Contents lists available atScienceDirect
Linear Algebra and its Applications
j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / l a a
On the largest eigenvalues of bipartite graphs which are
nearly complete
Yi-Fan Chen
a, Hung-Lin Fu
a, In-Jae Kim
b,∗, Eryn Stehr
b, Brendon Watts
caDepartment of Applied Mathematics, National Chiao Tung University, Hsin Chu 30050, Taiwan, ROC bDepartment of Mathematics and Statistics, Minnesota State University, Mankato, MN 56001, United States cDepartment of Mathematics, University of Oklahoma, Norman, OK 73019-0315, United States
A R T I C L E I N F O A B S T R A C T Article history:
Received 10 June 2009 Accepted 4 September 2009 Available online 14 October 2009
Submitted by S. Kirkland AMS classification: 05C50 15A18 Keywords: Bipartite graph Eigenvector Largest eigenvalue
For positive integers p, q, r, s and t satisfying rtp and stq, let G(p, q;r, s;t)be the bipartite graph with partite sets{u1,. . ., up}
and{v1,. . ., vq}such that uiand vjare not adjacent if and only if
there exists a positive integer k with 1kt such that(k−1)r+
1ikr and(k−1)s+1jks. In this paper we study the
largest eigenvalues of bipartite graphs which are nearly complete. We first compute the largest eigenvalue (and all other eigenvalues) of G(p, q;r, s;t), and then list nearly complete bipartite graphs ac-cording to the magnitudes of their largest eigenvalues. These results give an affirmative answer to [1, Conjecture 1.2] when the number of edges of a bipartite graph with partite sets U and V , having |U| =p and|V| =q for pq, is pq−2. Furthermore, we refine [1, Conjecture 1.2] for the case when the number of edges is at least
pq−p+1.
© 2009 Elsevier Inc. All rights reserved.
1. Introduction and preliminary
Let G be a (simple) graph with vertex set V
(
G) = {
v1,. . .
, vn}
and edge set E(
G) = {
vi, vj|
viand vj are adjacent}
. If H is a graph with V(
H) ⊆
V(
G)
and E(
H) ⊆
E(
G)
, then H is a subgraph of G. If H/=
G, then H is a proper subgraph. If H is a proper subgraph of G, and there is no proper subgraph Hof G such that H is a proper subgraph of H, then H is a maximal proper subgraph of G.∗Corresponding author.
E-mail address: in-jae.kim@mnsu.edu (I.-J. Kim).
0024-3795/$ - see front matter © 2009 Elsevier Inc. All rights reserved. doi:10.1016/j.laa.2009.09.008
The adjacency matrix of G on n vertices is the n by n matrix A
(
G)
whose entries aijare given by aij=
1, if vi, vj
∈
E(
G);
0, otherwise.
Note that A
(
G)
is symmetric. Hence, all the eigenvalues of A(
G)
are real. The eigenvalues of A(
G)
are called eigenvalues of the graph G. We can list the eigenvalues of graph G in non-increasing orderλ
1(
G)
λ
2(
G)
· · ·
λ
n(
G).
By [4, Theorem 8.4.5], we have the following result.
Proposition 1. If H is a subgraph of G, then
λ
1(
H)
λ
1(
G).
A graph G is bipartite if its vertex set can be partitioned into two parts U and V , so called partite sets, such that every edge has one end in U and the other in V . Let G be a bipartite graph with partite sets U
= {
u1,. . .
, up}
and V= {
v1,. . .
, vq}
, and A be the(
p+
q)
by(
p+
q)
adjacency matrix of G of the form O B BT O , (1)where B
= [
bij]
is the p by q matrix such that bij=
1, if uiand vjare adjacent
;
0, otherwise.
If B is the p by q matrix with each entry equal to 1, then the bipartite graph G is a complete bipartite graph, and denoted by Kp,q. The following two results describe spectral properties of bipartite graphs (Theorem2; see [8, Theorem 8.6.9]) and the matrix product of the form BBT(Proposition3; see [4]).
Theorem 2. Let G be a bipartite graph, and A be its adjacency matrix
.
