graphs. The result is the immediate consequences of
Theorem 2.15 and Theorem 2.17.

Theorem 3.5. Let P = {p1 , p2 , . . . , pk } = {q1k1 , q2k2 , . . . , qsks } be non-decreasing sequence of

natural numbers with p1 + p2 + · · · + pk = n. If λ1 is the positive adjacency eigenvalue of
Kp1 ,p2 ,...,pk , then
s
k−1 n 1X sp
(k − 1)(p1 p2 . . . pk ) k ≥ λ1 ≥ − qi + s (k − 1)q1 q2 · · · qs .
2 2 i=1 2

Proof. Since k1 q1 + k2 q2 + · · · + ks qs = p1 + p2 + · · · + pk = n. Hence the result follows from

Theorem 2.15 and Theorem 2.17.

Remark 3.6. The equality for the bound in Theorem 3.5 holds if G is a regular complete
multiprtite graph or a complete bipartite graph.

3.2 Normalized adjacency eigenvalues of complete multipartite

graphs
Let V = V1 ∪ V2 ∪ · · · ∪ Vk be the equitable partition for the vertex set of G = Kp1 ,p2 ,...,pk
consisting of independent subsets of G. Then the quotient matrix for A is given by

Q = D−1 Q,

69
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

where Q is the quotient matrix for A and D = diag[n − p1 , n − p2 , . . . , n − pk ]. The matrix Q

is similar to the matrix QD−1 . We observe that the matrix QD−1 is a multipart matrix for
p1
the finite sequence { n−p , p2 , . . . , n−p
1 n−p2
pk
k
}. Note that any eigenvalue of Q is also an eigenvalue
of A. Thus, as nullity of A is n − k, all nonzero eigenvalues of A is also eigenvalues of the
multipart matrix QD−1 . Now we have the following result:
Theorem 3.7. Let P = {p1 , p2 , . . . , pk } = {q1k1 , q2k2 , . . . , qsks } be non-decreasing sequence of
natural numbers with p1 +p2 +· · ·+pk = n. Then the eigenvalues of the normalized adjacency
matrix A of Kp1 ,p2 ,...,pk has the following properties:
(a) 0 is an eigenvalue with multiplicity n − k and 1 is a simple eigenvalue.
qi
(b) − is a eigenvalue of A with multiplicity ki − 1.
n − qi
(c) All non-regular eigenvalues are distinct.
(d) If 1 = µ1 > µ2 > · · · > µs are the non-regular eigenvalues of A, then
−qs −qs−1 −qs−2 −q2 −q1
> µs > > µs−1 > > ··· > µ2 > .
n − qs n − qs−1 n − qs−2 n − q2 n − q1

Proof. Since A is a row stochastic matrix, 1 is an eigenvalue of A. Note that all the nonzero
eigenvalues of A are eigenvalues of A(P), where
n p p2 pk o n q1 k1  q2 k2  q k s o
1 s
P= , ,..., = , ..., .
n − p1 n − p2 n − pk n − q1 n − q2 n − qs
Therefore, the result follows from Theorem 2.6, Theorem 2.9, Theorem 2.14, and Sylvestar
law of inertia.
Example 3.8 (Biregular complete multipartite graphs). Let Kp1 ,p2 ,...,pk be a biregular com-
plete multipartite graph. Then P = {pk1 , q k2 }. In that case, the normalized adjacency
p q
eigenvalues are 0 (with multiplicity n − k), − (with multiplicity k1 − 1), − (with
n−p n−q
(k − 1)pq
multiplicity k2 − 1), − (with multiplicity 1), and 1 (with multiplicity 1).
(n − p)(n − q)
Example 3.9 (Triregular complete multipartite graphs). Let Kp1 ,p2 ,...,pk be a triregular com-
plete multipartite graph. Then P = {pk1 , q k2 , rk3 }. In that case, the normalized adjacency
p q
eigenvalues are 0 (with multiplicity n − k), − (with multiplicity k1 − 1), − (with
n−p n−q
r
multiplicity k2 − 1), − (with multiplicity k3 − 1), 1 (with multiplicity 1), and two
n−r
negative non-regular eigenvalues are roots of the equation
 (k1 − 1)p (k2 − 1)q (k3 − 1)r  2pqr
x2 + 1 − − − x+ .
n−p n−q n−r (n − p)(n − q)(n − r)
In particular, qif k1 = k2 = k3 = 1, i.e., G = kp,q,r , then its normalized eigenvalues are
8pqr
0n−3 , 1, − 12 ± 21 1 − (n−p)(n−q)(n−r) .

70
 Mehatari/ American Journal of Combinatorics 2 (2023) 59–71

Acknowledgments
I thank the anonymous referee(s) whose comments and suggestions helped to improve this
manuscript.

References
[1] R. B. Bapat and M. Karimi, Integral complete multipartite graphs, Linear Algebra Appl.,
549 (2018) 1–11.

[2] Ş.B. Bozkurt, A.D. Güngör, I. Gutman, and A.S. Çevik, Randić matrix and Randić
energy, MATCH Commun. Math. Comput. Chem., 64 (2010) 239–250.

[3] C. Delorme, Eigenvalues of complete multipartite graphs, Discrete Math. 312 (2012)
2532–2535.

[4] F. Esser and F. Harary, On the spectrum of a complete multipartite graph, European J.
Combin., 1(3) (1980) 211–218.

[5] C. Godsil and G. Royle, Algebraic Graph Theory, Springer, 2001.

[6] R. A. Horn and C. R. Johnson, Matrix analysis, Cambridge University Press, 2012.

[7] D. Stevanović, I. Gutman, and M. U. Rehman, On spectral radius and energy of complete
multipartite graphs, Ars Mathematica Contemporanea 9 (2015) 109–113.

[8] L.G. Wang, X.L. Li, and C. Hoede, Integral complete r-partite graphs, Discrete Math.
283 (2004) 231–241.

Contact Information
Ranjit Mehatari Department of Mathematics
[email protected] National Institute of Technology Rourkela
Sundargarh, Odisha, India 769008

71