Then the eigenvalues of A are symmetric about the origin, i.
e.
, ifλ ∈ σ(
A)
, then−λ ∈ σ (
A).
Moreover,μ(
0)
is an eigenvalue of A2if and only if±√μ
are eigenvalues of A.
For each x
∈
R
n, if an n by n real symmetric matrix M satisfies xTMx 0, then M is a positive semidef-inite matrix. It is well known that each eigenvalue of a positive semidefinite matrix is a nonnegative real number (see [4]).Proposition 3. Let B be a p by q matrix
.
Then(
a)
The rank of B is equal to that of the p by p matrix BBT.
(
b)
The matrix BBTof order p is positive semidefinite.
(
c)
The number of nonzero eigenvalues of BBTis equal to the rank of BBT.
(
d)
If p q, then the q by q matrix BTB has the same eigenvalues as BBTof order p, counting multiplicity, together with additional(
q−
p)
eigenvalues equal to 0.
For positive integers p, q, r, s and t satisfying rt p and st q, we define G
(
p, q;
r, s;
t)
to be the bipartite graph with partite sets{
u1,. . .
, up}
and{
v1,. . .
, vq}
such that uiand vjare not adjacent if and only if there exists a positive integer k with 1 k t such that(
k−
1)
r+
1 i kr and(
k−
1)
s+
1 j ks. In Fig. 1 there is an edge between uiand vjif and only if they are not connected by a dotted line.In [2,3] the largest and the second largest eigenvalues of certain types of trees are obtained. This motivates our study of the largest eigenvalues of bipartite graphs which are close to a complete bipartite graph. In this paper we find all the eigenvalues of the bipartite graph G
(
p, q;
r, s;
t)
by computing eigen-values of the square of the adjacency matrix of G(
p, q;
r, s;
t)
, and list its eigenvalues according to theirFig. 1. G(5, 6;1, 2;2).
magnitudes. Using this result, we list nearly complete bipartite graphs according to the magnitudes of their largest eigenvalues. These results give an affirmative answer to [1, Conjecture 1.2] when the number of edges of a bipartite graph with partite sets U and V , having
|
U| =
p and|
V| =
q for p q, is pq−
2. Furthermore, by Theorem7, we refine [1, Conjecture 1.2] when the number of edges is at least pq−
p+
1 (see Conjecture11).2. Eigenvalues of G
(
p, q;
r, s;
t)
Let A be the adjacency matrix of G
(
p, q;
r, s;
t)
of the form (1). If we can compute eigenvalues of A2, then, by Theorem2, we can find the eigenvalues of A. Note thatA2
=
O B BT O O B BT O=
BBT O O BTB.
By Proposition3(d), it suffices to compute eigenvalues of BBT.
Let
(
a)
r×sbe the r by s matrix each of whose entries is a. When r=
s, we use(
a)
rto denote(
a)
r×r. Then B=
⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣(
0)
r×s(
1)
r×s· · ·
(
1)
r×s(
1)
r×s(
0)
r×s . .....
...
. .. . ..(
1)
r×s(
1)
rt×(q−st)(
1)
r×s· · ·
(
1)
r×s(
0)
r×s(
1)
(p−rt)×st(
1)
(p−rt)×(q−st) ⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦ , and BBT=
⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣(
q−
s)
r(
q−
2s)
r· · ·
(
q−
2s)
r(
q−
2s)
r(
q−
s)
r . .....
...
. .. . ..(
q−
2s)
r(
q−
s)
rt×(p−rt)(
q−
2s)
r· · ·
(
q−
2s)
r(
q−
s)
r(
q−
s)
(p−rt)×rt(
q)
(p−rt) ⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦.
Assume that p
>
rt and q>
st. Then it can be shown that, by elementary row and column opera-tions, B has the same rank as that of the(
t+
1)
by(
t+
1)
matrixC
=
⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣ 0 1· · ·
1 1 1 0 . ..... ...
...
. .. . .. 1...
1· · ·
1 0 1 1· · ·
1 1 ⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦.
(2)If t
=
1, then C=
0 1 1 1 ,which is nonsingular. For t 2, note that the
(
1, 1)
-block of C is(
1)
t−
It, which is nonsingular. Let Ri be the ith row of the matrix (2). By using the row operation(
t−
1)
Rt+1and then Rt+1−
Rifor each i=
1, 2,. . .
, t, we can obtain the following matrix whose rank is equal to that of C:⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣ 0 1
· · ·
1 1 1 0 . ..... ...
...
. .. . .. 1...
1· · ·
1 0 1 0· · ·
0−
1 ⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦.
This implies that the rank of C (and hence BBT) is t
+
1.If p
=
rt (resp. q=
st), then the matrix C is obtained by deleting the last row (resp. the last column) from the matrix in the form (2). By Proposition3(a) and (c), we get the following result.Proposition 4. If p
>
rt and q>
st, then BBThas t+
1 nonzero eigenvalues, and if p=
rt or q=
st, then BBThas t nonzero eigenvalues.
The p by p matrix BBTcan be rewritten as follows:
BBT
= (
q−
2s)(
1)
p+
sM, (3) where M=
⎡ ⎢ ⎢ ⎢ ⎢ ⎣(
1)
r O . ..(
1)
rt×(p−rt) O(
1)
r(
1)
(p−rt)×rt(
2)
p−rt ⎤ ⎥ ⎥ ⎥ ⎥ ⎦ p×p.
(4)In the following we find some eigenvalues and their corresponding eigenvectors of M and use them to find the eigenvalues of BBT. The vector eidenotes the vector with exactly one nonzero entry that is 1 and located in the ith position. For a square matrix M of order p and a p by 1 vector x, if Mx
=
0, then x is a nullvector of M.Proposition 5. For p
>
rt and q>
st, let BBTbe of the form(
3)
, and M be the p by p matrix of the form(
4).
Then the following hold:
(
a)
The p− (
t+
1)
nonzero vectors in the set{
erk+1−
erk+j|
k=
0, 1,. . .
, t−
1 and j=
2,. . .
, r} ∪ {
ert+1−
ej|
j=
rt+
2,. . .
, p}
(5) are linearly independent nullvectors of M.
(
b)
The p− (
t+
1)
linearly independent vectors in(
5)
are nonzero nullvectors of BBT.
(
c)
For t 2, the scalar r is an eigenvalue of M with multiplicity t−
1, and the t−
1 nonzero vectors in the set ⎧ ⎨ ⎩xk xk=
r i=1 ei−
r j=1 erk+j, k=
1,. . .
, t−
1 ⎫ ⎬ ⎭ (6)are linearly independent eigenvectors corresponding to the eigenvalue r
.
(
d)
When t 2, the(
t−
1)
nonzero vectors in(
6)
are linearly independent eigenvectors of BBT corre-sponding to the eigenvalue rs.
Proof. A direct computation proves (a) and (c). Since the nonzero vectors in
(
5)
and (6) are nullvectors of the matrix(
1)
pin (3), (b) and (d) follow.By Proposition4, when p
>
rt and q>
st, there are t+
1 nonzero eigenvalues of BBT. Hence, by Proposition5, there are two more nonzero eigenvalues of BBT to be computed. Let W1and W2 be subspaces ofR
p. We say that W1and W2are perpendicular (denoted by W1⊥
W2) provided that for any vectors w1∈
W1and w2∈
W2, w1is perpendicular to w2, i.e., w1Tw2=
0. For a subspace W ofR
p, we define W⊥as follows:W⊥
= {
z|
zTw=
0 for each w∈
W}.
Let S be a p by p matrix. We say that W is invariant under S if Sw
∈
W for each w∈
W . The following result can be found in [6, Theorems 4.2 and 4.3].Theorem 6. Let S be a p by p real symmetric matrix
.
Then the following hold:
(
a)
If E1and E2are the eigenspaces of S corresponding to distinct eigenvalues, then E1⊥
E2.
(
b)
If W is a subspace ofR
pwhich is invariant under S, then W⊥is also invariant under S.
Let W be the vector space spanned by the vectors in
(
5)
and (6). It can be verified that W⊥is spanned by z1and z2, i.e.,W⊥
=
z1, z2, where z1=
(
1)
rt×1(
0)
(p−rt)×1 and z2=
(
0)
rt×1(
1)
(p−rt)×1.
Moreover, by Theorem6, the eigenvectors of BBTcorresponding to the remaining two nonzero eigen-values are in W⊥, and
(
BBT)
z∈
W⊥for every z∈
W⊥. LetZ= {
z1, z2}
. Then the eigenvalues of the 2 by 2Z-matrix for BBT(see [7, p. 329]) are the remaining two eigenvalues of BBT(see [5, Proposition 1.5.4]). To find theZ-matrix for BBT, we compute(
BBT)
zifor each i=
1, 2:(
BBT)
z1= [
r(
q−
s) +
r(
t−
1)(
q−
2s)]
z1+ [
rt(
q−
s)]
z2,(
BBT)
z2= [(
p−
rt)(
q−
s)]
z1+ [
q(
p−
rt)]
z2.
Hence, the 2 by 2Z-matrix for BBTis
r
(
q−
s) +
r(
t−
1)(
q−
2s) (
p−
rt)(
q−
s)
rt
(
q−
s)
q(
p−
rt)
.
By computing the eigenvalues of theZ-matrix for BBT, we get the eigenvalues of BBT:
pq
−
2rst+
rs±
(
pq−
2rst+
rs)
2−
4rs(
p−
rt)(
q−
st)
2
.
(7)By Theorem2, Propositions3,4and5along with the eigenvalues of BBTin (7), we have found all the eigenvalues of G
(
p, q;
r, s;
t)
.Theorem 7. Let A be the adjacency matrix of the bipartite graph G
(
p, q;
r, s;
t).
Then the following hold:
(
a)
For p>
rt, q>
st and t 2, the eigenvalues of A areλ
1=
pq−
2rst+
rs+
(
pq−
2rst+
rs)
2−
4rs(
p−
rt)(
q−
st)
2λ
2=
√
rsλ
3=
pq−
2rst+
rs−
(
pq−
2rst+
rs)
2−
4rs(
p−
rt)(
q−
st)
2λ
4=
0− λ
3−λ
2−λ
1.
Moreover, the multiplicities of
±λ
1 and±λ
3are 1, the multiplicities of±λ
2 are t−
1, and the multiplicity ofλ
4=
0 is(
p+
q) −
2(
t+
1).
(
b)
If p=
rt or q=
st with t 2, then the eigenvalues of A areλ
1=
√
pq
−
2rst+
rsλ
2=
√
rs
λ
3=
0−λ
2−λ
1.
(
c)
If t=
1, then the eigenvalues of A areλ
1=
pq−
rs+
(
pq−
rs)
2−
4rs(
p−
r)(
q−
s)
2λ
2=
pq−
rs−
(
pq−
rs)
2−
4rs(
p−
r)(
q−
s)
2 0−λ
2−λ
1.
Proof. We here show that for nonnegative
λ
1,λ
2andλ
3in (a),λ
21λ
2 2λ
2
3when p rt, q st and t 2. Then the orders of eigenvalues according to their magnitudes in (a), (b) and (c) follow.
We first show that
λ
21λ
22. Since p rt and q rt, it follows that pq−
2rst+
rs rst2−
2rst+
rs=
rs(
t2−
2t+
1)
=
rs(
t−
1)
2.
Hence, for t 3,
λ
21λ
22. Let t=
2. Considerpq
−
3rs+
(
pq−
3rst)
2−
4rs(
p−
2r)(
q−
2s)
2
−
rs.
(8)If pq 5rs, then (8) is nonnegative and hence
λ
21λ
2 2. Suppose that pq<
5rs, i.e., 5rs−
pq>
0. We show that(
pq−
3rs)
2−
4rs(
p−
2r)(
q−
2s) − (
5rs−
pq)
2 (9)is nonnegative. By a simple calculation, it can be shown that (9) is equal to 8rs2p
+
8r2sq−
32r2s2.
Since p 2r and q 2s, we have 8rs2p
+
8r2sq−
32r2s2 16r2s2
+
16r2s2−
32r2s2=
0.
Hence,λ
21λ
22.Next, we show that
λ
22λ
23when p rt, q st and t 2. The difference
λ
2 2− λ
2 3is equal to rs+
2rst−
pq+
(
pq−
2rst+
rs)
2−
4rs(
p−
rt)(
q−
st)
2.
If rs+
2rst−
pq 0,λ
2 2λ
2 3.Suppose that rs
+
2rst−
pq<
0, i.e., pq−
2rst−
rs>
0. Then, in order to showλ
22λ
23, it suffices to show that
(
pq−
2rst+
rs)
2−
4rs(
p−
rt)(
q−
st) − (
pq−
2rst−
rs)
2 (10)is nonnegative. By a simple calculation, it can be shown that (10) is equal to 4rs2tp
+
4r2stq−
4r2s2t2−
8r2s2t.
Since p rt, q st and t 2, we have
4rs2tp
+
4r2stq−
4r2s2t2−
8r2s2t 4r2s2t2+
4r2s2t2−
4r2s2t2−
8r2s2t=
4r2s2t(
t−
2)
0
.
Hence, the result follows.
3. List of nearly complete bipartite graphs
We now list bipartite graphs, missing at most two edges from a complete bipartite graph, according to the magnitudes of their largest eigenvalues
λ
1. We denote by G(i)the bipartite graph with the ith largestλ
1among all bipartite graphs with 2n vertices.Theorem 8. For n 3, G(1)
=
Kn,n, G(2)
=
Kn−1,n+1, G(3)=
G(
n, n;
1, 1;
1)
, G(4)=
G(
n−
1, n+
1;
1, 1;
1)
, G(5)=
G(
n, n;
2, 1;
1)
, G(6)=
G(
n, n;
1, 1;
2)
, G(7)=
Kn−2,n+2, G(8)=
G(
n−
1, n+
1;
2, 1;
1)
, G(9)=
G(
n−
1, n+
1;
1, 2;
1)
and G(10)=
G(
n−
1, n+
1;
1, 1;
2).
Proof. Let H be a subgraph of Kp,qwith 1 p q and p
+
q=
2n. Note thatλ
1(
Kp,q) =
√
pq (11)and
√
pq n. Furthermore, by Proposition1,λ
1(
H)
√
pq. Hence, G(1)=
Kn,n.
By Proposition1and the fact that
√
pq with p+
q=
2n is increasing as the value of p grows from 1 to n, it is sufficient to consider G(
n, n;
1, 1;
1)
and Kn−1,n+1for G(2). By Theorem7and (11), we haveλ
1(
G(2)) =
max 1 √ 2[(
n 2−
1) + (
n4−
6n2+
8n−
3)
1/2]
1/2,(
n2−
1)
1/2 . Note that √1 2[(
n 2−
1)
+ (
n4−
6n2+
8n−
3)
1/2]
1/2√1 2[(
n 2−
1) + (
n4−
6n2+
8n−
3+
4n2−
8n+
4)
1/2]
1/2(
n2−
1)
1/2. Hence, G(2)=
Kn−1,n+1.
Similarly, in order to find G(3), it suffices to consider G
(
n, n;
1, 1;
1)
, G(
n−
1, n+
1;
1, 1;
1)
and Kn−2,n+2. We compute the largest largest eigenvalues of these three bipartite graphs by Theorem7and (11), and then use the facts, n4−
6n2+
8n−
3+
4n2−
8n+
4 n4−
8n2+
8n+
4 for n 3 and(
n−
4)
1/2=
√1 2[(
n2
−
1) + (
n4−
14n2+
49)
1/2]
1/2 in order to compare the largest eigenvalues. This givesNext, we consider the maximal proper subgraphs of G(1), G(2), G(3), and the complete bipartite graph Kn−2,n+2for G(4), using Theorem7and (11). By repeating this process, considering the maximal subgraphs of G(1),
. . .
, G(i), and a complete bipartite graph Kp,q with p+
q=
2n (which was not considered in the previous steps), we can get G(i+1)for i=
3,. . .
, 9.Example 9. The following is the list of bipartite graphs with 40 vertices according to the magnitudes of the largest eigenvalues:
Graph
λ
1 1. K20,20 20 2. K19,20 19.97498436 3. G(
20, 20;
1, 1;
1)
19.95227248 4. G(
19, 21;
1, 1;
1)
19.92720282 5. G(
20, 20;
2, 1;
1)
19.90663008 6. G(
20, 20;
1, 1;
2)
19.90432602 7. K18,20 19.89974874 8. G(
19, 21;
2, 1;
1)
19.88164226 9. G(
19, 21;
1, 2;
1)
19.88138664 10. G(
19, 21;
1, 1;
2)
19.87920160The computation of
λ
1 can be done by the open source mathematical software SAGE (seehttp://www.sagemath.org). The following is a SAGE code for
λ
1(
G(
20, 20;
2, 1;
1))
:p
=
20;
q=
20;
r=
2;
s=
1;
t=
1 def a(
i, j)
: for k in[
1..
t]
: if (i in[
r∗ (
k−
1) +
1..
r∗
k]
) and (j in[
s∗ (
k−
1) +
1..
s∗
k]
): return 0 else: return 1B
=
matrix([[
a(
i, j)
for j in[
1..
q]]
for i in[
1..
p]])
E= (
B∗
B.
transpose()).
eigenvalues()
print sqrt
(
E[
p−
1])
From Theorem8it follows thatλ
1(
G(
p, q;
2, 1;
1))
λ
1(
G(
p, q;
1, 2;
1))
λ
1(
G(
p, q;
1, 1;
2))
for
(
p, q) ∈ {(
n, n)
,(
n−
1, n+
1)}
. This can be generalized to the case for any positive integers p, q with 2 p q.Proposition 10. Let p and q be positive integers
.
If 2 p q, thenλ
1(
G(
p, q;
2, 1;
1))
λ
1(
G(
p, q;
1, 2;
1))
λ
1(
G(
p, q;
1, 1;
2)).
Proof. This can be shown by first computing the eigenvalues, using Theorem7, and then using a direct comparison.
Proposition10gives an affirmative answer to [1, Conjecture 1.2] when the number of edges of a bipartite graph with partite sets U and V , having
|
U| =
p and|
V| =
q for p q, is pq−
2. By TheoremConjecture 11. For positive integers p, q and k satisfying p q and k
<
p, let G be a bipartite graph with partite sets U and V , having|
U| =
p and|
V| =
q, and|
E(
G)| =
pq−
k. Thenλ
1(
G)
λ
1(
G(
p, q;
k, 1;
1)) =
pq−
k+
p2q2−
6pqk+
4pk+
4qk2−
3k2 2.
AcknowledgementsThe work of Yi-Fan Chen and Hung-Lin Fu was supported in part by NSC 94-2115-M-009-017. The work of In-Jae Kim was supported in part by the Presidential Teaching Scholar Fellowship from the Minnesota State University, Mankato, and the work of Brendon Watts was done as a part of SAMER activity while he was a PSEO student at the Minnesota State University, Mankato.
References
[1] A. Bhattacharya, S. Friedland, U.N. Peled, On the first eigenvalue of bipartite graphs, Electron. J. Combin. 15 (2008) #R144. [2] Xiang En Chen, On the largest eigenvalues of trees, Discrete Math. 285 (2004) 47–55.
[3] M. Hofmeister, On the two largest eigenvalues of trees, Linear Algebra Appl. 260 (1997) 43–59. [4] R. Horn, C.R. Johnson, Matrix Analysis, Cambridge Univeristy Press, 1985.
[5] I. Gohberg, P. Lancaster, L. Rodman, Invariant Subspaces of Matrices with Applications, SIAM, 2006. [6] S. Lang, Linear Algebra, Springer, 2001.
[7] D.C. Lay, Linear Algebra and its Applications, Addison Wesley, 2006. [8] D.B. West, Introduction to Graph Theory, Prentice-Hall, 1996.